ffs_softdep.c revision 207741
1/*- 2 * Copyright 1998, 2000 Marshall Kirk McKusick. 3 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org> 4 * All rights reserved. 5 * 6 * The soft updates code is derived from the appendix of a University 7 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 8 * "Soft Updates: A Solution to the Metadata Update Problem in File 9 * Systems", CSE-TR-254-95, August 1995). 10 * 11 * Further information about soft updates can be obtained from: 12 * 13 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 14 * 1614 Oxford Street mckusick@mckusick.com 15 * Berkeley, CA 94709-1608 +1-510-843-9542 16 * USA 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions 20 * are met: 21 * 22 * 1. Redistributions of source code must retain the above copyright 23 * notice, this list of conditions and the following disclaimer. 24 * 2. Redistributions in binary form must reproduce the above copyright 25 * notice, this list of conditions and the following disclaimer in the 26 * documentation and/or other materials provided with the distribution. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 29 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 30 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 31 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, 32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 34 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 35 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR 36 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE 37 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 * 39 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 40 */ 41 42#include <sys/cdefs.h> 43__FBSDID("$FreeBSD: head/sys/ufs/ffs/ffs_softdep.c 207741 2010-05-07 08:20:56Z jeff $"); 44 45#include "opt_ffs.h" 46#include "opt_ddb.h" 47 48/* 49 * For now we want the safety net that the DEBUG flag provides. 50 */ 51#ifndef DEBUG 52#define DEBUG 53#endif 54#define SUJ_DEBUG 55 56#include <sys/param.h> 57#include <sys/kernel.h> 58#include <sys/systm.h> 59#include <sys/bio.h> 60#include <sys/buf.h> 61#include <sys/kdb.h> 62#include <sys/kthread.h> 63#include <sys/lock.h> 64#include <sys/malloc.h> 65#include <sys/mount.h> 66#include <sys/mutex.h> 67#include <sys/namei.h> 68#include <sys/proc.h> 69#include <sys/stat.h> 70#include <sys/sysctl.h> 71#include <sys/syslog.h> 72#include <sys/vnode.h> 73#include <sys/conf.h> 74#include <ufs/ufs/dir.h> 75#include <ufs/ufs/extattr.h> 76#include <ufs/ufs/quota.h> 77#include <ufs/ufs/inode.h> 78#include <ufs/ufs/ufsmount.h> 79#include <ufs/ffs/fs.h> 80#include <ufs/ffs/softdep.h> 81#include <ufs/ffs/ffs_extern.h> 82#include <ufs/ufs/ufs_extern.h> 83 84#include <vm/vm.h> 85 86#include <ddb/ddb.h> 87 88#ifndef SOFTUPDATES 89 90int 91softdep_flushfiles(oldmnt, flags, td) 92 struct mount *oldmnt; 93 int flags; 94 struct thread *td; 95{ 96 97 panic("softdep_flushfiles called"); 98} 99 100int 101softdep_mount(devvp, mp, fs, cred) 102 struct vnode *devvp; 103 struct mount *mp; 104 struct fs *fs; 105 struct ucred *cred; 106{ 107 108 return (0); 109} 110 111void 112softdep_initialize() 113{ 114 115 return; 116} 117 118void 119softdep_uninitialize() 120{ 121 122 return; 123} 124 125void 126softdep_unmount(mp) 127 struct mount *mp; 128{ 129 130} 131 132void 133softdep_setup_sbupdate(ump, fs, bp) 134 struct ufsmount *ump; 135 struct fs *fs; 136 struct buf *bp; 137{ 138} 139 140void 141softdep_setup_inomapdep(bp, ip, newinum) 142 struct buf *bp; 143 struct inode *ip; 144 ino_t newinum; 145{ 146 147 panic("softdep_setup_inomapdep called"); 148} 149 150void 151softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 152 struct buf *bp; 153 struct mount *mp; 154 ufs2_daddr_t newblkno; 155 int frags; 156 int oldfrags; 157{ 158 159 panic("softdep_setup_blkmapdep called"); 160} 161 162void 163softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 164 struct inode *ip; 165 ufs_lbn_t lbn; 166 ufs2_daddr_t newblkno; 167 ufs2_daddr_t oldblkno; 168 long newsize; 169 long oldsize; 170 struct buf *bp; 171{ 172 173 panic("softdep_setup_allocdirect called"); 174} 175 176void 177softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 178 struct inode *ip; 179 ufs_lbn_t lbn; 180 ufs2_daddr_t newblkno; 181 ufs2_daddr_t oldblkno; 182 long newsize; 183 long oldsize; 184 struct buf *bp; 185{ 186 187 panic("softdep_setup_allocext called"); 188} 189 190void 191softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 192 struct inode *ip; 193 ufs_lbn_t lbn; 194 struct buf *bp; 195 int ptrno; 196 ufs2_daddr_t newblkno; 197 ufs2_daddr_t oldblkno; 198 struct buf *nbp; 199{ 200 201 panic("softdep_setup_allocindir_page called"); 202} 203 204void 205softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 206 struct buf *nbp; 207 struct inode *ip; 208 struct buf *bp; 209 int ptrno; 210 ufs2_daddr_t newblkno; 211{ 212 213 panic("softdep_setup_allocindir_meta called"); 214} 215 216void 217softdep_setup_freeblocks(ip, length, flags) 218 struct inode *ip; 219 off_t length; 220 int flags; 221{ 222 223 panic("softdep_setup_freeblocks called"); 224} 225 226void 227softdep_freefile(pvp, ino, mode) 228 struct vnode *pvp; 229 ino_t ino; 230 int mode; 231{ 232 233 panic("softdep_freefile called"); 234} 235 236int 237softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 238 struct buf *bp; 239 struct inode *dp; 240 off_t diroffset; 241 ino_t newinum; 242 struct buf *newdirbp; 243 int isnewblk; 244{ 245 246 panic("softdep_setup_directory_add called"); 247} 248 249void 250softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 251 struct buf *bp; 252 struct inode *dp; 253 caddr_t base; 254 caddr_t oldloc; 255 caddr_t newloc; 256 int entrysize; 257{ 258 259 panic("softdep_change_directoryentry_offset called"); 260} 261 262void 263softdep_setup_remove(bp, dp, ip, isrmdir) 264 struct buf *bp; 265 struct inode *dp; 266 struct inode *ip; 267 int isrmdir; 268{ 269 270 panic("softdep_setup_remove called"); 271} 272 273void 274softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 275 struct buf *bp; 276 struct inode *dp; 277 struct inode *ip; 278 ino_t newinum; 279 int isrmdir; 280{ 281 282 panic("softdep_setup_directory_change called"); 283} 284 285void * 286softdep_setup_trunc(vp, length, flags) 287 struct vnode *vp; 288 off_t length; 289 int flags; 290{ 291 292 panic("%s called", __FUNCTION__); 293 294 return (NULL); 295} 296 297int 298softdep_complete_trunc(vp, cookie) 299 struct vnode *vp; 300 void *cookie; 301{ 302 303 panic("%s called", __FUNCTION__); 304 305 return (0); 306} 307 308void 309softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 310 struct mount *mp; 311 struct buf *bp; 312 ufs2_daddr_t blkno; 313 int frags; 314 struct workhead *wkhd; 315{ 316 317 panic("%s called", __FUNCTION__); 318} 319 320void 321softdep_setup_inofree(mp, bp, ino, wkhd) 322 struct mount *mp; 323 struct buf *bp; 324 ino_t ino; 325 struct workhead *wkhd; 326{ 327 328 panic("%s called", __FUNCTION__); 329} 330 331void 332softdep_setup_unlink(dp, ip) 333 struct inode *dp; 334 struct inode *ip; 335{ 336 337 panic("%s called", __FUNCTION__); 338} 339 340void 341softdep_setup_link(dp, ip) 342 struct inode *dp; 343 struct inode *ip; 344{ 345 346 panic("%s called", __FUNCTION__); 347} 348 349void 350softdep_revert_link(dp, ip) 351 struct inode *dp; 352 struct inode *ip; 353{ 354 355 panic("%s called", __FUNCTION__); 356} 357 358void 359softdep_setup_rmdir(dp, ip) 360 struct inode *dp; 361 struct inode *ip; 362{ 363 364 panic("%s called", __FUNCTION__); 365} 366 367void 368softdep_revert_rmdir(dp, ip) 369 struct inode *dp; 370 struct inode *ip; 371{ 372 373 panic("%s called", __FUNCTION__); 374} 375 376void 377softdep_setup_create(dp, ip) 378 struct inode *dp; 379 struct inode *ip; 380{ 381 382 panic("%s called", __FUNCTION__); 383} 384 385void 386softdep_revert_create(dp, ip) 387 struct inode *dp; 388 struct inode *ip; 389{ 390 391 panic("%s called", __FUNCTION__); 392} 393 394void 395softdep_setup_mkdir(dp, ip) 396 struct inode *dp; 397 struct inode *ip; 398{ 399 400 panic("%s called", __FUNCTION__); 401} 402 403void 404softdep_revert_mkdir(dp, ip) 405 struct inode *dp; 406 struct inode *ip; 407{ 408 409 panic("%s called", __FUNCTION__); 410} 411 412void 413softdep_setup_dotdot_link(dp, ip) 414 struct inode *dp; 415 struct inode *ip; 416{ 417 418 panic("%s called", __FUNCTION__); 419} 420 421int 422softdep_prealloc(vp, waitok) 423 struct vnode *vp; 424 int waitok; 425{ 426 427 panic("%s called", __FUNCTION__); 428 429 return (0); 430} 431 432int 433softdep_journal_lookup(mp, vpp) 434 struct mount *mp; 435 struct vnode **vpp; 436{ 437 438 return (ENOENT); 439} 440 441void 442softdep_change_linkcnt(ip) 443 struct inode *ip; 444{ 445 446 panic("softdep_change_linkcnt called"); 447} 448 449void 450softdep_load_inodeblock(ip) 451 struct inode *ip; 452{ 453 454 panic("softdep_load_inodeblock called"); 455} 456 457void 458softdep_update_inodeblock(ip, bp, waitfor) 459 struct inode *ip; 460 struct buf *bp; 461 int waitfor; 462{ 463 464 panic("softdep_update_inodeblock called"); 465} 466 467int 468softdep_fsync(vp) 469 struct vnode *vp; /* the "in_core" copy of the inode */ 470{ 471 472 return (0); 473} 474 475void 476softdep_fsync_mountdev(vp) 477 struct vnode *vp; 478{ 479 480 return; 481} 482 483int 484softdep_flushworklist(oldmnt, countp, td) 485 struct mount *oldmnt; 486 int *countp; 487 struct thread *td; 488{ 489 490 *countp = 0; 491 return (0); 492} 493 494int 495softdep_sync_metadata(struct vnode *vp) 496{ 497 498 return (0); 499} 500 501int 502softdep_slowdown(vp) 503 struct vnode *vp; 504{ 505 506 panic("softdep_slowdown called"); 507} 508 509void 510softdep_releasefile(ip) 511 struct inode *ip; /* inode with the zero effective link count */ 512{ 513 514 panic("softdep_releasefile called"); 515} 516 517int 518softdep_request_cleanup(fs, vp) 519 struct fs *fs; 520 struct vnode *vp; 521{ 522 523 return (0); 524} 525 526int 527softdep_check_suspend(struct mount *mp, 528 struct vnode *devvp, 529 int softdep_deps, 530 int softdep_accdeps, 531 int secondary_writes, 532 int secondary_accwrites) 533{ 534 struct bufobj *bo; 535 int error; 536 537 (void) softdep_deps, 538 (void) softdep_accdeps; 539 540 bo = &devvp->v_bufobj; 541 ASSERT_BO_LOCKED(bo); 542 543 MNT_ILOCK(mp); 544 while (mp->mnt_secondary_writes != 0) { 545 BO_UNLOCK(bo); 546 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 547 (PUSER - 1) | PDROP, "secwr", 0); 548 BO_LOCK(bo); 549 MNT_ILOCK(mp); 550 } 551 552 /* 553 * Reasons for needing more work before suspend: 554 * - Dirty buffers on devvp. 555 * - Secondary writes occurred after start of vnode sync loop 556 */ 557 error = 0; 558 if (bo->bo_numoutput > 0 || 559 bo->bo_dirty.bv_cnt > 0 || 560 secondary_writes != 0 || 561 mp->mnt_secondary_writes != 0 || 562 secondary_accwrites != mp->mnt_secondary_accwrites) 563 error = EAGAIN; 564 BO_UNLOCK(bo); 565 return (error); 566} 567 568void 569softdep_get_depcounts(struct mount *mp, 570 int *softdepactivep, 571 int *softdepactiveaccp) 572{ 573 (void) mp; 574 *softdepactivep = 0; 575 *softdepactiveaccp = 0; 576} 577 578#else 579/* 580 * These definitions need to be adapted to the system to which 581 * this file is being ported. 582 */ 583 584#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE) 585 586#define D_PAGEDEP 0 587#define D_INODEDEP 1 588#define D_BMSAFEMAP 2 589#define D_NEWBLK 3 590#define D_ALLOCDIRECT 4 591#define D_INDIRDEP 5 592#define D_ALLOCINDIR 6 593#define D_FREEFRAG 7 594#define D_FREEBLKS 8 595#define D_FREEFILE 9 596#define D_DIRADD 10 597#define D_MKDIR 11 598#define D_DIRREM 12 599#define D_NEWDIRBLK 13 600#define D_FREEWORK 14 601#define D_FREEDEP 15 602#define D_JADDREF 16 603#define D_JREMREF 17 604#define D_JMVREF 18 605#define D_JNEWBLK 19 606#define D_JFREEBLK 20 607#define D_JFREEFRAG 21 608#define D_JSEG 22 609#define D_JSEGDEP 23 610#define D_SBDEP 24 611#define D_JTRUNC 25 612#define D_LAST D_JTRUNC 613 614unsigned long dep_current[D_LAST + 1]; 615unsigned long dep_total[D_LAST + 1]; 616 617 618SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, "soft updates stats"); 619SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 620 "total dependencies allocated"); 621SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 622 "current dependencies allocated"); 623 624#define SOFTDEP_TYPE(type, str, long) \ 625 static MALLOC_DEFINE(M_ ## type, #str, long); \ 626 SYSCTL_LONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 627 &dep_total[D_ ## type], 0, ""); \ 628 SYSCTL_LONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 629 &dep_current[D_ ## type], 0, ""); 630 631SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 632SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 633SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 634 "Block or frag allocated from cyl group map"); 635SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 636SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 637SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 638SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 639SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 640SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 641SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 642SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 643SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 644SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 645SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 646SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 647SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 648SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 649SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 650SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 651SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 652SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 653SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 654SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 655SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 656SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 657SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 658 659static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 660static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 661 662/* 663 * translate from workitem type to memory type 664 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 665 */ 666static struct malloc_type *memtype[] = { 667 M_PAGEDEP, 668 M_INODEDEP, 669 M_BMSAFEMAP, 670 M_NEWBLK, 671 M_ALLOCDIRECT, 672 M_INDIRDEP, 673 M_ALLOCINDIR, 674 M_FREEFRAG, 675 M_FREEBLKS, 676 M_FREEFILE, 677 M_DIRADD, 678 M_MKDIR, 679 M_DIRREM, 680 M_NEWDIRBLK, 681 M_FREEWORK, 682 M_FREEDEP, 683 M_JADDREF, 684 M_JREMREF, 685 M_JMVREF, 686 M_JNEWBLK, 687 M_JFREEBLK, 688 M_JFREEFRAG, 689 M_JSEG, 690 M_JSEGDEP, 691 M_SBDEP, 692 M_JTRUNC 693}; 694 695#define DtoM(type) (memtype[type]) 696 697/* 698 * Names of malloc types. 699 */ 700#define TYPENAME(type) \ 701 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 702/* 703 * End system adaptation definitions. 704 */ 705 706#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 707#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 708 709/* 710 * Forward declarations. 711 */ 712struct inodedep_hashhead; 713struct newblk_hashhead; 714struct pagedep_hashhead; 715struct bmsafemap_hashhead; 716 717/* 718 * Internal function prototypes. 719 */ 720static void softdep_error(char *, int); 721static void drain_output(struct vnode *); 722static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 723static void clear_remove(struct thread *); 724static void clear_inodedeps(struct thread *); 725static void unlinked_inodedep(struct mount *, struct inodedep *); 726static void clear_unlinked_inodedep(struct inodedep *); 727static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 728static int flush_pagedep_deps(struct vnode *, struct mount *, 729 struct diraddhd *); 730static void free_pagedep(struct pagedep *); 731static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 732static int flush_inodedep_deps(struct mount *, ino_t); 733static int flush_deplist(struct allocdirectlst *, int, int *); 734static int handle_written_filepage(struct pagedep *, struct buf *); 735static int handle_written_sbdep(struct sbdep *, struct buf *); 736static void initiate_write_sbdep(struct sbdep *); 737static void diradd_inode_written(struct diradd *, struct inodedep *); 738static int handle_written_indirdep(struct indirdep *, struct buf *, 739 struct buf**); 740static int handle_written_inodeblock(struct inodedep *, struct buf *); 741static int handle_written_bmsafemap(struct bmsafemap *, struct buf *); 742static void handle_written_jaddref(struct jaddref *); 743static void handle_written_jremref(struct jremref *); 744static void handle_written_jseg(struct jseg *, struct buf *); 745static void handle_written_jnewblk(struct jnewblk *); 746static void handle_written_jfreeblk(struct jfreeblk *); 747static void handle_written_jfreefrag(struct jfreefrag *); 748static void complete_jseg(struct jseg *); 749static void jseg_write(struct fs *, struct jblocks *, struct jseg *, 750 uint8_t *); 751static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 752static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 753static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 754static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 755static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 756static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 757static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 758static inline void inoref_write(struct inoref *, struct jseg *, 759 struct jrefrec *); 760static void handle_allocdirect_partdone(struct allocdirect *, 761 struct workhead *); 762static void cancel_newblk(struct newblk *, struct workhead *); 763static void indirdep_complete(struct indirdep *); 764static void handle_allocindir_partdone(struct allocindir *); 765static void initiate_write_filepage(struct pagedep *, struct buf *); 766static void initiate_write_indirdep(struct indirdep*, struct buf *); 767static void handle_written_mkdir(struct mkdir *, int); 768static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 769static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 770static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 771static void handle_workitem_freefile(struct freefile *); 772static void handle_workitem_remove(struct dirrem *, struct vnode *); 773static struct dirrem *newdirrem(struct buf *, struct inode *, 774 struct inode *, int, struct dirrem **); 775static void cancel_indirdep(struct indirdep *, struct buf *, struct inodedep *, 776 struct freeblks *); 777static void free_indirdep(struct indirdep *); 778static void free_diradd(struct diradd *, struct workhead *); 779static void merge_diradd(struct inodedep *, struct diradd *); 780static void complete_diradd(struct diradd *); 781static struct diradd *diradd_lookup(struct pagedep *, int); 782static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 783 struct jremref *); 784static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 785 struct jremref *); 786static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 787 struct jremref *, struct jremref *); 788static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 789 struct jremref *); 790static void cancel_allocindir(struct allocindir *, struct inodedep *, 791 struct freeblks *); 792static void complete_mkdir(struct mkdir *); 793static void free_newdirblk(struct newdirblk *); 794static void free_jremref(struct jremref *); 795static void free_jaddref(struct jaddref *); 796static void free_jsegdep(struct jsegdep *); 797static void free_jseg(struct jseg *); 798static void free_jnewblk(struct jnewblk *); 799static void free_jfreeblk(struct jfreeblk *); 800static void free_jfreefrag(struct jfreefrag *); 801static void free_freedep(struct freedep *); 802static void journal_jremref(struct dirrem *, struct jremref *, 803 struct inodedep *); 804static void cancel_jnewblk(struct jnewblk *, struct workhead *); 805static int cancel_jaddref(struct jaddref *, struct inodedep *, 806 struct workhead *); 807static void cancel_jfreefrag(struct jfreefrag *); 808static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 809static int deallocate_dependencies(struct buf *, struct inodedep *, 810 struct freeblks *); 811static void free_newblk(struct newblk *); 812static void cancel_allocdirect(struct allocdirectlst *, 813 struct allocdirect *, struct freeblks *, int); 814static int check_inode_unwritten(struct inodedep *); 815static int free_inodedep(struct inodedep *); 816static void freework_freeblock(struct freework *); 817static void handle_workitem_freeblocks(struct freeblks *, int); 818static void handle_complete_freeblocks(struct freeblks *); 819static void handle_workitem_indirblk(struct freework *); 820static void handle_written_freework(struct freework *); 821static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 822static void setup_allocindir_phase2(struct buf *, struct inode *, 823 struct inodedep *, struct allocindir *, ufs_lbn_t); 824static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 825 ufs2_daddr_t, ufs_lbn_t); 826static void handle_workitem_freefrag(struct freefrag *); 827static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 828 ufs_lbn_t); 829static void allocdirect_merge(struct allocdirectlst *, 830 struct allocdirect *, struct allocdirect *); 831static struct freefrag *allocindir_merge(struct allocindir *, 832 struct allocindir *); 833static int bmsafemap_find(struct bmsafemap_hashhead *, struct mount *, int, 834 struct bmsafemap **); 835static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 836 int cg); 837static int newblk_find(struct newblk_hashhead *, struct mount *, ufs2_daddr_t, 838 int, struct newblk **); 839static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 840static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 841 struct inodedep **); 842static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 843static int pagedep_lookup(struct mount *, ino_t, ufs_lbn_t, int, 844 struct pagedep **); 845static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 846 struct mount *mp, int, struct pagedep **); 847static void pause_timer(void *); 848static int request_cleanup(struct mount *, int); 849static int process_worklist_item(struct mount *, int); 850static void process_removes(struct vnode *); 851static void jwork_move(struct workhead *, struct workhead *); 852static void add_to_worklist(struct worklist *, int); 853static void remove_from_worklist(struct worklist *); 854static void softdep_flush(void); 855static int softdep_speedup(void); 856static void worklist_speedup(void); 857static int journal_mount(struct mount *, struct fs *, struct ucred *); 858static void journal_unmount(struct mount *); 859static int journal_space(struct ufsmount *, int); 860static void journal_suspend(struct ufsmount *); 861static void softdep_prelink(struct vnode *, struct vnode *); 862static void add_to_journal(struct worklist *); 863static void remove_from_journal(struct worklist *); 864static void softdep_process_journal(struct mount *, int); 865static struct jremref *newjremref(struct dirrem *, struct inode *, 866 struct inode *ip, off_t, nlink_t); 867static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 868 uint16_t); 869static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 870 uint16_t); 871static inline struct jsegdep *inoref_jseg(struct inoref *); 872static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 873static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 874 ufs2_daddr_t, int); 875static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 876 ufs2_daddr_t, long, ufs_lbn_t); 877static struct freework *newfreework(struct freeblks *, struct freework *, 878 ufs_lbn_t, ufs2_daddr_t, int, int); 879static void jwait(struct worklist *wk); 880static struct inodedep *inodedep_lookup_ip(struct inode *); 881static int bmsafemap_rollbacks(struct bmsafemap *); 882static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 883static void handle_jwork(struct workhead *); 884static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 885 struct mkdir **); 886static struct jblocks *jblocks_create(void); 887static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 888static void jblocks_free(struct jblocks *, struct mount *, int); 889static void jblocks_destroy(struct jblocks *); 890static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 891 892/* 893 * Exported softdep operations. 894 */ 895static void softdep_disk_io_initiation(struct buf *); 896static void softdep_disk_write_complete(struct buf *); 897static void softdep_deallocate_dependencies(struct buf *); 898static int softdep_count_dependencies(struct buf *bp, int); 899 900static struct mtx lk; 901MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 902 903#define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) 904#define ACQUIRE_LOCK(lk) mtx_lock(lk) 905#define FREE_LOCK(lk) mtx_unlock(lk) 906 907#define BUF_AREC(bp) ((bp)->b_lock.lock_object.lo_flags |= LO_RECURSABLE) 908#define BUF_NOREC(bp) ((bp)->b_lock.lock_object.lo_flags &= ~LO_RECURSABLE) 909 910/* 911 * Worklist queue management. 912 * These routines require that the lock be held. 913 */ 914#ifndef /* NOT */ DEBUG 915#define WORKLIST_INSERT(head, item) do { \ 916 (item)->wk_state |= ONWORKLIST; \ 917 LIST_INSERT_HEAD(head, item, wk_list); \ 918} while (0) 919#define WORKLIST_REMOVE(item) do { \ 920 (item)->wk_state &= ~ONWORKLIST; \ 921 LIST_REMOVE(item, wk_list); \ 922} while (0) 923#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 924#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 925 926#else /* DEBUG */ 927static void worklist_insert(struct workhead *, struct worklist *, int); 928static void worklist_remove(struct worklist *, int); 929 930#define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 931#define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 932#define WORKLIST_REMOVE(item) worklist_remove(item, 1) 933#define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 934 935static void 936worklist_insert(head, item, locked) 937 struct workhead *head; 938 struct worklist *item; 939 int locked; 940{ 941 942 if (locked) 943 mtx_assert(&lk, MA_OWNED); 944 if (item->wk_state & ONWORKLIST) 945 panic("worklist_insert: %p %s(0x%X) already on list", 946 item, TYPENAME(item->wk_type), item->wk_state); 947 item->wk_state |= ONWORKLIST; 948 LIST_INSERT_HEAD(head, item, wk_list); 949} 950 951static void 952worklist_remove(item, locked) 953 struct worklist *item; 954 int locked; 955{ 956 957 if (locked) 958 mtx_assert(&lk, MA_OWNED); 959 if ((item->wk_state & ONWORKLIST) == 0) 960 panic("worklist_remove: %p %s(0x%X) not on list", 961 item, TYPENAME(item->wk_type), item->wk_state); 962 item->wk_state &= ~ONWORKLIST; 963 LIST_REMOVE(item, wk_list); 964} 965#endif /* DEBUG */ 966 967/* 968 * Merge two jsegdeps keeping only the oldest one as newer references 969 * can't be discarded until after older references. 970 */ 971static inline struct jsegdep * 972jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 973{ 974 struct jsegdep *swp; 975 976 if (two == NULL) 977 return (one); 978 979 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 980 swp = one; 981 one = two; 982 two = swp; 983 } 984 WORKLIST_REMOVE(&two->jd_list); 985 free_jsegdep(two); 986 987 return (one); 988} 989 990/* 991 * If two freedeps are compatible free one to reduce list size. 992 */ 993static inline struct freedep * 994freedep_merge(struct freedep *one, struct freedep *two) 995{ 996 if (two == NULL) 997 return (one); 998 999 if (one->fd_freework == two->fd_freework) { 1000 WORKLIST_REMOVE(&two->fd_list); 1001 free_freedep(two); 1002 } 1003 return (one); 1004} 1005 1006/* 1007 * Move journal work from one list to another. Duplicate freedeps and 1008 * jsegdeps are coalesced to keep the lists as small as possible. 1009 */ 1010static void 1011jwork_move(dst, src) 1012 struct workhead *dst; 1013 struct workhead *src; 1014{ 1015 struct freedep *freedep; 1016 struct jsegdep *jsegdep; 1017 struct worklist *wkn; 1018 struct worklist *wk; 1019 1020 KASSERT(dst != src, 1021 ("jwork_move: dst == src")); 1022 freedep = NULL; 1023 jsegdep = NULL; 1024 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1025 if (wk->wk_type == D_JSEGDEP) 1026 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1027 if (wk->wk_type == D_FREEDEP) 1028 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1029 } 1030 1031 mtx_assert(&lk, MA_OWNED); 1032 while ((wk = LIST_FIRST(src)) != NULL) { 1033 WORKLIST_REMOVE(wk); 1034 WORKLIST_INSERT(dst, wk); 1035 if (wk->wk_type == D_JSEGDEP) { 1036 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1037 continue; 1038 } 1039 if (wk->wk_type == D_FREEDEP) 1040 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1041 } 1042} 1043 1044/* 1045 * Routines for tracking and managing workitems. 1046 */ 1047static void workitem_free(struct worklist *, int); 1048static void workitem_alloc(struct worklist *, int, struct mount *); 1049 1050#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 1051 1052static void 1053workitem_free(item, type) 1054 struct worklist *item; 1055 int type; 1056{ 1057 struct ufsmount *ump; 1058 mtx_assert(&lk, MA_OWNED); 1059 1060#ifdef DEBUG 1061 if (item->wk_state & ONWORKLIST) 1062 panic("workitem_free: %s(0x%X) still on list", 1063 TYPENAME(item->wk_type), item->wk_state); 1064 if (item->wk_type != type) 1065 panic("workitem_free: type mismatch %s != %s", 1066 TYPENAME(item->wk_type), TYPENAME(type)); 1067#endif 1068 ump = VFSTOUFS(item->wk_mp); 1069 if (--ump->softdep_deps == 0 && ump->softdep_req) 1070 wakeup(&ump->softdep_deps); 1071 dep_current[type]--; 1072 free(item, DtoM(type)); 1073} 1074 1075static void 1076workitem_alloc(item, type, mp) 1077 struct worklist *item; 1078 int type; 1079 struct mount *mp; 1080{ 1081 item->wk_type = type; 1082 item->wk_mp = mp; 1083 item->wk_state = 0; 1084 ACQUIRE_LOCK(&lk); 1085 dep_current[type]++; 1086 dep_total[type]++; 1087 VFSTOUFS(mp)->softdep_deps++; 1088 VFSTOUFS(mp)->softdep_accdeps++; 1089 FREE_LOCK(&lk); 1090} 1091 1092/* 1093 * Workitem queue management 1094 */ 1095static int max_softdeps; /* maximum number of structs before slowdown */ 1096static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 1097static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1098static int proc_waiting; /* tracks whether we have a timeout posted */ 1099static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1100static struct callout softdep_callout; 1101static int req_pending; 1102static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1103#define FLUSH_INODES 1 1104static int req_clear_remove; /* syncer process flush some freeblks */ 1105#define FLUSH_REMOVE 2 1106#define FLUSH_REMOVE_WAIT 3 1107static long num_freeblkdep; /* number of freeblks workitems allocated */ 1108 1109/* 1110 * runtime statistics 1111 */ 1112static int stat_worklist_push; /* number of worklist cleanups */ 1113static int stat_blk_limit_push; /* number of times block limit neared */ 1114static int stat_ino_limit_push; /* number of times inode limit neared */ 1115static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1116static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1117static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1118static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1119static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1120static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1121static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1122static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1123static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1124static int stat_journal_min; /* Times hit journal min threshold */ 1125static int stat_journal_low; /* Times hit journal low threshold */ 1126static int stat_journal_wait; /* Times blocked in jwait(). */ 1127static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1128static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1129static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1130static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1131 1132SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1133 &max_softdeps, 0, ""); 1134SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1135 &tickdelay, 0, ""); 1136SYSCTL_INT(_debug_softdep, OID_AUTO, maxindirdeps, CTLFLAG_RW, 1137 &maxindirdeps, 0, ""); 1138SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1139 &stat_worklist_push, 0,""); 1140SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1141 &stat_blk_limit_push, 0,""); 1142SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1143 &stat_ino_limit_push, 0,""); 1144SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1145 &stat_blk_limit_hit, 0, ""); 1146SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1147 &stat_ino_limit_hit, 0, ""); 1148SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1149 &stat_sync_limit_hit, 0, ""); 1150SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1151 &stat_indir_blk_ptrs, 0, ""); 1152SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1153 &stat_inode_bitmap, 0, ""); 1154SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1155 &stat_direct_blk_ptrs, 0, ""); 1156SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1157 &stat_dir_entry, 0, ""); 1158SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1159 &stat_jaddref, 0, ""); 1160SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1161 &stat_jnewblk, 0, ""); 1162SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1163 &stat_journal_low, 0, ""); 1164SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1165 &stat_journal_min, 0, ""); 1166SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1167 &stat_journal_wait, 0, ""); 1168SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1169 &stat_jwait_filepage, 0, ""); 1170SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1171 &stat_jwait_freeblks, 0, ""); 1172SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1173 &stat_jwait_inode, 0, ""); 1174SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1175 &stat_jwait_newblk, 0, ""); 1176 1177SYSCTL_DECL(_vfs_ffs); 1178 1179LIST_HEAD(bmsafemap_hashhead, bmsafemap) *bmsafemap_hashtbl; 1180static u_long bmsafemap_hash; /* size of hash table - 1 */ 1181 1182static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 1183SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1184 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1185 1186static struct proc *softdepproc; 1187static struct kproc_desc softdep_kp = { 1188 "softdepflush", 1189 softdep_flush, 1190 &softdepproc 1191}; 1192SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, 1193 &softdep_kp); 1194 1195static void 1196softdep_flush(void) 1197{ 1198 struct mount *nmp; 1199 struct mount *mp; 1200 struct ufsmount *ump; 1201 struct thread *td; 1202 int remaining; 1203 int vfslocked; 1204 1205 td = curthread; 1206 td->td_pflags |= TDP_NORUNNINGBUF; 1207 1208 for (;;) { 1209 kproc_suspend_check(softdepproc); 1210 vfslocked = VFS_LOCK_GIANT((struct mount *)NULL); 1211 ACQUIRE_LOCK(&lk); 1212 /* 1213 * If requested, try removing inode or removal dependencies. 1214 */ 1215 if (req_clear_inodedeps) { 1216 clear_inodedeps(td); 1217 req_clear_inodedeps -= 1; 1218 wakeup_one(&proc_waiting); 1219 } 1220 if (req_clear_remove) { 1221 clear_remove(td); 1222 req_clear_remove -= 1; 1223 wakeup_one(&proc_waiting); 1224 } 1225 FREE_LOCK(&lk); 1226 VFS_UNLOCK_GIANT(vfslocked); 1227 remaining = 0; 1228 mtx_lock(&mountlist_mtx); 1229 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1230 nmp = TAILQ_NEXT(mp, mnt_list); 1231 if ((mp->mnt_flag & MNT_SOFTDEP) == 0) 1232 continue; 1233 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 1234 continue; 1235 vfslocked = VFS_LOCK_GIANT(mp); 1236 softdep_process_worklist(mp, 0); 1237 ump = VFSTOUFS(mp); 1238 remaining += ump->softdep_on_worklist - 1239 ump->softdep_on_worklist_inprogress; 1240 VFS_UNLOCK_GIANT(vfslocked); 1241 mtx_lock(&mountlist_mtx); 1242 nmp = TAILQ_NEXT(mp, mnt_list); 1243 vfs_unbusy(mp); 1244 } 1245 mtx_unlock(&mountlist_mtx); 1246 if (remaining) 1247 continue; 1248 ACQUIRE_LOCK(&lk); 1249 if (!req_pending) 1250 msleep(&req_pending, &lk, PVM, "sdflush", hz); 1251 req_pending = 0; 1252 FREE_LOCK(&lk); 1253 } 1254} 1255 1256static void 1257worklist_speedup(void) 1258{ 1259 mtx_assert(&lk, MA_OWNED); 1260 if (req_pending == 0) { 1261 req_pending = 1; 1262 wakeup(&req_pending); 1263 } 1264} 1265 1266static int 1267softdep_speedup(void) 1268{ 1269 1270 worklist_speedup(); 1271 bd_speedup(); 1272 return speedup_syncer(); 1273} 1274 1275/* 1276 * Add an item to the end of the work queue. 1277 * This routine requires that the lock be held. 1278 * This is the only routine that adds items to the list. 1279 * The following routine is the only one that removes items 1280 * and does so in order from first to last. 1281 */ 1282static void 1283add_to_worklist(wk, nodelay) 1284 struct worklist *wk; 1285 int nodelay; 1286{ 1287 struct ufsmount *ump; 1288 1289 mtx_assert(&lk, MA_OWNED); 1290 ump = VFSTOUFS(wk->wk_mp); 1291 if (wk->wk_state & ONWORKLIST) 1292 panic("add_to_worklist: %s(0x%X) already on list", 1293 TYPENAME(wk->wk_type), wk->wk_state); 1294 wk->wk_state |= ONWORKLIST; 1295 if (LIST_EMPTY(&ump->softdep_workitem_pending)) 1296 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1297 else 1298 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1299 ump->softdep_worklist_tail = wk; 1300 ump->softdep_on_worklist += 1; 1301 if (nodelay) 1302 worklist_speedup(); 1303} 1304 1305/* 1306 * Remove the item to be processed. If we are removing the last 1307 * item on the list, we need to recalculate the tail pointer. 1308 */ 1309static void 1310remove_from_worklist(wk) 1311 struct worklist *wk; 1312{ 1313 struct ufsmount *ump; 1314 struct worklist *wkend; 1315 1316 ump = VFSTOUFS(wk->wk_mp); 1317 WORKLIST_REMOVE(wk); 1318 if (wk == ump->softdep_worklist_tail) { 1319 LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list) 1320 if (LIST_NEXT(wkend, wk_list) == NULL) 1321 break; 1322 ump->softdep_worklist_tail = wkend; 1323 } 1324 ump->softdep_on_worklist -= 1; 1325} 1326 1327/* 1328 * Process that runs once per second to handle items in the background queue. 1329 * 1330 * Note that we ensure that everything is done in the order in which they 1331 * appear in the queue. The code below depends on this property to ensure 1332 * that blocks of a file are freed before the inode itself is freed. This 1333 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1334 * until all the old ones have been purged from the dependency lists. 1335 */ 1336int 1337softdep_process_worklist(mp, full) 1338 struct mount *mp; 1339 int full; 1340{ 1341 struct thread *td = curthread; 1342 int cnt, matchcnt, loopcount; 1343 struct ufsmount *ump; 1344 long starttime; 1345 1346 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1347 /* 1348 * Record the process identifier of our caller so that we can give 1349 * this process preferential treatment in request_cleanup below. 1350 */ 1351 matchcnt = 0; 1352 ump = VFSTOUFS(mp); 1353 ACQUIRE_LOCK(&lk); 1354 loopcount = 1; 1355 starttime = time_second; 1356 softdep_process_journal(mp, full?MNT_WAIT:0); 1357 while (ump->softdep_on_worklist > 0) { 1358 if ((cnt = process_worklist_item(mp, LK_NOWAIT)) == -1) 1359 break; 1360 else 1361 matchcnt += cnt; 1362 /* 1363 * If requested, try removing inode or removal dependencies. 1364 */ 1365 if (req_clear_inodedeps) { 1366 clear_inodedeps(td); 1367 req_clear_inodedeps -= 1; 1368 wakeup_one(&proc_waiting); 1369 } 1370 if (req_clear_remove) { 1371 clear_remove(td); 1372 req_clear_remove -= 1; 1373 wakeup_one(&proc_waiting); 1374 } 1375 /* 1376 * We do not generally want to stop for buffer space, but if 1377 * we are really being a buffer hog, we will stop and wait. 1378 */ 1379 if (loopcount++ % 128 == 0) { 1380 FREE_LOCK(&lk); 1381 uio_yield(); 1382 bwillwrite(); 1383 ACQUIRE_LOCK(&lk); 1384 } 1385 /* 1386 * Never allow processing to run for more than one 1387 * second. Otherwise the other mountpoints may get 1388 * excessively backlogged. 1389 */ 1390 if (!full && starttime != time_second) 1391 break; 1392 } 1393 FREE_LOCK(&lk); 1394 return (matchcnt); 1395} 1396 1397/* 1398 * Process all removes associated with a vnode if we are running out of 1399 * journal space. Any other process which attempts to flush these will 1400 * be unable as we have the vnodes locked. 1401 */ 1402static void 1403process_removes(vp) 1404 struct vnode *vp; 1405{ 1406 struct inodedep *inodedep; 1407 struct dirrem *dirrem; 1408 struct mount *mp; 1409 ino_t inum; 1410 1411 mtx_assert(&lk, MA_OWNED); 1412 1413 mp = vp->v_mount; 1414 inum = VTOI(vp)->i_number; 1415 for (;;) { 1416 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1417 return; 1418 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) 1419 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1420 (COMPLETE | ONWORKLIST)) 1421 break; 1422 if (dirrem == NULL) 1423 return; 1424 /* 1425 * If another thread is trying to lock this vnode it will 1426 * fail but we must wait for it to do so before we can 1427 * proceed. 1428 */ 1429 if (dirrem->dm_state & INPROGRESS) { 1430 dirrem->dm_state |= IOWAITING; 1431 msleep(&dirrem->dm_list, &lk, PVM, "pwrwait", 0); 1432 continue; 1433 } 1434 remove_from_worklist(&dirrem->dm_list); 1435 FREE_LOCK(&lk); 1436 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1437 panic("process_removes: suspended filesystem"); 1438 handle_workitem_remove(dirrem, vp); 1439 vn_finished_secondary_write(mp); 1440 ACQUIRE_LOCK(&lk); 1441 } 1442} 1443 1444/* 1445 * Process one item on the worklist. 1446 */ 1447static int 1448process_worklist_item(mp, flags) 1449 struct mount *mp; 1450 int flags; 1451{ 1452 struct worklist *wk, *wkXXX; 1453 struct ufsmount *ump; 1454 struct vnode *vp; 1455 int matchcnt = 0; 1456 1457 mtx_assert(&lk, MA_OWNED); 1458 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1459 /* 1460 * If we are being called because of a process doing a 1461 * copy-on-write, then it is not safe to write as we may 1462 * recurse into the copy-on-write routine. 1463 */ 1464 if (curthread->td_pflags & TDP_COWINPROGRESS) 1465 return (-1); 1466 /* 1467 * Normally we just process each item on the worklist in order. 1468 * However, if we are in a situation where we cannot lock any 1469 * inodes, we have to skip over any dirrem requests whose 1470 * vnodes are resident and locked. 1471 */ 1472 vp = NULL; 1473 ump = VFSTOUFS(mp); 1474 LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) { 1475 if (wk->wk_state & INPROGRESS) { 1476 wkXXX = wk; 1477 continue; 1478 } 1479 wkXXX = wk; /* Record the last valid wk pointer. */ 1480 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM) 1481 break; 1482 wk->wk_state |= INPROGRESS; 1483 ump->softdep_on_worklist_inprogress++; 1484 FREE_LOCK(&lk); 1485 ffs_vgetf(mp, WK_DIRREM(wk)->dm_oldinum, 1486 LK_NOWAIT | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ); 1487 ACQUIRE_LOCK(&lk); 1488 if (wk->wk_state & IOWAITING) { 1489 wk->wk_state &= ~IOWAITING; 1490 wakeup(wk); 1491 } 1492 wk->wk_state &= ~INPROGRESS; 1493 ump->softdep_on_worklist_inprogress--; 1494 if (vp != NULL) 1495 break; 1496 } 1497 if (wk == 0) 1498 return (-1); 1499 remove_from_worklist(wk); 1500 FREE_LOCK(&lk); 1501 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1502 panic("process_worklist_item: suspended filesystem"); 1503 matchcnt++; 1504 switch (wk->wk_type) { 1505 1506 case D_DIRREM: 1507 /* removal of a directory entry */ 1508 handle_workitem_remove(WK_DIRREM(wk), vp); 1509 if (vp) 1510 vput(vp); 1511 break; 1512 1513 case D_FREEBLKS: 1514 /* releasing blocks and/or fragments from a file */ 1515 handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT); 1516 break; 1517 1518 case D_FREEFRAG: 1519 /* releasing a fragment when replaced as a file grows */ 1520 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1521 break; 1522 1523 case D_FREEFILE: 1524 /* releasing an inode when its link count drops to 0 */ 1525 handle_workitem_freefile(WK_FREEFILE(wk)); 1526 break; 1527 1528 case D_FREEWORK: 1529 /* Final block in an indirect was freed. */ 1530 handle_workitem_indirblk(WK_FREEWORK(wk)); 1531 break; 1532 1533 default: 1534 panic("%s_process_worklist: Unknown type %s", 1535 "softdep", TYPENAME(wk->wk_type)); 1536 /* NOTREACHED */ 1537 } 1538 vn_finished_secondary_write(mp); 1539 ACQUIRE_LOCK(&lk); 1540 return (matchcnt); 1541} 1542 1543/* 1544 * Move dependencies from one buffer to another. 1545 */ 1546int 1547softdep_move_dependencies(oldbp, newbp) 1548 struct buf *oldbp; 1549 struct buf *newbp; 1550{ 1551 struct worklist *wk, *wktail; 1552 int dirty; 1553 1554 dirty = 0; 1555 wktail = NULL; 1556 ACQUIRE_LOCK(&lk); 1557 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1558 LIST_REMOVE(wk, wk_list); 1559 if (wk->wk_type == D_BMSAFEMAP && 1560 bmsafemap_rollbacks(WK_BMSAFEMAP(wk))) 1561 dirty = 1; 1562 if (wktail == 0) 1563 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1564 else 1565 LIST_INSERT_AFTER(wktail, wk, wk_list); 1566 wktail = wk; 1567 } 1568 FREE_LOCK(&lk); 1569 1570 return (dirty); 1571} 1572 1573/* 1574 * Purge the work list of all items associated with a particular mount point. 1575 */ 1576int 1577softdep_flushworklist(oldmnt, countp, td) 1578 struct mount *oldmnt; 1579 int *countp; 1580 struct thread *td; 1581{ 1582 struct vnode *devvp; 1583 int count, error = 0; 1584 struct ufsmount *ump; 1585 1586 /* 1587 * Alternately flush the block device associated with the mount 1588 * point and process any dependencies that the flushing 1589 * creates. We continue until no more worklist dependencies 1590 * are found. 1591 */ 1592 *countp = 0; 1593 ump = VFSTOUFS(oldmnt); 1594 devvp = ump->um_devvp; 1595 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1596 *countp += count; 1597 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1598 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1599 VOP_UNLOCK(devvp, 0); 1600 if (error) 1601 break; 1602 } 1603 return (error); 1604} 1605 1606int 1607softdep_waitidle(struct mount *mp) 1608{ 1609 struct ufsmount *ump; 1610 int error; 1611 int i; 1612 1613 ump = VFSTOUFS(mp); 1614 ACQUIRE_LOCK(&lk); 1615 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1616 ump->softdep_req = 1; 1617 if (ump->softdep_on_worklist) 1618 panic("softdep_waitidle: work added after flush."); 1619 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 1620 } 1621 ump->softdep_req = 0; 1622 FREE_LOCK(&lk); 1623 error = 0; 1624 if (i == 10) { 1625 error = EBUSY; 1626 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1627 mp); 1628 } 1629 1630 return (error); 1631} 1632 1633/* 1634 * Flush all vnodes and worklist items associated with a specified mount point. 1635 */ 1636int 1637softdep_flushfiles(oldmnt, flags, td) 1638 struct mount *oldmnt; 1639 int flags; 1640 struct thread *td; 1641{ 1642 int error, depcount, loopcnt, retry_flush_count, retry; 1643 1644 loopcnt = 10; 1645 retry_flush_count = 3; 1646retry_flush: 1647 error = 0; 1648 1649 /* 1650 * Alternately flush the vnodes associated with the mount 1651 * point and process any dependencies that the flushing 1652 * creates. In theory, this loop can happen at most twice, 1653 * but we give it a few extra just to be sure. 1654 */ 1655 for (; loopcnt > 0; loopcnt--) { 1656 /* 1657 * Do another flush in case any vnodes were brought in 1658 * as part of the cleanup operations. 1659 */ 1660 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) 1661 break; 1662 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1663 depcount == 0) 1664 break; 1665 } 1666 /* 1667 * If we are unmounting then it is an error to fail. If we 1668 * are simply trying to downgrade to read-only, then filesystem 1669 * activity can keep us busy forever, so we just fail with EBUSY. 1670 */ 1671 if (loopcnt == 0) { 1672 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1673 panic("softdep_flushfiles: looping"); 1674 error = EBUSY; 1675 } 1676 if (!error) 1677 error = softdep_waitidle(oldmnt); 1678 if (!error) { 1679 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 1680 retry = 0; 1681 MNT_ILOCK(oldmnt); 1682 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 1683 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 1684 if (oldmnt->mnt_nvnodelistsize > 0) { 1685 if (--retry_flush_count > 0) { 1686 retry = 1; 1687 loopcnt = 3; 1688 } else 1689 error = EBUSY; 1690 } 1691 MNT_IUNLOCK(oldmnt); 1692 if (retry) 1693 goto retry_flush; 1694 } 1695 } 1696 return (error); 1697} 1698 1699/* 1700 * Structure hashing. 1701 * 1702 * There are three types of structures that can be looked up: 1703 * 1) pagedep structures identified by mount point, inode number, 1704 * and logical block. 1705 * 2) inodedep structures identified by mount point and inode number. 1706 * 3) newblk structures identified by mount point and 1707 * physical block number. 1708 * 1709 * The "pagedep" and "inodedep" dependency structures are hashed 1710 * separately from the file blocks and inodes to which they correspond. 1711 * This separation helps when the in-memory copy of an inode or 1712 * file block must be replaced. It also obviates the need to access 1713 * an inode or file page when simply updating (or de-allocating) 1714 * dependency structures. Lookup of newblk structures is needed to 1715 * find newly allocated blocks when trying to associate them with 1716 * their allocdirect or allocindir structure. 1717 * 1718 * The lookup routines optionally create and hash a new instance when 1719 * an existing entry is not found. 1720 */ 1721#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 1722#define NODELAY 0x0002 /* cannot do background work */ 1723 1724/* 1725 * Structures and routines associated with pagedep caching. 1726 */ 1727LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 1728u_long pagedep_hash; /* size of hash table - 1 */ 1729#define PAGEDEP_HASH(mp, inum, lbn) \ 1730 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 1731 pagedep_hash]) 1732 1733static int 1734pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 1735 struct pagedep_hashhead *pagedephd; 1736 ino_t ino; 1737 ufs_lbn_t lbn; 1738 struct mount *mp; 1739 int flags; 1740 struct pagedep **pagedeppp; 1741{ 1742 struct pagedep *pagedep; 1743 1744 LIST_FOREACH(pagedep, pagedephd, pd_hash) 1745 if (ino == pagedep->pd_ino && 1746 lbn == pagedep->pd_lbn && 1747 mp == pagedep->pd_list.wk_mp) 1748 break; 1749 if (pagedep) { 1750 *pagedeppp = pagedep; 1751 if ((flags & DEPALLOC) != 0 && 1752 (pagedep->pd_state & ONWORKLIST) == 0) 1753 return (0); 1754 return (1); 1755 } 1756 *pagedeppp = NULL; 1757 return (0); 1758} 1759/* 1760 * Look up a pagedep. Return 1 if found, 0 if not found or found 1761 * when asked to allocate but not associated with any buffer. 1762 * If not found, allocate if DEPALLOC flag is passed. 1763 * Found or allocated entry is returned in pagedeppp. 1764 * This routine must be called with splbio interrupts blocked. 1765 */ 1766static int 1767pagedep_lookup(mp, ino, lbn, flags, pagedeppp) 1768 struct mount *mp; 1769 ino_t ino; 1770 ufs_lbn_t lbn; 1771 int flags; 1772 struct pagedep **pagedeppp; 1773{ 1774 struct pagedep *pagedep; 1775 struct pagedep_hashhead *pagedephd; 1776 int ret; 1777 int i; 1778 1779 mtx_assert(&lk, MA_OWNED); 1780 pagedephd = PAGEDEP_HASH(mp, ino, lbn); 1781 1782 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 1783 if (*pagedeppp || (flags & DEPALLOC) == 0) 1784 return (ret); 1785 FREE_LOCK(&lk); 1786 pagedep = malloc(sizeof(struct pagedep), 1787 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 1788 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 1789 ACQUIRE_LOCK(&lk); 1790 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 1791 if (*pagedeppp) { 1792 WORKITEM_FREE(pagedep, D_PAGEDEP); 1793 return (ret); 1794 } 1795 pagedep->pd_ino = ino; 1796 pagedep->pd_lbn = lbn; 1797 LIST_INIT(&pagedep->pd_dirremhd); 1798 LIST_INIT(&pagedep->pd_pendinghd); 1799 for (i = 0; i < DAHASHSZ; i++) 1800 LIST_INIT(&pagedep->pd_diraddhd[i]); 1801 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 1802 *pagedeppp = pagedep; 1803 return (0); 1804} 1805 1806/* 1807 * Structures and routines associated with inodedep caching. 1808 */ 1809LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 1810static u_long inodedep_hash; /* size of hash table - 1 */ 1811static long num_inodedep; /* number of inodedep allocated */ 1812#define INODEDEP_HASH(fs, inum) \ 1813 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 1814 1815static int 1816inodedep_find(inodedephd, fs, inum, inodedeppp) 1817 struct inodedep_hashhead *inodedephd; 1818 struct fs *fs; 1819 ino_t inum; 1820 struct inodedep **inodedeppp; 1821{ 1822 struct inodedep *inodedep; 1823 1824 LIST_FOREACH(inodedep, inodedephd, id_hash) 1825 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 1826 break; 1827 if (inodedep) { 1828 *inodedeppp = inodedep; 1829 return (1); 1830 } 1831 *inodedeppp = NULL; 1832 1833 return (0); 1834} 1835/* 1836 * Look up an inodedep. Return 1 if found, 0 if not found. 1837 * If not found, allocate if DEPALLOC flag is passed. 1838 * Found or allocated entry is returned in inodedeppp. 1839 * This routine must be called with splbio interrupts blocked. 1840 */ 1841static int 1842inodedep_lookup(mp, inum, flags, inodedeppp) 1843 struct mount *mp; 1844 ino_t inum; 1845 int flags; 1846 struct inodedep **inodedeppp; 1847{ 1848 struct inodedep *inodedep; 1849 struct inodedep_hashhead *inodedephd; 1850 struct fs *fs; 1851 1852 mtx_assert(&lk, MA_OWNED); 1853 fs = VFSTOUFS(mp)->um_fs; 1854 inodedephd = INODEDEP_HASH(fs, inum); 1855 1856 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 1857 return (1); 1858 if ((flags & DEPALLOC) == 0) 1859 return (0); 1860 /* 1861 * If we are over our limit, try to improve the situation. 1862 */ 1863 if (num_inodedep > max_softdeps && (flags & NODELAY) == 0) 1864 request_cleanup(mp, FLUSH_INODES); 1865 FREE_LOCK(&lk); 1866 inodedep = malloc(sizeof(struct inodedep), 1867 M_INODEDEP, M_SOFTDEP_FLAGS); 1868 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 1869 ACQUIRE_LOCK(&lk); 1870 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 1871 WORKITEM_FREE(inodedep, D_INODEDEP); 1872 return (1); 1873 } 1874 num_inodedep += 1; 1875 inodedep->id_fs = fs; 1876 inodedep->id_ino = inum; 1877 inodedep->id_state = ALLCOMPLETE; 1878 inodedep->id_nlinkdelta = 0; 1879 inodedep->id_savedino1 = NULL; 1880 inodedep->id_savedsize = -1; 1881 inodedep->id_savedextsize = -1; 1882 inodedep->id_savednlink = -1; 1883 inodedep->id_bmsafemap = NULL; 1884 inodedep->id_mkdiradd = NULL; 1885 LIST_INIT(&inodedep->id_dirremhd); 1886 LIST_INIT(&inodedep->id_pendinghd); 1887 LIST_INIT(&inodedep->id_inowait); 1888 LIST_INIT(&inodedep->id_bufwait); 1889 TAILQ_INIT(&inodedep->id_inoreflst); 1890 TAILQ_INIT(&inodedep->id_inoupdt); 1891 TAILQ_INIT(&inodedep->id_newinoupdt); 1892 TAILQ_INIT(&inodedep->id_extupdt); 1893 TAILQ_INIT(&inodedep->id_newextupdt); 1894 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 1895 *inodedeppp = inodedep; 1896 return (0); 1897} 1898 1899/* 1900 * Structures and routines associated with newblk caching. 1901 */ 1902LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 1903u_long newblk_hash; /* size of hash table - 1 */ 1904#define NEWBLK_HASH(fs, inum) \ 1905 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 1906 1907static int 1908newblk_find(newblkhd, mp, newblkno, flags, newblkpp) 1909 struct newblk_hashhead *newblkhd; 1910 struct mount *mp; 1911 ufs2_daddr_t newblkno; 1912 int flags; 1913 struct newblk **newblkpp; 1914{ 1915 struct newblk *newblk; 1916 1917 LIST_FOREACH(newblk, newblkhd, nb_hash) { 1918 if (newblkno != newblk->nb_newblkno) 1919 continue; 1920 if (mp != newblk->nb_list.wk_mp) 1921 continue; 1922 /* 1923 * If we're creating a new dependency don't match those that 1924 * have already been converted to allocdirects. This is for 1925 * a frag extend. 1926 */ 1927 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 1928 continue; 1929 break; 1930 } 1931 if (newblk) { 1932 *newblkpp = newblk; 1933 return (1); 1934 } 1935 *newblkpp = NULL; 1936 return (0); 1937} 1938 1939/* 1940 * Look up a newblk. Return 1 if found, 0 if not found. 1941 * If not found, allocate if DEPALLOC flag is passed. 1942 * Found or allocated entry is returned in newblkpp. 1943 */ 1944static int 1945newblk_lookup(mp, newblkno, flags, newblkpp) 1946 struct mount *mp; 1947 ufs2_daddr_t newblkno; 1948 int flags; 1949 struct newblk **newblkpp; 1950{ 1951 struct newblk *newblk; 1952 struct newblk_hashhead *newblkhd; 1953 1954 newblkhd = NEWBLK_HASH(VFSTOUFS(mp)->um_fs, newblkno); 1955 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) 1956 return (1); 1957 if ((flags & DEPALLOC) == 0) 1958 return (0); 1959 FREE_LOCK(&lk); 1960 newblk = malloc(sizeof(union allblk), M_NEWBLK, 1961 M_SOFTDEP_FLAGS | M_ZERO); 1962 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 1963 ACQUIRE_LOCK(&lk); 1964 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) { 1965 WORKITEM_FREE(newblk, D_NEWBLK); 1966 return (1); 1967 } 1968 newblk->nb_freefrag = NULL; 1969 LIST_INIT(&newblk->nb_indirdeps); 1970 LIST_INIT(&newblk->nb_newdirblk); 1971 LIST_INIT(&newblk->nb_jwork); 1972 newblk->nb_state = ATTACHED; 1973 newblk->nb_newblkno = newblkno; 1974 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 1975 *newblkpp = newblk; 1976 return (0); 1977} 1978 1979/* 1980 * Executed during filesystem system initialization before 1981 * mounting any filesystems. 1982 */ 1983void 1984softdep_initialize() 1985{ 1986 1987 LIST_INIT(&mkdirlisthd); 1988 max_softdeps = desiredvnodes * 4; 1989 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash); 1990 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 1991 newblk_hashtbl = hashinit(desiredvnodes / 5, M_NEWBLK, &newblk_hash); 1992 bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, &bmsafemap_hash); 1993 1994 /* initialise bioops hack */ 1995 bioops.io_start = softdep_disk_io_initiation; 1996 bioops.io_complete = softdep_disk_write_complete; 1997 bioops.io_deallocate = softdep_deallocate_dependencies; 1998 bioops.io_countdeps = softdep_count_dependencies; 1999 2000 /* Initialize the callout with an mtx. */ 2001 callout_init_mtx(&softdep_callout, &lk, 0); 2002} 2003 2004/* 2005 * Executed after all filesystems have been unmounted during 2006 * filesystem module unload. 2007 */ 2008void 2009softdep_uninitialize() 2010{ 2011 2012 callout_drain(&softdep_callout); 2013 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 2014 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 2015 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 2016 hashdestroy(bmsafemap_hashtbl, M_BMSAFEMAP, bmsafemap_hash); 2017} 2018 2019/* 2020 * Called at mount time to notify the dependency code that a 2021 * filesystem wishes to use it. 2022 */ 2023int 2024softdep_mount(devvp, mp, fs, cred) 2025 struct vnode *devvp; 2026 struct mount *mp; 2027 struct fs *fs; 2028 struct ucred *cred; 2029{ 2030 struct csum_total cstotal; 2031 struct ufsmount *ump; 2032 struct cg *cgp; 2033 struct buf *bp; 2034 int error, cyl; 2035 2036 MNT_ILOCK(mp); 2037 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2038 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2039 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2040 MNTK_SOFTDEP; 2041 mp->mnt_noasync++; 2042 } 2043 MNT_IUNLOCK(mp); 2044 ump = VFSTOUFS(mp); 2045 LIST_INIT(&ump->softdep_workitem_pending); 2046 LIST_INIT(&ump->softdep_journal_pending); 2047 TAILQ_INIT(&ump->softdep_unlinked); 2048 ump->softdep_worklist_tail = NULL; 2049 ump->softdep_on_worklist = 0; 2050 ump->softdep_deps = 0; 2051 if ((fs->fs_flags & FS_SUJ) && 2052 (error = journal_mount(mp, fs, cred)) != 0) { 2053 printf("Failed to start journal: %d\n", error); 2054 return (error); 2055 } 2056 /* 2057 * When doing soft updates, the counters in the 2058 * superblock may have gotten out of sync. Recomputation 2059 * can take a long time and can be deferred for background 2060 * fsck. However, the old behavior of scanning the cylinder 2061 * groups and recalculating them at mount time is available 2062 * by setting vfs.ffs.compute_summary_at_mount to one. 2063 */ 2064 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2065 return (0); 2066 bzero(&cstotal, sizeof cstotal); 2067 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2068 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2069 fs->fs_cgsize, cred, &bp)) != 0) { 2070 brelse(bp); 2071 return (error); 2072 } 2073 cgp = (struct cg *)bp->b_data; 2074 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2075 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2076 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2077 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2078 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2079 brelse(bp); 2080 } 2081#ifdef DEBUG 2082 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2083 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2084#endif 2085 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2086 return (0); 2087} 2088 2089void 2090softdep_unmount(mp) 2091 struct mount *mp; 2092{ 2093 2094 if (mp->mnt_kern_flag & MNTK_SUJ) 2095 journal_unmount(mp); 2096} 2097 2098struct jblocks { 2099 struct jseglst jb_segs; /* TAILQ of current segments. */ 2100 struct jseg *jb_writeseg; /* Next write to complete. */ 2101 struct jextent *jb_extent; /* Extent array. */ 2102 uint64_t jb_nextseq; /* Next sequence number. */ 2103 uint64_t jb_oldestseq; /* Oldest active sequence number. */ 2104 int jb_avail; /* Available extents. */ 2105 int jb_used; /* Last used extent. */ 2106 int jb_head; /* Allocator head. */ 2107 int jb_off; /* Allocator extent offset. */ 2108 int jb_blocks; /* Total disk blocks covered. */ 2109 int jb_free; /* Total disk blocks free. */ 2110 int jb_min; /* Minimum free space. */ 2111 int jb_low; /* Low on space. */ 2112 int jb_age; /* Insertion time of oldest rec. */ 2113 int jb_suspended; /* Did journal suspend writes? */ 2114}; 2115 2116struct jextent { 2117 ufs2_daddr_t je_daddr; /* Disk block address. */ 2118 int je_blocks; /* Disk block count. */ 2119}; 2120 2121static struct jblocks * 2122jblocks_create(void) 2123{ 2124 struct jblocks *jblocks; 2125 2126 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2127 TAILQ_INIT(&jblocks->jb_segs); 2128 jblocks->jb_avail = 10; 2129 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2130 M_JBLOCKS, M_WAITOK | M_ZERO); 2131 2132 return (jblocks); 2133} 2134 2135static ufs2_daddr_t 2136jblocks_alloc(jblocks, bytes, actual) 2137 struct jblocks *jblocks; 2138 int bytes; 2139 int *actual; 2140{ 2141 ufs2_daddr_t daddr; 2142 struct jextent *jext; 2143 int freecnt; 2144 int blocks; 2145 2146 blocks = bytes / DEV_BSIZE; 2147 jext = &jblocks->jb_extent[jblocks->jb_head]; 2148 freecnt = jext->je_blocks - jblocks->jb_off; 2149 if (freecnt == 0) { 2150 jblocks->jb_off = 0; 2151 if (++jblocks->jb_head > jblocks->jb_used) 2152 jblocks->jb_head = 0; 2153 jext = &jblocks->jb_extent[jblocks->jb_head]; 2154 freecnt = jext->je_blocks; 2155 } 2156 if (freecnt > blocks) 2157 freecnt = blocks; 2158 *actual = freecnt * DEV_BSIZE; 2159 daddr = jext->je_daddr + jblocks->jb_off; 2160 jblocks->jb_off += freecnt; 2161 jblocks->jb_free -= freecnt; 2162 2163 return (daddr); 2164} 2165 2166static void 2167jblocks_free(jblocks, mp, bytes) 2168 struct jblocks *jblocks; 2169 struct mount *mp; 2170 int bytes; 2171{ 2172 2173 jblocks->jb_free += bytes / DEV_BSIZE; 2174 if (jblocks->jb_suspended) 2175 worklist_speedup(); 2176 wakeup(jblocks); 2177} 2178 2179static void 2180jblocks_destroy(jblocks) 2181 struct jblocks *jblocks; 2182{ 2183 2184 if (jblocks->jb_extent) 2185 free(jblocks->jb_extent, M_JBLOCKS); 2186 free(jblocks, M_JBLOCKS); 2187} 2188 2189static void 2190jblocks_add(jblocks, daddr, blocks) 2191 struct jblocks *jblocks; 2192 ufs2_daddr_t daddr; 2193 int blocks; 2194{ 2195 struct jextent *jext; 2196 2197 jblocks->jb_blocks += blocks; 2198 jblocks->jb_free += blocks; 2199 jext = &jblocks->jb_extent[jblocks->jb_used]; 2200 /* Adding the first block. */ 2201 if (jext->je_daddr == 0) { 2202 jext->je_daddr = daddr; 2203 jext->je_blocks = blocks; 2204 return; 2205 } 2206 /* Extending the last extent. */ 2207 if (jext->je_daddr + jext->je_blocks == daddr) { 2208 jext->je_blocks += blocks; 2209 return; 2210 } 2211 /* Adding a new extent. */ 2212 if (++jblocks->jb_used == jblocks->jb_avail) { 2213 jblocks->jb_avail *= 2; 2214 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2215 M_JBLOCKS, M_WAITOK | M_ZERO); 2216 memcpy(jext, jblocks->jb_extent, 2217 sizeof(struct jextent) * jblocks->jb_used); 2218 free(jblocks->jb_extent, M_JBLOCKS); 2219 jblocks->jb_extent = jext; 2220 } 2221 jext = &jblocks->jb_extent[jblocks->jb_used]; 2222 jext->je_daddr = daddr; 2223 jext->je_blocks = blocks; 2224 return; 2225} 2226 2227int 2228softdep_journal_lookup(mp, vpp) 2229 struct mount *mp; 2230 struct vnode **vpp; 2231{ 2232 struct componentname cnp; 2233 struct vnode *dvp; 2234 ino_t sujournal; 2235 int error; 2236 2237 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2238 if (error) 2239 return (error); 2240 bzero(&cnp, sizeof(cnp)); 2241 cnp.cn_nameiop = LOOKUP; 2242 cnp.cn_flags = ISLASTCN; 2243 cnp.cn_thread = curthread; 2244 cnp.cn_cred = curthread->td_ucred; 2245 cnp.cn_pnbuf = SUJ_FILE; 2246 cnp.cn_nameptr = SUJ_FILE; 2247 cnp.cn_namelen = strlen(SUJ_FILE); 2248 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2249 vput(dvp); 2250 if (error != 0) 2251 return (error); 2252 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2253 return (error); 2254} 2255 2256/* 2257 * Open and verify the journal file. 2258 */ 2259static int 2260journal_mount(mp, fs, cred) 2261 struct mount *mp; 2262 struct fs *fs; 2263 struct ucred *cred; 2264{ 2265 struct jblocks *jblocks; 2266 struct vnode *vp; 2267 struct inode *ip; 2268 ufs2_daddr_t blkno; 2269 int bcount; 2270 int error; 2271 int i; 2272 2273 mp->mnt_kern_flag |= MNTK_SUJ; 2274 error = softdep_journal_lookup(mp, &vp); 2275 if (error != 0) { 2276 printf("Failed to find journal. Use tunefs to create one\n"); 2277 return (error); 2278 } 2279 ip = VTOI(vp); 2280 if (ip->i_size < SUJ_MIN) { 2281 error = ENOSPC; 2282 goto out; 2283 } 2284 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2285 jblocks = jblocks_create(); 2286 for (i = 0; i < bcount; i++) { 2287 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2288 if (error) 2289 break; 2290 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2291 } 2292 if (error) { 2293 jblocks_destroy(jblocks); 2294 goto out; 2295 } 2296 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2297 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2298 /* 2299 * Only validate the journal contents if the filesystem is clean, 2300 * otherwise we write the logs but they'll never be used. If the 2301 * filesystem was still dirty when we mounted it the journal is 2302 * invalid and a new journal can only be valid if it starts from a 2303 * clean mount. 2304 */ 2305 if (fs->fs_clean) { 2306 DIP_SET(ip, i_modrev, fs->fs_mtime); 2307 ip->i_flags |= IN_MODIFIED; 2308 ffs_update(vp, 1); 2309 } 2310 VFSTOUFS(mp)->softdep_jblocks = jblocks; 2311out: 2312 vput(vp); 2313 return (error); 2314} 2315 2316static void 2317journal_unmount(mp) 2318 struct mount *mp; 2319{ 2320 struct ufsmount *ump; 2321 2322 ump = VFSTOUFS(mp); 2323 if (ump->softdep_jblocks) 2324 jblocks_destroy(ump->softdep_jblocks); 2325 ump->softdep_jblocks = NULL; 2326} 2327 2328/* 2329 * Called when a journal record is ready to be written. Space is allocated 2330 * and the journal entry is created when the journal is flushed to stable 2331 * store. 2332 */ 2333static void 2334add_to_journal(wk) 2335 struct worklist *wk; 2336{ 2337 struct ufsmount *ump; 2338 2339 mtx_assert(&lk, MA_OWNED); 2340 ump = VFSTOUFS(wk->wk_mp); 2341 if (wk->wk_state & ONWORKLIST) 2342 panic("add_to_journal: %s(0x%X) already on list", 2343 TYPENAME(wk->wk_type), wk->wk_state); 2344 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2345 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2346 ump->softdep_jblocks->jb_age = ticks; 2347 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2348 } else 2349 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2350 ump->softdep_journal_tail = wk; 2351 ump->softdep_on_journal += 1; 2352} 2353 2354/* 2355 * Remove an arbitrary item for the journal worklist maintain the tail 2356 * pointer. This happens when a new operation obviates the need to 2357 * journal an old operation. 2358 */ 2359static void 2360remove_from_journal(wk) 2361 struct worklist *wk; 2362{ 2363 struct ufsmount *ump; 2364 2365 mtx_assert(&lk, MA_OWNED); 2366 ump = VFSTOUFS(wk->wk_mp); 2367#ifdef DEBUG /* XXX Expensive, temporary. */ 2368 { 2369 struct worklist *wkn; 2370 2371 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2372 if (wkn == wk) 2373 break; 2374 if (wkn == NULL) 2375 panic("remove_from_journal: %p is not in journal", wk); 2376 } 2377#endif 2378 /* 2379 * We emulate a TAILQ to save space in most structures which do not 2380 * require TAILQ semantics. Here we must update the tail position 2381 * when removing the tail which is not the final entry. 2382 */ 2383 if (ump->softdep_journal_tail == wk) 2384 ump->softdep_journal_tail = 2385 (struct worklist *)wk->wk_list.le_prev; 2386 2387 WORKLIST_REMOVE(wk); 2388 ump->softdep_on_journal -= 1; 2389} 2390 2391/* 2392 * Check for journal space as well as dependency limits so the prelink 2393 * code can throttle both journaled and non-journaled filesystems. 2394 * Threshold is 0 for low and 1 for min. 2395 */ 2396static int 2397journal_space(ump, thresh) 2398 struct ufsmount *ump; 2399 int thresh; 2400{ 2401 struct jblocks *jblocks; 2402 int avail; 2403 2404 /* 2405 * We use a tighter restriction here to prevent request_cleanup() 2406 * running in threads from running into locks we currently hold. 2407 */ 2408 if (num_inodedep > (max_softdeps / 10) * 9) 2409 return (0); 2410 2411 jblocks = ump->softdep_jblocks; 2412 if (jblocks == NULL) 2413 return (1); 2414 if (thresh) 2415 thresh = jblocks->jb_min; 2416 else 2417 thresh = jblocks->jb_low; 2418 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2419 avail = jblocks->jb_free - avail; 2420 2421 return (avail > thresh); 2422} 2423 2424static void 2425journal_suspend(ump) 2426 struct ufsmount *ump; 2427{ 2428 struct jblocks *jblocks; 2429 struct mount *mp; 2430 2431 mp = UFSTOVFS(ump); 2432 jblocks = ump->softdep_jblocks; 2433 MNT_ILOCK(mp); 2434 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2435 stat_journal_min++; 2436 mp->mnt_kern_flag |= MNTK_SUSPEND; 2437 mp->mnt_susp_owner = FIRST_THREAD_IN_PROC(softdepproc); 2438 } 2439 jblocks->jb_suspended = 1; 2440 MNT_IUNLOCK(mp); 2441} 2442 2443/* 2444 * Called before any allocation function to be certain that there is 2445 * sufficient space in the journal prior to creating any new records. 2446 * Since in the case of block allocation we may have multiple locked 2447 * buffers at the time of the actual allocation we can not block 2448 * when the journal records are created. Doing so would create a deadlock 2449 * if any of these buffers needed to be flushed to reclaim space. Instead 2450 * we require a sufficiently large amount of available space such that 2451 * each thread in the system could have passed this allocation check and 2452 * still have sufficient free space. With 20% of a minimum journal size 2453 * of 1MB we have 6553 records available. 2454 */ 2455int 2456softdep_prealloc(vp, waitok) 2457 struct vnode *vp; 2458 int waitok; 2459{ 2460 struct ufsmount *ump; 2461 2462 if (DOINGSUJ(vp) == 0) 2463 return (0); 2464 ump = VFSTOUFS(vp->v_mount); 2465 ACQUIRE_LOCK(&lk); 2466 if (journal_space(ump, 0)) { 2467 FREE_LOCK(&lk); 2468 return (0); 2469 } 2470 stat_journal_low++; 2471 FREE_LOCK(&lk); 2472 if (waitok == MNT_NOWAIT) 2473 return (ENOSPC); 2474 /* 2475 * Attempt to sync this vnode once to flush any journal 2476 * work attached to it. 2477 */ 2478 ffs_syncvnode(vp, waitok); 2479 ACQUIRE_LOCK(&lk); 2480 process_removes(vp); 2481 if (journal_space(ump, 0) == 0) { 2482 softdep_speedup(); 2483 if (journal_space(ump, 1) == 0) 2484 journal_suspend(ump); 2485 } 2486 FREE_LOCK(&lk); 2487 2488 return (0); 2489} 2490 2491/* 2492 * Before adjusting a link count on a vnode verify that we have sufficient 2493 * journal space. If not, process operations that depend on the currently 2494 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 2495 * and softdep flush threads can not acquire these locks to reclaim space. 2496 */ 2497static void 2498softdep_prelink(dvp, vp) 2499 struct vnode *dvp; 2500 struct vnode *vp; 2501{ 2502 struct ufsmount *ump; 2503 2504 ump = VFSTOUFS(dvp->v_mount); 2505 mtx_assert(&lk, MA_OWNED); 2506 if (journal_space(ump, 0)) 2507 return; 2508 stat_journal_low++; 2509 FREE_LOCK(&lk); 2510 if (vp) 2511 ffs_syncvnode(vp, MNT_NOWAIT); 2512 ffs_syncvnode(dvp, MNT_WAIT); 2513 ACQUIRE_LOCK(&lk); 2514 /* Process vp before dvp as it may create .. removes. */ 2515 if (vp) 2516 process_removes(vp); 2517 process_removes(dvp); 2518 softdep_speedup(); 2519 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT); 2520 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT); 2521 if (journal_space(ump, 0) == 0) { 2522 softdep_speedup(); 2523 if (journal_space(ump, 1) == 0) 2524 journal_suspend(ump); 2525 } 2526} 2527 2528static void 2529jseg_write(fs, jblocks, jseg, data) 2530 struct fs *fs; 2531 struct jblocks *jblocks; 2532 struct jseg *jseg; 2533 uint8_t *data; 2534{ 2535 struct jsegrec *rec; 2536 2537 rec = (struct jsegrec *)data; 2538 rec->jsr_seq = jseg->js_seq; 2539 rec->jsr_oldest = jblocks->jb_oldestseq; 2540 rec->jsr_cnt = jseg->js_cnt; 2541 rec->jsr_blocks = jseg->js_size / DEV_BSIZE; 2542 rec->jsr_crc = 0; 2543 rec->jsr_time = fs->fs_mtime; 2544} 2545 2546static inline void 2547inoref_write(inoref, jseg, rec) 2548 struct inoref *inoref; 2549 struct jseg *jseg; 2550 struct jrefrec *rec; 2551{ 2552 2553 inoref->if_jsegdep->jd_seg = jseg; 2554 rec->jr_ino = inoref->if_ino; 2555 rec->jr_parent = inoref->if_parent; 2556 rec->jr_nlink = inoref->if_nlink; 2557 rec->jr_mode = inoref->if_mode; 2558 rec->jr_diroff = inoref->if_diroff; 2559} 2560 2561static void 2562jaddref_write(jaddref, jseg, data) 2563 struct jaddref *jaddref; 2564 struct jseg *jseg; 2565 uint8_t *data; 2566{ 2567 struct jrefrec *rec; 2568 2569 rec = (struct jrefrec *)data; 2570 rec->jr_op = JOP_ADDREF; 2571 inoref_write(&jaddref->ja_ref, jseg, rec); 2572} 2573 2574static void 2575jremref_write(jremref, jseg, data) 2576 struct jremref *jremref; 2577 struct jseg *jseg; 2578 uint8_t *data; 2579{ 2580 struct jrefrec *rec; 2581 2582 rec = (struct jrefrec *)data; 2583 rec->jr_op = JOP_REMREF; 2584 inoref_write(&jremref->jr_ref, jseg, rec); 2585} 2586 2587static void 2588jmvref_write(jmvref, jseg, data) 2589 struct jmvref *jmvref; 2590 struct jseg *jseg; 2591 uint8_t *data; 2592{ 2593 struct jmvrec *rec; 2594 2595 rec = (struct jmvrec *)data; 2596 rec->jm_op = JOP_MVREF; 2597 rec->jm_ino = jmvref->jm_ino; 2598 rec->jm_parent = jmvref->jm_parent; 2599 rec->jm_oldoff = jmvref->jm_oldoff; 2600 rec->jm_newoff = jmvref->jm_newoff; 2601} 2602 2603static void 2604jnewblk_write(jnewblk, jseg, data) 2605 struct jnewblk *jnewblk; 2606 struct jseg *jseg; 2607 uint8_t *data; 2608{ 2609 struct jblkrec *rec; 2610 2611 jnewblk->jn_jsegdep->jd_seg = jseg; 2612 rec = (struct jblkrec *)data; 2613 rec->jb_op = JOP_NEWBLK; 2614 rec->jb_ino = jnewblk->jn_ino; 2615 rec->jb_blkno = jnewblk->jn_blkno; 2616 rec->jb_lbn = jnewblk->jn_lbn; 2617 rec->jb_frags = jnewblk->jn_frags; 2618 rec->jb_oldfrags = jnewblk->jn_oldfrags; 2619} 2620 2621static void 2622jfreeblk_write(jfreeblk, jseg, data) 2623 struct jfreeblk *jfreeblk; 2624 struct jseg *jseg; 2625 uint8_t *data; 2626{ 2627 struct jblkrec *rec; 2628 2629 jfreeblk->jf_jsegdep->jd_seg = jseg; 2630 rec = (struct jblkrec *)data; 2631 rec->jb_op = JOP_FREEBLK; 2632 rec->jb_ino = jfreeblk->jf_ino; 2633 rec->jb_blkno = jfreeblk->jf_blkno; 2634 rec->jb_lbn = jfreeblk->jf_lbn; 2635 rec->jb_frags = jfreeblk->jf_frags; 2636 rec->jb_oldfrags = 0; 2637} 2638 2639static void 2640jfreefrag_write(jfreefrag, jseg, data) 2641 struct jfreefrag *jfreefrag; 2642 struct jseg *jseg; 2643 uint8_t *data; 2644{ 2645 struct jblkrec *rec; 2646 2647 jfreefrag->fr_jsegdep->jd_seg = jseg; 2648 rec = (struct jblkrec *)data; 2649 rec->jb_op = JOP_FREEBLK; 2650 rec->jb_ino = jfreefrag->fr_ino; 2651 rec->jb_blkno = jfreefrag->fr_blkno; 2652 rec->jb_lbn = jfreefrag->fr_lbn; 2653 rec->jb_frags = jfreefrag->fr_frags; 2654 rec->jb_oldfrags = 0; 2655} 2656 2657static void 2658jtrunc_write(jtrunc, jseg, data) 2659 struct jtrunc *jtrunc; 2660 struct jseg *jseg; 2661 uint8_t *data; 2662{ 2663 struct jtrncrec *rec; 2664 2665 rec = (struct jtrncrec *)data; 2666 rec->jt_op = JOP_TRUNC; 2667 rec->jt_ino = jtrunc->jt_ino; 2668 rec->jt_size = jtrunc->jt_size; 2669 rec->jt_extsize = jtrunc->jt_extsize; 2670} 2671 2672/* 2673 * Flush some journal records to disk. 2674 */ 2675static void 2676softdep_process_journal(mp, flags) 2677 struct mount *mp; 2678 int flags; 2679{ 2680 struct jblocks *jblocks; 2681 struct ufsmount *ump; 2682 struct worklist *wk; 2683 struct jseg *jseg; 2684 struct buf *bp; 2685 uint8_t *data; 2686 struct fs *fs; 2687 int segwritten; 2688 int jrecmin; /* Minimum records per block. */ 2689 int jrecmax; /* Maximum records per block. */ 2690 int size; 2691 int cnt; 2692 int off; 2693 2694 if ((mp->mnt_kern_flag & MNTK_SUJ) == 0) 2695 return; 2696 ump = VFSTOUFS(mp); 2697 fs = ump->um_fs; 2698 jblocks = ump->softdep_jblocks; 2699 /* 2700 * We write anywhere between a disk block and fs block. The upper 2701 * bound is picked to prevent buffer cache fragmentation and limit 2702 * processing time per I/O. 2703 */ 2704 jrecmin = (DEV_BSIZE / JREC_SIZE) - 1; /* -1 for seg header */ 2705 jrecmax = (fs->fs_bsize / DEV_BSIZE) * jrecmin; 2706 segwritten = 0; 2707 while ((cnt = ump->softdep_on_journal) != 0) { 2708 /* 2709 * Create a new segment to hold as many as 'cnt' journal 2710 * entries and add them to the segment. Notice cnt is 2711 * off by one to account for the space required by the 2712 * jsegrec. If we don't have a full block to log skip it 2713 * unless we haven't written anything. 2714 */ 2715 cnt++; 2716 if (cnt < jrecmax && segwritten) 2717 break; 2718 /* 2719 * Verify some free journal space. softdep_prealloc() should 2720 * guarantee that we don't run out so this is indicative of 2721 * a problem with the flow control. Try to recover 2722 * gracefully in any event. 2723 */ 2724 while (jblocks->jb_free == 0) { 2725 if (flags != MNT_WAIT) 2726 break; 2727 printf("softdep: Out of journal space!\n"); 2728 softdep_speedup(); 2729 msleep(jblocks, &lk, PRIBIO, "jblocks", 1); 2730 } 2731 FREE_LOCK(&lk); 2732 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 2733 workitem_alloc(&jseg->js_list, D_JSEG, mp); 2734 LIST_INIT(&jseg->js_entries); 2735 jseg->js_state = ATTACHED; 2736 jseg->js_jblocks = jblocks; 2737 bp = geteblk(fs->fs_bsize, 0); 2738 ACQUIRE_LOCK(&lk); 2739 /* 2740 * If there was a race while we were allocating the block 2741 * and jseg the entry we care about was likely written. 2742 * We bail out in both the WAIT and NOWAIT case and assume 2743 * the caller will loop if the entry it cares about is 2744 * not written. 2745 */ 2746 if (ump->softdep_on_journal == 0 || jblocks->jb_free == 0) { 2747 bp->b_flags |= B_INVAL | B_NOCACHE; 2748 WORKITEM_FREE(jseg, D_JSEG); 2749 FREE_LOCK(&lk); 2750 brelse(bp); 2751 ACQUIRE_LOCK(&lk); 2752 break; 2753 } 2754 /* 2755 * Calculate the disk block size required for the available 2756 * records rounded to the min size. 2757 */ 2758 cnt = ump->softdep_on_journal; 2759 if (cnt < jrecmax) 2760 size = howmany(cnt, jrecmin) * DEV_BSIZE; 2761 else 2762 size = fs->fs_bsize; 2763 /* 2764 * Allocate a disk block for this journal data and account 2765 * for truncation of the requested size if enough contiguous 2766 * space was not available. 2767 */ 2768 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 2769 bp->b_lblkno = bp->b_blkno; 2770 bp->b_offset = bp->b_blkno * DEV_BSIZE; 2771 bp->b_bcount = size; 2772 bp->b_bufobj = &ump->um_devvp->v_bufobj; 2773 bp->b_flags &= ~B_INVAL; 2774 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 2775 /* 2776 * Initialize our jseg with cnt records. Assign the next 2777 * sequence number to it and link it in-order. 2778 */ 2779 cnt = MIN(ump->softdep_on_journal, 2780 (size / DEV_BSIZE) * jrecmin); 2781 jseg->js_buf = bp; 2782 jseg->js_cnt = cnt; 2783 jseg->js_refs = cnt + 1; /* Self ref. */ 2784 jseg->js_size = size; 2785 jseg->js_seq = jblocks->jb_nextseq++; 2786 if (TAILQ_EMPTY(&jblocks->jb_segs)) 2787 jblocks->jb_oldestseq = jseg->js_seq; 2788 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 2789 if (jblocks->jb_writeseg == NULL) 2790 jblocks->jb_writeseg = jseg; 2791 /* 2792 * Start filling in records from the pending list. 2793 */ 2794 data = bp->b_data; 2795 off = 0; 2796 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 2797 != NULL) { 2798 /* Place a segment header on every device block. */ 2799 if ((off % DEV_BSIZE) == 0) { 2800 jseg_write(fs, jblocks, jseg, data); 2801 off += JREC_SIZE; 2802 data = bp->b_data + off; 2803 } 2804 remove_from_journal(wk); 2805 wk->wk_state |= IOSTARTED; 2806 WORKLIST_INSERT(&jseg->js_entries, wk); 2807 switch (wk->wk_type) { 2808 case D_JADDREF: 2809 jaddref_write(WK_JADDREF(wk), jseg, data); 2810 break; 2811 case D_JREMREF: 2812 jremref_write(WK_JREMREF(wk), jseg, data); 2813 break; 2814 case D_JMVREF: 2815 jmvref_write(WK_JMVREF(wk), jseg, data); 2816 break; 2817 case D_JNEWBLK: 2818 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 2819 break; 2820 case D_JFREEBLK: 2821 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 2822 break; 2823 case D_JFREEFRAG: 2824 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 2825 break; 2826 case D_JTRUNC: 2827 jtrunc_write(WK_JTRUNC(wk), jseg, data); 2828 break; 2829 default: 2830 panic("process_journal: Unknown type %s", 2831 TYPENAME(wk->wk_type)); 2832 /* NOTREACHED */ 2833 } 2834 if (--cnt == 0) 2835 break; 2836 off += JREC_SIZE; 2837 data = bp->b_data + off; 2838 } 2839 /* 2840 * Write this one buffer and continue. 2841 */ 2842 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 2843 FREE_LOCK(&lk); 2844 BO_LOCK(bp->b_bufobj); 2845 bgetvp(ump->um_devvp, bp); 2846 BO_UNLOCK(bp->b_bufobj); 2847 if (flags == MNT_NOWAIT) 2848 bawrite(bp); 2849 else 2850 bwrite(bp); 2851 ACQUIRE_LOCK(&lk); 2852 } 2853 /* 2854 * If we've suspended the filesystem because we ran out of journal 2855 * space either try to sync it here to make some progress or 2856 * unsuspend it if we already have. 2857 */ 2858 if (flags == 0 && jblocks && jblocks->jb_suspended) { 2859 if (journal_space(ump, jblocks->jb_min)) { 2860 FREE_LOCK(&lk); 2861 jblocks->jb_suspended = 0; 2862 mp->mnt_susp_owner = curthread; 2863 vfs_write_resume(mp); 2864 ACQUIRE_LOCK(&lk); 2865 return; 2866 } 2867 FREE_LOCK(&lk); 2868 VFS_SYNC(mp, MNT_NOWAIT); 2869 ffs_sbupdate(ump, MNT_WAIT, 0); 2870 ACQUIRE_LOCK(&lk); 2871 } 2872} 2873 2874/* 2875 * Complete a jseg, allowing all dependencies awaiting journal writes 2876 * to proceed. Each journal dependency also attaches a jsegdep to dependent 2877 * structures so that the journal segment can be freed to reclaim space. 2878 */ 2879static void 2880complete_jseg(jseg) 2881 struct jseg *jseg; 2882{ 2883 struct worklist *wk; 2884 struct jmvref *jmvref; 2885 int waiting; 2886 int i; 2887 2888 i = 0; 2889 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 2890 WORKLIST_REMOVE(wk); 2891 waiting = wk->wk_state & IOWAITING; 2892 wk->wk_state &= ~(IOSTARTED | IOWAITING); 2893 wk->wk_state |= COMPLETE; 2894 KASSERT(i < jseg->js_cnt, 2895 ("handle_written_jseg: overflow %d >= %d", 2896 i, jseg->js_cnt)); 2897 switch (wk->wk_type) { 2898 case D_JADDREF: 2899 handle_written_jaddref(WK_JADDREF(wk)); 2900 break; 2901 case D_JREMREF: 2902 handle_written_jremref(WK_JREMREF(wk)); 2903 break; 2904 case D_JMVREF: 2905 /* No jsegdep here. */ 2906 free_jseg(jseg); 2907 jmvref = WK_JMVREF(wk); 2908 LIST_REMOVE(jmvref, jm_deps); 2909 free_pagedep(jmvref->jm_pagedep); 2910 WORKITEM_FREE(jmvref, D_JMVREF); 2911 break; 2912 case D_JNEWBLK: 2913 handle_written_jnewblk(WK_JNEWBLK(wk)); 2914 break; 2915 case D_JFREEBLK: 2916 handle_written_jfreeblk(WK_JFREEBLK(wk)); 2917 break; 2918 case D_JFREEFRAG: 2919 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 2920 break; 2921 case D_JTRUNC: 2922 WK_JTRUNC(wk)->jt_jsegdep->jd_seg = jseg; 2923 WORKITEM_FREE(wk, D_JTRUNC); 2924 break; 2925 default: 2926 panic("handle_written_jseg: Unknown type %s", 2927 TYPENAME(wk->wk_type)); 2928 /* NOTREACHED */ 2929 } 2930 if (waiting) 2931 wakeup(wk); 2932 } 2933 /* Release the self reference so the structure may be freed. */ 2934 free_jseg(jseg); 2935} 2936 2937/* 2938 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Handle jseg 2939 * completions in order only. 2940 */ 2941static void 2942handle_written_jseg(jseg, bp) 2943 struct jseg *jseg; 2944 struct buf *bp; 2945{ 2946 struct jblocks *jblocks; 2947 struct jseg *jsegn; 2948 2949 if (jseg->js_refs == 0) 2950 panic("handle_written_jseg: No self-reference on %p", jseg); 2951 jseg->js_state |= DEPCOMPLETE; 2952 /* 2953 * We'll never need this buffer again, set flags so it will be 2954 * discarded. 2955 */ 2956 bp->b_flags |= B_INVAL | B_NOCACHE; 2957 jblocks = jseg->js_jblocks; 2958 /* 2959 * Don't allow out of order completions. If this isn't the first 2960 * block wait for it to write before we're done. 2961 */ 2962 if (jseg != jblocks->jb_writeseg) 2963 return; 2964 /* Iterate through available jsegs processing their entries. */ 2965 do { 2966 jsegn = TAILQ_NEXT(jseg, js_next); 2967 complete_jseg(jseg); 2968 jseg = jsegn; 2969 } while (jseg && jseg->js_state & DEPCOMPLETE); 2970 jblocks->jb_writeseg = jseg; 2971} 2972 2973static inline struct jsegdep * 2974inoref_jseg(inoref) 2975 struct inoref *inoref; 2976{ 2977 struct jsegdep *jsegdep; 2978 2979 jsegdep = inoref->if_jsegdep; 2980 inoref->if_jsegdep = NULL; 2981 2982 return (jsegdep); 2983} 2984 2985/* 2986 * Called once a jremref has made it to stable store. The jremref is marked 2987 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 2988 * for the jremref to complete will be awoken by free_jremref. 2989 */ 2990static void 2991handle_written_jremref(jremref) 2992 struct jremref *jremref; 2993{ 2994 struct inodedep *inodedep; 2995 struct jsegdep *jsegdep; 2996 struct dirrem *dirrem; 2997 2998 /* Grab the jsegdep. */ 2999 jsegdep = inoref_jseg(&jremref->jr_ref); 3000 /* 3001 * Remove us from the inoref list. 3002 */ 3003 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3004 0, &inodedep) == 0) 3005 panic("handle_written_jremref: Lost inodedep"); 3006 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3007 /* 3008 * Complete the dirrem. 3009 */ 3010 dirrem = jremref->jr_dirrem; 3011 jremref->jr_dirrem = NULL; 3012 LIST_REMOVE(jremref, jr_deps); 3013 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3014 WORKLIST_INSERT(&dirrem->dm_jwork, &jsegdep->jd_list); 3015 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3016 (dirrem->dm_state & COMPLETE) != 0) 3017 add_to_worklist(&dirrem->dm_list, 0); 3018 free_jremref(jremref); 3019} 3020 3021/* 3022 * Called once a jaddref has made it to stable store. The dependency is 3023 * marked complete and any dependent structures are added to the inode 3024 * bufwait list to be completed as soon as it is written. If a bitmap write 3025 * depends on this entry we move the inode into the inodedephd of the 3026 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3027 */ 3028static void 3029handle_written_jaddref(jaddref) 3030 struct jaddref *jaddref; 3031{ 3032 struct jsegdep *jsegdep; 3033 struct inodedep *inodedep; 3034 struct diradd *diradd; 3035 struct mkdir *mkdir; 3036 3037 /* Grab the jsegdep. */ 3038 jsegdep = inoref_jseg(&jaddref->ja_ref); 3039 mkdir = NULL; 3040 diradd = NULL; 3041 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3042 0, &inodedep) == 0) 3043 panic("handle_written_jaddref: Lost inodedep."); 3044 if (jaddref->ja_diradd == NULL) 3045 panic("handle_written_jaddref: No dependency"); 3046 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3047 diradd = jaddref->ja_diradd; 3048 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3049 } else if (jaddref->ja_state & MKDIR_PARENT) { 3050 mkdir = jaddref->ja_mkdir; 3051 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3052 } else if (jaddref->ja_state & MKDIR_BODY) 3053 mkdir = jaddref->ja_mkdir; 3054 else 3055 panic("handle_written_jaddref: Unknown dependency %p", 3056 jaddref->ja_diradd); 3057 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3058 /* 3059 * Remove us from the inode list. 3060 */ 3061 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3062 /* 3063 * The mkdir may be waiting on the jaddref to clear before freeing. 3064 */ 3065 if (mkdir) { 3066 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3067 ("handle_written_jaddref: Incorrect type for mkdir %s", 3068 TYPENAME(mkdir->md_list.wk_type))); 3069 mkdir->md_jaddref = NULL; 3070 diradd = mkdir->md_diradd; 3071 mkdir->md_state |= DEPCOMPLETE; 3072 complete_mkdir(mkdir); 3073 } 3074 WORKLIST_INSERT(&diradd->da_jwork, &jsegdep->jd_list); 3075 if (jaddref->ja_state & NEWBLOCK) { 3076 inodedep->id_state |= ONDEPLIST; 3077 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3078 inodedep, id_deps); 3079 } 3080 free_jaddref(jaddref); 3081} 3082 3083/* 3084 * Called once a jnewblk journal is written. The allocdirect or allocindir 3085 * is placed in the bmsafemap to await notification of a written bitmap. 3086 */ 3087static void 3088handle_written_jnewblk(jnewblk) 3089 struct jnewblk *jnewblk; 3090{ 3091 struct bmsafemap *bmsafemap; 3092 struct jsegdep *jsegdep; 3093 struct newblk *newblk; 3094 3095 /* Grab the jsegdep. */ 3096 jsegdep = jnewblk->jn_jsegdep; 3097 jnewblk->jn_jsegdep = NULL; 3098 /* 3099 * Add the written block to the bmsafemap so it can be notified when 3100 * the bitmap is on disk. 3101 */ 3102 newblk = jnewblk->jn_newblk; 3103 jnewblk->jn_newblk = NULL; 3104 if (newblk == NULL) 3105 panic("handle_written_jnewblk: No dependency for the segdep."); 3106 3107 newblk->nb_jnewblk = NULL; 3108 bmsafemap = newblk->nb_bmsafemap; 3109 WORKLIST_INSERT(&newblk->nb_jwork, &jsegdep->jd_list); 3110 newblk->nb_state |= ONDEPLIST; 3111 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 3112 free_jnewblk(jnewblk); 3113} 3114 3115/* 3116 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3117 * an in-flight allocation that has not yet been committed. Divorce us 3118 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3119 * to the worklist. 3120 */ 3121static void 3122cancel_jfreefrag(jfreefrag) 3123 struct jfreefrag *jfreefrag; 3124{ 3125 struct freefrag *freefrag; 3126 3127 if (jfreefrag->fr_jsegdep) { 3128 free_jsegdep(jfreefrag->fr_jsegdep); 3129 jfreefrag->fr_jsegdep = NULL; 3130 } 3131 freefrag = jfreefrag->fr_freefrag; 3132 jfreefrag->fr_freefrag = NULL; 3133 freefrag->ff_jfreefrag = NULL; 3134 free_jfreefrag(jfreefrag); 3135 freefrag->ff_state |= DEPCOMPLETE; 3136} 3137 3138/* 3139 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3140 */ 3141static void 3142free_jfreefrag(jfreefrag) 3143 struct jfreefrag *jfreefrag; 3144{ 3145 3146 if (jfreefrag->fr_state & IOSTARTED) 3147 WORKLIST_REMOVE(&jfreefrag->fr_list); 3148 else if (jfreefrag->fr_state & ONWORKLIST) 3149 remove_from_journal(&jfreefrag->fr_list); 3150 if (jfreefrag->fr_freefrag != NULL) 3151 panic("free_jfreefrag: Still attached to a freefrag."); 3152 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3153} 3154 3155/* 3156 * Called when the journal write for a jfreefrag completes. The parent 3157 * freefrag is added to the worklist if this completes its dependencies. 3158 */ 3159static void 3160handle_written_jfreefrag(jfreefrag) 3161 struct jfreefrag *jfreefrag; 3162{ 3163 struct jsegdep *jsegdep; 3164 struct freefrag *freefrag; 3165 3166 /* Grab the jsegdep. */ 3167 jsegdep = jfreefrag->fr_jsegdep; 3168 jfreefrag->fr_jsegdep = NULL; 3169 freefrag = jfreefrag->fr_freefrag; 3170 if (freefrag == NULL) 3171 panic("handle_written_jfreefrag: No freefrag."); 3172 freefrag->ff_state |= DEPCOMPLETE; 3173 freefrag->ff_jfreefrag = NULL; 3174 WORKLIST_INSERT(&freefrag->ff_jwork, &jsegdep->jd_list); 3175 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3176 add_to_worklist(&freefrag->ff_list, 0); 3177 jfreefrag->fr_freefrag = NULL; 3178 free_jfreefrag(jfreefrag); 3179} 3180 3181/* 3182 * Called when the journal write for a jfreeblk completes. The jfreeblk 3183 * is removed from the freeblks list of pending journal writes and the 3184 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3185 * have been reclaimed. 3186 */ 3187static void 3188handle_written_jfreeblk(jfreeblk) 3189 struct jfreeblk *jfreeblk; 3190{ 3191 struct freeblks *freeblks; 3192 struct jsegdep *jsegdep; 3193 3194 /* Grab the jsegdep. */ 3195 jsegdep = jfreeblk->jf_jsegdep; 3196 jfreeblk->jf_jsegdep = NULL; 3197 freeblks = jfreeblk->jf_freeblks; 3198 LIST_REMOVE(jfreeblk, jf_deps); 3199 WORKLIST_INSERT(&freeblks->fb_jwork, &jsegdep->jd_list); 3200 /* 3201 * If the freeblks is all journaled, we can add it to the worklist. 3202 */ 3203 if (LIST_EMPTY(&freeblks->fb_jfreeblkhd) && 3204 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) { 3205 /* Remove from the b_dep that is waiting on this write. */ 3206 if (freeblks->fb_state & ONWORKLIST) 3207 WORKLIST_REMOVE(&freeblks->fb_list); 3208 add_to_worklist(&freeblks->fb_list, 1); 3209 } 3210 3211 free_jfreeblk(jfreeblk); 3212} 3213 3214static struct jsegdep * 3215newjsegdep(struct worklist *wk) 3216{ 3217 struct jsegdep *jsegdep; 3218 3219 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3220 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3221 jsegdep->jd_seg = NULL; 3222 3223 return (jsegdep); 3224} 3225 3226static struct jmvref * 3227newjmvref(dp, ino, oldoff, newoff) 3228 struct inode *dp; 3229 ino_t ino; 3230 off_t oldoff; 3231 off_t newoff; 3232{ 3233 struct jmvref *jmvref; 3234 3235 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3236 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump)); 3237 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 3238 jmvref->jm_parent = dp->i_number; 3239 jmvref->jm_ino = ino; 3240 jmvref->jm_oldoff = oldoff; 3241 jmvref->jm_newoff = newoff; 3242 3243 return (jmvref); 3244} 3245 3246/* 3247 * Allocate a new jremref that tracks the removal of ip from dp with the 3248 * directory entry offset of diroff. Mark the entry as ATTACHED and 3249 * DEPCOMPLETE as we have all the information required for the journal write 3250 * and the directory has already been removed from the buffer. The caller 3251 * is responsible for linking the jremref into the pagedep and adding it 3252 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 3253 * a DOTDOT addition so handle_workitem_remove() can properly assign 3254 * the jsegdep when we're done. 3255 */ 3256static struct jremref * 3257newjremref(dirrem, dp, ip, diroff, nlink) 3258 struct dirrem *dirrem; 3259 struct inode *dp; 3260 struct inode *ip; 3261 off_t diroff; 3262 nlink_t nlink; 3263{ 3264 struct jremref *jremref; 3265 3266 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 3267 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump)); 3268 jremref->jr_state = ATTACHED; 3269 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 3270 nlink, ip->i_mode); 3271 jremref->jr_dirrem = dirrem; 3272 3273 return (jremref); 3274} 3275 3276static inline void 3277newinoref(inoref, ino, parent, diroff, nlink, mode) 3278 struct inoref *inoref; 3279 ino_t ino; 3280 ino_t parent; 3281 off_t diroff; 3282 nlink_t nlink; 3283 uint16_t mode; 3284{ 3285 3286 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 3287 inoref->if_diroff = diroff; 3288 inoref->if_ino = ino; 3289 inoref->if_parent = parent; 3290 inoref->if_nlink = nlink; 3291 inoref->if_mode = mode; 3292} 3293 3294/* 3295 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 3296 * directory offset may not be known until later. The caller is responsible 3297 * adding the entry to the journal when this information is available. nlink 3298 * should be the link count prior to the addition and mode is only required 3299 * to have the correct FMT. 3300 */ 3301static struct jaddref * 3302newjaddref(dp, ino, diroff, nlink, mode) 3303 struct inode *dp; 3304 ino_t ino; 3305 off_t diroff; 3306 int16_t nlink; 3307 uint16_t mode; 3308{ 3309 struct jaddref *jaddref; 3310 3311 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 3312 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump)); 3313 jaddref->ja_state = ATTACHED; 3314 jaddref->ja_mkdir = NULL; 3315 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 3316 3317 return (jaddref); 3318} 3319 3320/* 3321 * Create a new free dependency for a freework. The caller is responsible 3322 * for adjusting the reference count when it has the lock held. The freedep 3323 * will track an outstanding bitmap write that will ultimately clear the 3324 * freework to continue. 3325 */ 3326static struct freedep * 3327newfreedep(struct freework *freework) 3328{ 3329 struct freedep *freedep; 3330 3331 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 3332 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 3333 freedep->fd_freework = freework; 3334 3335 return (freedep); 3336} 3337 3338/* 3339 * Free a freedep structure once the buffer it is linked to is written. If 3340 * this is the last reference to the freework schedule it for completion. 3341 */ 3342static void 3343free_freedep(freedep) 3344 struct freedep *freedep; 3345{ 3346 3347 if (--freedep->fd_freework->fw_ref == 0) 3348 add_to_worklist(&freedep->fd_freework->fw_list, 1); 3349 WORKITEM_FREE(freedep, D_FREEDEP); 3350} 3351 3352/* 3353 * Allocate a new freework structure that may be a level in an indirect 3354 * when parent is not NULL or a top level block when it is. The top level 3355 * freework structures are allocated without lk held and before the freeblks 3356 * is visible outside of softdep_setup_freeblocks(). 3357 */ 3358static struct freework * 3359newfreework(freeblks, parent, lbn, nb, frags, journal) 3360 struct freeblks *freeblks; 3361 struct freework *parent; 3362 ufs_lbn_t lbn; 3363 ufs2_daddr_t nb; 3364 int frags; 3365 int journal; 3366{ 3367 struct freework *freework; 3368 3369 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 3370 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 3371 freework->fw_freeblks = freeblks; 3372 freework->fw_parent = parent; 3373 freework->fw_lbn = lbn; 3374 freework->fw_blkno = nb; 3375 freework->fw_frags = frags; 3376 freework->fw_ref = 0; 3377 freework->fw_off = 0; 3378 LIST_INIT(&freework->fw_jwork); 3379 3380 if (parent == NULL) { 3381 WORKLIST_INSERT_UNLOCKED(&freeblks->fb_freeworkhd, 3382 &freework->fw_list); 3383 freeblks->fb_ref++; 3384 } 3385 if (journal) 3386 newjfreeblk(freeblks, lbn, nb, frags); 3387 3388 return (freework); 3389} 3390 3391/* 3392 * Allocate a new jfreeblk to journal top level block pointer when truncating 3393 * a file. The caller must add this to the worklist when lk is held. 3394 */ 3395static struct jfreeblk * 3396newjfreeblk(freeblks, lbn, blkno, frags) 3397 struct freeblks *freeblks; 3398 ufs_lbn_t lbn; 3399 ufs2_daddr_t blkno; 3400 int frags; 3401{ 3402 struct jfreeblk *jfreeblk; 3403 3404 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 3405 workitem_alloc(&jfreeblk->jf_list, D_JFREEBLK, freeblks->fb_list.wk_mp); 3406 jfreeblk->jf_jsegdep = newjsegdep(&jfreeblk->jf_list); 3407 jfreeblk->jf_state = ATTACHED | DEPCOMPLETE; 3408 jfreeblk->jf_ino = freeblks->fb_previousinum; 3409 jfreeblk->jf_lbn = lbn; 3410 jfreeblk->jf_blkno = blkno; 3411 jfreeblk->jf_frags = frags; 3412 jfreeblk->jf_freeblks = freeblks; 3413 LIST_INSERT_HEAD(&freeblks->fb_jfreeblkhd, jfreeblk, jf_deps); 3414 3415 return (jfreeblk); 3416} 3417 3418static void move_newblock_dep(struct jaddref *, struct inodedep *); 3419/* 3420 * If we're canceling a new bitmap we have to search for another ref 3421 * to move into the bmsafemap dep. This might be better expressed 3422 * with another structure. 3423 */ 3424static void 3425move_newblock_dep(jaddref, inodedep) 3426 struct jaddref *jaddref; 3427 struct inodedep *inodedep; 3428{ 3429 struct inoref *inoref; 3430 struct jaddref *jaddrefn; 3431 3432 jaddrefn = NULL; 3433 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 3434 inoref = TAILQ_NEXT(inoref, if_deps)) { 3435 if ((jaddref->ja_state & NEWBLOCK) && 3436 inoref->if_list.wk_type == D_JADDREF) { 3437 jaddrefn = (struct jaddref *)inoref; 3438 break; 3439 } 3440 } 3441 if (jaddrefn == NULL) 3442 return; 3443 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 3444 jaddrefn->ja_state |= jaddref->ja_state & 3445 (ATTACHED | UNDONE | NEWBLOCK); 3446 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 3447 jaddref->ja_state |= ATTACHED; 3448 LIST_REMOVE(jaddref, ja_bmdeps); 3449 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 3450 ja_bmdeps); 3451} 3452 3453/* 3454 * Cancel a jaddref either before it has been written or while it is being 3455 * written. This happens when a link is removed before the add reaches 3456 * the disk. The jaddref dependency is kept linked into the bmsafemap 3457 * and inode to prevent the link count or bitmap from reaching the disk 3458 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 3459 * required. 3460 * 3461 * Returns 1 if the canceled addref requires journaling of the remove and 3462 * 0 otherwise. 3463 */ 3464static int 3465cancel_jaddref(jaddref, inodedep, wkhd) 3466 struct jaddref *jaddref; 3467 struct inodedep *inodedep; 3468 struct workhead *wkhd; 3469{ 3470 struct inoref *inoref; 3471 struct jsegdep *jsegdep; 3472 int needsj; 3473 3474 KASSERT((jaddref->ja_state & COMPLETE) == 0, 3475 ("cancel_jaddref: Canceling complete jaddref")); 3476 if (jaddref->ja_state & (IOSTARTED | COMPLETE)) 3477 needsj = 1; 3478 else 3479 needsj = 0; 3480 if (inodedep == NULL) 3481 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3482 0, &inodedep) == 0) 3483 panic("cancel_jaddref: Lost inodedep"); 3484 /* 3485 * We must adjust the nlink of any reference operation that follows 3486 * us so that it is consistent with the in-memory reference. This 3487 * ensures that inode nlink rollbacks always have the correct link. 3488 */ 3489 if (needsj == 0) 3490 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 3491 inoref = TAILQ_NEXT(inoref, if_deps)) 3492 inoref->if_nlink--; 3493 jsegdep = inoref_jseg(&jaddref->ja_ref); 3494 if (jaddref->ja_state & NEWBLOCK) 3495 move_newblock_dep(jaddref, inodedep); 3496 if (jaddref->ja_state & IOWAITING) { 3497 jaddref->ja_state &= ~IOWAITING; 3498 wakeup(&jaddref->ja_list); 3499 } 3500 jaddref->ja_mkdir = NULL; 3501 if (jaddref->ja_state & IOSTARTED) { 3502 jaddref->ja_state &= ~IOSTARTED; 3503 WORKLIST_REMOVE(&jaddref->ja_list); 3504 WORKLIST_INSERT(wkhd, &jsegdep->jd_list); 3505 } else { 3506 free_jsegdep(jsegdep); 3507 if (jaddref->ja_state & DEPCOMPLETE) 3508 remove_from_journal(&jaddref->ja_list); 3509 } 3510 /* 3511 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 3512 * can arrange for them to be freed with the bitmap. Otherwise we 3513 * no longer need this addref attached to the inoreflst and it 3514 * will incorrectly adjust nlink if we leave it. 3515 */ 3516 if ((jaddref->ja_state & NEWBLOCK) == 0) { 3517 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 3518 if_deps); 3519 jaddref->ja_state |= COMPLETE; 3520 free_jaddref(jaddref); 3521 return (needsj); 3522 } 3523 jaddref->ja_state |= GOINGAWAY; 3524 /* 3525 * Leave the head of the list for jsegdeps for fast merging. 3526 */ 3527 if (LIST_FIRST(wkhd) != NULL) { 3528 jaddref->ja_state |= ONWORKLIST; 3529 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 3530 } else 3531 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 3532 3533 return (needsj); 3534} 3535 3536/* 3537 * Attempt to free a jaddref structure when some work completes. This 3538 * should only succeed once the entry is written and all dependencies have 3539 * been notified. 3540 */ 3541static void 3542free_jaddref(jaddref) 3543 struct jaddref *jaddref; 3544{ 3545 3546 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 3547 return; 3548 if (jaddref->ja_ref.if_jsegdep) 3549 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 3550 jaddref, jaddref->ja_state); 3551 if (jaddref->ja_state & NEWBLOCK) 3552 LIST_REMOVE(jaddref, ja_bmdeps); 3553 if (jaddref->ja_state & (IOSTARTED | ONWORKLIST)) 3554 panic("free_jaddref: Bad state %p(0x%X)", 3555 jaddref, jaddref->ja_state); 3556 if (jaddref->ja_mkdir != NULL) 3557 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 3558 WORKITEM_FREE(jaddref, D_JADDREF); 3559} 3560 3561/* 3562 * Free a jremref structure once it has been written or discarded. 3563 */ 3564static void 3565free_jremref(jremref) 3566 struct jremref *jremref; 3567{ 3568 3569 if (jremref->jr_ref.if_jsegdep) 3570 free_jsegdep(jremref->jr_ref.if_jsegdep); 3571 if (jremref->jr_state & IOSTARTED) 3572 panic("free_jremref: IO still pending"); 3573 WORKITEM_FREE(jremref, D_JREMREF); 3574} 3575 3576/* 3577 * Free a jnewblk structure. 3578 */ 3579static void 3580free_jnewblk(jnewblk) 3581 struct jnewblk *jnewblk; 3582{ 3583 3584 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 3585 return; 3586 LIST_REMOVE(jnewblk, jn_deps); 3587 if (jnewblk->jn_newblk != NULL) 3588 panic("free_jnewblk: Dependency still attached."); 3589 WORKITEM_FREE(jnewblk, D_JNEWBLK); 3590} 3591 3592/* 3593 * Cancel a jnewblk which has been superseded by a freeblk. The jnewblk 3594 * is kept linked into the bmsafemap until the free completes, thus 3595 * preventing the modified state from ever reaching disk. The free 3596 * routine must pass this structure via ffs_blkfree() to 3597 * softdep_setup_freeblks() so there is no race in releasing the space. 3598 */ 3599static void 3600cancel_jnewblk(jnewblk, wkhd) 3601 struct jnewblk *jnewblk; 3602 struct workhead *wkhd; 3603{ 3604 struct jsegdep *jsegdep; 3605 3606 jsegdep = jnewblk->jn_jsegdep; 3607 jnewblk->jn_jsegdep = NULL; 3608 free_jsegdep(jsegdep); 3609 jnewblk->jn_newblk = NULL; 3610 jnewblk->jn_state |= GOINGAWAY; 3611 if (jnewblk->jn_state & IOSTARTED) { 3612 jnewblk->jn_state &= ~IOSTARTED; 3613 WORKLIST_REMOVE(&jnewblk->jn_list); 3614 } else 3615 remove_from_journal(&jnewblk->jn_list); 3616 /* 3617 * Leave the head of the list for jsegdeps for fast merging. 3618 */ 3619 if (LIST_FIRST(wkhd) != NULL) { 3620 jnewblk->jn_state |= ONWORKLIST; 3621 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jnewblk->jn_list, wk_list); 3622 } else 3623 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 3624 if (jnewblk->jn_state & IOWAITING) { 3625 jnewblk->jn_state &= ~IOWAITING; 3626 wakeup(&jnewblk->jn_list); 3627 } 3628} 3629 3630static void 3631free_jfreeblk(jfreeblk) 3632 struct jfreeblk *jfreeblk; 3633{ 3634 3635 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 3636} 3637 3638/* 3639 * Release one reference to a jseg and free it if the count reaches 0. This 3640 * should eventually reclaim journal space as well. 3641 */ 3642static void 3643free_jseg(jseg) 3644 struct jseg *jseg; 3645{ 3646 struct jblocks *jblocks; 3647 3648 KASSERT(jseg->js_refs > 0, 3649 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 3650 if (--jseg->js_refs != 0) 3651 return; 3652 /* 3653 * Free only those jsegs which have none allocated before them to 3654 * preserve the journal space ordering. 3655 */ 3656 jblocks = jseg->js_jblocks; 3657 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 3658 jblocks->jb_oldestseq = jseg->js_seq; 3659 if (jseg->js_refs != 0) 3660 break; 3661 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 3662 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 3663 KASSERT(LIST_EMPTY(&jseg->js_entries), 3664 ("free_jseg: Freed jseg has valid entries.")); 3665 WORKITEM_FREE(jseg, D_JSEG); 3666 } 3667} 3668 3669/* 3670 * Release a jsegdep and decrement the jseg count. 3671 */ 3672static void 3673free_jsegdep(jsegdep) 3674 struct jsegdep *jsegdep; 3675{ 3676 3677 if (jsegdep->jd_seg) 3678 free_jseg(jsegdep->jd_seg); 3679 WORKITEM_FREE(jsegdep, D_JSEGDEP); 3680} 3681 3682/* 3683 * Wait for a journal item to make it to disk. Initiate journal processing 3684 * if required. 3685 */ 3686static void 3687jwait(wk) 3688 struct worklist *wk; 3689{ 3690 3691 stat_journal_wait++; 3692 /* 3693 * If IO has not started we process the journal. We can't mark the 3694 * worklist item as IOWAITING because we drop the lock while 3695 * processing the journal and the worklist entry may be freed after 3696 * this point. The caller may call back in and re-issue the request. 3697 */ 3698 if ((wk->wk_state & IOSTARTED) == 0) { 3699 softdep_process_journal(wk->wk_mp, MNT_WAIT); 3700 return; 3701 } 3702 wk->wk_state |= IOWAITING; 3703 msleep(wk, &lk, PRIBIO, "jwait", 0); 3704} 3705 3706/* 3707 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 3708 * appropriate. This is a convenience function to reduce duplicate code 3709 * for the setup and revert functions below. 3710 */ 3711static struct inodedep * 3712inodedep_lookup_ip(ip) 3713 struct inode *ip; 3714{ 3715 struct inodedep *inodedep; 3716 3717 KASSERT(ip->i_nlink >= ip->i_effnlink, 3718 ("inodedep_lookup_ip: bad delta")); 3719 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 3720 DEPALLOC, &inodedep); 3721 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 3722 3723 return (inodedep); 3724} 3725 3726/* 3727 * Create a journal entry that describes a truncate that we're about to 3728 * perform. The inode allocations and frees between here and the completion 3729 * of the operation are done asynchronously and without journaling. At 3730 * the end of the operation the vnode is sync'd and the journal space 3731 * is released. Recovery will discover the partially completed truncate 3732 * and complete it. 3733 */ 3734void * 3735softdep_setup_trunc(vp, length, flags) 3736 struct vnode *vp; 3737 off_t length; 3738 int flags; 3739{ 3740 struct jsegdep *jsegdep; 3741 struct jtrunc *jtrunc; 3742 struct ufsmount *ump; 3743 struct inode *ip; 3744 3745 softdep_prealloc(vp, MNT_WAIT); 3746 ip = VTOI(vp); 3747 ump = VFSTOUFS(vp->v_mount); 3748 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 3749 workitem_alloc(&jtrunc->jt_list, D_JTRUNC, vp->v_mount); 3750 jsegdep = jtrunc->jt_jsegdep = newjsegdep(&jtrunc->jt_list); 3751 jtrunc->jt_ino = ip->i_number; 3752 jtrunc->jt_extsize = 0; 3753 jtrunc->jt_size = length; 3754 if ((flags & IO_EXT) == 0 && ump->um_fstype == UFS2) 3755 jtrunc->jt_extsize = ip->i_din2->di_extsize; 3756 if ((flags & IO_NORMAL) == 0) 3757 jtrunc->jt_size = DIP(ip, i_size); 3758 ACQUIRE_LOCK(&lk); 3759 add_to_journal(&jtrunc->jt_list); 3760 while (jsegdep->jd_seg == NULL) { 3761 stat_jwait_freeblks++; 3762 jwait(&jtrunc->jt_list); 3763 } 3764 FREE_LOCK(&lk); 3765 3766 return (jsegdep); 3767} 3768 3769/* 3770 * After synchronous truncation is complete we free sync the vnode and 3771 * release the jsegdep so the journal space can be freed. 3772 */ 3773int 3774softdep_complete_trunc(vp, cookie) 3775 struct vnode *vp; 3776 void *cookie; 3777{ 3778 int error; 3779 3780 error = ffs_syncvnode(vp, MNT_WAIT); 3781 ACQUIRE_LOCK(&lk); 3782 free_jsegdep((struct jsegdep *)cookie); 3783 FREE_LOCK(&lk); 3784 3785 return (error); 3786} 3787 3788/* 3789 * Called prior to creating a new inode and linking it to a directory. The 3790 * jaddref structure must already be allocated by softdep_setup_inomapdep 3791 * and it is discovered here so we can initialize the mode and update 3792 * nlinkdelta. 3793 */ 3794void 3795softdep_setup_create(dp, ip) 3796 struct inode *dp; 3797 struct inode *ip; 3798{ 3799 struct inodedep *inodedep; 3800 struct jaddref *jaddref; 3801 struct vnode *dvp; 3802 3803 KASSERT(ip->i_nlink == 1, 3804 ("softdep_setup_create: Invalid link count.")); 3805 dvp = ITOV(dp); 3806 ACQUIRE_LOCK(&lk); 3807 inodedep = inodedep_lookup_ip(ip); 3808 if (DOINGSUJ(dvp)) { 3809 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 3810 inoreflst); 3811 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 3812 ("softdep_setup_create: No addref structure present.")); 3813 jaddref->ja_mode = ip->i_mode; 3814 } 3815 softdep_prelink(dvp, NULL); 3816 FREE_LOCK(&lk); 3817} 3818 3819/* 3820 * Create a jaddref structure to track the addition of a DOTDOT link when 3821 * we are reparenting an inode as part of a rename. This jaddref will be 3822 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 3823 * non-journaling softdep. 3824 */ 3825void 3826softdep_setup_dotdot_link(dp, ip) 3827 struct inode *dp; 3828 struct inode *ip; 3829{ 3830 struct inodedep *inodedep; 3831 struct jaddref *jaddref; 3832 struct vnode *dvp; 3833 struct vnode *vp; 3834 3835 dvp = ITOV(dp); 3836 vp = ITOV(ip); 3837 jaddref = NULL; 3838 /* 3839 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 3840 * is used as a normal link would be. 3841 */ 3842 if (DOINGSUJ(dvp)) 3843 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 3844 dp->i_effnlink - 1, dp->i_mode); 3845 ACQUIRE_LOCK(&lk); 3846 inodedep = inodedep_lookup_ip(dp); 3847 if (jaddref) 3848 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 3849 if_deps); 3850 softdep_prelink(dvp, ITOV(ip)); 3851 FREE_LOCK(&lk); 3852} 3853 3854/* 3855 * Create a jaddref structure to track a new link to an inode. The directory 3856 * offset is not known until softdep_setup_directory_add or 3857 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 3858 * softdep. 3859 */ 3860void 3861softdep_setup_link(dp, ip) 3862 struct inode *dp; 3863 struct inode *ip; 3864{ 3865 struct inodedep *inodedep; 3866 struct jaddref *jaddref; 3867 struct vnode *dvp; 3868 3869 dvp = ITOV(dp); 3870 jaddref = NULL; 3871 if (DOINGSUJ(dvp)) 3872 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 3873 ip->i_mode); 3874 ACQUIRE_LOCK(&lk); 3875 inodedep = inodedep_lookup_ip(ip); 3876 if (jaddref) 3877 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 3878 if_deps); 3879 softdep_prelink(dvp, ITOV(ip)); 3880 FREE_LOCK(&lk); 3881} 3882 3883/* 3884 * Called to create the jaddref structures to track . and .. references as 3885 * well as lookup and further initialize the incomplete jaddref created 3886 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 3887 * nlinkdelta for non-journaling softdep. 3888 */ 3889void 3890softdep_setup_mkdir(dp, ip) 3891 struct inode *dp; 3892 struct inode *ip; 3893{ 3894 struct inodedep *inodedep; 3895 struct jaddref *dotdotaddref; 3896 struct jaddref *dotaddref; 3897 struct jaddref *jaddref; 3898 struct vnode *dvp; 3899 3900 dvp = ITOV(dp); 3901 dotaddref = dotdotaddref = NULL; 3902 if (DOINGSUJ(dvp)) { 3903 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 3904 ip->i_mode); 3905 dotaddref->ja_state |= MKDIR_BODY; 3906 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 3907 dp->i_effnlink - 1, dp->i_mode); 3908 dotdotaddref->ja_state |= MKDIR_PARENT; 3909 } 3910 ACQUIRE_LOCK(&lk); 3911 inodedep = inodedep_lookup_ip(ip); 3912 if (DOINGSUJ(dvp)) { 3913 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 3914 inoreflst); 3915 KASSERT(jaddref != NULL, 3916 ("softdep_setup_mkdir: No addref structure present.")); 3917 KASSERT(jaddref->ja_parent == dp->i_number, 3918 ("softdep_setup_mkdir: bad parent %d", 3919 jaddref->ja_parent)); 3920 jaddref->ja_mode = ip->i_mode; 3921 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 3922 if_deps); 3923 } 3924 inodedep = inodedep_lookup_ip(dp); 3925 if (DOINGSUJ(dvp)) 3926 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 3927 &dotdotaddref->ja_ref, if_deps); 3928 softdep_prelink(ITOV(dp), NULL); 3929 FREE_LOCK(&lk); 3930} 3931 3932/* 3933 * Called to track nlinkdelta of the inode and parent directories prior to 3934 * unlinking a directory. 3935 */ 3936void 3937softdep_setup_rmdir(dp, ip) 3938 struct inode *dp; 3939 struct inode *ip; 3940{ 3941 struct vnode *dvp; 3942 3943 dvp = ITOV(dp); 3944 ACQUIRE_LOCK(&lk); 3945 (void) inodedep_lookup_ip(ip); 3946 (void) inodedep_lookup_ip(dp); 3947 softdep_prelink(dvp, ITOV(ip)); 3948 FREE_LOCK(&lk); 3949} 3950 3951/* 3952 * Called to track nlinkdelta of the inode and parent directories prior to 3953 * unlink. 3954 */ 3955void 3956softdep_setup_unlink(dp, ip) 3957 struct inode *dp; 3958 struct inode *ip; 3959{ 3960 struct vnode *dvp; 3961 3962 dvp = ITOV(dp); 3963 ACQUIRE_LOCK(&lk); 3964 (void) inodedep_lookup_ip(ip); 3965 (void) inodedep_lookup_ip(dp); 3966 softdep_prelink(dvp, ITOV(ip)); 3967 FREE_LOCK(&lk); 3968} 3969 3970/* 3971 * Called to release the journal structures created by a failed non-directory 3972 * creation. Adjusts nlinkdelta for non-journaling softdep. 3973 */ 3974void 3975softdep_revert_create(dp, ip) 3976 struct inode *dp; 3977 struct inode *ip; 3978{ 3979 struct inodedep *inodedep; 3980 struct jaddref *jaddref; 3981 struct vnode *dvp; 3982 3983 dvp = ITOV(dp); 3984 ACQUIRE_LOCK(&lk); 3985 inodedep = inodedep_lookup_ip(ip); 3986 if (DOINGSUJ(dvp)) { 3987 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 3988 inoreflst); 3989 KASSERT(jaddref->ja_parent == dp->i_number, 3990 ("softdep_revert_create: addref parent mismatch")); 3991 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 3992 } 3993 FREE_LOCK(&lk); 3994} 3995 3996/* 3997 * Called to release the journal structures created by a failed dotdot link 3998 * creation. Adjusts nlinkdelta for non-journaling softdep. 3999 */ 4000void 4001softdep_revert_dotdot_link(dp, ip) 4002 struct inode *dp; 4003 struct inode *ip; 4004{ 4005 struct inodedep *inodedep; 4006 struct jaddref *jaddref; 4007 struct vnode *dvp; 4008 4009 dvp = ITOV(dp); 4010 ACQUIRE_LOCK(&lk); 4011 inodedep = inodedep_lookup_ip(dp); 4012 if (DOINGSUJ(dvp)) { 4013 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4014 inoreflst); 4015 KASSERT(jaddref->ja_parent == ip->i_number, 4016 ("softdep_revert_dotdot_link: addref parent mismatch")); 4017 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4018 } 4019 FREE_LOCK(&lk); 4020} 4021 4022/* 4023 * Called to release the journal structures created by a failed link 4024 * addition. Adjusts nlinkdelta for non-journaling softdep. 4025 */ 4026void 4027softdep_revert_link(dp, ip) 4028 struct inode *dp; 4029 struct inode *ip; 4030{ 4031 struct inodedep *inodedep; 4032 struct jaddref *jaddref; 4033 struct vnode *dvp; 4034 4035 dvp = ITOV(dp); 4036 ACQUIRE_LOCK(&lk); 4037 inodedep = inodedep_lookup_ip(ip); 4038 if (DOINGSUJ(dvp)) { 4039 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4040 inoreflst); 4041 KASSERT(jaddref->ja_parent == dp->i_number, 4042 ("softdep_revert_link: addref parent mismatch")); 4043 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4044 } 4045 FREE_LOCK(&lk); 4046} 4047 4048/* 4049 * Called to release the journal structures created by a failed mkdir 4050 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4051 */ 4052void 4053softdep_revert_mkdir(dp, ip) 4054 struct inode *dp; 4055 struct inode *ip; 4056{ 4057 struct inodedep *inodedep; 4058 struct jaddref *jaddref; 4059 struct vnode *dvp; 4060 4061 dvp = ITOV(dp); 4062 4063 ACQUIRE_LOCK(&lk); 4064 inodedep = inodedep_lookup_ip(dp); 4065 if (DOINGSUJ(dvp)) { 4066 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4067 inoreflst); 4068 KASSERT(jaddref->ja_parent == ip->i_number, 4069 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4070 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4071 } 4072 inodedep = inodedep_lookup_ip(ip); 4073 if (DOINGSUJ(dvp)) { 4074 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4075 inoreflst); 4076 KASSERT(jaddref->ja_parent == dp->i_number, 4077 ("softdep_revert_mkdir: addref parent mismatch")); 4078 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4079 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4080 inoreflst); 4081 KASSERT(jaddref->ja_parent == ip->i_number, 4082 ("softdep_revert_mkdir: dot addref parent mismatch")); 4083 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4084 } 4085 FREE_LOCK(&lk); 4086} 4087 4088/* 4089 * Called to correct nlinkdelta after a failed rmdir. 4090 */ 4091void 4092softdep_revert_rmdir(dp, ip) 4093 struct inode *dp; 4094 struct inode *ip; 4095{ 4096 4097 ACQUIRE_LOCK(&lk); 4098 (void) inodedep_lookup_ip(ip); 4099 (void) inodedep_lookup_ip(dp); 4100 FREE_LOCK(&lk); 4101} 4102 4103/* 4104 * Protecting the freemaps (or bitmaps). 4105 * 4106 * To eliminate the need to execute fsck before mounting a filesystem 4107 * after a power failure, one must (conservatively) guarantee that the 4108 * on-disk copy of the bitmaps never indicate that a live inode or block is 4109 * free. So, when a block or inode is allocated, the bitmap should be 4110 * updated (on disk) before any new pointers. When a block or inode is 4111 * freed, the bitmap should not be updated until all pointers have been 4112 * reset. The latter dependency is handled by the delayed de-allocation 4113 * approach described below for block and inode de-allocation. The former 4114 * dependency is handled by calling the following procedure when a block or 4115 * inode is allocated. When an inode is allocated an "inodedep" is created 4116 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4117 * Each "inodedep" is also inserted into the hash indexing structure so 4118 * that any additional link additions can be made dependent on the inode 4119 * allocation. 4120 * 4121 * The ufs filesystem maintains a number of free block counts (e.g., per 4122 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4123 * in addition to the bitmaps. These counts are used to improve efficiency 4124 * during allocation and therefore must be consistent with the bitmaps. 4125 * There is no convenient way to guarantee post-crash consistency of these 4126 * counts with simple update ordering, for two main reasons: (1) The counts 4127 * and bitmaps for a single cylinder group block are not in the same disk 4128 * sector. If a disk write is interrupted (e.g., by power failure), one may 4129 * be written and the other not. (2) Some of the counts are located in the 4130 * superblock rather than the cylinder group block. So, we focus our soft 4131 * updates implementation on protecting the bitmaps. When mounting a 4132 * filesystem, we recompute the auxiliary counts from the bitmaps. 4133 */ 4134 4135/* 4136 * Called just after updating the cylinder group block to allocate an inode. 4137 */ 4138void 4139softdep_setup_inomapdep(bp, ip, newinum) 4140 struct buf *bp; /* buffer for cylgroup block with inode map */ 4141 struct inode *ip; /* inode related to allocation */ 4142 ino_t newinum; /* new inode number being allocated */ 4143{ 4144 struct inodedep *inodedep; 4145 struct bmsafemap *bmsafemap; 4146 struct jaddref *jaddref; 4147 struct mount *mp; 4148 struct fs *fs; 4149 4150 mp = UFSTOVFS(ip->i_ump); 4151 fs = ip->i_ump->um_fs; 4152 jaddref = NULL; 4153 4154 /* 4155 * Allocate the journal reference add structure so that the bitmap 4156 * can be dependent on it. 4157 */ 4158 if (mp->mnt_kern_flag & MNTK_SUJ) { 4159 jaddref = newjaddref(ip, newinum, 0, 0, 0); 4160 jaddref->ja_state |= NEWBLOCK; 4161 } 4162 4163 /* 4164 * Create a dependency for the newly allocated inode. 4165 * Panic if it already exists as something is seriously wrong. 4166 * Otherwise add it to the dependency list for the buffer holding 4167 * the cylinder group map from which it was allocated. 4168 */ 4169 ACQUIRE_LOCK(&lk); 4170 if ((inodedep_lookup(mp, newinum, DEPALLOC|NODELAY, &inodedep))) 4171 panic("softdep_setup_inomapdep: dependency %p for new" 4172 "inode already exists", inodedep); 4173 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum)); 4174 if (jaddref) { 4175 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 4176 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4177 if_deps); 4178 } else { 4179 inodedep->id_state |= ONDEPLIST; 4180 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 4181 } 4182 inodedep->id_bmsafemap = bmsafemap; 4183 inodedep->id_state &= ~DEPCOMPLETE; 4184 FREE_LOCK(&lk); 4185} 4186 4187/* 4188 * Called just after updating the cylinder group block to 4189 * allocate block or fragment. 4190 */ 4191void 4192softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 4193 struct buf *bp; /* buffer for cylgroup block with block map */ 4194 struct mount *mp; /* filesystem doing allocation */ 4195 ufs2_daddr_t newblkno; /* number of newly allocated block */ 4196 int frags; /* Number of fragments. */ 4197 int oldfrags; /* Previous number of fragments for extend. */ 4198{ 4199 struct newblk *newblk; 4200 struct bmsafemap *bmsafemap; 4201 struct jnewblk *jnewblk; 4202 struct fs *fs; 4203 4204 fs = VFSTOUFS(mp)->um_fs; 4205 jnewblk = NULL; 4206 /* 4207 * Create a dependency for the newly allocated block. 4208 * Add it to the dependency list for the buffer holding 4209 * the cylinder group map from which it was allocated. 4210 */ 4211 if (mp->mnt_kern_flag & MNTK_SUJ) { 4212 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 4213 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 4214 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 4215 jnewblk->jn_state = ATTACHED; 4216 jnewblk->jn_blkno = newblkno; 4217 jnewblk->jn_frags = frags; 4218 jnewblk->jn_oldfrags = oldfrags; 4219#ifdef SUJ_DEBUG 4220 { 4221 struct cg *cgp; 4222 uint8_t *blksfree; 4223 long bno; 4224 int i; 4225 4226 cgp = (struct cg *)bp->b_data; 4227 blksfree = cg_blksfree(cgp); 4228 bno = dtogd(fs, jnewblk->jn_blkno); 4229 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 4230 i++) { 4231 if (isset(blksfree, bno + i)) 4232 panic("softdep_setup_blkmapdep: " 4233 "free fragment %d from %d-%d " 4234 "state 0x%X dep %p", i, 4235 jnewblk->jn_oldfrags, 4236 jnewblk->jn_frags, 4237 jnewblk->jn_state, 4238 jnewblk->jn_newblk); 4239 } 4240 } 4241#endif 4242 } 4243 ACQUIRE_LOCK(&lk); 4244 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 4245 panic("softdep_setup_blkmapdep: found block"); 4246 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 4247 dtog(fs, newblkno)); 4248 if (jnewblk) { 4249 jnewblk->jn_newblk = newblk; 4250 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 4251 } else { 4252 newblk->nb_state |= ONDEPLIST; 4253 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 4254 } 4255 newblk->nb_bmsafemap = bmsafemap; 4256 newblk->nb_jnewblk = jnewblk; 4257 FREE_LOCK(&lk); 4258} 4259 4260#define BMSAFEMAP_HASH(fs, cg) \ 4261 (&bmsafemap_hashtbl[((((register_t)(fs)) >> 13) + (cg)) & bmsafemap_hash]) 4262 4263static int 4264bmsafemap_find(bmsafemaphd, mp, cg, bmsafemapp) 4265 struct bmsafemap_hashhead *bmsafemaphd; 4266 struct mount *mp; 4267 int cg; 4268 struct bmsafemap **bmsafemapp; 4269{ 4270 struct bmsafemap *bmsafemap; 4271 4272 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 4273 if (bmsafemap->sm_list.wk_mp == mp && bmsafemap->sm_cg == cg) 4274 break; 4275 if (bmsafemap) { 4276 *bmsafemapp = bmsafemap; 4277 return (1); 4278 } 4279 *bmsafemapp = NULL; 4280 4281 return (0); 4282} 4283 4284/* 4285 * Find the bmsafemap associated with a cylinder group buffer. 4286 * If none exists, create one. The buffer must be locked when 4287 * this routine is called and this routine must be called with 4288 * splbio interrupts blocked. 4289 */ 4290static struct bmsafemap * 4291bmsafemap_lookup(mp, bp, cg) 4292 struct mount *mp; 4293 struct buf *bp; 4294 int cg; 4295{ 4296 struct bmsafemap_hashhead *bmsafemaphd; 4297 struct bmsafemap *bmsafemap, *collision; 4298 struct worklist *wk; 4299 struct fs *fs; 4300 4301 mtx_assert(&lk, MA_OWNED); 4302 if (bp) 4303 LIST_FOREACH(wk, &bp->b_dep, wk_list) 4304 if (wk->wk_type == D_BMSAFEMAP) 4305 return (WK_BMSAFEMAP(wk)); 4306 fs = VFSTOUFS(mp)->um_fs; 4307 bmsafemaphd = BMSAFEMAP_HASH(fs, cg); 4308 if (bmsafemap_find(bmsafemaphd, mp, cg, &bmsafemap) == 1) 4309 return (bmsafemap); 4310 FREE_LOCK(&lk); 4311 bmsafemap = malloc(sizeof(struct bmsafemap), 4312 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 4313 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 4314 bmsafemap->sm_buf = bp; 4315 LIST_INIT(&bmsafemap->sm_inodedephd); 4316 LIST_INIT(&bmsafemap->sm_inodedepwr); 4317 LIST_INIT(&bmsafemap->sm_newblkhd); 4318 LIST_INIT(&bmsafemap->sm_newblkwr); 4319 LIST_INIT(&bmsafemap->sm_jaddrefhd); 4320 LIST_INIT(&bmsafemap->sm_jnewblkhd); 4321 ACQUIRE_LOCK(&lk); 4322 if (bmsafemap_find(bmsafemaphd, mp, cg, &collision) == 1) { 4323 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 4324 return (collision); 4325 } 4326 bmsafemap->sm_cg = cg; 4327 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 4328 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 4329 return (bmsafemap); 4330} 4331 4332/* 4333 * Direct block allocation dependencies. 4334 * 4335 * When a new block is allocated, the corresponding disk locations must be 4336 * initialized (with zeros or new data) before the on-disk inode points to 4337 * them. Also, the freemap from which the block was allocated must be 4338 * updated (on disk) before the inode's pointer. These two dependencies are 4339 * independent of each other and are needed for all file blocks and indirect 4340 * blocks that are pointed to directly by the inode. Just before the 4341 * "in-core" version of the inode is updated with a newly allocated block 4342 * number, a procedure (below) is called to setup allocation dependency 4343 * structures. These structures are removed when the corresponding 4344 * dependencies are satisfied or when the block allocation becomes obsolete 4345 * (i.e., the file is deleted, the block is de-allocated, or the block is a 4346 * fragment that gets upgraded). All of these cases are handled in 4347 * procedures described later. 4348 * 4349 * When a file extension causes a fragment to be upgraded, either to a larger 4350 * fragment or to a full block, the on-disk location may change (if the 4351 * previous fragment could not simply be extended). In this case, the old 4352 * fragment must be de-allocated, but not until after the inode's pointer has 4353 * been updated. In most cases, this is handled by later procedures, which 4354 * will construct a "freefrag" structure to be added to the workitem queue 4355 * when the inode update is complete (or obsolete). The main exception to 4356 * this is when an allocation occurs while a pending allocation dependency 4357 * (for the same block pointer) remains. This case is handled in the main 4358 * allocation dependency setup procedure by immediately freeing the 4359 * unreferenced fragments. 4360 */ 4361void 4362softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 4363 struct inode *ip; /* inode to which block is being added */ 4364 ufs_lbn_t off; /* block pointer within inode */ 4365 ufs2_daddr_t newblkno; /* disk block number being added */ 4366 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 4367 long newsize; /* size of new block */ 4368 long oldsize; /* size of new block */ 4369 struct buf *bp; /* bp for allocated block */ 4370{ 4371 struct allocdirect *adp, *oldadp; 4372 struct allocdirectlst *adphead; 4373 struct freefrag *freefrag; 4374 struct inodedep *inodedep; 4375 struct pagedep *pagedep; 4376 struct jnewblk *jnewblk; 4377 struct newblk *newblk; 4378 struct mount *mp; 4379 ufs_lbn_t lbn; 4380 4381 lbn = bp->b_lblkno; 4382 mp = UFSTOVFS(ip->i_ump); 4383 if (oldblkno && oldblkno != newblkno) 4384 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 4385 else 4386 freefrag = NULL; 4387 4388 ACQUIRE_LOCK(&lk); 4389 if (off >= NDADDR) { 4390 if (lbn > 0) 4391 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 4392 lbn, off); 4393 /* allocating an indirect block */ 4394 if (oldblkno != 0) 4395 panic("softdep_setup_allocdirect: non-zero indir"); 4396 } else { 4397 if (off != lbn) 4398 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 4399 lbn, off); 4400 /* 4401 * Allocating a direct block. 4402 * 4403 * If we are allocating a directory block, then we must 4404 * allocate an associated pagedep to track additions and 4405 * deletions. 4406 */ 4407 if ((ip->i_mode & IFMT) == IFDIR && 4408 pagedep_lookup(mp, ip->i_number, off, DEPALLOC, 4409 &pagedep) == 0) 4410 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 4411 } 4412 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 4413 panic("softdep_setup_allocdirect: lost block"); 4414 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 4415 ("softdep_setup_allocdirect: newblk already initialized")); 4416 /* 4417 * Convert the newblk to an allocdirect. 4418 */ 4419 newblk->nb_list.wk_type = D_ALLOCDIRECT; 4420 adp = (struct allocdirect *)newblk; 4421 newblk->nb_freefrag = freefrag; 4422 adp->ad_offset = off; 4423 adp->ad_oldblkno = oldblkno; 4424 adp->ad_newsize = newsize; 4425 adp->ad_oldsize = oldsize; 4426 4427 /* 4428 * Finish initializing the journal. 4429 */ 4430 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 4431 jnewblk->jn_ino = ip->i_number; 4432 jnewblk->jn_lbn = lbn; 4433 add_to_journal(&jnewblk->jn_list); 4434 } 4435 if (freefrag && freefrag->ff_jfreefrag != NULL) 4436 add_to_journal(&freefrag->ff_jfreefrag->fr_list); 4437 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 4438 adp->ad_inodedep = inodedep; 4439 4440 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 4441 /* 4442 * The list of allocdirects must be kept in sorted and ascending 4443 * order so that the rollback routines can quickly determine the 4444 * first uncommitted block (the size of the file stored on disk 4445 * ends at the end of the lowest committed fragment, or if there 4446 * are no fragments, at the end of the highest committed block). 4447 * Since files generally grow, the typical case is that the new 4448 * block is to be added at the end of the list. We speed this 4449 * special case by checking against the last allocdirect in the 4450 * list before laboriously traversing the list looking for the 4451 * insertion point. 4452 */ 4453 adphead = &inodedep->id_newinoupdt; 4454 oldadp = TAILQ_LAST(adphead, allocdirectlst); 4455 if (oldadp == NULL || oldadp->ad_offset <= off) { 4456 /* insert at end of list */ 4457 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 4458 if (oldadp != NULL && oldadp->ad_offset == off) 4459 allocdirect_merge(adphead, adp, oldadp); 4460 FREE_LOCK(&lk); 4461 return; 4462 } 4463 TAILQ_FOREACH(oldadp, adphead, ad_next) { 4464 if (oldadp->ad_offset >= off) 4465 break; 4466 } 4467 if (oldadp == NULL) 4468 panic("softdep_setup_allocdirect: lost entry"); 4469 /* insert in middle of list */ 4470 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 4471 if (oldadp->ad_offset == off) 4472 allocdirect_merge(adphead, adp, oldadp); 4473 4474 FREE_LOCK(&lk); 4475} 4476 4477/* 4478 * Replace an old allocdirect dependency with a newer one. 4479 * This routine must be called with splbio interrupts blocked. 4480 */ 4481static void 4482allocdirect_merge(adphead, newadp, oldadp) 4483 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 4484 struct allocdirect *newadp; /* allocdirect being added */ 4485 struct allocdirect *oldadp; /* existing allocdirect being checked */ 4486{ 4487 struct worklist *wk; 4488 struct freefrag *freefrag; 4489 struct newdirblk *newdirblk; 4490 4491 freefrag = NULL; 4492 mtx_assert(&lk, MA_OWNED); 4493 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 4494 newadp->ad_oldsize != oldadp->ad_newsize || 4495 newadp->ad_offset >= NDADDR) 4496 panic("%s %jd != new %jd || old size %ld != new %ld", 4497 "allocdirect_merge: old blkno", 4498 (intmax_t)newadp->ad_oldblkno, 4499 (intmax_t)oldadp->ad_newblkno, 4500 newadp->ad_oldsize, oldadp->ad_newsize); 4501 newadp->ad_oldblkno = oldadp->ad_oldblkno; 4502 newadp->ad_oldsize = oldadp->ad_oldsize; 4503 /* 4504 * If the old dependency had a fragment to free or had never 4505 * previously had a block allocated, then the new dependency 4506 * can immediately post its freefrag and adopt the old freefrag. 4507 * This action is done by swapping the freefrag dependencies. 4508 * The new dependency gains the old one's freefrag, and the 4509 * old one gets the new one and then immediately puts it on 4510 * the worklist when it is freed by free_newblk. It is 4511 * not possible to do this swap when the old dependency had a 4512 * non-zero size but no previous fragment to free. This condition 4513 * arises when the new block is an extension of the old block. 4514 * Here, the first part of the fragment allocated to the new 4515 * dependency is part of the block currently claimed on disk by 4516 * the old dependency, so cannot legitimately be freed until the 4517 * conditions for the new dependency are fulfilled. 4518 */ 4519 freefrag = newadp->ad_freefrag; 4520 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 4521 newadp->ad_freefrag = oldadp->ad_freefrag; 4522 oldadp->ad_freefrag = freefrag; 4523 } 4524 /* 4525 * If we are tracking a new directory-block allocation, 4526 * move it from the old allocdirect to the new allocdirect. 4527 */ 4528 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 4529 newdirblk = WK_NEWDIRBLK(wk); 4530 WORKLIST_REMOVE(&newdirblk->db_list); 4531 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 4532 panic("allocdirect_merge: extra newdirblk"); 4533 WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list); 4534 } 4535 TAILQ_REMOVE(adphead, oldadp, ad_next); 4536 /* 4537 * We need to move any journal dependencies over to the freefrag 4538 * that releases this block if it exists. Otherwise we are 4539 * extending an existing block and we'll wait until that is 4540 * complete to release the journal space and extend the 4541 * new journal to cover this old space as well. 4542 */ 4543 if (freefrag == NULL) { 4544 struct jnewblk *jnewblk; 4545 struct jnewblk *njnewblk; 4546 4547 if (oldadp->ad_newblkno != newadp->ad_newblkno) 4548 panic("allocdirect_merge: %jd != %jd", 4549 oldadp->ad_newblkno, newadp->ad_newblkno); 4550 jnewblk = oldadp->ad_block.nb_jnewblk; 4551 cancel_newblk(&oldadp->ad_block, &newadp->ad_block.nb_jwork); 4552 /* 4553 * We have an unwritten jnewblk, we need to merge the 4554 * frag bits with our own. The newer adp's journal can not 4555 * be written prior to the old one so no need to check for 4556 * it here. 4557 */ 4558 if (jnewblk) { 4559 njnewblk = newadp->ad_block.nb_jnewblk; 4560 if (njnewblk == NULL) 4561 panic("allocdirect_merge: No jnewblk"); 4562 if (jnewblk->jn_state & UNDONE) { 4563 njnewblk->jn_state |= UNDONE | NEWBLOCK; 4564 njnewblk->jn_state &= ~ATTACHED; 4565 jnewblk->jn_state &= ~UNDONE; 4566 } 4567 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 4568 WORKLIST_REMOVE(&jnewblk->jn_list); 4569 jnewblk->jn_state |= ATTACHED | COMPLETE; 4570 free_jnewblk(jnewblk); 4571 } 4572 } else { 4573 /* 4574 * We can skip journaling for this freefrag and just complete 4575 * any pending journal work for the allocdirect that is being 4576 * removed after the freefrag completes. 4577 */ 4578 if (freefrag->ff_jfreefrag) 4579 cancel_jfreefrag(freefrag->ff_jfreefrag); 4580 cancel_newblk(&oldadp->ad_block, &freefrag->ff_jwork); 4581 } 4582 free_newblk(&oldadp->ad_block); 4583} 4584 4585/* 4586 * Allocate a jfreefrag structure to journal a single block free. 4587 */ 4588static struct jfreefrag * 4589newjfreefrag(freefrag, ip, blkno, size, lbn) 4590 struct freefrag *freefrag; 4591 struct inode *ip; 4592 ufs2_daddr_t blkno; 4593 long size; 4594 ufs_lbn_t lbn; 4595{ 4596 struct jfreefrag *jfreefrag; 4597 struct fs *fs; 4598 4599 fs = ip->i_fs; 4600 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 4601 M_SOFTDEP_FLAGS); 4602 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump)); 4603 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 4604 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 4605 jfreefrag->fr_ino = ip->i_number; 4606 jfreefrag->fr_lbn = lbn; 4607 jfreefrag->fr_blkno = blkno; 4608 jfreefrag->fr_frags = numfrags(fs, size); 4609 jfreefrag->fr_freefrag = freefrag; 4610 4611 return (jfreefrag); 4612} 4613 4614/* 4615 * Allocate a new freefrag structure. 4616 */ 4617static struct freefrag * 4618newfreefrag(ip, blkno, size, lbn) 4619 struct inode *ip; 4620 ufs2_daddr_t blkno; 4621 long size; 4622 ufs_lbn_t lbn; 4623{ 4624 struct freefrag *freefrag; 4625 struct fs *fs; 4626 4627 fs = ip->i_fs; 4628 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 4629 panic("newfreefrag: frag size"); 4630 freefrag = malloc(sizeof(struct freefrag), 4631 M_FREEFRAG, M_SOFTDEP_FLAGS); 4632 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 4633 freefrag->ff_state = ATTACHED; 4634 LIST_INIT(&freefrag->ff_jwork); 4635 freefrag->ff_inum = ip->i_number; 4636 freefrag->ff_blkno = blkno; 4637 freefrag->ff_fragsize = size; 4638 4639 if (fs->fs_flags & FS_SUJ) { 4640 freefrag->ff_jfreefrag = 4641 newjfreefrag(freefrag, ip, blkno, size, lbn); 4642 } else { 4643 freefrag->ff_state |= DEPCOMPLETE; 4644 freefrag->ff_jfreefrag = NULL; 4645 } 4646 4647 return (freefrag); 4648} 4649 4650/* 4651 * This workitem de-allocates fragments that were replaced during 4652 * file block allocation. 4653 */ 4654static void 4655handle_workitem_freefrag(freefrag) 4656 struct freefrag *freefrag; 4657{ 4658 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 4659 struct workhead wkhd; 4660 4661 /* 4662 * It would be illegal to add new completion items to the 4663 * freefrag after it was schedule to be done so it must be 4664 * safe to modify the list head here. 4665 */ 4666 LIST_INIT(&wkhd); 4667 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 4668 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 4669 freefrag->ff_fragsize, freefrag->ff_inum, &wkhd); 4670 ACQUIRE_LOCK(&lk); 4671 WORKITEM_FREE(freefrag, D_FREEFRAG); 4672 FREE_LOCK(&lk); 4673} 4674 4675/* 4676 * Set up a dependency structure for an external attributes data block. 4677 * This routine follows much of the structure of softdep_setup_allocdirect. 4678 * See the description of softdep_setup_allocdirect above for details. 4679 */ 4680void 4681softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 4682 struct inode *ip; 4683 ufs_lbn_t off; 4684 ufs2_daddr_t newblkno; 4685 ufs2_daddr_t oldblkno; 4686 long newsize; 4687 long oldsize; 4688 struct buf *bp; 4689{ 4690 struct allocdirect *adp, *oldadp; 4691 struct allocdirectlst *adphead; 4692 struct freefrag *freefrag; 4693 struct inodedep *inodedep; 4694 struct jnewblk *jnewblk; 4695 struct newblk *newblk; 4696 struct mount *mp; 4697 ufs_lbn_t lbn; 4698 4699 if (off >= NXADDR) 4700 panic("softdep_setup_allocext: lbn %lld > NXADDR", 4701 (long long)off); 4702 4703 lbn = bp->b_lblkno; 4704 mp = UFSTOVFS(ip->i_ump); 4705 if (oldblkno && oldblkno != newblkno) 4706 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 4707 else 4708 freefrag = NULL; 4709 4710 ACQUIRE_LOCK(&lk); 4711 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 4712 panic("softdep_setup_allocext: lost block"); 4713 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 4714 ("softdep_setup_allocext: newblk already initialized")); 4715 /* 4716 * Convert the newblk to an allocdirect. 4717 */ 4718 newblk->nb_list.wk_type = D_ALLOCDIRECT; 4719 adp = (struct allocdirect *)newblk; 4720 newblk->nb_freefrag = freefrag; 4721 adp->ad_offset = off; 4722 adp->ad_oldblkno = oldblkno; 4723 adp->ad_newsize = newsize; 4724 adp->ad_oldsize = oldsize; 4725 adp->ad_state |= EXTDATA; 4726 4727 /* 4728 * Finish initializing the journal. 4729 */ 4730 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 4731 jnewblk->jn_ino = ip->i_number; 4732 jnewblk->jn_lbn = lbn; 4733 add_to_journal(&jnewblk->jn_list); 4734 } 4735 if (freefrag && freefrag->ff_jfreefrag != NULL) 4736 add_to_journal(&freefrag->ff_jfreefrag->fr_list); 4737 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 4738 adp->ad_inodedep = inodedep; 4739 4740 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 4741 /* 4742 * The list of allocdirects must be kept in sorted and ascending 4743 * order so that the rollback routines can quickly determine the 4744 * first uncommitted block (the size of the file stored on disk 4745 * ends at the end of the lowest committed fragment, or if there 4746 * are no fragments, at the end of the highest committed block). 4747 * Since files generally grow, the typical case is that the new 4748 * block is to be added at the end of the list. We speed this 4749 * special case by checking against the last allocdirect in the 4750 * list before laboriously traversing the list looking for the 4751 * insertion point. 4752 */ 4753 adphead = &inodedep->id_newextupdt; 4754 oldadp = TAILQ_LAST(adphead, allocdirectlst); 4755 if (oldadp == NULL || oldadp->ad_offset <= off) { 4756 /* insert at end of list */ 4757 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 4758 if (oldadp != NULL && oldadp->ad_offset == off) 4759 allocdirect_merge(adphead, adp, oldadp); 4760 FREE_LOCK(&lk); 4761 return; 4762 } 4763 TAILQ_FOREACH(oldadp, adphead, ad_next) { 4764 if (oldadp->ad_offset >= off) 4765 break; 4766 } 4767 if (oldadp == NULL) 4768 panic("softdep_setup_allocext: lost entry"); 4769 /* insert in middle of list */ 4770 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 4771 if (oldadp->ad_offset == off) 4772 allocdirect_merge(adphead, adp, oldadp); 4773 FREE_LOCK(&lk); 4774} 4775 4776/* 4777 * Indirect block allocation dependencies. 4778 * 4779 * The same dependencies that exist for a direct block also exist when 4780 * a new block is allocated and pointed to by an entry in a block of 4781 * indirect pointers. The undo/redo states described above are also 4782 * used here. Because an indirect block contains many pointers that 4783 * may have dependencies, a second copy of the entire in-memory indirect 4784 * block is kept. The buffer cache copy is always completely up-to-date. 4785 * The second copy, which is used only as a source for disk writes, 4786 * contains only the safe pointers (i.e., those that have no remaining 4787 * update dependencies). The second copy is freed when all pointers 4788 * are safe. The cache is not allowed to replace indirect blocks with 4789 * pending update dependencies. If a buffer containing an indirect 4790 * block with dependencies is written, these routines will mark it 4791 * dirty again. It can only be successfully written once all the 4792 * dependencies are removed. The ffs_fsync routine in conjunction with 4793 * softdep_sync_metadata work together to get all the dependencies 4794 * removed so that a file can be successfully written to disk. Three 4795 * procedures are used when setting up indirect block pointer 4796 * dependencies. The division is necessary because of the organization 4797 * of the "balloc" routine and because of the distinction between file 4798 * pages and file metadata blocks. 4799 */ 4800 4801/* 4802 * Allocate a new allocindir structure. 4803 */ 4804static struct allocindir * 4805newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 4806 struct inode *ip; /* inode for file being extended */ 4807 int ptrno; /* offset of pointer in indirect block */ 4808 ufs2_daddr_t newblkno; /* disk block number being added */ 4809 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 4810 ufs_lbn_t lbn; 4811{ 4812 struct newblk *newblk; 4813 struct allocindir *aip; 4814 struct freefrag *freefrag; 4815 struct jnewblk *jnewblk; 4816 4817 if (oldblkno) 4818 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn); 4819 else 4820 freefrag = NULL; 4821 ACQUIRE_LOCK(&lk); 4822 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0) 4823 panic("new_allocindir: lost block"); 4824 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 4825 ("newallocindir: newblk already initialized")); 4826 newblk->nb_list.wk_type = D_ALLOCINDIR; 4827 newblk->nb_freefrag = freefrag; 4828 aip = (struct allocindir *)newblk; 4829 aip->ai_offset = ptrno; 4830 aip->ai_oldblkno = oldblkno; 4831 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 4832 jnewblk->jn_ino = ip->i_number; 4833 jnewblk->jn_lbn = lbn; 4834 add_to_journal(&jnewblk->jn_list); 4835 } 4836 if (freefrag && freefrag->ff_jfreefrag != NULL) 4837 add_to_journal(&freefrag->ff_jfreefrag->fr_list); 4838 return (aip); 4839} 4840 4841/* 4842 * Called just before setting an indirect block pointer 4843 * to a newly allocated file page. 4844 */ 4845void 4846softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 4847 struct inode *ip; /* inode for file being extended */ 4848 ufs_lbn_t lbn; /* allocated block number within file */ 4849 struct buf *bp; /* buffer with indirect blk referencing page */ 4850 int ptrno; /* offset of pointer in indirect block */ 4851 ufs2_daddr_t newblkno; /* disk block number being added */ 4852 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 4853 struct buf *nbp; /* buffer holding allocated page */ 4854{ 4855 struct inodedep *inodedep; 4856 struct allocindir *aip; 4857 struct pagedep *pagedep; 4858 struct mount *mp; 4859 4860 if (lbn != nbp->b_lblkno) 4861 panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 4862 lbn, bp->b_lblkno); 4863 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 4864 mp = UFSTOVFS(ip->i_ump); 4865 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 4866 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 4867 /* 4868 * If we are allocating a directory page, then we must 4869 * allocate an associated pagedep to track additions and 4870 * deletions. 4871 */ 4872 if ((ip->i_mode & IFMT) == IFDIR && 4873 pagedep_lookup(mp, ip->i_number, lbn, DEPALLOC, &pagedep) == 0) 4874 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list); 4875 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 4876 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 4877 FREE_LOCK(&lk); 4878} 4879 4880/* 4881 * Called just before setting an indirect block pointer to a 4882 * newly allocated indirect block. 4883 */ 4884void 4885softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 4886 struct buf *nbp; /* newly allocated indirect block */ 4887 struct inode *ip; /* inode for file being extended */ 4888 struct buf *bp; /* indirect block referencing allocated block */ 4889 int ptrno; /* offset of pointer in indirect block */ 4890 ufs2_daddr_t newblkno; /* disk block number being added */ 4891{ 4892 struct inodedep *inodedep; 4893 struct allocindir *aip; 4894 ufs_lbn_t lbn; 4895 4896 lbn = nbp->b_lblkno; 4897 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 4898 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 4899 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep); 4900 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 4901 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 4902 FREE_LOCK(&lk); 4903} 4904 4905static void 4906indirdep_complete(indirdep) 4907 struct indirdep *indirdep; 4908{ 4909 struct allocindir *aip; 4910 4911 LIST_REMOVE(indirdep, ir_next); 4912 indirdep->ir_state &= ~ONDEPLIST; 4913 4914 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 4915 LIST_REMOVE(aip, ai_next); 4916 free_newblk(&aip->ai_block); 4917 } 4918 /* 4919 * If this indirdep is not attached to a buf it was simply waiting 4920 * on completion to clear completehd. free_indirdep() asserts 4921 * that nothing is dangling. 4922 */ 4923 if ((indirdep->ir_state & ONWORKLIST) == 0) 4924 free_indirdep(indirdep); 4925} 4926 4927/* 4928 * Called to finish the allocation of the "aip" allocated 4929 * by one of the two routines above. 4930 */ 4931static void 4932setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 4933 struct buf *bp; /* in-memory copy of the indirect block */ 4934 struct inode *ip; /* inode for file being extended */ 4935 struct inodedep *inodedep; /* Inodedep for ip */ 4936 struct allocindir *aip; /* allocindir allocated by the above routines */ 4937 ufs_lbn_t lbn; /* Logical block number for this block. */ 4938{ 4939 struct worklist *wk; 4940 struct fs *fs; 4941 struct newblk *newblk; 4942 struct indirdep *indirdep, *newindirdep; 4943 struct allocindir *oldaip; 4944 struct freefrag *freefrag; 4945 struct mount *mp; 4946 ufs2_daddr_t blkno; 4947 4948 mp = UFSTOVFS(ip->i_ump); 4949 fs = ip->i_fs; 4950 mtx_assert(&lk, MA_OWNED); 4951 if (bp->b_lblkno >= 0) 4952 panic("setup_allocindir_phase2: not indir blk"); 4953 for (freefrag = NULL, indirdep = NULL, newindirdep = NULL; ; ) { 4954 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 4955 if (wk->wk_type != D_INDIRDEP) 4956 continue; 4957 indirdep = WK_INDIRDEP(wk); 4958 break; 4959 } 4960 if (indirdep == NULL && newindirdep) { 4961 indirdep = newindirdep; 4962 newindirdep = NULL; 4963 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 4964 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, 4965 &newblk)) { 4966 indirdep->ir_state |= ONDEPLIST; 4967 LIST_INSERT_HEAD(&newblk->nb_indirdeps, 4968 indirdep, ir_next); 4969 } else 4970 indirdep->ir_state |= DEPCOMPLETE; 4971 } 4972 if (indirdep) { 4973 aip->ai_indirdep = indirdep; 4974 /* 4975 * Check to see if there is an existing dependency 4976 * for this block. If there is, merge the old 4977 * dependency into the new one. This happens 4978 * as a result of reallocblk only. 4979 */ 4980 if (aip->ai_oldblkno == 0) 4981 oldaip = NULL; 4982 else 4983 4984 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, 4985 ai_next) 4986 if (oldaip->ai_offset == aip->ai_offset) 4987 break; 4988 if (oldaip != NULL) 4989 freefrag = allocindir_merge(aip, oldaip); 4990 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 4991 KASSERT(aip->ai_offset >= 0 && 4992 aip->ai_offset < NINDIR(ip->i_ump->um_fs), 4993 ("setup_allocindir_phase2: Bad offset %d", 4994 aip->ai_offset)); 4995 KASSERT(indirdep->ir_savebp != NULL, 4996 ("setup_allocindir_phase2 NULL ir_savebp")); 4997 if (ip->i_ump->um_fstype == UFS1) 4998 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data) 4999 [aip->ai_offset] = aip->ai_oldblkno; 5000 else 5001 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data) 5002 [aip->ai_offset] = aip->ai_oldblkno; 5003 FREE_LOCK(&lk); 5004 if (freefrag != NULL) 5005 handle_workitem_freefrag(freefrag); 5006 } else 5007 FREE_LOCK(&lk); 5008 if (newindirdep) { 5009 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE; 5010 brelse(newindirdep->ir_savebp); 5011 ACQUIRE_LOCK(&lk); 5012 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP); 5013 if (indirdep) 5014 break; 5015 FREE_LOCK(&lk); 5016 } 5017 if (indirdep) { 5018 ACQUIRE_LOCK(&lk); 5019 break; 5020 } 5021 newindirdep = malloc(sizeof(struct indirdep), 5022 M_INDIRDEP, M_SOFTDEP_FLAGS); 5023 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5024 newindirdep->ir_state = ATTACHED; 5025 if (ip->i_ump->um_fstype == UFS1) 5026 newindirdep->ir_state |= UFS1FMT; 5027 newindirdep->ir_saveddata = NULL; 5028 LIST_INIT(&newindirdep->ir_deplisthd); 5029 LIST_INIT(&newindirdep->ir_donehd); 5030 LIST_INIT(&newindirdep->ir_writehd); 5031 LIST_INIT(&newindirdep->ir_completehd); 5032 LIST_INIT(&newindirdep->ir_jwork); 5033 if (bp->b_blkno == bp->b_lblkno) { 5034 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5035 NULL, NULL); 5036 bp->b_blkno = blkno; 5037 } 5038 newindirdep->ir_savebp = 5039 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5040 BUF_KERNPROC(newindirdep->ir_savebp); 5041 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5042 ACQUIRE_LOCK(&lk); 5043 } 5044} 5045 5046/* 5047 * Merge two allocindirs which refer to the same block. Move newblock 5048 * dependencies and setup the freefrags appropriately. 5049 */ 5050static struct freefrag * 5051allocindir_merge(aip, oldaip) 5052 struct allocindir *aip; 5053 struct allocindir *oldaip; 5054{ 5055 struct newdirblk *newdirblk; 5056 struct freefrag *freefrag; 5057 struct worklist *wk; 5058 5059 if (oldaip->ai_newblkno != aip->ai_oldblkno) 5060 panic("allocindir_merge: blkno"); 5061 aip->ai_oldblkno = oldaip->ai_oldblkno; 5062 freefrag = aip->ai_freefrag; 5063 aip->ai_freefrag = oldaip->ai_freefrag; 5064 oldaip->ai_freefrag = NULL; 5065 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 5066 /* 5067 * If we are tracking a new directory-block allocation, 5068 * move it from the old allocindir to the new allocindir. 5069 */ 5070 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 5071 newdirblk = WK_NEWDIRBLK(wk); 5072 WORKLIST_REMOVE(&newdirblk->db_list); 5073 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 5074 panic("allocindir_merge: extra newdirblk"); 5075 WORKLIST_INSERT(&aip->ai_newdirblk, &newdirblk->db_list); 5076 } 5077 /* 5078 * We can skip journaling for this freefrag and just complete 5079 * any pending journal work for the allocindir that is being 5080 * removed after the freefrag completes. 5081 */ 5082 if (freefrag->ff_jfreefrag) 5083 cancel_jfreefrag(freefrag->ff_jfreefrag); 5084 LIST_REMOVE(oldaip, ai_next); 5085 cancel_newblk(&oldaip->ai_block, &freefrag->ff_jwork); 5086 free_newblk(&oldaip->ai_block); 5087 5088 return (freefrag); 5089} 5090 5091/* 5092 * Block de-allocation dependencies. 5093 * 5094 * When blocks are de-allocated, the on-disk pointers must be nullified before 5095 * the blocks are made available for use by other files. (The true 5096 * requirement is that old pointers must be nullified before new on-disk 5097 * pointers are set. We chose this slightly more stringent requirement to 5098 * reduce complexity.) Our implementation handles this dependency by updating 5099 * the inode (or indirect block) appropriately but delaying the actual block 5100 * de-allocation (i.e., freemap and free space count manipulation) until 5101 * after the updated versions reach stable storage. After the disk is 5102 * updated, the blocks can be safely de-allocated whenever it is convenient. 5103 * This implementation handles only the common case of reducing a file's 5104 * length to zero. Other cases are handled by the conventional synchronous 5105 * write approach. 5106 * 5107 * The ffs implementation with which we worked double-checks 5108 * the state of the block pointers and file size as it reduces 5109 * a file's length. Some of this code is replicated here in our 5110 * soft updates implementation. The freeblks->fb_chkcnt field is 5111 * used to transfer a part of this information to the procedure 5112 * that eventually de-allocates the blocks. 5113 * 5114 * This routine should be called from the routine that shortens 5115 * a file's length, before the inode's size or block pointers 5116 * are modified. It will save the block pointer information for 5117 * later release and zero the inode so that the calling routine 5118 * can release it. 5119 */ 5120void 5121softdep_setup_freeblocks(ip, length, flags) 5122 struct inode *ip; /* The inode whose length is to be reduced */ 5123 off_t length; /* The new length for the file */ 5124 int flags; /* IO_EXT and/or IO_NORMAL */ 5125{ 5126 struct ufs1_dinode *dp1; 5127 struct ufs2_dinode *dp2; 5128 struct freeblks *freeblks; 5129 struct inodedep *inodedep; 5130 struct allocdirect *adp; 5131 struct jfreeblk *jfreeblk; 5132 struct bufobj *bo; 5133 struct vnode *vp; 5134 struct buf *bp; 5135 struct fs *fs; 5136 ufs2_daddr_t extblocks, datablocks; 5137 struct mount *mp; 5138 int i, delay, error; 5139 ufs2_daddr_t blkno; 5140 ufs_lbn_t tmpval; 5141 ufs_lbn_t lbn; 5142 long oldextsize; 5143 long oldsize; 5144 int frags; 5145 int needj; 5146 5147 fs = ip->i_fs; 5148 mp = UFSTOVFS(ip->i_ump); 5149 if (length != 0) 5150 panic("softdep_setup_freeblocks: non-zero length"); 5151 freeblks = malloc(sizeof(struct freeblks), 5152 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 5153 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 5154 LIST_INIT(&freeblks->fb_jfreeblkhd); 5155 LIST_INIT(&freeblks->fb_jwork); 5156 freeblks->fb_state = ATTACHED; 5157 freeblks->fb_uid = ip->i_uid; 5158 freeblks->fb_previousinum = ip->i_number; 5159 freeblks->fb_devvp = ip->i_devvp; 5160 freeblks->fb_chkcnt = 0; 5161 ACQUIRE_LOCK(&lk); 5162 /* 5163 * If we're truncating a removed file that will never be written 5164 * we don't need to journal the block frees. The canceled journals 5165 * for the allocations will suffice. 5166 */ 5167 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5168 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED || 5169 (fs->fs_flags & FS_SUJ) == 0) 5170 needj = 0; 5171 else 5172 needj = 1; 5173 num_freeblkdep++; 5174 FREE_LOCK(&lk); 5175 extblocks = 0; 5176 if (fs->fs_magic == FS_UFS2_MAGIC) 5177 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 5178 datablocks = DIP(ip, i_blocks) - extblocks; 5179 if ((flags & IO_NORMAL) != 0) { 5180 oldsize = ip->i_size; 5181 ip->i_size = 0; 5182 DIP_SET(ip, i_size, 0); 5183 freeblks->fb_chkcnt = datablocks; 5184 for (i = 0; i < NDADDR; i++) { 5185 blkno = DIP(ip, i_db[i]); 5186 DIP_SET(ip, i_db[i], 0); 5187 if (blkno == 0) 5188 continue; 5189 frags = sblksize(fs, oldsize, i); 5190 frags = numfrags(fs, frags); 5191 newfreework(freeblks, NULL, i, blkno, frags, needj); 5192 } 5193 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 5194 i++, tmpval *= NINDIR(fs)) { 5195 blkno = DIP(ip, i_ib[i]); 5196 DIP_SET(ip, i_ib[i], 0); 5197 if (blkno) 5198 newfreework(freeblks, NULL, -lbn - i, blkno, 5199 fs->fs_frag, needj); 5200 lbn += tmpval; 5201 } 5202 /* 5203 * If the file was removed, then the space being freed was 5204 * accounted for then (see softdep_releasefile()). If the 5205 * file is merely being truncated, then we account for it now. 5206 */ 5207 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 5208 UFS_LOCK(ip->i_ump); 5209 fs->fs_pendingblocks += datablocks; 5210 UFS_UNLOCK(ip->i_ump); 5211 } 5212 } 5213 if ((flags & IO_EXT) != 0) { 5214 oldextsize = ip->i_din2->di_extsize; 5215 ip->i_din2->di_extsize = 0; 5216 freeblks->fb_chkcnt += extblocks; 5217 for (i = 0; i < NXADDR; i++) { 5218 blkno = ip->i_din2->di_extb[i]; 5219 ip->i_din2->di_extb[i] = 0; 5220 if (blkno == 0) 5221 continue; 5222 frags = sblksize(fs, oldextsize, i); 5223 frags = numfrags(fs, frags); 5224 newfreework(freeblks, NULL, -1 - i, blkno, frags, 5225 needj); 5226 } 5227 } 5228 if (LIST_EMPTY(&freeblks->fb_jfreeblkhd)) 5229 needj = 0; 5230 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt); 5231 /* 5232 * Push the zero'ed inode to to its disk buffer so that we are free 5233 * to delete its dependencies below. Once the dependencies are gone 5234 * the buffer can be safely released. 5235 */ 5236 if ((error = bread(ip->i_devvp, 5237 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 5238 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 5239 brelse(bp); 5240 softdep_error("softdep_setup_freeblocks", error); 5241 } 5242 if (ip->i_ump->um_fstype == UFS1) { 5243 dp1 = ((struct ufs1_dinode *)bp->b_data + 5244 ino_to_fsbo(fs, ip->i_number)); 5245 ip->i_din1->di_freelink = dp1->di_freelink; 5246 *dp1 = *ip->i_din1; 5247 } else { 5248 dp2 = ((struct ufs2_dinode *)bp->b_data + 5249 ino_to_fsbo(fs, ip->i_number)); 5250 ip->i_din2->di_freelink = dp2->di_freelink; 5251 *dp2 = *ip->i_din2; 5252 } 5253 /* 5254 * Find and eliminate any inode dependencies. 5255 */ 5256 ACQUIRE_LOCK(&lk); 5257 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5258 if ((inodedep->id_state & IOSTARTED) != 0) 5259 panic("softdep_setup_freeblocks: inode busy"); 5260 /* 5261 * Add the freeblks structure to the list of operations that 5262 * must await the zero'ed inode being written to disk. If we 5263 * still have a bitmap dependency (delay == 0), then the inode 5264 * has never been written to disk, so we can process the 5265 * freeblks below once we have deleted the dependencies. 5266 */ 5267 delay = (inodedep->id_state & DEPCOMPLETE); 5268 if (delay) 5269 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 5270 else if (needj) 5271 freeblks->fb_state |= DEPCOMPLETE | COMPLETE; 5272 /* 5273 * Because the file length has been truncated to zero, any 5274 * pending block allocation dependency structures associated 5275 * with this inode are obsolete and can simply be de-allocated. 5276 * We must first merge the two dependency lists to get rid of 5277 * any duplicate freefrag structures, then purge the merged list. 5278 * If we still have a bitmap dependency, then the inode has never 5279 * been written to disk, so we can free any fragments without delay. 5280 */ 5281 if (flags & IO_NORMAL) { 5282 merge_inode_lists(&inodedep->id_newinoupdt, 5283 &inodedep->id_inoupdt); 5284 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 5285 cancel_allocdirect(&inodedep->id_inoupdt, adp, 5286 freeblks, delay); 5287 } 5288 if (flags & IO_EXT) { 5289 merge_inode_lists(&inodedep->id_newextupdt, 5290 &inodedep->id_extupdt); 5291 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 5292 cancel_allocdirect(&inodedep->id_extupdt, adp, 5293 freeblks, delay); 5294 } 5295 LIST_FOREACH(jfreeblk, &freeblks->fb_jfreeblkhd, jf_deps) 5296 add_to_journal(&jfreeblk->jf_list); 5297 5298 FREE_LOCK(&lk); 5299 bdwrite(bp); 5300 /* 5301 * We must wait for any I/O in progress to finish so that 5302 * all potential buffers on the dirty list will be visible. 5303 * Once they are all there, walk the list and get rid of 5304 * any dependencies. 5305 */ 5306 vp = ITOV(ip); 5307 bo = &vp->v_bufobj; 5308 BO_LOCK(bo); 5309 drain_output(vp); 5310restart: 5311 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 5312 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 5313 ((flags & IO_NORMAL) == 0 && 5314 (bp->b_xflags & BX_ALTDATA) == 0)) 5315 continue; 5316 if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL) 5317 goto restart; 5318 BO_UNLOCK(bo); 5319 ACQUIRE_LOCK(&lk); 5320 (void) inodedep_lookup(mp, ip->i_number, 0, &inodedep); 5321 if (deallocate_dependencies(bp, inodedep, freeblks)) 5322 bp->b_flags |= B_INVAL | B_NOCACHE; 5323 FREE_LOCK(&lk); 5324 brelse(bp); 5325 BO_LOCK(bo); 5326 goto restart; 5327 } 5328 BO_UNLOCK(bo); 5329 ACQUIRE_LOCK(&lk); 5330 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 5331 (void) free_inodedep(inodedep); 5332 5333 if (delay) { 5334 freeblks->fb_state |= DEPCOMPLETE; 5335 /* 5336 * If the inode with zeroed block pointers is now on disk 5337 * we can start freeing blocks. Add freeblks to the worklist 5338 * instead of calling handle_workitem_freeblocks directly as 5339 * it is more likely that additional IO is needed to complete 5340 * the request here than in the !delay case. 5341 */ 5342 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 5343 add_to_worklist(&freeblks->fb_list, 1); 5344 } 5345 5346 FREE_LOCK(&lk); 5347 /* 5348 * If the inode has never been written to disk (delay == 0) and 5349 * we're not waiting on any journal writes, then we can process the 5350 * freeblks now that we have deleted the dependencies. 5351 */ 5352 if (!delay && !needj) 5353 handle_workitem_freeblocks(freeblks, 0); 5354} 5355 5356/* 5357 * Reclaim any dependency structures from a buffer that is about to 5358 * be reallocated to a new vnode. The buffer must be locked, thus, 5359 * no I/O completion operations can occur while we are manipulating 5360 * its associated dependencies. The mutex is held so that other I/O's 5361 * associated with related dependencies do not occur. Returns 1 if 5362 * all dependencies were cleared, 0 otherwise. 5363 */ 5364static int 5365deallocate_dependencies(bp, inodedep, freeblks) 5366 struct buf *bp; 5367 struct inodedep *inodedep; 5368 struct freeblks *freeblks; 5369{ 5370 struct worklist *wk; 5371 struct indirdep *indirdep; 5372 struct newdirblk *newdirblk; 5373 struct allocindir *aip; 5374 struct pagedep *pagedep; 5375 struct jremref *jremref; 5376 struct jmvref *jmvref; 5377 struct dirrem *dirrem; 5378 int i; 5379 5380 mtx_assert(&lk, MA_OWNED); 5381 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 5382 switch (wk->wk_type) { 5383 5384 case D_INDIRDEP: 5385 indirdep = WK_INDIRDEP(wk); 5386 if (bp->b_lblkno >= 0 || 5387 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 5388 panic("deallocate_dependencies: not indir"); 5389 cancel_indirdep(indirdep, bp, inodedep, freeblks); 5390 continue; 5391 5392 case D_PAGEDEP: 5393 pagedep = WK_PAGEDEP(wk); 5394 /* 5395 * There should be no directory add dependencies present 5396 * as the directory could not be truncated until all 5397 * children were removed. 5398 */ 5399 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 5400 ("deallocate_dependencies: pendinghd != NULL")); 5401 for (i = 0; i < DAHASHSZ; i++) 5402 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 5403 ("deallocate_dependencies: diraddhd != NULL")); 5404 /* 5405 * Copy any directory remove dependencies to the list 5406 * to be processed after the zero'ed inode is written. 5407 * If the inode has already been written, then they 5408 * can be dumped directly onto the work list. 5409 */ 5410 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 5411 /* 5412 * If there are any dirrems we wait for 5413 * the journal write to complete and 5414 * then restart the buf scan as the lock 5415 * has been dropped. 5416 */ 5417 while ((jremref = 5418 LIST_FIRST(&dirrem->dm_jremrefhd)) 5419 != NULL) { 5420 stat_jwait_filepage++; 5421 jwait(&jremref->jr_list); 5422 return (0); 5423 } 5424 LIST_REMOVE(dirrem, dm_next); 5425 dirrem->dm_dirinum = pagedep->pd_ino; 5426 if (inodedep == NULL || 5427 (inodedep->id_state & ALLCOMPLETE) == 5428 ALLCOMPLETE) { 5429 dirrem->dm_state |= COMPLETE; 5430 add_to_worklist(&dirrem->dm_list, 0); 5431 } else 5432 WORKLIST_INSERT(&inodedep->id_bufwait, 5433 &dirrem->dm_list); 5434 } 5435 if ((pagedep->pd_state & NEWBLOCK) != 0) { 5436 newdirblk = pagedep->pd_newdirblk; 5437 WORKLIST_REMOVE(&newdirblk->db_list); 5438 free_newdirblk(newdirblk); 5439 } 5440 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) 5441 != NULL) { 5442 stat_jwait_filepage++; 5443 jwait(&jmvref->jm_list); 5444 return (0); 5445 } 5446 WORKLIST_REMOVE(&pagedep->pd_list); 5447 LIST_REMOVE(pagedep, pd_hash); 5448 WORKITEM_FREE(pagedep, D_PAGEDEP); 5449 continue; 5450 5451 case D_ALLOCINDIR: 5452 aip = WK_ALLOCINDIR(wk); 5453 cancel_allocindir(aip, inodedep, freeblks); 5454 continue; 5455 5456 case D_ALLOCDIRECT: 5457 case D_INODEDEP: 5458 panic("deallocate_dependencies: Unexpected type %s", 5459 TYPENAME(wk->wk_type)); 5460 /* NOTREACHED */ 5461 5462 default: 5463 panic("deallocate_dependencies: Unknown type %s", 5464 TYPENAME(wk->wk_type)); 5465 /* NOTREACHED */ 5466 } 5467 } 5468 5469 return (1); 5470} 5471 5472/* 5473 * An allocdirect is being canceled due to a truncate. We must make sure 5474 * the journal entry is released in concert with the blkfree that releases 5475 * the storage. Completed journal entries must not be released until the 5476 * space is no longer pointed to by the inode or in the bitmap. 5477 */ 5478static void 5479cancel_allocdirect(adphead, adp, freeblks, delay) 5480 struct allocdirectlst *adphead; 5481 struct allocdirect *adp; 5482 struct freeblks *freeblks; 5483 int delay; 5484{ 5485 struct freework *freework; 5486 struct newblk *newblk; 5487 struct worklist *wk; 5488 ufs_lbn_t lbn; 5489 5490 TAILQ_REMOVE(adphead, adp, ad_next); 5491 newblk = (struct newblk *)adp; 5492 /* 5493 * If the journal hasn't been written the jnewblk must be passed 5494 * to the call to ffs_freeblk that reclaims the space. We accomplish 5495 * this by linking the journal dependency into the freework to be 5496 * freed when freework_freeblock() is called. If the journal has 5497 * been written we can simply reclaim the journal space when the 5498 * freeblks work is complete. 5499 */ 5500 if (newblk->nb_jnewblk == NULL) { 5501 cancel_newblk(newblk, &freeblks->fb_jwork); 5502 goto found; 5503 } 5504 lbn = newblk->nb_jnewblk->jn_lbn; 5505 /* 5506 * Find the correct freework structure so it releases the canceled 5507 * journal when the bitmap is cleared. This preserves rollback 5508 * until the allocation is reverted. 5509 */ 5510 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 5511 freework = WK_FREEWORK(wk); 5512 if (freework->fw_lbn != lbn) 5513 continue; 5514 cancel_newblk(newblk, &freework->fw_jwork); 5515 goto found; 5516 } 5517 panic("cancel_allocdirect: Freework not found for lbn %jd\n", lbn); 5518found: 5519 if (delay) 5520 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait, 5521 &newblk->nb_list); 5522 else 5523 free_newblk(newblk); 5524 return; 5525} 5526 5527 5528static void 5529cancel_newblk(newblk, wkhd) 5530 struct newblk *newblk; 5531 struct workhead *wkhd; 5532{ 5533 struct indirdep *indirdep; 5534 struct allocindir *aip; 5535 5536 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) { 5537 indirdep->ir_state &= ~ONDEPLIST; 5538 LIST_REMOVE(indirdep, ir_next); 5539 /* 5540 * If an indirdep is not on the buf worklist we need to 5541 * free it here as deallocate_dependencies() will never 5542 * find it. These pointers were never visible on disk and 5543 * can be discarded immediately. 5544 */ 5545 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5546 LIST_REMOVE(aip, ai_next); 5547 cancel_newblk(&aip->ai_block, wkhd); 5548 free_newblk(&aip->ai_block); 5549 } 5550 /* 5551 * If this indirdep is not attached to a buf it was simply 5552 * waiting on completion to clear completehd. free_indirdep() 5553 * asserts that nothing is dangling. 5554 */ 5555 if ((indirdep->ir_state & ONWORKLIST) == 0) 5556 free_indirdep(indirdep); 5557 } 5558 if (newblk->nb_state & ONDEPLIST) { 5559 newblk->nb_state &= ~ONDEPLIST; 5560 LIST_REMOVE(newblk, nb_deps); 5561 } 5562 if (newblk->nb_state & ONWORKLIST) 5563 WORKLIST_REMOVE(&newblk->nb_list); 5564 /* 5565 * If the journal entry hasn't been written we hold onto the dep 5566 * until it is safe to free along with the other journal work. 5567 */ 5568 if (newblk->nb_jnewblk != NULL) { 5569 cancel_jnewblk(newblk->nb_jnewblk, wkhd); 5570 newblk->nb_jnewblk = NULL; 5571 } 5572 if (!LIST_EMPTY(&newblk->nb_jwork)) 5573 jwork_move(wkhd, &newblk->nb_jwork); 5574} 5575 5576/* 5577 * Free a newblk. Generate a new freefrag work request if appropriate. 5578 * This must be called after the inode pointer and any direct block pointers 5579 * are valid or fully removed via truncate or frag extension. 5580 */ 5581static void 5582free_newblk(newblk) 5583 struct newblk *newblk; 5584{ 5585 struct indirdep *indirdep; 5586 struct newdirblk *newdirblk; 5587 struct freefrag *freefrag; 5588 struct worklist *wk; 5589 5590 mtx_assert(&lk, MA_OWNED); 5591 if (newblk->nb_state & ONDEPLIST) 5592 LIST_REMOVE(newblk, nb_deps); 5593 if (newblk->nb_state & ONWORKLIST) 5594 WORKLIST_REMOVE(&newblk->nb_list); 5595 LIST_REMOVE(newblk, nb_hash); 5596 if ((freefrag = newblk->nb_freefrag) != NULL) { 5597 freefrag->ff_state |= COMPLETE; 5598 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 5599 add_to_worklist(&freefrag->ff_list, 0); 5600 } 5601 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) { 5602 newdirblk = WK_NEWDIRBLK(wk); 5603 WORKLIST_REMOVE(&newdirblk->db_list); 5604 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 5605 panic("free_newblk: extra newdirblk"); 5606 free_newdirblk(newdirblk); 5607 } 5608 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) { 5609 indirdep->ir_state |= DEPCOMPLETE; 5610 indirdep_complete(indirdep); 5611 } 5612 KASSERT(newblk->nb_jnewblk == NULL, 5613 ("free_newblk; jnewblk %p still attached", newblk->nb_jnewblk)); 5614 handle_jwork(&newblk->nb_jwork); 5615 newblk->nb_list.wk_type = D_NEWBLK; 5616 WORKITEM_FREE(newblk, D_NEWBLK); 5617} 5618 5619/* 5620 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 5621 * This routine must be called with splbio interrupts blocked. 5622 */ 5623static void 5624free_newdirblk(newdirblk) 5625 struct newdirblk *newdirblk; 5626{ 5627 struct pagedep *pagedep; 5628 struct diradd *dap; 5629 struct worklist *wk; 5630 int i; 5631 5632 mtx_assert(&lk, MA_OWNED); 5633 /* 5634 * If the pagedep is still linked onto the directory buffer 5635 * dependency chain, then some of the entries on the 5636 * pd_pendinghd list may not be committed to disk yet. In 5637 * this case, we will simply clear the NEWBLOCK flag and 5638 * let the pd_pendinghd list be processed when the pagedep 5639 * is next written. If the pagedep is no longer on the buffer 5640 * dependency chain, then all the entries on the pd_pending 5641 * list are committed to disk and we can free them here. 5642 */ 5643 pagedep = newdirblk->db_pagedep; 5644 pagedep->pd_state &= ~NEWBLOCK; 5645 if ((pagedep->pd_state & ONWORKLIST) == 0) 5646 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 5647 free_diradd(dap, NULL); 5648 /* 5649 * If no dependencies remain, the pagedep will be freed. 5650 */ 5651 for (i = 0; i < DAHASHSZ; i++) 5652 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 5653 break; 5654 if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0 && 5655 LIST_EMPTY(&pagedep->pd_jmvrefhd)) { 5656 KASSERT(LIST_FIRST(&pagedep->pd_dirremhd) == NULL, 5657 ("free_newdirblk: Freeing non-free pagedep %p", pagedep)); 5658 LIST_REMOVE(pagedep, pd_hash); 5659 WORKITEM_FREE(pagedep, D_PAGEDEP); 5660 } 5661 /* Should only ever be one item in the list. */ 5662 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 5663 WORKLIST_REMOVE(wk); 5664 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 5665 } 5666 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 5667} 5668 5669/* 5670 * Prepare an inode to be freed. The actual free operation is not 5671 * done until the zero'ed inode has been written to disk. 5672 */ 5673void 5674softdep_freefile(pvp, ino, mode) 5675 struct vnode *pvp; 5676 ino_t ino; 5677 int mode; 5678{ 5679 struct inode *ip = VTOI(pvp); 5680 struct inodedep *inodedep; 5681 struct freefile *freefile; 5682 5683 /* 5684 * This sets up the inode de-allocation dependency. 5685 */ 5686 freefile = malloc(sizeof(struct freefile), 5687 M_FREEFILE, M_SOFTDEP_FLAGS); 5688 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 5689 freefile->fx_mode = mode; 5690 freefile->fx_oldinum = ino; 5691 freefile->fx_devvp = ip->i_devvp; 5692 LIST_INIT(&freefile->fx_jwork); 5693 if ((ip->i_flag & IN_SPACECOUNTED) == 0) { 5694 UFS_LOCK(ip->i_ump); 5695 ip->i_fs->fs_pendinginodes += 1; 5696 UFS_UNLOCK(ip->i_ump); 5697 } 5698 5699 /* 5700 * If the inodedep does not exist, then the zero'ed inode has 5701 * been written to disk. If the allocated inode has never been 5702 * written to disk, then the on-disk inode is zero'ed. In either 5703 * case we can free the file immediately. If the journal was 5704 * canceled before being written the inode will never make it to 5705 * disk and we must send the canceled journal entrys to 5706 * ffs_freefile() to be cleared in conjunction with the bitmap. 5707 * Any blocks waiting on the inode to write can be safely freed 5708 * here as it will never been written. 5709 */ 5710 ACQUIRE_LOCK(&lk); 5711 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 5712 /* 5713 * Remove this inode from the unlinked list and set 5714 * GOINGAWAY as appropriate to indicate that this inode 5715 * will never be written. 5716 */ 5717 if (inodedep && inodedep->id_state & UNLINKED) { 5718 /* 5719 * Save the journal work to be freed with the bitmap 5720 * before we clear UNLINKED. Otherwise it can be lost 5721 * if the inode block is written. 5722 */ 5723 handle_bufwait(inodedep, &freefile->fx_jwork); 5724 clear_unlinked_inodedep(inodedep); 5725 /* Re-acquire inodedep as we've dropped lk. */ 5726 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 5727 if (inodedep && (inodedep->id_state & DEPCOMPLETE) == 0) 5728 inodedep->id_state |= GOINGAWAY; 5729 } 5730 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 5731 FREE_LOCK(&lk); 5732 handle_workitem_freefile(freefile); 5733 return; 5734 } 5735 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 5736 FREE_LOCK(&lk); 5737 if (ip->i_number == ino) 5738 ip->i_flag |= IN_MODIFIED; 5739} 5740 5741/* 5742 * Check to see if an inode has never been written to disk. If 5743 * so free the inodedep and return success, otherwise return failure. 5744 * This routine must be called with splbio interrupts blocked. 5745 * 5746 * If we still have a bitmap dependency, then the inode has never 5747 * been written to disk. Drop the dependency as it is no longer 5748 * necessary since the inode is being deallocated. We set the 5749 * ALLCOMPLETE flags since the bitmap now properly shows that the 5750 * inode is not allocated. Even if the inode is actively being 5751 * written, it has been rolled back to its zero'ed state, so we 5752 * are ensured that a zero inode is what is on the disk. For short 5753 * lived files, this change will usually result in removing all the 5754 * dependencies from the inode so that it can be freed immediately. 5755 */ 5756static int 5757check_inode_unwritten(inodedep) 5758 struct inodedep *inodedep; 5759{ 5760 5761 mtx_assert(&lk, MA_OWNED); 5762 5763 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 5764 !LIST_EMPTY(&inodedep->id_pendinghd) || 5765 !LIST_EMPTY(&inodedep->id_bufwait) || 5766 !LIST_EMPTY(&inodedep->id_inowait) || 5767 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 5768 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 5769 !TAILQ_EMPTY(&inodedep->id_extupdt) || 5770 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 5771 inodedep->id_mkdiradd != NULL || 5772 inodedep->id_nlinkdelta != 0) 5773 return (0); 5774 /* 5775 * Another process might be in initiate_write_inodeblock_ufs[12] 5776 * trying to allocate memory without holding "Softdep Lock". 5777 */ 5778 if ((inodedep->id_state & IOSTARTED) != 0 && 5779 inodedep->id_savedino1 == NULL) 5780 return (0); 5781 5782 if (inodedep->id_state & ONDEPLIST) 5783 LIST_REMOVE(inodedep, id_deps); 5784 inodedep->id_state &= ~ONDEPLIST; 5785 inodedep->id_state |= ALLCOMPLETE; 5786 inodedep->id_bmsafemap = NULL; 5787 if (inodedep->id_state & ONWORKLIST) 5788 WORKLIST_REMOVE(&inodedep->id_list); 5789 if (inodedep->id_savedino1 != NULL) { 5790 free(inodedep->id_savedino1, M_SAVEDINO); 5791 inodedep->id_savedino1 = NULL; 5792 } 5793 if (free_inodedep(inodedep) == 0) 5794 panic("check_inode_unwritten: busy inode"); 5795 return (1); 5796} 5797 5798/* 5799 * Try to free an inodedep structure. Return 1 if it could be freed. 5800 */ 5801static int 5802free_inodedep(inodedep) 5803 struct inodedep *inodedep; 5804{ 5805 5806 mtx_assert(&lk, MA_OWNED); 5807 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 5808 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 5809 !LIST_EMPTY(&inodedep->id_dirremhd) || 5810 !LIST_EMPTY(&inodedep->id_pendinghd) || 5811 !LIST_EMPTY(&inodedep->id_bufwait) || 5812 !LIST_EMPTY(&inodedep->id_inowait) || 5813 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 5814 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 5815 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 5816 !TAILQ_EMPTY(&inodedep->id_extupdt) || 5817 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 5818 inodedep->id_mkdiradd != NULL || 5819 inodedep->id_nlinkdelta != 0 || 5820 inodedep->id_savedino1 != NULL) 5821 return (0); 5822 if (inodedep->id_state & ONDEPLIST) 5823 LIST_REMOVE(inodedep, id_deps); 5824 LIST_REMOVE(inodedep, id_hash); 5825 WORKITEM_FREE(inodedep, D_INODEDEP); 5826 num_inodedep -= 1; 5827 return (1); 5828} 5829 5830/* 5831 * Free the block referenced by a freework structure. The parent freeblks 5832 * structure is released and completed when the final cg bitmap reaches 5833 * the disk. This routine may be freeing a jnewblk which never made it to 5834 * disk in which case we do not have to wait as the operation is undone 5835 * in memory immediately. 5836 */ 5837static void 5838freework_freeblock(freework) 5839 struct freework *freework; 5840{ 5841 struct freeblks *freeblks; 5842 struct ufsmount *ump; 5843 struct workhead wkhd; 5844 struct fs *fs; 5845 int complete; 5846 int pending; 5847 int bsize; 5848 int needj; 5849 5850 freeblks = freework->fw_freeblks; 5851 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 5852 fs = ump->um_fs; 5853 needj = freeblks->fb_list.wk_mp->mnt_kern_flag & MNTK_SUJ; 5854 complete = 0; 5855 LIST_INIT(&wkhd); 5856 /* 5857 * If we are canceling an existing jnewblk pass it to the free 5858 * routine, otherwise pass the freeblk which will ultimately 5859 * release the freeblks. If we're not journaling, we can just 5860 * free the freeblks immediately. 5861 */ 5862 if (!LIST_EMPTY(&freework->fw_jwork)) { 5863 LIST_SWAP(&wkhd, &freework->fw_jwork, worklist, wk_list); 5864 complete = 1; 5865 } else if (needj) 5866 WORKLIST_INSERT_UNLOCKED(&wkhd, &freework->fw_list); 5867 bsize = lfragtosize(fs, freework->fw_frags); 5868 pending = btodb(bsize); 5869 ACQUIRE_LOCK(&lk); 5870 freeblks->fb_chkcnt -= pending; 5871 FREE_LOCK(&lk); 5872 /* 5873 * extattr blocks don't show up in pending blocks. XXX why? 5874 */ 5875 if (freework->fw_lbn >= 0 || freework->fw_lbn <= -NDADDR) { 5876 UFS_LOCK(ump); 5877 fs->fs_pendingblocks -= pending; 5878 UFS_UNLOCK(ump); 5879 } 5880 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, 5881 bsize, freeblks->fb_previousinum, &wkhd); 5882 if (complete == 0 && needj) 5883 return; 5884 /* 5885 * The jnewblk will be discarded and the bits in the map never 5886 * made it to disk. We can immediately free the freeblk. 5887 */ 5888 ACQUIRE_LOCK(&lk); 5889 handle_written_freework(freework); 5890 FREE_LOCK(&lk); 5891} 5892 5893/* 5894 * Start, continue, or finish the process of freeing an indirect block tree. 5895 * The free operation may be paused at any point with fw_off containing the 5896 * offset to restart from. This enables us to implement some flow control 5897 * for large truncates which may fan out and generate a huge number of 5898 * dependencies. 5899 */ 5900static void 5901handle_workitem_indirblk(freework) 5902 struct freework *freework; 5903{ 5904 struct freeblks *freeblks; 5905 struct ufsmount *ump; 5906 struct fs *fs; 5907 5908 5909 freeblks = freework->fw_freeblks; 5910 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 5911 fs = ump->um_fs; 5912 if (freework->fw_off == NINDIR(fs)) 5913 freework_freeblock(freework); 5914 else 5915 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 5916 freework->fw_lbn); 5917} 5918 5919/* 5920 * Called when a freework structure attached to a cg buf is written. The 5921 * ref on either the parent or the freeblks structure is released and 5922 * either may be added to the worklist if it is the final ref. 5923 */ 5924static void 5925handle_written_freework(freework) 5926 struct freework *freework; 5927{ 5928 struct freeblks *freeblks; 5929 struct freework *parent; 5930 5931 freeblks = freework->fw_freeblks; 5932 parent = freework->fw_parent; 5933 if (parent) { 5934 if (--parent->fw_ref != 0) 5935 parent = NULL; 5936 freeblks = NULL; 5937 } else if (--freeblks->fb_ref != 0) 5938 freeblks = NULL; 5939 WORKITEM_FREE(freework, D_FREEWORK); 5940 /* 5941 * Don't delay these block frees or it takes an intolerable amount 5942 * of time to process truncates and free their journal entries. 5943 */ 5944 if (freeblks) 5945 add_to_worklist(&freeblks->fb_list, 1); 5946 if (parent) 5947 add_to_worklist(&parent->fw_list, 1); 5948} 5949 5950/* 5951 * This workitem routine performs the block de-allocation. 5952 * The workitem is added to the pending list after the updated 5953 * inode block has been written to disk. As mentioned above, 5954 * checks regarding the number of blocks de-allocated (compared 5955 * to the number of blocks allocated for the file) are also 5956 * performed in this function. 5957 */ 5958static void 5959handle_workitem_freeblocks(freeblks, flags) 5960 struct freeblks *freeblks; 5961 int flags; 5962{ 5963 struct freework *freework; 5964 struct worklist *wk; 5965 5966 KASSERT(LIST_EMPTY(&freeblks->fb_jfreeblkhd), 5967 ("handle_workitem_freeblocks: Journal entries not written.")); 5968 if (LIST_EMPTY(&freeblks->fb_freeworkhd)) { 5969 handle_complete_freeblocks(freeblks); 5970 return; 5971 } 5972 freeblks->fb_ref++; 5973 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 5974 KASSERT(wk->wk_type == D_FREEWORK, 5975 ("handle_workitem_freeblocks: Unknown type %s", 5976 TYPENAME(wk->wk_type))); 5977 WORKLIST_REMOVE_UNLOCKED(wk); 5978 freework = WK_FREEWORK(wk); 5979 if (freework->fw_lbn <= -NDADDR) 5980 handle_workitem_indirblk(freework); 5981 else 5982 freework_freeblock(freework); 5983 } 5984 ACQUIRE_LOCK(&lk); 5985 if (--freeblks->fb_ref != 0) 5986 freeblks = NULL; 5987 FREE_LOCK(&lk); 5988 if (freeblks) 5989 handle_complete_freeblocks(freeblks); 5990} 5991 5992/* 5993 * Once all of the freework workitems are complete we can retire the 5994 * freeblocks dependency and any journal work awaiting completion. This 5995 * can not be called until all other dependencies are stable on disk. 5996 */ 5997static void 5998handle_complete_freeblocks(freeblks) 5999 struct freeblks *freeblks; 6000{ 6001 struct inode *ip; 6002 struct vnode *vp; 6003 struct fs *fs; 6004 struct ufsmount *ump; 6005 int flags; 6006 6007 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 6008 fs = ump->um_fs; 6009 flags = LK_NOWAIT; 6010 6011 /* 6012 * If we still have not finished background cleanup, then check 6013 * to see if the block count needs to be adjusted. 6014 */ 6015 if (freeblks->fb_chkcnt != 0 && (fs->fs_flags & FS_UNCLEAN) != 0 && 6016 ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_previousinum, 6017 (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ) == 0) { 6018 ip = VTOI(vp); 6019 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + freeblks->fb_chkcnt); 6020 ip->i_flag |= IN_CHANGE; 6021 vput(vp); 6022 } 6023 6024#ifdef INVARIANTS 6025 if (freeblks->fb_chkcnt != 0 && 6026 ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0)) 6027 printf("handle_workitem_freeblocks: block count\n"); 6028#endif /* INVARIANTS */ 6029 6030 ACQUIRE_LOCK(&lk); 6031 /* 6032 * All of the freeblock deps must be complete prior to this call 6033 * so it's now safe to complete earlier outstanding journal entries. 6034 */ 6035 handle_jwork(&freeblks->fb_jwork); 6036 WORKITEM_FREE(freeblks, D_FREEBLKS); 6037 num_freeblkdep--; 6038 FREE_LOCK(&lk); 6039} 6040 6041/* 6042 * Release blocks associated with the inode ip and stored in the indirect 6043 * block dbn. If level is greater than SINGLE, the block is an indirect block 6044 * and recursive calls to indirtrunc must be used to cleanse other indirect 6045 * blocks. 6046 */ 6047static void 6048indir_trunc(freework, dbn, lbn) 6049 struct freework *freework; 6050 ufs2_daddr_t dbn; 6051 ufs_lbn_t lbn; 6052{ 6053 struct freework *nfreework; 6054 struct workhead wkhd; 6055 struct jnewblk *jnewblk; 6056 struct freeblks *freeblks; 6057 struct buf *bp; 6058 struct fs *fs; 6059 struct worklist *wkn; 6060 struct worklist *wk; 6061 struct indirdep *indirdep; 6062 struct ufsmount *ump; 6063 ufs1_daddr_t *bap1 = 0; 6064 ufs2_daddr_t nb, nnb, *bap2 = 0; 6065 ufs_lbn_t lbnadd; 6066 int i, nblocks, ufs1fmt; 6067 int fs_pendingblocks; 6068 int freedeps; 6069 int needj; 6070 int level; 6071 int cnt; 6072 6073 LIST_INIT(&wkhd); 6074 level = lbn_level(lbn); 6075 if (level == -1) 6076 panic("indir_trunc: Invalid lbn %jd\n", lbn); 6077 freeblks = freework->fw_freeblks; 6078 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 6079 fs = ump->um_fs; 6080 fs_pendingblocks = 0; 6081 freedeps = 0; 6082 needj = UFSTOVFS(ump)->mnt_kern_flag & MNTK_SUJ; 6083 lbnadd = 1; 6084 for (i = level; i > 0; i--) 6085 lbnadd *= NINDIR(fs); 6086 /* 6087 * Get buffer of block pointers to be freed. This routine is not 6088 * called until the zero'ed inode has been written, so it is safe 6089 * to free blocks as they are encountered. Because the inode has 6090 * been zero'ed, calls to bmap on these blocks will fail. So, we 6091 * have to use the on-disk address and the block device for the 6092 * filesystem to look them up. If the file was deleted before its 6093 * indirect blocks were all written to disk, the routine that set 6094 * us up (deallocate_dependencies) will have arranged to leave 6095 * a complete copy of the indirect block in memory for our use. 6096 * Otherwise we have to read the blocks in from the disk. 6097 */ 6098#ifdef notyet 6099 bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0, 6100 GB_NOCREAT); 6101#else 6102 bp = incore(&freeblks->fb_devvp->v_bufobj, dbn); 6103#endif 6104 ACQUIRE_LOCK(&lk); 6105 if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) { 6106 if (wk->wk_type != D_INDIRDEP || 6107 (wk->wk_state & GOINGAWAY) == 0) 6108 panic("indir_trunc: lost indirdep %p", wk); 6109 indirdep = WK_INDIRDEP(wk); 6110 LIST_SWAP(&wkhd, &indirdep->ir_jwork, worklist, wk_list); 6111 free_indirdep(indirdep); 6112 if (!LIST_EMPTY(&bp->b_dep)) 6113 panic("indir_trunc: dangling dep %p", 6114 LIST_FIRST(&bp->b_dep)); 6115 ump->um_numindirdeps -= 1; 6116 FREE_LOCK(&lk); 6117 } else { 6118#ifdef notyet 6119 if (bp) 6120 brelse(bp); 6121#endif 6122 FREE_LOCK(&lk); 6123 if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 6124 NOCRED, &bp) != 0) { 6125 brelse(bp); 6126 return; 6127 } 6128 } 6129 /* 6130 * Recursively free indirect blocks. 6131 */ 6132 if (ump->um_fstype == UFS1) { 6133 ufs1fmt = 1; 6134 bap1 = (ufs1_daddr_t *)bp->b_data; 6135 } else { 6136 ufs1fmt = 0; 6137 bap2 = (ufs2_daddr_t *)bp->b_data; 6138 } 6139 /* 6140 * Reclaim indirect blocks which never made it to disk. 6141 */ 6142 cnt = 0; 6143 LIST_FOREACH_SAFE(wk, &wkhd, wk_list, wkn) { 6144 struct workhead freewk; 6145 if (wk->wk_type != D_JNEWBLK) 6146 continue; 6147 WORKLIST_REMOVE_UNLOCKED(wk); 6148 LIST_INIT(&freewk); 6149 WORKLIST_INSERT_UNLOCKED(&freewk, wk); 6150 jnewblk = WK_JNEWBLK(wk); 6151 if (jnewblk->jn_lbn > 0) 6152 i = (jnewblk->jn_lbn - -lbn) / lbnadd; 6153 else 6154 i = (jnewblk->jn_lbn - (lbn + 1)) / lbnadd; 6155 KASSERT(i >= 0 && i < NINDIR(fs), 6156 ("indir_trunc: Index out of range %d parent %jd lbn %jd", 6157 i, lbn, jnewblk->jn_lbn)); 6158 /* Clear the pointer so it isn't found below. */ 6159 if (ufs1fmt) { 6160 nb = bap1[i]; 6161 bap1[i] = 0; 6162 } else { 6163 nb = bap2[i]; 6164 bap2[i] = 0; 6165 } 6166 KASSERT(nb == jnewblk->jn_blkno, 6167 ("indir_trunc: Block mismatch %jd != %jd", 6168 nb, jnewblk->jn_blkno)); 6169 ffs_blkfree(ump, fs, freeblks->fb_devvp, jnewblk->jn_blkno, 6170 fs->fs_bsize, freeblks->fb_previousinum, &freewk); 6171 cnt++; 6172 } 6173 ACQUIRE_LOCK(&lk); 6174 if (needj) 6175 freework->fw_ref += NINDIR(fs) + 1; 6176 /* Any remaining journal work can be completed with freeblks. */ 6177 jwork_move(&freeblks->fb_jwork, &wkhd); 6178 FREE_LOCK(&lk); 6179 nblocks = btodb(fs->fs_bsize); 6180 if (ufs1fmt) 6181 nb = bap1[0]; 6182 else 6183 nb = bap2[0]; 6184 nfreework = freework; 6185 /* 6186 * Reclaim on disk blocks. 6187 */ 6188 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 6189 if (i != NINDIR(fs) - 1) { 6190 if (ufs1fmt) 6191 nnb = bap1[i+1]; 6192 else 6193 nnb = bap2[i+1]; 6194 } else 6195 nnb = 0; 6196 if (nb == 0) 6197 continue; 6198 cnt++; 6199 if (level != 0) { 6200 ufs_lbn_t nlbn; 6201 6202 nlbn = (lbn + 1) - (i * lbnadd); 6203 if (needj != 0) { 6204 nfreework = newfreework(freeblks, freework, 6205 nlbn, nb, fs->fs_frag, 0); 6206 freedeps++; 6207 } 6208 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 6209 } else { 6210 struct freedep *freedep; 6211 6212 /* 6213 * Attempt to aggregate freedep dependencies for 6214 * all blocks being released to the same CG. 6215 */ 6216 LIST_INIT(&wkhd); 6217 if (needj != 0 && 6218 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 6219 freedep = newfreedep(freework); 6220 WORKLIST_INSERT_UNLOCKED(&wkhd, 6221 &freedep->fd_list); 6222 freedeps++; 6223 } 6224 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 6225 fs->fs_bsize, freeblks->fb_previousinum, &wkhd); 6226 } 6227 } 6228 if (level == 0) 6229 fs_pendingblocks = (nblocks * cnt); 6230 /* 6231 * If we're not journaling we can free the indirect now. Otherwise 6232 * setup the ref counts and offset so this indirect can be completed 6233 * when its children are free. 6234 */ 6235 if (needj == 0) { 6236 fs_pendingblocks += nblocks; 6237 dbn = dbtofsb(fs, dbn); 6238 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 6239 freeblks->fb_previousinum, NULL); 6240 ACQUIRE_LOCK(&lk); 6241 freeblks->fb_chkcnt -= fs_pendingblocks; 6242 if (freework->fw_blkno == dbn) 6243 handle_written_freework(freework); 6244 FREE_LOCK(&lk); 6245 freework = NULL; 6246 } else { 6247 ACQUIRE_LOCK(&lk); 6248 freework->fw_off = i; 6249 freework->fw_ref += freedeps; 6250 freework->fw_ref -= NINDIR(fs) + 1; 6251 if (freework->fw_ref != 0) 6252 freework = NULL; 6253 freeblks->fb_chkcnt -= fs_pendingblocks; 6254 FREE_LOCK(&lk); 6255 } 6256 if (fs_pendingblocks) { 6257 UFS_LOCK(ump); 6258 fs->fs_pendingblocks -= fs_pendingblocks; 6259 UFS_UNLOCK(ump); 6260 } 6261 bp->b_flags |= B_INVAL | B_NOCACHE; 6262 brelse(bp); 6263 if (freework) 6264 handle_workitem_indirblk(freework); 6265 return; 6266} 6267 6268/* 6269 * Cancel an allocindir when it is removed via truncation. 6270 */ 6271static void 6272cancel_allocindir(aip, inodedep, freeblks) 6273 struct allocindir *aip; 6274 struct inodedep *inodedep; 6275 struct freeblks *freeblks; 6276{ 6277 struct newblk *newblk; 6278 6279 /* 6280 * If the journal hasn't been written the jnewblk must be passed 6281 * to the call to ffs_freeblk that reclaims the space. We accomplish 6282 * this by linking the journal dependency into the indirdep to be 6283 * freed when indir_trunc() is called. If the journal has already 6284 * been written we can simply reclaim the journal space when the 6285 * freeblks work is complete. 6286 */ 6287 LIST_REMOVE(aip, ai_next); 6288 newblk = (struct newblk *)aip; 6289 if (newblk->nb_jnewblk == NULL) 6290 cancel_newblk(newblk, &freeblks->fb_jwork); 6291 else 6292 cancel_newblk(newblk, &aip->ai_indirdep->ir_jwork); 6293 if (inodedep && inodedep->id_state & DEPCOMPLETE) 6294 WORKLIST_INSERT(&inodedep->id_bufwait, &newblk->nb_list); 6295 else 6296 free_newblk(newblk); 6297} 6298 6299/* 6300 * Create the mkdir dependencies for . and .. in a new directory. Link them 6301 * in to a newdirblk so any subsequent additions are tracked properly. The 6302 * caller is responsible for adding the mkdir1 dependency to the journal 6303 * and updating id_mkdiradd. This function returns with lk held. 6304 */ 6305static struct mkdir * 6306setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 6307 struct diradd *dap; 6308 ino_t newinum; 6309 ino_t dinum; 6310 struct buf *newdirbp; 6311 struct mkdir **mkdirp; 6312{ 6313 struct newblk *newblk; 6314 struct pagedep *pagedep; 6315 struct inodedep *inodedep; 6316 struct newdirblk *newdirblk = 0; 6317 struct mkdir *mkdir1, *mkdir2; 6318 struct worklist *wk; 6319 struct jaddref *jaddref; 6320 struct mount *mp; 6321 6322 mp = dap->da_list.wk_mp; 6323 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 6324 M_SOFTDEP_FLAGS); 6325 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 6326 LIST_INIT(&newdirblk->db_mkdir); 6327 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 6328 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 6329 mkdir1->md_state = ATTACHED | MKDIR_BODY; 6330 mkdir1->md_diradd = dap; 6331 mkdir1->md_jaddref = NULL; 6332 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 6333 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 6334 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 6335 mkdir2->md_diradd = dap; 6336 mkdir2->md_jaddref = NULL; 6337 if ((mp->mnt_kern_flag & MNTK_SUJ) == 0) { 6338 mkdir1->md_state |= DEPCOMPLETE; 6339 mkdir2->md_state |= DEPCOMPLETE; 6340 } 6341 /* 6342 * Dependency on "." and ".." being written to disk. 6343 */ 6344 mkdir1->md_buf = newdirbp; 6345 ACQUIRE_LOCK(&lk); 6346 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 6347 /* 6348 * We must link the pagedep, allocdirect, and newdirblk for 6349 * the initial file page so the pointer to the new directory 6350 * is not written until the directory contents are live and 6351 * any subsequent additions are not marked live until the 6352 * block is reachable via the inode. 6353 */ 6354 if (pagedep_lookup(mp, newinum, 0, 0, &pagedep) == 0) 6355 panic("setup_newdir: lost pagedep"); 6356 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 6357 if (wk->wk_type == D_ALLOCDIRECT) 6358 break; 6359 if (wk == NULL) 6360 panic("setup_newdir: lost allocdirect"); 6361 newblk = WK_NEWBLK(wk); 6362 pagedep->pd_state |= NEWBLOCK; 6363 pagedep->pd_newdirblk = newdirblk; 6364 newdirblk->db_pagedep = pagedep; 6365 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 6366 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 6367 /* 6368 * Look up the inodedep for the parent directory so that we 6369 * can link mkdir2 into the pending dotdot jaddref or 6370 * the inode write if there is none. If the inode is 6371 * ALLCOMPLETE and no jaddref is present all dependencies have 6372 * been satisfied and mkdir2 can be freed. 6373 */ 6374 inodedep_lookup(mp, dinum, 0, &inodedep); 6375 if (mp->mnt_kern_flag & MNTK_SUJ) { 6376 if (inodedep == NULL) 6377 panic("setup_newdir: Lost parent."); 6378 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 6379 inoreflst); 6380 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 6381 (jaddref->ja_state & MKDIR_PARENT), 6382 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 6383 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 6384 mkdir2->md_jaddref = jaddref; 6385 jaddref->ja_mkdir = mkdir2; 6386 } else if (inodedep == NULL || 6387 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 6388 dap->da_state &= ~MKDIR_PARENT; 6389 WORKITEM_FREE(mkdir2, D_MKDIR); 6390 } else { 6391 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 6392 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list); 6393 } 6394 *mkdirp = mkdir2; 6395 6396 return (mkdir1); 6397} 6398 6399/* 6400 * Directory entry addition dependencies. 6401 * 6402 * When adding a new directory entry, the inode (with its incremented link 6403 * count) must be written to disk before the directory entry's pointer to it. 6404 * Also, if the inode is newly allocated, the corresponding freemap must be 6405 * updated (on disk) before the directory entry's pointer. These requirements 6406 * are met via undo/redo on the directory entry's pointer, which consists 6407 * simply of the inode number. 6408 * 6409 * As directory entries are added and deleted, the free space within a 6410 * directory block can become fragmented. The ufs filesystem will compact 6411 * a fragmented directory block to make space for a new entry. When this 6412 * occurs, the offsets of previously added entries change. Any "diradd" 6413 * dependency structures corresponding to these entries must be updated with 6414 * the new offsets. 6415 */ 6416 6417/* 6418 * This routine is called after the in-memory inode's link 6419 * count has been incremented, but before the directory entry's 6420 * pointer to the inode has been set. 6421 */ 6422int 6423softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 6424 struct buf *bp; /* buffer containing directory block */ 6425 struct inode *dp; /* inode for directory */ 6426 off_t diroffset; /* offset of new entry in directory */ 6427 ino_t newinum; /* inode referenced by new directory entry */ 6428 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 6429 int isnewblk; /* entry is in a newly allocated block */ 6430{ 6431 int offset; /* offset of new entry within directory block */ 6432 ufs_lbn_t lbn; /* block in directory containing new entry */ 6433 struct fs *fs; 6434 struct diradd *dap; 6435 struct newblk *newblk; 6436 struct pagedep *pagedep; 6437 struct inodedep *inodedep; 6438 struct newdirblk *newdirblk = 0; 6439 struct mkdir *mkdir1, *mkdir2; 6440 struct jaddref *jaddref; 6441 struct mount *mp; 6442 int isindir; 6443 6444 /* 6445 * Whiteouts have no dependencies. 6446 */ 6447 if (newinum == WINO) { 6448 if (newdirbp != NULL) 6449 bdwrite(newdirbp); 6450 return (0); 6451 } 6452 jaddref = NULL; 6453 mkdir1 = mkdir2 = NULL; 6454 mp = UFSTOVFS(dp->i_ump); 6455 fs = dp->i_fs; 6456 lbn = lblkno(fs, diroffset); 6457 offset = blkoff(fs, diroffset); 6458 dap = malloc(sizeof(struct diradd), M_DIRADD, 6459 M_SOFTDEP_FLAGS|M_ZERO); 6460 workitem_alloc(&dap->da_list, D_DIRADD, mp); 6461 dap->da_offset = offset; 6462 dap->da_newinum = newinum; 6463 dap->da_state = ATTACHED; 6464 LIST_INIT(&dap->da_jwork); 6465 isindir = bp->b_lblkno >= NDADDR; 6466 if (isnewblk && 6467 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 6468 newdirblk = malloc(sizeof(struct newdirblk), 6469 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 6470 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 6471 LIST_INIT(&newdirblk->db_mkdir); 6472 } 6473 /* 6474 * If we're creating a new directory setup the dependencies and set 6475 * the dap state to wait for them. Otherwise it's COMPLETE and 6476 * we can move on. 6477 */ 6478 if (newdirbp == NULL) { 6479 dap->da_state |= DEPCOMPLETE; 6480 ACQUIRE_LOCK(&lk); 6481 } else { 6482 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 6483 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 6484 &mkdir2); 6485 } 6486 /* 6487 * Link into parent directory pagedep to await its being written. 6488 */ 6489 if (pagedep_lookup(mp, dp->i_number, lbn, DEPALLOC, &pagedep) == 0) 6490 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 6491#ifdef DEBUG 6492 if (diradd_lookup(pagedep, offset) != NULL) 6493 panic("softdep_setup_directory_add: %p already at off %d\n", 6494 diradd_lookup(pagedep, offset), offset); 6495#endif 6496 dap->da_pagedep = pagedep; 6497 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 6498 da_pdlist); 6499 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 6500 /* 6501 * If we're journaling, link the diradd into the jaddref so it 6502 * may be completed after the journal entry is written. Otherwise, 6503 * link the diradd into its inodedep. If the inode is not yet 6504 * written place it on the bufwait list, otherwise do the post-inode 6505 * write processing to put it on the id_pendinghd list. 6506 */ 6507 if (mp->mnt_kern_flag & MNTK_SUJ) { 6508 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 6509 inoreflst); 6510 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 6511 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 6512 jaddref->ja_diroff = diroffset; 6513 jaddref->ja_diradd = dap; 6514 add_to_journal(&jaddref->ja_list); 6515 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 6516 diradd_inode_written(dap, inodedep); 6517 else 6518 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 6519 /* 6520 * Add the journal entries for . and .. links now that the primary 6521 * link is written. 6522 */ 6523 if (mkdir1 != NULL && mp->mnt_kern_flag & MNTK_SUJ) { 6524 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 6525 inoreflst, if_deps); 6526 KASSERT(jaddref != NULL && 6527 jaddref->ja_ino == jaddref->ja_parent && 6528 (jaddref->ja_state & MKDIR_BODY), 6529 ("softdep_setup_directory_add: bad dot jaddref %p", 6530 jaddref)); 6531 mkdir1->md_jaddref = jaddref; 6532 jaddref->ja_mkdir = mkdir1; 6533 /* 6534 * It is important that the dotdot journal entry 6535 * is added prior to the dot entry since dot writes 6536 * both the dot and dotdot links. These both must 6537 * be added after the primary link for the journal 6538 * to remain consistent. 6539 */ 6540 add_to_journal(&mkdir2->md_jaddref->ja_list); 6541 add_to_journal(&jaddref->ja_list); 6542 } 6543 /* 6544 * If we are adding a new directory remember this diradd so that if 6545 * we rename it we can keep the dot and dotdot dependencies. If 6546 * we are adding a new name for an inode that has a mkdiradd we 6547 * must be in rename and we have to move the dot and dotdot 6548 * dependencies to this new name. The old name is being orphaned 6549 * soon. 6550 */ 6551 if (mkdir1 != NULL) { 6552 if (inodedep->id_mkdiradd != NULL) 6553 panic("softdep_setup_directory_add: Existing mkdir"); 6554 inodedep->id_mkdiradd = dap; 6555 } else if (inodedep->id_mkdiradd) 6556 merge_diradd(inodedep, dap); 6557 if (newdirblk) { 6558 /* 6559 * There is nothing to do if we are already tracking 6560 * this block. 6561 */ 6562 if ((pagedep->pd_state & NEWBLOCK) != 0) { 6563 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 6564 FREE_LOCK(&lk); 6565 return (0); 6566 } 6567 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 6568 == 0) 6569 panic("softdep_setup_directory_add: lost entry"); 6570 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 6571 pagedep->pd_state |= NEWBLOCK; 6572 pagedep->pd_newdirblk = newdirblk; 6573 newdirblk->db_pagedep = pagedep; 6574 FREE_LOCK(&lk); 6575 /* 6576 * If we extended into an indirect signal direnter to sync. 6577 */ 6578 if (isindir) 6579 return (1); 6580 return (0); 6581 } 6582 FREE_LOCK(&lk); 6583 return (0); 6584} 6585 6586/* 6587 * This procedure is called to change the offset of a directory 6588 * entry when compacting a directory block which must be owned 6589 * exclusively by the caller. Note that the actual entry movement 6590 * must be done in this procedure to ensure that no I/O completions 6591 * occur while the move is in progress. 6592 */ 6593void 6594softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 6595 struct buf *bp; /* Buffer holding directory block. */ 6596 struct inode *dp; /* inode for directory */ 6597 caddr_t base; /* address of dp->i_offset */ 6598 caddr_t oldloc; /* address of old directory location */ 6599 caddr_t newloc; /* address of new directory location */ 6600 int entrysize; /* size of directory entry */ 6601{ 6602 int offset, oldoffset, newoffset; 6603 struct pagedep *pagedep; 6604 struct jmvref *jmvref; 6605 struct diradd *dap; 6606 struct direct *de; 6607 struct mount *mp; 6608 ufs_lbn_t lbn; 6609 int flags; 6610 6611 mp = UFSTOVFS(dp->i_ump); 6612 de = (struct direct *)oldloc; 6613 jmvref = NULL; 6614 flags = 0; 6615 /* 6616 * Moves are always journaled as it would be too complex to 6617 * determine if any affected adds or removes are present in the 6618 * journal. 6619 */ 6620 if (mp->mnt_kern_flag & MNTK_SUJ) { 6621 flags = DEPALLOC; 6622 jmvref = newjmvref(dp, de->d_ino, 6623 dp->i_offset + (oldloc - base), 6624 dp->i_offset + (newloc - base)); 6625 } 6626 lbn = lblkno(dp->i_fs, dp->i_offset); 6627 offset = blkoff(dp->i_fs, dp->i_offset); 6628 oldoffset = offset + (oldloc - base); 6629 newoffset = offset + (newloc - base); 6630 ACQUIRE_LOCK(&lk); 6631 if (pagedep_lookup(mp, dp->i_number, lbn, flags, &pagedep) == 0) { 6632 if (pagedep) 6633 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 6634 goto done; 6635 } 6636 dap = diradd_lookup(pagedep, oldoffset); 6637 if (dap) { 6638 dap->da_offset = newoffset; 6639 newoffset = DIRADDHASH(newoffset); 6640 oldoffset = DIRADDHASH(oldoffset); 6641 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 6642 newoffset != oldoffset) { 6643 LIST_REMOVE(dap, da_pdlist); 6644 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 6645 dap, da_pdlist); 6646 } 6647 } 6648done: 6649 if (jmvref) { 6650 jmvref->jm_pagedep = pagedep; 6651 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 6652 add_to_journal(&jmvref->jm_list); 6653 } 6654 bcopy(oldloc, newloc, entrysize); 6655 FREE_LOCK(&lk); 6656} 6657 6658/* 6659 * Move the mkdir dependencies and journal work from one diradd to another 6660 * when renaming a directory. The new name must depend on the mkdir deps 6661 * completing as the old name did. Directories can only have one valid link 6662 * at a time so one must be canonical. 6663 */ 6664static void 6665merge_diradd(inodedep, newdap) 6666 struct inodedep *inodedep; 6667 struct diradd *newdap; 6668{ 6669 struct diradd *olddap; 6670 struct mkdir *mkdir, *nextmd; 6671 short state; 6672 6673 olddap = inodedep->id_mkdiradd; 6674 inodedep->id_mkdiradd = newdap; 6675 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 6676 newdap->da_state &= ~DEPCOMPLETE; 6677 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 6678 nextmd = LIST_NEXT(mkdir, md_mkdirs); 6679 if (mkdir->md_diradd != olddap) 6680 continue; 6681 mkdir->md_diradd = newdap; 6682 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 6683 newdap->da_state |= state; 6684 olddap->da_state &= ~state; 6685 if ((olddap->da_state & 6686 (MKDIR_PARENT | MKDIR_BODY)) == 0) 6687 break; 6688 } 6689 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 6690 panic("merge_diradd: unfound ref"); 6691 } 6692 /* 6693 * Any mkdir related journal items are not safe to be freed until 6694 * the new name is stable. 6695 */ 6696 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 6697 olddap->da_state |= DEPCOMPLETE; 6698 complete_diradd(olddap); 6699} 6700 6701/* 6702 * Move the diradd to the pending list when all diradd dependencies are 6703 * complete. 6704 */ 6705static void 6706complete_diradd(dap) 6707 struct diradd *dap; 6708{ 6709 struct pagedep *pagedep; 6710 6711 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 6712 if (dap->da_state & DIRCHG) 6713 pagedep = dap->da_previous->dm_pagedep; 6714 else 6715 pagedep = dap->da_pagedep; 6716 LIST_REMOVE(dap, da_pdlist); 6717 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 6718 } 6719} 6720 6721/* 6722 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 6723 * add entries and conditonally journal the remove. 6724 */ 6725static void 6726cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 6727 struct diradd *dap; 6728 struct dirrem *dirrem; 6729 struct jremref *jremref; 6730 struct jremref *dotremref; 6731 struct jremref *dotdotremref; 6732{ 6733 struct inodedep *inodedep; 6734 struct jaddref *jaddref; 6735 struct inoref *inoref; 6736 struct mkdir *mkdir; 6737 6738 /* 6739 * If no remove references were allocated we're on a non-journaled 6740 * filesystem and can skip the cancel step. 6741 */ 6742 if (jremref == NULL) { 6743 free_diradd(dap, NULL); 6744 return; 6745 } 6746 /* 6747 * Cancel the primary name an free it if it does not require 6748 * journaling. 6749 */ 6750 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 6751 0, &inodedep) != 0) { 6752 /* Abort the addref that reference this diradd. */ 6753 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 6754 if (inoref->if_list.wk_type != D_JADDREF) 6755 continue; 6756 jaddref = (struct jaddref *)inoref; 6757 if (jaddref->ja_diradd != dap) 6758 continue; 6759 if (cancel_jaddref(jaddref, inodedep, 6760 &dirrem->dm_jwork) == 0) { 6761 free_jremref(jremref); 6762 jremref = NULL; 6763 } 6764 break; 6765 } 6766 } 6767 /* 6768 * Cancel subordinate names and free them if they do not require 6769 * journaling. 6770 */ 6771 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 6772 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 6773 if (mkdir->md_diradd != dap) 6774 continue; 6775 if ((jaddref = mkdir->md_jaddref) == NULL) 6776 continue; 6777 mkdir->md_jaddref = NULL; 6778 if (mkdir->md_state & MKDIR_PARENT) { 6779 if (cancel_jaddref(jaddref, NULL, 6780 &dirrem->dm_jwork) == 0) { 6781 free_jremref(dotdotremref); 6782 dotdotremref = NULL; 6783 } 6784 } else { 6785 if (cancel_jaddref(jaddref, inodedep, 6786 &dirrem->dm_jwork) == 0) { 6787 free_jremref(dotremref); 6788 dotremref = NULL; 6789 } 6790 } 6791 } 6792 } 6793 6794 if (jremref) 6795 journal_jremref(dirrem, jremref, inodedep); 6796 if (dotremref) 6797 journal_jremref(dirrem, dotremref, inodedep); 6798 if (dotdotremref) 6799 journal_jremref(dirrem, dotdotremref, NULL); 6800 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 6801 free_diradd(dap, &dirrem->dm_jwork); 6802} 6803 6804/* 6805 * Free a diradd dependency structure. This routine must be called 6806 * with splbio interrupts blocked. 6807 */ 6808static void 6809free_diradd(dap, wkhd) 6810 struct diradd *dap; 6811 struct workhead *wkhd; 6812{ 6813 struct dirrem *dirrem; 6814 struct pagedep *pagedep; 6815 struct inodedep *inodedep; 6816 struct mkdir *mkdir, *nextmd; 6817 6818 mtx_assert(&lk, MA_OWNED); 6819 LIST_REMOVE(dap, da_pdlist); 6820 if (dap->da_state & ONWORKLIST) 6821 WORKLIST_REMOVE(&dap->da_list); 6822 if ((dap->da_state & DIRCHG) == 0) { 6823 pagedep = dap->da_pagedep; 6824 } else { 6825 dirrem = dap->da_previous; 6826 pagedep = dirrem->dm_pagedep; 6827 dirrem->dm_dirinum = pagedep->pd_ino; 6828 dirrem->dm_state |= COMPLETE; 6829 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 6830 add_to_worklist(&dirrem->dm_list, 0); 6831 } 6832 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 6833 0, &inodedep) != 0) 6834 if (inodedep->id_mkdiradd == dap) 6835 inodedep->id_mkdiradd = NULL; 6836 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 6837 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 6838 nextmd = LIST_NEXT(mkdir, md_mkdirs); 6839 if (mkdir->md_diradd != dap) 6840 continue; 6841 dap->da_state &= 6842 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 6843 LIST_REMOVE(mkdir, md_mkdirs); 6844 if (mkdir->md_state & ONWORKLIST) 6845 WORKLIST_REMOVE(&mkdir->md_list); 6846 if (mkdir->md_jaddref != NULL) 6847 panic("free_diradd: Unexpected jaddref"); 6848 WORKITEM_FREE(mkdir, D_MKDIR); 6849 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 6850 break; 6851 } 6852 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 6853 panic("free_diradd: unfound ref"); 6854 } 6855 if (inodedep) 6856 free_inodedep(inodedep); 6857 /* 6858 * Free any journal segments waiting for the directory write. 6859 */ 6860 handle_jwork(&dap->da_jwork); 6861 WORKITEM_FREE(dap, D_DIRADD); 6862} 6863 6864/* 6865 * Directory entry removal dependencies. 6866 * 6867 * When removing a directory entry, the entry's inode pointer must be 6868 * zero'ed on disk before the corresponding inode's link count is decremented 6869 * (possibly freeing the inode for re-use). This dependency is handled by 6870 * updating the directory entry but delaying the inode count reduction until 6871 * after the directory block has been written to disk. After this point, the 6872 * inode count can be decremented whenever it is convenient. 6873 */ 6874 6875/* 6876 * This routine should be called immediately after removing 6877 * a directory entry. The inode's link count should not be 6878 * decremented by the calling procedure -- the soft updates 6879 * code will do this task when it is safe. 6880 */ 6881void 6882softdep_setup_remove(bp, dp, ip, isrmdir) 6883 struct buf *bp; /* buffer containing directory block */ 6884 struct inode *dp; /* inode for the directory being modified */ 6885 struct inode *ip; /* inode for directory entry being removed */ 6886 int isrmdir; /* indicates if doing RMDIR */ 6887{ 6888 struct dirrem *dirrem, *prevdirrem; 6889 struct inodedep *inodedep; 6890 int direct; 6891 6892 /* 6893 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 6894 * newdirrem() to setup the full directory remove which requires 6895 * isrmdir > 1. 6896 */ 6897 dirrem = newdirrem(bp, dp, ip, isrmdir?2:0, &prevdirrem); 6898 /* 6899 * Add the dirrem to the inodedep's pending remove list for quick 6900 * discovery later. 6901 */ 6902 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 6903 &inodedep) == 0) 6904 panic("softdep_setup_remove: Lost inodedep."); 6905 dirrem->dm_state |= ONDEPLIST; 6906 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 6907 6908 /* 6909 * If the COMPLETE flag is clear, then there were no active 6910 * entries and we want to roll back to a zeroed entry until 6911 * the new inode is committed to disk. If the COMPLETE flag is 6912 * set then we have deleted an entry that never made it to 6913 * disk. If the entry we deleted resulted from a name change, 6914 * then the old name still resides on disk. We cannot delete 6915 * its inode (returned to us in prevdirrem) until the zeroed 6916 * directory entry gets to disk. The new inode has never been 6917 * referenced on the disk, so can be deleted immediately. 6918 */ 6919 if ((dirrem->dm_state & COMPLETE) == 0) { 6920 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 6921 dm_next); 6922 FREE_LOCK(&lk); 6923 } else { 6924 if (prevdirrem != NULL) 6925 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 6926 prevdirrem, dm_next); 6927 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 6928 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 6929 FREE_LOCK(&lk); 6930 if (direct) 6931 handle_workitem_remove(dirrem, NULL); 6932 } 6933} 6934 6935/* 6936 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 6937 * pd_pendinghd list of a pagedep. 6938 */ 6939static struct diradd * 6940diradd_lookup(pagedep, offset) 6941 struct pagedep *pagedep; 6942 int offset; 6943{ 6944 struct diradd *dap; 6945 6946 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 6947 if (dap->da_offset == offset) 6948 return (dap); 6949 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 6950 if (dap->da_offset == offset) 6951 return (dap); 6952 return (NULL); 6953} 6954 6955/* 6956 * Search for a .. diradd dependency in a directory that is being removed. 6957 * If the directory was renamed to a new parent we have a diradd rather 6958 * than a mkdir for the .. entry. We need to cancel it now before 6959 * it is found in truncate(). 6960 */ 6961static struct jremref * 6962cancel_diradd_dotdot(ip, dirrem, jremref) 6963 struct inode *ip; 6964 struct dirrem *dirrem; 6965 struct jremref *jremref; 6966{ 6967 struct pagedep *pagedep; 6968 struct diradd *dap; 6969 struct worklist *wk; 6970 6971 if (pagedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 0, 6972 &pagedep) == 0) 6973 return (jremref); 6974 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 6975 if (dap == NULL) 6976 return (jremref); 6977 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 6978 /* 6979 * Mark any journal work as belonging to the parent so it is freed 6980 * with the .. reference. 6981 */ 6982 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 6983 wk->wk_state |= MKDIR_PARENT; 6984 return (NULL); 6985} 6986 6987/* 6988 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 6989 * replace it with a dirrem/diradd pair as a result of re-parenting a 6990 * directory. This ensures that we don't simultaneously have a mkdir and 6991 * a diradd for the same .. entry. 6992 */ 6993static struct jremref * 6994cancel_mkdir_dotdot(ip, dirrem, jremref) 6995 struct inode *ip; 6996 struct dirrem *dirrem; 6997 struct jremref *jremref; 6998{ 6999 struct inodedep *inodedep; 7000 struct jaddref *jaddref; 7001 struct mkdir *mkdir; 7002 struct diradd *dap; 7003 7004 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 7005 &inodedep) == 0) 7006 panic("cancel_mkdir_dotdot: Lost inodedep"); 7007 dap = inodedep->id_mkdiradd; 7008 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 7009 return (jremref); 7010 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; 7011 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 7012 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 7013 break; 7014 if (mkdir == NULL) 7015 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 7016 if ((jaddref = mkdir->md_jaddref) != NULL) { 7017 mkdir->md_jaddref = NULL; 7018 jaddref->ja_state &= ~MKDIR_PARENT; 7019 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0, 7020 &inodedep) == 0) 7021 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 7022 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 7023 journal_jremref(dirrem, jremref, inodedep); 7024 jremref = NULL; 7025 } 7026 } 7027 if (mkdir->md_state & ONWORKLIST) 7028 WORKLIST_REMOVE(&mkdir->md_list); 7029 mkdir->md_state |= ALLCOMPLETE; 7030 complete_mkdir(mkdir); 7031 return (jremref); 7032} 7033 7034static void 7035journal_jremref(dirrem, jremref, inodedep) 7036 struct dirrem *dirrem; 7037 struct jremref *jremref; 7038 struct inodedep *inodedep; 7039{ 7040 7041 if (inodedep == NULL) 7042 if (inodedep_lookup(jremref->jr_list.wk_mp, 7043 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 7044 panic("journal_jremref: Lost inodedep"); 7045 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 7046 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 7047 add_to_journal(&jremref->jr_list); 7048} 7049 7050static void 7051dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 7052 struct dirrem *dirrem; 7053 struct jremref *jremref; 7054 struct jremref *dotremref; 7055 struct jremref *dotdotremref; 7056{ 7057 struct inodedep *inodedep; 7058 7059 7060 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 7061 &inodedep) == 0) 7062 panic("dirrem_journal: Lost inodedep"); 7063 journal_jremref(dirrem, jremref, inodedep); 7064 if (dotremref) 7065 journal_jremref(dirrem, dotremref, inodedep); 7066 if (dotdotremref) 7067 journal_jremref(dirrem, dotdotremref, NULL); 7068} 7069 7070/* 7071 * Allocate a new dirrem if appropriate and return it along with 7072 * its associated pagedep. Called without a lock, returns with lock. 7073 */ 7074static long num_dirrem; /* number of dirrem allocated */ 7075static struct dirrem * 7076newdirrem(bp, dp, ip, isrmdir, prevdirremp) 7077 struct buf *bp; /* buffer containing directory block */ 7078 struct inode *dp; /* inode for the directory being modified */ 7079 struct inode *ip; /* inode for directory entry being removed */ 7080 int isrmdir; /* indicates if doing RMDIR */ 7081 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 7082{ 7083 int offset; 7084 ufs_lbn_t lbn; 7085 struct diradd *dap; 7086 struct dirrem *dirrem; 7087 struct pagedep *pagedep; 7088 struct jremref *jremref; 7089 struct jremref *dotremref; 7090 struct jremref *dotdotremref; 7091 struct vnode *dvp; 7092 7093 /* 7094 * Whiteouts have no deletion dependencies. 7095 */ 7096 if (ip == NULL) 7097 panic("newdirrem: whiteout"); 7098 dvp = ITOV(dp); 7099 /* 7100 * If we are over our limit, try to improve the situation. 7101 * Limiting the number of dirrem structures will also limit 7102 * the number of freefile and freeblks structures. 7103 */ 7104 ACQUIRE_LOCK(&lk); 7105 if (!(ip->i_flags & SF_SNAPSHOT) && num_dirrem > max_softdeps / 2) 7106 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE); 7107 num_dirrem += 1; 7108 FREE_LOCK(&lk); 7109 dirrem = malloc(sizeof(struct dirrem), 7110 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 7111 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 7112 LIST_INIT(&dirrem->dm_jremrefhd); 7113 LIST_INIT(&dirrem->dm_jwork); 7114 dirrem->dm_state = isrmdir ? RMDIR : 0; 7115 dirrem->dm_oldinum = ip->i_number; 7116 *prevdirremp = NULL; 7117 /* 7118 * Allocate remove reference structures to track journal write 7119 * dependencies. We will always have one for the link and 7120 * when doing directories we will always have one more for dot. 7121 * When renaming a directory we skip the dotdot link change so 7122 * this is not needed. 7123 */ 7124 jremref = dotremref = dotdotremref = NULL; 7125 if (DOINGSUJ(dvp)) { 7126 if (isrmdir) { 7127 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 7128 ip->i_effnlink + 2); 7129 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 7130 ip->i_effnlink + 1); 7131 } else 7132 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 7133 ip->i_effnlink + 1); 7134 if (isrmdir > 1) { 7135 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 7136 dp->i_effnlink + 1); 7137 dotdotremref->jr_state |= MKDIR_PARENT; 7138 } 7139 } 7140 ACQUIRE_LOCK(&lk); 7141 lbn = lblkno(dp->i_fs, dp->i_offset); 7142 offset = blkoff(dp->i_fs, dp->i_offset); 7143 if (pagedep_lookup(UFSTOVFS(dp->i_ump), dp->i_number, lbn, DEPALLOC, 7144 &pagedep) == 0) 7145 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 7146 dirrem->dm_pagedep = pagedep; 7147 /* 7148 * If we're renaming a .. link to a new directory, cancel any 7149 * existing MKDIR_PARENT mkdir. If it has already been canceled 7150 * the jremref is preserved for any potential diradd in this 7151 * location. This can not coincide with a rmdir. 7152 */ 7153 if (dp->i_offset == DOTDOT_OFFSET) { 7154 if (isrmdir) 7155 panic("newdirrem: .. directory change during remove?"); 7156 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 7157 } 7158 /* 7159 * If we're removing a directory search for the .. dependency now and 7160 * cancel it. Any pending journal work will be added to the dirrem 7161 * to be completed when the workitem remove completes. 7162 */ 7163 if (isrmdir > 1) 7164 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 7165 /* 7166 * Check for a diradd dependency for the same directory entry. 7167 * If present, then both dependencies become obsolete and can 7168 * be de-allocated. 7169 */ 7170 dap = diradd_lookup(pagedep, offset); 7171 if (dap == NULL) { 7172 /* 7173 * Link the jremref structures into the dirrem so they are 7174 * written prior to the pagedep. 7175 */ 7176 if (jremref) 7177 dirrem_journal(dirrem, jremref, dotremref, 7178 dotdotremref); 7179 return (dirrem); 7180 } 7181 /* 7182 * Must be ATTACHED at this point. 7183 */ 7184 if ((dap->da_state & ATTACHED) == 0) 7185 panic("newdirrem: not ATTACHED"); 7186 if (dap->da_newinum != ip->i_number) 7187 panic("newdirrem: inum %d should be %d", 7188 ip->i_number, dap->da_newinum); 7189 /* 7190 * If we are deleting a changed name that never made it to disk, 7191 * then return the dirrem describing the previous inode (which 7192 * represents the inode currently referenced from this entry on disk). 7193 */ 7194 if ((dap->da_state & DIRCHG) != 0) { 7195 *prevdirremp = dap->da_previous; 7196 dap->da_state &= ~DIRCHG; 7197 dap->da_pagedep = pagedep; 7198 } 7199 /* 7200 * We are deleting an entry that never made it to disk. 7201 * Mark it COMPLETE so we can delete its inode immediately. 7202 */ 7203 dirrem->dm_state |= COMPLETE; 7204 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 7205#ifdef SUJ_DEBUG 7206 if (isrmdir == 0) { 7207 struct worklist *wk; 7208 7209 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 7210 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 7211 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 7212 } 7213#endif 7214 7215 return (dirrem); 7216} 7217 7218/* 7219 * Directory entry change dependencies. 7220 * 7221 * Changing an existing directory entry requires that an add operation 7222 * be completed first followed by a deletion. The semantics for the addition 7223 * are identical to the description of adding a new entry above except 7224 * that the rollback is to the old inode number rather than zero. Once 7225 * the addition dependency is completed, the removal is done as described 7226 * in the removal routine above. 7227 */ 7228 7229/* 7230 * This routine should be called immediately after changing 7231 * a directory entry. The inode's link count should not be 7232 * decremented by the calling procedure -- the soft updates 7233 * code will perform this task when it is safe. 7234 */ 7235void 7236softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 7237 struct buf *bp; /* buffer containing directory block */ 7238 struct inode *dp; /* inode for the directory being modified */ 7239 struct inode *ip; /* inode for directory entry being removed */ 7240 ino_t newinum; /* new inode number for changed entry */ 7241 int isrmdir; /* indicates if doing RMDIR */ 7242{ 7243 int offset; 7244 struct diradd *dap = NULL; 7245 struct dirrem *dirrem, *prevdirrem; 7246 struct pagedep *pagedep; 7247 struct inodedep *inodedep; 7248 struct jaddref *jaddref; 7249 struct mount *mp; 7250 7251 offset = blkoff(dp->i_fs, dp->i_offset); 7252 mp = UFSTOVFS(dp->i_ump); 7253 7254 /* 7255 * Whiteouts do not need diradd dependencies. 7256 */ 7257 if (newinum != WINO) { 7258 dap = malloc(sizeof(struct diradd), 7259 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 7260 workitem_alloc(&dap->da_list, D_DIRADD, mp); 7261 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 7262 dap->da_offset = offset; 7263 dap->da_newinum = newinum; 7264 LIST_INIT(&dap->da_jwork); 7265 } 7266 7267 /* 7268 * Allocate a new dirrem and ACQUIRE_LOCK. 7269 */ 7270 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 7271 pagedep = dirrem->dm_pagedep; 7272 /* 7273 * The possible values for isrmdir: 7274 * 0 - non-directory file rename 7275 * 1 - directory rename within same directory 7276 * inum - directory rename to new directory of given inode number 7277 * When renaming to a new directory, we are both deleting and 7278 * creating a new directory entry, so the link count on the new 7279 * directory should not change. Thus we do not need the followup 7280 * dirrem which is usually done in handle_workitem_remove. We set 7281 * the DIRCHG flag to tell handle_workitem_remove to skip the 7282 * followup dirrem. 7283 */ 7284 if (isrmdir > 1) 7285 dirrem->dm_state |= DIRCHG; 7286 7287 /* 7288 * Whiteouts have no additional dependencies, 7289 * so just put the dirrem on the correct list. 7290 */ 7291 if (newinum == WINO) { 7292 if ((dirrem->dm_state & COMPLETE) == 0) { 7293 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 7294 dm_next); 7295 } else { 7296 dirrem->dm_dirinum = pagedep->pd_ino; 7297 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 7298 add_to_worklist(&dirrem->dm_list, 0); 7299 } 7300 FREE_LOCK(&lk); 7301 return; 7302 } 7303 /* 7304 * Add the dirrem to the inodedep's pending remove list for quick 7305 * discovery later. A valid nlinkdelta ensures that this lookup 7306 * will not fail. 7307 */ 7308 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 7309 panic("softdep_setup_directory_change: Lost inodedep."); 7310 dirrem->dm_state |= ONDEPLIST; 7311 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 7312 7313 /* 7314 * If the COMPLETE flag is clear, then there were no active 7315 * entries and we want to roll back to the previous inode until 7316 * the new inode is committed to disk. If the COMPLETE flag is 7317 * set, then we have deleted an entry that never made it to disk. 7318 * If the entry we deleted resulted from a name change, then the old 7319 * inode reference still resides on disk. Any rollback that we do 7320 * needs to be to that old inode (returned to us in prevdirrem). If 7321 * the entry we deleted resulted from a create, then there is 7322 * no entry on the disk, so we want to roll back to zero rather 7323 * than the uncommitted inode. In either of the COMPLETE cases we 7324 * want to immediately free the unwritten and unreferenced inode. 7325 */ 7326 if ((dirrem->dm_state & COMPLETE) == 0) { 7327 dap->da_previous = dirrem; 7328 } else { 7329 if (prevdirrem != NULL) { 7330 dap->da_previous = prevdirrem; 7331 } else { 7332 dap->da_state &= ~DIRCHG; 7333 dap->da_pagedep = pagedep; 7334 } 7335 dirrem->dm_dirinum = pagedep->pd_ino; 7336 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 7337 add_to_worklist(&dirrem->dm_list, 0); 7338 } 7339 /* 7340 * Lookup the jaddref for this journal entry. We must finish 7341 * initializing it and make the diradd write dependent on it. 7342 * If we're not journaling Put it on the id_bufwait list if the inode 7343 * is not yet written. If it is written, do the post-inode write 7344 * processing to put it on the id_pendinghd list. 7345 */ 7346 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 7347 if (mp->mnt_kern_flag & MNTK_SUJ) { 7348 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 7349 inoreflst); 7350 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 7351 ("softdep_setup_directory_change: bad jaddref %p", 7352 jaddref)); 7353 jaddref->ja_diroff = dp->i_offset; 7354 jaddref->ja_diradd = dap; 7355 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 7356 dap, da_pdlist); 7357 add_to_journal(&jaddref->ja_list); 7358 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 7359 dap->da_state |= COMPLETE; 7360 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 7361 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 7362 } else { 7363 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 7364 dap, da_pdlist); 7365 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 7366 } 7367 /* 7368 * If we're making a new name for a directory that has not been 7369 * committed when need to move the dot and dotdot references to 7370 * this new name. 7371 */ 7372 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 7373 merge_diradd(inodedep, dap); 7374 FREE_LOCK(&lk); 7375} 7376 7377/* 7378 * Called whenever the link count on an inode is changed. 7379 * It creates an inode dependency so that the new reference(s) 7380 * to the inode cannot be committed to disk until the updated 7381 * inode has been written. 7382 */ 7383void 7384softdep_change_linkcnt(ip) 7385 struct inode *ip; /* the inode with the increased link count */ 7386{ 7387 struct inodedep *inodedep; 7388 7389 ACQUIRE_LOCK(&lk); 7390 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep); 7391 if (ip->i_nlink < ip->i_effnlink) 7392 panic("softdep_change_linkcnt: bad delta"); 7393 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 7394 FREE_LOCK(&lk); 7395} 7396 7397/* 7398 * Called when the effective link count and the reference count 7399 * on an inode drops to zero. At this point there are no names 7400 * referencing the file in the filesystem and no active file 7401 * references. The space associated with the file will be freed 7402 * as soon as the necessary soft dependencies are cleared. 7403 */ 7404void 7405softdep_releasefile(ip) 7406 struct inode *ip; /* inode with the zero effective link count */ 7407{ 7408 struct inodedep *inodedep; 7409 struct fs *fs; 7410 int extblocks; 7411 7412 if (ip->i_effnlink > 0) 7413 panic("softdep_releasefile: file still referenced"); 7414 /* 7415 * We may be called several times as the on-disk link count 7416 * drops to zero. We only want to account for the space once. 7417 */ 7418 if (ip->i_flag & IN_SPACECOUNTED) 7419 return; 7420 /* 7421 * We have to deactivate a snapshot otherwise copyonwrites may 7422 * add blocks and the cleanup may remove blocks after we have 7423 * tried to account for them. 7424 */ 7425 if ((ip->i_flags & SF_SNAPSHOT) != 0) 7426 ffs_snapremove(ITOV(ip)); 7427 /* 7428 * If we are tracking an nlinkdelta, we have to also remember 7429 * whether we accounted for the freed space yet. 7430 */ 7431 ACQUIRE_LOCK(&lk); 7432 if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep))) 7433 inodedep->id_state |= SPACECOUNTED; 7434 FREE_LOCK(&lk); 7435 fs = ip->i_fs; 7436 extblocks = 0; 7437 if (fs->fs_magic == FS_UFS2_MAGIC) 7438 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 7439 UFS_LOCK(ip->i_ump); 7440 ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks; 7441 ip->i_fs->fs_pendinginodes += 1; 7442 UFS_UNLOCK(ip->i_ump); 7443 ip->i_flag |= IN_SPACECOUNTED; 7444} 7445 7446/* 7447 * Attach a sbdep dependency to the superblock buf so that we can keep 7448 * track of the head of the linked list of referenced but unlinked inodes. 7449 */ 7450void 7451softdep_setup_sbupdate(ump, fs, bp) 7452 struct ufsmount *ump; 7453 struct fs *fs; 7454 struct buf *bp; 7455{ 7456 struct sbdep *sbdep; 7457 struct worklist *wk; 7458 7459 if ((fs->fs_flags & FS_SUJ) == 0) 7460 return; 7461 LIST_FOREACH(wk, &bp->b_dep, wk_list) 7462 if (wk->wk_type == D_SBDEP) 7463 break; 7464 if (wk != NULL) 7465 return; 7466 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 7467 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 7468 sbdep->sb_fs = fs; 7469 sbdep->sb_ump = ump; 7470 ACQUIRE_LOCK(&lk); 7471 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 7472 FREE_LOCK(&lk); 7473} 7474 7475/* 7476 * Return the first unlinked inodedep which is ready to be the head of the 7477 * list. The inodedep and all those after it must have valid next pointers. 7478 */ 7479static struct inodedep * 7480first_unlinked_inodedep(ump) 7481 struct ufsmount *ump; 7482{ 7483 struct inodedep *inodedep; 7484 struct inodedep *idp; 7485 7486 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 7487 inodedep; inodedep = idp) { 7488 if ((inodedep->id_state & UNLINKNEXT) == 0) 7489 return (NULL); 7490 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 7491 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 7492 break; 7493 if ((inodedep->id_state & UNLINKPREV) == 0) 7494 panic("first_unlinked_inodedep: prev != next"); 7495 } 7496 if (inodedep == NULL) 7497 return (NULL); 7498 7499 return (inodedep); 7500} 7501 7502/* 7503 * Set the sujfree unlinked head pointer prior to writing a superblock. 7504 */ 7505static void 7506initiate_write_sbdep(sbdep) 7507 struct sbdep *sbdep; 7508{ 7509 struct inodedep *inodedep; 7510 struct fs *bpfs; 7511 struct fs *fs; 7512 7513 bpfs = sbdep->sb_fs; 7514 fs = sbdep->sb_ump->um_fs; 7515 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 7516 if (inodedep) { 7517 fs->fs_sujfree = inodedep->id_ino; 7518 inodedep->id_state |= UNLINKPREV; 7519 } else 7520 fs->fs_sujfree = 0; 7521 bpfs->fs_sujfree = fs->fs_sujfree; 7522} 7523 7524/* 7525 * After a superblock is written determine whether it must be written again 7526 * due to a changing unlinked list head. 7527 */ 7528static int 7529handle_written_sbdep(sbdep, bp) 7530 struct sbdep *sbdep; 7531 struct buf *bp; 7532{ 7533 struct inodedep *inodedep; 7534 struct mount *mp; 7535 struct fs *fs; 7536 7537 fs = sbdep->sb_fs; 7538 mp = UFSTOVFS(sbdep->sb_ump); 7539 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 7540 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 7541 (inodedep == NULL && fs->fs_sujfree != 0)) { 7542 bdirty(bp); 7543 return (1); 7544 } 7545 WORKITEM_FREE(sbdep, D_SBDEP); 7546 if (fs->fs_sujfree == 0) 7547 return (0); 7548 if (inodedep_lookup(mp, fs->fs_sujfree, 0, &inodedep) == 0) 7549 panic("handle_written_sbdep: lost inodedep"); 7550 /* 7551 * Now that we have a record of this indode in stable store allow it 7552 * to be written to free up pending work. Inodes may see a lot of 7553 * write activity after they are unlinked which we must not hold up. 7554 */ 7555 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 7556 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 7557 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 7558 inodedep, inodedep->id_state); 7559 if (inodedep->id_state & UNLINKONLIST) 7560 break; 7561 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 7562 } 7563 7564 return (0); 7565} 7566 7567/* 7568 * Mark an inodedep has unlinked and insert it into the in-memory unlinked 7569 * list. 7570 */ 7571static void 7572unlinked_inodedep(mp, inodedep) 7573 struct mount *mp; 7574 struct inodedep *inodedep; 7575{ 7576 struct ufsmount *ump; 7577 7578 if ((mp->mnt_kern_flag & MNTK_SUJ) == 0) 7579 return; 7580 ump = VFSTOUFS(mp); 7581 ump->um_fs->fs_fmod = 1; 7582 inodedep->id_state |= UNLINKED; 7583 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 7584} 7585 7586/* 7587 * Remove an inodedep from the unlinked inodedep list. This may require 7588 * disk writes if the inode has made it that far. 7589 */ 7590static void 7591clear_unlinked_inodedep(inodedep) 7592 struct inodedep *inodedep; 7593{ 7594 struct ufsmount *ump; 7595 struct inodedep *idp; 7596 struct inodedep *idn; 7597 struct fs *fs; 7598 struct buf *bp; 7599 ino_t ino; 7600 ino_t nino; 7601 ino_t pino; 7602 int error; 7603 7604 ump = VFSTOUFS(inodedep->id_list.wk_mp); 7605 fs = ump->um_fs; 7606 ino = inodedep->id_ino; 7607 error = 0; 7608 for (;;) { 7609 /* 7610 * If nothing has yet been written simply remove us from 7611 * the in memory list and return. This is the most common 7612 * case where handle_workitem_remove() loses the final 7613 * reference. 7614 */ 7615 if ((inodedep->id_state & UNLINKLINKS) == 0) 7616 break; 7617 /* 7618 * If we have a NEXT pointer and no PREV pointer we can simply 7619 * clear NEXT's PREV and remove ourselves from the list. Be 7620 * careful not to clear PREV if the superblock points at 7621 * next as well. 7622 */ 7623 idn = TAILQ_NEXT(inodedep, id_unlinked); 7624 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 7625 if (idn && fs->fs_sujfree != idn->id_ino) 7626 idn->id_state &= ~UNLINKPREV; 7627 break; 7628 } 7629 /* 7630 * Here we have an inodedep which is actually linked into 7631 * the list. We must remove it by forcing a write to the 7632 * link before us, whether it be the superblock or an inode. 7633 * Unfortunately the list may change while we're waiting 7634 * on the buf lock for either resource so we must loop until 7635 * we lock. the right one. If both the superblock and an 7636 * inode point to this inode we must clear the inode first 7637 * followed by the superblock. 7638 */ 7639 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 7640 pino = 0; 7641 if (idp && (idp->id_state & UNLINKNEXT)) 7642 pino = idp->id_ino; 7643 FREE_LOCK(&lk); 7644 if (pino == 0) 7645 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 7646 (int)fs->fs_sbsize, 0, 0, 0); 7647 else 7648 error = bread(ump->um_devvp, 7649 fsbtodb(fs, ino_to_fsba(fs, pino)), 7650 (int)fs->fs_bsize, NOCRED, &bp); 7651 ACQUIRE_LOCK(&lk); 7652 if (error) 7653 break; 7654 /* If the list has changed restart the loop. */ 7655 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 7656 nino = 0; 7657 if (idp && (idp->id_state & UNLINKNEXT)) 7658 nino = idp->id_ino; 7659 if (nino != pino || 7660 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 7661 FREE_LOCK(&lk); 7662 brelse(bp); 7663 ACQUIRE_LOCK(&lk); 7664 continue; 7665 } 7666 /* 7667 * Remove us from the in memory list. After this we cannot 7668 * access the inodedep. 7669 */ 7670 idn = TAILQ_NEXT(inodedep, id_unlinked); 7671 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS); 7672 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 7673 /* 7674 * Determine the next inode number. 7675 */ 7676 nino = 0; 7677 if (idn) { 7678 /* 7679 * If next isn't on the list we can just clear prev's 7680 * state and schedule it to be fixed later. No need 7681 * to synchronously write if we're not in the real 7682 * list. 7683 */ 7684 if ((idn->id_state & UNLINKPREV) == 0 && pino != 0) { 7685 idp->id_state &= ~UNLINKNEXT; 7686 if ((idp->id_state & ONWORKLIST) == 0) 7687 WORKLIST_INSERT(&bp->b_dep, 7688 &idp->id_list); 7689 FREE_LOCK(&lk); 7690 bawrite(bp); 7691 ACQUIRE_LOCK(&lk); 7692 return; 7693 } 7694 nino = idn->id_ino; 7695 } 7696 FREE_LOCK(&lk); 7697 /* 7698 * The predecessor's next pointer is manually updated here 7699 * so that the NEXT flag is never cleared for an element 7700 * that is in the list. 7701 */ 7702 if (pino == 0) { 7703 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 7704 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 7705 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 7706 bp); 7707 } else if (fs->fs_magic == FS_UFS1_MAGIC) 7708 ((struct ufs1_dinode *)bp->b_data + 7709 ino_to_fsbo(fs, pino))->di_freelink = nino; 7710 else 7711 ((struct ufs2_dinode *)bp->b_data + 7712 ino_to_fsbo(fs, pino))->di_freelink = nino; 7713 /* 7714 * If the bwrite fails we have no recourse to recover. The 7715 * filesystem is corrupted already. 7716 */ 7717 bwrite(bp); 7718 ACQUIRE_LOCK(&lk); 7719 /* 7720 * If the superblock pointer still needs to be cleared force 7721 * a write here. 7722 */ 7723 if (fs->fs_sujfree == ino) { 7724 FREE_LOCK(&lk); 7725 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 7726 (int)fs->fs_sbsize, 0, 0, 0); 7727 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 7728 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 7729 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 7730 bp); 7731 bwrite(bp); 7732 ACQUIRE_LOCK(&lk); 7733 } 7734 if (fs->fs_sujfree != ino) 7735 return; 7736 panic("clear_unlinked_inodedep: Failed to clear free head"); 7737 } 7738 if (inodedep->id_ino == fs->fs_sujfree) 7739 panic("clear_unlinked_inodedep: Freeing head of free list"); 7740 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS); 7741 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 7742 return; 7743} 7744 7745/* 7746 * This workitem decrements the inode's link count. 7747 * If the link count reaches zero, the file is removed. 7748 */ 7749static void 7750handle_workitem_remove(dirrem, xp) 7751 struct dirrem *dirrem; 7752 struct vnode *xp; 7753{ 7754 struct thread *td = curthread; 7755 struct inodedep *inodedep; 7756 struct workhead dotdotwk; 7757 struct worklist *wk; 7758 struct ufsmount *ump; 7759 struct mount *mp; 7760 struct vnode *vp; 7761 struct inode *ip; 7762 ino_t oldinum; 7763 int error; 7764 7765 if (dirrem->dm_state & ONWORKLIST) 7766 panic("handle_workitem_remove: dirrem %p still on worklist", 7767 dirrem); 7768 oldinum = dirrem->dm_oldinum; 7769 mp = dirrem->dm_list.wk_mp; 7770 ump = VFSTOUFS(mp); 7771 if ((vp = xp) == NULL && 7772 (error = ffs_vgetf(mp, oldinum, LK_EXCLUSIVE, &vp, 7773 FFSV_FORCEINSMQ)) != 0) { 7774 softdep_error("handle_workitem_remove: vget", error); 7775 return; 7776 } 7777 ip = VTOI(vp); 7778 ACQUIRE_LOCK(&lk); 7779 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 7780 panic("handle_workitem_remove: lost inodedep"); 7781 if (dirrem->dm_state & ONDEPLIST) 7782 LIST_REMOVE(dirrem, dm_inonext); 7783 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 7784 ("handle_workitem_remove: Journal entries not written.")); 7785 7786 /* 7787 * Move all dependencies waiting on the remove to complete 7788 * from the dirrem to the inode inowait list to be completed 7789 * after the inode has been updated and written to disk. Any 7790 * marked MKDIR_PARENT are saved to be completed when the .. ref 7791 * is removed. 7792 */ 7793 LIST_INIT(&dotdotwk); 7794 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 7795 WORKLIST_REMOVE(wk); 7796 if (wk->wk_state & MKDIR_PARENT) { 7797 wk->wk_state &= ~MKDIR_PARENT; 7798 WORKLIST_INSERT(&dotdotwk, wk); 7799 continue; 7800 } 7801 WORKLIST_INSERT(&inodedep->id_inowait, wk); 7802 } 7803 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 7804 /* 7805 * Normal file deletion. 7806 */ 7807 if ((dirrem->dm_state & RMDIR) == 0) { 7808 ip->i_nlink--; 7809 DIP_SET(ip, i_nlink, ip->i_nlink); 7810 ip->i_flag |= IN_CHANGE; 7811 if (ip->i_nlink < ip->i_effnlink) 7812 panic("handle_workitem_remove: bad file delta"); 7813 if (ip->i_nlink == 0) 7814 unlinked_inodedep(mp, inodedep); 7815 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 7816 num_dirrem -= 1; 7817 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 7818 ("handle_workitem_remove: worklist not empty. %s", 7819 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 7820 WORKITEM_FREE(dirrem, D_DIRREM); 7821 FREE_LOCK(&lk); 7822 goto out; 7823 } 7824 /* 7825 * Directory deletion. Decrement reference count for both the 7826 * just deleted parent directory entry and the reference for ".". 7827 * Next truncate the directory to length zero. When the 7828 * truncation completes, arrange to have the reference count on 7829 * the parent decremented to account for the loss of "..". 7830 */ 7831 ip->i_nlink -= 2; 7832 DIP_SET(ip, i_nlink, ip->i_nlink); 7833 ip->i_flag |= IN_CHANGE; 7834 if (ip->i_nlink < ip->i_effnlink) 7835 panic("handle_workitem_remove: bad dir delta"); 7836 if (ip->i_nlink == 0) 7837 unlinked_inodedep(mp, inodedep); 7838 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 7839 FREE_LOCK(&lk); 7840 if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0) 7841 softdep_error("handle_workitem_remove: truncate", error); 7842 ACQUIRE_LOCK(&lk); 7843 /* 7844 * Rename a directory to a new parent. Since, we are both deleting 7845 * and creating a new directory entry, the link count on the new 7846 * directory should not change. Thus we skip the followup dirrem. 7847 */ 7848 if (dirrem->dm_state & DIRCHG) { 7849 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 7850 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 7851 num_dirrem -= 1; 7852 WORKITEM_FREE(dirrem, D_DIRREM); 7853 FREE_LOCK(&lk); 7854 goto out; 7855 } 7856 dirrem->dm_state = ONDEPLIST; 7857 dirrem->dm_oldinum = dirrem->dm_dirinum; 7858 /* 7859 * Place the dirrem on the parent's diremhd list. 7860 */ 7861 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 7862 panic("handle_workitem_remove: lost dir inodedep"); 7863 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 7864 /* 7865 * If the allocated inode has never been written to disk, then 7866 * the on-disk inode is zero'ed and we can remove the file 7867 * immediately. When journaling if the inode has been marked 7868 * unlinked and not DEPCOMPLETE we know it can never be written. 7869 */ 7870 inodedep_lookup(mp, oldinum, 0, &inodedep); 7871 if (inodedep == NULL || 7872 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 7873 check_inode_unwritten(inodedep)) { 7874 if (xp != NULL) 7875 add_to_worklist(&dirrem->dm_list, 0); 7876 FREE_LOCK(&lk); 7877 if (xp == NULL) { 7878 vput(vp); 7879 handle_workitem_remove(dirrem, NULL); 7880 } 7881 return; 7882 } 7883 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 7884 FREE_LOCK(&lk); 7885 ip->i_flag |= IN_CHANGE; 7886out: 7887 ffs_update(vp, 0); 7888 if (xp == NULL) 7889 vput(vp); 7890} 7891 7892/* 7893 * Inode de-allocation dependencies. 7894 * 7895 * When an inode's link count is reduced to zero, it can be de-allocated. We 7896 * found it convenient to postpone de-allocation until after the inode is 7897 * written to disk with its new link count (zero). At this point, all of the 7898 * on-disk inode's block pointers are nullified and, with careful dependency 7899 * list ordering, all dependencies related to the inode will be satisfied and 7900 * the corresponding dependency structures de-allocated. So, if/when the 7901 * inode is reused, there will be no mixing of old dependencies with new 7902 * ones. This artificial dependency is set up by the block de-allocation 7903 * procedure above (softdep_setup_freeblocks) and completed by the 7904 * following procedure. 7905 */ 7906static void 7907handle_workitem_freefile(freefile) 7908 struct freefile *freefile; 7909{ 7910 struct workhead wkhd; 7911 struct fs *fs; 7912 struct inodedep *idp; 7913 struct ufsmount *ump; 7914 int error; 7915 7916 ump = VFSTOUFS(freefile->fx_list.wk_mp); 7917 fs = ump->um_fs; 7918#ifdef DEBUG 7919 ACQUIRE_LOCK(&lk); 7920 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 7921 FREE_LOCK(&lk); 7922 if (error) 7923 panic("handle_workitem_freefile: inodedep %p survived", idp); 7924#endif 7925 UFS_LOCK(ump); 7926 fs->fs_pendinginodes -= 1; 7927 UFS_UNLOCK(ump); 7928 LIST_INIT(&wkhd); 7929 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 7930 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 7931 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 7932 softdep_error("handle_workitem_freefile", error); 7933 ACQUIRE_LOCK(&lk); 7934 WORKITEM_FREE(freefile, D_FREEFILE); 7935 FREE_LOCK(&lk); 7936} 7937 7938 7939/* 7940 * Helper function which unlinks marker element from work list and returns 7941 * the next element on the list. 7942 */ 7943static __inline struct worklist * 7944markernext(struct worklist *marker) 7945{ 7946 struct worklist *next; 7947 7948 next = LIST_NEXT(marker, wk_list); 7949 LIST_REMOVE(marker, wk_list); 7950 return next; 7951} 7952 7953/* 7954 * Disk writes. 7955 * 7956 * The dependency structures constructed above are most actively used when file 7957 * system blocks are written to disk. No constraints are placed on when a 7958 * block can be written, but unsatisfied update dependencies are made safe by 7959 * modifying (or replacing) the source memory for the duration of the disk 7960 * write. When the disk write completes, the memory block is again brought 7961 * up-to-date. 7962 * 7963 * In-core inode structure reclamation. 7964 * 7965 * Because there are a finite number of "in-core" inode structures, they are 7966 * reused regularly. By transferring all inode-related dependencies to the 7967 * in-memory inode block and indexing them separately (via "inodedep"s), we 7968 * can allow "in-core" inode structures to be reused at any time and avoid 7969 * any increase in contention. 7970 * 7971 * Called just before entering the device driver to initiate a new disk I/O. 7972 * The buffer must be locked, thus, no I/O completion operations can occur 7973 * while we are manipulating its associated dependencies. 7974 */ 7975static void 7976softdep_disk_io_initiation(bp) 7977 struct buf *bp; /* structure describing disk write to occur */ 7978{ 7979 struct worklist *wk; 7980 struct worklist marker; 7981 struct inodedep *inodedep; 7982 struct freeblks *freeblks; 7983 struct jfreeblk *jfreeblk; 7984 struct newblk *newblk; 7985 7986 /* 7987 * We only care about write operations. There should never 7988 * be dependencies for reads. 7989 */ 7990 if (bp->b_iocmd != BIO_WRITE) 7991 panic("softdep_disk_io_initiation: not write"); 7992 7993 if (bp->b_vflags & BV_BKGRDINPROG) 7994 panic("softdep_disk_io_initiation: Writing buffer with " 7995 "background write in progress: %p", bp); 7996 7997 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 7998 PHOLD(curproc); /* Don't swap out kernel stack */ 7999 8000 ACQUIRE_LOCK(&lk); 8001 /* 8002 * Do any necessary pre-I/O processing. 8003 */ 8004 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 8005 wk = markernext(&marker)) { 8006 LIST_INSERT_AFTER(wk, &marker, wk_list); 8007 switch (wk->wk_type) { 8008 8009 case D_PAGEDEP: 8010 initiate_write_filepage(WK_PAGEDEP(wk), bp); 8011 continue; 8012 8013 case D_INODEDEP: 8014 inodedep = WK_INODEDEP(wk); 8015 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 8016 initiate_write_inodeblock_ufs1(inodedep, bp); 8017 else 8018 initiate_write_inodeblock_ufs2(inodedep, bp); 8019 continue; 8020 8021 case D_INDIRDEP: 8022 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 8023 continue; 8024 8025 case D_BMSAFEMAP: 8026 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 8027 continue; 8028 8029 case D_JSEG: 8030 WK_JSEG(wk)->js_buf = NULL; 8031 continue; 8032 8033 case D_FREEBLKS: 8034 freeblks = WK_FREEBLKS(wk); 8035 jfreeblk = LIST_FIRST(&freeblks->fb_jfreeblkhd); 8036 /* 8037 * We have to wait for the jfreeblks to be journaled 8038 * before we can write an inodeblock with updated 8039 * pointers. Be careful to arrange the marker so 8040 * we revisit the jfreeblk if it's not removed by 8041 * the first jwait(). 8042 */ 8043 if (jfreeblk != NULL) { 8044 LIST_REMOVE(&marker, wk_list); 8045 LIST_INSERT_BEFORE(wk, &marker, wk_list); 8046 jwait(&jfreeblk->jf_list); 8047 } 8048 continue; 8049 case D_ALLOCDIRECT: 8050 case D_ALLOCINDIR: 8051 /* 8052 * We have to wait for the jnewblk to be journaled 8053 * before we can write to a block otherwise the 8054 * contents may be confused with an earlier file 8055 * at recovery time. Handle the marker as described 8056 * above. 8057 */ 8058 newblk = WK_NEWBLK(wk); 8059 if (newblk->nb_jnewblk != NULL) { 8060 LIST_REMOVE(&marker, wk_list); 8061 LIST_INSERT_BEFORE(wk, &marker, wk_list); 8062 jwait(&newblk->nb_jnewblk->jn_list); 8063 } 8064 continue; 8065 8066 case D_SBDEP: 8067 initiate_write_sbdep(WK_SBDEP(wk)); 8068 continue; 8069 8070 case D_MKDIR: 8071 case D_FREEWORK: 8072 case D_FREEDEP: 8073 case D_JSEGDEP: 8074 continue; 8075 8076 default: 8077 panic("handle_disk_io_initiation: Unexpected type %s", 8078 TYPENAME(wk->wk_type)); 8079 /* NOTREACHED */ 8080 } 8081 } 8082 FREE_LOCK(&lk); 8083 PRELE(curproc); /* Allow swapout of kernel stack */ 8084} 8085 8086/* 8087 * Called from within the procedure above to deal with unsatisfied 8088 * allocation dependencies in a directory. The buffer must be locked, 8089 * thus, no I/O completion operations can occur while we are 8090 * manipulating its associated dependencies. 8091 */ 8092static void 8093initiate_write_filepage(pagedep, bp) 8094 struct pagedep *pagedep; 8095 struct buf *bp; 8096{ 8097 struct jremref *jremref; 8098 struct jmvref *jmvref; 8099 struct dirrem *dirrem; 8100 struct diradd *dap; 8101 struct direct *ep; 8102 int i; 8103 8104 if (pagedep->pd_state & IOSTARTED) { 8105 /* 8106 * This can only happen if there is a driver that does not 8107 * understand chaining. Here biodone will reissue the call 8108 * to strategy for the incomplete buffers. 8109 */ 8110 printf("initiate_write_filepage: already started\n"); 8111 return; 8112 } 8113 pagedep->pd_state |= IOSTARTED; 8114 /* 8115 * Wait for all journal remove dependencies to hit the disk. 8116 * We can not allow any potentially conflicting directory adds 8117 * to be visible before removes and rollback is too difficult. 8118 * lk may be dropped and re-acquired, however we hold the buf 8119 * locked so the dependency can not go away. 8120 */ 8121 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 8122 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 8123 stat_jwait_filepage++; 8124 jwait(&jremref->jr_list); 8125 } 8126 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 8127 stat_jwait_filepage++; 8128 jwait(&jmvref->jm_list); 8129 } 8130 for (i = 0; i < DAHASHSZ; i++) { 8131 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 8132 ep = (struct direct *) 8133 ((char *)bp->b_data + dap->da_offset); 8134 if (ep->d_ino != dap->da_newinum) 8135 panic("%s: dir inum %d != new %d", 8136 "initiate_write_filepage", 8137 ep->d_ino, dap->da_newinum); 8138 if (dap->da_state & DIRCHG) 8139 ep->d_ino = dap->da_previous->dm_oldinum; 8140 else 8141 ep->d_ino = 0; 8142 dap->da_state &= ~ATTACHED; 8143 dap->da_state |= UNDONE; 8144 } 8145 } 8146} 8147 8148/* 8149 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 8150 * Note that any bug fixes made to this routine must be done in the 8151 * version found below. 8152 * 8153 * Called from within the procedure above to deal with unsatisfied 8154 * allocation dependencies in an inodeblock. The buffer must be 8155 * locked, thus, no I/O completion operations can occur while we 8156 * are manipulating its associated dependencies. 8157 */ 8158static void 8159initiate_write_inodeblock_ufs1(inodedep, bp) 8160 struct inodedep *inodedep; 8161 struct buf *bp; /* The inode block */ 8162{ 8163 struct allocdirect *adp, *lastadp; 8164 struct ufs1_dinode *dp; 8165 struct ufs1_dinode *sip; 8166 struct inoref *inoref; 8167 struct fs *fs; 8168 ufs_lbn_t i; 8169#ifdef INVARIANTS 8170 ufs_lbn_t prevlbn = 0; 8171#endif 8172 int deplist; 8173 8174 if (inodedep->id_state & IOSTARTED) 8175 panic("initiate_write_inodeblock_ufs1: already started"); 8176 inodedep->id_state |= IOSTARTED; 8177 fs = inodedep->id_fs; 8178 dp = (struct ufs1_dinode *)bp->b_data + 8179 ino_to_fsbo(fs, inodedep->id_ino); 8180 8181 /* 8182 * If we're on the unlinked list but have not yet written our 8183 * next pointer initialize it here. 8184 */ 8185 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 8186 struct inodedep *inon; 8187 8188 inon = TAILQ_NEXT(inodedep, id_unlinked); 8189 dp->di_freelink = inon ? inon->id_ino : 0; 8190 } 8191 /* 8192 * If the bitmap is not yet written, then the allocated 8193 * inode cannot be written to disk. 8194 */ 8195 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 8196 if (inodedep->id_savedino1 != NULL) 8197 panic("initiate_write_inodeblock_ufs1: I/O underway"); 8198 FREE_LOCK(&lk); 8199 sip = malloc(sizeof(struct ufs1_dinode), 8200 M_SAVEDINO, M_SOFTDEP_FLAGS); 8201 ACQUIRE_LOCK(&lk); 8202 inodedep->id_savedino1 = sip; 8203 *inodedep->id_savedino1 = *dp; 8204 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 8205 dp->di_gen = inodedep->id_savedino1->di_gen; 8206 dp->di_freelink = inodedep->id_savedino1->di_freelink; 8207 return; 8208 } 8209 /* 8210 * If no dependencies, then there is nothing to roll back. 8211 */ 8212 inodedep->id_savedsize = dp->di_size; 8213 inodedep->id_savedextsize = 0; 8214 inodedep->id_savednlink = dp->di_nlink; 8215 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 8216 TAILQ_EMPTY(&inodedep->id_inoreflst)) 8217 return; 8218 /* 8219 * Revert the link count to that of the first unwritten journal entry. 8220 */ 8221 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 8222 if (inoref) 8223 dp->di_nlink = inoref->if_nlink; 8224 /* 8225 * Set the dependencies to busy. 8226 */ 8227 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 8228 adp = TAILQ_NEXT(adp, ad_next)) { 8229#ifdef INVARIANTS 8230 if (deplist != 0 && prevlbn >= adp->ad_offset) 8231 panic("softdep_write_inodeblock: lbn order"); 8232 prevlbn = adp->ad_offset; 8233 if (adp->ad_offset < NDADDR && 8234 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 8235 panic("%s: direct pointer #%jd mismatch %d != %jd", 8236 "softdep_write_inodeblock", 8237 (intmax_t)adp->ad_offset, 8238 dp->di_db[adp->ad_offset], 8239 (intmax_t)adp->ad_newblkno); 8240 if (adp->ad_offset >= NDADDR && 8241 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 8242 panic("%s: indirect pointer #%jd mismatch %d != %jd", 8243 "softdep_write_inodeblock", 8244 (intmax_t)adp->ad_offset - NDADDR, 8245 dp->di_ib[adp->ad_offset - NDADDR], 8246 (intmax_t)adp->ad_newblkno); 8247 deplist |= 1 << adp->ad_offset; 8248 if ((adp->ad_state & ATTACHED) == 0) 8249 panic("softdep_write_inodeblock: Unknown state 0x%x", 8250 adp->ad_state); 8251#endif /* INVARIANTS */ 8252 adp->ad_state &= ~ATTACHED; 8253 adp->ad_state |= UNDONE; 8254 } 8255 /* 8256 * The on-disk inode cannot claim to be any larger than the last 8257 * fragment that has been written. Otherwise, the on-disk inode 8258 * might have fragments that were not the last block in the file 8259 * which would corrupt the filesystem. 8260 */ 8261 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 8262 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 8263 if (adp->ad_offset >= NDADDR) 8264 break; 8265 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 8266 /* keep going until hitting a rollback to a frag */ 8267 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 8268 continue; 8269 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 8270 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 8271#ifdef INVARIANTS 8272 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 8273 panic("softdep_write_inodeblock: lost dep1"); 8274#endif /* INVARIANTS */ 8275 dp->di_db[i] = 0; 8276 } 8277 for (i = 0; i < NIADDR; i++) { 8278#ifdef INVARIANTS 8279 if (dp->di_ib[i] != 0 && 8280 (deplist & ((1 << NDADDR) << i)) == 0) 8281 panic("softdep_write_inodeblock: lost dep2"); 8282#endif /* INVARIANTS */ 8283 dp->di_ib[i] = 0; 8284 } 8285 return; 8286 } 8287 /* 8288 * If we have zero'ed out the last allocated block of the file, 8289 * roll back the size to the last currently allocated block. 8290 * We know that this last allocated block is a full-sized as 8291 * we already checked for fragments in the loop above. 8292 */ 8293 if (lastadp != NULL && 8294 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 8295 for (i = lastadp->ad_offset; i >= 0; i--) 8296 if (dp->di_db[i] != 0) 8297 break; 8298 dp->di_size = (i + 1) * fs->fs_bsize; 8299 } 8300 /* 8301 * The only dependencies are for indirect blocks. 8302 * 8303 * The file size for indirect block additions is not guaranteed. 8304 * Such a guarantee would be non-trivial to achieve. The conventional 8305 * synchronous write implementation also does not make this guarantee. 8306 * Fsck should catch and fix discrepancies. Arguably, the file size 8307 * can be over-estimated without destroying integrity when the file 8308 * moves into the indirect blocks (i.e., is large). If we want to 8309 * postpone fsck, we are stuck with this argument. 8310 */ 8311 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 8312 dp->di_ib[adp->ad_offset - NDADDR] = 0; 8313} 8314 8315/* 8316 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 8317 * Note that any bug fixes made to this routine must be done in the 8318 * version found above. 8319 * 8320 * Called from within the procedure above to deal with unsatisfied 8321 * allocation dependencies in an inodeblock. The buffer must be 8322 * locked, thus, no I/O completion operations can occur while we 8323 * are manipulating its associated dependencies. 8324 */ 8325static void 8326initiate_write_inodeblock_ufs2(inodedep, bp) 8327 struct inodedep *inodedep; 8328 struct buf *bp; /* The inode block */ 8329{ 8330 struct allocdirect *adp, *lastadp; 8331 struct ufs2_dinode *dp; 8332 struct ufs2_dinode *sip; 8333 struct inoref *inoref; 8334 struct fs *fs; 8335 ufs_lbn_t i; 8336#ifdef INVARIANTS 8337 ufs_lbn_t prevlbn = 0; 8338#endif 8339 int deplist; 8340 8341 if (inodedep->id_state & IOSTARTED) 8342 panic("initiate_write_inodeblock_ufs2: already started"); 8343 inodedep->id_state |= IOSTARTED; 8344 fs = inodedep->id_fs; 8345 dp = (struct ufs2_dinode *)bp->b_data + 8346 ino_to_fsbo(fs, inodedep->id_ino); 8347 8348 /* 8349 * If we're on the unlinked list but have not yet written our 8350 * next pointer initialize it here. 8351 */ 8352 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 8353 struct inodedep *inon; 8354 8355 inon = TAILQ_NEXT(inodedep, id_unlinked); 8356 dp->di_freelink = inon ? inon->id_ino : 0; 8357 } 8358 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == 8359 (UNLINKED | UNLINKNEXT)) { 8360 struct inodedep *inon; 8361 ino_t freelink; 8362 8363 inon = TAILQ_NEXT(inodedep, id_unlinked); 8364 freelink = inon ? inon->id_ino : 0; 8365 if (freelink != dp->di_freelink) 8366 panic("ino %p(0x%X) %d, %d != %d", 8367 inodedep, inodedep->id_state, inodedep->id_ino, 8368 freelink, dp->di_freelink); 8369 } 8370 /* 8371 * If the bitmap is not yet written, then the allocated 8372 * inode cannot be written to disk. 8373 */ 8374 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 8375 if (inodedep->id_savedino2 != NULL) 8376 panic("initiate_write_inodeblock_ufs2: I/O underway"); 8377 FREE_LOCK(&lk); 8378 sip = malloc(sizeof(struct ufs2_dinode), 8379 M_SAVEDINO, M_SOFTDEP_FLAGS); 8380 ACQUIRE_LOCK(&lk); 8381 inodedep->id_savedino2 = sip; 8382 *inodedep->id_savedino2 = *dp; 8383 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 8384 dp->di_gen = inodedep->id_savedino2->di_gen; 8385 dp->di_freelink = inodedep->id_savedino2->di_freelink; 8386 return; 8387 } 8388 /* 8389 * If no dependencies, then there is nothing to roll back. 8390 */ 8391 inodedep->id_savedsize = dp->di_size; 8392 inodedep->id_savedextsize = dp->di_extsize; 8393 inodedep->id_savednlink = dp->di_nlink; 8394 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 8395 TAILQ_EMPTY(&inodedep->id_extupdt) && 8396 TAILQ_EMPTY(&inodedep->id_inoreflst)) 8397 return; 8398 /* 8399 * Revert the link count to that of the first unwritten journal entry. 8400 */ 8401 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 8402 if (inoref) 8403 dp->di_nlink = inoref->if_nlink; 8404 8405 /* 8406 * Set the ext data dependencies to busy. 8407 */ 8408 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 8409 adp = TAILQ_NEXT(adp, ad_next)) { 8410#ifdef INVARIANTS 8411 if (deplist != 0 && prevlbn >= adp->ad_offset) 8412 panic("softdep_write_inodeblock: lbn order"); 8413 prevlbn = adp->ad_offset; 8414 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 8415 panic("%s: direct pointer #%jd mismatch %jd != %jd", 8416 "softdep_write_inodeblock", 8417 (intmax_t)adp->ad_offset, 8418 (intmax_t)dp->di_extb[adp->ad_offset], 8419 (intmax_t)adp->ad_newblkno); 8420 deplist |= 1 << adp->ad_offset; 8421 if ((adp->ad_state & ATTACHED) == 0) 8422 panic("softdep_write_inodeblock: Unknown state 0x%x", 8423 adp->ad_state); 8424#endif /* INVARIANTS */ 8425 adp->ad_state &= ~ATTACHED; 8426 adp->ad_state |= UNDONE; 8427 } 8428 /* 8429 * The on-disk inode cannot claim to be any larger than the last 8430 * fragment that has been written. Otherwise, the on-disk inode 8431 * might have fragments that were not the last block in the ext 8432 * data which would corrupt the filesystem. 8433 */ 8434 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 8435 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 8436 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 8437 /* keep going until hitting a rollback to a frag */ 8438 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 8439 continue; 8440 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 8441 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 8442#ifdef INVARIANTS 8443 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 8444 panic("softdep_write_inodeblock: lost dep1"); 8445#endif /* INVARIANTS */ 8446 dp->di_extb[i] = 0; 8447 } 8448 lastadp = NULL; 8449 break; 8450 } 8451 /* 8452 * If we have zero'ed out the last allocated block of the ext 8453 * data, roll back the size to the last currently allocated block. 8454 * We know that this last allocated block is a full-sized as 8455 * we already checked for fragments in the loop above. 8456 */ 8457 if (lastadp != NULL && 8458 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 8459 for (i = lastadp->ad_offset; i >= 0; i--) 8460 if (dp->di_extb[i] != 0) 8461 break; 8462 dp->di_extsize = (i + 1) * fs->fs_bsize; 8463 } 8464 /* 8465 * Set the file data dependencies to busy. 8466 */ 8467 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 8468 adp = TAILQ_NEXT(adp, ad_next)) { 8469#ifdef INVARIANTS 8470 if (deplist != 0 && prevlbn >= adp->ad_offset) 8471 panic("softdep_write_inodeblock: lbn order"); 8472 prevlbn = adp->ad_offset; 8473 if (adp->ad_offset < NDADDR && 8474 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 8475 panic("%s: direct pointer #%jd mismatch %jd != %jd", 8476 "softdep_write_inodeblock", 8477 (intmax_t)adp->ad_offset, 8478 (intmax_t)dp->di_db[adp->ad_offset], 8479 (intmax_t)adp->ad_newblkno); 8480 if (adp->ad_offset >= NDADDR && 8481 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 8482 panic("%s indirect pointer #%jd mismatch %jd != %jd", 8483 "softdep_write_inodeblock:", 8484 (intmax_t)adp->ad_offset - NDADDR, 8485 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 8486 (intmax_t)adp->ad_newblkno); 8487 deplist |= 1 << adp->ad_offset; 8488 if ((adp->ad_state & ATTACHED) == 0) 8489 panic("softdep_write_inodeblock: Unknown state 0x%x", 8490 adp->ad_state); 8491#endif /* INVARIANTS */ 8492 adp->ad_state &= ~ATTACHED; 8493 adp->ad_state |= UNDONE; 8494 } 8495 /* 8496 * The on-disk inode cannot claim to be any larger than the last 8497 * fragment that has been written. Otherwise, the on-disk inode 8498 * might have fragments that were not the last block in the file 8499 * which would corrupt the filesystem. 8500 */ 8501 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 8502 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 8503 if (adp->ad_offset >= NDADDR) 8504 break; 8505 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 8506 /* keep going until hitting a rollback to a frag */ 8507 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 8508 continue; 8509 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 8510 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 8511#ifdef INVARIANTS 8512 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 8513 panic("softdep_write_inodeblock: lost dep2"); 8514#endif /* INVARIANTS */ 8515 dp->di_db[i] = 0; 8516 } 8517 for (i = 0; i < NIADDR; i++) { 8518#ifdef INVARIANTS 8519 if (dp->di_ib[i] != 0 && 8520 (deplist & ((1 << NDADDR) << i)) == 0) 8521 panic("softdep_write_inodeblock: lost dep3"); 8522#endif /* INVARIANTS */ 8523 dp->di_ib[i] = 0; 8524 } 8525 return; 8526 } 8527 /* 8528 * If we have zero'ed out the last allocated block of the file, 8529 * roll back the size to the last currently allocated block. 8530 * We know that this last allocated block is a full-sized as 8531 * we already checked for fragments in the loop above. 8532 */ 8533 if (lastadp != NULL && 8534 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 8535 for (i = lastadp->ad_offset; i >= 0; i--) 8536 if (dp->di_db[i] != 0) 8537 break; 8538 dp->di_size = (i + 1) * fs->fs_bsize; 8539 } 8540 /* 8541 * The only dependencies are for indirect blocks. 8542 * 8543 * The file size for indirect block additions is not guaranteed. 8544 * Such a guarantee would be non-trivial to achieve. The conventional 8545 * synchronous write implementation also does not make this guarantee. 8546 * Fsck should catch and fix discrepancies. Arguably, the file size 8547 * can be over-estimated without destroying integrity when the file 8548 * moves into the indirect blocks (i.e., is large). If we want to 8549 * postpone fsck, we are stuck with this argument. 8550 */ 8551 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 8552 dp->di_ib[adp->ad_offset - NDADDR] = 0; 8553} 8554 8555/* 8556 * Cancel an indirdep as a result of truncation. Release all of the 8557 * children allocindirs and place their journal work on the appropriate 8558 * list. 8559 */ 8560static void 8561cancel_indirdep(indirdep, bp, inodedep, freeblks) 8562 struct indirdep *indirdep; 8563 struct buf *bp; 8564 struct inodedep *inodedep; 8565 struct freeblks *freeblks; 8566{ 8567 struct allocindir *aip; 8568 8569 /* 8570 * None of the indirect pointers will ever be visible, 8571 * so they can simply be tossed. GOINGAWAY ensures 8572 * that allocated pointers will be saved in the buffer 8573 * cache until they are freed. Note that they will 8574 * only be able to be found by their physical address 8575 * since the inode mapping the logical address will 8576 * be gone. The save buffer used for the safe copy 8577 * was allocated in setup_allocindir_phase2 using 8578 * the physical address so it could be used for this 8579 * purpose. Hence we swap the safe copy with the real 8580 * copy, allowing the safe copy to be freed and holding 8581 * on to the real copy for later use in indir_trunc. 8582 */ 8583 if (indirdep->ir_state & GOINGAWAY) 8584 panic("cancel_indirdep: already gone"); 8585 if (indirdep->ir_state & ONDEPLIST) { 8586 indirdep->ir_state &= ~ONDEPLIST; 8587 LIST_REMOVE(indirdep, ir_next); 8588 } 8589 indirdep->ir_state |= GOINGAWAY; 8590 VFSTOUFS(indirdep->ir_list.wk_mp)->um_numindirdeps += 1; 8591 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 8592 cancel_allocindir(aip, inodedep, freeblks); 8593 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) 8594 cancel_allocindir(aip, inodedep, freeblks); 8595 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) 8596 cancel_allocindir(aip, inodedep, freeblks); 8597 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0) 8598 cancel_allocindir(aip, inodedep, freeblks); 8599 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 8600 WORKLIST_REMOVE(&indirdep->ir_list); 8601 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 8602 indirdep->ir_savebp = NULL; 8603} 8604 8605/* 8606 * Free an indirdep once it no longer has new pointers to track. 8607 */ 8608static void 8609free_indirdep(indirdep) 8610 struct indirdep *indirdep; 8611{ 8612 8613 KASSERT(LIST_EMPTY(&indirdep->ir_jwork), 8614 ("free_indirdep: Journal work not empty.")); 8615 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 8616 ("free_indirdep: Complete head not empty.")); 8617 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 8618 ("free_indirdep: write head not empty.")); 8619 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 8620 ("free_indirdep: done head not empty.")); 8621 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 8622 ("free_indirdep: deplist head not empty.")); 8623 KASSERT(indirdep->ir_savebp == NULL, 8624 ("free_indirdep: %p ir_savebp != NULL", indirdep)); 8625 KASSERT((indirdep->ir_state & ONDEPLIST) == 0, 8626 ("free_indirdep: %p still on deplist.", indirdep)); 8627 if (indirdep->ir_state & ONWORKLIST) 8628 WORKLIST_REMOVE(&indirdep->ir_list); 8629 WORKITEM_FREE(indirdep, D_INDIRDEP); 8630} 8631 8632/* 8633 * Called before a write to an indirdep. This routine is responsible for 8634 * rolling back pointers to a safe state which includes only those 8635 * allocindirs which have been completed. 8636 */ 8637static void 8638initiate_write_indirdep(indirdep, bp) 8639 struct indirdep *indirdep; 8640 struct buf *bp; 8641{ 8642 8643 if (indirdep->ir_state & GOINGAWAY) 8644 panic("disk_io_initiation: indirdep gone"); 8645 8646 /* 8647 * If there are no remaining dependencies, this will be writing 8648 * the real pointers. 8649 */ 8650 if (LIST_EMPTY(&indirdep->ir_deplisthd)) 8651 return; 8652 /* 8653 * Replace up-to-date version with safe version. 8654 */ 8655 FREE_LOCK(&lk); 8656 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 8657 M_SOFTDEP_FLAGS); 8658 ACQUIRE_LOCK(&lk); 8659 indirdep->ir_state &= ~ATTACHED; 8660 indirdep->ir_state |= UNDONE; 8661 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 8662 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 8663 bp->b_bcount); 8664} 8665 8666/* 8667 * Called when an inode has been cleared in a cg bitmap. This finally 8668 * eliminates any canceled jaddrefs 8669 */ 8670void 8671softdep_setup_inofree(mp, bp, ino, wkhd) 8672 struct mount *mp; 8673 struct buf *bp; 8674 ino_t ino; 8675 struct workhead *wkhd; 8676{ 8677 struct worklist *wk, *wkn; 8678 struct inodedep *inodedep; 8679 uint8_t *inosused; 8680 struct cg *cgp; 8681 struct fs *fs; 8682 8683 ACQUIRE_LOCK(&lk); 8684 fs = VFSTOUFS(mp)->um_fs; 8685 cgp = (struct cg *)bp->b_data; 8686 inosused = cg_inosused(cgp); 8687 if (isset(inosused, ino % fs->fs_ipg)) 8688 panic("softdep_setup_inofree: inode %d not freed.", ino); 8689 if (inodedep_lookup(mp, ino, 0, &inodedep)) 8690 panic("softdep_setup_inofree: ino %d has existing inodedep %p", 8691 ino, inodedep); 8692 if (wkhd) { 8693 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 8694 if (wk->wk_type != D_JADDREF) 8695 continue; 8696 WORKLIST_REMOVE(wk); 8697 /* 8698 * We can free immediately even if the jaddref 8699 * isn't attached in a background write as now 8700 * the bitmaps are reconciled. 8701 */ 8702 wk->wk_state |= COMPLETE | ATTACHED; 8703 free_jaddref(WK_JADDREF(wk)); 8704 } 8705 jwork_move(&bp->b_dep, wkhd); 8706 } 8707 FREE_LOCK(&lk); 8708} 8709 8710 8711/* 8712 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 8713 * map. Any dependencies waiting for the write to clear are added to the 8714 * buf's list and any jnewblks that are being canceled are discarded 8715 * immediately. 8716 */ 8717void 8718softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 8719 struct mount *mp; 8720 struct buf *bp; 8721 ufs2_daddr_t blkno; 8722 int frags; 8723 struct workhead *wkhd; 8724{ 8725 struct jnewblk *jnewblk; 8726 struct worklist *wk, *wkn; 8727#ifdef SUJ_DEBUG 8728 struct bmsafemap *bmsafemap; 8729 struct fs *fs; 8730 uint8_t *blksfree; 8731 struct cg *cgp; 8732 ufs2_daddr_t jstart; 8733 ufs2_daddr_t jend; 8734 ufs2_daddr_t end; 8735 long bno; 8736 int i; 8737#endif 8738 8739 ACQUIRE_LOCK(&lk); 8740 /* 8741 * Detach any jnewblks which have been canceled. They must linger 8742 * until the bitmap is cleared again by ffs_blkfree() to prevent 8743 * an unjournaled allocation from hitting the disk. 8744 */ 8745 if (wkhd) { 8746 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 8747 if (wk->wk_type != D_JNEWBLK) 8748 continue; 8749 jnewblk = WK_JNEWBLK(wk); 8750 KASSERT(jnewblk->jn_state & GOINGAWAY, 8751 ("softdep_setup_blkfree: jnewblk not canceled.")); 8752 WORKLIST_REMOVE(wk); 8753#ifdef SUJ_DEBUG 8754 /* 8755 * Assert that this block is free in the bitmap 8756 * before we discard the jnewblk. 8757 */ 8758 fs = VFSTOUFS(mp)->um_fs; 8759 cgp = (struct cg *)bp->b_data; 8760 blksfree = cg_blksfree(cgp); 8761 bno = dtogd(fs, jnewblk->jn_blkno); 8762 for (i = jnewblk->jn_oldfrags; 8763 i < jnewblk->jn_frags; i++) { 8764 if (isset(blksfree, bno + i)) 8765 continue; 8766 panic("softdep_setup_blkfree: not free"); 8767 } 8768#endif 8769 /* 8770 * Even if it's not attached we can free immediately 8771 * as the new bitmap is correct. 8772 */ 8773 wk->wk_state |= COMPLETE | ATTACHED; 8774 free_jnewblk(jnewblk); 8775 } 8776 /* 8777 * The buf must be locked by the caller otherwise these could 8778 * be added while it's being written and the write would 8779 * complete them before they made it to disk. 8780 */ 8781 jwork_move(&bp->b_dep, wkhd); 8782 } 8783 8784#ifdef SUJ_DEBUG 8785 /* 8786 * Assert that we are not freeing a block which has an outstanding 8787 * allocation dependency. 8788 */ 8789 fs = VFSTOUFS(mp)->um_fs; 8790 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno)); 8791 end = blkno + frags; 8792 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 8793 /* 8794 * Don't match against blocks that will be freed when the 8795 * background write is done. 8796 */ 8797 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 8798 (COMPLETE | DEPCOMPLETE)) 8799 continue; 8800 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 8801 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 8802 if ((blkno >= jstart && blkno < jend) || 8803 (end > jstart && end <= jend)) { 8804 printf("state 0x%X %jd - %d %d dep %p\n", 8805 jnewblk->jn_state, jnewblk->jn_blkno, 8806 jnewblk->jn_oldfrags, jnewblk->jn_frags, 8807 jnewblk->jn_newblk); 8808 panic("softdep_setup_blkfree: " 8809 "%jd-%jd(%d) overlaps with %jd-%jd", 8810 blkno, end, frags, jstart, jend); 8811 } 8812 } 8813#endif 8814 FREE_LOCK(&lk); 8815} 8816 8817static void 8818initiate_write_bmsafemap(bmsafemap, bp) 8819 struct bmsafemap *bmsafemap; 8820 struct buf *bp; /* The cg block. */ 8821{ 8822 struct jaddref *jaddref; 8823 struct jnewblk *jnewblk; 8824 uint8_t *inosused; 8825 uint8_t *blksfree; 8826 struct cg *cgp; 8827 struct fs *fs; 8828 int cleared; 8829 ino_t ino; 8830 long bno; 8831 int i; 8832 8833 if (bmsafemap->sm_state & IOSTARTED) 8834 panic("initiate_write_bmsafemap: Already started\n"); 8835 bmsafemap->sm_state |= IOSTARTED; 8836 /* 8837 * Clear any inode allocations which are pending journal writes. 8838 */ 8839 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 8840 cgp = (struct cg *)bp->b_data; 8841 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 8842 inosused = cg_inosused(cgp); 8843 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 8844 ino = jaddref->ja_ino % fs->fs_ipg; 8845 /* 8846 * If this is a background copy the inode may not 8847 * be marked used yet. 8848 */ 8849 if (isset(inosused, ino)) { 8850 if ((jaddref->ja_mode & IFMT) == IFDIR) 8851 cgp->cg_cs.cs_ndir--; 8852 cgp->cg_cs.cs_nifree++; 8853 clrbit(inosused, ino); 8854 jaddref->ja_state &= ~ATTACHED; 8855 jaddref->ja_state |= UNDONE; 8856 stat_jaddref++; 8857 } else if ((bp->b_xflags & BX_BKGRDMARKER) == 0) 8858 panic("initiate_write_bmsafemap: inode %d " 8859 "marked free", jaddref->ja_ino); 8860 } 8861 } 8862 /* 8863 * Clear any block allocations which are pending journal writes. 8864 */ 8865 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 8866 cgp = (struct cg *)bp->b_data; 8867 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 8868 blksfree = cg_blksfree(cgp); 8869 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 8870 bno = dtogd(fs, jnewblk->jn_blkno); 8871 cleared = 0; 8872 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 8873 i++) { 8874 if (isclr(blksfree, bno + i)) { 8875 cleared = 1; 8876 setbit(blksfree, bno + i); 8877 } 8878 } 8879 /* 8880 * We may not clear the block if it's a background 8881 * copy. In that case there is no reason to detach 8882 * it. 8883 */ 8884 if (cleared) { 8885 stat_jnewblk++; 8886 jnewblk->jn_state &= ~ATTACHED; 8887 jnewblk->jn_state |= UNDONE; 8888 } else if ((bp->b_xflags & BX_BKGRDMARKER) == 0) 8889 panic("initiate_write_bmsafemap: block %jd " 8890 "marked free", jnewblk->jn_blkno); 8891 } 8892 } 8893 /* 8894 * Move allocation lists to the written lists so they can be 8895 * cleared once the block write is complete. 8896 */ 8897 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 8898 inodedep, id_deps); 8899 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 8900 newblk, nb_deps); 8901} 8902 8903/* 8904 * This routine is called during the completion interrupt 8905 * service routine for a disk write (from the procedure called 8906 * by the device driver to inform the filesystem caches of 8907 * a request completion). It should be called early in this 8908 * procedure, before the block is made available to other 8909 * processes or other routines are called. 8910 * 8911 */ 8912static void 8913softdep_disk_write_complete(bp) 8914 struct buf *bp; /* describes the completed disk write */ 8915{ 8916 struct worklist *wk; 8917 struct worklist *owk; 8918 struct workhead reattach; 8919 struct buf *sbp; 8920 8921 /* 8922 * If an error occurred while doing the write, then the data 8923 * has not hit the disk and the dependencies cannot be unrolled. 8924 */ 8925 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 8926 return; 8927 LIST_INIT(&reattach); 8928 /* 8929 * This lock must not be released anywhere in this code segment. 8930 */ 8931 sbp = NULL; 8932 owk = NULL; 8933 ACQUIRE_LOCK(&lk); 8934 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 8935 WORKLIST_REMOVE(wk); 8936 if (wk == owk) 8937 panic("duplicate worklist: %p\n", wk); 8938 owk = wk; 8939 switch (wk->wk_type) { 8940 8941 case D_PAGEDEP: 8942 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 8943 WORKLIST_INSERT(&reattach, wk); 8944 continue; 8945 8946 case D_INODEDEP: 8947 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 8948 WORKLIST_INSERT(&reattach, wk); 8949 continue; 8950 8951 case D_BMSAFEMAP: 8952 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp)) 8953 WORKLIST_INSERT(&reattach, wk); 8954 continue; 8955 8956 case D_MKDIR: 8957 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 8958 continue; 8959 8960 case D_ALLOCDIRECT: 8961 wk->wk_state |= COMPLETE; 8962 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 8963 continue; 8964 8965 case D_ALLOCINDIR: 8966 wk->wk_state |= COMPLETE; 8967 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 8968 continue; 8969 8970 case D_INDIRDEP: 8971 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp)) 8972 WORKLIST_INSERT(&reattach, wk); 8973 continue; 8974 8975 case D_FREEBLKS: 8976 wk->wk_state |= COMPLETE; 8977 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 8978 add_to_worklist(wk, 1); 8979 continue; 8980 8981 case D_FREEWORK: 8982 handle_written_freework(WK_FREEWORK(wk)); 8983 break; 8984 8985 case D_FREEDEP: 8986 free_freedep(WK_FREEDEP(wk)); 8987 continue; 8988 8989 case D_JSEGDEP: 8990 free_jsegdep(WK_JSEGDEP(wk)); 8991 continue; 8992 8993 case D_JSEG: 8994 handle_written_jseg(WK_JSEG(wk), bp); 8995 continue; 8996 8997 case D_SBDEP: 8998 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 8999 WORKLIST_INSERT(&reattach, wk); 9000 continue; 9001 9002 default: 9003 panic("handle_disk_write_complete: Unknown type %s", 9004 TYPENAME(wk->wk_type)); 9005 /* NOTREACHED */ 9006 } 9007 } 9008 /* 9009 * Reattach any requests that must be redone. 9010 */ 9011 while ((wk = LIST_FIRST(&reattach)) != NULL) { 9012 WORKLIST_REMOVE(wk); 9013 WORKLIST_INSERT(&bp->b_dep, wk); 9014 } 9015 FREE_LOCK(&lk); 9016 if (sbp) 9017 brelse(sbp); 9018} 9019 9020/* 9021 * Called from within softdep_disk_write_complete above. Note that 9022 * this routine is always called from interrupt level with further 9023 * splbio interrupts blocked. 9024 */ 9025static void 9026handle_allocdirect_partdone(adp, wkhd) 9027 struct allocdirect *adp; /* the completed allocdirect */ 9028 struct workhead *wkhd; /* Work to do when inode is writtne. */ 9029{ 9030 struct allocdirectlst *listhead; 9031 struct allocdirect *listadp; 9032 struct inodedep *inodedep; 9033 long bsize; 9034 9035 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 9036 return; 9037 /* 9038 * The on-disk inode cannot claim to be any larger than the last 9039 * fragment that has been written. Otherwise, the on-disk inode 9040 * might have fragments that were not the last block in the file 9041 * which would corrupt the filesystem. Thus, we cannot free any 9042 * allocdirects after one whose ad_oldblkno claims a fragment as 9043 * these blocks must be rolled back to zero before writing the inode. 9044 * We check the currently active set of allocdirects in id_inoupdt 9045 * or id_extupdt as appropriate. 9046 */ 9047 inodedep = adp->ad_inodedep; 9048 bsize = inodedep->id_fs->fs_bsize; 9049 if (adp->ad_state & EXTDATA) 9050 listhead = &inodedep->id_extupdt; 9051 else 9052 listhead = &inodedep->id_inoupdt; 9053 TAILQ_FOREACH(listadp, listhead, ad_next) { 9054 /* found our block */ 9055 if (listadp == adp) 9056 break; 9057 /* continue if ad_oldlbn is not a fragment */ 9058 if (listadp->ad_oldsize == 0 || 9059 listadp->ad_oldsize == bsize) 9060 continue; 9061 /* hit a fragment */ 9062 return; 9063 } 9064 /* 9065 * If we have reached the end of the current list without 9066 * finding the just finished dependency, then it must be 9067 * on the future dependency list. Future dependencies cannot 9068 * be freed until they are moved to the current list. 9069 */ 9070 if (listadp == NULL) { 9071#ifdef DEBUG 9072 if (adp->ad_state & EXTDATA) 9073 listhead = &inodedep->id_newextupdt; 9074 else 9075 listhead = &inodedep->id_newinoupdt; 9076 TAILQ_FOREACH(listadp, listhead, ad_next) 9077 /* found our block */ 9078 if (listadp == adp) 9079 break; 9080 if (listadp == NULL) 9081 panic("handle_allocdirect_partdone: lost dep"); 9082#endif /* DEBUG */ 9083 return; 9084 } 9085 /* 9086 * If we have found the just finished dependency, then queue 9087 * it along with anything that follows it that is complete. 9088 * Since the pointer has not yet been written in the inode 9089 * as the dependency prevents it, place the allocdirect on the 9090 * bufwait list where it will be freed once the pointer is 9091 * valid. 9092 */ 9093 if (wkhd == NULL) 9094 wkhd = &inodedep->id_bufwait; 9095 for (; adp; adp = listadp) { 9096 listadp = TAILQ_NEXT(adp, ad_next); 9097 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 9098 return; 9099 TAILQ_REMOVE(listhead, adp, ad_next); 9100 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 9101 } 9102} 9103 9104/* 9105 * Called from within softdep_disk_write_complete above. This routine 9106 * completes successfully written allocindirs. 9107 */ 9108static void 9109handle_allocindir_partdone(aip) 9110 struct allocindir *aip; /* the completed allocindir */ 9111{ 9112 struct indirdep *indirdep; 9113 9114 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 9115 return; 9116 indirdep = aip->ai_indirdep; 9117 LIST_REMOVE(aip, ai_next); 9118 if (indirdep->ir_state & UNDONE) { 9119 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 9120 return; 9121 } 9122 if (indirdep->ir_state & UFS1FMT) 9123 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 9124 aip->ai_newblkno; 9125 else 9126 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 9127 aip->ai_newblkno; 9128 /* 9129 * Await the pointer write before freeing the allocindir. 9130 */ 9131 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 9132} 9133 9134/* 9135 * Release segments held on a jwork list. 9136 */ 9137static void 9138handle_jwork(wkhd) 9139 struct workhead *wkhd; 9140{ 9141 struct worklist *wk; 9142 9143 while ((wk = LIST_FIRST(wkhd)) != NULL) { 9144 WORKLIST_REMOVE(wk); 9145 switch (wk->wk_type) { 9146 case D_JSEGDEP: 9147 free_jsegdep(WK_JSEGDEP(wk)); 9148 continue; 9149 default: 9150 panic("handle_jwork: Unknown type %s\n", 9151 TYPENAME(wk->wk_type)); 9152 } 9153 } 9154} 9155 9156/* 9157 * Handle the bufwait list on an inode when it is safe to release items 9158 * held there. This normally happens after an inode block is written but 9159 * may be delayed and handle later if there are pending journal items that 9160 * are not yet safe to be released. 9161 */ 9162static struct freefile * 9163handle_bufwait(inodedep, refhd) 9164 struct inodedep *inodedep; 9165 struct workhead *refhd; 9166{ 9167 struct jaddref *jaddref; 9168 struct freefile *freefile; 9169 struct worklist *wk; 9170 9171 freefile = NULL; 9172 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 9173 WORKLIST_REMOVE(wk); 9174 switch (wk->wk_type) { 9175 case D_FREEFILE: 9176 /* 9177 * We defer adding freefile to the worklist 9178 * until all other additions have been made to 9179 * ensure that it will be done after all the 9180 * old blocks have been freed. 9181 */ 9182 if (freefile != NULL) 9183 panic("handle_bufwait: freefile"); 9184 freefile = WK_FREEFILE(wk); 9185 continue; 9186 9187 case D_MKDIR: 9188 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 9189 continue; 9190 9191 case D_DIRADD: 9192 diradd_inode_written(WK_DIRADD(wk), inodedep); 9193 continue; 9194 9195 case D_FREEFRAG: 9196 wk->wk_state |= COMPLETE; 9197 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 9198 add_to_worklist(wk, 0); 9199 continue; 9200 9201 case D_DIRREM: 9202 wk->wk_state |= COMPLETE; 9203 add_to_worklist(wk, 0); 9204 continue; 9205 9206 case D_ALLOCDIRECT: 9207 case D_ALLOCINDIR: 9208 free_newblk(WK_NEWBLK(wk)); 9209 continue; 9210 9211 case D_JNEWBLK: 9212 wk->wk_state |= COMPLETE; 9213 free_jnewblk(WK_JNEWBLK(wk)); 9214 continue; 9215 9216 /* 9217 * Save freed journal segments and add references on 9218 * the supplied list which will delay their release 9219 * until the cg bitmap is cleared on disk. 9220 */ 9221 case D_JSEGDEP: 9222 if (refhd == NULL) 9223 free_jsegdep(WK_JSEGDEP(wk)); 9224 else 9225 WORKLIST_INSERT(refhd, wk); 9226 continue; 9227 9228 case D_JADDREF: 9229 jaddref = WK_JADDREF(wk); 9230 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 9231 if_deps); 9232 /* 9233 * Transfer any jaddrefs to the list to be freed with 9234 * the bitmap if we're handling a removed file. 9235 */ 9236 if (refhd == NULL) { 9237 wk->wk_state |= COMPLETE; 9238 free_jaddref(jaddref); 9239 } else 9240 WORKLIST_INSERT(refhd, wk); 9241 continue; 9242 9243 default: 9244 panic("handle_bufwait: Unknown type %p(%s)", 9245 wk, TYPENAME(wk->wk_type)); 9246 /* NOTREACHED */ 9247 } 9248 } 9249 return (freefile); 9250} 9251/* 9252 * Called from within softdep_disk_write_complete above to restore 9253 * in-memory inode block contents to their most up-to-date state. Note 9254 * that this routine is always called from interrupt level with further 9255 * splbio interrupts blocked. 9256 */ 9257static int 9258handle_written_inodeblock(inodedep, bp) 9259 struct inodedep *inodedep; 9260 struct buf *bp; /* buffer containing the inode block */ 9261{ 9262 struct freefile *freefile; 9263 struct allocdirect *adp, *nextadp; 9264 struct ufs1_dinode *dp1 = NULL; 9265 struct ufs2_dinode *dp2 = NULL; 9266 struct workhead wkhd; 9267 int hadchanges, fstype; 9268 ino_t freelink; 9269 9270 LIST_INIT(&wkhd); 9271 hadchanges = 0; 9272 freefile = NULL; 9273 if ((inodedep->id_state & IOSTARTED) == 0) 9274 panic("handle_written_inodeblock: not started"); 9275 inodedep->id_state &= ~IOSTARTED; 9276 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 9277 fstype = UFS1; 9278 dp1 = (struct ufs1_dinode *)bp->b_data + 9279 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 9280 freelink = dp1->di_freelink; 9281 } else { 9282 fstype = UFS2; 9283 dp2 = (struct ufs2_dinode *)bp->b_data + 9284 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 9285 freelink = dp2->di_freelink; 9286 } 9287 /* 9288 * If we wrote a valid freelink pointer during the last write 9289 * record it here. 9290 */ 9291 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 9292 struct inodedep *inon; 9293 9294 inon = TAILQ_NEXT(inodedep, id_unlinked); 9295 if ((inon == NULL && freelink == 0) || 9296 (inon && inon->id_ino == freelink)) { 9297 if (inon) 9298 inon->id_state |= UNLINKPREV; 9299 inodedep->id_state |= UNLINKNEXT; 9300 } else 9301 hadchanges = 1; 9302 } 9303 /* Leave this inodeblock dirty until it's in the list. */ 9304 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) 9305 hadchanges = 1; 9306 /* 9307 * If we had to rollback the inode allocation because of 9308 * bitmaps being incomplete, then simply restore it. 9309 * Keep the block dirty so that it will not be reclaimed until 9310 * all associated dependencies have been cleared and the 9311 * corresponding updates written to disk. 9312 */ 9313 if (inodedep->id_savedino1 != NULL) { 9314 hadchanges = 1; 9315 if (fstype == UFS1) 9316 *dp1 = *inodedep->id_savedino1; 9317 else 9318 *dp2 = *inodedep->id_savedino2; 9319 free(inodedep->id_savedino1, M_SAVEDINO); 9320 inodedep->id_savedino1 = NULL; 9321 if ((bp->b_flags & B_DELWRI) == 0) 9322 stat_inode_bitmap++; 9323 bdirty(bp); 9324 /* 9325 * If the inode is clear here and GOINGAWAY it will never 9326 * be written. Process the bufwait and clear any pending 9327 * work which may include the freefile. 9328 */ 9329 if (inodedep->id_state & GOINGAWAY) 9330 goto bufwait; 9331 return (1); 9332 } 9333 inodedep->id_state |= COMPLETE; 9334 /* 9335 * Roll forward anything that had to be rolled back before 9336 * the inode could be updated. 9337 */ 9338 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 9339 nextadp = TAILQ_NEXT(adp, ad_next); 9340 if (adp->ad_state & ATTACHED) 9341 panic("handle_written_inodeblock: new entry"); 9342 if (fstype == UFS1) { 9343 if (adp->ad_offset < NDADDR) { 9344 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 9345 panic("%s %s #%jd mismatch %d != %jd", 9346 "handle_written_inodeblock:", 9347 "direct pointer", 9348 (intmax_t)adp->ad_offset, 9349 dp1->di_db[adp->ad_offset], 9350 (intmax_t)adp->ad_oldblkno); 9351 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 9352 } else { 9353 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 9354 panic("%s: %s #%jd allocated as %d", 9355 "handle_written_inodeblock", 9356 "indirect pointer", 9357 (intmax_t)adp->ad_offset - NDADDR, 9358 dp1->di_ib[adp->ad_offset - NDADDR]); 9359 dp1->di_ib[adp->ad_offset - NDADDR] = 9360 adp->ad_newblkno; 9361 } 9362 } else { 9363 if (adp->ad_offset < NDADDR) { 9364 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 9365 panic("%s: %s #%jd %s %jd != %jd", 9366 "handle_written_inodeblock", 9367 "direct pointer", 9368 (intmax_t)adp->ad_offset, "mismatch", 9369 (intmax_t)dp2->di_db[adp->ad_offset], 9370 (intmax_t)adp->ad_oldblkno); 9371 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 9372 } else { 9373 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 9374 panic("%s: %s #%jd allocated as %jd", 9375 "handle_written_inodeblock", 9376 "indirect pointer", 9377 (intmax_t)adp->ad_offset - NDADDR, 9378 (intmax_t) 9379 dp2->di_ib[adp->ad_offset - NDADDR]); 9380 dp2->di_ib[adp->ad_offset - NDADDR] = 9381 adp->ad_newblkno; 9382 } 9383 } 9384 adp->ad_state &= ~UNDONE; 9385 adp->ad_state |= ATTACHED; 9386 hadchanges = 1; 9387 } 9388 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 9389 nextadp = TAILQ_NEXT(adp, ad_next); 9390 if (adp->ad_state & ATTACHED) 9391 panic("handle_written_inodeblock: new entry"); 9392 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 9393 panic("%s: direct pointers #%jd %s %jd != %jd", 9394 "handle_written_inodeblock", 9395 (intmax_t)adp->ad_offset, "mismatch", 9396 (intmax_t)dp2->di_extb[adp->ad_offset], 9397 (intmax_t)adp->ad_oldblkno); 9398 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 9399 adp->ad_state &= ~UNDONE; 9400 adp->ad_state |= ATTACHED; 9401 hadchanges = 1; 9402 } 9403 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 9404 stat_direct_blk_ptrs++; 9405 /* 9406 * Reset the file size to its most up-to-date value. 9407 */ 9408 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 9409 panic("handle_written_inodeblock: bad size"); 9410 if (inodedep->id_savednlink > LINK_MAX) 9411 panic("handle_written_inodeblock: Invalid link count " 9412 "%d for inodedep %p", inodedep->id_savednlink, inodedep); 9413 if (fstype == UFS1) { 9414 if (dp1->di_nlink != inodedep->id_savednlink) { 9415 dp1->di_nlink = inodedep->id_savednlink; 9416 hadchanges = 1; 9417 } 9418 if (dp1->di_size != inodedep->id_savedsize) { 9419 dp1->di_size = inodedep->id_savedsize; 9420 hadchanges = 1; 9421 } 9422 } else { 9423 if (dp2->di_nlink != inodedep->id_savednlink) { 9424 dp2->di_nlink = inodedep->id_savednlink; 9425 hadchanges = 1; 9426 } 9427 if (dp2->di_size != inodedep->id_savedsize) { 9428 dp2->di_size = inodedep->id_savedsize; 9429 hadchanges = 1; 9430 } 9431 if (dp2->di_extsize != inodedep->id_savedextsize) { 9432 dp2->di_extsize = inodedep->id_savedextsize; 9433 hadchanges = 1; 9434 } 9435 } 9436 inodedep->id_savedsize = -1; 9437 inodedep->id_savedextsize = -1; 9438 inodedep->id_savednlink = -1; 9439 /* 9440 * If there were any rollbacks in the inode block, then it must be 9441 * marked dirty so that its will eventually get written back in 9442 * its correct form. 9443 */ 9444 if (hadchanges) 9445 bdirty(bp); 9446bufwait: 9447 /* 9448 * Process any allocdirects that completed during the update. 9449 */ 9450 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 9451 handle_allocdirect_partdone(adp, &wkhd); 9452 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 9453 handle_allocdirect_partdone(adp, &wkhd); 9454 /* 9455 * Process deallocations that were held pending until the 9456 * inode had been written to disk. Freeing of the inode 9457 * is delayed until after all blocks have been freed to 9458 * avoid creation of new <vfsid, inum, lbn> triples 9459 * before the old ones have been deleted. Completely 9460 * unlinked inodes are not processed until the unlinked 9461 * inode list is written or the last reference is removed. 9462 */ 9463 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 9464 freefile = handle_bufwait(inodedep, NULL); 9465 if (freefile && !LIST_EMPTY(&wkhd)) { 9466 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 9467 freefile = NULL; 9468 } 9469 } 9470 /* 9471 * Move rolled forward dependency completions to the bufwait list 9472 * now that those that were already written have been processed. 9473 */ 9474 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 9475 panic("handle_written_inodeblock: bufwait but no changes"); 9476 jwork_move(&inodedep->id_bufwait, &wkhd); 9477 9478 if (freefile != NULL) { 9479 /* 9480 * If the inode is goingaway it was never written. Fake up 9481 * the state here so free_inodedep() can succeed. 9482 */ 9483 if (inodedep->id_state & GOINGAWAY) 9484 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 9485 if (free_inodedep(inodedep) == 0) 9486 panic("handle_written_inodeblock: live inodedep %p", 9487 inodedep); 9488 add_to_worklist(&freefile->fx_list, 0); 9489 return (0); 9490 } 9491 9492 /* 9493 * If no outstanding dependencies, free it. 9494 */ 9495 if (free_inodedep(inodedep) || 9496 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 9497 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 9498 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 9499 LIST_FIRST(&inodedep->id_bufwait) == 0)) 9500 return (0); 9501 return (hadchanges); 9502} 9503 9504static int 9505handle_written_indirdep(indirdep, bp, bpp) 9506 struct indirdep *indirdep; 9507 struct buf *bp; 9508 struct buf **bpp; 9509{ 9510 struct allocindir *aip; 9511 int chgs; 9512 9513 if (indirdep->ir_state & GOINGAWAY) 9514 panic("disk_write_complete: indirdep gone"); 9515 chgs = 0; 9516 /* 9517 * If there were rollbacks revert them here. 9518 */ 9519 if (indirdep->ir_saveddata) { 9520 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 9521 free(indirdep->ir_saveddata, M_INDIRDEP); 9522 indirdep->ir_saveddata = 0; 9523 chgs = 1; 9524 } 9525 indirdep->ir_state &= ~UNDONE; 9526 indirdep->ir_state |= ATTACHED; 9527 /* 9528 * Move allocindirs with written pointers to the completehd if 9529 * the the indirdep's pointer is not yet written. Otherwise 9530 * free them here. 9531 */ 9532 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) { 9533 LIST_REMOVE(aip, ai_next); 9534 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 9535 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 9536 ai_next); 9537 continue; 9538 } 9539 free_newblk(&aip->ai_block); 9540 } 9541 /* 9542 * Move allocindirs that have finished dependency processing from 9543 * the done list to the write list after updating the pointers. 9544 */ 9545 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 9546 handle_allocindir_partdone(aip); 9547 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 9548 panic("disk_write_complete: not gone"); 9549 chgs = 1; 9550 } 9551 /* 9552 * If this indirdep has been detached from its newblk during 9553 * I/O we need to keep this dep attached to the buffer so 9554 * deallocate_dependencies can find it and properly resolve 9555 * any outstanding dependencies. 9556 */ 9557 if ((indirdep->ir_state & (ONDEPLIST | DEPCOMPLETE)) == 0) 9558 chgs = 1; 9559 if ((bp->b_flags & B_DELWRI) == 0) 9560 stat_indir_blk_ptrs++; 9561 /* 9562 * If there were no changes we can discard the savedbp and detach 9563 * ourselves from the buf. We are only carrying completed pointers 9564 * in this case. 9565 */ 9566 if (chgs == 0) { 9567 struct buf *sbp; 9568 9569 sbp = indirdep->ir_savebp; 9570 sbp->b_flags |= B_INVAL | B_NOCACHE; 9571 indirdep->ir_savebp = NULL; 9572 if (*bpp != NULL) 9573 panic("handle_written_indirdep: bp already exists."); 9574 *bpp = sbp; 9575 } else 9576 bdirty(bp); 9577 /* 9578 * If there are no fresh dependencies and none waiting on writes 9579 * we can free the indirdep. 9580 */ 9581 if ((indirdep->ir_state & DEPCOMPLETE) && chgs == 0) { 9582 if (indirdep->ir_state & ONDEPLIST) 9583 LIST_REMOVE(indirdep, ir_next); 9584 free_indirdep(indirdep); 9585 return (0); 9586 } 9587 9588 return (chgs); 9589} 9590 9591/* 9592 * Process a diradd entry after its dependent inode has been written. 9593 * This routine must be called with splbio interrupts blocked. 9594 */ 9595static void 9596diradd_inode_written(dap, inodedep) 9597 struct diradd *dap; 9598 struct inodedep *inodedep; 9599{ 9600 9601 dap->da_state |= COMPLETE; 9602 complete_diradd(dap); 9603 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9604} 9605 9606/* 9607 * Returns true if the bmsafemap will have rollbacks when written. Must 9608 * only be called with lk and the buf lock on the cg held. 9609 */ 9610static int 9611bmsafemap_rollbacks(bmsafemap) 9612 struct bmsafemap *bmsafemap; 9613{ 9614 9615 return (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 9616 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd)); 9617} 9618 9619/* 9620 * Complete a write to a bmsafemap structure. Roll forward any bitmap 9621 * changes if it's not a background write. Set all written dependencies 9622 * to DEPCOMPLETE and free the structure if possible. 9623 */ 9624static int 9625handle_written_bmsafemap(bmsafemap, bp) 9626 struct bmsafemap *bmsafemap; 9627 struct buf *bp; 9628{ 9629 struct newblk *newblk; 9630 struct inodedep *inodedep; 9631 struct jaddref *jaddref, *jatmp; 9632 struct jnewblk *jnewblk, *jntmp; 9633 uint8_t *inosused; 9634 uint8_t *blksfree; 9635 struct cg *cgp; 9636 struct fs *fs; 9637 ino_t ino; 9638 long bno; 9639 int chgs; 9640 int i; 9641 9642 if ((bmsafemap->sm_state & IOSTARTED) == 0) 9643 panic("initiate_write_bmsafemap: Not started\n"); 9644 chgs = 0; 9645 bmsafemap->sm_state &= ~IOSTARTED; 9646 /* 9647 * Restore unwritten inode allocation pending jaddref writes. 9648 */ 9649 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 9650 cgp = (struct cg *)bp->b_data; 9651 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 9652 inosused = cg_inosused(cgp); 9653 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 9654 ja_bmdeps, jatmp) { 9655 if ((jaddref->ja_state & UNDONE) == 0) 9656 continue; 9657 ino = jaddref->ja_ino % fs->fs_ipg; 9658 if (isset(inosused, ino)) 9659 panic("handle_written_bmsafemap: " 9660 "re-allocated inode"); 9661 if ((bp->b_xflags & BX_BKGRDMARKER) == 0) { 9662 if ((jaddref->ja_mode & IFMT) == IFDIR) 9663 cgp->cg_cs.cs_ndir++; 9664 cgp->cg_cs.cs_nifree--; 9665 setbit(inosused, ino); 9666 chgs = 1; 9667 } 9668 jaddref->ja_state &= ~UNDONE; 9669 jaddref->ja_state |= ATTACHED; 9670 free_jaddref(jaddref); 9671 } 9672 } 9673 /* 9674 * Restore any block allocations which are pending journal writes. 9675 */ 9676 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 9677 cgp = (struct cg *)bp->b_data; 9678 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 9679 blksfree = cg_blksfree(cgp); 9680 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 9681 jntmp) { 9682 if ((jnewblk->jn_state & UNDONE) == 0) 9683 continue; 9684 bno = dtogd(fs, jnewblk->jn_blkno); 9685 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 9686 i++) { 9687 if (bp->b_xflags & BX_BKGRDMARKER) 9688 break; 9689 if ((jnewblk->jn_state & NEWBLOCK) == 0 && 9690 isclr(blksfree, bno + i)) 9691 panic("handle_written_bmsafemap: " 9692 "re-allocated fragment"); 9693 clrbit(blksfree, bno + i); 9694 chgs = 1; 9695 } 9696 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 9697 jnewblk->jn_state |= ATTACHED; 9698 free_jnewblk(jnewblk); 9699 } 9700 } 9701 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 9702 newblk->nb_state |= DEPCOMPLETE; 9703 newblk->nb_state &= ~ONDEPLIST; 9704 newblk->nb_bmsafemap = NULL; 9705 LIST_REMOVE(newblk, nb_deps); 9706 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 9707 handle_allocdirect_partdone( 9708 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 9709 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 9710 handle_allocindir_partdone( 9711 WK_ALLOCINDIR(&newblk->nb_list)); 9712 else if (newblk->nb_list.wk_type != D_NEWBLK) 9713 panic("handle_written_bmsafemap: Unexpected type: %s", 9714 TYPENAME(newblk->nb_list.wk_type)); 9715 } 9716 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 9717 inodedep->id_state |= DEPCOMPLETE; 9718 inodedep->id_state &= ~ONDEPLIST; 9719 LIST_REMOVE(inodedep, id_deps); 9720 inodedep->id_bmsafemap = NULL; 9721 } 9722 if (LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 9723 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 9724 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 9725 LIST_EMPTY(&bmsafemap->sm_inodedephd)) { 9726 if (chgs) 9727 bdirty(bp); 9728 LIST_REMOVE(bmsafemap, sm_hash); 9729 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 9730 return (0); 9731 } 9732 bdirty(bp); 9733 return (1); 9734} 9735 9736/* 9737 * Try to free a mkdir dependency. 9738 */ 9739static void 9740complete_mkdir(mkdir) 9741 struct mkdir *mkdir; 9742{ 9743 struct diradd *dap; 9744 9745 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 9746 return; 9747 LIST_REMOVE(mkdir, md_mkdirs); 9748 dap = mkdir->md_diradd; 9749 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 9750 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 9751 dap->da_state |= DEPCOMPLETE; 9752 complete_diradd(dap); 9753 } 9754 WORKITEM_FREE(mkdir, D_MKDIR); 9755} 9756 9757/* 9758 * Handle the completion of a mkdir dependency. 9759 */ 9760static void 9761handle_written_mkdir(mkdir, type) 9762 struct mkdir *mkdir; 9763 int type; 9764{ 9765 9766 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 9767 panic("handle_written_mkdir: bad type"); 9768 mkdir->md_state |= COMPLETE; 9769 complete_mkdir(mkdir); 9770} 9771 9772static void 9773free_pagedep(pagedep) 9774 struct pagedep *pagedep; 9775{ 9776 int i; 9777 9778 if (pagedep->pd_state & (NEWBLOCK | ONWORKLIST)) 9779 return; 9780 for (i = 0; i < DAHASHSZ; i++) 9781 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 9782 return; 9783 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 9784 return; 9785 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 9786 return; 9787 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 9788 return; 9789 LIST_REMOVE(pagedep, pd_hash); 9790 WORKITEM_FREE(pagedep, D_PAGEDEP); 9791} 9792 9793/* 9794 * Called from within softdep_disk_write_complete above. 9795 * A write operation was just completed. Removed inodes can 9796 * now be freed and associated block pointers may be committed. 9797 * Note that this routine is always called from interrupt level 9798 * with further splbio interrupts blocked. 9799 */ 9800static int 9801handle_written_filepage(pagedep, bp) 9802 struct pagedep *pagedep; 9803 struct buf *bp; /* buffer containing the written page */ 9804{ 9805 struct dirrem *dirrem; 9806 struct diradd *dap, *nextdap; 9807 struct direct *ep; 9808 int i, chgs; 9809 9810 if ((pagedep->pd_state & IOSTARTED) == 0) 9811 panic("handle_written_filepage: not started"); 9812 pagedep->pd_state &= ~IOSTARTED; 9813 /* 9814 * Process any directory removals that have been committed. 9815 */ 9816 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 9817 LIST_REMOVE(dirrem, dm_next); 9818 dirrem->dm_state |= COMPLETE; 9819 dirrem->dm_dirinum = pagedep->pd_ino; 9820 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9821 ("handle_written_filepage: Journal entries not written.")); 9822 add_to_worklist(&dirrem->dm_list, 0); 9823 } 9824 /* 9825 * Free any directory additions that have been committed. 9826 * If it is a newly allocated block, we have to wait until 9827 * the on-disk directory inode claims the new block. 9828 */ 9829 if ((pagedep->pd_state & NEWBLOCK) == 0) 9830 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 9831 free_diradd(dap, NULL); 9832 /* 9833 * Uncommitted directory entries must be restored. 9834 */ 9835 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 9836 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 9837 dap = nextdap) { 9838 nextdap = LIST_NEXT(dap, da_pdlist); 9839 if (dap->da_state & ATTACHED) 9840 panic("handle_written_filepage: attached"); 9841 ep = (struct direct *) 9842 ((char *)bp->b_data + dap->da_offset); 9843 ep->d_ino = dap->da_newinum; 9844 dap->da_state &= ~UNDONE; 9845 dap->da_state |= ATTACHED; 9846 chgs = 1; 9847 /* 9848 * If the inode referenced by the directory has 9849 * been written out, then the dependency can be 9850 * moved to the pending list. 9851 */ 9852 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 9853 LIST_REMOVE(dap, da_pdlist); 9854 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 9855 da_pdlist); 9856 } 9857 } 9858 } 9859 /* 9860 * If there were any rollbacks in the directory, then it must be 9861 * marked dirty so that its will eventually get written back in 9862 * its correct form. 9863 */ 9864 if (chgs) { 9865 if ((bp->b_flags & B_DELWRI) == 0) 9866 stat_dir_entry++; 9867 bdirty(bp); 9868 return (1); 9869 } 9870 /* 9871 * If we are not waiting for a new directory block to be 9872 * claimed by its inode, then the pagedep will be freed. 9873 * Otherwise it will remain to track any new entries on 9874 * the page in case they are fsync'ed. 9875 */ 9876 if ((pagedep->pd_state & NEWBLOCK) == 0 && 9877 LIST_EMPTY(&pagedep->pd_jmvrefhd)) { 9878 LIST_REMOVE(pagedep, pd_hash); 9879 WORKITEM_FREE(pagedep, D_PAGEDEP); 9880 } 9881 return (0); 9882} 9883 9884/* 9885 * Writing back in-core inode structures. 9886 * 9887 * The filesystem only accesses an inode's contents when it occupies an 9888 * "in-core" inode structure. These "in-core" structures are separate from 9889 * the page frames used to cache inode blocks. Only the latter are 9890 * transferred to/from the disk. So, when the updated contents of the 9891 * "in-core" inode structure are copied to the corresponding in-memory inode 9892 * block, the dependencies are also transferred. The following procedure is 9893 * called when copying a dirty "in-core" inode to a cached inode block. 9894 */ 9895 9896/* 9897 * Called when an inode is loaded from disk. If the effective link count 9898 * differed from the actual link count when it was last flushed, then we 9899 * need to ensure that the correct effective link count is put back. 9900 */ 9901void 9902softdep_load_inodeblock(ip) 9903 struct inode *ip; /* the "in_core" copy of the inode */ 9904{ 9905 struct inodedep *inodedep; 9906 9907 /* 9908 * Check for alternate nlink count. 9909 */ 9910 ip->i_effnlink = ip->i_nlink; 9911 ACQUIRE_LOCK(&lk); 9912 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 9913 &inodedep) == 0) { 9914 FREE_LOCK(&lk); 9915 return; 9916 } 9917 ip->i_effnlink -= inodedep->id_nlinkdelta; 9918 if (inodedep->id_state & SPACECOUNTED) 9919 ip->i_flag |= IN_SPACECOUNTED; 9920 FREE_LOCK(&lk); 9921} 9922 9923/* 9924 * This routine is called just before the "in-core" inode 9925 * information is to be copied to the in-memory inode block. 9926 * Recall that an inode block contains several inodes. If 9927 * the force flag is set, then the dependencies will be 9928 * cleared so that the update can always be made. Note that 9929 * the buffer is locked when this routine is called, so we 9930 * will never be in the middle of writing the inode block 9931 * to disk. 9932 */ 9933void 9934softdep_update_inodeblock(ip, bp, waitfor) 9935 struct inode *ip; /* the "in_core" copy of the inode */ 9936 struct buf *bp; /* the buffer containing the inode block */ 9937 int waitfor; /* nonzero => update must be allowed */ 9938{ 9939 struct inodedep *inodedep; 9940 struct inoref *inoref; 9941 struct worklist *wk; 9942 struct mount *mp; 9943 struct buf *ibp; 9944 struct fs *fs; 9945 int error; 9946 9947 mp = UFSTOVFS(ip->i_ump); 9948 fs = ip->i_fs; 9949 /* 9950 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 9951 * does not have access to the in-core ip so must write directly into 9952 * the inode block buffer when setting freelink. 9953 */ 9954 if (fs->fs_magic == FS_UFS1_MAGIC) 9955 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 9956 ino_to_fsbo(fs, ip->i_number))->di_freelink); 9957 else 9958 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 9959 ino_to_fsbo(fs, ip->i_number))->di_freelink); 9960 /* 9961 * If the effective link count is not equal to the actual link 9962 * count, then we must track the difference in an inodedep while 9963 * the inode is (potentially) tossed out of the cache. Otherwise, 9964 * if there is no existing inodedep, then there are no dependencies 9965 * to track. 9966 */ 9967 ACQUIRE_LOCK(&lk); 9968again: 9969 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 9970 FREE_LOCK(&lk); 9971 if (ip->i_effnlink != ip->i_nlink) 9972 panic("softdep_update_inodeblock: bad link count"); 9973 return; 9974 } 9975 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 9976 panic("softdep_update_inodeblock: bad delta"); 9977 /* 9978 * If we're flushing all dependencies we must also move any waiting 9979 * for journal writes onto the bufwait list prior to I/O. 9980 */ 9981 if (waitfor) { 9982 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 9983 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 9984 == DEPCOMPLETE) { 9985 stat_jwait_inode++; 9986 jwait(&inoref->if_list); 9987 goto again; 9988 } 9989 } 9990 } 9991 /* 9992 * Changes have been initiated. Anything depending on these 9993 * changes cannot occur until this inode has been written. 9994 */ 9995 inodedep->id_state &= ~COMPLETE; 9996 if ((inodedep->id_state & ONWORKLIST) == 0) 9997 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 9998 /* 9999 * Any new dependencies associated with the incore inode must 10000 * now be moved to the list associated with the buffer holding 10001 * the in-memory copy of the inode. Once merged process any 10002 * allocdirects that are completed by the merger. 10003 */ 10004 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 10005 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 10006 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 10007 NULL); 10008 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 10009 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 10010 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 10011 NULL); 10012 /* 10013 * Now that the inode has been pushed into the buffer, the 10014 * operations dependent on the inode being written to disk 10015 * can be moved to the id_bufwait so that they will be 10016 * processed when the buffer I/O completes. 10017 */ 10018 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 10019 WORKLIST_REMOVE(wk); 10020 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 10021 } 10022 /* 10023 * Newly allocated inodes cannot be written until the bitmap 10024 * that allocates them have been written (indicated by 10025 * DEPCOMPLETE being set in id_state). If we are doing a 10026 * forced sync (e.g., an fsync on a file), we force the bitmap 10027 * to be written so that the update can be done. 10028 */ 10029 if (waitfor == 0) { 10030 FREE_LOCK(&lk); 10031 return; 10032 } 10033retry: 10034 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 10035 FREE_LOCK(&lk); 10036 return; 10037 } 10038 ibp = inodedep->id_bmsafemap->sm_buf; 10039 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 10040 if (ibp == NULL) { 10041 /* 10042 * If ibp came back as NULL, the dependency could have been 10043 * freed while we slept. Look it up again, and check to see 10044 * that it has completed. 10045 */ 10046 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 10047 goto retry; 10048 FREE_LOCK(&lk); 10049 return; 10050 } 10051 FREE_LOCK(&lk); 10052 if ((error = bwrite(ibp)) != 0) 10053 softdep_error("softdep_update_inodeblock: bwrite", error); 10054} 10055 10056/* 10057 * Merge the a new inode dependency list (such as id_newinoupdt) into an 10058 * old inode dependency list (such as id_inoupdt). This routine must be 10059 * called with splbio interrupts blocked. 10060 */ 10061static void 10062merge_inode_lists(newlisthead, oldlisthead) 10063 struct allocdirectlst *newlisthead; 10064 struct allocdirectlst *oldlisthead; 10065{ 10066 struct allocdirect *listadp, *newadp; 10067 10068 newadp = TAILQ_FIRST(newlisthead); 10069 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 10070 if (listadp->ad_offset < newadp->ad_offset) { 10071 listadp = TAILQ_NEXT(listadp, ad_next); 10072 continue; 10073 } 10074 TAILQ_REMOVE(newlisthead, newadp, ad_next); 10075 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 10076 if (listadp->ad_offset == newadp->ad_offset) { 10077 allocdirect_merge(oldlisthead, newadp, 10078 listadp); 10079 listadp = newadp; 10080 } 10081 newadp = TAILQ_FIRST(newlisthead); 10082 } 10083 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 10084 TAILQ_REMOVE(newlisthead, newadp, ad_next); 10085 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 10086 } 10087} 10088 10089/* 10090 * If we are doing an fsync, then we must ensure that any directory 10091 * entries for the inode have been written after the inode gets to disk. 10092 */ 10093int 10094softdep_fsync(vp) 10095 struct vnode *vp; /* the "in_core" copy of the inode */ 10096{ 10097 struct inodedep *inodedep; 10098 struct pagedep *pagedep; 10099 struct inoref *inoref; 10100 struct worklist *wk; 10101 struct diradd *dap; 10102 struct mount *mp; 10103 struct vnode *pvp; 10104 struct inode *ip; 10105 struct buf *bp; 10106 struct fs *fs; 10107 struct thread *td = curthread; 10108 int error, flushparent, pagedep_new_block; 10109 ino_t parentino; 10110 ufs_lbn_t lbn; 10111 10112 ip = VTOI(vp); 10113 fs = ip->i_fs; 10114 mp = vp->v_mount; 10115 ACQUIRE_LOCK(&lk); 10116restart: 10117 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 10118 FREE_LOCK(&lk); 10119 return (0); 10120 } 10121 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 10122 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 10123 == DEPCOMPLETE) { 10124 stat_jwait_inode++; 10125 jwait(&inoref->if_list); 10126 goto restart; 10127 } 10128 } 10129 if (!LIST_EMPTY(&inodedep->id_inowait) || 10130 !TAILQ_EMPTY(&inodedep->id_extupdt) || 10131 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 10132 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 10133 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 10134 panic("softdep_fsync: pending ops %p", inodedep); 10135 for (error = 0, flushparent = 0; ; ) { 10136 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 10137 break; 10138 if (wk->wk_type != D_DIRADD) 10139 panic("softdep_fsync: Unexpected type %s", 10140 TYPENAME(wk->wk_type)); 10141 dap = WK_DIRADD(wk); 10142 /* 10143 * Flush our parent if this directory entry has a MKDIR_PARENT 10144 * dependency or is contained in a newly allocated block. 10145 */ 10146 if (dap->da_state & DIRCHG) 10147 pagedep = dap->da_previous->dm_pagedep; 10148 else 10149 pagedep = dap->da_pagedep; 10150 parentino = pagedep->pd_ino; 10151 lbn = pagedep->pd_lbn; 10152 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 10153 panic("softdep_fsync: dirty"); 10154 if ((dap->da_state & MKDIR_PARENT) || 10155 (pagedep->pd_state & NEWBLOCK)) 10156 flushparent = 1; 10157 else 10158 flushparent = 0; 10159 /* 10160 * If we are being fsync'ed as part of vgone'ing this vnode, 10161 * then we will not be able to release and recover the 10162 * vnode below, so we just have to give up on writing its 10163 * directory entry out. It will eventually be written, just 10164 * not now, but then the user was not asking to have it 10165 * written, so we are not breaking any promises. 10166 */ 10167 if (vp->v_iflag & VI_DOOMED) 10168 break; 10169 /* 10170 * We prevent deadlock by always fetching inodes from the 10171 * root, moving down the directory tree. Thus, when fetching 10172 * our parent directory, we first try to get the lock. If 10173 * that fails, we must unlock ourselves before requesting 10174 * the lock on our parent. See the comment in ufs_lookup 10175 * for details on possible races. 10176 */ 10177 FREE_LOCK(&lk); 10178 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 10179 FFSV_FORCEINSMQ)) { 10180 error = vfs_busy(mp, MBF_NOWAIT); 10181 if (error != 0) { 10182 vfs_ref(mp); 10183 VOP_UNLOCK(vp, 0); 10184 error = vfs_busy(mp, 0); 10185 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 10186 vfs_rel(mp); 10187 if (error != 0) 10188 return (ENOENT); 10189 if (vp->v_iflag & VI_DOOMED) { 10190 vfs_unbusy(mp); 10191 return (ENOENT); 10192 } 10193 } 10194 VOP_UNLOCK(vp, 0); 10195 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 10196 &pvp, FFSV_FORCEINSMQ); 10197 vfs_unbusy(mp); 10198 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 10199 if (vp->v_iflag & VI_DOOMED) { 10200 if (error == 0) 10201 vput(pvp); 10202 error = ENOENT; 10203 } 10204 if (error != 0) 10205 return (error); 10206 } 10207 /* 10208 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 10209 * that are contained in direct blocks will be resolved by 10210 * doing a ffs_update. Pagedeps contained in indirect blocks 10211 * may require a complete sync'ing of the directory. So, we 10212 * try the cheap and fast ffs_update first, and if that fails, 10213 * then we do the slower ffs_syncvnode of the directory. 10214 */ 10215 if (flushparent) { 10216 int locked; 10217 10218 if ((error = ffs_update(pvp, 1)) != 0) { 10219 vput(pvp); 10220 return (error); 10221 } 10222 ACQUIRE_LOCK(&lk); 10223 locked = 1; 10224 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 10225 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 10226 if (wk->wk_type != D_DIRADD) 10227 panic("softdep_fsync: Unexpected type %s", 10228 TYPENAME(wk->wk_type)); 10229 dap = WK_DIRADD(wk); 10230 if (dap->da_state & DIRCHG) 10231 pagedep = dap->da_previous->dm_pagedep; 10232 else 10233 pagedep = dap->da_pagedep; 10234 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 10235 FREE_LOCK(&lk); 10236 locked = 0; 10237 if (pagedep_new_block && 10238 (error = ffs_syncvnode(pvp, MNT_WAIT))) { 10239 vput(pvp); 10240 return (error); 10241 } 10242 } 10243 } 10244 if (locked) 10245 FREE_LOCK(&lk); 10246 } 10247 /* 10248 * Flush directory page containing the inode's name. 10249 */ 10250 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 10251 &bp); 10252 if (error == 0) 10253 error = bwrite(bp); 10254 else 10255 brelse(bp); 10256 vput(pvp); 10257 if (error != 0) 10258 return (error); 10259 ACQUIRE_LOCK(&lk); 10260 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 10261 break; 10262 } 10263 FREE_LOCK(&lk); 10264 return (0); 10265} 10266 10267/* 10268 * Flush all the dirty bitmaps associated with the block device 10269 * before flushing the rest of the dirty blocks so as to reduce 10270 * the number of dependencies that will have to be rolled back. 10271 */ 10272void 10273softdep_fsync_mountdev(vp) 10274 struct vnode *vp; 10275{ 10276 struct buf *bp, *nbp; 10277 struct worklist *wk; 10278 struct bufobj *bo; 10279 10280 if (!vn_isdisk(vp, NULL)) 10281 panic("softdep_fsync_mountdev: vnode not a disk"); 10282 bo = &vp->v_bufobj; 10283restart: 10284 BO_LOCK(bo); 10285 ACQUIRE_LOCK(&lk); 10286 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 10287 /* 10288 * If it is already scheduled, skip to the next buffer. 10289 */ 10290 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 10291 continue; 10292 10293 if ((bp->b_flags & B_DELWRI) == 0) 10294 panic("softdep_fsync_mountdev: not dirty"); 10295 /* 10296 * We are only interested in bitmaps with outstanding 10297 * dependencies. 10298 */ 10299 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 10300 wk->wk_type != D_BMSAFEMAP || 10301 (bp->b_vflags & BV_BKGRDINPROG)) { 10302 BUF_UNLOCK(bp); 10303 continue; 10304 } 10305 FREE_LOCK(&lk); 10306 BO_UNLOCK(bo); 10307 bremfree(bp); 10308 (void) bawrite(bp); 10309 goto restart; 10310 } 10311 FREE_LOCK(&lk); 10312 drain_output(vp); 10313 BO_UNLOCK(bo); 10314} 10315 10316/* 10317 * This routine is called when we are trying to synchronously flush a 10318 * file. This routine must eliminate any filesystem metadata dependencies 10319 * so that the syncing routine can succeed by pushing the dirty blocks 10320 * associated with the file. If any I/O errors occur, they are returned. 10321 */ 10322int 10323softdep_sync_metadata(struct vnode *vp) 10324{ 10325 struct pagedep *pagedep; 10326 struct allocindir *aip; 10327 struct newblk *newblk; 10328 struct buf *bp, *nbp; 10329 struct worklist *wk; 10330 struct bufobj *bo; 10331 int i, error, waitfor; 10332 10333 if (!DOINGSOFTDEP(vp)) 10334 return (0); 10335 /* 10336 * Ensure that any direct block dependencies have been cleared. 10337 */ 10338 ACQUIRE_LOCK(&lk); 10339 if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) { 10340 FREE_LOCK(&lk); 10341 return (error); 10342 } 10343 FREE_LOCK(&lk); 10344 /* 10345 * For most files, the only metadata dependencies are the 10346 * cylinder group maps that allocate their inode or blocks. 10347 * The block allocation dependencies can be found by traversing 10348 * the dependency lists for any buffers that remain on their 10349 * dirty buffer list. The inode allocation dependency will 10350 * be resolved when the inode is updated with MNT_WAIT. 10351 * This work is done in two passes. The first pass grabs most 10352 * of the buffers and begins asynchronously writing them. The 10353 * only way to wait for these asynchronous writes is to sleep 10354 * on the filesystem vnode which may stay busy for a long time 10355 * if the filesystem is active. So, instead, we make a second 10356 * pass over the dependencies blocking on each write. In the 10357 * usual case we will be blocking against a write that we 10358 * initiated, so when it is done the dependency will have been 10359 * resolved. Thus the second pass is expected to end quickly. 10360 */ 10361 waitfor = MNT_NOWAIT; 10362 bo = &vp->v_bufobj; 10363 10364top: 10365 /* 10366 * We must wait for any I/O in progress to finish so that 10367 * all potential buffers on the dirty list will be visible. 10368 */ 10369 BO_LOCK(bo); 10370 drain_output(vp); 10371 while ((bp = TAILQ_FIRST(&bo->bo_dirty.bv_hd)) != NULL) { 10372 bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT); 10373 if (bp) 10374 break; 10375 } 10376 BO_UNLOCK(bo); 10377 if (bp == NULL) 10378 return (0); 10379loop: 10380 /* While syncing snapshots, we must allow recursive lookups */ 10381 BUF_AREC(bp); 10382 ACQUIRE_LOCK(&lk); 10383 /* 10384 * As we hold the buffer locked, none of its dependencies 10385 * will disappear. 10386 */ 10387 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 10388 switch (wk->wk_type) { 10389 10390 case D_ALLOCDIRECT: 10391 case D_ALLOCINDIR: 10392 newblk = WK_NEWBLK(wk); 10393 if (newblk->nb_jnewblk != NULL) { 10394 stat_jwait_newblk++; 10395 jwait(&newblk->nb_jnewblk->jn_list); 10396 goto restart; 10397 } 10398 if (newblk->nb_state & DEPCOMPLETE) 10399 continue; 10400 nbp = newblk->nb_bmsafemap->sm_buf; 10401 nbp = getdirtybuf(nbp, &lk, waitfor); 10402 if (nbp == NULL) 10403 continue; 10404 FREE_LOCK(&lk); 10405 if (waitfor == MNT_NOWAIT) { 10406 bawrite(nbp); 10407 } else if ((error = bwrite(nbp)) != 0) { 10408 break; 10409 } 10410 ACQUIRE_LOCK(&lk); 10411 continue; 10412 10413 case D_INDIRDEP: 10414 restart: 10415 10416 LIST_FOREACH(aip, 10417 &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) { 10418 newblk = (struct newblk *)aip; 10419 if (newblk->nb_jnewblk != NULL) { 10420 stat_jwait_newblk++; 10421 jwait(&newblk->nb_jnewblk->jn_list); 10422 goto restart; 10423 } 10424 if (newblk->nb_state & DEPCOMPLETE) 10425 continue; 10426 nbp = newblk->nb_bmsafemap->sm_buf; 10427 nbp = getdirtybuf(nbp, &lk, MNT_WAIT); 10428 if (nbp == NULL) 10429 goto restart; 10430 FREE_LOCK(&lk); 10431 if ((error = bwrite(nbp)) != 0) { 10432 goto loop_end; 10433 } 10434 ACQUIRE_LOCK(&lk); 10435 goto restart; 10436 } 10437 continue; 10438 10439 case D_PAGEDEP: 10440 /* 10441 * We are trying to sync a directory that may 10442 * have dependencies on both its own metadata 10443 * and/or dependencies on the inodes of any 10444 * recently allocated files. We walk its diradd 10445 * lists pushing out the associated inode. 10446 */ 10447 pagedep = WK_PAGEDEP(wk); 10448 for (i = 0; i < DAHASHSZ; i++) { 10449 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 10450 continue; 10451 if ((error = 10452 flush_pagedep_deps(vp, wk->wk_mp, 10453 &pagedep->pd_diraddhd[i]))) { 10454 FREE_LOCK(&lk); 10455 goto loop_end; 10456 } 10457 } 10458 continue; 10459 10460 default: 10461 panic("softdep_sync_metadata: Unknown type %s", 10462 TYPENAME(wk->wk_type)); 10463 /* NOTREACHED */ 10464 } 10465 loop_end: 10466 /* We reach here only in error and unlocked */ 10467 if (error == 0) 10468 panic("softdep_sync_metadata: zero error"); 10469 BUF_NOREC(bp); 10470 bawrite(bp); 10471 return (error); 10472 } 10473 FREE_LOCK(&lk); 10474 BO_LOCK(bo); 10475 while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) { 10476 nbp = getdirtybuf(nbp, BO_MTX(bo), MNT_WAIT); 10477 if (nbp) 10478 break; 10479 } 10480 BO_UNLOCK(bo); 10481 BUF_NOREC(bp); 10482 bawrite(bp); 10483 if (nbp != NULL) { 10484 bp = nbp; 10485 goto loop; 10486 } 10487 /* 10488 * The brief unlock is to allow any pent up dependency 10489 * processing to be done. Then proceed with the second pass. 10490 */ 10491 if (waitfor == MNT_NOWAIT) { 10492 waitfor = MNT_WAIT; 10493 goto top; 10494 } 10495 10496 /* 10497 * If we have managed to get rid of all the dirty buffers, 10498 * then we are done. For certain directories and block 10499 * devices, we may need to do further work. 10500 * 10501 * We must wait for any I/O in progress to finish so that 10502 * all potential buffers on the dirty list will be visible. 10503 */ 10504 BO_LOCK(bo); 10505 drain_output(vp); 10506 BO_UNLOCK(bo); 10507 return ffs_update(vp, 1); 10508 /* return (0); */ 10509} 10510 10511/* 10512 * Flush the dependencies associated with an inodedep. 10513 * Called with splbio blocked. 10514 */ 10515static int 10516flush_inodedep_deps(mp, ino) 10517 struct mount *mp; 10518 ino_t ino; 10519{ 10520 struct inodedep *inodedep; 10521 struct inoref *inoref; 10522 int error, waitfor; 10523 10524 /* 10525 * This work is done in two passes. The first pass grabs most 10526 * of the buffers and begins asynchronously writing them. The 10527 * only way to wait for these asynchronous writes is to sleep 10528 * on the filesystem vnode which may stay busy for a long time 10529 * if the filesystem is active. So, instead, we make a second 10530 * pass over the dependencies blocking on each write. In the 10531 * usual case we will be blocking against a write that we 10532 * initiated, so when it is done the dependency will have been 10533 * resolved. Thus the second pass is expected to end quickly. 10534 * We give a brief window at the top of the loop to allow 10535 * any pending I/O to complete. 10536 */ 10537 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 10538 if (error) 10539 return (error); 10540 FREE_LOCK(&lk); 10541 ACQUIRE_LOCK(&lk); 10542restart: 10543 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 10544 return (0); 10545 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 10546 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 10547 == DEPCOMPLETE) { 10548 stat_jwait_inode++; 10549 jwait(&inoref->if_list); 10550 goto restart; 10551 } 10552 } 10553 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 10554 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 10555 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 10556 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 10557 continue; 10558 /* 10559 * If pass2, we are done, otherwise do pass 2. 10560 */ 10561 if (waitfor == MNT_WAIT) 10562 break; 10563 waitfor = MNT_WAIT; 10564 } 10565 /* 10566 * Try freeing inodedep in case all dependencies have been removed. 10567 */ 10568 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 10569 (void) free_inodedep(inodedep); 10570 return (0); 10571} 10572 10573/* 10574 * Flush an inode dependency list. 10575 * Called with splbio blocked. 10576 */ 10577static int 10578flush_deplist(listhead, waitfor, errorp) 10579 struct allocdirectlst *listhead; 10580 int waitfor; 10581 int *errorp; 10582{ 10583 struct allocdirect *adp; 10584 struct newblk *newblk; 10585 struct buf *bp; 10586 10587 mtx_assert(&lk, MA_OWNED); 10588 TAILQ_FOREACH(adp, listhead, ad_next) { 10589 newblk = (struct newblk *)adp; 10590 if (newblk->nb_jnewblk != NULL) { 10591 stat_jwait_newblk++; 10592 jwait(&newblk->nb_jnewblk->jn_list); 10593 return (1); 10594 } 10595 if (newblk->nb_state & DEPCOMPLETE) 10596 continue; 10597 bp = newblk->nb_bmsafemap->sm_buf; 10598 bp = getdirtybuf(bp, &lk, waitfor); 10599 if (bp == NULL) { 10600 if (waitfor == MNT_NOWAIT) 10601 continue; 10602 return (1); 10603 } 10604 FREE_LOCK(&lk); 10605 if (waitfor == MNT_NOWAIT) { 10606 bawrite(bp); 10607 } else if ((*errorp = bwrite(bp)) != 0) { 10608 ACQUIRE_LOCK(&lk); 10609 return (1); 10610 } 10611 ACQUIRE_LOCK(&lk); 10612 return (1); 10613 } 10614 return (0); 10615} 10616 10617/* 10618 * Flush dependencies associated with an allocdirect block. 10619 */ 10620static int 10621flush_newblk_dep(vp, mp, lbn) 10622 struct vnode *vp; 10623 struct mount *mp; 10624 ufs_lbn_t lbn; 10625{ 10626 struct newblk *newblk; 10627 struct bufobj *bo; 10628 struct inode *ip; 10629 struct buf *bp; 10630 ufs2_daddr_t blkno; 10631 int error; 10632 10633 error = 0; 10634 bo = &vp->v_bufobj; 10635 ip = VTOI(vp); 10636 blkno = DIP(ip, i_db[lbn]); 10637 if (blkno == 0) 10638 panic("flush_newblk_dep: Missing block"); 10639 ACQUIRE_LOCK(&lk); 10640 /* 10641 * Loop until all dependencies related to this block are satisfied. 10642 * We must be careful to restart after each sleep in case a write 10643 * completes some part of this process for us. 10644 */ 10645 for (;;) { 10646 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 10647 FREE_LOCK(&lk); 10648 break; 10649 } 10650 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 10651 panic("flush_newblk_deps: Bad newblk %p", newblk); 10652 /* 10653 * Flush the journal. 10654 */ 10655 if (newblk->nb_jnewblk != NULL) { 10656 stat_jwait_newblk++; 10657 jwait(&newblk->nb_jnewblk->jn_list); 10658 continue; 10659 } 10660 /* 10661 * Write the bitmap dependency. 10662 */ 10663 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 10664 bp = newblk->nb_bmsafemap->sm_buf; 10665 bp = getdirtybuf(bp, &lk, MNT_WAIT); 10666 if (bp == NULL) 10667 continue; 10668 FREE_LOCK(&lk); 10669 error = bwrite(bp); 10670 if (error) 10671 break; 10672 ACQUIRE_LOCK(&lk); 10673 continue; 10674 } 10675 /* 10676 * Write the buffer. 10677 */ 10678 FREE_LOCK(&lk); 10679 BO_LOCK(bo); 10680 bp = gbincore(bo, lbn); 10681 if (bp != NULL) { 10682 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 10683 LK_INTERLOCK, BO_MTX(bo)); 10684 if (error == ENOLCK) { 10685 ACQUIRE_LOCK(&lk); 10686 continue; /* Slept, retry */ 10687 } 10688 if (error != 0) 10689 break; /* Failed */ 10690 if (bp->b_flags & B_DELWRI) { 10691 bremfree(bp); 10692 error = bwrite(bp); 10693 if (error) 10694 break; 10695 } else 10696 BUF_UNLOCK(bp); 10697 } else 10698 BO_UNLOCK(bo); 10699 /* 10700 * We have to wait for the direct pointers to 10701 * point at the newdirblk before the dependency 10702 * will go away. 10703 */ 10704 error = ffs_update(vp, MNT_WAIT); 10705 if (error) 10706 break; 10707 ACQUIRE_LOCK(&lk); 10708 } 10709 return (error); 10710} 10711 10712/* 10713 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 10714 * Called with splbio blocked. 10715 */ 10716static int 10717flush_pagedep_deps(pvp, mp, diraddhdp) 10718 struct vnode *pvp; 10719 struct mount *mp; 10720 struct diraddhd *diraddhdp; 10721{ 10722 struct inodedep *inodedep; 10723 struct inoref *inoref; 10724 struct ufsmount *ump; 10725 struct diradd *dap; 10726 struct vnode *vp; 10727 int error = 0; 10728 struct buf *bp; 10729 ino_t inum; 10730 10731 ump = VFSTOUFS(mp); 10732restart: 10733 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 10734 /* 10735 * Flush ourselves if this directory entry 10736 * has a MKDIR_PARENT dependency. 10737 */ 10738 if (dap->da_state & MKDIR_PARENT) { 10739 FREE_LOCK(&lk); 10740 if ((error = ffs_update(pvp, MNT_WAIT)) != 0) 10741 break; 10742 ACQUIRE_LOCK(&lk); 10743 /* 10744 * If that cleared dependencies, go on to next. 10745 */ 10746 if (dap != LIST_FIRST(diraddhdp)) 10747 continue; 10748 if (dap->da_state & MKDIR_PARENT) 10749 panic("flush_pagedep_deps: MKDIR_PARENT"); 10750 } 10751 /* 10752 * A newly allocated directory must have its "." and 10753 * ".." entries written out before its name can be 10754 * committed in its parent. 10755 */ 10756 inum = dap->da_newinum; 10757 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 10758 panic("flush_pagedep_deps: lost inode1"); 10759 /* 10760 * Wait for any pending journal adds to complete so we don't 10761 * cause rollbacks while syncing. 10762 */ 10763 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 10764 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 10765 == DEPCOMPLETE) { 10766 stat_jwait_inode++; 10767 jwait(&inoref->if_list); 10768 goto restart; 10769 } 10770 } 10771 if (dap->da_state & MKDIR_BODY) { 10772 FREE_LOCK(&lk); 10773 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 10774 FFSV_FORCEINSMQ))) 10775 break; 10776 error = flush_newblk_dep(vp, mp, 0); 10777 /* 10778 * If we still have the dependency we might need to 10779 * update the vnode to sync the new link count to 10780 * disk. 10781 */ 10782 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 10783 error = ffs_update(vp, MNT_WAIT); 10784 vput(vp); 10785 if (error != 0) 10786 break; 10787 ACQUIRE_LOCK(&lk); 10788 /* 10789 * If that cleared dependencies, go on to next. 10790 */ 10791 if (dap != LIST_FIRST(diraddhdp)) 10792 continue; 10793 if (dap->da_state & MKDIR_BODY) { 10794 inodedep_lookup(UFSTOVFS(ump), inum, 0, 10795 &inodedep); 10796 panic("flush_pagedep_deps: MKDIR_BODY " 10797 "inodedep %p dap %p vp %p", 10798 inodedep, dap, vp); 10799 } 10800 } 10801 /* 10802 * Flush the inode on which the directory entry depends. 10803 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 10804 * the only remaining dependency is that the updated inode 10805 * count must get pushed to disk. The inode has already 10806 * been pushed into its inode buffer (via VOP_UPDATE) at 10807 * the time of the reference count change. So we need only 10808 * locate that buffer, ensure that there will be no rollback 10809 * caused by a bitmap dependency, then write the inode buffer. 10810 */ 10811retry: 10812 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 10813 panic("flush_pagedep_deps: lost inode"); 10814 /* 10815 * If the inode still has bitmap dependencies, 10816 * push them to disk. 10817 */ 10818 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 10819 bp = inodedep->id_bmsafemap->sm_buf; 10820 bp = getdirtybuf(bp, &lk, MNT_WAIT); 10821 if (bp == NULL) 10822 goto retry; 10823 FREE_LOCK(&lk); 10824 if ((error = bwrite(bp)) != 0) 10825 break; 10826 ACQUIRE_LOCK(&lk); 10827 if (dap != LIST_FIRST(diraddhdp)) 10828 continue; 10829 } 10830 /* 10831 * If the inode is still sitting in a buffer waiting 10832 * to be written or waiting for the link count to be 10833 * adjusted update it here to flush it to disk. 10834 */ 10835 if (dap == LIST_FIRST(diraddhdp)) { 10836 FREE_LOCK(&lk); 10837 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 10838 FFSV_FORCEINSMQ))) 10839 break; 10840 error = ffs_update(vp, MNT_WAIT); 10841 vput(vp); 10842 if (error) 10843 break; 10844 ACQUIRE_LOCK(&lk); 10845 } 10846 /* 10847 * If we have failed to get rid of all the dependencies 10848 * then something is seriously wrong. 10849 */ 10850 if (dap == LIST_FIRST(diraddhdp)) { 10851 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 10852 panic("flush_pagedep_deps: failed to flush " 10853 "inodedep %p ino %d dap %p", inodedep, inum, dap); 10854 } 10855 } 10856 if (error) 10857 ACQUIRE_LOCK(&lk); 10858 return (error); 10859} 10860 10861/* 10862 * A large burst of file addition or deletion activity can drive the 10863 * memory load excessively high. First attempt to slow things down 10864 * using the techniques below. If that fails, this routine requests 10865 * the offending operations to fall back to running synchronously 10866 * until the memory load returns to a reasonable level. 10867 */ 10868int 10869softdep_slowdown(vp) 10870 struct vnode *vp; 10871{ 10872 int max_softdeps_hard; 10873 10874 ACQUIRE_LOCK(&lk); 10875 max_softdeps_hard = max_softdeps * 11 / 10; 10876 if (num_dirrem < max_softdeps_hard / 2 && 10877 num_inodedep < max_softdeps_hard && 10878 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps && 10879 num_freeblkdep < max_softdeps_hard) { 10880 FREE_LOCK(&lk); 10881 return (0); 10882 } 10883 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps) 10884 softdep_speedup(); 10885 stat_sync_limit_hit += 1; 10886 FREE_LOCK(&lk); 10887 return (1); 10888} 10889 10890/* 10891 * Called by the allocation routines when they are about to fail 10892 * in the hope that we can free up some disk space. 10893 * 10894 * First check to see if the work list has anything on it. If it has, 10895 * clean up entries until we successfully free some space. Because this 10896 * process holds inodes locked, we cannot handle any remove requests 10897 * that might block on a locked inode as that could lead to deadlock. 10898 * If the worklist yields no free space, encourage the syncer daemon 10899 * to help us. In no event will we try for longer than tickdelay seconds. 10900 */ 10901int 10902softdep_request_cleanup(fs, vp) 10903 struct fs *fs; 10904 struct vnode *vp; 10905{ 10906 struct ufsmount *ump; 10907 long starttime; 10908 ufs2_daddr_t needed; 10909 int error; 10910 10911 ump = VTOI(vp)->i_ump; 10912 mtx_assert(UFS_MTX(ump), MA_OWNED); 10913 needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize; 10914 starttime = time_second + tickdelay; 10915 /* 10916 * If we are being called because of a process doing a 10917 * copy-on-write, then it is not safe to update the vnode 10918 * as we may recurse into the copy-on-write routine. 10919 */ 10920 if (!(curthread->td_pflags & TDP_COWINPROGRESS)) { 10921 UFS_UNLOCK(ump); 10922 error = ffs_update(vp, 1); 10923 UFS_LOCK(ump); 10924 if (error != 0) 10925 return (0); 10926 } 10927 while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) { 10928 if (time_second > starttime) 10929 return (0); 10930 UFS_UNLOCK(ump); 10931 ACQUIRE_LOCK(&lk); 10932 process_removes(vp); 10933 if (ump->softdep_on_worklist > 0 && 10934 process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) { 10935 stat_worklist_push += 1; 10936 FREE_LOCK(&lk); 10937 UFS_LOCK(ump); 10938 continue; 10939 } 10940 request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT); 10941 FREE_LOCK(&lk); 10942 UFS_LOCK(ump); 10943 } 10944 return (1); 10945} 10946 10947/* 10948 * If memory utilization has gotten too high, deliberately slow things 10949 * down and speed up the I/O processing. 10950 */ 10951extern struct thread *syncertd; 10952static int 10953request_cleanup(mp, resource) 10954 struct mount *mp; 10955 int resource; 10956{ 10957 struct thread *td = curthread; 10958 struct ufsmount *ump; 10959 10960 mtx_assert(&lk, MA_OWNED); 10961 /* 10962 * We never hold up the filesystem syncer or buf daemon. 10963 */ 10964 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 10965 return (0); 10966 ump = VFSTOUFS(mp); 10967 /* 10968 * First check to see if the work list has gotten backlogged. 10969 * If it has, co-opt this process to help clean up two entries. 10970 * Because this process may hold inodes locked, we cannot 10971 * handle any remove requests that might block on a locked 10972 * inode as that could lead to deadlock. We set TDP_SOFTDEP 10973 * to avoid recursively processing the worklist. 10974 */ 10975 if (ump->softdep_on_worklist > max_softdeps / 10) { 10976 td->td_pflags |= TDP_SOFTDEP; 10977 process_worklist_item(mp, LK_NOWAIT); 10978 process_worklist_item(mp, LK_NOWAIT); 10979 td->td_pflags &= ~TDP_SOFTDEP; 10980 stat_worklist_push += 2; 10981 return(1); 10982 } 10983 /* 10984 * Next, we attempt to speed up the syncer process. If that 10985 * is successful, then we allow the process to continue. 10986 */ 10987 if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT) 10988 return(0); 10989 /* 10990 * If we are resource constrained on inode dependencies, try 10991 * flushing some dirty inodes. Otherwise, we are constrained 10992 * by file deletions, so try accelerating flushes of directories 10993 * with removal dependencies. We would like to do the cleanup 10994 * here, but we probably hold an inode locked at this point and 10995 * that might deadlock against one that we try to clean. So, 10996 * the best that we can do is request the syncer daemon to do 10997 * the cleanup for us. 10998 */ 10999 switch (resource) { 11000 11001 case FLUSH_INODES: 11002 stat_ino_limit_push += 1; 11003 req_clear_inodedeps += 1; 11004 stat_countp = &stat_ino_limit_hit; 11005 break; 11006 11007 case FLUSH_REMOVE: 11008 case FLUSH_REMOVE_WAIT: 11009 stat_blk_limit_push += 1; 11010 req_clear_remove += 1; 11011 stat_countp = &stat_blk_limit_hit; 11012 break; 11013 11014 default: 11015 panic("request_cleanup: unknown type"); 11016 } 11017 /* 11018 * Hopefully the syncer daemon will catch up and awaken us. 11019 * We wait at most tickdelay before proceeding in any case. 11020 */ 11021 proc_waiting += 1; 11022 if (callout_pending(&softdep_callout) == FALSE) 11023 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 11024 pause_timer, 0); 11025 11026 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 11027 proc_waiting -= 1; 11028 return (1); 11029} 11030 11031/* 11032 * Awaken processes pausing in request_cleanup and clear proc_waiting 11033 * to indicate that there is no longer a timer running. 11034 */ 11035static void 11036pause_timer(arg) 11037 void *arg; 11038{ 11039 11040 /* 11041 * The callout_ API has acquired mtx and will hold it around this 11042 * function call. 11043 */ 11044 *stat_countp += 1; 11045 wakeup_one(&proc_waiting); 11046 if (proc_waiting > 0) 11047 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 11048 pause_timer, 0); 11049} 11050 11051/* 11052 * Flush out a directory with at least one removal dependency in an effort to 11053 * reduce the number of dirrem, freefile, and freeblks dependency structures. 11054 */ 11055static void 11056clear_remove(td) 11057 struct thread *td; 11058{ 11059 struct pagedep_hashhead *pagedephd; 11060 struct pagedep *pagedep; 11061 static int next = 0; 11062 struct mount *mp; 11063 struct vnode *vp; 11064 struct bufobj *bo; 11065 int error, cnt; 11066 ino_t ino; 11067 11068 mtx_assert(&lk, MA_OWNED); 11069 11070 for (cnt = 0; cnt < pagedep_hash; cnt++) { 11071 pagedephd = &pagedep_hashtbl[next++]; 11072 if (next >= pagedep_hash) 11073 next = 0; 11074 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 11075 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 11076 continue; 11077 mp = pagedep->pd_list.wk_mp; 11078 ino = pagedep->pd_ino; 11079 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 11080 continue; 11081 FREE_LOCK(&lk); 11082 11083 /* 11084 * Let unmount clear deps 11085 */ 11086 error = vfs_busy(mp, MBF_NOWAIT); 11087 if (error != 0) 11088 goto finish_write; 11089 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 11090 FFSV_FORCEINSMQ); 11091 vfs_unbusy(mp); 11092 if (error != 0) { 11093 softdep_error("clear_remove: vget", error); 11094 goto finish_write; 11095 } 11096 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 11097 softdep_error("clear_remove: fsync", error); 11098 bo = &vp->v_bufobj; 11099 BO_LOCK(bo); 11100 drain_output(vp); 11101 BO_UNLOCK(bo); 11102 vput(vp); 11103 finish_write: 11104 vn_finished_write(mp); 11105 ACQUIRE_LOCK(&lk); 11106 return; 11107 } 11108 } 11109} 11110 11111/* 11112 * Clear out a block of dirty inodes in an effort to reduce 11113 * the number of inodedep dependency structures. 11114 */ 11115static void 11116clear_inodedeps(td) 11117 struct thread *td; 11118{ 11119 struct inodedep_hashhead *inodedephd; 11120 struct inodedep *inodedep; 11121 static int next = 0; 11122 struct mount *mp; 11123 struct vnode *vp; 11124 struct fs *fs; 11125 int error, cnt; 11126 ino_t firstino, lastino, ino; 11127 11128 mtx_assert(&lk, MA_OWNED); 11129 /* 11130 * Pick a random inode dependency to be cleared. 11131 * We will then gather up all the inodes in its block 11132 * that have dependencies and flush them out. 11133 */ 11134 for (cnt = 0; cnt < inodedep_hash; cnt++) { 11135 inodedephd = &inodedep_hashtbl[next++]; 11136 if (next >= inodedep_hash) 11137 next = 0; 11138 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 11139 break; 11140 } 11141 if (inodedep == NULL) 11142 return; 11143 fs = inodedep->id_fs; 11144 mp = inodedep->id_list.wk_mp; 11145 /* 11146 * Find the last inode in the block with dependencies. 11147 */ 11148 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 11149 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 11150 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 11151 break; 11152 /* 11153 * Asynchronously push all but the last inode with dependencies. 11154 * Synchronously push the last inode with dependencies to ensure 11155 * that the inode block gets written to free up the inodedeps. 11156 */ 11157 for (ino = firstino; ino <= lastino; ino++) { 11158 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 11159 continue; 11160 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 11161 continue; 11162 FREE_LOCK(&lk); 11163 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 11164 if (error != 0) { 11165 vn_finished_write(mp); 11166 ACQUIRE_LOCK(&lk); 11167 return; 11168 } 11169 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 11170 FFSV_FORCEINSMQ)) != 0) { 11171 softdep_error("clear_inodedeps: vget", error); 11172 vfs_unbusy(mp); 11173 vn_finished_write(mp); 11174 ACQUIRE_LOCK(&lk); 11175 return; 11176 } 11177 vfs_unbusy(mp); 11178 if (ino == lastino) { 11179 if ((error = ffs_syncvnode(vp, MNT_WAIT))) 11180 softdep_error("clear_inodedeps: fsync1", error); 11181 } else { 11182 if ((error = ffs_syncvnode(vp, MNT_NOWAIT))) 11183 softdep_error("clear_inodedeps: fsync2", error); 11184 BO_LOCK(&vp->v_bufobj); 11185 drain_output(vp); 11186 BO_UNLOCK(&vp->v_bufobj); 11187 } 11188 vput(vp); 11189 vn_finished_write(mp); 11190 ACQUIRE_LOCK(&lk); 11191 } 11192} 11193 11194/* 11195 * Function to determine if the buffer has outstanding dependencies 11196 * that will cause a roll-back if the buffer is written. If wantcount 11197 * is set, return number of dependencies, otherwise just yes or no. 11198 */ 11199static int 11200softdep_count_dependencies(bp, wantcount) 11201 struct buf *bp; 11202 int wantcount; 11203{ 11204 struct worklist *wk; 11205 struct bmsafemap *bmsafemap; 11206 struct inodedep *inodedep; 11207 struct indirdep *indirdep; 11208 struct freeblks *freeblks; 11209 struct allocindir *aip; 11210 struct pagedep *pagedep; 11211 struct dirrem *dirrem; 11212 struct newblk *newblk; 11213 struct mkdir *mkdir; 11214 struct diradd *dap; 11215 int i, retval; 11216 11217 retval = 0; 11218 ACQUIRE_LOCK(&lk); 11219 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11220 switch (wk->wk_type) { 11221 11222 case D_INODEDEP: 11223 inodedep = WK_INODEDEP(wk); 11224 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 11225 /* bitmap allocation dependency */ 11226 retval += 1; 11227 if (!wantcount) 11228 goto out; 11229 } 11230 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 11231 /* direct block pointer dependency */ 11232 retval += 1; 11233 if (!wantcount) 11234 goto out; 11235 } 11236 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 11237 /* direct block pointer dependency */ 11238 retval += 1; 11239 if (!wantcount) 11240 goto out; 11241 } 11242 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 11243 /* Add reference dependency. */ 11244 retval += 1; 11245 if (!wantcount) 11246 goto out; 11247 } 11248 continue; 11249 11250 case D_INDIRDEP: 11251 indirdep = WK_INDIRDEP(wk); 11252 11253 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 11254 /* indirect block pointer dependency */ 11255 retval += 1; 11256 if (!wantcount) 11257 goto out; 11258 } 11259 continue; 11260 11261 case D_PAGEDEP: 11262 pagedep = WK_PAGEDEP(wk); 11263 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 11264 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 11265 /* Journal remove ref dependency. */ 11266 retval += 1; 11267 if (!wantcount) 11268 goto out; 11269 } 11270 } 11271 for (i = 0; i < DAHASHSZ; i++) { 11272 11273 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 11274 /* directory entry dependency */ 11275 retval += 1; 11276 if (!wantcount) 11277 goto out; 11278 } 11279 } 11280 continue; 11281 11282 case D_BMSAFEMAP: 11283 bmsafemap = WK_BMSAFEMAP(wk); 11284 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 11285 /* Add reference dependency. */ 11286 retval += 1; 11287 if (!wantcount) 11288 goto out; 11289 } 11290 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 11291 /* Allocate block dependency. */ 11292 retval += 1; 11293 if (!wantcount) 11294 goto out; 11295 } 11296 continue; 11297 11298 case D_FREEBLKS: 11299 freeblks = WK_FREEBLKS(wk); 11300 if (LIST_FIRST(&freeblks->fb_jfreeblkhd)) { 11301 /* Freeblk journal dependency. */ 11302 retval += 1; 11303 if (!wantcount) 11304 goto out; 11305 } 11306 continue; 11307 11308 case D_ALLOCDIRECT: 11309 case D_ALLOCINDIR: 11310 newblk = WK_NEWBLK(wk); 11311 if (newblk->nb_jnewblk) { 11312 /* Journal allocate dependency. */ 11313 retval += 1; 11314 if (!wantcount) 11315 goto out; 11316 } 11317 continue; 11318 11319 case D_MKDIR: 11320 mkdir = WK_MKDIR(wk); 11321 if (mkdir->md_jaddref) { 11322 /* Journal reference dependency. */ 11323 retval += 1; 11324 if (!wantcount) 11325 goto out; 11326 } 11327 continue; 11328 11329 case D_FREEWORK: 11330 case D_FREEDEP: 11331 case D_JSEGDEP: 11332 case D_JSEG: 11333 case D_SBDEP: 11334 /* never a dependency on these blocks */ 11335 continue; 11336 11337 default: 11338 panic("softdep_count_dependencies: Unexpected type %s", 11339 TYPENAME(wk->wk_type)); 11340 /* NOTREACHED */ 11341 } 11342 } 11343out: 11344 FREE_LOCK(&lk); 11345 return retval; 11346} 11347 11348/* 11349 * Acquire exclusive access to a buffer. 11350 * Must be called with a locked mtx parameter. 11351 * Return acquired buffer or NULL on failure. 11352 */ 11353static struct buf * 11354getdirtybuf(bp, mtx, waitfor) 11355 struct buf *bp; 11356 struct mtx *mtx; 11357 int waitfor; 11358{ 11359 int error; 11360 11361 mtx_assert(mtx, MA_OWNED); 11362 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 11363 if (waitfor != MNT_WAIT) 11364 return (NULL); 11365 error = BUF_LOCK(bp, 11366 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 11367 /* 11368 * Even if we sucessfully acquire bp here, we have dropped 11369 * mtx, which may violates our guarantee. 11370 */ 11371 if (error == 0) 11372 BUF_UNLOCK(bp); 11373 else if (error != ENOLCK) 11374 panic("getdirtybuf: inconsistent lock: %d", error); 11375 mtx_lock(mtx); 11376 return (NULL); 11377 } 11378 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 11379 if (mtx == &lk && waitfor == MNT_WAIT) { 11380 mtx_unlock(mtx); 11381 BO_LOCK(bp->b_bufobj); 11382 BUF_UNLOCK(bp); 11383 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 11384 bp->b_vflags |= BV_BKGRDWAIT; 11385 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 11386 PRIBIO | PDROP, "getbuf", 0); 11387 } else 11388 BO_UNLOCK(bp->b_bufobj); 11389 mtx_lock(mtx); 11390 return (NULL); 11391 } 11392 BUF_UNLOCK(bp); 11393 if (waitfor != MNT_WAIT) 11394 return (NULL); 11395 /* 11396 * The mtx argument must be bp->b_vp's mutex in 11397 * this case. 11398 */ 11399#ifdef DEBUG_VFS_LOCKS 11400 if (bp->b_vp->v_type != VCHR) 11401 ASSERT_BO_LOCKED(bp->b_bufobj); 11402#endif 11403 bp->b_vflags |= BV_BKGRDWAIT; 11404 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 11405 return (NULL); 11406 } 11407 if ((bp->b_flags & B_DELWRI) == 0) { 11408 BUF_UNLOCK(bp); 11409 return (NULL); 11410 } 11411 bremfree(bp); 11412 return (bp); 11413} 11414 11415 11416/* 11417 * Check if it is safe to suspend the file system now. On entry, 11418 * the vnode interlock for devvp should be held. Return 0 with 11419 * the mount interlock held if the file system can be suspended now, 11420 * otherwise return EAGAIN with the mount interlock held. 11421 */ 11422int 11423softdep_check_suspend(struct mount *mp, 11424 struct vnode *devvp, 11425 int softdep_deps, 11426 int softdep_accdeps, 11427 int secondary_writes, 11428 int secondary_accwrites) 11429{ 11430 struct bufobj *bo; 11431 struct ufsmount *ump; 11432 int error; 11433 11434 ump = VFSTOUFS(mp); 11435 bo = &devvp->v_bufobj; 11436 ASSERT_BO_LOCKED(bo); 11437 11438 for (;;) { 11439 if (!TRY_ACQUIRE_LOCK(&lk)) { 11440 BO_UNLOCK(bo); 11441 ACQUIRE_LOCK(&lk); 11442 FREE_LOCK(&lk); 11443 BO_LOCK(bo); 11444 continue; 11445 } 11446 MNT_ILOCK(mp); 11447 if (mp->mnt_secondary_writes != 0) { 11448 FREE_LOCK(&lk); 11449 BO_UNLOCK(bo); 11450 msleep(&mp->mnt_secondary_writes, 11451 MNT_MTX(mp), 11452 (PUSER - 1) | PDROP, "secwr", 0); 11453 BO_LOCK(bo); 11454 continue; 11455 } 11456 break; 11457 } 11458 11459 /* 11460 * Reasons for needing more work before suspend: 11461 * - Dirty buffers on devvp. 11462 * - Softdep activity occurred after start of vnode sync loop 11463 * - Secondary writes occurred after start of vnode sync loop 11464 */ 11465 error = 0; 11466 if (bo->bo_numoutput > 0 || 11467 bo->bo_dirty.bv_cnt > 0 || 11468 softdep_deps != 0 || 11469 ump->softdep_deps != 0 || 11470 softdep_accdeps != ump->softdep_accdeps || 11471 secondary_writes != 0 || 11472 mp->mnt_secondary_writes != 0 || 11473 secondary_accwrites != mp->mnt_secondary_accwrites) 11474 error = EAGAIN; 11475 FREE_LOCK(&lk); 11476 BO_UNLOCK(bo); 11477 return (error); 11478} 11479 11480 11481/* 11482 * Get the number of dependency structures for the file system, both 11483 * the current number and the total number allocated. These will 11484 * later be used to detect that softdep processing has occurred. 11485 */ 11486void 11487softdep_get_depcounts(struct mount *mp, 11488 int *softdep_depsp, 11489 int *softdep_accdepsp) 11490{ 11491 struct ufsmount *ump; 11492 11493 ump = VFSTOUFS(mp); 11494 ACQUIRE_LOCK(&lk); 11495 *softdep_depsp = ump->softdep_deps; 11496 *softdep_accdepsp = ump->softdep_accdeps; 11497 FREE_LOCK(&lk); 11498} 11499 11500/* 11501 * Wait for pending output on a vnode to complete. 11502 * Must be called with vnode lock and interlock locked. 11503 * 11504 * XXX: Should just be a call to bufobj_wwait(). 11505 */ 11506static void 11507drain_output(vp) 11508 struct vnode *vp; 11509{ 11510 struct bufobj *bo; 11511 11512 bo = &vp->v_bufobj; 11513 ASSERT_VOP_LOCKED(vp, "drain_output"); 11514 ASSERT_BO_LOCKED(bo); 11515 11516 while (bo->bo_numoutput) { 11517 bo->bo_flag |= BO_WWAIT; 11518 msleep((caddr_t)&bo->bo_numoutput, 11519 BO_MTX(bo), PRIBIO + 1, "drainvp", 0); 11520 } 11521} 11522 11523/* 11524 * Called whenever a buffer that is being invalidated or reallocated 11525 * contains dependencies. This should only happen if an I/O error has 11526 * occurred. The routine is called with the buffer locked. 11527 */ 11528static void 11529softdep_deallocate_dependencies(bp) 11530 struct buf *bp; 11531{ 11532 11533 if ((bp->b_ioflags & BIO_ERROR) == 0) 11534 panic("softdep_deallocate_dependencies: dangling deps"); 11535 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 11536 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 11537} 11538 11539/* 11540 * Function to handle asynchronous write errors in the filesystem. 11541 */ 11542static void 11543softdep_error(func, error) 11544 char *func; 11545 int error; 11546{ 11547 11548 /* XXX should do something better! */ 11549 printf("%s: got error %d while accessing filesystem\n", func, error); 11550} 11551 11552#ifdef DDB 11553 11554static void 11555inodedep_print(struct inodedep *inodedep, int verbose) 11556{ 11557 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d" 11558 " saveino %p\n", 11559 inodedep, inodedep->id_fs, inodedep->id_state, 11560 (intmax_t)inodedep->id_ino, 11561 (intmax_t)fsbtodb(inodedep->id_fs, 11562 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 11563 inodedep->id_nlinkdelta, inodedep->id_savednlink, 11564 inodedep->id_savedino1); 11565 11566 if (verbose == 0) 11567 return; 11568 11569 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 11570 "mkdiradd %p\n", 11571 LIST_FIRST(&inodedep->id_pendinghd), 11572 LIST_FIRST(&inodedep->id_bufwait), 11573 LIST_FIRST(&inodedep->id_inowait), 11574 TAILQ_FIRST(&inodedep->id_inoreflst), 11575 inodedep->id_mkdiradd); 11576 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 11577 TAILQ_FIRST(&inodedep->id_inoupdt), 11578 TAILQ_FIRST(&inodedep->id_newinoupdt), 11579 TAILQ_FIRST(&inodedep->id_extupdt), 11580 TAILQ_FIRST(&inodedep->id_newextupdt)); 11581} 11582 11583DB_SHOW_COMMAND(inodedep, db_show_inodedep) 11584{ 11585 11586 if (have_addr == 0) { 11587 db_printf("Address required\n"); 11588 return; 11589 } 11590 inodedep_print((struct inodedep*)addr, 1); 11591} 11592 11593DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 11594{ 11595 struct inodedep_hashhead *inodedephd; 11596 struct inodedep *inodedep; 11597 struct fs *fs; 11598 int cnt; 11599 11600 fs = have_addr ? (struct fs *)addr : NULL; 11601 for (cnt = 0; cnt < inodedep_hash; cnt++) { 11602 inodedephd = &inodedep_hashtbl[cnt]; 11603 LIST_FOREACH(inodedep, inodedephd, id_hash) { 11604 if (fs != NULL && fs != inodedep->id_fs) 11605 continue; 11606 inodedep_print(inodedep, 0); 11607 } 11608 } 11609} 11610 11611DB_SHOW_COMMAND(worklist, db_show_worklist) 11612{ 11613 struct worklist *wk; 11614 11615 if (have_addr == 0) { 11616 db_printf("Address required\n"); 11617 return; 11618 } 11619 wk = (struct worklist *)addr; 11620 printf("worklist: %p type %s state 0x%X\n", 11621 wk, TYPENAME(wk->wk_type), wk->wk_state); 11622} 11623 11624DB_SHOW_COMMAND(workhead, db_show_workhead) 11625{ 11626 struct workhead *wkhd; 11627 struct worklist *wk; 11628 int i; 11629 11630 if (have_addr == 0) { 11631 db_printf("Address required\n"); 11632 return; 11633 } 11634 wkhd = (struct workhead *)addr; 11635 wk = LIST_FIRST(wkhd); 11636 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 11637 db_printf("worklist: %p type %s state 0x%X", 11638 wk, TYPENAME(wk->wk_type), wk->wk_state); 11639 if (i == 100) 11640 db_printf("workhead overflow"); 11641 printf("\n"); 11642} 11643 11644 11645DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 11646{ 11647 struct jaddref *jaddref; 11648 struct diradd *diradd; 11649 struct mkdir *mkdir; 11650 11651 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 11652 diradd = mkdir->md_diradd; 11653 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 11654 mkdir, mkdir->md_state, diradd, diradd->da_state); 11655 if ((jaddref = mkdir->md_jaddref) != NULL) 11656 db_printf(" jaddref %p jaddref state 0x%X", 11657 jaddref, jaddref->ja_state); 11658 db_printf("\n"); 11659 } 11660} 11661 11662#endif /* DDB */ 11663 11664#endif /* SOFTUPDATES */ 11665