ffs_softdep.c revision 249218
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 249218 2013-04-06 22:21:23Z jeff $"); 44 45#include "opt_ffs.h" 46#include "opt_quota.h" 47#include "opt_ddb.h" 48 49/* 50 * For now we want the safety net that the DEBUG flag provides. 51 */ 52#ifndef DEBUG 53#define DEBUG 54#endif 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/ktr.h> 64#include <sys/limits.h> 65#include <sys/lock.h> 66#include <sys/malloc.h> 67#include <sys/mount.h> 68#include <sys/mutex.h> 69#include <sys/namei.h> 70#include <sys/priv.h> 71#include <sys/proc.h> 72#include <sys/stat.h> 73#include <sys/sysctl.h> 74#include <sys/syslog.h> 75#include <sys/vnode.h> 76#include <sys/conf.h> 77 78#include <ufs/ufs/dir.h> 79#include <ufs/ufs/extattr.h> 80#include <ufs/ufs/quota.h> 81#include <ufs/ufs/inode.h> 82#include <ufs/ufs/ufsmount.h> 83#include <ufs/ffs/fs.h> 84#include <ufs/ffs/softdep.h> 85#include <ufs/ffs/ffs_extern.h> 86#include <ufs/ufs/ufs_extern.h> 87 88#include <vm/vm.h> 89#include <vm/vm_extern.h> 90#include <vm/vm_object.h> 91 92#include <geom/geom.h> 93 94#include <ddb/ddb.h> 95 96#define KTR_SUJ 0 /* Define to KTR_SPARE. */ 97 98#ifndef SOFTUPDATES 99 100int 101softdep_flushfiles(oldmnt, flags, td) 102 struct mount *oldmnt; 103 int flags; 104 struct thread *td; 105{ 106 107 panic("softdep_flushfiles called"); 108} 109 110int 111softdep_mount(devvp, mp, fs, cred) 112 struct vnode *devvp; 113 struct mount *mp; 114 struct fs *fs; 115 struct ucred *cred; 116{ 117 118 return (0); 119} 120 121void 122softdep_initialize() 123{ 124 125 return; 126} 127 128void 129softdep_uninitialize() 130{ 131 132 return; 133} 134 135void 136softdep_unmount(mp) 137 struct mount *mp; 138{ 139 140} 141 142void 143softdep_setup_sbupdate(ump, fs, bp) 144 struct ufsmount *ump; 145 struct fs *fs; 146 struct buf *bp; 147{ 148} 149 150void 151softdep_setup_inomapdep(bp, ip, newinum, mode) 152 struct buf *bp; 153 struct inode *ip; 154 ino_t newinum; 155 int mode; 156{ 157 158 panic("softdep_setup_inomapdep called"); 159} 160 161void 162softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 163 struct buf *bp; 164 struct mount *mp; 165 ufs2_daddr_t newblkno; 166 int frags; 167 int oldfrags; 168{ 169 170 panic("softdep_setup_blkmapdep called"); 171} 172 173void 174softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 175 struct inode *ip; 176 ufs_lbn_t lbn; 177 ufs2_daddr_t newblkno; 178 ufs2_daddr_t oldblkno; 179 long newsize; 180 long oldsize; 181 struct buf *bp; 182{ 183 184 panic("softdep_setup_allocdirect called"); 185} 186 187void 188softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 189 struct inode *ip; 190 ufs_lbn_t lbn; 191 ufs2_daddr_t newblkno; 192 ufs2_daddr_t oldblkno; 193 long newsize; 194 long oldsize; 195 struct buf *bp; 196{ 197 198 panic("softdep_setup_allocext called"); 199} 200 201void 202softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 203 struct inode *ip; 204 ufs_lbn_t lbn; 205 struct buf *bp; 206 int ptrno; 207 ufs2_daddr_t newblkno; 208 ufs2_daddr_t oldblkno; 209 struct buf *nbp; 210{ 211 212 panic("softdep_setup_allocindir_page called"); 213} 214 215void 216softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 217 struct buf *nbp; 218 struct inode *ip; 219 struct buf *bp; 220 int ptrno; 221 ufs2_daddr_t newblkno; 222{ 223 224 panic("softdep_setup_allocindir_meta called"); 225} 226 227void 228softdep_journal_freeblocks(ip, cred, length, flags) 229 struct inode *ip; 230 struct ucred *cred; 231 off_t length; 232 int flags; 233{ 234 235 panic("softdep_journal_freeblocks called"); 236} 237 238void 239softdep_journal_fsync(ip) 240 struct inode *ip; 241{ 242 243 panic("softdep_journal_fsync called"); 244} 245 246void 247softdep_setup_freeblocks(ip, length, flags) 248 struct inode *ip; 249 off_t length; 250 int flags; 251{ 252 253 panic("softdep_setup_freeblocks called"); 254} 255 256void 257softdep_freefile(pvp, ino, mode) 258 struct vnode *pvp; 259 ino_t ino; 260 int mode; 261{ 262 263 panic("softdep_freefile called"); 264} 265 266int 267softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 268 struct buf *bp; 269 struct inode *dp; 270 off_t diroffset; 271 ino_t newinum; 272 struct buf *newdirbp; 273 int isnewblk; 274{ 275 276 panic("softdep_setup_directory_add called"); 277} 278 279void 280softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 281 struct buf *bp; 282 struct inode *dp; 283 caddr_t base; 284 caddr_t oldloc; 285 caddr_t newloc; 286 int entrysize; 287{ 288 289 panic("softdep_change_directoryentry_offset called"); 290} 291 292void 293softdep_setup_remove(bp, dp, ip, isrmdir) 294 struct buf *bp; 295 struct inode *dp; 296 struct inode *ip; 297 int isrmdir; 298{ 299 300 panic("softdep_setup_remove called"); 301} 302 303void 304softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 305 struct buf *bp; 306 struct inode *dp; 307 struct inode *ip; 308 ino_t newinum; 309 int isrmdir; 310{ 311 312 panic("softdep_setup_directory_change called"); 313} 314 315void 316softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 317 struct mount *mp; 318 struct buf *bp; 319 ufs2_daddr_t blkno; 320 int frags; 321 struct workhead *wkhd; 322{ 323 324 panic("%s called", __FUNCTION__); 325} 326 327void 328softdep_setup_inofree(mp, bp, ino, wkhd) 329 struct mount *mp; 330 struct buf *bp; 331 ino_t ino; 332 struct workhead *wkhd; 333{ 334 335 panic("%s called", __FUNCTION__); 336} 337 338void 339softdep_setup_unlink(dp, ip) 340 struct inode *dp; 341 struct inode *ip; 342{ 343 344 panic("%s called", __FUNCTION__); 345} 346 347void 348softdep_setup_link(dp, ip) 349 struct inode *dp; 350 struct inode *ip; 351{ 352 353 panic("%s called", __FUNCTION__); 354} 355 356void 357softdep_revert_link(dp, ip) 358 struct inode *dp; 359 struct inode *ip; 360{ 361 362 panic("%s called", __FUNCTION__); 363} 364 365void 366softdep_setup_rmdir(dp, ip) 367 struct inode *dp; 368 struct inode *ip; 369{ 370 371 panic("%s called", __FUNCTION__); 372} 373 374void 375softdep_revert_rmdir(dp, ip) 376 struct inode *dp; 377 struct inode *ip; 378{ 379 380 panic("%s called", __FUNCTION__); 381} 382 383void 384softdep_setup_create(dp, ip) 385 struct inode *dp; 386 struct inode *ip; 387{ 388 389 panic("%s called", __FUNCTION__); 390} 391 392void 393softdep_revert_create(dp, ip) 394 struct inode *dp; 395 struct inode *ip; 396{ 397 398 panic("%s called", __FUNCTION__); 399} 400 401void 402softdep_setup_mkdir(dp, ip) 403 struct inode *dp; 404 struct inode *ip; 405{ 406 407 panic("%s called", __FUNCTION__); 408} 409 410void 411softdep_revert_mkdir(dp, ip) 412 struct inode *dp; 413 struct inode *ip; 414{ 415 416 panic("%s called", __FUNCTION__); 417} 418 419void 420softdep_setup_dotdot_link(dp, ip) 421 struct inode *dp; 422 struct inode *ip; 423{ 424 425 panic("%s called", __FUNCTION__); 426} 427 428int 429softdep_prealloc(vp, waitok) 430 struct vnode *vp; 431 int waitok; 432{ 433 434 panic("%s called", __FUNCTION__); 435 436 return (0); 437} 438 439int 440softdep_journal_lookup(mp, vpp) 441 struct mount *mp; 442 struct vnode **vpp; 443{ 444 445 return (ENOENT); 446} 447 448void 449softdep_change_linkcnt(ip) 450 struct inode *ip; 451{ 452 453 panic("softdep_change_linkcnt called"); 454} 455 456void 457softdep_load_inodeblock(ip) 458 struct inode *ip; 459{ 460 461 panic("softdep_load_inodeblock called"); 462} 463 464void 465softdep_update_inodeblock(ip, bp, waitfor) 466 struct inode *ip; 467 struct buf *bp; 468 int waitfor; 469{ 470 471 panic("softdep_update_inodeblock called"); 472} 473 474int 475softdep_fsync(vp) 476 struct vnode *vp; /* the "in_core" copy of the inode */ 477{ 478 479 return (0); 480} 481 482void 483softdep_fsync_mountdev(vp) 484 struct vnode *vp; 485{ 486 487 return; 488} 489 490int 491softdep_flushworklist(oldmnt, countp, td) 492 struct mount *oldmnt; 493 int *countp; 494 struct thread *td; 495{ 496 497 *countp = 0; 498 return (0); 499} 500 501int 502softdep_sync_metadata(struct vnode *vp) 503{ 504 505 return (0); 506} 507 508int 509softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 510{ 511 512 return (0); 513} 514 515int 516softdep_slowdown(vp) 517 struct vnode *vp; 518{ 519 520 panic("softdep_slowdown called"); 521} 522 523void 524softdep_releasefile(ip) 525 struct inode *ip; /* inode with the zero effective link count */ 526{ 527 528 panic("softdep_releasefile called"); 529} 530 531int 532softdep_request_cleanup(fs, vp, cred, resource) 533 struct fs *fs; 534 struct vnode *vp; 535 struct ucred *cred; 536 int resource; 537{ 538 539 return (0); 540} 541 542int 543softdep_check_suspend(struct mount *mp, 544 struct vnode *devvp, 545 int softdep_deps, 546 int softdep_accdeps, 547 int secondary_writes, 548 int secondary_accwrites) 549{ 550 struct bufobj *bo; 551 int error; 552 553 (void) softdep_deps, 554 (void) softdep_accdeps; 555 556 bo = &devvp->v_bufobj; 557 ASSERT_BO_LOCKED(bo); 558 559 MNT_ILOCK(mp); 560 while (mp->mnt_secondary_writes != 0) { 561 BO_UNLOCK(bo); 562 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 563 (PUSER - 1) | PDROP, "secwr", 0); 564 BO_LOCK(bo); 565 MNT_ILOCK(mp); 566 } 567 568 /* 569 * Reasons for needing more work before suspend: 570 * - Dirty buffers on devvp. 571 * - Secondary writes occurred after start of vnode sync loop 572 */ 573 error = 0; 574 if (bo->bo_numoutput > 0 || 575 bo->bo_dirty.bv_cnt > 0 || 576 secondary_writes != 0 || 577 mp->mnt_secondary_writes != 0 || 578 secondary_accwrites != mp->mnt_secondary_accwrites) 579 error = EAGAIN; 580 BO_UNLOCK(bo); 581 return (error); 582} 583 584void 585softdep_get_depcounts(struct mount *mp, 586 int *softdepactivep, 587 int *softdepactiveaccp) 588{ 589 (void) mp; 590 *softdepactivep = 0; 591 *softdepactiveaccp = 0; 592} 593 594void 595softdep_buf_append(bp, wkhd) 596 struct buf *bp; 597 struct workhead *wkhd; 598{ 599 600 panic("softdep_buf_appendwork called"); 601} 602 603void 604softdep_inode_append(ip, cred, wkhd) 605 struct inode *ip; 606 struct ucred *cred; 607 struct workhead *wkhd; 608{ 609 610 panic("softdep_inode_appendwork called"); 611} 612 613void 614softdep_freework(wkhd) 615 struct workhead *wkhd; 616{ 617 618 panic("softdep_freework called"); 619} 620 621#else 622 623FEATURE(softupdates, "FFS soft-updates support"); 624 625/* 626 * These definitions need to be adapted to the system to which 627 * this file is being ported. 628 */ 629 630#define M_SOFTDEP_FLAGS (M_WAITOK) 631 632#define D_PAGEDEP 0 633#define D_INODEDEP 1 634#define D_BMSAFEMAP 2 635#define D_NEWBLK 3 636#define D_ALLOCDIRECT 4 637#define D_INDIRDEP 5 638#define D_ALLOCINDIR 6 639#define D_FREEFRAG 7 640#define D_FREEBLKS 8 641#define D_FREEFILE 9 642#define D_DIRADD 10 643#define D_MKDIR 11 644#define D_DIRREM 12 645#define D_NEWDIRBLK 13 646#define D_FREEWORK 14 647#define D_FREEDEP 15 648#define D_JADDREF 16 649#define D_JREMREF 17 650#define D_JMVREF 18 651#define D_JNEWBLK 19 652#define D_JFREEBLK 20 653#define D_JFREEFRAG 21 654#define D_JSEG 22 655#define D_JSEGDEP 23 656#define D_SBDEP 24 657#define D_JTRUNC 25 658#define D_JFSYNC 26 659#define D_SENTINAL 27 660#define D_LAST D_SENTINAL 661 662unsigned long dep_current[D_LAST + 1]; 663unsigned long dep_total[D_LAST + 1]; 664unsigned long dep_write[D_LAST + 1]; 665 666 667static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 668 "soft updates stats"); 669static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 670 "total dependencies allocated"); 671static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 672 "current dependencies allocated"); 673static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 674 "current dependencies written"); 675 676#define SOFTDEP_TYPE(type, str, long) \ 677 static MALLOC_DEFINE(M_ ## type, #str, long); \ 678 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 679 &dep_total[D_ ## type], 0, ""); \ 680 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 681 &dep_current[D_ ## type], 0, ""); \ 682 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 683 &dep_write[D_ ## type], 0, ""); 684 685SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 686SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 687SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 688 "Block or frag allocated from cyl group map"); 689SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 690SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 691SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 692SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 693SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 694SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 695SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 696SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 697SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 698SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 699SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 700SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 701SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 702SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 703SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 704SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 705SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 706SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 707SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 708SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 709SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 710SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 711SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 712SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 713 714static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 715static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 716 717/* 718 * translate from workitem type to memory type 719 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 720 */ 721static struct malloc_type *memtype[] = { 722 M_PAGEDEP, 723 M_INODEDEP, 724 M_BMSAFEMAP, 725 M_NEWBLK, 726 M_ALLOCDIRECT, 727 M_INDIRDEP, 728 M_ALLOCINDIR, 729 M_FREEFRAG, 730 M_FREEBLKS, 731 M_FREEFILE, 732 M_DIRADD, 733 M_MKDIR, 734 M_DIRREM, 735 M_NEWDIRBLK, 736 M_FREEWORK, 737 M_FREEDEP, 738 M_JADDREF, 739 M_JREMREF, 740 M_JMVREF, 741 M_JNEWBLK, 742 M_JFREEBLK, 743 M_JFREEFRAG, 744 M_JSEG, 745 M_JSEGDEP, 746 M_SBDEP, 747 M_JTRUNC, 748 M_JFSYNC 749}; 750 751static LIST_HEAD(mkdirlist, mkdir) mkdirlisthd; 752 753#define DtoM(type) (memtype[type]) 754 755/* 756 * Names of malloc types. 757 */ 758#define TYPENAME(type) \ 759 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 760/* 761 * End system adaptation definitions. 762 */ 763 764#define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 765#define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 766 767/* 768 * Forward declarations. 769 */ 770struct inodedep_hashhead; 771struct newblk_hashhead; 772struct pagedep_hashhead; 773struct bmsafemap_hashhead; 774 775/* 776 * Private journaling structures. 777 */ 778struct jblocks { 779 struct jseglst jb_segs; /* TAILQ of current segments. */ 780 struct jseg *jb_writeseg; /* Next write to complete. */ 781 struct jseg *jb_oldestseg; /* Oldest segment with valid entries. */ 782 struct jextent *jb_extent; /* Extent array. */ 783 uint64_t jb_nextseq; /* Next sequence number. */ 784 uint64_t jb_oldestwrseq; /* Oldest written sequence number. */ 785 uint8_t jb_needseg; /* Need a forced segment. */ 786 uint8_t jb_suspended; /* Did journal suspend writes? */ 787 int jb_avail; /* Available extents. */ 788 int jb_used; /* Last used extent. */ 789 int jb_head; /* Allocator head. */ 790 int jb_off; /* Allocator extent offset. */ 791 int jb_blocks; /* Total disk blocks covered. */ 792 int jb_free; /* Total disk blocks free. */ 793 int jb_min; /* Minimum free space. */ 794 int jb_low; /* Low on space. */ 795 int jb_age; /* Insertion time of oldest rec. */ 796}; 797 798struct jextent { 799 ufs2_daddr_t je_daddr; /* Disk block address. */ 800 int je_blocks; /* Disk block count. */ 801}; 802 803/* 804 * Internal function prototypes. 805 */ 806static void softdep_error(char *, int); 807static void drain_output(struct vnode *); 808static struct buf *getdirtybuf(struct buf *, struct mtx *, int); 809static void clear_remove(void); 810static void clear_inodedeps(void); 811static void unlinked_inodedep(struct mount *, struct inodedep *); 812static void clear_unlinked_inodedep(struct inodedep *); 813static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 814static int flush_pagedep_deps(struct vnode *, struct mount *, 815 struct diraddhd *); 816static int free_pagedep(struct pagedep *); 817static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 818static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 819static int flush_deplist(struct allocdirectlst *, int, int *); 820static int sync_cgs(struct mount *, int); 821static int handle_written_filepage(struct pagedep *, struct buf *); 822static int handle_written_sbdep(struct sbdep *, struct buf *); 823static void initiate_write_sbdep(struct sbdep *); 824static void diradd_inode_written(struct diradd *, struct inodedep *); 825static int handle_written_indirdep(struct indirdep *, struct buf *, 826 struct buf**); 827static int handle_written_inodeblock(struct inodedep *, struct buf *); 828static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 829 uint8_t *); 830static int handle_written_bmsafemap(struct bmsafemap *, struct buf *); 831static void handle_written_jaddref(struct jaddref *); 832static void handle_written_jremref(struct jremref *); 833static void handle_written_jseg(struct jseg *, struct buf *); 834static void handle_written_jnewblk(struct jnewblk *); 835static void handle_written_jblkdep(struct jblkdep *); 836static void handle_written_jfreefrag(struct jfreefrag *); 837static void complete_jseg(struct jseg *); 838static void complete_jsegs(struct jseg *); 839static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 840static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 841static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 842static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 843static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 844static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 845static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 846static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 847static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 848static inline void inoref_write(struct inoref *, struct jseg *, 849 struct jrefrec *); 850static void handle_allocdirect_partdone(struct allocdirect *, 851 struct workhead *); 852static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 853 struct workhead *); 854static void indirdep_complete(struct indirdep *); 855static int indirblk_lookup(struct mount *, ufs2_daddr_t); 856static void indirblk_insert(struct freework *); 857static void indirblk_remove(struct freework *); 858static void handle_allocindir_partdone(struct allocindir *); 859static void initiate_write_filepage(struct pagedep *, struct buf *); 860static void initiate_write_indirdep(struct indirdep*, struct buf *); 861static void handle_written_mkdir(struct mkdir *, int); 862static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 863 uint8_t *); 864static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 865static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 866static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 867static void handle_workitem_freefile(struct freefile *); 868static int handle_workitem_remove(struct dirrem *, int); 869static struct dirrem *newdirrem(struct buf *, struct inode *, 870 struct inode *, int, struct dirrem **); 871static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 872 struct buf *); 873static void cancel_indirdep(struct indirdep *, struct buf *, 874 struct freeblks *); 875static void free_indirdep(struct indirdep *); 876static void free_diradd(struct diradd *, struct workhead *); 877static void merge_diradd(struct inodedep *, struct diradd *); 878static void complete_diradd(struct diradd *); 879static struct diradd *diradd_lookup(struct pagedep *, int); 880static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 881 struct jremref *); 882static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 883 struct jremref *); 884static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 885 struct jremref *, struct jremref *); 886static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 887 struct jremref *); 888static void cancel_allocindir(struct allocindir *, struct buf *bp, 889 struct freeblks *, int); 890static int setup_trunc_indir(struct freeblks *, struct inode *, 891 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 892static void complete_trunc_indir(struct freework *); 893static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 894 int); 895static void complete_mkdir(struct mkdir *); 896static void free_newdirblk(struct newdirblk *); 897static void free_jremref(struct jremref *); 898static void free_jaddref(struct jaddref *); 899static void free_jsegdep(struct jsegdep *); 900static void free_jsegs(struct jblocks *); 901static void rele_jseg(struct jseg *); 902static void free_jseg(struct jseg *, struct jblocks *); 903static void free_jnewblk(struct jnewblk *); 904static void free_jblkdep(struct jblkdep *); 905static void free_jfreefrag(struct jfreefrag *); 906static void free_freedep(struct freedep *); 907static void journal_jremref(struct dirrem *, struct jremref *, 908 struct inodedep *); 909static void cancel_jnewblk(struct jnewblk *, struct workhead *); 910static int cancel_jaddref(struct jaddref *, struct inodedep *, 911 struct workhead *); 912static void cancel_jfreefrag(struct jfreefrag *); 913static inline void setup_freedirect(struct freeblks *, struct inode *, 914 int, int); 915static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 916static inline void setup_freeindir(struct freeblks *, struct inode *, int, 917 ufs_lbn_t, int); 918static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 919static void freeblks_free(struct ufsmount *, struct freeblks *, int); 920static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 921ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 922static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 923static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 924 int, int); 925static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 926static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 927static int deallocate_dependencies(struct buf *, struct freeblks *, int); 928static void newblk_freefrag(struct newblk*); 929static void free_newblk(struct newblk *); 930static void cancel_allocdirect(struct allocdirectlst *, 931 struct allocdirect *, struct freeblks *); 932static int check_inode_unwritten(struct inodedep *); 933static int free_inodedep(struct inodedep *); 934static void freework_freeblock(struct freework *); 935static void freework_enqueue(struct freework *); 936static int handle_workitem_freeblocks(struct freeblks *, int); 937static int handle_complete_freeblocks(struct freeblks *, int); 938static void handle_workitem_indirblk(struct freework *); 939static void handle_written_freework(struct freework *); 940static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 941static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 942 struct workhead *); 943static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 944 struct inodedep *, struct allocindir *, ufs_lbn_t); 945static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 946 ufs2_daddr_t, ufs_lbn_t); 947static void handle_workitem_freefrag(struct freefrag *); 948static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 949 ufs_lbn_t); 950static void allocdirect_merge(struct allocdirectlst *, 951 struct allocdirect *, struct allocdirect *); 952static struct freefrag *allocindir_merge(struct allocindir *, 953 struct allocindir *); 954static int bmsafemap_find(struct bmsafemap_hashhead *, struct mount *, int, 955 struct bmsafemap **); 956static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 957 int cg, struct bmsafemap *); 958static int newblk_find(struct newblk_hashhead *, struct mount *, ufs2_daddr_t, 959 int, struct newblk **); 960static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 961static int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t, 962 struct inodedep **); 963static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 964static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 965 int, struct pagedep **); 966static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 967 struct mount *mp, int, struct pagedep **); 968static void pause_timer(void *); 969static int request_cleanup(struct mount *, int); 970static int process_worklist_item(struct mount *, int, int); 971static void process_removes(struct vnode *); 972static void process_truncates(struct vnode *); 973static void jwork_move(struct workhead *, struct workhead *); 974static void jwork_insert(struct workhead *, struct jsegdep *); 975static void add_to_worklist(struct worklist *, int); 976static void wake_worklist(struct worklist *); 977static void wait_worklist(struct worklist *, char *); 978static void remove_from_worklist(struct worklist *); 979static void softdep_flush(void); 980static void softdep_flushjournal(struct mount *); 981static int softdep_speedup(void); 982static void worklist_speedup(void); 983static int journal_mount(struct mount *, struct fs *, struct ucred *); 984static void journal_unmount(struct mount *); 985static int journal_space(struct ufsmount *, int); 986static void journal_suspend(struct ufsmount *); 987static int journal_unsuspend(struct ufsmount *ump); 988static void softdep_prelink(struct vnode *, struct vnode *); 989static void add_to_journal(struct worklist *); 990static void remove_from_journal(struct worklist *); 991static void softdep_process_journal(struct mount *, struct worklist *, int); 992static struct jremref *newjremref(struct dirrem *, struct inode *, 993 struct inode *ip, off_t, nlink_t); 994static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 995 uint16_t); 996static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 997 uint16_t); 998static inline struct jsegdep *inoref_jseg(struct inoref *); 999static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 1000static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 1001 ufs2_daddr_t, int); 1002static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 1003static void move_newblock_dep(struct jaddref *, struct inodedep *); 1004static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 1005static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 1006 ufs2_daddr_t, long, ufs_lbn_t); 1007static struct freework *newfreework(struct ufsmount *, struct freeblks *, 1008 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 1009static int jwait(struct worklist *, int); 1010static struct inodedep *inodedep_lookup_ip(struct inode *); 1011static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 1012static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 1013static void handle_jwork(struct workhead *); 1014static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 1015 struct mkdir **); 1016static struct jblocks *jblocks_create(void); 1017static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 1018static void jblocks_free(struct jblocks *, struct mount *, int); 1019static void jblocks_destroy(struct jblocks *); 1020static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 1021 1022/* 1023 * Exported softdep operations. 1024 */ 1025static void softdep_disk_io_initiation(struct buf *); 1026static void softdep_disk_write_complete(struct buf *); 1027static void softdep_deallocate_dependencies(struct buf *); 1028static int softdep_count_dependencies(struct buf *bp, int); 1029 1030static struct mtx lk; 1031MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF); 1032 1033#define TRY_ACQUIRE_LOCK(lk) mtx_trylock(lk) 1034#define ACQUIRE_LOCK(lk) mtx_lock(lk) 1035#define FREE_LOCK(lk) mtx_unlock(lk) 1036 1037#define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 1038#define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 1039 1040/* 1041 * Worklist queue management. 1042 * These routines require that the lock be held. 1043 */ 1044#ifndef /* NOT */ DEBUG 1045#define WORKLIST_INSERT(head, item) do { \ 1046 (item)->wk_state |= ONWORKLIST; \ 1047 LIST_INSERT_HEAD(head, item, wk_list); \ 1048} while (0) 1049#define WORKLIST_REMOVE(item) do { \ 1050 (item)->wk_state &= ~ONWORKLIST; \ 1051 LIST_REMOVE(item, wk_list); \ 1052} while (0) 1053#define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1054#define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1055 1056#else /* DEBUG */ 1057static void worklist_insert(struct workhead *, struct worklist *, int); 1058static void worklist_remove(struct worklist *, int); 1059 1060#define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1061#define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1062#define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1063#define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1064 1065static void 1066worklist_insert(head, item, locked) 1067 struct workhead *head; 1068 struct worklist *item; 1069 int locked; 1070{ 1071 1072 if (locked) 1073 mtx_assert(&lk, MA_OWNED); 1074 if (item->wk_state & ONWORKLIST) 1075 panic("worklist_insert: %p %s(0x%X) already on list", 1076 item, TYPENAME(item->wk_type), item->wk_state); 1077 item->wk_state |= ONWORKLIST; 1078 LIST_INSERT_HEAD(head, item, wk_list); 1079} 1080 1081static void 1082worklist_remove(item, locked) 1083 struct worklist *item; 1084 int locked; 1085{ 1086 1087 if (locked) 1088 mtx_assert(&lk, MA_OWNED); 1089 if ((item->wk_state & ONWORKLIST) == 0) 1090 panic("worklist_remove: %p %s(0x%X) not on list", 1091 item, TYPENAME(item->wk_type), item->wk_state); 1092 item->wk_state &= ~ONWORKLIST; 1093 LIST_REMOVE(item, wk_list); 1094} 1095#endif /* DEBUG */ 1096 1097/* 1098 * Merge two jsegdeps keeping only the oldest one as newer references 1099 * can't be discarded until after older references. 1100 */ 1101static inline struct jsegdep * 1102jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1103{ 1104 struct jsegdep *swp; 1105 1106 if (two == NULL) 1107 return (one); 1108 1109 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1110 swp = one; 1111 one = two; 1112 two = swp; 1113 } 1114 WORKLIST_REMOVE(&two->jd_list); 1115 free_jsegdep(two); 1116 1117 return (one); 1118} 1119 1120/* 1121 * If two freedeps are compatible free one to reduce list size. 1122 */ 1123static inline struct freedep * 1124freedep_merge(struct freedep *one, struct freedep *two) 1125{ 1126 if (two == NULL) 1127 return (one); 1128 1129 if (one->fd_freework == two->fd_freework) { 1130 WORKLIST_REMOVE(&two->fd_list); 1131 free_freedep(two); 1132 } 1133 return (one); 1134} 1135 1136/* 1137 * Move journal work from one list to another. Duplicate freedeps and 1138 * jsegdeps are coalesced to keep the lists as small as possible. 1139 */ 1140static void 1141jwork_move(dst, src) 1142 struct workhead *dst; 1143 struct workhead *src; 1144{ 1145 struct freedep *freedep; 1146 struct jsegdep *jsegdep; 1147 struct worklist *wkn; 1148 struct worklist *wk; 1149 1150 KASSERT(dst != src, 1151 ("jwork_move: dst == src")); 1152 freedep = NULL; 1153 jsegdep = NULL; 1154 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1155 if (wk->wk_type == D_JSEGDEP) 1156 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1157 if (wk->wk_type == D_FREEDEP) 1158 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1159 } 1160 1161 mtx_assert(&lk, MA_OWNED); 1162 while ((wk = LIST_FIRST(src)) != NULL) { 1163 WORKLIST_REMOVE(wk); 1164 WORKLIST_INSERT(dst, wk); 1165 if (wk->wk_type == D_JSEGDEP) { 1166 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1167 continue; 1168 } 1169 if (wk->wk_type == D_FREEDEP) 1170 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1171 } 1172} 1173 1174static void 1175jwork_insert(dst, jsegdep) 1176 struct workhead *dst; 1177 struct jsegdep *jsegdep; 1178{ 1179 struct jsegdep *jsegdepn; 1180 struct worklist *wk; 1181 1182 LIST_FOREACH(wk, dst, wk_list) 1183 if (wk->wk_type == D_JSEGDEP) 1184 break; 1185 if (wk == NULL) { 1186 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1187 return; 1188 } 1189 jsegdepn = WK_JSEGDEP(wk); 1190 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1191 WORKLIST_REMOVE(wk); 1192 free_jsegdep(jsegdepn); 1193 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1194 } else 1195 free_jsegdep(jsegdep); 1196} 1197 1198/* 1199 * Routines for tracking and managing workitems. 1200 */ 1201static void workitem_free(struct worklist *, int); 1202static void workitem_alloc(struct worklist *, int, struct mount *); 1203 1204#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type)) 1205 1206static void 1207workitem_free(item, type) 1208 struct worklist *item; 1209 int type; 1210{ 1211 struct ufsmount *ump; 1212 mtx_assert(&lk, MA_OWNED); 1213 1214#ifdef DEBUG 1215 if (item->wk_state & ONWORKLIST) 1216 panic("workitem_free: %s(0x%X) still on list", 1217 TYPENAME(item->wk_type), item->wk_state); 1218 if (item->wk_type != type) 1219 panic("workitem_free: type mismatch %s != %s", 1220 TYPENAME(item->wk_type), TYPENAME(type)); 1221#endif 1222 if (item->wk_state & IOWAITING) 1223 wakeup(item); 1224 ump = VFSTOUFS(item->wk_mp); 1225 if (--ump->softdep_deps == 0 && ump->softdep_req) 1226 wakeup(&ump->softdep_deps); 1227 dep_current[type]--; 1228 free(item, DtoM(type)); 1229} 1230 1231static void 1232workitem_alloc(item, type, mp) 1233 struct worklist *item; 1234 int type; 1235 struct mount *mp; 1236{ 1237 struct ufsmount *ump; 1238 1239 item->wk_type = type; 1240 item->wk_mp = mp; 1241 item->wk_state = 0; 1242 1243 ump = VFSTOUFS(mp); 1244 ACQUIRE_LOCK(&lk); 1245 dep_current[type]++; 1246 dep_total[type]++; 1247 ump->softdep_deps++; 1248 ump->softdep_accdeps++; 1249 FREE_LOCK(&lk); 1250} 1251 1252/* 1253 * Workitem queue management 1254 */ 1255static int max_softdeps; /* maximum number of structs before slowdown */ 1256static int maxindirdeps = 50; /* max number of indirdeps before slowdown */ 1257static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1258static int proc_waiting; /* tracks whether we have a timeout posted */ 1259static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1260static struct callout softdep_callout; 1261static int req_pending; 1262static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1263static int req_clear_remove; /* syncer process flush some freeblks */ 1264static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1265 1266/* 1267 * runtime statistics 1268 */ 1269static int stat_worklist_push; /* number of worklist cleanups */ 1270static int stat_blk_limit_push; /* number of times block limit neared */ 1271static int stat_ino_limit_push; /* number of times inode limit neared */ 1272static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1273static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1274static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1275static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1276static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1277static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1278static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1279static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1280static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1281static int stat_journal_min; /* Times hit journal min threshold */ 1282static int stat_journal_low; /* Times hit journal low threshold */ 1283static int stat_journal_wait; /* Times blocked in jwait(). */ 1284static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1285static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1286static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1287static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1288static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1289static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1290static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1291static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1292static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1293 1294SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1295 &max_softdeps, 0, ""); 1296SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1297 &tickdelay, 0, ""); 1298SYSCTL_INT(_debug_softdep, OID_AUTO, maxindirdeps, CTLFLAG_RW, 1299 &maxindirdeps, 0, ""); 1300SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1301 &stat_worklist_push, 0,""); 1302SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1303 &stat_blk_limit_push, 0,""); 1304SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1305 &stat_ino_limit_push, 0,""); 1306SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1307 &stat_blk_limit_hit, 0, ""); 1308SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1309 &stat_ino_limit_hit, 0, ""); 1310SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1311 &stat_sync_limit_hit, 0, ""); 1312SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1313 &stat_indir_blk_ptrs, 0, ""); 1314SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1315 &stat_inode_bitmap, 0, ""); 1316SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1317 &stat_direct_blk_ptrs, 0, ""); 1318SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1319 &stat_dir_entry, 0, ""); 1320SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1321 &stat_jaddref, 0, ""); 1322SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1323 &stat_jnewblk, 0, ""); 1324SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1325 &stat_journal_low, 0, ""); 1326SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1327 &stat_journal_min, 0, ""); 1328SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1329 &stat_journal_wait, 0, ""); 1330SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1331 &stat_jwait_filepage, 0, ""); 1332SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1333 &stat_jwait_freeblks, 0, ""); 1334SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1335 &stat_jwait_inode, 0, ""); 1336SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1337 &stat_jwait_newblk, 0, ""); 1338SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1339 &stat_cleanup_blkrequests, 0, ""); 1340SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1341 &stat_cleanup_inorequests, 0, ""); 1342SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1343 &stat_cleanup_high_delay, 0, ""); 1344SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1345 &stat_cleanup_retries, 0, ""); 1346SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1347 &stat_cleanup_failures, 0, ""); 1348SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1349 &softdep_flushcache, 0, ""); 1350 1351SYSCTL_DECL(_vfs_ffs); 1352 1353LIST_HEAD(bmsafemap_hashhead, bmsafemap) *bmsafemap_hashtbl; 1354static u_long bmsafemap_hash; /* size of hash table - 1 */ 1355 1356static int compute_summary_at_mount = 0; /* Whether to recompute the summary at mount time */ 1357SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1358 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1359 1360static struct proc *softdepproc; 1361static struct kproc_desc softdep_kp = { 1362 "softdepflush", 1363 softdep_flush, 1364 &softdepproc 1365}; 1366SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, 1367 &softdep_kp); 1368 1369static void 1370softdep_flush(void) 1371{ 1372 struct mount *nmp; 1373 struct mount *mp; 1374 struct ufsmount *ump; 1375 struct thread *td; 1376 int remaining; 1377 int progress; 1378 1379 td = curthread; 1380 td->td_pflags |= TDP_NORUNNINGBUF; 1381 1382 for (;;) { 1383 kproc_suspend_check(softdepproc); 1384 ACQUIRE_LOCK(&lk); 1385 /* 1386 * If requested, try removing inode or removal dependencies. 1387 */ 1388 if (req_clear_inodedeps) { 1389 clear_inodedeps(); 1390 req_clear_inodedeps -= 1; 1391 wakeup_one(&proc_waiting); 1392 } 1393 if (req_clear_remove) { 1394 clear_remove(); 1395 req_clear_remove -= 1; 1396 wakeup_one(&proc_waiting); 1397 } 1398 FREE_LOCK(&lk); 1399 remaining = progress = 0; 1400 mtx_lock(&mountlist_mtx); 1401 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) { 1402 nmp = TAILQ_NEXT(mp, mnt_list); 1403 if (MOUNTEDSOFTDEP(mp) == 0) 1404 continue; 1405 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) 1406 continue; 1407 progress += softdep_process_worklist(mp, 0); 1408 ump = VFSTOUFS(mp); 1409 remaining += ump->softdep_on_worklist; 1410 mtx_lock(&mountlist_mtx); 1411 nmp = TAILQ_NEXT(mp, mnt_list); 1412 vfs_unbusy(mp); 1413 } 1414 mtx_unlock(&mountlist_mtx); 1415 if (remaining && progress) 1416 continue; 1417 ACQUIRE_LOCK(&lk); 1418 if (!req_pending) 1419 msleep(&req_pending, &lk, PVM, "sdflush", hz); 1420 req_pending = 0; 1421 FREE_LOCK(&lk); 1422 } 1423} 1424 1425static void 1426worklist_speedup(void) 1427{ 1428 mtx_assert(&lk, MA_OWNED); 1429 if (req_pending == 0) { 1430 req_pending = 1; 1431 wakeup(&req_pending); 1432 } 1433} 1434 1435static int 1436softdep_speedup(void) 1437{ 1438 1439 worklist_speedup(); 1440 bd_speedup(); 1441 return speedup_syncer(); 1442} 1443 1444/* 1445 * Add an item to the end of the work queue. 1446 * This routine requires that the lock be held. 1447 * This is the only routine that adds items to the list. 1448 * The following routine is the only one that removes items 1449 * and does so in order from first to last. 1450 */ 1451 1452#define WK_HEAD 0x0001 /* Add to HEAD. */ 1453#define WK_NODELAY 0x0002 /* Process immediately. */ 1454 1455static void 1456add_to_worklist(wk, flags) 1457 struct worklist *wk; 1458 int flags; 1459{ 1460 struct ufsmount *ump; 1461 1462 mtx_assert(&lk, MA_OWNED); 1463 ump = VFSTOUFS(wk->wk_mp); 1464 if (wk->wk_state & ONWORKLIST) 1465 panic("add_to_worklist: %s(0x%X) already on list", 1466 TYPENAME(wk->wk_type), wk->wk_state); 1467 wk->wk_state |= ONWORKLIST; 1468 if (ump->softdep_on_worklist == 0) { 1469 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1470 ump->softdep_worklist_tail = wk; 1471 } else if (flags & WK_HEAD) { 1472 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1473 } else { 1474 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1475 ump->softdep_worklist_tail = wk; 1476 } 1477 ump->softdep_on_worklist += 1; 1478 if (flags & WK_NODELAY) 1479 worklist_speedup(); 1480} 1481 1482/* 1483 * Remove the item to be processed. If we are removing the last 1484 * item on the list, we need to recalculate the tail pointer. 1485 */ 1486static void 1487remove_from_worklist(wk) 1488 struct worklist *wk; 1489{ 1490 struct ufsmount *ump; 1491 1492 ump = VFSTOUFS(wk->wk_mp); 1493 WORKLIST_REMOVE(wk); 1494 if (ump->softdep_worklist_tail == wk) 1495 ump->softdep_worklist_tail = 1496 (struct worklist *)wk->wk_list.le_prev; 1497 ump->softdep_on_worklist -= 1; 1498} 1499 1500static void 1501wake_worklist(wk) 1502 struct worklist *wk; 1503{ 1504 if (wk->wk_state & IOWAITING) { 1505 wk->wk_state &= ~IOWAITING; 1506 wakeup(wk); 1507 } 1508} 1509 1510static void 1511wait_worklist(wk, wmesg) 1512 struct worklist *wk; 1513 char *wmesg; 1514{ 1515 1516 wk->wk_state |= IOWAITING; 1517 msleep(wk, &lk, PVM, wmesg, 0); 1518} 1519 1520/* 1521 * Process that runs once per second to handle items in the background queue. 1522 * 1523 * Note that we ensure that everything is done in the order in which they 1524 * appear in the queue. The code below depends on this property to ensure 1525 * that blocks of a file are freed before the inode itself is freed. This 1526 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1527 * until all the old ones have been purged from the dependency lists. 1528 */ 1529int 1530softdep_process_worklist(mp, full) 1531 struct mount *mp; 1532 int full; 1533{ 1534 int cnt, matchcnt; 1535 struct ufsmount *ump; 1536 long starttime; 1537 1538 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1539 /* 1540 * Record the process identifier of our caller so that we can give 1541 * this process preferential treatment in request_cleanup below. 1542 */ 1543 matchcnt = 0; 1544 ump = VFSTOUFS(mp); 1545 ACQUIRE_LOCK(&lk); 1546 starttime = time_second; 1547 softdep_process_journal(mp, NULL, full?MNT_WAIT:0); 1548 while (ump->softdep_on_worklist > 0) { 1549 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1550 break; 1551 else 1552 matchcnt += cnt; 1553 /* 1554 * If requested, try removing inode or removal dependencies. 1555 */ 1556 if (req_clear_inodedeps) { 1557 clear_inodedeps(); 1558 req_clear_inodedeps -= 1; 1559 wakeup_one(&proc_waiting); 1560 } 1561 if (req_clear_remove) { 1562 clear_remove(); 1563 req_clear_remove -= 1; 1564 wakeup_one(&proc_waiting); 1565 } 1566 /* 1567 * We do not generally want to stop for buffer space, but if 1568 * we are really being a buffer hog, we will stop and wait. 1569 */ 1570 if (should_yield()) { 1571 FREE_LOCK(&lk); 1572 kern_yield(PRI_USER); 1573 bwillwrite(); 1574 ACQUIRE_LOCK(&lk); 1575 } 1576 /* 1577 * Never allow processing to run for more than one 1578 * second. Otherwise the other mountpoints may get 1579 * excessively backlogged. 1580 */ 1581 if (!full && starttime != time_second) 1582 break; 1583 } 1584 if (full == 0) 1585 journal_unsuspend(ump); 1586 FREE_LOCK(&lk); 1587 return (matchcnt); 1588} 1589 1590/* 1591 * Process all removes associated with a vnode if we are running out of 1592 * journal space. Any other process which attempts to flush these will 1593 * be unable as we have the vnodes locked. 1594 */ 1595static void 1596process_removes(vp) 1597 struct vnode *vp; 1598{ 1599 struct inodedep *inodedep; 1600 struct dirrem *dirrem; 1601 struct mount *mp; 1602 ino_t inum; 1603 1604 mtx_assert(&lk, MA_OWNED); 1605 1606 mp = vp->v_mount; 1607 inum = VTOI(vp)->i_number; 1608 for (;;) { 1609top: 1610 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1611 return; 1612 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1613 /* 1614 * If another thread is trying to lock this vnode 1615 * it will fail but we must wait for it to do so 1616 * before we can proceed. 1617 */ 1618 if (dirrem->dm_state & INPROGRESS) { 1619 wait_worklist(&dirrem->dm_list, "pwrwait"); 1620 goto top; 1621 } 1622 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1623 (COMPLETE | ONWORKLIST)) 1624 break; 1625 } 1626 if (dirrem == NULL) 1627 return; 1628 remove_from_worklist(&dirrem->dm_list); 1629 FREE_LOCK(&lk); 1630 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1631 panic("process_removes: suspended filesystem"); 1632 handle_workitem_remove(dirrem, 0); 1633 vn_finished_secondary_write(mp); 1634 ACQUIRE_LOCK(&lk); 1635 } 1636} 1637 1638/* 1639 * Process all truncations associated with a vnode if we are running out 1640 * of journal space. This is called when the vnode lock is already held 1641 * and no other process can clear the truncation. This function returns 1642 * a value greater than zero if it did any work. 1643 */ 1644static void 1645process_truncates(vp) 1646 struct vnode *vp; 1647{ 1648 struct inodedep *inodedep; 1649 struct freeblks *freeblks; 1650 struct mount *mp; 1651 ino_t inum; 1652 int cgwait; 1653 1654 mtx_assert(&lk, MA_OWNED); 1655 1656 mp = vp->v_mount; 1657 inum = VTOI(vp)->i_number; 1658 for (;;) { 1659 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1660 return; 1661 cgwait = 0; 1662 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1663 /* Journal entries not yet written. */ 1664 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1665 jwait(&LIST_FIRST( 1666 &freeblks->fb_jblkdephd)->jb_list, 1667 MNT_WAIT); 1668 break; 1669 } 1670 /* Another thread is executing this item. */ 1671 if (freeblks->fb_state & INPROGRESS) { 1672 wait_worklist(&freeblks->fb_list, "ptrwait"); 1673 break; 1674 } 1675 /* Freeblks is waiting on a inode write. */ 1676 if ((freeblks->fb_state & COMPLETE) == 0) { 1677 FREE_LOCK(&lk); 1678 ffs_update(vp, 1); 1679 ACQUIRE_LOCK(&lk); 1680 break; 1681 } 1682 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1683 (ALLCOMPLETE | ONWORKLIST)) { 1684 remove_from_worklist(&freeblks->fb_list); 1685 freeblks->fb_state |= INPROGRESS; 1686 FREE_LOCK(&lk); 1687 if (vn_start_secondary_write(NULL, &mp, 1688 V_NOWAIT)) 1689 panic("process_truncates: " 1690 "suspended filesystem"); 1691 handle_workitem_freeblocks(freeblks, 0); 1692 vn_finished_secondary_write(mp); 1693 ACQUIRE_LOCK(&lk); 1694 break; 1695 } 1696 if (freeblks->fb_cgwait) 1697 cgwait++; 1698 } 1699 if (cgwait) { 1700 FREE_LOCK(&lk); 1701 sync_cgs(mp, MNT_WAIT); 1702 ffs_sync_snap(mp, MNT_WAIT); 1703 ACQUIRE_LOCK(&lk); 1704 continue; 1705 } 1706 if (freeblks == NULL) 1707 break; 1708 } 1709 return; 1710} 1711 1712/* 1713 * Process one item on the worklist. 1714 */ 1715static int 1716process_worklist_item(mp, target, flags) 1717 struct mount *mp; 1718 int target; 1719 int flags; 1720{ 1721 struct worklist sintenel; 1722 struct worklist *wk; 1723 struct ufsmount *ump; 1724 int matchcnt; 1725 int error; 1726 1727 mtx_assert(&lk, MA_OWNED); 1728 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1729 /* 1730 * If we are being called because of a process doing a 1731 * copy-on-write, then it is not safe to write as we may 1732 * recurse into the copy-on-write routine. 1733 */ 1734 if (curthread->td_pflags & TDP_COWINPROGRESS) 1735 return (-1); 1736 PHOLD(curproc); /* Don't let the stack go away. */ 1737 ump = VFSTOUFS(mp); 1738 matchcnt = 0; 1739 sintenel.wk_mp = NULL; 1740 sintenel.wk_type = D_SENTINAL; 1741 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sintenel, wk_list); 1742 for (wk = LIST_NEXT(&sintenel, wk_list); wk != NULL; 1743 wk = LIST_NEXT(&sintenel, wk_list)) { 1744 if (wk->wk_type == D_SENTINAL) { 1745 LIST_REMOVE(&sintenel, wk_list); 1746 LIST_INSERT_AFTER(wk, &sintenel, wk_list); 1747 continue; 1748 } 1749 if (wk->wk_state & INPROGRESS) 1750 panic("process_worklist_item: %p already in progress.", 1751 wk); 1752 wk->wk_state |= INPROGRESS; 1753 remove_from_worklist(wk); 1754 FREE_LOCK(&lk); 1755 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1756 panic("process_worklist_item: suspended filesystem"); 1757 switch (wk->wk_type) { 1758 case D_DIRREM: 1759 /* removal of a directory entry */ 1760 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1761 break; 1762 1763 case D_FREEBLKS: 1764 /* releasing blocks and/or fragments from a file */ 1765 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1766 flags); 1767 break; 1768 1769 case D_FREEFRAG: 1770 /* releasing a fragment when replaced as a file grows */ 1771 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1772 error = 0; 1773 break; 1774 1775 case D_FREEFILE: 1776 /* releasing an inode when its link count drops to 0 */ 1777 handle_workitem_freefile(WK_FREEFILE(wk)); 1778 error = 0; 1779 break; 1780 1781 default: 1782 panic("%s_process_worklist: Unknown type %s", 1783 "softdep", TYPENAME(wk->wk_type)); 1784 /* NOTREACHED */ 1785 } 1786 vn_finished_secondary_write(mp); 1787 ACQUIRE_LOCK(&lk); 1788 if (error == 0) { 1789 if (++matchcnt == target) 1790 break; 1791 continue; 1792 } 1793 /* 1794 * We have to retry the worklist item later. Wake up any 1795 * waiters who may be able to complete it immediately and 1796 * add the item back to the head so we don't try to execute 1797 * it again. 1798 */ 1799 wk->wk_state &= ~INPROGRESS; 1800 wake_worklist(wk); 1801 add_to_worklist(wk, WK_HEAD); 1802 } 1803 LIST_REMOVE(&sintenel, wk_list); 1804 /* Sentinal could've become the tail from remove_from_worklist. */ 1805 if (ump->softdep_worklist_tail == &sintenel) 1806 ump->softdep_worklist_tail = 1807 (struct worklist *)sintenel.wk_list.le_prev; 1808 PRELE(curproc); 1809 return (matchcnt); 1810} 1811 1812/* 1813 * Move dependencies from one buffer to another. 1814 */ 1815int 1816softdep_move_dependencies(oldbp, newbp) 1817 struct buf *oldbp; 1818 struct buf *newbp; 1819{ 1820 struct worklist *wk, *wktail; 1821 int dirty; 1822 1823 dirty = 0; 1824 wktail = NULL; 1825 ACQUIRE_LOCK(&lk); 1826 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1827 LIST_REMOVE(wk, wk_list); 1828 if (wk->wk_type == D_BMSAFEMAP && 1829 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1830 dirty = 1; 1831 if (wktail == 0) 1832 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1833 else 1834 LIST_INSERT_AFTER(wktail, wk, wk_list); 1835 wktail = wk; 1836 } 1837 FREE_LOCK(&lk); 1838 1839 return (dirty); 1840} 1841 1842/* 1843 * Purge the work list of all items associated with a particular mount point. 1844 */ 1845int 1846softdep_flushworklist(oldmnt, countp, td) 1847 struct mount *oldmnt; 1848 int *countp; 1849 struct thread *td; 1850{ 1851 struct vnode *devvp; 1852 int count, error = 0; 1853 struct ufsmount *ump; 1854 1855 /* 1856 * Alternately flush the block device associated with the mount 1857 * point and process any dependencies that the flushing 1858 * creates. We continue until no more worklist dependencies 1859 * are found. 1860 */ 1861 *countp = 0; 1862 ump = VFSTOUFS(oldmnt); 1863 devvp = ump->um_devvp; 1864 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1865 *countp += count; 1866 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1867 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1868 VOP_UNLOCK(devvp, 0); 1869 if (error) 1870 break; 1871 } 1872 return (error); 1873} 1874 1875int 1876softdep_waitidle(struct mount *mp) 1877{ 1878 struct ufsmount *ump; 1879 int error; 1880 int i; 1881 1882 ump = VFSTOUFS(mp); 1883 ACQUIRE_LOCK(&lk); 1884 for (i = 0; i < 10 && ump->softdep_deps; i++) { 1885 ump->softdep_req = 1; 1886 if (ump->softdep_on_worklist) 1887 panic("softdep_waitidle: work added after flush."); 1888 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1); 1889 } 1890 ump->softdep_req = 0; 1891 FREE_LOCK(&lk); 1892 error = 0; 1893 if (i == 10) { 1894 error = EBUSY; 1895 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1896 mp); 1897 } 1898 1899 return (error); 1900} 1901 1902/* 1903 * Flush all vnodes and worklist items associated with a specified mount point. 1904 */ 1905int 1906softdep_flushfiles(oldmnt, flags, td) 1907 struct mount *oldmnt; 1908 int flags; 1909 struct thread *td; 1910{ 1911#ifdef QUOTA 1912 struct ufsmount *ump; 1913 int i; 1914#endif 1915 int error, early, depcount, loopcnt, retry_flush_count, retry; 1916 int morework; 1917 1918 loopcnt = 10; 1919 retry_flush_count = 3; 1920retry_flush: 1921 error = 0; 1922 1923 /* 1924 * Alternately flush the vnodes associated with the mount 1925 * point and process any dependencies that the flushing 1926 * creates. In theory, this loop can happen at most twice, 1927 * but we give it a few extra just to be sure. 1928 */ 1929 for (; loopcnt > 0; loopcnt--) { 1930 /* 1931 * Do another flush in case any vnodes were brought in 1932 * as part of the cleanup operations. 1933 */ 1934 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1935 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1936 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1937 break; 1938 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1939 depcount == 0) 1940 break; 1941 } 1942 /* 1943 * If we are unmounting then it is an error to fail. If we 1944 * are simply trying to downgrade to read-only, then filesystem 1945 * activity can keep us busy forever, so we just fail with EBUSY. 1946 */ 1947 if (loopcnt == 0) { 1948 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 1949 panic("softdep_flushfiles: looping"); 1950 error = EBUSY; 1951 } 1952 if (!error) 1953 error = softdep_waitidle(oldmnt); 1954 if (!error) { 1955 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 1956 retry = 0; 1957 MNT_ILOCK(oldmnt); 1958 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 1959 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 1960 morework = oldmnt->mnt_nvnodelistsize > 0; 1961#ifdef QUOTA 1962 ump = VFSTOUFS(oldmnt); 1963 UFS_LOCK(ump); 1964 for (i = 0; i < MAXQUOTAS; i++) { 1965 if (ump->um_quotas[i] != NULLVP) 1966 morework = 1; 1967 } 1968 UFS_UNLOCK(ump); 1969#endif 1970 if (morework) { 1971 if (--retry_flush_count > 0) { 1972 retry = 1; 1973 loopcnt = 3; 1974 } else 1975 error = EBUSY; 1976 } 1977 MNT_IUNLOCK(oldmnt); 1978 if (retry) 1979 goto retry_flush; 1980 } 1981 } 1982 return (error); 1983} 1984 1985/* 1986 * Structure hashing. 1987 * 1988 * There are three types of structures that can be looked up: 1989 * 1) pagedep structures identified by mount point, inode number, 1990 * and logical block. 1991 * 2) inodedep structures identified by mount point and inode number. 1992 * 3) newblk structures identified by mount point and 1993 * physical block number. 1994 * 1995 * The "pagedep" and "inodedep" dependency structures are hashed 1996 * separately from the file blocks and inodes to which they correspond. 1997 * This separation helps when the in-memory copy of an inode or 1998 * file block must be replaced. It also obviates the need to access 1999 * an inode or file page when simply updating (or de-allocating) 2000 * dependency structures. Lookup of newblk structures is needed to 2001 * find newly allocated blocks when trying to associate them with 2002 * their allocdirect or allocindir structure. 2003 * 2004 * The lookup routines optionally create and hash a new instance when 2005 * an existing entry is not found. 2006 */ 2007#define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2008#define NODELAY 0x0002 /* cannot do background work */ 2009 2010/* 2011 * Structures and routines associated with pagedep caching. 2012 */ 2013LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl; 2014u_long pagedep_hash; /* size of hash table - 1 */ 2015#define PAGEDEP_HASH(mp, inum, lbn) \ 2016 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \ 2017 pagedep_hash]) 2018 2019static int 2020pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp) 2021 struct pagedep_hashhead *pagedephd; 2022 ino_t ino; 2023 ufs_lbn_t lbn; 2024 struct mount *mp; 2025 int flags; 2026 struct pagedep **pagedeppp; 2027{ 2028 struct pagedep *pagedep; 2029 2030 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2031 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn && 2032 mp == pagedep->pd_list.wk_mp) { 2033 *pagedeppp = pagedep; 2034 return (1); 2035 } 2036 } 2037 *pagedeppp = NULL; 2038 return (0); 2039} 2040/* 2041 * Look up a pagedep. Return 1 if found, 0 otherwise. 2042 * If not found, allocate if DEPALLOC flag is passed. 2043 * Found or allocated entry is returned in pagedeppp. 2044 * This routine must be called with splbio interrupts blocked. 2045 */ 2046static int 2047pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2048 struct mount *mp; 2049 struct buf *bp; 2050 ino_t ino; 2051 ufs_lbn_t lbn; 2052 int flags; 2053 struct pagedep **pagedeppp; 2054{ 2055 struct pagedep *pagedep; 2056 struct pagedep_hashhead *pagedephd; 2057 struct worklist *wk; 2058 int ret; 2059 int i; 2060 2061 mtx_assert(&lk, MA_OWNED); 2062 if (bp) { 2063 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2064 if (wk->wk_type == D_PAGEDEP) { 2065 *pagedeppp = WK_PAGEDEP(wk); 2066 return (1); 2067 } 2068 } 2069 } 2070 pagedephd = PAGEDEP_HASH(mp, ino, lbn); 2071 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 2072 if (ret) { 2073 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2074 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2075 return (1); 2076 } 2077 if ((flags & DEPALLOC) == 0) 2078 return (0); 2079 FREE_LOCK(&lk); 2080 pagedep = malloc(sizeof(struct pagedep), 2081 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2082 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2083 ACQUIRE_LOCK(&lk); 2084 ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp); 2085 if (*pagedeppp) { 2086 /* 2087 * This should never happen since we only create pagedeps 2088 * with the vnode lock held. Could be an assert. 2089 */ 2090 WORKITEM_FREE(pagedep, D_PAGEDEP); 2091 return (ret); 2092 } 2093 pagedep->pd_ino = ino; 2094 pagedep->pd_lbn = lbn; 2095 LIST_INIT(&pagedep->pd_dirremhd); 2096 LIST_INIT(&pagedep->pd_pendinghd); 2097 for (i = 0; i < DAHASHSZ; i++) 2098 LIST_INIT(&pagedep->pd_diraddhd[i]); 2099 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2100 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2101 *pagedeppp = pagedep; 2102 return (0); 2103} 2104 2105/* 2106 * Structures and routines associated with inodedep caching. 2107 */ 2108LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl; 2109static u_long inodedep_hash; /* size of hash table - 1 */ 2110#define INODEDEP_HASH(fs, inum) \ 2111 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash]) 2112 2113static int 2114inodedep_find(inodedephd, fs, inum, inodedeppp) 2115 struct inodedep_hashhead *inodedephd; 2116 struct fs *fs; 2117 ino_t inum; 2118 struct inodedep **inodedeppp; 2119{ 2120 struct inodedep *inodedep; 2121 2122 LIST_FOREACH(inodedep, inodedephd, id_hash) 2123 if (inum == inodedep->id_ino && fs == inodedep->id_fs) 2124 break; 2125 if (inodedep) { 2126 *inodedeppp = inodedep; 2127 return (1); 2128 } 2129 *inodedeppp = NULL; 2130 2131 return (0); 2132} 2133/* 2134 * Look up an inodedep. Return 1 if found, 0 if not found. 2135 * If not found, allocate if DEPALLOC flag is passed. 2136 * Found or allocated entry is returned in inodedeppp. 2137 * This routine must be called with splbio interrupts blocked. 2138 */ 2139static int 2140inodedep_lookup(mp, inum, flags, inodedeppp) 2141 struct mount *mp; 2142 ino_t inum; 2143 int flags; 2144 struct inodedep **inodedeppp; 2145{ 2146 struct inodedep *inodedep; 2147 struct inodedep_hashhead *inodedephd; 2148 struct fs *fs; 2149 2150 mtx_assert(&lk, MA_OWNED); 2151 fs = VFSTOUFS(mp)->um_fs; 2152 inodedephd = INODEDEP_HASH(fs, inum); 2153 2154 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) 2155 return (1); 2156 if ((flags & DEPALLOC) == 0) 2157 return (0); 2158 /* 2159 * If we are over our limit, try to improve the situation. 2160 */ 2161 if (dep_current[D_INODEDEP] > max_softdeps && (flags & NODELAY) == 0) 2162 request_cleanup(mp, FLUSH_INODES); 2163 FREE_LOCK(&lk); 2164 inodedep = malloc(sizeof(struct inodedep), 2165 M_INODEDEP, M_SOFTDEP_FLAGS); 2166 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2167 ACQUIRE_LOCK(&lk); 2168 if (inodedep_find(inodedephd, fs, inum, inodedeppp)) { 2169 WORKITEM_FREE(inodedep, D_INODEDEP); 2170 return (1); 2171 } 2172 inodedep->id_fs = fs; 2173 inodedep->id_ino = inum; 2174 inodedep->id_state = ALLCOMPLETE; 2175 inodedep->id_nlinkdelta = 0; 2176 inodedep->id_savedino1 = NULL; 2177 inodedep->id_savedsize = -1; 2178 inodedep->id_savedextsize = -1; 2179 inodedep->id_savednlink = -1; 2180 inodedep->id_bmsafemap = NULL; 2181 inodedep->id_mkdiradd = NULL; 2182 LIST_INIT(&inodedep->id_dirremhd); 2183 LIST_INIT(&inodedep->id_pendinghd); 2184 LIST_INIT(&inodedep->id_inowait); 2185 LIST_INIT(&inodedep->id_bufwait); 2186 TAILQ_INIT(&inodedep->id_inoreflst); 2187 TAILQ_INIT(&inodedep->id_inoupdt); 2188 TAILQ_INIT(&inodedep->id_newinoupdt); 2189 TAILQ_INIT(&inodedep->id_extupdt); 2190 TAILQ_INIT(&inodedep->id_newextupdt); 2191 TAILQ_INIT(&inodedep->id_freeblklst); 2192 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2193 *inodedeppp = inodedep; 2194 return (0); 2195} 2196 2197/* 2198 * Structures and routines associated with newblk caching. 2199 */ 2200LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl; 2201u_long newblk_hash; /* size of hash table - 1 */ 2202#define NEWBLK_HASH(fs, inum) \ 2203 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash]) 2204 2205static int 2206newblk_find(newblkhd, mp, newblkno, flags, newblkpp) 2207 struct newblk_hashhead *newblkhd; 2208 struct mount *mp; 2209 ufs2_daddr_t newblkno; 2210 int flags; 2211 struct newblk **newblkpp; 2212{ 2213 struct newblk *newblk; 2214 2215 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2216 if (newblkno != newblk->nb_newblkno) 2217 continue; 2218 if (mp != newblk->nb_list.wk_mp) 2219 continue; 2220 /* 2221 * If we're creating a new dependency don't match those that 2222 * have already been converted to allocdirects. This is for 2223 * a frag extend. 2224 */ 2225 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2226 continue; 2227 break; 2228 } 2229 if (newblk) { 2230 *newblkpp = newblk; 2231 return (1); 2232 } 2233 *newblkpp = NULL; 2234 return (0); 2235} 2236 2237/* 2238 * Look up a newblk. Return 1 if found, 0 if not found. 2239 * If not found, allocate if DEPALLOC flag is passed. 2240 * Found or allocated entry is returned in newblkpp. 2241 */ 2242static int 2243newblk_lookup(mp, newblkno, flags, newblkpp) 2244 struct mount *mp; 2245 ufs2_daddr_t newblkno; 2246 int flags; 2247 struct newblk **newblkpp; 2248{ 2249 struct newblk *newblk; 2250 struct newblk_hashhead *newblkhd; 2251 2252 newblkhd = NEWBLK_HASH(VFSTOUFS(mp)->um_fs, newblkno); 2253 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) 2254 return (1); 2255 if ((flags & DEPALLOC) == 0) 2256 return (0); 2257 FREE_LOCK(&lk); 2258 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2259 M_SOFTDEP_FLAGS | M_ZERO); 2260 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2261 ACQUIRE_LOCK(&lk); 2262 if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) { 2263 WORKITEM_FREE(newblk, D_NEWBLK); 2264 return (1); 2265 } 2266 newblk->nb_freefrag = NULL; 2267 LIST_INIT(&newblk->nb_indirdeps); 2268 LIST_INIT(&newblk->nb_newdirblk); 2269 LIST_INIT(&newblk->nb_jwork); 2270 newblk->nb_state = ATTACHED; 2271 newblk->nb_newblkno = newblkno; 2272 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2273 *newblkpp = newblk; 2274 return (0); 2275} 2276 2277/* 2278 * Structures and routines associated with freed indirect block caching. 2279 */ 2280struct freeworklst *indir_hashtbl; 2281u_long indir_hash; /* size of hash table - 1 */ 2282#define INDIR_HASH(mp, blkno) \ 2283 (&indir_hashtbl[((((register_t)(mp)) >> 13) + (blkno)) & indir_hash]) 2284 2285/* 2286 * Lookup an indirect block in the indir hash table. The freework is 2287 * removed and potentially freed. The caller must do a blocking journal 2288 * write before writing to the blkno. 2289 */ 2290static int 2291indirblk_lookup(mp, blkno) 2292 struct mount *mp; 2293 ufs2_daddr_t blkno; 2294{ 2295 struct freework *freework; 2296 struct freeworklst *wkhd; 2297 2298 wkhd = INDIR_HASH(mp, blkno); 2299 TAILQ_FOREACH(freework, wkhd, fw_next) { 2300 if (freework->fw_blkno != blkno) 2301 continue; 2302 if (freework->fw_list.wk_mp != mp) 2303 continue; 2304 indirblk_remove(freework); 2305 return (1); 2306 } 2307 return (0); 2308} 2309 2310/* 2311 * Insert an indirect block represented by freework into the indirblk 2312 * hash table so that it may prevent the block from being re-used prior 2313 * to the journal being written. 2314 */ 2315static void 2316indirblk_insert(freework) 2317 struct freework *freework; 2318{ 2319 struct jblocks *jblocks; 2320 struct jseg *jseg; 2321 2322 jblocks = VFSTOUFS(freework->fw_list.wk_mp)->softdep_jblocks; 2323 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2324 if (jseg == NULL) 2325 return; 2326 2327 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2328 TAILQ_INSERT_HEAD(INDIR_HASH(freework->fw_list.wk_mp, 2329 freework->fw_blkno), freework, fw_next); 2330 freework->fw_state &= ~DEPCOMPLETE; 2331} 2332 2333static void 2334indirblk_remove(freework) 2335 struct freework *freework; 2336{ 2337 2338 LIST_REMOVE(freework, fw_segs); 2339 TAILQ_REMOVE(INDIR_HASH(freework->fw_list.wk_mp, 2340 freework->fw_blkno), freework, fw_next); 2341 freework->fw_state |= DEPCOMPLETE; 2342 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2343 WORKITEM_FREE(freework, D_FREEWORK); 2344} 2345 2346/* 2347 * Executed during filesystem system initialization before 2348 * mounting any filesystems. 2349 */ 2350void 2351softdep_initialize() 2352{ 2353 int i; 2354 2355 LIST_INIT(&mkdirlisthd); 2356 max_softdeps = desiredvnodes * 4; 2357 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash); 2358 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash); 2359 newblk_hashtbl = hashinit(desiredvnodes / 5, M_NEWBLK, &newblk_hash); 2360 bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, &bmsafemap_hash); 2361 i = 1 << (ffs(desiredvnodes / 10) - 1); 2362 indir_hashtbl = malloc(i * sizeof(indir_hashtbl[0]), M_FREEWORK, 2363 M_WAITOK); 2364 indir_hash = i - 1; 2365 for (i = 0; i <= indir_hash; i++) 2366 TAILQ_INIT(&indir_hashtbl[i]); 2367 2368 /* initialise bioops hack */ 2369 bioops.io_start = softdep_disk_io_initiation; 2370 bioops.io_complete = softdep_disk_write_complete; 2371 bioops.io_deallocate = softdep_deallocate_dependencies; 2372 bioops.io_countdeps = softdep_count_dependencies; 2373 2374 /* Initialize the callout with an mtx. */ 2375 callout_init_mtx(&softdep_callout, &lk, 0); 2376} 2377 2378/* 2379 * Executed after all filesystems have been unmounted during 2380 * filesystem module unload. 2381 */ 2382void 2383softdep_uninitialize() 2384{ 2385 2386 callout_drain(&softdep_callout); 2387 hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash); 2388 hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash); 2389 hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash); 2390 hashdestroy(bmsafemap_hashtbl, M_BMSAFEMAP, bmsafemap_hash); 2391 free(indir_hashtbl, M_FREEWORK); 2392} 2393 2394/* 2395 * Called at mount time to notify the dependency code that a 2396 * filesystem wishes to use it. 2397 */ 2398int 2399softdep_mount(devvp, mp, fs, cred) 2400 struct vnode *devvp; 2401 struct mount *mp; 2402 struct fs *fs; 2403 struct ucred *cred; 2404{ 2405 struct csum_total cstotal; 2406 struct ufsmount *ump; 2407 struct cg *cgp; 2408 struct buf *bp; 2409 int error, cyl; 2410 2411 MNT_ILOCK(mp); 2412 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2413 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2414 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2415 MNTK_SOFTDEP | MNTK_NOASYNC; 2416 } 2417 MNT_IUNLOCK(mp); 2418 ump = VFSTOUFS(mp); 2419 LIST_INIT(&ump->softdep_workitem_pending); 2420 LIST_INIT(&ump->softdep_journal_pending); 2421 TAILQ_INIT(&ump->softdep_unlinked); 2422 LIST_INIT(&ump->softdep_dirtycg); 2423 ump->softdep_worklist_tail = NULL; 2424 ump->softdep_on_worklist = 0; 2425 ump->softdep_deps = 0; 2426 if ((fs->fs_flags & FS_SUJ) && 2427 (error = journal_mount(mp, fs, cred)) != 0) { 2428 printf("Failed to start journal: %d\n", error); 2429 return (error); 2430 } 2431 /* 2432 * When doing soft updates, the counters in the 2433 * superblock may have gotten out of sync. Recomputation 2434 * can take a long time and can be deferred for background 2435 * fsck. However, the old behavior of scanning the cylinder 2436 * groups and recalculating them at mount time is available 2437 * by setting vfs.ffs.compute_summary_at_mount to one. 2438 */ 2439 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2440 return (0); 2441 bzero(&cstotal, sizeof cstotal); 2442 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2443 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2444 fs->fs_cgsize, cred, &bp)) != 0) { 2445 brelse(bp); 2446 return (error); 2447 } 2448 cgp = (struct cg *)bp->b_data; 2449 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2450 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2451 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2452 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2453 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2454 brelse(bp); 2455 } 2456#ifdef DEBUG 2457 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2458 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2459#endif 2460 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2461 return (0); 2462} 2463 2464void 2465softdep_unmount(mp) 2466 struct mount *mp; 2467{ 2468 2469 MNT_ILOCK(mp); 2470 mp->mnt_flag &= ~MNT_SOFTDEP; 2471 if (MOUNTEDSUJ(mp) == 0) { 2472 MNT_IUNLOCK(mp); 2473 return; 2474 } 2475 mp->mnt_flag &= ~MNT_SUJ; 2476 MNT_IUNLOCK(mp); 2477 journal_unmount(mp); 2478} 2479 2480static struct jblocks * 2481jblocks_create(void) 2482{ 2483 struct jblocks *jblocks; 2484 2485 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2486 TAILQ_INIT(&jblocks->jb_segs); 2487 jblocks->jb_avail = 10; 2488 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2489 M_JBLOCKS, M_WAITOK | M_ZERO); 2490 2491 return (jblocks); 2492} 2493 2494static ufs2_daddr_t 2495jblocks_alloc(jblocks, bytes, actual) 2496 struct jblocks *jblocks; 2497 int bytes; 2498 int *actual; 2499{ 2500 ufs2_daddr_t daddr; 2501 struct jextent *jext; 2502 int freecnt; 2503 int blocks; 2504 2505 blocks = bytes / DEV_BSIZE; 2506 jext = &jblocks->jb_extent[jblocks->jb_head]; 2507 freecnt = jext->je_blocks - jblocks->jb_off; 2508 if (freecnt == 0) { 2509 jblocks->jb_off = 0; 2510 if (++jblocks->jb_head > jblocks->jb_used) 2511 jblocks->jb_head = 0; 2512 jext = &jblocks->jb_extent[jblocks->jb_head]; 2513 freecnt = jext->je_blocks; 2514 } 2515 if (freecnt > blocks) 2516 freecnt = blocks; 2517 *actual = freecnt * DEV_BSIZE; 2518 daddr = jext->je_daddr + jblocks->jb_off; 2519 jblocks->jb_off += freecnt; 2520 jblocks->jb_free -= freecnt; 2521 2522 return (daddr); 2523} 2524 2525static void 2526jblocks_free(jblocks, mp, bytes) 2527 struct jblocks *jblocks; 2528 struct mount *mp; 2529 int bytes; 2530{ 2531 2532 jblocks->jb_free += bytes / DEV_BSIZE; 2533 if (jblocks->jb_suspended) 2534 worklist_speedup(); 2535 wakeup(jblocks); 2536} 2537 2538static void 2539jblocks_destroy(jblocks) 2540 struct jblocks *jblocks; 2541{ 2542 2543 if (jblocks->jb_extent) 2544 free(jblocks->jb_extent, M_JBLOCKS); 2545 free(jblocks, M_JBLOCKS); 2546} 2547 2548static void 2549jblocks_add(jblocks, daddr, blocks) 2550 struct jblocks *jblocks; 2551 ufs2_daddr_t daddr; 2552 int blocks; 2553{ 2554 struct jextent *jext; 2555 2556 jblocks->jb_blocks += blocks; 2557 jblocks->jb_free += blocks; 2558 jext = &jblocks->jb_extent[jblocks->jb_used]; 2559 /* Adding the first block. */ 2560 if (jext->je_daddr == 0) { 2561 jext->je_daddr = daddr; 2562 jext->je_blocks = blocks; 2563 return; 2564 } 2565 /* Extending the last extent. */ 2566 if (jext->je_daddr + jext->je_blocks == daddr) { 2567 jext->je_blocks += blocks; 2568 return; 2569 } 2570 /* Adding a new extent. */ 2571 if (++jblocks->jb_used == jblocks->jb_avail) { 2572 jblocks->jb_avail *= 2; 2573 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2574 M_JBLOCKS, M_WAITOK | M_ZERO); 2575 memcpy(jext, jblocks->jb_extent, 2576 sizeof(struct jextent) * jblocks->jb_used); 2577 free(jblocks->jb_extent, M_JBLOCKS); 2578 jblocks->jb_extent = jext; 2579 } 2580 jext = &jblocks->jb_extent[jblocks->jb_used]; 2581 jext->je_daddr = daddr; 2582 jext->je_blocks = blocks; 2583 return; 2584} 2585 2586int 2587softdep_journal_lookup(mp, vpp) 2588 struct mount *mp; 2589 struct vnode **vpp; 2590{ 2591 struct componentname cnp; 2592 struct vnode *dvp; 2593 ino_t sujournal; 2594 int error; 2595 2596 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2597 if (error) 2598 return (error); 2599 bzero(&cnp, sizeof(cnp)); 2600 cnp.cn_nameiop = LOOKUP; 2601 cnp.cn_flags = ISLASTCN; 2602 cnp.cn_thread = curthread; 2603 cnp.cn_cred = curthread->td_ucred; 2604 cnp.cn_pnbuf = SUJ_FILE; 2605 cnp.cn_nameptr = SUJ_FILE; 2606 cnp.cn_namelen = strlen(SUJ_FILE); 2607 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2608 vput(dvp); 2609 if (error != 0) 2610 return (error); 2611 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2612 return (error); 2613} 2614 2615/* 2616 * Open and verify the journal file. 2617 */ 2618static int 2619journal_mount(mp, fs, cred) 2620 struct mount *mp; 2621 struct fs *fs; 2622 struct ucred *cred; 2623{ 2624 struct jblocks *jblocks; 2625 struct vnode *vp; 2626 struct inode *ip; 2627 ufs2_daddr_t blkno; 2628 int bcount; 2629 int error; 2630 int i; 2631 2632 error = softdep_journal_lookup(mp, &vp); 2633 if (error != 0) { 2634 printf("Failed to find journal. Use tunefs to create one\n"); 2635 return (error); 2636 } 2637 ip = VTOI(vp); 2638 if (ip->i_size < SUJ_MIN) { 2639 error = ENOSPC; 2640 goto out; 2641 } 2642 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2643 jblocks = jblocks_create(); 2644 for (i = 0; i < bcount; i++) { 2645 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2646 if (error) 2647 break; 2648 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2649 } 2650 if (error) { 2651 jblocks_destroy(jblocks); 2652 goto out; 2653 } 2654 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2655 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2656 VFSTOUFS(mp)->softdep_jblocks = jblocks; 2657out: 2658 if (error == 0) { 2659 MNT_ILOCK(mp); 2660 mp->mnt_flag |= MNT_SUJ; 2661 mp->mnt_flag &= ~MNT_SOFTDEP; 2662 MNT_IUNLOCK(mp); 2663 /* 2664 * Only validate the journal contents if the 2665 * filesystem is clean, otherwise we write the logs 2666 * but they'll never be used. If the filesystem was 2667 * still dirty when we mounted it the journal is 2668 * invalid and a new journal can only be valid if it 2669 * starts from a clean mount. 2670 */ 2671 if (fs->fs_clean) { 2672 DIP_SET(ip, i_modrev, fs->fs_mtime); 2673 ip->i_flags |= IN_MODIFIED; 2674 ffs_update(vp, 1); 2675 } 2676 } 2677 vput(vp); 2678 return (error); 2679} 2680 2681static void 2682journal_unmount(mp) 2683 struct mount *mp; 2684{ 2685 struct ufsmount *ump; 2686 2687 ump = VFSTOUFS(mp); 2688 if (ump->softdep_jblocks) 2689 jblocks_destroy(ump->softdep_jblocks); 2690 ump->softdep_jblocks = NULL; 2691} 2692 2693/* 2694 * Called when a journal record is ready to be written. Space is allocated 2695 * and the journal entry is created when the journal is flushed to stable 2696 * store. 2697 */ 2698static void 2699add_to_journal(wk) 2700 struct worklist *wk; 2701{ 2702 struct ufsmount *ump; 2703 2704 mtx_assert(&lk, MA_OWNED); 2705 ump = VFSTOUFS(wk->wk_mp); 2706 if (wk->wk_state & ONWORKLIST) 2707 panic("add_to_journal: %s(0x%X) already on list", 2708 TYPENAME(wk->wk_type), wk->wk_state); 2709 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2710 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2711 ump->softdep_jblocks->jb_age = ticks; 2712 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2713 } else 2714 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2715 ump->softdep_journal_tail = wk; 2716 ump->softdep_on_journal += 1; 2717} 2718 2719/* 2720 * Remove an arbitrary item for the journal worklist maintain the tail 2721 * pointer. This happens when a new operation obviates the need to 2722 * journal an old operation. 2723 */ 2724static void 2725remove_from_journal(wk) 2726 struct worklist *wk; 2727{ 2728 struct ufsmount *ump; 2729 2730 mtx_assert(&lk, MA_OWNED); 2731 ump = VFSTOUFS(wk->wk_mp); 2732#ifdef SUJ_DEBUG 2733 { 2734 struct worklist *wkn; 2735 2736 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2737 if (wkn == wk) 2738 break; 2739 if (wkn == NULL) 2740 panic("remove_from_journal: %p is not in journal", wk); 2741 } 2742#endif 2743 /* 2744 * We emulate a TAILQ to save space in most structures which do not 2745 * require TAILQ semantics. Here we must update the tail position 2746 * when removing the tail which is not the final entry. This works 2747 * only if the worklist linkage are at the beginning of the structure. 2748 */ 2749 if (ump->softdep_journal_tail == wk) 2750 ump->softdep_journal_tail = 2751 (struct worklist *)wk->wk_list.le_prev; 2752 2753 WORKLIST_REMOVE(wk); 2754 ump->softdep_on_journal -= 1; 2755} 2756 2757/* 2758 * Check for journal space as well as dependency limits so the prelink 2759 * code can throttle both journaled and non-journaled filesystems. 2760 * Threshold is 0 for low and 1 for min. 2761 */ 2762static int 2763journal_space(ump, thresh) 2764 struct ufsmount *ump; 2765 int thresh; 2766{ 2767 struct jblocks *jblocks; 2768 int avail; 2769 2770 jblocks = ump->softdep_jblocks; 2771 if (jblocks == NULL) 2772 return (1); 2773 /* 2774 * We use a tighter restriction here to prevent request_cleanup() 2775 * running in threads from running into locks we currently hold. 2776 */ 2777 if (dep_current[D_INODEDEP] > (max_softdeps / 10) * 9) 2778 return (0); 2779 if (thresh) 2780 thresh = jblocks->jb_min; 2781 else 2782 thresh = jblocks->jb_low; 2783 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2784 avail = jblocks->jb_free - avail; 2785 2786 return (avail > thresh); 2787} 2788 2789static void 2790journal_suspend(ump) 2791 struct ufsmount *ump; 2792{ 2793 struct jblocks *jblocks; 2794 struct mount *mp; 2795 2796 mp = UFSTOVFS(ump); 2797 jblocks = ump->softdep_jblocks; 2798 MNT_ILOCK(mp); 2799 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2800 stat_journal_min++; 2801 mp->mnt_kern_flag |= MNTK_SUSPEND; 2802 mp->mnt_susp_owner = FIRST_THREAD_IN_PROC(softdepproc); 2803 } 2804 jblocks->jb_suspended = 1; 2805 MNT_IUNLOCK(mp); 2806} 2807 2808static int 2809journal_unsuspend(struct ufsmount *ump) 2810{ 2811 struct jblocks *jblocks; 2812 struct mount *mp; 2813 2814 mp = UFSTOVFS(ump); 2815 jblocks = ump->softdep_jblocks; 2816 2817 if (jblocks != NULL && jblocks->jb_suspended && 2818 journal_space(ump, jblocks->jb_min)) { 2819 jblocks->jb_suspended = 0; 2820 FREE_LOCK(&lk); 2821 mp->mnt_susp_owner = curthread; 2822 vfs_write_resume(mp, 0); 2823 ACQUIRE_LOCK(&lk); 2824 return (1); 2825 } 2826 return (0); 2827} 2828 2829/* 2830 * Called before any allocation function to be certain that there is 2831 * sufficient space in the journal prior to creating any new records. 2832 * Since in the case of block allocation we may have multiple locked 2833 * buffers at the time of the actual allocation we can not block 2834 * when the journal records are created. Doing so would create a deadlock 2835 * if any of these buffers needed to be flushed to reclaim space. Instead 2836 * we require a sufficiently large amount of available space such that 2837 * each thread in the system could have passed this allocation check and 2838 * still have sufficient free space. With 20% of a minimum journal size 2839 * of 1MB we have 6553 records available. 2840 */ 2841int 2842softdep_prealloc(vp, waitok) 2843 struct vnode *vp; 2844 int waitok; 2845{ 2846 struct ufsmount *ump; 2847 2848 /* 2849 * Nothing to do if we are not running journaled soft updates. 2850 * If we currently hold the snapshot lock, we must avoid handling 2851 * other resources that could cause deadlock. 2852 */ 2853 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp))) 2854 return (0); 2855 ump = VFSTOUFS(vp->v_mount); 2856 ACQUIRE_LOCK(&lk); 2857 if (journal_space(ump, 0)) { 2858 FREE_LOCK(&lk); 2859 return (0); 2860 } 2861 stat_journal_low++; 2862 FREE_LOCK(&lk); 2863 if (waitok == MNT_NOWAIT) 2864 return (ENOSPC); 2865 /* 2866 * Attempt to sync this vnode once to flush any journal 2867 * work attached to it. 2868 */ 2869 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 2870 ffs_syncvnode(vp, waitok, 0); 2871 ACQUIRE_LOCK(&lk); 2872 process_removes(vp); 2873 process_truncates(vp); 2874 if (journal_space(ump, 0) == 0) { 2875 softdep_speedup(); 2876 if (journal_space(ump, 1) == 0) 2877 journal_suspend(ump); 2878 } 2879 FREE_LOCK(&lk); 2880 2881 return (0); 2882} 2883 2884/* 2885 * Before adjusting a link count on a vnode verify that we have sufficient 2886 * journal space. If not, process operations that depend on the currently 2887 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 2888 * and softdep flush threads can not acquire these locks to reclaim space. 2889 */ 2890static void 2891softdep_prelink(dvp, vp) 2892 struct vnode *dvp; 2893 struct vnode *vp; 2894{ 2895 struct ufsmount *ump; 2896 2897 ump = VFSTOUFS(dvp->v_mount); 2898 mtx_assert(&lk, MA_OWNED); 2899 /* 2900 * Nothing to do if we have sufficient journal space. 2901 * If we currently hold the snapshot lock, we must avoid 2902 * handling other resources that could cause deadlock. 2903 */ 2904 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 2905 return; 2906 stat_journal_low++; 2907 FREE_LOCK(&lk); 2908 if (vp) 2909 ffs_syncvnode(vp, MNT_NOWAIT, 0); 2910 ffs_syncvnode(dvp, MNT_WAIT, 0); 2911 ACQUIRE_LOCK(&lk); 2912 /* Process vp before dvp as it may create .. removes. */ 2913 if (vp) { 2914 process_removes(vp); 2915 process_truncates(vp); 2916 } 2917 process_removes(dvp); 2918 process_truncates(dvp); 2919 softdep_speedup(); 2920 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 2921 if (journal_space(ump, 0) == 0) { 2922 softdep_speedup(); 2923 if (journal_space(ump, 1) == 0) 2924 journal_suspend(ump); 2925 } 2926} 2927 2928static void 2929jseg_write(ump, jseg, data) 2930 struct ufsmount *ump; 2931 struct jseg *jseg; 2932 uint8_t *data; 2933{ 2934 struct jsegrec *rec; 2935 2936 rec = (struct jsegrec *)data; 2937 rec->jsr_seq = jseg->js_seq; 2938 rec->jsr_oldest = jseg->js_oldseq; 2939 rec->jsr_cnt = jseg->js_cnt; 2940 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 2941 rec->jsr_crc = 0; 2942 rec->jsr_time = ump->um_fs->fs_mtime; 2943} 2944 2945static inline void 2946inoref_write(inoref, jseg, rec) 2947 struct inoref *inoref; 2948 struct jseg *jseg; 2949 struct jrefrec *rec; 2950{ 2951 2952 inoref->if_jsegdep->jd_seg = jseg; 2953 rec->jr_ino = inoref->if_ino; 2954 rec->jr_parent = inoref->if_parent; 2955 rec->jr_nlink = inoref->if_nlink; 2956 rec->jr_mode = inoref->if_mode; 2957 rec->jr_diroff = inoref->if_diroff; 2958} 2959 2960static void 2961jaddref_write(jaddref, jseg, data) 2962 struct jaddref *jaddref; 2963 struct jseg *jseg; 2964 uint8_t *data; 2965{ 2966 struct jrefrec *rec; 2967 2968 rec = (struct jrefrec *)data; 2969 rec->jr_op = JOP_ADDREF; 2970 inoref_write(&jaddref->ja_ref, jseg, rec); 2971} 2972 2973static void 2974jremref_write(jremref, jseg, data) 2975 struct jremref *jremref; 2976 struct jseg *jseg; 2977 uint8_t *data; 2978{ 2979 struct jrefrec *rec; 2980 2981 rec = (struct jrefrec *)data; 2982 rec->jr_op = JOP_REMREF; 2983 inoref_write(&jremref->jr_ref, jseg, rec); 2984} 2985 2986static void 2987jmvref_write(jmvref, jseg, data) 2988 struct jmvref *jmvref; 2989 struct jseg *jseg; 2990 uint8_t *data; 2991{ 2992 struct jmvrec *rec; 2993 2994 rec = (struct jmvrec *)data; 2995 rec->jm_op = JOP_MVREF; 2996 rec->jm_ino = jmvref->jm_ino; 2997 rec->jm_parent = jmvref->jm_parent; 2998 rec->jm_oldoff = jmvref->jm_oldoff; 2999 rec->jm_newoff = jmvref->jm_newoff; 3000} 3001 3002static void 3003jnewblk_write(jnewblk, jseg, data) 3004 struct jnewblk *jnewblk; 3005 struct jseg *jseg; 3006 uint8_t *data; 3007{ 3008 struct jblkrec *rec; 3009 3010 jnewblk->jn_jsegdep->jd_seg = jseg; 3011 rec = (struct jblkrec *)data; 3012 rec->jb_op = JOP_NEWBLK; 3013 rec->jb_ino = jnewblk->jn_ino; 3014 rec->jb_blkno = jnewblk->jn_blkno; 3015 rec->jb_lbn = jnewblk->jn_lbn; 3016 rec->jb_frags = jnewblk->jn_frags; 3017 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3018} 3019 3020static void 3021jfreeblk_write(jfreeblk, jseg, data) 3022 struct jfreeblk *jfreeblk; 3023 struct jseg *jseg; 3024 uint8_t *data; 3025{ 3026 struct jblkrec *rec; 3027 3028 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3029 rec = (struct jblkrec *)data; 3030 rec->jb_op = JOP_FREEBLK; 3031 rec->jb_ino = jfreeblk->jf_ino; 3032 rec->jb_blkno = jfreeblk->jf_blkno; 3033 rec->jb_lbn = jfreeblk->jf_lbn; 3034 rec->jb_frags = jfreeblk->jf_frags; 3035 rec->jb_oldfrags = 0; 3036} 3037 3038static void 3039jfreefrag_write(jfreefrag, jseg, data) 3040 struct jfreefrag *jfreefrag; 3041 struct jseg *jseg; 3042 uint8_t *data; 3043{ 3044 struct jblkrec *rec; 3045 3046 jfreefrag->fr_jsegdep->jd_seg = jseg; 3047 rec = (struct jblkrec *)data; 3048 rec->jb_op = JOP_FREEBLK; 3049 rec->jb_ino = jfreefrag->fr_ino; 3050 rec->jb_blkno = jfreefrag->fr_blkno; 3051 rec->jb_lbn = jfreefrag->fr_lbn; 3052 rec->jb_frags = jfreefrag->fr_frags; 3053 rec->jb_oldfrags = 0; 3054} 3055 3056static void 3057jtrunc_write(jtrunc, jseg, data) 3058 struct jtrunc *jtrunc; 3059 struct jseg *jseg; 3060 uint8_t *data; 3061{ 3062 struct jtrncrec *rec; 3063 3064 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3065 rec = (struct jtrncrec *)data; 3066 rec->jt_op = JOP_TRUNC; 3067 rec->jt_ino = jtrunc->jt_ino; 3068 rec->jt_size = jtrunc->jt_size; 3069 rec->jt_extsize = jtrunc->jt_extsize; 3070} 3071 3072static void 3073jfsync_write(jfsync, jseg, data) 3074 struct jfsync *jfsync; 3075 struct jseg *jseg; 3076 uint8_t *data; 3077{ 3078 struct jtrncrec *rec; 3079 3080 rec = (struct jtrncrec *)data; 3081 rec->jt_op = JOP_SYNC; 3082 rec->jt_ino = jfsync->jfs_ino; 3083 rec->jt_size = jfsync->jfs_size; 3084 rec->jt_extsize = jfsync->jfs_extsize; 3085} 3086 3087static void 3088softdep_flushjournal(mp) 3089 struct mount *mp; 3090{ 3091 struct jblocks *jblocks; 3092 struct ufsmount *ump; 3093 3094 if (MOUNTEDSUJ(mp) == 0) 3095 return; 3096 ump = VFSTOUFS(mp); 3097 jblocks = ump->softdep_jblocks; 3098 ACQUIRE_LOCK(&lk); 3099 while (ump->softdep_on_journal) { 3100 jblocks->jb_needseg = 1; 3101 softdep_process_journal(mp, NULL, MNT_WAIT); 3102 } 3103 FREE_LOCK(&lk); 3104} 3105 3106static void softdep_synchronize_completed(struct bio *); 3107static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3108 3109static void 3110softdep_synchronize_completed(bp) 3111 struct bio *bp; 3112{ 3113 struct jseg *oldest; 3114 struct jseg *jseg; 3115 3116 /* 3117 * caller1 marks the last segment written before we issued the 3118 * synchronize cache. 3119 */ 3120 jseg = bp->bio_caller1; 3121 oldest = NULL; 3122 ACQUIRE_LOCK(&lk); 3123 /* 3124 * Mark all the journal entries waiting on the synchronize cache 3125 * as completed so they may continue on. 3126 */ 3127 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3128 jseg->js_state |= COMPLETE; 3129 oldest = jseg; 3130 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3131 } 3132 /* 3133 * Restart deferred journal entry processing from the oldest 3134 * completed jseg. 3135 */ 3136 if (oldest) 3137 complete_jsegs(oldest); 3138 3139 FREE_LOCK(&lk); 3140 g_destroy_bio(bp); 3141} 3142 3143/* 3144 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3145 * barriers. The journal must be written prior to any blocks that depend 3146 * on it and the journal can not be released until the blocks have be 3147 * written. This code handles both barriers simultaneously. 3148 */ 3149static void 3150softdep_synchronize(bp, ump, caller1) 3151 struct bio *bp; 3152 struct ufsmount *ump; 3153 void *caller1; 3154{ 3155 3156 bp->bio_cmd = BIO_FLUSH; 3157 bp->bio_flags |= BIO_ORDERED; 3158 bp->bio_data = NULL; 3159 bp->bio_offset = ump->um_cp->provider->mediasize; 3160 bp->bio_length = 0; 3161 bp->bio_done = softdep_synchronize_completed; 3162 bp->bio_caller1 = caller1; 3163 g_io_request(bp, 3164 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3165} 3166 3167/* 3168 * Flush some journal records to disk. 3169 */ 3170static void 3171softdep_process_journal(mp, needwk, flags) 3172 struct mount *mp; 3173 struct worklist *needwk; 3174 int flags; 3175{ 3176 struct jblocks *jblocks; 3177 struct ufsmount *ump; 3178 struct worklist *wk; 3179 struct jseg *jseg; 3180 struct buf *bp; 3181 struct bio *bio; 3182 uint8_t *data; 3183 struct fs *fs; 3184 int shouldflush; 3185 int segwritten; 3186 int jrecmin; /* Minimum records per block. */ 3187 int jrecmax; /* Maximum records per block. */ 3188 int size; 3189 int cnt; 3190 int off; 3191 int devbsize; 3192 3193 if (MOUNTEDSUJ(mp) == 0) 3194 return; 3195 shouldflush = softdep_flushcache; 3196 bio = NULL; 3197 jseg = NULL; 3198 ump = VFSTOUFS(mp); 3199 fs = ump->um_fs; 3200 jblocks = ump->softdep_jblocks; 3201 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3202 /* 3203 * We write anywhere between a disk block and fs block. The upper 3204 * bound is picked to prevent buffer cache fragmentation and limit 3205 * processing time per I/O. 3206 */ 3207 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3208 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3209 segwritten = 0; 3210 for (;;) { 3211 cnt = ump->softdep_on_journal; 3212 /* 3213 * Criteria for writing a segment: 3214 * 1) We have a full block. 3215 * 2) We're called from jwait() and haven't found the 3216 * journal item yet. 3217 * 3) Always write if needseg is set. 3218 * 4) If we are called from process_worklist and have 3219 * not yet written anything we write a partial block 3220 * to enforce a 1 second maximum latency on journal 3221 * entries. 3222 */ 3223 if (cnt < (jrecmax - 1) && needwk == NULL && 3224 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3225 break; 3226 cnt++; 3227 /* 3228 * Verify some free journal space. softdep_prealloc() should 3229 * guarantee that we don't run out so this is indicative of 3230 * a problem with the flow control. Try to recover 3231 * gracefully in any event. 3232 */ 3233 while (jblocks->jb_free == 0) { 3234 if (flags != MNT_WAIT) 3235 break; 3236 printf("softdep: Out of journal space!\n"); 3237 softdep_speedup(); 3238 msleep(jblocks, &lk, PRIBIO, "jblocks", hz); 3239 } 3240 FREE_LOCK(&lk); 3241 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3242 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3243 LIST_INIT(&jseg->js_entries); 3244 LIST_INIT(&jseg->js_indirs); 3245 jseg->js_state = ATTACHED; 3246 if (shouldflush == 0) 3247 jseg->js_state |= COMPLETE; 3248 else if (bio == NULL) 3249 bio = g_alloc_bio(); 3250 jseg->js_jblocks = jblocks; 3251 bp = geteblk(fs->fs_bsize, 0); 3252 ACQUIRE_LOCK(&lk); 3253 /* 3254 * If there was a race while we were allocating the block 3255 * and jseg the entry we care about was likely written. 3256 * We bail out in both the WAIT and NOWAIT case and assume 3257 * the caller will loop if the entry it cares about is 3258 * not written. 3259 */ 3260 cnt = ump->softdep_on_journal; 3261 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3262 bp->b_flags |= B_INVAL | B_NOCACHE; 3263 WORKITEM_FREE(jseg, D_JSEG); 3264 FREE_LOCK(&lk); 3265 brelse(bp); 3266 ACQUIRE_LOCK(&lk); 3267 break; 3268 } 3269 /* 3270 * Calculate the disk block size required for the available 3271 * records rounded to the min size. 3272 */ 3273 if (cnt == 0) 3274 size = devbsize; 3275 else if (cnt < jrecmax) 3276 size = howmany(cnt, jrecmin) * devbsize; 3277 else 3278 size = fs->fs_bsize; 3279 /* 3280 * Allocate a disk block for this journal data and account 3281 * for truncation of the requested size if enough contiguous 3282 * space was not available. 3283 */ 3284 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3285 bp->b_lblkno = bp->b_blkno; 3286 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3287 bp->b_bcount = size; 3288 bp->b_flags &= ~B_INVAL; 3289 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3290 /* 3291 * Initialize our jseg with cnt records. Assign the next 3292 * sequence number to it and link it in-order. 3293 */ 3294 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3295 jseg->js_buf = bp; 3296 jseg->js_cnt = cnt; 3297 jseg->js_refs = cnt + 1; /* Self ref. */ 3298 jseg->js_size = size; 3299 jseg->js_seq = jblocks->jb_nextseq++; 3300 if (jblocks->jb_oldestseg == NULL) 3301 jblocks->jb_oldestseg = jseg; 3302 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3303 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3304 if (jblocks->jb_writeseg == NULL) 3305 jblocks->jb_writeseg = jseg; 3306 /* 3307 * Start filling in records from the pending list. 3308 */ 3309 data = bp->b_data; 3310 off = 0; 3311 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3312 != NULL) { 3313 if (cnt == 0) 3314 break; 3315 /* Place a segment header on every device block. */ 3316 if ((off % devbsize) == 0) { 3317 jseg_write(ump, jseg, data); 3318 off += JREC_SIZE; 3319 data = bp->b_data + off; 3320 } 3321 if (wk == needwk) 3322 needwk = NULL; 3323 remove_from_journal(wk); 3324 wk->wk_state |= INPROGRESS; 3325 WORKLIST_INSERT(&jseg->js_entries, wk); 3326 switch (wk->wk_type) { 3327 case D_JADDREF: 3328 jaddref_write(WK_JADDREF(wk), jseg, data); 3329 break; 3330 case D_JREMREF: 3331 jremref_write(WK_JREMREF(wk), jseg, data); 3332 break; 3333 case D_JMVREF: 3334 jmvref_write(WK_JMVREF(wk), jseg, data); 3335 break; 3336 case D_JNEWBLK: 3337 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3338 break; 3339 case D_JFREEBLK: 3340 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3341 break; 3342 case D_JFREEFRAG: 3343 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3344 break; 3345 case D_JTRUNC: 3346 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3347 break; 3348 case D_JFSYNC: 3349 jfsync_write(WK_JFSYNC(wk), jseg, data); 3350 break; 3351 default: 3352 panic("process_journal: Unknown type %s", 3353 TYPENAME(wk->wk_type)); 3354 /* NOTREACHED */ 3355 } 3356 off += JREC_SIZE; 3357 data = bp->b_data + off; 3358 cnt--; 3359 } 3360 /* 3361 * Write this one buffer and continue. 3362 */ 3363 segwritten = 1; 3364 jblocks->jb_needseg = 0; 3365 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3366 FREE_LOCK(&lk); 3367 pbgetvp(ump->um_devvp, bp); 3368 /* 3369 * We only do the blocking wait once we find the journal 3370 * entry we're looking for. 3371 */ 3372 if (needwk == NULL && flags == MNT_WAIT) 3373 bwrite(bp); 3374 else 3375 bawrite(bp); 3376 ACQUIRE_LOCK(&lk); 3377 } 3378 /* 3379 * If we wrote a segment issue a synchronize cache so the journal 3380 * is reflected on disk before the data is written. Since reclaiming 3381 * journal space also requires writing a journal record this 3382 * process also enforces a barrier before reclamation. 3383 */ 3384 if (segwritten && shouldflush) { 3385 softdep_synchronize(bio, ump, 3386 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3387 } else if (bio) 3388 g_destroy_bio(bio); 3389 /* 3390 * If we've suspended the filesystem because we ran out of journal 3391 * space either try to sync it here to make some progress or 3392 * unsuspend it if we already have. 3393 */ 3394 if (flags == 0 && jblocks->jb_suspended) { 3395 if (journal_unsuspend(ump)) 3396 return; 3397 FREE_LOCK(&lk); 3398 VFS_SYNC(mp, MNT_NOWAIT); 3399 ffs_sbupdate(ump, MNT_WAIT, 0); 3400 ACQUIRE_LOCK(&lk); 3401 } 3402} 3403 3404/* 3405 * Complete a jseg, allowing all dependencies awaiting journal writes 3406 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3407 * structures so that the journal segment can be freed to reclaim space. 3408 */ 3409static void 3410complete_jseg(jseg) 3411 struct jseg *jseg; 3412{ 3413 struct worklist *wk; 3414 struct jmvref *jmvref; 3415 int waiting; 3416#ifdef INVARIANTS 3417 int i = 0; 3418#endif 3419 3420 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3421 WORKLIST_REMOVE(wk); 3422 waiting = wk->wk_state & IOWAITING; 3423 wk->wk_state &= ~(INPROGRESS | IOWAITING); 3424 wk->wk_state |= COMPLETE; 3425 KASSERT(i++ < jseg->js_cnt, 3426 ("handle_written_jseg: overflow %d >= %d", 3427 i - 1, jseg->js_cnt)); 3428 switch (wk->wk_type) { 3429 case D_JADDREF: 3430 handle_written_jaddref(WK_JADDREF(wk)); 3431 break; 3432 case D_JREMREF: 3433 handle_written_jremref(WK_JREMREF(wk)); 3434 break; 3435 case D_JMVREF: 3436 rele_jseg(jseg); /* No jsegdep. */ 3437 jmvref = WK_JMVREF(wk); 3438 LIST_REMOVE(jmvref, jm_deps); 3439 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3440 free_pagedep(jmvref->jm_pagedep); 3441 WORKITEM_FREE(jmvref, D_JMVREF); 3442 break; 3443 case D_JNEWBLK: 3444 handle_written_jnewblk(WK_JNEWBLK(wk)); 3445 break; 3446 case D_JFREEBLK: 3447 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3448 break; 3449 case D_JTRUNC: 3450 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3451 break; 3452 case D_JFSYNC: 3453 rele_jseg(jseg); /* No jsegdep. */ 3454 WORKITEM_FREE(wk, D_JFSYNC); 3455 break; 3456 case D_JFREEFRAG: 3457 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3458 break; 3459 default: 3460 panic("handle_written_jseg: Unknown type %s", 3461 TYPENAME(wk->wk_type)); 3462 /* NOTREACHED */ 3463 } 3464 if (waiting) 3465 wakeup(wk); 3466 } 3467 /* Release the self reference so the structure may be freed. */ 3468 rele_jseg(jseg); 3469} 3470 3471/* 3472 * Determine which jsegs are ready for completion processing. Waits for 3473 * synchronize cache to complete as well as forcing in-order completion 3474 * of journal entries. 3475 */ 3476static void 3477complete_jsegs(jseg) 3478 struct jseg *jseg; 3479{ 3480 struct jblocks *jblocks; 3481 struct jseg *jsegn; 3482 3483 jblocks = jseg->js_jblocks; 3484 /* 3485 * Don't allow out of order completions. If this isn't the first 3486 * block wait for it to write before we're done. 3487 */ 3488 if (jseg != jblocks->jb_writeseg) 3489 return; 3490 /* Iterate through available jsegs processing their entries. */ 3491 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3492 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3493 jsegn = TAILQ_NEXT(jseg, js_next); 3494 complete_jseg(jseg); 3495 jseg = jsegn; 3496 } 3497 jblocks->jb_writeseg = jseg; 3498 /* 3499 * Attempt to free jsegs now that oldestwrseq may have advanced. 3500 */ 3501 free_jsegs(jblocks); 3502} 3503 3504/* 3505 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3506 * the final completions. 3507 */ 3508static void 3509handle_written_jseg(jseg, bp) 3510 struct jseg *jseg; 3511 struct buf *bp; 3512{ 3513 3514 if (jseg->js_refs == 0) 3515 panic("handle_written_jseg: No self-reference on %p", jseg); 3516 jseg->js_state |= DEPCOMPLETE; 3517 /* 3518 * We'll never need this buffer again, set flags so it will be 3519 * discarded. 3520 */ 3521 bp->b_flags |= B_INVAL | B_NOCACHE; 3522 pbrelvp(bp); 3523 complete_jsegs(jseg); 3524} 3525 3526static inline struct jsegdep * 3527inoref_jseg(inoref) 3528 struct inoref *inoref; 3529{ 3530 struct jsegdep *jsegdep; 3531 3532 jsegdep = inoref->if_jsegdep; 3533 inoref->if_jsegdep = NULL; 3534 3535 return (jsegdep); 3536} 3537 3538/* 3539 * Called once a jremref has made it to stable store. The jremref is marked 3540 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3541 * for the jremref to complete will be awoken by free_jremref. 3542 */ 3543static void 3544handle_written_jremref(jremref) 3545 struct jremref *jremref; 3546{ 3547 struct inodedep *inodedep; 3548 struct jsegdep *jsegdep; 3549 struct dirrem *dirrem; 3550 3551 /* Grab the jsegdep. */ 3552 jsegdep = inoref_jseg(&jremref->jr_ref); 3553 /* 3554 * Remove us from the inoref list. 3555 */ 3556 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3557 0, &inodedep) == 0) 3558 panic("handle_written_jremref: Lost inodedep"); 3559 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3560 /* 3561 * Complete the dirrem. 3562 */ 3563 dirrem = jremref->jr_dirrem; 3564 jremref->jr_dirrem = NULL; 3565 LIST_REMOVE(jremref, jr_deps); 3566 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3567 jwork_insert(&dirrem->dm_jwork, jsegdep); 3568 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3569 (dirrem->dm_state & COMPLETE) != 0) 3570 add_to_worklist(&dirrem->dm_list, 0); 3571 free_jremref(jremref); 3572} 3573 3574/* 3575 * Called once a jaddref has made it to stable store. The dependency is 3576 * marked complete and any dependent structures are added to the inode 3577 * bufwait list to be completed as soon as it is written. If a bitmap write 3578 * depends on this entry we move the inode into the inodedephd of the 3579 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3580 */ 3581static void 3582handle_written_jaddref(jaddref) 3583 struct jaddref *jaddref; 3584{ 3585 struct jsegdep *jsegdep; 3586 struct inodedep *inodedep; 3587 struct diradd *diradd; 3588 struct mkdir *mkdir; 3589 3590 /* Grab the jsegdep. */ 3591 jsegdep = inoref_jseg(&jaddref->ja_ref); 3592 mkdir = NULL; 3593 diradd = NULL; 3594 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3595 0, &inodedep) == 0) 3596 panic("handle_written_jaddref: Lost inodedep."); 3597 if (jaddref->ja_diradd == NULL) 3598 panic("handle_written_jaddref: No dependency"); 3599 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3600 diradd = jaddref->ja_diradd; 3601 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3602 } else if (jaddref->ja_state & MKDIR_PARENT) { 3603 mkdir = jaddref->ja_mkdir; 3604 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3605 } else if (jaddref->ja_state & MKDIR_BODY) 3606 mkdir = jaddref->ja_mkdir; 3607 else 3608 panic("handle_written_jaddref: Unknown dependency %p", 3609 jaddref->ja_diradd); 3610 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3611 /* 3612 * Remove us from the inode list. 3613 */ 3614 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3615 /* 3616 * The mkdir may be waiting on the jaddref to clear before freeing. 3617 */ 3618 if (mkdir) { 3619 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3620 ("handle_written_jaddref: Incorrect type for mkdir %s", 3621 TYPENAME(mkdir->md_list.wk_type))); 3622 mkdir->md_jaddref = NULL; 3623 diradd = mkdir->md_diradd; 3624 mkdir->md_state |= DEPCOMPLETE; 3625 complete_mkdir(mkdir); 3626 } 3627 jwork_insert(&diradd->da_jwork, jsegdep); 3628 if (jaddref->ja_state & NEWBLOCK) { 3629 inodedep->id_state |= ONDEPLIST; 3630 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3631 inodedep, id_deps); 3632 } 3633 free_jaddref(jaddref); 3634} 3635 3636/* 3637 * Called once a jnewblk journal is written. The allocdirect or allocindir 3638 * is placed in the bmsafemap to await notification of a written bitmap. If 3639 * the operation was canceled we add the segdep to the appropriate 3640 * dependency to free the journal space once the canceling operation 3641 * completes. 3642 */ 3643static void 3644handle_written_jnewblk(jnewblk) 3645 struct jnewblk *jnewblk; 3646{ 3647 struct bmsafemap *bmsafemap; 3648 struct freefrag *freefrag; 3649 struct freework *freework; 3650 struct jsegdep *jsegdep; 3651 struct newblk *newblk; 3652 3653 /* Grab the jsegdep. */ 3654 jsegdep = jnewblk->jn_jsegdep; 3655 jnewblk->jn_jsegdep = NULL; 3656 if (jnewblk->jn_dep == NULL) 3657 panic("handle_written_jnewblk: No dependency for the segdep."); 3658 switch (jnewblk->jn_dep->wk_type) { 3659 case D_NEWBLK: 3660 case D_ALLOCDIRECT: 3661 case D_ALLOCINDIR: 3662 /* 3663 * Add the written block to the bmsafemap so it can 3664 * be notified when the bitmap is on disk. 3665 */ 3666 newblk = WK_NEWBLK(jnewblk->jn_dep); 3667 newblk->nb_jnewblk = NULL; 3668 if ((newblk->nb_state & GOINGAWAY) == 0) { 3669 bmsafemap = newblk->nb_bmsafemap; 3670 newblk->nb_state |= ONDEPLIST; 3671 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3672 nb_deps); 3673 } 3674 jwork_insert(&newblk->nb_jwork, jsegdep); 3675 break; 3676 case D_FREEFRAG: 3677 /* 3678 * A newblock being removed by a freefrag when replaced by 3679 * frag extension. 3680 */ 3681 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3682 freefrag->ff_jdep = NULL; 3683 jwork_insert(&freefrag->ff_jwork, jsegdep); 3684 break; 3685 case D_FREEWORK: 3686 /* 3687 * A direct block was removed by truncate. 3688 */ 3689 freework = WK_FREEWORK(jnewblk->jn_dep); 3690 freework->fw_jnewblk = NULL; 3691 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3692 break; 3693 default: 3694 panic("handle_written_jnewblk: Unknown type %d.", 3695 jnewblk->jn_dep->wk_type); 3696 } 3697 jnewblk->jn_dep = NULL; 3698 free_jnewblk(jnewblk); 3699} 3700 3701/* 3702 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3703 * an in-flight allocation that has not yet been committed. Divorce us 3704 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3705 * to the worklist. 3706 */ 3707static void 3708cancel_jfreefrag(jfreefrag) 3709 struct jfreefrag *jfreefrag; 3710{ 3711 struct freefrag *freefrag; 3712 3713 if (jfreefrag->fr_jsegdep) { 3714 free_jsegdep(jfreefrag->fr_jsegdep); 3715 jfreefrag->fr_jsegdep = NULL; 3716 } 3717 freefrag = jfreefrag->fr_freefrag; 3718 jfreefrag->fr_freefrag = NULL; 3719 free_jfreefrag(jfreefrag); 3720 freefrag->ff_state |= DEPCOMPLETE; 3721 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3722} 3723 3724/* 3725 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3726 */ 3727static void 3728free_jfreefrag(jfreefrag) 3729 struct jfreefrag *jfreefrag; 3730{ 3731 3732 if (jfreefrag->fr_state & INPROGRESS) 3733 WORKLIST_REMOVE(&jfreefrag->fr_list); 3734 else if (jfreefrag->fr_state & ONWORKLIST) 3735 remove_from_journal(&jfreefrag->fr_list); 3736 if (jfreefrag->fr_freefrag != NULL) 3737 panic("free_jfreefrag: Still attached to a freefrag."); 3738 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3739} 3740 3741/* 3742 * Called when the journal write for a jfreefrag completes. The parent 3743 * freefrag is added to the worklist if this completes its dependencies. 3744 */ 3745static void 3746handle_written_jfreefrag(jfreefrag) 3747 struct jfreefrag *jfreefrag; 3748{ 3749 struct jsegdep *jsegdep; 3750 struct freefrag *freefrag; 3751 3752 /* Grab the jsegdep. */ 3753 jsegdep = jfreefrag->fr_jsegdep; 3754 jfreefrag->fr_jsegdep = NULL; 3755 freefrag = jfreefrag->fr_freefrag; 3756 if (freefrag == NULL) 3757 panic("handle_written_jfreefrag: No freefrag."); 3758 freefrag->ff_state |= DEPCOMPLETE; 3759 freefrag->ff_jdep = NULL; 3760 jwork_insert(&freefrag->ff_jwork, jsegdep); 3761 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3762 add_to_worklist(&freefrag->ff_list, 0); 3763 jfreefrag->fr_freefrag = NULL; 3764 free_jfreefrag(jfreefrag); 3765} 3766 3767/* 3768 * Called when the journal write for a jfreeblk completes. The jfreeblk 3769 * is removed from the freeblks list of pending journal writes and the 3770 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3771 * have been reclaimed. 3772 */ 3773static void 3774handle_written_jblkdep(jblkdep) 3775 struct jblkdep *jblkdep; 3776{ 3777 struct freeblks *freeblks; 3778 struct jsegdep *jsegdep; 3779 3780 /* Grab the jsegdep. */ 3781 jsegdep = jblkdep->jb_jsegdep; 3782 jblkdep->jb_jsegdep = NULL; 3783 freeblks = jblkdep->jb_freeblks; 3784 LIST_REMOVE(jblkdep, jb_deps); 3785 jwork_insert(&freeblks->fb_jwork, jsegdep); 3786 /* 3787 * If the freeblks is all journaled, we can add it to the worklist. 3788 */ 3789 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3790 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3791 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3792 3793 free_jblkdep(jblkdep); 3794} 3795 3796static struct jsegdep * 3797newjsegdep(struct worklist *wk) 3798{ 3799 struct jsegdep *jsegdep; 3800 3801 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3802 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3803 jsegdep->jd_seg = NULL; 3804 3805 return (jsegdep); 3806} 3807 3808static struct jmvref * 3809newjmvref(dp, ino, oldoff, newoff) 3810 struct inode *dp; 3811 ino_t ino; 3812 off_t oldoff; 3813 off_t newoff; 3814{ 3815 struct jmvref *jmvref; 3816 3817 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3818 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump)); 3819 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 3820 jmvref->jm_parent = dp->i_number; 3821 jmvref->jm_ino = ino; 3822 jmvref->jm_oldoff = oldoff; 3823 jmvref->jm_newoff = newoff; 3824 3825 return (jmvref); 3826} 3827 3828/* 3829 * Allocate a new jremref that tracks the removal of ip from dp with the 3830 * directory entry offset of diroff. Mark the entry as ATTACHED and 3831 * DEPCOMPLETE as we have all the information required for the journal write 3832 * and the directory has already been removed from the buffer. The caller 3833 * is responsible for linking the jremref into the pagedep and adding it 3834 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 3835 * a DOTDOT addition so handle_workitem_remove() can properly assign 3836 * the jsegdep when we're done. 3837 */ 3838static struct jremref * 3839newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 3840 off_t diroff, nlink_t nlink) 3841{ 3842 struct jremref *jremref; 3843 3844 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 3845 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump)); 3846 jremref->jr_state = ATTACHED; 3847 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 3848 nlink, ip->i_mode); 3849 jremref->jr_dirrem = dirrem; 3850 3851 return (jremref); 3852} 3853 3854static inline void 3855newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 3856 nlink_t nlink, uint16_t mode) 3857{ 3858 3859 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 3860 inoref->if_diroff = diroff; 3861 inoref->if_ino = ino; 3862 inoref->if_parent = parent; 3863 inoref->if_nlink = nlink; 3864 inoref->if_mode = mode; 3865} 3866 3867/* 3868 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 3869 * directory offset may not be known until later. The caller is responsible 3870 * adding the entry to the journal when this information is available. nlink 3871 * should be the link count prior to the addition and mode is only required 3872 * to have the correct FMT. 3873 */ 3874static struct jaddref * 3875newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 3876 uint16_t mode) 3877{ 3878 struct jaddref *jaddref; 3879 3880 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 3881 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump)); 3882 jaddref->ja_state = ATTACHED; 3883 jaddref->ja_mkdir = NULL; 3884 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 3885 3886 return (jaddref); 3887} 3888 3889/* 3890 * Create a new free dependency for a freework. The caller is responsible 3891 * for adjusting the reference count when it has the lock held. The freedep 3892 * will track an outstanding bitmap write that will ultimately clear the 3893 * freework to continue. 3894 */ 3895static struct freedep * 3896newfreedep(struct freework *freework) 3897{ 3898 struct freedep *freedep; 3899 3900 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 3901 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 3902 freedep->fd_freework = freework; 3903 3904 return (freedep); 3905} 3906 3907/* 3908 * Free a freedep structure once the buffer it is linked to is written. If 3909 * this is the last reference to the freework schedule it for completion. 3910 */ 3911static void 3912free_freedep(freedep) 3913 struct freedep *freedep; 3914{ 3915 struct freework *freework; 3916 3917 freework = freedep->fd_freework; 3918 freework->fw_freeblks->fb_cgwait--; 3919 if (--freework->fw_ref == 0) 3920 freework_enqueue(freework); 3921 WORKITEM_FREE(freedep, D_FREEDEP); 3922} 3923 3924/* 3925 * Allocate a new freework structure that may be a level in an indirect 3926 * when parent is not NULL or a top level block when it is. The top level 3927 * freework structures are allocated without lk held and before the freeblks 3928 * is visible outside of softdep_setup_freeblocks(). 3929 */ 3930static struct freework * 3931newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 3932 struct ufsmount *ump; 3933 struct freeblks *freeblks; 3934 struct freework *parent; 3935 ufs_lbn_t lbn; 3936 ufs2_daddr_t nb; 3937 int frags; 3938 int off; 3939 int journal; 3940{ 3941 struct freework *freework; 3942 3943 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 3944 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 3945 freework->fw_state = ATTACHED; 3946 freework->fw_jnewblk = NULL; 3947 freework->fw_freeblks = freeblks; 3948 freework->fw_parent = parent; 3949 freework->fw_lbn = lbn; 3950 freework->fw_blkno = nb; 3951 freework->fw_frags = frags; 3952 freework->fw_indir = NULL; 3953 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR) 3954 ? 0 : NINDIR(ump->um_fs) + 1; 3955 freework->fw_start = freework->fw_off = off; 3956 if (journal) 3957 newjfreeblk(freeblks, lbn, nb, frags); 3958 if (parent == NULL) { 3959 ACQUIRE_LOCK(&lk); 3960 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 3961 freeblks->fb_ref++; 3962 FREE_LOCK(&lk); 3963 } 3964 3965 return (freework); 3966} 3967 3968/* 3969 * Eliminate a jfreeblk for a block that does not need journaling. 3970 */ 3971static void 3972cancel_jfreeblk(freeblks, blkno) 3973 struct freeblks *freeblks; 3974 ufs2_daddr_t blkno; 3975{ 3976 struct jfreeblk *jfreeblk; 3977 struct jblkdep *jblkdep; 3978 3979 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 3980 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 3981 continue; 3982 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 3983 if (jfreeblk->jf_blkno == blkno) 3984 break; 3985 } 3986 if (jblkdep == NULL) 3987 return; 3988 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 3989 free_jsegdep(jblkdep->jb_jsegdep); 3990 LIST_REMOVE(jblkdep, jb_deps); 3991 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 3992} 3993 3994/* 3995 * Allocate a new jfreeblk to journal top level block pointer when truncating 3996 * a file. The caller must add this to the worklist when lk is held. 3997 */ 3998static struct jfreeblk * 3999newjfreeblk(freeblks, lbn, blkno, frags) 4000 struct freeblks *freeblks; 4001 ufs_lbn_t lbn; 4002 ufs2_daddr_t blkno; 4003 int frags; 4004{ 4005 struct jfreeblk *jfreeblk; 4006 4007 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4008 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4009 freeblks->fb_list.wk_mp); 4010 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4011 jfreeblk->jf_dep.jb_freeblks = freeblks; 4012 jfreeblk->jf_ino = freeblks->fb_inum; 4013 jfreeblk->jf_lbn = lbn; 4014 jfreeblk->jf_blkno = blkno; 4015 jfreeblk->jf_frags = frags; 4016 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4017 4018 return (jfreeblk); 4019} 4020 4021/* 4022 * Allocate a new jtrunc to track a partial truncation. 4023 */ 4024static struct jtrunc * 4025newjtrunc(freeblks, size, extsize) 4026 struct freeblks *freeblks; 4027 off_t size; 4028 int extsize; 4029{ 4030 struct jtrunc *jtrunc; 4031 4032 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4033 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4034 freeblks->fb_list.wk_mp); 4035 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4036 jtrunc->jt_dep.jb_freeblks = freeblks; 4037 jtrunc->jt_ino = freeblks->fb_inum; 4038 jtrunc->jt_size = size; 4039 jtrunc->jt_extsize = extsize; 4040 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4041 4042 return (jtrunc); 4043} 4044 4045/* 4046 * If we're canceling a new bitmap we have to search for another ref 4047 * to move into the bmsafemap dep. This might be better expressed 4048 * with another structure. 4049 */ 4050static void 4051move_newblock_dep(jaddref, inodedep) 4052 struct jaddref *jaddref; 4053 struct inodedep *inodedep; 4054{ 4055 struct inoref *inoref; 4056 struct jaddref *jaddrefn; 4057 4058 jaddrefn = NULL; 4059 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4060 inoref = TAILQ_NEXT(inoref, if_deps)) { 4061 if ((jaddref->ja_state & NEWBLOCK) && 4062 inoref->if_list.wk_type == D_JADDREF) { 4063 jaddrefn = (struct jaddref *)inoref; 4064 break; 4065 } 4066 } 4067 if (jaddrefn == NULL) 4068 return; 4069 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4070 jaddrefn->ja_state |= jaddref->ja_state & 4071 (ATTACHED | UNDONE | NEWBLOCK); 4072 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4073 jaddref->ja_state |= ATTACHED; 4074 LIST_REMOVE(jaddref, ja_bmdeps); 4075 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4076 ja_bmdeps); 4077} 4078 4079/* 4080 * Cancel a jaddref either before it has been written or while it is being 4081 * written. This happens when a link is removed before the add reaches 4082 * the disk. The jaddref dependency is kept linked into the bmsafemap 4083 * and inode to prevent the link count or bitmap from reaching the disk 4084 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4085 * required. 4086 * 4087 * Returns 1 if the canceled addref requires journaling of the remove and 4088 * 0 otherwise. 4089 */ 4090static int 4091cancel_jaddref(jaddref, inodedep, wkhd) 4092 struct jaddref *jaddref; 4093 struct inodedep *inodedep; 4094 struct workhead *wkhd; 4095{ 4096 struct inoref *inoref; 4097 struct jsegdep *jsegdep; 4098 int needsj; 4099 4100 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4101 ("cancel_jaddref: Canceling complete jaddref")); 4102 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4103 needsj = 1; 4104 else 4105 needsj = 0; 4106 if (inodedep == NULL) 4107 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4108 0, &inodedep) == 0) 4109 panic("cancel_jaddref: Lost inodedep"); 4110 /* 4111 * We must adjust the nlink of any reference operation that follows 4112 * us so that it is consistent with the in-memory reference. This 4113 * ensures that inode nlink rollbacks always have the correct link. 4114 */ 4115 if (needsj == 0) { 4116 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4117 inoref = TAILQ_NEXT(inoref, if_deps)) { 4118 if (inoref->if_state & GOINGAWAY) 4119 break; 4120 inoref->if_nlink--; 4121 } 4122 } 4123 jsegdep = inoref_jseg(&jaddref->ja_ref); 4124 if (jaddref->ja_state & NEWBLOCK) 4125 move_newblock_dep(jaddref, inodedep); 4126 wake_worklist(&jaddref->ja_list); 4127 jaddref->ja_mkdir = NULL; 4128 if (jaddref->ja_state & INPROGRESS) { 4129 jaddref->ja_state &= ~INPROGRESS; 4130 WORKLIST_REMOVE(&jaddref->ja_list); 4131 jwork_insert(wkhd, jsegdep); 4132 } else { 4133 free_jsegdep(jsegdep); 4134 if (jaddref->ja_state & DEPCOMPLETE) 4135 remove_from_journal(&jaddref->ja_list); 4136 } 4137 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4138 /* 4139 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4140 * can arrange for them to be freed with the bitmap. Otherwise we 4141 * no longer need this addref attached to the inoreflst and it 4142 * will incorrectly adjust nlink if we leave it. 4143 */ 4144 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4145 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4146 if_deps); 4147 jaddref->ja_state |= COMPLETE; 4148 free_jaddref(jaddref); 4149 return (needsj); 4150 } 4151 /* 4152 * Leave the head of the list for jsegdeps for fast merging. 4153 */ 4154 if (LIST_FIRST(wkhd) != NULL) { 4155 jaddref->ja_state |= ONWORKLIST; 4156 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4157 } else 4158 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4159 4160 return (needsj); 4161} 4162 4163/* 4164 * Attempt to free a jaddref structure when some work completes. This 4165 * should only succeed once the entry is written and all dependencies have 4166 * been notified. 4167 */ 4168static void 4169free_jaddref(jaddref) 4170 struct jaddref *jaddref; 4171{ 4172 4173 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4174 return; 4175 if (jaddref->ja_ref.if_jsegdep) 4176 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4177 jaddref, jaddref->ja_state); 4178 if (jaddref->ja_state & NEWBLOCK) 4179 LIST_REMOVE(jaddref, ja_bmdeps); 4180 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4181 panic("free_jaddref: Bad state %p(0x%X)", 4182 jaddref, jaddref->ja_state); 4183 if (jaddref->ja_mkdir != NULL) 4184 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4185 WORKITEM_FREE(jaddref, D_JADDREF); 4186} 4187 4188/* 4189 * Free a jremref structure once it has been written or discarded. 4190 */ 4191static void 4192free_jremref(jremref) 4193 struct jremref *jremref; 4194{ 4195 4196 if (jremref->jr_ref.if_jsegdep) 4197 free_jsegdep(jremref->jr_ref.if_jsegdep); 4198 if (jremref->jr_state & INPROGRESS) 4199 panic("free_jremref: IO still pending"); 4200 WORKITEM_FREE(jremref, D_JREMREF); 4201} 4202 4203/* 4204 * Free a jnewblk structure. 4205 */ 4206static void 4207free_jnewblk(jnewblk) 4208 struct jnewblk *jnewblk; 4209{ 4210 4211 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4212 return; 4213 LIST_REMOVE(jnewblk, jn_deps); 4214 if (jnewblk->jn_dep != NULL) 4215 panic("free_jnewblk: Dependency still attached."); 4216 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4217} 4218 4219/* 4220 * Cancel a jnewblk which has been been made redundant by frag extension. 4221 */ 4222static void 4223cancel_jnewblk(jnewblk, wkhd) 4224 struct jnewblk *jnewblk; 4225 struct workhead *wkhd; 4226{ 4227 struct jsegdep *jsegdep; 4228 4229 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4230 jsegdep = jnewblk->jn_jsegdep; 4231 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4232 panic("cancel_jnewblk: Invalid state"); 4233 jnewblk->jn_jsegdep = NULL; 4234 jnewblk->jn_dep = NULL; 4235 jnewblk->jn_state |= GOINGAWAY; 4236 if (jnewblk->jn_state & INPROGRESS) { 4237 jnewblk->jn_state &= ~INPROGRESS; 4238 WORKLIST_REMOVE(&jnewblk->jn_list); 4239 jwork_insert(wkhd, jsegdep); 4240 } else { 4241 free_jsegdep(jsegdep); 4242 remove_from_journal(&jnewblk->jn_list); 4243 } 4244 wake_worklist(&jnewblk->jn_list); 4245 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4246} 4247 4248static void 4249free_jblkdep(jblkdep) 4250 struct jblkdep *jblkdep; 4251{ 4252 4253 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4254 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4255 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4256 WORKITEM_FREE(jblkdep, D_JTRUNC); 4257 else 4258 panic("free_jblkdep: Unexpected type %s", 4259 TYPENAME(jblkdep->jb_list.wk_type)); 4260} 4261 4262/* 4263 * Free a single jseg once it is no longer referenced in memory or on 4264 * disk. Reclaim journal blocks and dependencies waiting for the segment 4265 * to disappear. 4266 */ 4267static void 4268free_jseg(jseg, jblocks) 4269 struct jseg *jseg; 4270 struct jblocks *jblocks; 4271{ 4272 struct freework *freework; 4273 4274 /* 4275 * Free freework structures that were lingering to indicate freed 4276 * indirect blocks that forced journal write ordering on reallocate. 4277 */ 4278 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4279 indirblk_remove(freework); 4280 if (jblocks->jb_oldestseg == jseg) 4281 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4282 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4283 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4284 KASSERT(LIST_EMPTY(&jseg->js_entries), 4285 ("free_jseg: Freed jseg has valid entries.")); 4286 WORKITEM_FREE(jseg, D_JSEG); 4287} 4288 4289/* 4290 * Free all jsegs that meet the criteria for being reclaimed and update 4291 * oldestseg. 4292 */ 4293static void 4294free_jsegs(jblocks) 4295 struct jblocks *jblocks; 4296{ 4297 struct jseg *jseg; 4298 4299 /* 4300 * Free only those jsegs which have none allocated before them to 4301 * preserve the journal space ordering. 4302 */ 4303 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4304 /* 4305 * Only reclaim space when nothing depends on this journal 4306 * set and another set has written that it is no longer 4307 * valid. 4308 */ 4309 if (jseg->js_refs != 0) { 4310 jblocks->jb_oldestseg = jseg; 4311 return; 4312 } 4313 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4314 break; 4315 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4316 break; 4317 /* 4318 * We can free jsegs that didn't write entries when 4319 * oldestwrseq == js_seq. 4320 */ 4321 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4322 jseg->js_cnt != 0) 4323 break; 4324 free_jseg(jseg, jblocks); 4325 } 4326 /* 4327 * If we exited the loop above we still must discover the 4328 * oldest valid segment. 4329 */ 4330 if (jseg) 4331 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4332 jseg = TAILQ_NEXT(jseg, js_next)) 4333 if (jseg->js_refs != 0) 4334 break; 4335 jblocks->jb_oldestseg = jseg; 4336 /* 4337 * The journal has no valid records but some jsegs may still be 4338 * waiting on oldestwrseq to advance. We force a small record 4339 * out to permit these lingering records to be reclaimed. 4340 */ 4341 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4342 jblocks->jb_needseg = 1; 4343} 4344 4345/* 4346 * Release one reference to a jseg and free it if the count reaches 0. This 4347 * should eventually reclaim journal space as well. 4348 */ 4349static void 4350rele_jseg(jseg) 4351 struct jseg *jseg; 4352{ 4353 4354 KASSERT(jseg->js_refs > 0, 4355 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4356 if (--jseg->js_refs != 0) 4357 return; 4358 free_jsegs(jseg->js_jblocks); 4359} 4360 4361/* 4362 * Release a jsegdep and decrement the jseg count. 4363 */ 4364static void 4365free_jsegdep(jsegdep) 4366 struct jsegdep *jsegdep; 4367{ 4368 4369 if (jsegdep->jd_seg) 4370 rele_jseg(jsegdep->jd_seg); 4371 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4372} 4373 4374/* 4375 * Wait for a journal item to make it to disk. Initiate journal processing 4376 * if required. 4377 */ 4378static int 4379jwait(wk, waitfor) 4380 struct worklist *wk; 4381 int waitfor; 4382{ 4383 4384 /* 4385 * Blocking journal waits cause slow synchronous behavior. Record 4386 * stats on the frequency of these blocking operations. 4387 */ 4388 if (waitfor == MNT_WAIT) { 4389 stat_journal_wait++; 4390 switch (wk->wk_type) { 4391 case D_JREMREF: 4392 case D_JMVREF: 4393 stat_jwait_filepage++; 4394 break; 4395 case D_JTRUNC: 4396 case D_JFREEBLK: 4397 stat_jwait_freeblks++; 4398 break; 4399 case D_JNEWBLK: 4400 stat_jwait_newblk++; 4401 break; 4402 case D_JADDREF: 4403 stat_jwait_inode++; 4404 break; 4405 default: 4406 break; 4407 } 4408 } 4409 /* 4410 * If IO has not started we process the journal. We can't mark the 4411 * worklist item as IOWAITING because we drop the lock while 4412 * processing the journal and the worklist entry may be freed after 4413 * this point. The caller may call back in and re-issue the request. 4414 */ 4415 if ((wk->wk_state & INPROGRESS) == 0) { 4416 softdep_process_journal(wk->wk_mp, wk, waitfor); 4417 if (waitfor != MNT_WAIT) 4418 return (EBUSY); 4419 return (0); 4420 } 4421 if (waitfor != MNT_WAIT) 4422 return (EBUSY); 4423 wait_worklist(wk, "jwait"); 4424 return (0); 4425} 4426 4427/* 4428 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4429 * appropriate. This is a convenience function to reduce duplicate code 4430 * for the setup and revert functions below. 4431 */ 4432static struct inodedep * 4433inodedep_lookup_ip(ip) 4434 struct inode *ip; 4435{ 4436 struct inodedep *inodedep; 4437 int dflags; 4438 4439 KASSERT(ip->i_nlink >= ip->i_effnlink, 4440 ("inodedep_lookup_ip: bad delta")); 4441 dflags = DEPALLOC; 4442 if (IS_SNAPSHOT(ip)) 4443 dflags |= NODELAY; 4444 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, 4445 &inodedep); 4446 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4447 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4448 4449 return (inodedep); 4450} 4451 4452/* 4453 * Called prior to creating a new inode and linking it to a directory. The 4454 * jaddref structure must already be allocated by softdep_setup_inomapdep 4455 * and it is discovered here so we can initialize the mode and update 4456 * nlinkdelta. 4457 */ 4458void 4459softdep_setup_create(dp, ip) 4460 struct inode *dp; 4461 struct inode *ip; 4462{ 4463 struct inodedep *inodedep; 4464 struct jaddref *jaddref; 4465 struct vnode *dvp; 4466 4467 KASSERT(ip->i_nlink == 1, 4468 ("softdep_setup_create: Invalid link count.")); 4469 dvp = ITOV(dp); 4470 ACQUIRE_LOCK(&lk); 4471 inodedep = inodedep_lookup_ip(ip); 4472 if (DOINGSUJ(dvp)) { 4473 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4474 inoreflst); 4475 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4476 ("softdep_setup_create: No addref structure present.")); 4477 } 4478 softdep_prelink(dvp, NULL); 4479 FREE_LOCK(&lk); 4480} 4481 4482/* 4483 * Create a jaddref structure to track the addition of a DOTDOT link when 4484 * we are reparenting an inode as part of a rename. This jaddref will be 4485 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4486 * non-journaling softdep. 4487 */ 4488void 4489softdep_setup_dotdot_link(dp, ip) 4490 struct inode *dp; 4491 struct inode *ip; 4492{ 4493 struct inodedep *inodedep; 4494 struct jaddref *jaddref; 4495 struct vnode *dvp; 4496 struct vnode *vp; 4497 4498 dvp = ITOV(dp); 4499 vp = ITOV(ip); 4500 jaddref = NULL; 4501 /* 4502 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4503 * is used as a normal link would be. 4504 */ 4505 if (DOINGSUJ(dvp)) 4506 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4507 dp->i_effnlink - 1, dp->i_mode); 4508 ACQUIRE_LOCK(&lk); 4509 inodedep = inodedep_lookup_ip(dp); 4510 if (jaddref) 4511 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4512 if_deps); 4513 softdep_prelink(dvp, ITOV(ip)); 4514 FREE_LOCK(&lk); 4515} 4516 4517/* 4518 * Create a jaddref structure to track a new link to an inode. The directory 4519 * offset is not known until softdep_setup_directory_add or 4520 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4521 * softdep. 4522 */ 4523void 4524softdep_setup_link(dp, ip) 4525 struct inode *dp; 4526 struct inode *ip; 4527{ 4528 struct inodedep *inodedep; 4529 struct jaddref *jaddref; 4530 struct vnode *dvp; 4531 4532 dvp = ITOV(dp); 4533 jaddref = NULL; 4534 if (DOINGSUJ(dvp)) 4535 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4536 ip->i_mode); 4537 ACQUIRE_LOCK(&lk); 4538 inodedep = inodedep_lookup_ip(ip); 4539 if (jaddref) 4540 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4541 if_deps); 4542 softdep_prelink(dvp, ITOV(ip)); 4543 FREE_LOCK(&lk); 4544} 4545 4546/* 4547 * Called to create the jaddref structures to track . and .. references as 4548 * well as lookup and further initialize the incomplete jaddref created 4549 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4550 * nlinkdelta for non-journaling softdep. 4551 */ 4552void 4553softdep_setup_mkdir(dp, ip) 4554 struct inode *dp; 4555 struct inode *ip; 4556{ 4557 struct inodedep *inodedep; 4558 struct jaddref *dotdotaddref; 4559 struct jaddref *dotaddref; 4560 struct jaddref *jaddref; 4561 struct vnode *dvp; 4562 4563 dvp = ITOV(dp); 4564 dotaddref = dotdotaddref = NULL; 4565 if (DOINGSUJ(dvp)) { 4566 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4567 ip->i_mode); 4568 dotaddref->ja_state |= MKDIR_BODY; 4569 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4570 dp->i_effnlink - 1, dp->i_mode); 4571 dotdotaddref->ja_state |= MKDIR_PARENT; 4572 } 4573 ACQUIRE_LOCK(&lk); 4574 inodedep = inodedep_lookup_ip(ip); 4575 if (DOINGSUJ(dvp)) { 4576 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4577 inoreflst); 4578 KASSERT(jaddref != NULL, 4579 ("softdep_setup_mkdir: No addref structure present.")); 4580 KASSERT(jaddref->ja_parent == dp->i_number, 4581 ("softdep_setup_mkdir: bad parent %ju", 4582 (uintmax_t)jaddref->ja_parent)); 4583 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4584 if_deps); 4585 } 4586 inodedep = inodedep_lookup_ip(dp); 4587 if (DOINGSUJ(dvp)) 4588 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4589 &dotdotaddref->ja_ref, if_deps); 4590 softdep_prelink(ITOV(dp), NULL); 4591 FREE_LOCK(&lk); 4592} 4593 4594/* 4595 * Called to track nlinkdelta of the inode and parent directories prior to 4596 * unlinking a directory. 4597 */ 4598void 4599softdep_setup_rmdir(dp, ip) 4600 struct inode *dp; 4601 struct inode *ip; 4602{ 4603 struct vnode *dvp; 4604 4605 dvp = ITOV(dp); 4606 ACQUIRE_LOCK(&lk); 4607 (void) inodedep_lookup_ip(ip); 4608 (void) inodedep_lookup_ip(dp); 4609 softdep_prelink(dvp, ITOV(ip)); 4610 FREE_LOCK(&lk); 4611} 4612 4613/* 4614 * Called to track nlinkdelta of the inode and parent directories prior to 4615 * unlink. 4616 */ 4617void 4618softdep_setup_unlink(dp, ip) 4619 struct inode *dp; 4620 struct inode *ip; 4621{ 4622 struct vnode *dvp; 4623 4624 dvp = ITOV(dp); 4625 ACQUIRE_LOCK(&lk); 4626 (void) inodedep_lookup_ip(ip); 4627 (void) inodedep_lookup_ip(dp); 4628 softdep_prelink(dvp, ITOV(ip)); 4629 FREE_LOCK(&lk); 4630} 4631 4632/* 4633 * Called to release the journal structures created by a failed non-directory 4634 * creation. Adjusts nlinkdelta for non-journaling softdep. 4635 */ 4636void 4637softdep_revert_create(dp, ip) 4638 struct inode *dp; 4639 struct inode *ip; 4640{ 4641 struct inodedep *inodedep; 4642 struct jaddref *jaddref; 4643 struct vnode *dvp; 4644 4645 dvp = ITOV(dp); 4646 ACQUIRE_LOCK(&lk); 4647 inodedep = inodedep_lookup_ip(ip); 4648 if (DOINGSUJ(dvp)) { 4649 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4650 inoreflst); 4651 KASSERT(jaddref->ja_parent == dp->i_number, 4652 ("softdep_revert_create: addref parent mismatch")); 4653 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4654 } 4655 FREE_LOCK(&lk); 4656} 4657 4658/* 4659 * Called to release the journal structures created by a failed dotdot link 4660 * creation. Adjusts nlinkdelta for non-journaling softdep. 4661 */ 4662void 4663softdep_revert_dotdot_link(dp, ip) 4664 struct inode *dp; 4665 struct inode *ip; 4666{ 4667 struct inodedep *inodedep; 4668 struct jaddref *jaddref; 4669 struct vnode *dvp; 4670 4671 dvp = ITOV(dp); 4672 ACQUIRE_LOCK(&lk); 4673 inodedep = inodedep_lookup_ip(dp); 4674 if (DOINGSUJ(dvp)) { 4675 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4676 inoreflst); 4677 KASSERT(jaddref->ja_parent == ip->i_number, 4678 ("softdep_revert_dotdot_link: addref parent mismatch")); 4679 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4680 } 4681 FREE_LOCK(&lk); 4682} 4683 4684/* 4685 * Called to release the journal structures created by a failed link 4686 * addition. Adjusts nlinkdelta for non-journaling softdep. 4687 */ 4688void 4689softdep_revert_link(dp, ip) 4690 struct inode *dp; 4691 struct inode *ip; 4692{ 4693 struct inodedep *inodedep; 4694 struct jaddref *jaddref; 4695 struct vnode *dvp; 4696 4697 dvp = ITOV(dp); 4698 ACQUIRE_LOCK(&lk); 4699 inodedep = inodedep_lookup_ip(ip); 4700 if (DOINGSUJ(dvp)) { 4701 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4702 inoreflst); 4703 KASSERT(jaddref->ja_parent == dp->i_number, 4704 ("softdep_revert_link: addref parent mismatch")); 4705 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4706 } 4707 FREE_LOCK(&lk); 4708} 4709 4710/* 4711 * Called to release the journal structures created by a failed mkdir 4712 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4713 */ 4714void 4715softdep_revert_mkdir(dp, ip) 4716 struct inode *dp; 4717 struct inode *ip; 4718{ 4719 struct inodedep *inodedep; 4720 struct jaddref *jaddref; 4721 struct jaddref *dotaddref; 4722 struct vnode *dvp; 4723 4724 dvp = ITOV(dp); 4725 4726 ACQUIRE_LOCK(&lk); 4727 inodedep = inodedep_lookup_ip(dp); 4728 if (DOINGSUJ(dvp)) { 4729 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4730 inoreflst); 4731 KASSERT(jaddref->ja_parent == ip->i_number, 4732 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4733 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4734 } 4735 inodedep = inodedep_lookup_ip(ip); 4736 if (DOINGSUJ(dvp)) { 4737 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4738 inoreflst); 4739 KASSERT(jaddref->ja_parent == dp->i_number, 4740 ("softdep_revert_mkdir: addref parent mismatch")); 4741 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4742 inoreflst, if_deps); 4743 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4744 KASSERT(dotaddref->ja_parent == ip->i_number, 4745 ("softdep_revert_mkdir: dot addref parent mismatch")); 4746 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4747 } 4748 FREE_LOCK(&lk); 4749} 4750 4751/* 4752 * Called to correct nlinkdelta after a failed rmdir. 4753 */ 4754void 4755softdep_revert_rmdir(dp, ip) 4756 struct inode *dp; 4757 struct inode *ip; 4758{ 4759 4760 ACQUIRE_LOCK(&lk); 4761 (void) inodedep_lookup_ip(ip); 4762 (void) inodedep_lookup_ip(dp); 4763 FREE_LOCK(&lk); 4764} 4765 4766/* 4767 * Protecting the freemaps (or bitmaps). 4768 * 4769 * To eliminate the need to execute fsck before mounting a filesystem 4770 * after a power failure, one must (conservatively) guarantee that the 4771 * on-disk copy of the bitmaps never indicate that a live inode or block is 4772 * free. So, when a block or inode is allocated, the bitmap should be 4773 * updated (on disk) before any new pointers. When a block or inode is 4774 * freed, the bitmap should not be updated until all pointers have been 4775 * reset. The latter dependency is handled by the delayed de-allocation 4776 * approach described below for block and inode de-allocation. The former 4777 * dependency is handled by calling the following procedure when a block or 4778 * inode is allocated. When an inode is allocated an "inodedep" is created 4779 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4780 * Each "inodedep" is also inserted into the hash indexing structure so 4781 * that any additional link additions can be made dependent on the inode 4782 * allocation. 4783 * 4784 * The ufs filesystem maintains a number of free block counts (e.g., per 4785 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4786 * in addition to the bitmaps. These counts are used to improve efficiency 4787 * during allocation and therefore must be consistent with the bitmaps. 4788 * There is no convenient way to guarantee post-crash consistency of these 4789 * counts with simple update ordering, for two main reasons: (1) The counts 4790 * and bitmaps for a single cylinder group block are not in the same disk 4791 * sector. If a disk write is interrupted (e.g., by power failure), one may 4792 * be written and the other not. (2) Some of the counts are located in the 4793 * superblock rather than the cylinder group block. So, we focus our soft 4794 * updates implementation on protecting the bitmaps. When mounting a 4795 * filesystem, we recompute the auxiliary counts from the bitmaps. 4796 */ 4797 4798/* 4799 * Called just after updating the cylinder group block to allocate an inode. 4800 */ 4801void 4802softdep_setup_inomapdep(bp, ip, newinum, mode) 4803 struct buf *bp; /* buffer for cylgroup block with inode map */ 4804 struct inode *ip; /* inode related to allocation */ 4805 ino_t newinum; /* new inode number being allocated */ 4806 int mode; 4807{ 4808 struct inodedep *inodedep; 4809 struct bmsafemap *bmsafemap; 4810 struct jaddref *jaddref; 4811 struct mount *mp; 4812 struct fs *fs; 4813 4814 mp = UFSTOVFS(ip->i_ump); 4815 fs = ip->i_ump->um_fs; 4816 jaddref = NULL; 4817 4818 /* 4819 * Allocate the journal reference add structure so that the bitmap 4820 * can be dependent on it. 4821 */ 4822 if (MOUNTEDSUJ(mp)) { 4823 jaddref = newjaddref(ip, newinum, 0, 0, mode); 4824 jaddref->ja_state |= NEWBLOCK; 4825 } 4826 4827 /* 4828 * Create a dependency for the newly allocated inode. 4829 * Panic if it already exists as something is seriously wrong. 4830 * Otherwise add it to the dependency list for the buffer holding 4831 * the cylinder group map from which it was allocated. 4832 * 4833 * We have to preallocate a bmsafemap entry in case it is needed 4834 * in bmsafemap_lookup since once we allocate the inodedep, we 4835 * have to finish initializing it before we can FREE_LOCK(). 4836 * By preallocating, we avoid FREE_LOCK() while doing a malloc 4837 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 4838 * creating the inodedep as it can be freed during the time 4839 * that we FREE_LOCK() while allocating the inodedep. We must 4840 * call workitem_alloc() before entering the locked section as 4841 * it also acquires the lock and we must avoid trying doing so 4842 * recursively. 4843 */ 4844 bmsafemap = malloc(sizeof(struct bmsafemap), 4845 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 4846 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 4847 ACQUIRE_LOCK(&lk); 4848 if ((inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep))) 4849 panic("softdep_setup_inomapdep: dependency %p for new" 4850 "inode already exists", inodedep); 4851 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 4852 if (jaddref) { 4853 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 4854 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4855 if_deps); 4856 } else { 4857 inodedep->id_state |= ONDEPLIST; 4858 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 4859 } 4860 inodedep->id_bmsafemap = bmsafemap; 4861 inodedep->id_state &= ~DEPCOMPLETE; 4862 FREE_LOCK(&lk); 4863} 4864 4865/* 4866 * Called just after updating the cylinder group block to 4867 * allocate block or fragment. 4868 */ 4869void 4870softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 4871 struct buf *bp; /* buffer for cylgroup block with block map */ 4872 struct mount *mp; /* filesystem doing allocation */ 4873 ufs2_daddr_t newblkno; /* number of newly allocated block */ 4874 int frags; /* Number of fragments. */ 4875 int oldfrags; /* Previous number of fragments for extend. */ 4876{ 4877 struct newblk *newblk; 4878 struct bmsafemap *bmsafemap; 4879 struct jnewblk *jnewblk; 4880 struct fs *fs; 4881 4882 fs = VFSTOUFS(mp)->um_fs; 4883 jnewblk = NULL; 4884 /* 4885 * Create a dependency for the newly allocated block. 4886 * Add it to the dependency list for the buffer holding 4887 * the cylinder group map from which it was allocated. 4888 */ 4889 if (MOUNTEDSUJ(mp)) { 4890 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 4891 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 4892 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 4893 jnewblk->jn_state = ATTACHED; 4894 jnewblk->jn_blkno = newblkno; 4895 jnewblk->jn_frags = frags; 4896 jnewblk->jn_oldfrags = oldfrags; 4897#ifdef SUJ_DEBUG 4898 { 4899 struct cg *cgp; 4900 uint8_t *blksfree; 4901 long bno; 4902 int i; 4903 4904 cgp = (struct cg *)bp->b_data; 4905 blksfree = cg_blksfree(cgp); 4906 bno = dtogd(fs, jnewblk->jn_blkno); 4907 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 4908 i++) { 4909 if (isset(blksfree, bno + i)) 4910 panic("softdep_setup_blkmapdep: " 4911 "free fragment %d from %d-%d " 4912 "state 0x%X dep %p", i, 4913 jnewblk->jn_oldfrags, 4914 jnewblk->jn_frags, 4915 jnewblk->jn_state, 4916 jnewblk->jn_dep); 4917 } 4918 } 4919#endif 4920 } 4921 4922 CTR3(KTR_SUJ, 4923 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 4924 newblkno, frags, oldfrags); 4925 ACQUIRE_LOCK(&lk); 4926 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 4927 panic("softdep_setup_blkmapdep: found block"); 4928 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 4929 dtog(fs, newblkno), NULL); 4930 if (jnewblk) { 4931 jnewblk->jn_dep = (struct worklist *)newblk; 4932 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 4933 } else { 4934 newblk->nb_state |= ONDEPLIST; 4935 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 4936 } 4937 newblk->nb_bmsafemap = bmsafemap; 4938 newblk->nb_jnewblk = jnewblk; 4939 FREE_LOCK(&lk); 4940} 4941 4942#define BMSAFEMAP_HASH(fs, cg) \ 4943 (&bmsafemap_hashtbl[((((register_t)(fs)) >> 13) + (cg)) & bmsafemap_hash]) 4944 4945static int 4946bmsafemap_find(bmsafemaphd, mp, cg, bmsafemapp) 4947 struct bmsafemap_hashhead *bmsafemaphd; 4948 struct mount *mp; 4949 int cg; 4950 struct bmsafemap **bmsafemapp; 4951{ 4952 struct bmsafemap *bmsafemap; 4953 4954 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 4955 if (bmsafemap->sm_list.wk_mp == mp && bmsafemap->sm_cg == cg) 4956 break; 4957 if (bmsafemap) { 4958 *bmsafemapp = bmsafemap; 4959 return (1); 4960 } 4961 *bmsafemapp = NULL; 4962 4963 return (0); 4964} 4965 4966/* 4967 * Find the bmsafemap associated with a cylinder group buffer. 4968 * If none exists, create one. The buffer must be locked when 4969 * this routine is called and this routine must be called with 4970 * the softdep lock held. To avoid giving up the lock while 4971 * allocating a new bmsafemap, a preallocated bmsafemap may be 4972 * provided. If it is provided but not needed, it is freed. 4973 */ 4974static struct bmsafemap * 4975bmsafemap_lookup(mp, bp, cg, newbmsafemap) 4976 struct mount *mp; 4977 struct buf *bp; 4978 int cg; 4979 struct bmsafemap *newbmsafemap; 4980{ 4981 struct bmsafemap_hashhead *bmsafemaphd; 4982 struct bmsafemap *bmsafemap, *collision; 4983 struct worklist *wk; 4984 struct fs *fs; 4985 4986 mtx_assert(&lk, MA_OWNED); 4987 if (bp) 4988 LIST_FOREACH(wk, &bp->b_dep, wk_list) 4989 if (wk->wk_type == D_BMSAFEMAP) { 4990 if (newbmsafemap) 4991 WORKITEM_FREE(newbmsafemap,D_BMSAFEMAP); 4992 return (WK_BMSAFEMAP(wk)); 4993 } 4994 fs = VFSTOUFS(mp)->um_fs; 4995 bmsafemaphd = BMSAFEMAP_HASH(fs, cg); 4996 if (bmsafemap_find(bmsafemaphd, mp, cg, &bmsafemap) == 1) { 4997 if (newbmsafemap) 4998 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 4999 return (bmsafemap); 5000 } 5001 if (newbmsafemap) { 5002 bmsafemap = newbmsafemap; 5003 } else { 5004 FREE_LOCK(&lk); 5005 bmsafemap = malloc(sizeof(struct bmsafemap), 5006 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5007 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5008 ACQUIRE_LOCK(&lk); 5009 } 5010 bmsafemap->sm_buf = bp; 5011 LIST_INIT(&bmsafemap->sm_inodedephd); 5012 LIST_INIT(&bmsafemap->sm_inodedepwr); 5013 LIST_INIT(&bmsafemap->sm_newblkhd); 5014 LIST_INIT(&bmsafemap->sm_newblkwr); 5015 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5016 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5017 LIST_INIT(&bmsafemap->sm_freehd); 5018 LIST_INIT(&bmsafemap->sm_freewr); 5019 if (bmsafemap_find(bmsafemaphd, mp, cg, &collision) == 1) { 5020 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5021 return (collision); 5022 } 5023 bmsafemap->sm_cg = cg; 5024 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5025 LIST_INSERT_HEAD(&VFSTOUFS(mp)->softdep_dirtycg, bmsafemap, sm_next); 5026 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5027 return (bmsafemap); 5028} 5029 5030/* 5031 * Direct block allocation dependencies. 5032 * 5033 * When a new block is allocated, the corresponding disk locations must be 5034 * initialized (with zeros or new data) before the on-disk inode points to 5035 * them. Also, the freemap from which the block was allocated must be 5036 * updated (on disk) before the inode's pointer. These two dependencies are 5037 * independent of each other and are needed for all file blocks and indirect 5038 * blocks that are pointed to directly by the inode. Just before the 5039 * "in-core" version of the inode is updated with a newly allocated block 5040 * number, a procedure (below) is called to setup allocation dependency 5041 * structures. These structures are removed when the corresponding 5042 * dependencies are satisfied or when the block allocation becomes obsolete 5043 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5044 * fragment that gets upgraded). All of these cases are handled in 5045 * procedures described later. 5046 * 5047 * When a file extension causes a fragment to be upgraded, either to a larger 5048 * fragment or to a full block, the on-disk location may change (if the 5049 * previous fragment could not simply be extended). In this case, the old 5050 * fragment must be de-allocated, but not until after the inode's pointer has 5051 * been updated. In most cases, this is handled by later procedures, which 5052 * will construct a "freefrag" structure to be added to the workitem queue 5053 * when the inode update is complete (or obsolete). The main exception to 5054 * this is when an allocation occurs while a pending allocation dependency 5055 * (for the same block pointer) remains. This case is handled in the main 5056 * allocation dependency setup procedure by immediately freeing the 5057 * unreferenced fragments. 5058 */ 5059void 5060softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5061 struct inode *ip; /* inode to which block is being added */ 5062 ufs_lbn_t off; /* block pointer within inode */ 5063 ufs2_daddr_t newblkno; /* disk block number being added */ 5064 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5065 long newsize; /* size of new block */ 5066 long oldsize; /* size of new block */ 5067 struct buf *bp; /* bp for allocated block */ 5068{ 5069 struct allocdirect *adp, *oldadp; 5070 struct allocdirectlst *adphead; 5071 struct freefrag *freefrag; 5072 struct inodedep *inodedep; 5073 struct pagedep *pagedep; 5074 struct jnewblk *jnewblk; 5075 struct newblk *newblk; 5076 struct mount *mp; 5077 ufs_lbn_t lbn; 5078 5079 lbn = bp->b_lblkno; 5080 mp = UFSTOVFS(ip->i_ump); 5081 if (oldblkno && oldblkno != newblkno) 5082 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5083 else 5084 freefrag = NULL; 5085 5086 CTR6(KTR_SUJ, 5087 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5088 "off %jd newsize %ld oldsize %d", 5089 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5090 ACQUIRE_LOCK(&lk); 5091 if (off >= NDADDR) { 5092 if (lbn > 0) 5093 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5094 lbn, off); 5095 /* allocating an indirect block */ 5096 if (oldblkno != 0) 5097 panic("softdep_setup_allocdirect: non-zero indir"); 5098 } else { 5099 if (off != lbn) 5100 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5101 lbn, off); 5102 /* 5103 * Allocating a direct block. 5104 * 5105 * If we are allocating a directory block, then we must 5106 * allocate an associated pagedep to track additions and 5107 * deletions. 5108 */ 5109 if ((ip->i_mode & IFMT) == IFDIR) 5110 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5111 &pagedep); 5112 } 5113 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5114 panic("softdep_setup_allocdirect: lost block"); 5115 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5116 ("softdep_setup_allocdirect: newblk already initialized")); 5117 /* 5118 * Convert the newblk to an allocdirect. 5119 */ 5120 newblk->nb_list.wk_type = D_ALLOCDIRECT; 5121 adp = (struct allocdirect *)newblk; 5122 newblk->nb_freefrag = freefrag; 5123 adp->ad_offset = off; 5124 adp->ad_oldblkno = oldblkno; 5125 adp->ad_newsize = newsize; 5126 adp->ad_oldsize = oldsize; 5127 5128 /* 5129 * Finish initializing the journal. 5130 */ 5131 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5132 jnewblk->jn_ino = ip->i_number; 5133 jnewblk->jn_lbn = lbn; 5134 add_to_journal(&jnewblk->jn_list); 5135 } 5136 if (freefrag && freefrag->ff_jdep != NULL && 5137 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5138 add_to_journal(freefrag->ff_jdep); 5139 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5140 adp->ad_inodedep = inodedep; 5141 5142 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5143 /* 5144 * The list of allocdirects must be kept in sorted and ascending 5145 * order so that the rollback routines can quickly determine the 5146 * first uncommitted block (the size of the file stored on disk 5147 * ends at the end of the lowest committed fragment, or if there 5148 * are no fragments, at the end of the highest committed block). 5149 * Since files generally grow, the typical case is that the new 5150 * block is to be added at the end of the list. We speed this 5151 * special case by checking against the last allocdirect in the 5152 * list before laboriously traversing the list looking for the 5153 * insertion point. 5154 */ 5155 adphead = &inodedep->id_newinoupdt; 5156 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5157 if (oldadp == NULL || oldadp->ad_offset <= off) { 5158 /* insert at end of list */ 5159 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5160 if (oldadp != NULL && oldadp->ad_offset == off) 5161 allocdirect_merge(adphead, adp, oldadp); 5162 FREE_LOCK(&lk); 5163 return; 5164 } 5165 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5166 if (oldadp->ad_offset >= off) 5167 break; 5168 } 5169 if (oldadp == NULL) 5170 panic("softdep_setup_allocdirect: lost entry"); 5171 /* insert in middle of list */ 5172 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5173 if (oldadp->ad_offset == off) 5174 allocdirect_merge(adphead, adp, oldadp); 5175 5176 FREE_LOCK(&lk); 5177} 5178 5179/* 5180 * Merge a newer and older journal record to be stored either in a 5181 * newblock or freefrag. This handles aggregating journal records for 5182 * fragment allocation into a second record as well as replacing a 5183 * journal free with an aborted journal allocation. A segment for the 5184 * oldest record will be placed on wkhd if it has been written. If not 5185 * the segment for the newer record will suffice. 5186 */ 5187static struct worklist * 5188jnewblk_merge(new, old, wkhd) 5189 struct worklist *new; 5190 struct worklist *old; 5191 struct workhead *wkhd; 5192{ 5193 struct jnewblk *njnewblk; 5194 struct jnewblk *jnewblk; 5195 5196 /* Handle NULLs to simplify callers. */ 5197 if (new == NULL) 5198 return (old); 5199 if (old == NULL) 5200 return (new); 5201 /* Replace a jfreefrag with a jnewblk. */ 5202 if (new->wk_type == D_JFREEFRAG) { 5203 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5204 panic("jnewblk_merge: blkno mismatch: %p, %p", 5205 old, new); 5206 cancel_jfreefrag(WK_JFREEFRAG(new)); 5207 return (old); 5208 } 5209 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5210 panic("jnewblk_merge: Bad type: old %d new %d\n", 5211 old->wk_type, new->wk_type); 5212 /* 5213 * Handle merging of two jnewblk records that describe 5214 * different sets of fragments in the same block. 5215 */ 5216 jnewblk = WK_JNEWBLK(old); 5217 njnewblk = WK_JNEWBLK(new); 5218 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5219 panic("jnewblk_merge: Merging disparate blocks."); 5220 /* 5221 * The record may be rolled back in the cg. 5222 */ 5223 if (jnewblk->jn_state & UNDONE) { 5224 jnewblk->jn_state &= ~UNDONE; 5225 njnewblk->jn_state |= UNDONE; 5226 njnewblk->jn_state &= ~ATTACHED; 5227 } 5228 /* 5229 * We modify the newer addref and free the older so that if neither 5230 * has been written the most up-to-date copy will be on disk. If 5231 * both have been written but rolled back we only temporarily need 5232 * one of them to fix the bits when the cg write completes. 5233 */ 5234 jnewblk->jn_state |= ATTACHED | COMPLETE; 5235 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5236 cancel_jnewblk(jnewblk, wkhd); 5237 WORKLIST_REMOVE(&jnewblk->jn_list); 5238 free_jnewblk(jnewblk); 5239 return (new); 5240} 5241 5242/* 5243 * Replace an old allocdirect dependency with a newer one. 5244 * This routine must be called with splbio interrupts blocked. 5245 */ 5246static void 5247allocdirect_merge(adphead, newadp, oldadp) 5248 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5249 struct allocdirect *newadp; /* allocdirect being added */ 5250 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5251{ 5252 struct worklist *wk; 5253 struct freefrag *freefrag; 5254 5255 freefrag = NULL; 5256 mtx_assert(&lk, MA_OWNED); 5257 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5258 newadp->ad_oldsize != oldadp->ad_newsize || 5259 newadp->ad_offset >= NDADDR) 5260 panic("%s %jd != new %jd || old size %ld != new %ld", 5261 "allocdirect_merge: old blkno", 5262 (intmax_t)newadp->ad_oldblkno, 5263 (intmax_t)oldadp->ad_newblkno, 5264 newadp->ad_oldsize, oldadp->ad_newsize); 5265 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5266 newadp->ad_oldsize = oldadp->ad_oldsize; 5267 /* 5268 * If the old dependency had a fragment to free or had never 5269 * previously had a block allocated, then the new dependency 5270 * can immediately post its freefrag and adopt the old freefrag. 5271 * This action is done by swapping the freefrag dependencies. 5272 * The new dependency gains the old one's freefrag, and the 5273 * old one gets the new one and then immediately puts it on 5274 * the worklist when it is freed by free_newblk. It is 5275 * not possible to do this swap when the old dependency had a 5276 * non-zero size but no previous fragment to free. This condition 5277 * arises when the new block is an extension of the old block. 5278 * Here, the first part of the fragment allocated to the new 5279 * dependency is part of the block currently claimed on disk by 5280 * the old dependency, so cannot legitimately be freed until the 5281 * conditions for the new dependency are fulfilled. 5282 */ 5283 freefrag = newadp->ad_freefrag; 5284 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5285 newadp->ad_freefrag = oldadp->ad_freefrag; 5286 oldadp->ad_freefrag = freefrag; 5287 } 5288 /* 5289 * If we are tracking a new directory-block allocation, 5290 * move it from the old allocdirect to the new allocdirect. 5291 */ 5292 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5293 WORKLIST_REMOVE(wk); 5294 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5295 panic("allocdirect_merge: extra newdirblk"); 5296 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5297 } 5298 TAILQ_REMOVE(adphead, oldadp, ad_next); 5299 /* 5300 * We need to move any journal dependencies over to the freefrag 5301 * that releases this block if it exists. Otherwise we are 5302 * extending an existing block and we'll wait until that is 5303 * complete to release the journal space and extend the 5304 * new journal to cover this old space as well. 5305 */ 5306 if (freefrag == NULL) { 5307 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5308 panic("allocdirect_merge: %jd != %jd", 5309 oldadp->ad_newblkno, newadp->ad_newblkno); 5310 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5311 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5312 &oldadp->ad_block.nb_jnewblk->jn_list, 5313 &newadp->ad_block.nb_jwork); 5314 oldadp->ad_block.nb_jnewblk = NULL; 5315 cancel_newblk(&oldadp->ad_block, NULL, 5316 &newadp->ad_block.nb_jwork); 5317 } else { 5318 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5319 &freefrag->ff_list, &freefrag->ff_jwork); 5320 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5321 &freefrag->ff_jwork); 5322 } 5323 free_newblk(&oldadp->ad_block); 5324} 5325 5326/* 5327 * Allocate a jfreefrag structure to journal a single block free. 5328 */ 5329static struct jfreefrag * 5330newjfreefrag(freefrag, ip, blkno, size, lbn) 5331 struct freefrag *freefrag; 5332 struct inode *ip; 5333 ufs2_daddr_t blkno; 5334 long size; 5335 ufs_lbn_t lbn; 5336{ 5337 struct jfreefrag *jfreefrag; 5338 struct fs *fs; 5339 5340 fs = ip->i_fs; 5341 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5342 M_SOFTDEP_FLAGS); 5343 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump)); 5344 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5345 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5346 jfreefrag->fr_ino = ip->i_number; 5347 jfreefrag->fr_lbn = lbn; 5348 jfreefrag->fr_blkno = blkno; 5349 jfreefrag->fr_frags = numfrags(fs, size); 5350 jfreefrag->fr_freefrag = freefrag; 5351 5352 return (jfreefrag); 5353} 5354 5355/* 5356 * Allocate a new freefrag structure. 5357 */ 5358static struct freefrag * 5359newfreefrag(ip, blkno, size, lbn) 5360 struct inode *ip; 5361 ufs2_daddr_t blkno; 5362 long size; 5363 ufs_lbn_t lbn; 5364{ 5365 struct freefrag *freefrag; 5366 struct fs *fs; 5367 5368 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5369 ip->i_number, blkno, size, lbn); 5370 fs = ip->i_fs; 5371 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5372 panic("newfreefrag: frag size"); 5373 freefrag = malloc(sizeof(struct freefrag), 5374 M_FREEFRAG, M_SOFTDEP_FLAGS); 5375 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 5376 freefrag->ff_state = ATTACHED; 5377 LIST_INIT(&freefrag->ff_jwork); 5378 freefrag->ff_inum = ip->i_number; 5379 freefrag->ff_vtype = ITOV(ip)->v_type; 5380 freefrag->ff_blkno = blkno; 5381 freefrag->ff_fragsize = size; 5382 5383 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) { 5384 freefrag->ff_jdep = (struct worklist *) 5385 newjfreefrag(freefrag, ip, blkno, size, lbn); 5386 } else { 5387 freefrag->ff_state |= DEPCOMPLETE; 5388 freefrag->ff_jdep = NULL; 5389 } 5390 5391 return (freefrag); 5392} 5393 5394/* 5395 * This workitem de-allocates fragments that were replaced during 5396 * file block allocation. 5397 */ 5398static void 5399handle_workitem_freefrag(freefrag) 5400 struct freefrag *freefrag; 5401{ 5402 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5403 struct workhead wkhd; 5404 5405 CTR3(KTR_SUJ, 5406 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5407 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5408 /* 5409 * It would be illegal to add new completion items to the 5410 * freefrag after it was schedule to be done so it must be 5411 * safe to modify the list head here. 5412 */ 5413 LIST_INIT(&wkhd); 5414 ACQUIRE_LOCK(&lk); 5415 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5416 /* 5417 * If the journal has not been written we must cancel it here. 5418 */ 5419 if (freefrag->ff_jdep) { 5420 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5421 panic("handle_workitem_freefrag: Unexpected type %d\n", 5422 freefrag->ff_jdep->wk_type); 5423 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5424 } 5425 FREE_LOCK(&lk); 5426 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5427 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5428 ACQUIRE_LOCK(&lk); 5429 WORKITEM_FREE(freefrag, D_FREEFRAG); 5430 FREE_LOCK(&lk); 5431} 5432 5433/* 5434 * Set up a dependency structure for an external attributes data block. 5435 * This routine follows much of the structure of softdep_setup_allocdirect. 5436 * See the description of softdep_setup_allocdirect above for details. 5437 */ 5438void 5439softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5440 struct inode *ip; 5441 ufs_lbn_t off; 5442 ufs2_daddr_t newblkno; 5443 ufs2_daddr_t oldblkno; 5444 long newsize; 5445 long oldsize; 5446 struct buf *bp; 5447{ 5448 struct allocdirect *adp, *oldadp; 5449 struct allocdirectlst *adphead; 5450 struct freefrag *freefrag; 5451 struct inodedep *inodedep; 5452 struct jnewblk *jnewblk; 5453 struct newblk *newblk; 5454 struct mount *mp; 5455 ufs_lbn_t lbn; 5456 5457 if (off >= NXADDR) 5458 panic("softdep_setup_allocext: lbn %lld > NXADDR", 5459 (long long)off); 5460 5461 lbn = bp->b_lblkno; 5462 mp = UFSTOVFS(ip->i_ump); 5463 if (oldblkno && oldblkno != newblkno) 5464 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5465 else 5466 freefrag = NULL; 5467 5468 ACQUIRE_LOCK(&lk); 5469 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5470 panic("softdep_setup_allocext: lost block"); 5471 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5472 ("softdep_setup_allocext: newblk already initialized")); 5473 /* 5474 * Convert the newblk to an allocdirect. 5475 */ 5476 newblk->nb_list.wk_type = D_ALLOCDIRECT; 5477 adp = (struct allocdirect *)newblk; 5478 newblk->nb_freefrag = freefrag; 5479 adp->ad_offset = off; 5480 adp->ad_oldblkno = oldblkno; 5481 adp->ad_newsize = newsize; 5482 adp->ad_oldsize = oldsize; 5483 adp->ad_state |= EXTDATA; 5484 5485 /* 5486 * Finish initializing the journal. 5487 */ 5488 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5489 jnewblk->jn_ino = ip->i_number; 5490 jnewblk->jn_lbn = lbn; 5491 add_to_journal(&jnewblk->jn_list); 5492 } 5493 if (freefrag && freefrag->ff_jdep != NULL && 5494 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5495 add_to_journal(freefrag->ff_jdep); 5496 inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep); 5497 adp->ad_inodedep = inodedep; 5498 5499 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5500 /* 5501 * The list of allocdirects must be kept in sorted and ascending 5502 * order so that the rollback routines can quickly determine the 5503 * first uncommitted block (the size of the file stored on disk 5504 * ends at the end of the lowest committed fragment, or if there 5505 * are no fragments, at the end of the highest committed block). 5506 * Since files generally grow, the typical case is that the new 5507 * block is to be added at the end of the list. We speed this 5508 * special case by checking against the last allocdirect in the 5509 * list before laboriously traversing the list looking for the 5510 * insertion point. 5511 */ 5512 adphead = &inodedep->id_newextupdt; 5513 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5514 if (oldadp == NULL || oldadp->ad_offset <= off) { 5515 /* insert at end of list */ 5516 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5517 if (oldadp != NULL && oldadp->ad_offset == off) 5518 allocdirect_merge(adphead, adp, oldadp); 5519 FREE_LOCK(&lk); 5520 return; 5521 } 5522 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5523 if (oldadp->ad_offset >= off) 5524 break; 5525 } 5526 if (oldadp == NULL) 5527 panic("softdep_setup_allocext: lost entry"); 5528 /* insert in middle of list */ 5529 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5530 if (oldadp->ad_offset == off) 5531 allocdirect_merge(adphead, adp, oldadp); 5532 FREE_LOCK(&lk); 5533} 5534 5535/* 5536 * Indirect block allocation dependencies. 5537 * 5538 * The same dependencies that exist for a direct block also exist when 5539 * a new block is allocated and pointed to by an entry in a block of 5540 * indirect pointers. The undo/redo states described above are also 5541 * used here. Because an indirect block contains many pointers that 5542 * may have dependencies, a second copy of the entire in-memory indirect 5543 * block is kept. The buffer cache copy is always completely up-to-date. 5544 * The second copy, which is used only as a source for disk writes, 5545 * contains only the safe pointers (i.e., those that have no remaining 5546 * update dependencies). The second copy is freed when all pointers 5547 * are safe. The cache is not allowed to replace indirect blocks with 5548 * pending update dependencies. If a buffer containing an indirect 5549 * block with dependencies is written, these routines will mark it 5550 * dirty again. It can only be successfully written once all the 5551 * dependencies are removed. The ffs_fsync routine in conjunction with 5552 * softdep_sync_metadata work together to get all the dependencies 5553 * removed so that a file can be successfully written to disk. Three 5554 * procedures are used when setting up indirect block pointer 5555 * dependencies. The division is necessary because of the organization 5556 * of the "balloc" routine and because of the distinction between file 5557 * pages and file metadata blocks. 5558 */ 5559 5560/* 5561 * Allocate a new allocindir structure. 5562 */ 5563static struct allocindir * 5564newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5565 struct inode *ip; /* inode for file being extended */ 5566 int ptrno; /* offset of pointer in indirect block */ 5567 ufs2_daddr_t newblkno; /* disk block number being added */ 5568 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5569 ufs_lbn_t lbn; 5570{ 5571 struct newblk *newblk; 5572 struct allocindir *aip; 5573 struct freefrag *freefrag; 5574 struct jnewblk *jnewblk; 5575 5576 if (oldblkno) 5577 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn); 5578 else 5579 freefrag = NULL; 5580 ACQUIRE_LOCK(&lk); 5581 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0) 5582 panic("new_allocindir: lost block"); 5583 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5584 ("newallocindir: newblk already initialized")); 5585 newblk->nb_list.wk_type = D_ALLOCINDIR; 5586 newblk->nb_freefrag = freefrag; 5587 aip = (struct allocindir *)newblk; 5588 aip->ai_offset = ptrno; 5589 aip->ai_oldblkno = oldblkno; 5590 aip->ai_lbn = lbn; 5591 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5592 jnewblk->jn_ino = ip->i_number; 5593 jnewblk->jn_lbn = lbn; 5594 add_to_journal(&jnewblk->jn_list); 5595 } 5596 if (freefrag && freefrag->ff_jdep != NULL && 5597 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5598 add_to_journal(freefrag->ff_jdep); 5599 return (aip); 5600} 5601 5602/* 5603 * Called just before setting an indirect block pointer 5604 * to a newly allocated file page. 5605 */ 5606void 5607softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5608 struct inode *ip; /* inode for file being extended */ 5609 ufs_lbn_t lbn; /* allocated block number within file */ 5610 struct buf *bp; /* buffer with indirect blk referencing page */ 5611 int ptrno; /* offset of pointer in indirect block */ 5612 ufs2_daddr_t newblkno; /* disk block number being added */ 5613 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5614 struct buf *nbp; /* buffer holding allocated page */ 5615{ 5616 struct inodedep *inodedep; 5617 struct freefrag *freefrag; 5618 struct allocindir *aip; 5619 struct pagedep *pagedep; 5620 struct mount *mp; 5621 int dflags; 5622 5623 if (lbn != nbp->b_lblkno) 5624 panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5625 lbn, bp->b_lblkno); 5626 CTR4(KTR_SUJ, 5627 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5628 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5629 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5630 mp = UFSTOVFS(ip->i_ump); 5631 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5632 dflags = DEPALLOC; 5633 if (IS_SNAPSHOT(ip)) 5634 dflags |= NODELAY; 5635 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 5636 /* 5637 * If we are allocating a directory page, then we must 5638 * allocate an associated pagedep to track additions and 5639 * deletions. 5640 */ 5641 if ((ip->i_mode & IFMT) == IFDIR) 5642 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5643 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5644 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5645 FREE_LOCK(&lk); 5646 if (freefrag) 5647 handle_workitem_freefrag(freefrag); 5648} 5649 5650/* 5651 * Called just before setting an indirect block pointer to a 5652 * newly allocated indirect block. 5653 */ 5654void 5655softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5656 struct buf *nbp; /* newly allocated indirect block */ 5657 struct inode *ip; /* inode for file being extended */ 5658 struct buf *bp; /* indirect block referencing allocated block */ 5659 int ptrno; /* offset of pointer in indirect block */ 5660 ufs2_daddr_t newblkno; /* disk block number being added */ 5661{ 5662 struct inodedep *inodedep; 5663 struct allocindir *aip; 5664 ufs_lbn_t lbn; 5665 int dflags; 5666 5667 CTR3(KTR_SUJ, 5668 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5669 ip->i_number, newblkno, ptrno); 5670 lbn = nbp->b_lblkno; 5671 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5672 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5673 dflags = DEPALLOC; 5674 if (IS_SNAPSHOT(ip)) 5675 dflags |= NODELAY; 5676 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 5677 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5678 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5679 panic("softdep_setup_allocindir_meta: Block already existed"); 5680 FREE_LOCK(&lk); 5681} 5682 5683static void 5684indirdep_complete(indirdep) 5685 struct indirdep *indirdep; 5686{ 5687 struct allocindir *aip; 5688 5689 LIST_REMOVE(indirdep, ir_next); 5690 indirdep->ir_state |= DEPCOMPLETE; 5691 5692 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5693 LIST_REMOVE(aip, ai_next); 5694 free_newblk(&aip->ai_block); 5695 } 5696 /* 5697 * If this indirdep is not attached to a buf it was simply waiting 5698 * on completion to clear completehd. free_indirdep() asserts 5699 * that nothing is dangling. 5700 */ 5701 if ((indirdep->ir_state & ONWORKLIST) == 0) 5702 free_indirdep(indirdep); 5703} 5704 5705static struct indirdep * 5706indirdep_lookup(mp, ip, bp) 5707 struct mount *mp; 5708 struct inode *ip; 5709 struct buf *bp; 5710{ 5711 struct indirdep *indirdep, *newindirdep; 5712 struct newblk *newblk; 5713 struct worklist *wk; 5714 struct fs *fs; 5715 ufs2_daddr_t blkno; 5716 5717 mtx_assert(&lk, MA_OWNED); 5718 indirdep = NULL; 5719 newindirdep = NULL; 5720 fs = ip->i_fs; 5721 for (;;) { 5722 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5723 if (wk->wk_type != D_INDIRDEP) 5724 continue; 5725 indirdep = WK_INDIRDEP(wk); 5726 break; 5727 } 5728 /* Found on the buffer worklist, no new structure to free. */ 5729 if (indirdep != NULL && newindirdep == NULL) 5730 return (indirdep); 5731 if (indirdep != NULL && newindirdep != NULL) 5732 panic("indirdep_lookup: simultaneous create"); 5733 /* None found on the buffer and a new structure is ready. */ 5734 if (indirdep == NULL && newindirdep != NULL) 5735 break; 5736 /* None found and no new structure available. */ 5737 FREE_LOCK(&lk); 5738 newindirdep = malloc(sizeof(struct indirdep), 5739 M_INDIRDEP, M_SOFTDEP_FLAGS); 5740 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5741 newindirdep->ir_state = ATTACHED; 5742 if (ip->i_ump->um_fstype == UFS1) 5743 newindirdep->ir_state |= UFS1FMT; 5744 TAILQ_INIT(&newindirdep->ir_trunc); 5745 newindirdep->ir_saveddata = NULL; 5746 LIST_INIT(&newindirdep->ir_deplisthd); 5747 LIST_INIT(&newindirdep->ir_donehd); 5748 LIST_INIT(&newindirdep->ir_writehd); 5749 LIST_INIT(&newindirdep->ir_completehd); 5750 if (bp->b_blkno == bp->b_lblkno) { 5751 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5752 NULL, NULL); 5753 bp->b_blkno = blkno; 5754 } 5755 newindirdep->ir_freeblks = NULL; 5756 newindirdep->ir_savebp = 5757 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5758 newindirdep->ir_bp = bp; 5759 BUF_KERNPROC(newindirdep->ir_savebp); 5760 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5761 ACQUIRE_LOCK(&lk); 5762 } 5763 indirdep = newindirdep; 5764 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5765 /* 5766 * If the block is not yet allocated we don't set DEPCOMPLETE so 5767 * that we don't free dependencies until the pointers are valid. 5768 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5769 * than using the hash. 5770 */ 5771 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5772 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5773 else 5774 indirdep->ir_state |= DEPCOMPLETE; 5775 return (indirdep); 5776} 5777 5778/* 5779 * Called to finish the allocation of the "aip" allocated 5780 * by one of the two routines above. 5781 */ 5782static struct freefrag * 5783setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 5784 struct buf *bp; /* in-memory copy of the indirect block */ 5785 struct inode *ip; /* inode for file being extended */ 5786 struct inodedep *inodedep; /* Inodedep for ip */ 5787 struct allocindir *aip; /* allocindir allocated by the above routines */ 5788 ufs_lbn_t lbn; /* Logical block number for this block. */ 5789{ 5790 struct fs *fs; 5791 struct indirdep *indirdep; 5792 struct allocindir *oldaip; 5793 struct freefrag *freefrag; 5794 struct mount *mp; 5795 5796 mtx_assert(&lk, MA_OWNED); 5797 mp = UFSTOVFS(ip->i_ump); 5798 fs = ip->i_fs; 5799 if (bp->b_lblkno >= 0) 5800 panic("setup_allocindir_phase2: not indir blk"); 5801 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 5802 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 5803 indirdep = indirdep_lookup(mp, ip, bp); 5804 KASSERT(indirdep->ir_savebp != NULL, 5805 ("setup_allocindir_phase2 NULL ir_savebp")); 5806 aip->ai_indirdep = indirdep; 5807 /* 5808 * Check for an unwritten dependency for this indirect offset. If 5809 * there is, merge the old dependency into the new one. This happens 5810 * as a result of reallocblk only. 5811 */ 5812 freefrag = NULL; 5813 if (aip->ai_oldblkno != 0) { 5814 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 5815 if (oldaip->ai_offset == aip->ai_offset) { 5816 freefrag = allocindir_merge(aip, oldaip); 5817 goto done; 5818 } 5819 } 5820 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 5821 if (oldaip->ai_offset == aip->ai_offset) { 5822 freefrag = allocindir_merge(aip, oldaip); 5823 goto done; 5824 } 5825 } 5826 } 5827done: 5828 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 5829 return (freefrag); 5830} 5831 5832/* 5833 * Merge two allocindirs which refer to the same block. Move newblock 5834 * dependencies and setup the freefrags appropriately. 5835 */ 5836static struct freefrag * 5837allocindir_merge(aip, oldaip) 5838 struct allocindir *aip; 5839 struct allocindir *oldaip; 5840{ 5841 struct freefrag *freefrag; 5842 struct worklist *wk; 5843 5844 if (oldaip->ai_newblkno != aip->ai_oldblkno) 5845 panic("allocindir_merge: blkno"); 5846 aip->ai_oldblkno = oldaip->ai_oldblkno; 5847 freefrag = aip->ai_freefrag; 5848 aip->ai_freefrag = oldaip->ai_freefrag; 5849 oldaip->ai_freefrag = NULL; 5850 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 5851 /* 5852 * If we are tracking a new directory-block allocation, 5853 * move it from the old allocindir to the new allocindir. 5854 */ 5855 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 5856 WORKLIST_REMOVE(wk); 5857 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 5858 panic("allocindir_merge: extra newdirblk"); 5859 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 5860 } 5861 /* 5862 * We can skip journaling for this freefrag and just complete 5863 * any pending journal work for the allocindir that is being 5864 * removed after the freefrag completes. 5865 */ 5866 if (freefrag->ff_jdep) 5867 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 5868 LIST_REMOVE(oldaip, ai_next); 5869 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 5870 &freefrag->ff_list, &freefrag->ff_jwork); 5871 free_newblk(&oldaip->ai_block); 5872 5873 return (freefrag); 5874} 5875 5876static inline void 5877setup_freedirect(freeblks, ip, i, needj) 5878 struct freeblks *freeblks; 5879 struct inode *ip; 5880 int i; 5881 int needj; 5882{ 5883 ufs2_daddr_t blkno; 5884 int frags; 5885 5886 blkno = DIP(ip, i_db[i]); 5887 if (blkno == 0) 5888 return; 5889 DIP_SET(ip, i_db[i], 0); 5890 frags = sblksize(ip->i_fs, ip->i_size, i); 5891 frags = numfrags(ip->i_fs, frags); 5892 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj); 5893} 5894 5895static inline void 5896setup_freeext(freeblks, ip, i, needj) 5897 struct freeblks *freeblks; 5898 struct inode *ip; 5899 int i; 5900 int needj; 5901{ 5902 ufs2_daddr_t blkno; 5903 int frags; 5904 5905 blkno = ip->i_din2->di_extb[i]; 5906 if (blkno == 0) 5907 return; 5908 ip->i_din2->di_extb[i] = 0; 5909 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i); 5910 frags = numfrags(ip->i_fs, frags); 5911 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 5912} 5913 5914static inline void 5915setup_freeindir(freeblks, ip, i, lbn, needj) 5916 struct freeblks *freeblks; 5917 struct inode *ip; 5918 int i; 5919 ufs_lbn_t lbn; 5920 int needj; 5921{ 5922 ufs2_daddr_t blkno; 5923 5924 blkno = DIP(ip, i_ib[i]); 5925 if (blkno == 0) 5926 return; 5927 DIP_SET(ip, i_ib[i], 0); 5928 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag, 5929 0, needj); 5930} 5931 5932static inline struct freeblks * 5933newfreeblks(mp, ip) 5934 struct mount *mp; 5935 struct inode *ip; 5936{ 5937 struct freeblks *freeblks; 5938 5939 freeblks = malloc(sizeof(struct freeblks), 5940 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 5941 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 5942 LIST_INIT(&freeblks->fb_jblkdephd); 5943 LIST_INIT(&freeblks->fb_jwork); 5944 freeblks->fb_ref = 0; 5945 freeblks->fb_cgwait = 0; 5946 freeblks->fb_state = ATTACHED; 5947 freeblks->fb_uid = ip->i_uid; 5948 freeblks->fb_inum = ip->i_number; 5949 freeblks->fb_vtype = ITOV(ip)->v_type; 5950 freeblks->fb_modrev = DIP(ip, i_modrev); 5951 freeblks->fb_devvp = ip->i_devvp; 5952 freeblks->fb_chkcnt = 0; 5953 freeblks->fb_len = 0; 5954 5955 return (freeblks); 5956} 5957 5958static void 5959trunc_indirdep(indirdep, freeblks, bp, off) 5960 struct indirdep *indirdep; 5961 struct freeblks *freeblks; 5962 struct buf *bp; 5963 int off; 5964{ 5965 struct allocindir *aip, *aipn; 5966 5967 /* 5968 * The first set of allocindirs won't be in savedbp. 5969 */ 5970 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 5971 if (aip->ai_offset > off) 5972 cancel_allocindir(aip, bp, freeblks, 1); 5973 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 5974 if (aip->ai_offset > off) 5975 cancel_allocindir(aip, bp, freeblks, 1); 5976 /* 5977 * These will exist in savedbp. 5978 */ 5979 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 5980 if (aip->ai_offset > off) 5981 cancel_allocindir(aip, NULL, freeblks, 0); 5982 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 5983 if (aip->ai_offset > off) 5984 cancel_allocindir(aip, NULL, freeblks, 0); 5985} 5986 5987/* 5988 * Follow the chain of indirects down to lastlbn creating a freework 5989 * structure for each. This will be used to start indir_trunc() at 5990 * the right offset and create the journal records for the parrtial 5991 * truncation. A second step will handle the truncated dependencies. 5992 */ 5993static int 5994setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 5995 struct freeblks *freeblks; 5996 struct inode *ip; 5997 ufs_lbn_t lbn; 5998 ufs_lbn_t lastlbn; 5999 ufs2_daddr_t blkno; 6000{ 6001 struct indirdep *indirdep; 6002 struct indirdep *indirn; 6003 struct freework *freework; 6004 struct newblk *newblk; 6005 struct mount *mp; 6006 struct buf *bp; 6007 uint8_t *start; 6008 uint8_t *end; 6009 ufs_lbn_t lbnadd; 6010 int level; 6011 int error; 6012 int off; 6013 6014 6015 freework = NULL; 6016 if (blkno == 0) 6017 return (0); 6018 mp = freeblks->fb_list.wk_mp; 6019 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6020 if ((bp->b_flags & B_CACHE) == 0) { 6021 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6022 bp->b_iocmd = BIO_READ; 6023 bp->b_flags &= ~B_INVAL; 6024 bp->b_ioflags &= ~BIO_ERROR; 6025 vfs_busy_pages(bp, 0); 6026 bp->b_iooffset = dbtob(bp->b_blkno); 6027 bstrategy(bp); 6028 curthread->td_ru.ru_inblock++; 6029 error = bufwait(bp); 6030 if (error) { 6031 brelse(bp); 6032 return (error); 6033 } 6034 } 6035 level = lbn_level(lbn); 6036 lbnadd = lbn_offset(ip->i_fs, level); 6037 /* 6038 * Compute the offset of the last block we want to keep. Store 6039 * in the freework the first block we want to completely free. 6040 */ 6041 off = (lastlbn - -(lbn + level)) / lbnadd; 6042 if (off + 1 == NINDIR(ip->i_fs)) 6043 goto nowork; 6044 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1, 6045 0); 6046 /* 6047 * Link the freework into the indirdep. This will prevent any new 6048 * allocations from proceeding until we are finished with the 6049 * truncate and the block is written. 6050 */ 6051 ACQUIRE_LOCK(&lk); 6052 indirdep = indirdep_lookup(mp, ip, bp); 6053 if (indirdep->ir_freeblks) 6054 panic("setup_trunc_indir: indirdep already truncated."); 6055 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6056 freework->fw_indir = indirdep; 6057 /* 6058 * Cancel any allocindirs that will not make it to disk. 6059 * We have to do this for all copies of the indirdep that 6060 * live on this newblk. 6061 */ 6062 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6063 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk); 6064 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6065 trunc_indirdep(indirn, freeblks, bp, off); 6066 } else 6067 trunc_indirdep(indirdep, freeblks, bp, off); 6068 FREE_LOCK(&lk); 6069 /* 6070 * Creation is protected by the buf lock. The saveddata is only 6071 * needed if a full truncation follows a partial truncation but it 6072 * is difficult to allocate in that case so we fetch it anyway. 6073 */ 6074 if (indirdep->ir_saveddata == NULL) 6075 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6076 M_SOFTDEP_FLAGS); 6077nowork: 6078 /* Fetch the blkno of the child and the zero start offset. */ 6079 if (ip->i_ump->um_fstype == UFS1) { 6080 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6081 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6082 } else { 6083 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6084 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6085 } 6086 if (freework) { 6087 /* Zero the truncated pointers. */ 6088 end = bp->b_data + bp->b_bcount; 6089 bzero(start, end - start); 6090 bdwrite(bp); 6091 } else 6092 bqrelse(bp); 6093 if (level == 0) 6094 return (0); 6095 lbn++; /* adjust level */ 6096 lbn -= (off * lbnadd); 6097 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6098} 6099 6100/* 6101 * Complete the partial truncation of an indirect block setup by 6102 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6103 * copy and writes them to disk before the freeblks is allowed to complete. 6104 */ 6105static void 6106complete_trunc_indir(freework) 6107 struct freework *freework; 6108{ 6109 struct freework *fwn; 6110 struct indirdep *indirdep; 6111 struct buf *bp; 6112 uintptr_t start; 6113 int count; 6114 6115 indirdep = freework->fw_indir; 6116 for (;;) { 6117 bp = indirdep->ir_bp; 6118 /* See if the block was discarded. */ 6119 if (bp == NULL) 6120 break; 6121 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6122 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6123 break; 6124 if (BUF_LOCK(bp, 6125 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, &lk) == 0) 6126 BUF_UNLOCK(bp); 6127 ACQUIRE_LOCK(&lk); 6128 } 6129 mtx_assert(&lk, MA_OWNED); 6130 freework->fw_state |= DEPCOMPLETE; 6131 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6132 /* 6133 * Zero the pointers in the saved copy. 6134 */ 6135 if (indirdep->ir_state & UFS1FMT) 6136 start = sizeof(ufs1_daddr_t); 6137 else 6138 start = sizeof(ufs2_daddr_t); 6139 start *= freework->fw_start; 6140 count = indirdep->ir_savebp->b_bcount - start; 6141 start += (uintptr_t)indirdep->ir_savebp->b_data; 6142 bzero((char *)start, count); 6143 /* 6144 * We need to start the next truncation in the list if it has not 6145 * been started yet. 6146 */ 6147 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6148 if (fwn != NULL) { 6149 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6150 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6151 if ((fwn->fw_state & ONWORKLIST) == 0) 6152 freework_enqueue(fwn); 6153 } 6154 /* 6155 * If bp is NULL the block was fully truncated, restore 6156 * the saved block list otherwise free it if it is no 6157 * longer needed. 6158 */ 6159 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6160 if (bp == NULL) 6161 bcopy(indirdep->ir_saveddata, 6162 indirdep->ir_savebp->b_data, 6163 indirdep->ir_savebp->b_bcount); 6164 free(indirdep->ir_saveddata, M_INDIRDEP); 6165 indirdep->ir_saveddata = NULL; 6166 } 6167 /* 6168 * When bp is NULL there is a full truncation pending. We 6169 * must wait for this full truncation to be journaled before 6170 * we can release this freework because the disk pointers will 6171 * never be written as zero. 6172 */ 6173 if (bp == NULL) { 6174 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6175 handle_written_freework(freework); 6176 else 6177 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6178 &freework->fw_list); 6179 } else { 6180 /* Complete when the real copy is written. */ 6181 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6182 BUF_UNLOCK(bp); 6183 } 6184} 6185 6186/* 6187 * Calculate the number of blocks we are going to release where datablocks 6188 * is the current total and length is the new file size. 6189 */ 6190ufs2_daddr_t 6191blkcount(fs, datablocks, length) 6192 struct fs *fs; 6193 ufs2_daddr_t datablocks; 6194 off_t length; 6195{ 6196 off_t totblks, numblks; 6197 6198 totblks = 0; 6199 numblks = howmany(length, fs->fs_bsize); 6200 if (numblks <= NDADDR) { 6201 totblks = howmany(length, fs->fs_fsize); 6202 goto out; 6203 } 6204 totblks = blkstofrags(fs, numblks); 6205 numblks -= NDADDR; 6206 /* 6207 * Count all single, then double, then triple indirects required. 6208 * Subtracting one indirects worth of blocks for each pass 6209 * acknowledges one of each pointed to by the inode. 6210 */ 6211 for (;;) { 6212 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6213 numblks -= NINDIR(fs); 6214 if (numblks <= 0) 6215 break; 6216 numblks = howmany(numblks, NINDIR(fs)); 6217 } 6218out: 6219 totblks = fsbtodb(fs, totblks); 6220 /* 6221 * Handle sparse files. We can't reclaim more blocks than the inode 6222 * references. We will correct it later in handle_complete_freeblks() 6223 * when we know the real count. 6224 */ 6225 if (totblks > datablocks) 6226 return (0); 6227 return (datablocks - totblks); 6228} 6229 6230/* 6231 * Handle freeblocks for journaled softupdate filesystems. 6232 * 6233 * Contrary to normal softupdates, we must preserve the block pointers in 6234 * indirects until their subordinates are free. This is to avoid journaling 6235 * every block that is freed which may consume more space than the journal 6236 * itself. The recovery program will see the free block journals at the 6237 * base of the truncated area and traverse them to reclaim space. The 6238 * pointers in the inode may be cleared immediately after the journal 6239 * records are written because each direct and indirect pointer in the 6240 * inode is recorded in a journal. This permits full truncation to proceed 6241 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6242 * 6243 * The algorithm is as follows: 6244 * 1) Traverse the in-memory state and create journal entries to release 6245 * the relevant blocks and full indirect trees. 6246 * 2) Traverse the indirect block chain adding partial truncation freework 6247 * records to indirects in the path to lastlbn. The freework will 6248 * prevent new allocation dependencies from being satisfied in this 6249 * indirect until the truncation completes. 6250 * 3) Read and lock the inode block, performing an update with the new size 6251 * and pointers. This prevents truncated data from becoming valid on 6252 * disk through step 4. 6253 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6254 * eliminate journal work for those records that do not require it. 6255 * 5) Schedule the journal records to be written followed by the inode block. 6256 * 6) Allocate any necessary frags for the end of file. 6257 * 7) Zero any partially truncated blocks. 6258 * 6259 * From this truncation proceeds asynchronously using the freework and 6260 * indir_trunc machinery. The file will not be extended again into a 6261 * partially truncated indirect block until all work is completed but 6262 * the normal dependency mechanism ensures that it is rolled back/forward 6263 * as appropriate. Further truncation may occur without delay and is 6264 * serialized in indir_trunc(). 6265 */ 6266void 6267softdep_journal_freeblocks(ip, cred, length, flags) 6268 struct inode *ip; /* The inode whose length is to be reduced */ 6269 struct ucred *cred; 6270 off_t length; /* The new length for the file */ 6271 int flags; /* IO_EXT and/or IO_NORMAL */ 6272{ 6273 struct freeblks *freeblks, *fbn; 6274 struct worklist *wk, *wkn; 6275 struct inodedep *inodedep; 6276 struct jblkdep *jblkdep; 6277 struct allocdirect *adp, *adpn; 6278 struct fs *fs; 6279 struct buf *bp; 6280 struct vnode *vp; 6281 struct mount *mp; 6282 ufs2_daddr_t extblocks, datablocks; 6283 ufs_lbn_t tmpval, lbn, lastlbn; 6284 int frags, lastoff, iboff, allocblock, needj, dflags, error, i; 6285 6286 fs = ip->i_fs; 6287 mp = UFSTOVFS(ip->i_ump); 6288 vp = ITOV(ip); 6289 needj = 1; 6290 iboff = -1; 6291 allocblock = 0; 6292 extblocks = 0; 6293 datablocks = 0; 6294 frags = 0; 6295 freeblks = newfreeblks(mp, ip); 6296 ACQUIRE_LOCK(&lk); 6297 /* 6298 * If we're truncating a removed file that will never be written 6299 * we don't need to journal the block frees. The canceled journals 6300 * for the allocations will suffice. 6301 */ 6302 dflags = DEPALLOC; 6303 if (IS_SNAPSHOT(ip)) 6304 dflags |= NODELAY; 6305 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6306 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6307 length == 0) 6308 needj = 0; 6309 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6310 ip->i_number, length, needj); 6311 FREE_LOCK(&lk); 6312 /* 6313 * Calculate the lbn that we are truncating to. This results in -1 6314 * if we're truncating the 0 bytes. So it is the last lbn we want 6315 * to keep, not the first lbn we want to truncate. 6316 */ 6317 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6318 lastoff = blkoff(fs, length); 6319 /* 6320 * Compute frags we are keeping in lastlbn. 0 means all. 6321 */ 6322 if (lastlbn >= 0 && lastlbn < NDADDR) { 6323 frags = fragroundup(fs, lastoff); 6324 /* adp offset of last valid allocdirect. */ 6325 iboff = lastlbn; 6326 } else if (lastlbn > 0) 6327 iboff = NDADDR; 6328 if (fs->fs_magic == FS_UFS2_MAGIC) 6329 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6330 /* 6331 * Handle normal data blocks and indirects. This section saves 6332 * values used after the inode update to complete frag and indirect 6333 * truncation. 6334 */ 6335 if ((flags & IO_NORMAL) != 0) { 6336 /* 6337 * Handle truncation of whole direct and indirect blocks. 6338 */ 6339 for (i = iboff + 1; i < NDADDR; i++) 6340 setup_freedirect(freeblks, ip, i, needj); 6341 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6342 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6343 /* Release a whole indirect tree. */ 6344 if (lbn > lastlbn) { 6345 setup_freeindir(freeblks, ip, i, -lbn -i, 6346 needj); 6347 continue; 6348 } 6349 iboff = i + NDADDR; 6350 /* 6351 * Traverse partially truncated indirect tree. 6352 */ 6353 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6354 setup_trunc_indir(freeblks, ip, -lbn - i, 6355 lastlbn, DIP(ip, i_ib[i])); 6356 } 6357 /* 6358 * Handle partial truncation to a frag boundary. 6359 */ 6360 if (frags) { 6361 ufs2_daddr_t blkno; 6362 long oldfrags; 6363 6364 oldfrags = blksize(fs, ip, lastlbn); 6365 blkno = DIP(ip, i_db[lastlbn]); 6366 if (blkno && oldfrags != frags) { 6367 oldfrags -= frags; 6368 oldfrags = numfrags(ip->i_fs, oldfrags); 6369 blkno += numfrags(ip->i_fs, frags); 6370 newfreework(ip->i_ump, freeblks, NULL, lastlbn, 6371 blkno, oldfrags, 0, needj); 6372 } else if (blkno == 0) 6373 allocblock = 1; 6374 } 6375 /* 6376 * Add a journal record for partial truncate if we are 6377 * handling indirect blocks. Non-indirects need no extra 6378 * journaling. 6379 */ 6380 if (length != 0 && lastlbn >= NDADDR) { 6381 ip->i_flag |= IN_TRUNCATED; 6382 newjtrunc(freeblks, length, 0); 6383 } 6384 ip->i_size = length; 6385 DIP_SET(ip, i_size, ip->i_size); 6386 datablocks = DIP(ip, i_blocks) - extblocks; 6387 if (length != 0) 6388 datablocks = blkcount(ip->i_fs, datablocks, length); 6389 freeblks->fb_len = length; 6390 } 6391 if ((flags & IO_EXT) != 0) { 6392 for (i = 0; i < NXADDR; i++) 6393 setup_freeext(freeblks, ip, i, needj); 6394 ip->i_din2->di_extsize = 0; 6395 datablocks += extblocks; 6396 } 6397#ifdef QUOTA 6398 /* Reference the quotas in case the block count is wrong in the end. */ 6399 quotaref(vp, freeblks->fb_quota); 6400 (void) chkdq(ip, -datablocks, NOCRED, 0); 6401#endif 6402 freeblks->fb_chkcnt = -datablocks; 6403 UFS_LOCK(ip->i_ump); 6404 fs->fs_pendingblocks += datablocks; 6405 UFS_UNLOCK(ip->i_ump); 6406 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6407 /* 6408 * Handle truncation of incomplete alloc direct dependencies. We 6409 * hold the inode block locked to prevent incomplete dependencies 6410 * from reaching the disk while we are eliminating those that 6411 * have been truncated. This is a partially inlined ffs_update(). 6412 */ 6413 ufs_itimes(vp); 6414 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6415 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6416 (int)fs->fs_bsize, cred, &bp); 6417 if (error) { 6418 brelse(bp); 6419 softdep_error("softdep_journal_freeblocks", error); 6420 return; 6421 } 6422 if (bp->b_bufsize == fs->fs_bsize) 6423 bp->b_flags |= B_CLUSTEROK; 6424 softdep_update_inodeblock(ip, bp, 0); 6425 if (ip->i_ump->um_fstype == UFS1) 6426 *((struct ufs1_dinode *)bp->b_data + 6427 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6428 else 6429 *((struct ufs2_dinode *)bp->b_data + 6430 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6431 ACQUIRE_LOCK(&lk); 6432 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6433 if ((inodedep->id_state & IOSTARTED) != 0) 6434 panic("softdep_setup_freeblocks: inode busy"); 6435 /* 6436 * Add the freeblks structure to the list of operations that 6437 * must await the zero'ed inode being written to disk. If we 6438 * still have a bitmap dependency (needj), then the inode 6439 * has never been written to disk, so we can process the 6440 * freeblks below once we have deleted the dependencies. 6441 */ 6442 if (needj) 6443 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6444 else 6445 freeblks->fb_state |= COMPLETE; 6446 if ((flags & IO_NORMAL) != 0) { 6447 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6448 if (adp->ad_offset > iboff) 6449 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6450 freeblks); 6451 /* 6452 * Truncate the allocdirect. We could eliminate 6453 * or modify journal records as well. 6454 */ 6455 else if (adp->ad_offset == iboff && frags) 6456 adp->ad_newsize = frags; 6457 } 6458 } 6459 if ((flags & IO_EXT) != 0) 6460 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6461 cancel_allocdirect(&inodedep->id_extupdt, adp, 6462 freeblks); 6463 /* 6464 * Scan the bufwait list for newblock dependencies that will never 6465 * make it to disk. 6466 */ 6467 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6468 if (wk->wk_type != D_ALLOCDIRECT) 6469 continue; 6470 adp = WK_ALLOCDIRECT(wk); 6471 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6472 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6473 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6474 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6475 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6476 } 6477 } 6478 /* 6479 * Add journal work. 6480 */ 6481 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6482 add_to_journal(&jblkdep->jb_list); 6483 FREE_LOCK(&lk); 6484 bdwrite(bp); 6485 /* 6486 * Truncate dependency structures beyond length. 6487 */ 6488 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6489 /* 6490 * This is only set when we need to allocate a fragment because 6491 * none existed at the end of a frag-sized file. It handles only 6492 * allocating a new, zero filled block. 6493 */ 6494 if (allocblock) { 6495 ip->i_size = length - lastoff; 6496 DIP_SET(ip, i_size, ip->i_size); 6497 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6498 if (error != 0) { 6499 softdep_error("softdep_journal_freeblks", error); 6500 return; 6501 } 6502 ip->i_size = length; 6503 DIP_SET(ip, i_size, length); 6504 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6505 allocbuf(bp, frags); 6506 ffs_update(vp, 0); 6507 bawrite(bp); 6508 } else if (lastoff != 0 && vp->v_type != VDIR) { 6509 int size; 6510 6511 /* 6512 * Zero the end of a truncated frag or block. 6513 */ 6514 size = sblksize(fs, length, lastlbn); 6515 error = bread(vp, lastlbn, size, cred, &bp); 6516 if (error) { 6517 softdep_error("softdep_journal_freeblks", error); 6518 return; 6519 } 6520 bzero((char *)bp->b_data + lastoff, size - lastoff); 6521 bawrite(bp); 6522 6523 } 6524 ACQUIRE_LOCK(&lk); 6525 inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6526 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6527 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6528 /* 6529 * We zero earlier truncations so they don't erroneously 6530 * update i_blocks. 6531 */ 6532 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6533 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6534 fbn->fb_len = 0; 6535 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6536 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6537 freeblks->fb_state |= INPROGRESS; 6538 else 6539 freeblks = NULL; 6540 FREE_LOCK(&lk); 6541 if (freeblks) 6542 handle_workitem_freeblocks(freeblks, 0); 6543 trunc_pages(ip, length, extblocks, flags); 6544 6545} 6546 6547/* 6548 * Flush a JOP_SYNC to the journal. 6549 */ 6550void 6551softdep_journal_fsync(ip) 6552 struct inode *ip; 6553{ 6554 struct jfsync *jfsync; 6555 6556 if ((ip->i_flag & IN_TRUNCATED) == 0) 6557 return; 6558 ip->i_flag &= ~IN_TRUNCATED; 6559 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6560 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump)); 6561 jfsync->jfs_size = ip->i_size; 6562 jfsync->jfs_ino = ip->i_number; 6563 ACQUIRE_LOCK(&lk); 6564 add_to_journal(&jfsync->jfs_list); 6565 jwait(&jfsync->jfs_list, MNT_WAIT); 6566 FREE_LOCK(&lk); 6567} 6568 6569/* 6570 * Block de-allocation dependencies. 6571 * 6572 * When blocks are de-allocated, the on-disk pointers must be nullified before 6573 * the blocks are made available for use by other files. (The true 6574 * requirement is that old pointers must be nullified before new on-disk 6575 * pointers are set. We chose this slightly more stringent requirement to 6576 * reduce complexity.) Our implementation handles this dependency by updating 6577 * the inode (or indirect block) appropriately but delaying the actual block 6578 * de-allocation (i.e., freemap and free space count manipulation) until 6579 * after the updated versions reach stable storage. After the disk is 6580 * updated, the blocks can be safely de-allocated whenever it is convenient. 6581 * This implementation handles only the common case of reducing a file's 6582 * length to zero. Other cases are handled by the conventional synchronous 6583 * write approach. 6584 * 6585 * The ffs implementation with which we worked double-checks 6586 * the state of the block pointers and file size as it reduces 6587 * a file's length. Some of this code is replicated here in our 6588 * soft updates implementation. The freeblks->fb_chkcnt field is 6589 * used to transfer a part of this information to the procedure 6590 * that eventually de-allocates the blocks. 6591 * 6592 * This routine should be called from the routine that shortens 6593 * a file's length, before the inode's size or block pointers 6594 * are modified. It will save the block pointer information for 6595 * later release and zero the inode so that the calling routine 6596 * can release it. 6597 */ 6598void 6599softdep_setup_freeblocks(ip, length, flags) 6600 struct inode *ip; /* The inode whose length is to be reduced */ 6601 off_t length; /* The new length for the file */ 6602 int flags; /* IO_EXT and/or IO_NORMAL */ 6603{ 6604 struct ufs1_dinode *dp1; 6605 struct ufs2_dinode *dp2; 6606 struct freeblks *freeblks; 6607 struct inodedep *inodedep; 6608 struct allocdirect *adp; 6609 struct buf *bp; 6610 struct fs *fs; 6611 ufs2_daddr_t extblocks, datablocks; 6612 struct mount *mp; 6613 int i, delay, error, dflags; 6614 ufs_lbn_t tmpval; 6615 ufs_lbn_t lbn; 6616 6617 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6618 ip->i_number, length); 6619 fs = ip->i_fs; 6620 mp = UFSTOVFS(ip->i_ump); 6621 if (length != 0) 6622 panic("softdep_setup_freeblocks: non-zero length"); 6623 freeblks = newfreeblks(mp, ip); 6624 extblocks = 0; 6625 datablocks = 0; 6626 if (fs->fs_magic == FS_UFS2_MAGIC) 6627 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6628 if ((flags & IO_NORMAL) != 0) { 6629 for (i = 0; i < NDADDR; i++) 6630 setup_freedirect(freeblks, ip, i, 0); 6631 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6632 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6633 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6634 ip->i_size = 0; 6635 DIP_SET(ip, i_size, 0); 6636 datablocks = DIP(ip, i_blocks) - extblocks; 6637 } 6638 if ((flags & IO_EXT) != 0) { 6639 for (i = 0; i < NXADDR; i++) 6640 setup_freeext(freeblks, ip, i, 0); 6641 ip->i_din2->di_extsize = 0; 6642 datablocks += extblocks; 6643 } 6644#ifdef QUOTA 6645 /* Reference the quotas in case the block count is wrong in the end. */ 6646 quotaref(ITOV(ip), freeblks->fb_quota); 6647 (void) chkdq(ip, -datablocks, NOCRED, 0); 6648#endif 6649 freeblks->fb_chkcnt = -datablocks; 6650 UFS_LOCK(ip->i_ump); 6651 fs->fs_pendingblocks += datablocks; 6652 UFS_UNLOCK(ip->i_ump); 6653 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6654 /* 6655 * Push the zero'ed inode to to its disk buffer so that we are free 6656 * to delete its dependencies below. Once the dependencies are gone 6657 * the buffer can be safely released. 6658 */ 6659 if ((error = bread(ip->i_devvp, 6660 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6661 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6662 brelse(bp); 6663 softdep_error("softdep_setup_freeblocks", error); 6664 } 6665 if (ip->i_ump->um_fstype == UFS1) { 6666 dp1 = ((struct ufs1_dinode *)bp->b_data + 6667 ino_to_fsbo(fs, ip->i_number)); 6668 ip->i_din1->di_freelink = dp1->di_freelink; 6669 *dp1 = *ip->i_din1; 6670 } else { 6671 dp2 = ((struct ufs2_dinode *)bp->b_data + 6672 ino_to_fsbo(fs, ip->i_number)); 6673 ip->i_din2->di_freelink = dp2->di_freelink; 6674 *dp2 = *ip->i_din2; 6675 } 6676 /* 6677 * Find and eliminate any inode dependencies. 6678 */ 6679 ACQUIRE_LOCK(&lk); 6680 dflags = DEPALLOC; 6681 if (IS_SNAPSHOT(ip)) 6682 dflags |= NODELAY; 6683 (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep); 6684 if ((inodedep->id_state & IOSTARTED) != 0) 6685 panic("softdep_setup_freeblocks: inode busy"); 6686 /* 6687 * Add the freeblks structure to the list of operations that 6688 * must await the zero'ed inode being written to disk. If we 6689 * still have a bitmap dependency (delay == 0), then the inode 6690 * has never been written to disk, so we can process the 6691 * freeblks below once we have deleted the dependencies. 6692 */ 6693 delay = (inodedep->id_state & DEPCOMPLETE); 6694 if (delay) 6695 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6696 else 6697 freeblks->fb_state |= COMPLETE; 6698 /* 6699 * Because the file length has been truncated to zero, any 6700 * pending block allocation dependency structures associated 6701 * with this inode are obsolete and can simply be de-allocated. 6702 * We must first merge the two dependency lists to get rid of 6703 * any duplicate freefrag structures, then purge the merged list. 6704 * If we still have a bitmap dependency, then the inode has never 6705 * been written to disk, so we can free any fragments without delay. 6706 */ 6707 if (flags & IO_NORMAL) { 6708 merge_inode_lists(&inodedep->id_newinoupdt, 6709 &inodedep->id_inoupdt); 6710 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 6711 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6712 freeblks); 6713 } 6714 if (flags & IO_EXT) { 6715 merge_inode_lists(&inodedep->id_newextupdt, 6716 &inodedep->id_extupdt); 6717 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6718 cancel_allocdirect(&inodedep->id_extupdt, adp, 6719 freeblks); 6720 } 6721 FREE_LOCK(&lk); 6722 bdwrite(bp); 6723 trunc_dependencies(ip, freeblks, -1, 0, flags); 6724 ACQUIRE_LOCK(&lk); 6725 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6726 (void) free_inodedep(inodedep); 6727 freeblks->fb_state |= DEPCOMPLETE; 6728 /* 6729 * If the inode with zeroed block pointers is now on disk 6730 * we can start freeing blocks. 6731 */ 6732 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6733 freeblks->fb_state |= INPROGRESS; 6734 else 6735 freeblks = NULL; 6736 FREE_LOCK(&lk); 6737 if (freeblks) 6738 handle_workitem_freeblocks(freeblks, 0); 6739 trunc_pages(ip, length, extblocks, flags); 6740} 6741 6742/* 6743 * Eliminate pages from the page cache that back parts of this inode and 6744 * adjust the vnode pager's idea of our size. This prevents stale data 6745 * from hanging around in the page cache. 6746 */ 6747static void 6748trunc_pages(ip, length, extblocks, flags) 6749 struct inode *ip; 6750 off_t length; 6751 ufs2_daddr_t extblocks; 6752 int flags; 6753{ 6754 struct vnode *vp; 6755 struct fs *fs; 6756 ufs_lbn_t lbn; 6757 off_t end, extend; 6758 6759 vp = ITOV(ip); 6760 fs = ip->i_fs; 6761 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 6762 if ((flags & IO_EXT) != 0) 6763 vn_pages_remove(vp, extend, 0); 6764 if ((flags & IO_NORMAL) == 0) 6765 return; 6766 BO_LOCK(&vp->v_bufobj); 6767 drain_output(vp); 6768 BO_UNLOCK(&vp->v_bufobj); 6769 /* 6770 * The vnode pager eliminates file pages we eliminate indirects 6771 * below. 6772 */ 6773 vnode_pager_setsize(vp, length); 6774 /* 6775 * Calculate the end based on the last indirect we want to keep. If 6776 * the block extends into indirects we can just use the negative of 6777 * its lbn. Doubles and triples exist at lower numbers so we must 6778 * be careful not to remove those, if they exist. double and triple 6779 * indirect lbns do not overlap with others so it is not important 6780 * to verify how many levels are required. 6781 */ 6782 lbn = lblkno(fs, length); 6783 if (lbn >= NDADDR) { 6784 /* Calculate the virtual lbn of the triple indirect. */ 6785 lbn = -lbn - (NIADDR - 1); 6786 end = OFF_TO_IDX(lblktosize(fs, lbn)); 6787 } else 6788 end = extend; 6789 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 6790} 6791 6792/* 6793 * See if the buf bp is in the range eliminated by truncation. 6794 */ 6795static int 6796trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 6797 struct buf *bp; 6798 int *blkoffp; 6799 ufs_lbn_t lastlbn; 6800 int lastoff; 6801 int flags; 6802{ 6803 ufs_lbn_t lbn; 6804 6805 *blkoffp = 0; 6806 /* Only match ext/normal blocks as appropriate. */ 6807 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 6808 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 6809 return (0); 6810 /* ALTDATA is always a full truncation. */ 6811 if ((bp->b_xflags & BX_ALTDATA) != 0) 6812 return (1); 6813 /* -1 is full truncation. */ 6814 if (lastlbn == -1) 6815 return (1); 6816 /* 6817 * If this is a partial truncate we only want those 6818 * blocks and indirect blocks that cover the range 6819 * we're after. 6820 */ 6821 lbn = bp->b_lblkno; 6822 if (lbn < 0) 6823 lbn = -(lbn + lbn_level(lbn)); 6824 if (lbn < lastlbn) 6825 return (0); 6826 /* Here we only truncate lblkno if it's partial. */ 6827 if (lbn == lastlbn) { 6828 if (lastoff == 0) 6829 return (0); 6830 *blkoffp = lastoff; 6831 } 6832 return (1); 6833} 6834 6835/* 6836 * Eliminate any dependencies that exist in memory beyond lblkno:off 6837 */ 6838static void 6839trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 6840 struct inode *ip; 6841 struct freeblks *freeblks; 6842 ufs_lbn_t lastlbn; 6843 int lastoff; 6844 int flags; 6845{ 6846 struct bufobj *bo; 6847 struct vnode *vp; 6848 struct buf *bp; 6849 struct fs *fs; 6850 int blkoff; 6851 6852 /* 6853 * We must wait for any I/O in progress to finish so that 6854 * all potential buffers on the dirty list will be visible. 6855 * Once they are all there, walk the list and get rid of 6856 * any dependencies. 6857 */ 6858 fs = ip->i_fs; 6859 vp = ITOV(ip); 6860 bo = &vp->v_bufobj; 6861 BO_LOCK(bo); 6862 drain_output(vp); 6863 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 6864 bp->b_vflags &= ~BV_SCANNED; 6865restart: 6866 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 6867 if (bp->b_vflags & BV_SCANNED) 6868 continue; 6869 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 6870 bp->b_vflags |= BV_SCANNED; 6871 continue; 6872 } 6873 if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL) 6874 goto restart; 6875 BO_UNLOCK(bo); 6876 if (deallocate_dependencies(bp, freeblks, blkoff)) 6877 bqrelse(bp); 6878 else 6879 brelse(bp); 6880 BO_LOCK(bo); 6881 goto restart; 6882 } 6883 /* 6884 * Now do the work of vtruncbuf while also matching indirect blocks. 6885 */ 6886 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 6887 bp->b_vflags &= ~BV_SCANNED; 6888cleanrestart: 6889 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 6890 if (bp->b_vflags & BV_SCANNED) 6891 continue; 6892 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 6893 bp->b_vflags |= BV_SCANNED; 6894 continue; 6895 } 6896 if (BUF_LOCK(bp, 6897 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6898 BO_MTX(bo)) == ENOLCK) { 6899 BO_LOCK(bo); 6900 goto cleanrestart; 6901 } 6902 bp->b_vflags |= BV_SCANNED; 6903 BO_LOCK(bo); 6904 bremfree(bp); 6905 BO_UNLOCK(bo); 6906 if (blkoff != 0) { 6907 allocbuf(bp, blkoff); 6908 bqrelse(bp); 6909 } else { 6910 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 6911 brelse(bp); 6912 } 6913 BO_LOCK(bo); 6914 goto cleanrestart; 6915 } 6916 drain_output(vp); 6917 BO_UNLOCK(bo); 6918} 6919 6920static int 6921cancel_pagedep(pagedep, freeblks, blkoff) 6922 struct pagedep *pagedep; 6923 struct freeblks *freeblks; 6924 int blkoff; 6925{ 6926 struct jremref *jremref; 6927 struct jmvref *jmvref; 6928 struct dirrem *dirrem, *tmp; 6929 int i; 6930 6931 /* 6932 * Copy any directory remove dependencies to the list 6933 * to be processed after the freeblks proceeds. If 6934 * directory entry never made it to disk they 6935 * can be dumped directly onto the work list. 6936 */ 6937 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 6938 /* Skip this directory removal if it is intended to remain. */ 6939 if (dirrem->dm_offset < blkoff) 6940 continue; 6941 /* 6942 * If there are any dirrems we wait for the journal write 6943 * to complete and then restart the buf scan as the lock 6944 * has been dropped. 6945 */ 6946 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 6947 jwait(&jremref->jr_list, MNT_WAIT); 6948 return (ERESTART); 6949 } 6950 LIST_REMOVE(dirrem, dm_next); 6951 dirrem->dm_dirinum = pagedep->pd_ino; 6952 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 6953 } 6954 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 6955 jwait(&jmvref->jm_list, MNT_WAIT); 6956 return (ERESTART); 6957 } 6958 /* 6959 * When we're partially truncating a pagedep we just want to flush 6960 * journal entries and return. There can not be any adds in the 6961 * truncated portion of the directory and newblk must remain if 6962 * part of the block remains. 6963 */ 6964 if (blkoff != 0) { 6965 struct diradd *dap; 6966 6967 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 6968 if (dap->da_offset > blkoff) 6969 panic("cancel_pagedep: diradd %p off %d > %d", 6970 dap, dap->da_offset, blkoff); 6971 for (i = 0; i < DAHASHSZ; i++) 6972 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 6973 if (dap->da_offset > blkoff) 6974 panic("cancel_pagedep: diradd %p off %d > %d", 6975 dap, dap->da_offset, blkoff); 6976 return (0); 6977 } 6978 /* 6979 * There should be no directory add dependencies present 6980 * as the directory could not be truncated until all 6981 * children were removed. 6982 */ 6983 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 6984 ("deallocate_dependencies: pendinghd != NULL")); 6985 for (i = 0; i < DAHASHSZ; i++) 6986 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 6987 ("deallocate_dependencies: diraddhd != NULL")); 6988 if ((pagedep->pd_state & NEWBLOCK) != 0) 6989 free_newdirblk(pagedep->pd_newdirblk); 6990 if (free_pagedep(pagedep) == 0) 6991 panic("Failed to free pagedep %p", pagedep); 6992 return (0); 6993} 6994 6995/* 6996 * Reclaim any dependency structures from a buffer that is about to 6997 * be reallocated to a new vnode. The buffer must be locked, thus, 6998 * no I/O completion operations can occur while we are manipulating 6999 * its associated dependencies. The mutex is held so that other I/O's 7000 * associated with related dependencies do not occur. 7001 */ 7002static int 7003deallocate_dependencies(bp, freeblks, off) 7004 struct buf *bp; 7005 struct freeblks *freeblks; 7006 int off; 7007{ 7008 struct indirdep *indirdep; 7009 struct pagedep *pagedep; 7010 struct allocdirect *adp; 7011 struct worklist *wk, *wkn; 7012 7013 ACQUIRE_LOCK(&lk); 7014 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7015 switch (wk->wk_type) { 7016 case D_INDIRDEP: 7017 indirdep = WK_INDIRDEP(wk); 7018 if (bp->b_lblkno >= 0 || 7019 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7020 panic("deallocate_dependencies: not indir"); 7021 cancel_indirdep(indirdep, bp, freeblks); 7022 continue; 7023 7024 case D_PAGEDEP: 7025 pagedep = WK_PAGEDEP(wk); 7026 if (cancel_pagedep(pagedep, freeblks, off)) { 7027 FREE_LOCK(&lk); 7028 return (ERESTART); 7029 } 7030 continue; 7031 7032 case D_ALLOCINDIR: 7033 /* 7034 * Simply remove the allocindir, we'll find it via 7035 * the indirdep where we can clear pointers if 7036 * needed. 7037 */ 7038 WORKLIST_REMOVE(wk); 7039 continue; 7040 7041 case D_FREEWORK: 7042 /* 7043 * A truncation is waiting for the zero'd pointers 7044 * to be written. It can be freed when the freeblks 7045 * is journaled. 7046 */ 7047 WORKLIST_REMOVE(wk); 7048 wk->wk_state |= ONDEPLIST; 7049 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7050 break; 7051 7052 case D_ALLOCDIRECT: 7053 adp = WK_ALLOCDIRECT(wk); 7054 if (off != 0) 7055 continue; 7056 /* FALLTHROUGH */ 7057 default: 7058 panic("deallocate_dependencies: Unexpected type %s", 7059 TYPENAME(wk->wk_type)); 7060 /* NOTREACHED */ 7061 } 7062 } 7063 FREE_LOCK(&lk); 7064 /* 7065 * Don't throw away this buf, we were partially truncating and 7066 * some deps may always remain. 7067 */ 7068 if (off) { 7069 allocbuf(bp, off); 7070 bp->b_vflags |= BV_SCANNED; 7071 return (EBUSY); 7072 } 7073 bp->b_flags |= B_INVAL | B_NOCACHE; 7074 7075 return (0); 7076} 7077 7078/* 7079 * An allocdirect is being canceled due to a truncate. We must make sure 7080 * the journal entry is released in concert with the blkfree that releases 7081 * the storage. Completed journal entries must not be released until the 7082 * space is no longer pointed to by the inode or in the bitmap. 7083 */ 7084static void 7085cancel_allocdirect(adphead, adp, freeblks) 7086 struct allocdirectlst *adphead; 7087 struct allocdirect *adp; 7088 struct freeblks *freeblks; 7089{ 7090 struct freework *freework; 7091 struct newblk *newblk; 7092 struct worklist *wk; 7093 7094 TAILQ_REMOVE(adphead, adp, ad_next); 7095 newblk = (struct newblk *)adp; 7096 freework = NULL; 7097 /* 7098 * Find the correct freework structure. 7099 */ 7100 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7101 if (wk->wk_type != D_FREEWORK) 7102 continue; 7103 freework = WK_FREEWORK(wk); 7104 if (freework->fw_blkno == newblk->nb_newblkno) 7105 break; 7106 } 7107 if (freework == NULL) 7108 panic("cancel_allocdirect: Freework not found"); 7109 /* 7110 * If a newblk exists at all we still have the journal entry that 7111 * initiated the allocation so we do not need to journal the free. 7112 */ 7113 cancel_jfreeblk(freeblks, freework->fw_blkno); 7114 /* 7115 * If the journal hasn't been written the jnewblk must be passed 7116 * to the call to ffs_blkfree that reclaims the space. We accomplish 7117 * this by linking the journal dependency into the freework to be 7118 * freed when freework_freeblock() is called. If the journal has 7119 * been written we can simply reclaim the journal space when the 7120 * freeblks work is complete. 7121 */ 7122 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7123 &freeblks->fb_jwork); 7124 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7125} 7126 7127 7128/* 7129 * Cancel a new block allocation. May be an indirect or direct block. We 7130 * remove it from various lists and return any journal record that needs to 7131 * be resolved by the caller. 7132 * 7133 * A special consideration is made for indirects which were never pointed 7134 * at on disk and will never be found once this block is released. 7135 */ 7136static struct jnewblk * 7137cancel_newblk(newblk, wk, wkhd) 7138 struct newblk *newblk; 7139 struct worklist *wk; 7140 struct workhead *wkhd; 7141{ 7142 struct jnewblk *jnewblk; 7143 7144 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7145 7146 newblk->nb_state |= GOINGAWAY; 7147 /* 7148 * Previously we traversed the completedhd on each indirdep 7149 * attached to this newblk to cancel them and gather journal 7150 * work. Since we need only the oldest journal segment and 7151 * the lowest point on the tree will always have the oldest 7152 * journal segment we are free to release the segments 7153 * of any subordinates and may leave the indirdep list to 7154 * indirdep_complete() when this newblk is freed. 7155 */ 7156 if (newblk->nb_state & ONDEPLIST) { 7157 newblk->nb_state &= ~ONDEPLIST; 7158 LIST_REMOVE(newblk, nb_deps); 7159 } 7160 if (newblk->nb_state & ONWORKLIST) 7161 WORKLIST_REMOVE(&newblk->nb_list); 7162 /* 7163 * If the journal entry hasn't been written we save a pointer to 7164 * the dependency that frees it until it is written or the 7165 * superseding operation completes. 7166 */ 7167 jnewblk = newblk->nb_jnewblk; 7168 if (jnewblk != NULL && wk != NULL) { 7169 newblk->nb_jnewblk = NULL; 7170 jnewblk->jn_dep = wk; 7171 } 7172 if (!LIST_EMPTY(&newblk->nb_jwork)) 7173 jwork_move(wkhd, &newblk->nb_jwork); 7174 /* 7175 * When truncating we must free the newdirblk early to remove 7176 * the pagedep from the hash before returning. 7177 */ 7178 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7179 free_newdirblk(WK_NEWDIRBLK(wk)); 7180 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7181 panic("cancel_newblk: extra newdirblk"); 7182 7183 return (jnewblk); 7184} 7185 7186/* 7187 * Schedule the freefrag associated with a newblk to be released once 7188 * the pointers are written and the previous block is no longer needed. 7189 */ 7190static void 7191newblk_freefrag(newblk) 7192 struct newblk *newblk; 7193{ 7194 struct freefrag *freefrag; 7195 7196 if (newblk->nb_freefrag == NULL) 7197 return; 7198 freefrag = newblk->nb_freefrag; 7199 newblk->nb_freefrag = NULL; 7200 freefrag->ff_state |= COMPLETE; 7201 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7202 add_to_worklist(&freefrag->ff_list, 0); 7203} 7204 7205/* 7206 * Free a newblk. Generate a new freefrag work request if appropriate. 7207 * This must be called after the inode pointer and any direct block pointers 7208 * are valid or fully removed via truncate or frag extension. 7209 */ 7210static void 7211free_newblk(newblk) 7212 struct newblk *newblk; 7213{ 7214 struct indirdep *indirdep; 7215 struct worklist *wk; 7216 7217 KASSERT(newblk->nb_jnewblk == NULL, 7218 ("free_newblk; jnewblk %p still attached", newblk->nb_jnewblk)); 7219 mtx_assert(&lk, MA_OWNED); 7220 newblk_freefrag(newblk); 7221 if (newblk->nb_state & ONDEPLIST) 7222 LIST_REMOVE(newblk, nb_deps); 7223 if (newblk->nb_state & ONWORKLIST) 7224 WORKLIST_REMOVE(&newblk->nb_list); 7225 LIST_REMOVE(newblk, nb_hash); 7226 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7227 free_newdirblk(WK_NEWDIRBLK(wk)); 7228 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7229 panic("free_newblk: extra newdirblk"); 7230 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7231 indirdep_complete(indirdep); 7232 handle_jwork(&newblk->nb_jwork); 7233 newblk->nb_list.wk_type = D_NEWBLK; 7234 WORKITEM_FREE(newblk, D_NEWBLK); 7235} 7236 7237/* 7238 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7239 * This routine must be called with splbio interrupts blocked. 7240 */ 7241static void 7242free_newdirblk(newdirblk) 7243 struct newdirblk *newdirblk; 7244{ 7245 struct pagedep *pagedep; 7246 struct diradd *dap; 7247 struct worklist *wk; 7248 7249 mtx_assert(&lk, MA_OWNED); 7250 WORKLIST_REMOVE(&newdirblk->db_list); 7251 /* 7252 * If the pagedep is still linked onto the directory buffer 7253 * dependency chain, then some of the entries on the 7254 * pd_pendinghd list may not be committed to disk yet. In 7255 * this case, we will simply clear the NEWBLOCK flag and 7256 * let the pd_pendinghd list be processed when the pagedep 7257 * is next written. If the pagedep is no longer on the buffer 7258 * dependency chain, then all the entries on the pd_pending 7259 * list are committed to disk and we can free them here. 7260 */ 7261 pagedep = newdirblk->db_pagedep; 7262 pagedep->pd_state &= ~NEWBLOCK; 7263 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7264 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7265 free_diradd(dap, NULL); 7266 /* 7267 * If no dependencies remain, the pagedep will be freed. 7268 */ 7269 free_pagedep(pagedep); 7270 } 7271 /* Should only ever be one item in the list. */ 7272 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7273 WORKLIST_REMOVE(wk); 7274 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7275 } 7276 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7277} 7278 7279/* 7280 * Prepare an inode to be freed. The actual free operation is not 7281 * done until the zero'ed inode has been written to disk. 7282 */ 7283void 7284softdep_freefile(pvp, ino, mode) 7285 struct vnode *pvp; 7286 ino_t ino; 7287 int mode; 7288{ 7289 struct inode *ip = VTOI(pvp); 7290 struct inodedep *inodedep; 7291 struct freefile *freefile; 7292 struct freeblks *freeblks; 7293 7294 /* 7295 * This sets up the inode de-allocation dependency. 7296 */ 7297 freefile = malloc(sizeof(struct freefile), 7298 M_FREEFILE, M_SOFTDEP_FLAGS); 7299 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7300 freefile->fx_mode = mode; 7301 freefile->fx_oldinum = ino; 7302 freefile->fx_devvp = ip->i_devvp; 7303 LIST_INIT(&freefile->fx_jwork); 7304 UFS_LOCK(ip->i_ump); 7305 ip->i_fs->fs_pendinginodes += 1; 7306 UFS_UNLOCK(ip->i_ump); 7307 7308 /* 7309 * If the inodedep does not exist, then the zero'ed inode has 7310 * been written to disk. If the allocated inode has never been 7311 * written to disk, then the on-disk inode is zero'ed. In either 7312 * case we can free the file immediately. If the journal was 7313 * canceled before being written the inode will never make it to 7314 * disk and we must send the canceled journal entrys to 7315 * ffs_freefile() to be cleared in conjunction with the bitmap. 7316 * Any blocks waiting on the inode to write can be safely freed 7317 * here as it will never been written. 7318 */ 7319 ACQUIRE_LOCK(&lk); 7320 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7321 if (inodedep) { 7322 /* 7323 * Clear out freeblks that no longer need to reference 7324 * this inode. 7325 */ 7326 while ((freeblks = 7327 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7328 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7329 fb_next); 7330 freeblks->fb_state &= ~ONDEPLIST; 7331 } 7332 /* 7333 * Remove this inode from the unlinked list. 7334 */ 7335 if (inodedep->id_state & UNLINKED) { 7336 /* 7337 * Save the journal work to be freed with the bitmap 7338 * before we clear UNLINKED. Otherwise it can be lost 7339 * if the inode block is written. 7340 */ 7341 handle_bufwait(inodedep, &freefile->fx_jwork); 7342 clear_unlinked_inodedep(inodedep); 7343 /* Re-acquire inodedep as we've dropped lk. */ 7344 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7345 } 7346 } 7347 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7348 FREE_LOCK(&lk); 7349 handle_workitem_freefile(freefile); 7350 return; 7351 } 7352 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7353 inodedep->id_state |= GOINGAWAY; 7354 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7355 FREE_LOCK(&lk); 7356 if (ip->i_number == ino) 7357 ip->i_flag |= IN_MODIFIED; 7358} 7359 7360/* 7361 * Check to see if an inode has never been written to disk. If 7362 * so free the inodedep and return success, otherwise return failure. 7363 * This routine must be called with splbio interrupts blocked. 7364 * 7365 * If we still have a bitmap dependency, then the inode has never 7366 * been written to disk. Drop the dependency as it is no longer 7367 * necessary since the inode is being deallocated. We set the 7368 * ALLCOMPLETE flags since the bitmap now properly shows that the 7369 * inode is not allocated. Even if the inode is actively being 7370 * written, it has been rolled back to its zero'ed state, so we 7371 * are ensured that a zero inode is what is on the disk. For short 7372 * lived files, this change will usually result in removing all the 7373 * dependencies from the inode so that it can be freed immediately. 7374 */ 7375static int 7376check_inode_unwritten(inodedep) 7377 struct inodedep *inodedep; 7378{ 7379 7380 mtx_assert(&lk, MA_OWNED); 7381 7382 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7383 !LIST_EMPTY(&inodedep->id_dirremhd) || 7384 !LIST_EMPTY(&inodedep->id_pendinghd) || 7385 !LIST_EMPTY(&inodedep->id_bufwait) || 7386 !LIST_EMPTY(&inodedep->id_inowait) || 7387 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7388 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7389 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7390 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7391 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7392 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7393 inodedep->id_mkdiradd != NULL || 7394 inodedep->id_nlinkdelta != 0) 7395 return (0); 7396 /* 7397 * Another process might be in initiate_write_inodeblock_ufs[12] 7398 * trying to allocate memory without holding "Softdep Lock". 7399 */ 7400 if ((inodedep->id_state & IOSTARTED) != 0 && 7401 inodedep->id_savedino1 == NULL) 7402 return (0); 7403 7404 if (inodedep->id_state & ONDEPLIST) 7405 LIST_REMOVE(inodedep, id_deps); 7406 inodedep->id_state &= ~ONDEPLIST; 7407 inodedep->id_state |= ALLCOMPLETE; 7408 inodedep->id_bmsafemap = NULL; 7409 if (inodedep->id_state & ONWORKLIST) 7410 WORKLIST_REMOVE(&inodedep->id_list); 7411 if (inodedep->id_savedino1 != NULL) { 7412 free(inodedep->id_savedino1, M_SAVEDINO); 7413 inodedep->id_savedino1 = NULL; 7414 } 7415 if (free_inodedep(inodedep) == 0) 7416 panic("check_inode_unwritten: busy inode"); 7417 return (1); 7418} 7419 7420/* 7421 * Try to free an inodedep structure. Return 1 if it could be freed. 7422 */ 7423static int 7424free_inodedep(inodedep) 7425 struct inodedep *inodedep; 7426{ 7427 7428 mtx_assert(&lk, MA_OWNED); 7429 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7430 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7431 !LIST_EMPTY(&inodedep->id_dirremhd) || 7432 !LIST_EMPTY(&inodedep->id_pendinghd) || 7433 !LIST_EMPTY(&inodedep->id_bufwait) || 7434 !LIST_EMPTY(&inodedep->id_inowait) || 7435 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7436 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7437 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7438 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7439 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7440 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7441 inodedep->id_mkdiradd != NULL || 7442 inodedep->id_nlinkdelta != 0 || 7443 inodedep->id_savedino1 != NULL) 7444 return (0); 7445 if (inodedep->id_state & ONDEPLIST) 7446 LIST_REMOVE(inodedep, id_deps); 7447 LIST_REMOVE(inodedep, id_hash); 7448 WORKITEM_FREE(inodedep, D_INODEDEP); 7449 return (1); 7450} 7451 7452/* 7453 * Free the block referenced by a freework structure. The parent freeblks 7454 * structure is released and completed when the final cg bitmap reaches 7455 * the disk. This routine may be freeing a jnewblk which never made it to 7456 * disk in which case we do not have to wait as the operation is undone 7457 * in memory immediately. 7458 */ 7459static void 7460freework_freeblock(freework) 7461 struct freework *freework; 7462{ 7463 struct freeblks *freeblks; 7464 struct jnewblk *jnewblk; 7465 struct ufsmount *ump; 7466 struct workhead wkhd; 7467 struct fs *fs; 7468 int bsize; 7469 int needj; 7470 7471 mtx_assert(&lk, MA_OWNED); 7472 /* 7473 * Handle partial truncate separately. 7474 */ 7475 if (freework->fw_indir) { 7476 complete_trunc_indir(freework); 7477 return; 7478 } 7479 freeblks = freework->fw_freeblks; 7480 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7481 fs = ump->um_fs; 7482 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7483 bsize = lfragtosize(fs, freework->fw_frags); 7484 LIST_INIT(&wkhd); 7485 /* 7486 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7487 * on the indirblk hashtable and prevents premature freeing. 7488 */ 7489 freework->fw_state |= DEPCOMPLETE; 7490 /* 7491 * SUJ needs to wait for the segment referencing freed indirect 7492 * blocks to expire so that we know the checker will not confuse 7493 * a re-allocated indirect block with its old contents. 7494 */ 7495 if (needj && freework->fw_lbn <= -NDADDR) 7496 indirblk_insert(freework); 7497 /* 7498 * If we are canceling an existing jnewblk pass it to the free 7499 * routine, otherwise pass the freeblk which will ultimately 7500 * release the freeblks. If we're not journaling, we can just 7501 * free the freeblks immediately. 7502 */ 7503 jnewblk = freework->fw_jnewblk; 7504 if (jnewblk != NULL) { 7505 cancel_jnewblk(jnewblk, &wkhd); 7506 needj = 0; 7507 } else if (needj) { 7508 freework->fw_state |= DELAYEDFREE; 7509 freeblks->fb_cgwait++; 7510 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7511 } 7512 FREE_LOCK(&lk); 7513 freeblks_free(ump, freeblks, btodb(bsize)); 7514 CTR4(KTR_SUJ, 7515 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld", 7516 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7517 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7518 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7519 ACQUIRE_LOCK(&lk); 7520 /* 7521 * The jnewblk will be discarded and the bits in the map never 7522 * made it to disk. We can immediately free the freeblk. 7523 */ 7524 if (needj == 0) 7525 handle_written_freework(freework); 7526} 7527 7528/* 7529 * We enqueue freework items that need processing back on the freeblks and 7530 * add the freeblks to the worklist. This makes it easier to find all work 7531 * required to flush a truncation in process_truncates(). 7532 */ 7533static void 7534freework_enqueue(freework) 7535 struct freework *freework; 7536{ 7537 struct freeblks *freeblks; 7538 7539 freeblks = freework->fw_freeblks; 7540 if ((freework->fw_state & INPROGRESS) == 0) 7541 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7542 if ((freeblks->fb_state & 7543 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7544 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7545 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7546} 7547 7548/* 7549 * Start, continue, or finish the process of freeing an indirect block tree. 7550 * The free operation may be paused at any point with fw_off containing the 7551 * offset to restart from. This enables us to implement some flow control 7552 * for large truncates which may fan out and generate a huge number of 7553 * dependencies. 7554 */ 7555static void 7556handle_workitem_indirblk(freework) 7557 struct freework *freework; 7558{ 7559 struct freeblks *freeblks; 7560 struct ufsmount *ump; 7561 struct fs *fs; 7562 7563 freeblks = freework->fw_freeblks; 7564 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7565 fs = ump->um_fs; 7566 if (freework->fw_state & DEPCOMPLETE) { 7567 handle_written_freework(freework); 7568 return; 7569 } 7570 if (freework->fw_off == NINDIR(fs)) { 7571 freework_freeblock(freework); 7572 return; 7573 } 7574 freework->fw_state |= INPROGRESS; 7575 FREE_LOCK(&lk); 7576 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7577 freework->fw_lbn); 7578 ACQUIRE_LOCK(&lk); 7579} 7580 7581/* 7582 * Called when a freework structure attached to a cg buf is written. The 7583 * ref on either the parent or the freeblks structure is released and 7584 * the freeblks is added back to the worklist if there is more work to do. 7585 */ 7586static void 7587handle_written_freework(freework) 7588 struct freework *freework; 7589{ 7590 struct freeblks *freeblks; 7591 struct freework *parent; 7592 7593 freeblks = freework->fw_freeblks; 7594 parent = freework->fw_parent; 7595 if (freework->fw_state & DELAYEDFREE) 7596 freeblks->fb_cgwait--; 7597 freework->fw_state |= COMPLETE; 7598 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7599 WORKITEM_FREE(freework, D_FREEWORK); 7600 if (parent) { 7601 if (--parent->fw_ref == 0) 7602 freework_enqueue(parent); 7603 return; 7604 } 7605 if (--freeblks->fb_ref != 0) 7606 return; 7607 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7608 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7609 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7610} 7611 7612/* 7613 * This workitem routine performs the block de-allocation. 7614 * The workitem is added to the pending list after the updated 7615 * inode block has been written to disk. As mentioned above, 7616 * checks regarding the number of blocks de-allocated (compared 7617 * to the number of blocks allocated for the file) are also 7618 * performed in this function. 7619 */ 7620static int 7621handle_workitem_freeblocks(freeblks, flags) 7622 struct freeblks *freeblks; 7623 int flags; 7624{ 7625 struct freework *freework; 7626 struct newblk *newblk; 7627 struct allocindir *aip; 7628 struct ufsmount *ump; 7629 struct worklist *wk; 7630 7631 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7632 ("handle_workitem_freeblocks: Journal entries not written.")); 7633 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7634 ACQUIRE_LOCK(&lk); 7635 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7636 WORKLIST_REMOVE(wk); 7637 switch (wk->wk_type) { 7638 case D_DIRREM: 7639 wk->wk_state |= COMPLETE; 7640 add_to_worklist(wk, 0); 7641 continue; 7642 7643 case D_ALLOCDIRECT: 7644 free_newblk(WK_NEWBLK(wk)); 7645 continue; 7646 7647 case D_ALLOCINDIR: 7648 aip = WK_ALLOCINDIR(wk); 7649 freework = NULL; 7650 if (aip->ai_state & DELAYEDFREE) { 7651 FREE_LOCK(&lk); 7652 freework = newfreework(ump, freeblks, NULL, 7653 aip->ai_lbn, aip->ai_newblkno, 7654 ump->um_fs->fs_frag, 0, 0); 7655 ACQUIRE_LOCK(&lk); 7656 } 7657 newblk = WK_NEWBLK(wk); 7658 if (newblk->nb_jnewblk) { 7659 freework->fw_jnewblk = newblk->nb_jnewblk; 7660 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7661 newblk->nb_jnewblk = NULL; 7662 } 7663 free_newblk(newblk); 7664 continue; 7665 7666 case D_FREEWORK: 7667 freework = WK_FREEWORK(wk); 7668 if (freework->fw_lbn <= -NDADDR) 7669 handle_workitem_indirblk(freework); 7670 else 7671 freework_freeblock(freework); 7672 continue; 7673 default: 7674 panic("handle_workitem_freeblocks: Unknown type %s", 7675 TYPENAME(wk->wk_type)); 7676 } 7677 } 7678 if (freeblks->fb_ref != 0) { 7679 freeblks->fb_state &= ~INPROGRESS; 7680 wake_worklist(&freeblks->fb_list); 7681 freeblks = NULL; 7682 } 7683 FREE_LOCK(&lk); 7684 if (freeblks) 7685 return handle_complete_freeblocks(freeblks, flags); 7686 return (0); 7687} 7688 7689/* 7690 * Handle completion of block free via truncate. This allows fs_pending 7691 * to track the actual free block count more closely than if we only updated 7692 * it at the end. We must be careful to handle cases where the block count 7693 * on free was incorrect. 7694 */ 7695static void 7696freeblks_free(ump, freeblks, blocks) 7697 struct ufsmount *ump; 7698 struct freeblks *freeblks; 7699 int blocks; 7700{ 7701 struct fs *fs; 7702 ufs2_daddr_t remain; 7703 7704 UFS_LOCK(ump); 7705 remain = -freeblks->fb_chkcnt; 7706 freeblks->fb_chkcnt += blocks; 7707 if (remain > 0) { 7708 if (remain < blocks) 7709 blocks = remain; 7710 fs = ump->um_fs; 7711 fs->fs_pendingblocks -= blocks; 7712 } 7713 UFS_UNLOCK(ump); 7714} 7715 7716/* 7717 * Once all of the freework workitems are complete we can retire the 7718 * freeblocks dependency and any journal work awaiting completion. This 7719 * can not be called until all other dependencies are stable on disk. 7720 */ 7721static int 7722handle_complete_freeblocks(freeblks, flags) 7723 struct freeblks *freeblks; 7724 int flags; 7725{ 7726 struct inodedep *inodedep; 7727 struct inode *ip; 7728 struct vnode *vp; 7729 struct fs *fs; 7730 struct ufsmount *ump; 7731 ufs2_daddr_t spare; 7732 7733 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7734 fs = ump->um_fs; 7735 flags = LK_EXCLUSIVE | flags; 7736 spare = freeblks->fb_chkcnt; 7737 7738 /* 7739 * If we did not release the expected number of blocks we may have 7740 * to adjust the inode block count here. Only do so if it wasn't 7741 * a truncation to zero and the modrev still matches. 7742 */ 7743 if (spare && freeblks->fb_len != 0) { 7744 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7745 flags, &vp, FFSV_FORCEINSMQ) != 0) 7746 return (EBUSY); 7747 ip = VTOI(vp); 7748 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 7749 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 7750 ip->i_flag |= IN_CHANGE; 7751 /* 7752 * We must wait so this happens before the 7753 * journal is reclaimed. 7754 */ 7755 ffs_update(vp, 1); 7756 } 7757 vput(vp); 7758 } 7759 if (spare < 0) { 7760 UFS_LOCK(ump); 7761 fs->fs_pendingblocks += spare; 7762 UFS_UNLOCK(ump); 7763 } 7764#ifdef QUOTA 7765 /* Handle spare. */ 7766 if (spare) 7767 quotaadj(freeblks->fb_quota, ump, -spare); 7768 quotarele(freeblks->fb_quota); 7769#endif 7770 ACQUIRE_LOCK(&lk); 7771 if (freeblks->fb_state & ONDEPLIST) { 7772 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7773 0, &inodedep); 7774 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 7775 freeblks->fb_state &= ~ONDEPLIST; 7776 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 7777 free_inodedep(inodedep); 7778 } 7779 /* 7780 * All of the freeblock deps must be complete prior to this call 7781 * so it's now safe to complete earlier outstanding journal entries. 7782 */ 7783 handle_jwork(&freeblks->fb_jwork); 7784 WORKITEM_FREE(freeblks, D_FREEBLKS); 7785 FREE_LOCK(&lk); 7786 return (0); 7787} 7788 7789/* 7790 * Release blocks associated with the freeblks and stored in the indirect 7791 * block dbn. If level is greater than SINGLE, the block is an indirect block 7792 * and recursive calls to indirtrunc must be used to cleanse other indirect 7793 * blocks. 7794 * 7795 * This handles partial and complete truncation of blocks. Partial is noted 7796 * with goingaway == 0. In this case the freework is completed after the 7797 * zero'd indirects are written to disk. For full truncation the freework 7798 * is completed after the block is freed. 7799 */ 7800static void 7801indir_trunc(freework, dbn, lbn) 7802 struct freework *freework; 7803 ufs2_daddr_t dbn; 7804 ufs_lbn_t lbn; 7805{ 7806 struct freework *nfreework; 7807 struct workhead wkhd; 7808 struct freeblks *freeblks; 7809 struct buf *bp; 7810 struct fs *fs; 7811 struct indirdep *indirdep; 7812 struct ufsmount *ump; 7813 ufs1_daddr_t *bap1 = 0; 7814 ufs2_daddr_t nb, nnb, *bap2 = 0; 7815 ufs_lbn_t lbnadd, nlbn; 7816 int i, nblocks, ufs1fmt; 7817 int freedblocks; 7818 int goingaway; 7819 int freedeps; 7820 int needj; 7821 int level; 7822 int cnt; 7823 7824 freeblks = freework->fw_freeblks; 7825 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7826 fs = ump->um_fs; 7827 /* 7828 * Get buffer of block pointers to be freed. There are three cases: 7829 * 7830 * 1) Partial truncate caches the indirdep pointer in the freework 7831 * which provides us a back copy to the save bp which holds the 7832 * pointers we want to clear. When this completes the zero 7833 * pointers are written to the real copy. 7834 * 2) The indirect is being completely truncated, cancel_indirdep() 7835 * eliminated the real copy and placed the indirdep on the saved 7836 * copy. The indirdep and buf are discarded when this completes. 7837 * 3) The indirect was not in memory, we read a copy off of the disk 7838 * using the devvp and drop and invalidate the buffer when we're 7839 * done. 7840 */ 7841 goingaway = 1; 7842 indirdep = NULL; 7843 if (freework->fw_indir != NULL) { 7844 goingaway = 0; 7845 indirdep = freework->fw_indir; 7846 bp = indirdep->ir_savebp; 7847 if (bp == NULL || bp->b_blkno != dbn) 7848 panic("indir_trunc: Bad saved buf %p blkno %jd", 7849 bp, (intmax_t)dbn); 7850 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 7851 /* 7852 * The lock prevents the buf dep list from changing and 7853 * indirects on devvp should only ever have one dependency. 7854 */ 7855 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 7856 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 7857 panic("indir_trunc: Bad indirdep %p from buf %p", 7858 indirdep, bp); 7859 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 7860 NOCRED, &bp) != 0) { 7861 brelse(bp); 7862 return; 7863 } 7864 ACQUIRE_LOCK(&lk); 7865 /* Protects against a race with complete_trunc_indir(). */ 7866 freework->fw_state &= ~INPROGRESS; 7867 /* 7868 * If we have an indirdep we need to enforce the truncation order 7869 * and discard it when it is complete. 7870 */ 7871 if (indirdep) { 7872 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 7873 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 7874 /* 7875 * Add the complete truncate to the list on the 7876 * indirdep to enforce in-order processing. 7877 */ 7878 if (freework->fw_indir == NULL) 7879 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 7880 freework, fw_next); 7881 FREE_LOCK(&lk); 7882 return; 7883 } 7884 /* 7885 * If we're goingaway, free the indirdep. Otherwise it will 7886 * linger until the write completes. 7887 */ 7888 if (goingaway) { 7889 free_indirdep(indirdep); 7890 ump->um_numindirdeps -= 1; 7891 } 7892 } 7893 FREE_LOCK(&lk); 7894 /* Initialize pointers depending on block size. */ 7895 if (ump->um_fstype == UFS1) { 7896 bap1 = (ufs1_daddr_t *)bp->b_data; 7897 nb = bap1[freework->fw_off]; 7898 ufs1fmt = 1; 7899 } else { 7900 bap2 = (ufs2_daddr_t *)bp->b_data; 7901 nb = bap2[freework->fw_off]; 7902 ufs1fmt = 0; 7903 } 7904 level = lbn_level(lbn); 7905 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 7906 lbnadd = lbn_offset(fs, level); 7907 nblocks = btodb(fs->fs_bsize); 7908 nfreework = freework; 7909 freedeps = 0; 7910 cnt = 0; 7911 /* 7912 * Reclaim blocks. Traverses into nested indirect levels and 7913 * arranges for the current level to be freed when subordinates 7914 * are free when journaling. 7915 */ 7916 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 7917 if (i != NINDIR(fs) - 1) { 7918 if (ufs1fmt) 7919 nnb = bap1[i+1]; 7920 else 7921 nnb = bap2[i+1]; 7922 } else 7923 nnb = 0; 7924 if (nb == 0) 7925 continue; 7926 cnt++; 7927 if (level != 0) { 7928 nlbn = (lbn + 1) - (i * lbnadd); 7929 if (needj != 0) { 7930 nfreework = newfreework(ump, freeblks, freework, 7931 nlbn, nb, fs->fs_frag, 0, 0); 7932 freedeps++; 7933 } 7934 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 7935 } else { 7936 struct freedep *freedep; 7937 7938 /* 7939 * Attempt to aggregate freedep dependencies for 7940 * all blocks being released to the same CG. 7941 */ 7942 LIST_INIT(&wkhd); 7943 if (needj != 0 && 7944 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 7945 freedep = newfreedep(freework); 7946 WORKLIST_INSERT_UNLOCKED(&wkhd, 7947 &freedep->fd_list); 7948 freedeps++; 7949 } 7950 CTR3(KTR_SUJ, 7951 "indir_trunc: ino %d blkno %jd size %ld", 7952 freeblks->fb_inum, nb, fs->fs_bsize); 7953 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 7954 fs->fs_bsize, freeblks->fb_inum, 7955 freeblks->fb_vtype, &wkhd); 7956 } 7957 } 7958 if (goingaway) { 7959 bp->b_flags |= B_INVAL | B_NOCACHE; 7960 brelse(bp); 7961 } 7962 freedblocks = 0; 7963 if (level == 0) 7964 freedblocks = (nblocks * cnt); 7965 if (needj == 0) 7966 freedblocks += nblocks; 7967 freeblks_free(ump, freeblks, freedblocks); 7968 /* 7969 * If we are journaling set up the ref counts and offset so this 7970 * indirect can be completed when its children are free. 7971 */ 7972 if (needj) { 7973 ACQUIRE_LOCK(&lk); 7974 freework->fw_off = i; 7975 freework->fw_ref += freedeps; 7976 freework->fw_ref -= NINDIR(fs) + 1; 7977 if (level == 0) 7978 freeblks->fb_cgwait += freedeps; 7979 if (freework->fw_ref == 0) 7980 freework_freeblock(freework); 7981 FREE_LOCK(&lk); 7982 return; 7983 } 7984 /* 7985 * If we're not journaling we can free the indirect now. 7986 */ 7987 dbn = dbtofsb(fs, dbn); 7988 CTR3(KTR_SUJ, 7989 "indir_trunc 2: ino %d blkno %jd size %ld", 7990 freeblks->fb_inum, dbn, fs->fs_bsize); 7991 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 7992 freeblks->fb_inum, freeblks->fb_vtype, NULL); 7993 /* Non SUJ softdep does single-threaded truncations. */ 7994 if (freework->fw_blkno == dbn) { 7995 freework->fw_state |= ALLCOMPLETE; 7996 ACQUIRE_LOCK(&lk); 7997 handle_written_freework(freework); 7998 FREE_LOCK(&lk); 7999 } 8000 return; 8001} 8002 8003/* 8004 * Cancel an allocindir when it is removed via truncation. When bp is not 8005 * NULL the indirect never appeared on disk and is scheduled to be freed 8006 * independently of the indir so we can more easily track journal work. 8007 */ 8008static void 8009cancel_allocindir(aip, bp, freeblks, trunc) 8010 struct allocindir *aip; 8011 struct buf *bp; 8012 struct freeblks *freeblks; 8013 int trunc; 8014{ 8015 struct indirdep *indirdep; 8016 struct freefrag *freefrag; 8017 struct newblk *newblk; 8018 8019 newblk = (struct newblk *)aip; 8020 LIST_REMOVE(aip, ai_next); 8021 /* 8022 * We must eliminate the pointer in bp if it must be freed on its 8023 * own due to partial truncate or pending journal work. 8024 */ 8025 if (bp && (trunc || newblk->nb_jnewblk)) { 8026 /* 8027 * Clear the pointer and mark the aip to be freed 8028 * directly if it never existed on disk. 8029 */ 8030 aip->ai_state |= DELAYEDFREE; 8031 indirdep = aip->ai_indirdep; 8032 if (indirdep->ir_state & UFS1FMT) 8033 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8034 else 8035 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8036 } 8037 /* 8038 * When truncating the previous pointer will be freed via 8039 * savedbp. Eliminate the freefrag which would dup free. 8040 */ 8041 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8042 newblk->nb_freefrag = NULL; 8043 if (freefrag->ff_jdep) 8044 cancel_jfreefrag( 8045 WK_JFREEFRAG(freefrag->ff_jdep)); 8046 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8047 WORKITEM_FREE(freefrag, D_FREEFRAG); 8048 } 8049 /* 8050 * If the journal hasn't been written the jnewblk must be passed 8051 * to the call to ffs_blkfree that reclaims the space. We accomplish 8052 * this by leaving the journal dependency on the newblk to be freed 8053 * when a freework is created in handle_workitem_freeblocks(). 8054 */ 8055 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8056 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8057} 8058 8059/* 8060 * Create the mkdir dependencies for . and .. in a new directory. Link them 8061 * in to a newdirblk so any subsequent additions are tracked properly. The 8062 * caller is responsible for adding the mkdir1 dependency to the journal 8063 * and updating id_mkdiradd. This function returns with lk held. 8064 */ 8065static struct mkdir * 8066setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8067 struct diradd *dap; 8068 ino_t newinum; 8069 ino_t dinum; 8070 struct buf *newdirbp; 8071 struct mkdir **mkdirp; 8072{ 8073 struct newblk *newblk; 8074 struct pagedep *pagedep; 8075 struct inodedep *inodedep; 8076 struct newdirblk *newdirblk = 0; 8077 struct mkdir *mkdir1, *mkdir2; 8078 struct worklist *wk; 8079 struct jaddref *jaddref; 8080 struct mount *mp; 8081 8082 mp = dap->da_list.wk_mp; 8083 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8084 M_SOFTDEP_FLAGS); 8085 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8086 LIST_INIT(&newdirblk->db_mkdir); 8087 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8088 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8089 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8090 mkdir1->md_diradd = dap; 8091 mkdir1->md_jaddref = NULL; 8092 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8093 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8094 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8095 mkdir2->md_diradd = dap; 8096 mkdir2->md_jaddref = NULL; 8097 if (MOUNTEDSUJ(mp) == 0) { 8098 mkdir1->md_state |= DEPCOMPLETE; 8099 mkdir2->md_state |= DEPCOMPLETE; 8100 } 8101 /* 8102 * Dependency on "." and ".." being written to disk. 8103 */ 8104 mkdir1->md_buf = newdirbp; 8105 ACQUIRE_LOCK(&lk); 8106 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs); 8107 /* 8108 * We must link the pagedep, allocdirect, and newdirblk for 8109 * the initial file page so the pointer to the new directory 8110 * is not written until the directory contents are live and 8111 * any subsequent additions are not marked live until the 8112 * block is reachable via the inode. 8113 */ 8114 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8115 panic("setup_newdir: lost pagedep"); 8116 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8117 if (wk->wk_type == D_ALLOCDIRECT) 8118 break; 8119 if (wk == NULL) 8120 panic("setup_newdir: lost allocdirect"); 8121 if (pagedep->pd_state & NEWBLOCK) 8122 panic("setup_newdir: NEWBLOCK already set"); 8123 newblk = WK_NEWBLK(wk); 8124 pagedep->pd_state |= NEWBLOCK; 8125 pagedep->pd_newdirblk = newdirblk; 8126 newdirblk->db_pagedep = pagedep; 8127 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8128 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8129 /* 8130 * Look up the inodedep for the parent directory so that we 8131 * can link mkdir2 into the pending dotdot jaddref or 8132 * the inode write if there is none. If the inode is 8133 * ALLCOMPLETE and no jaddref is present all dependencies have 8134 * been satisfied and mkdir2 can be freed. 8135 */ 8136 inodedep_lookup(mp, dinum, 0, &inodedep); 8137 if (MOUNTEDSUJ(mp)) { 8138 if (inodedep == NULL) 8139 panic("setup_newdir: Lost parent."); 8140 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8141 inoreflst); 8142 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8143 (jaddref->ja_state & MKDIR_PARENT), 8144 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8145 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 8146 mkdir2->md_jaddref = jaddref; 8147 jaddref->ja_mkdir = mkdir2; 8148 } else if (inodedep == NULL || 8149 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8150 dap->da_state &= ~MKDIR_PARENT; 8151 WORKITEM_FREE(mkdir2, D_MKDIR); 8152 } else { 8153 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs); 8154 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8155 } 8156 *mkdirp = mkdir2; 8157 8158 return (mkdir1); 8159} 8160 8161/* 8162 * Directory entry addition dependencies. 8163 * 8164 * When adding a new directory entry, the inode (with its incremented link 8165 * count) must be written to disk before the directory entry's pointer to it. 8166 * Also, if the inode is newly allocated, the corresponding freemap must be 8167 * updated (on disk) before the directory entry's pointer. These requirements 8168 * are met via undo/redo on the directory entry's pointer, which consists 8169 * simply of the inode number. 8170 * 8171 * As directory entries are added and deleted, the free space within a 8172 * directory block can become fragmented. The ufs filesystem will compact 8173 * a fragmented directory block to make space for a new entry. When this 8174 * occurs, the offsets of previously added entries change. Any "diradd" 8175 * dependency structures corresponding to these entries must be updated with 8176 * the new offsets. 8177 */ 8178 8179/* 8180 * This routine is called after the in-memory inode's link 8181 * count has been incremented, but before the directory entry's 8182 * pointer to the inode has been set. 8183 */ 8184int 8185softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8186 struct buf *bp; /* buffer containing directory block */ 8187 struct inode *dp; /* inode for directory */ 8188 off_t diroffset; /* offset of new entry in directory */ 8189 ino_t newinum; /* inode referenced by new directory entry */ 8190 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8191 int isnewblk; /* entry is in a newly allocated block */ 8192{ 8193 int offset; /* offset of new entry within directory block */ 8194 ufs_lbn_t lbn; /* block in directory containing new entry */ 8195 struct fs *fs; 8196 struct diradd *dap; 8197 struct newblk *newblk; 8198 struct pagedep *pagedep; 8199 struct inodedep *inodedep; 8200 struct newdirblk *newdirblk = 0; 8201 struct mkdir *mkdir1, *mkdir2; 8202 struct jaddref *jaddref; 8203 struct mount *mp; 8204 int isindir; 8205 8206 /* 8207 * Whiteouts have no dependencies. 8208 */ 8209 if (newinum == WINO) { 8210 if (newdirbp != NULL) 8211 bdwrite(newdirbp); 8212 return (0); 8213 } 8214 jaddref = NULL; 8215 mkdir1 = mkdir2 = NULL; 8216 mp = UFSTOVFS(dp->i_ump); 8217 fs = dp->i_fs; 8218 lbn = lblkno(fs, diroffset); 8219 offset = blkoff(fs, diroffset); 8220 dap = malloc(sizeof(struct diradd), M_DIRADD, 8221 M_SOFTDEP_FLAGS|M_ZERO); 8222 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8223 dap->da_offset = offset; 8224 dap->da_newinum = newinum; 8225 dap->da_state = ATTACHED; 8226 LIST_INIT(&dap->da_jwork); 8227 isindir = bp->b_lblkno >= NDADDR; 8228 if (isnewblk && 8229 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8230 newdirblk = malloc(sizeof(struct newdirblk), 8231 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8232 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8233 LIST_INIT(&newdirblk->db_mkdir); 8234 } 8235 /* 8236 * If we're creating a new directory setup the dependencies and set 8237 * the dap state to wait for them. Otherwise it's COMPLETE and 8238 * we can move on. 8239 */ 8240 if (newdirbp == NULL) { 8241 dap->da_state |= DEPCOMPLETE; 8242 ACQUIRE_LOCK(&lk); 8243 } else { 8244 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8245 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8246 &mkdir2); 8247 } 8248 /* 8249 * Link into parent directory pagedep to await its being written. 8250 */ 8251 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8252#ifdef DEBUG 8253 if (diradd_lookup(pagedep, offset) != NULL) 8254 panic("softdep_setup_directory_add: %p already at off %d\n", 8255 diradd_lookup(pagedep, offset), offset); 8256#endif 8257 dap->da_pagedep = pagedep; 8258 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8259 da_pdlist); 8260 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 8261 /* 8262 * If we're journaling, link the diradd into the jaddref so it 8263 * may be completed after the journal entry is written. Otherwise, 8264 * link the diradd into its inodedep. If the inode is not yet 8265 * written place it on the bufwait list, otherwise do the post-inode 8266 * write processing to put it on the id_pendinghd list. 8267 */ 8268 if (MOUNTEDSUJ(mp)) { 8269 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8270 inoreflst); 8271 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8272 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8273 jaddref->ja_diroff = diroffset; 8274 jaddref->ja_diradd = dap; 8275 add_to_journal(&jaddref->ja_list); 8276 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8277 diradd_inode_written(dap, inodedep); 8278 else 8279 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8280 /* 8281 * Add the journal entries for . and .. links now that the primary 8282 * link is written. 8283 */ 8284 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8285 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8286 inoreflst, if_deps); 8287 KASSERT(jaddref != NULL && 8288 jaddref->ja_ino == jaddref->ja_parent && 8289 (jaddref->ja_state & MKDIR_BODY), 8290 ("softdep_setup_directory_add: bad dot jaddref %p", 8291 jaddref)); 8292 mkdir1->md_jaddref = jaddref; 8293 jaddref->ja_mkdir = mkdir1; 8294 /* 8295 * It is important that the dotdot journal entry 8296 * is added prior to the dot entry since dot writes 8297 * both the dot and dotdot links. These both must 8298 * be added after the primary link for the journal 8299 * to remain consistent. 8300 */ 8301 add_to_journal(&mkdir2->md_jaddref->ja_list); 8302 add_to_journal(&jaddref->ja_list); 8303 } 8304 /* 8305 * If we are adding a new directory remember this diradd so that if 8306 * we rename it we can keep the dot and dotdot dependencies. If 8307 * we are adding a new name for an inode that has a mkdiradd we 8308 * must be in rename and we have to move the dot and dotdot 8309 * dependencies to this new name. The old name is being orphaned 8310 * soon. 8311 */ 8312 if (mkdir1 != NULL) { 8313 if (inodedep->id_mkdiradd != NULL) 8314 panic("softdep_setup_directory_add: Existing mkdir"); 8315 inodedep->id_mkdiradd = dap; 8316 } else if (inodedep->id_mkdiradd) 8317 merge_diradd(inodedep, dap); 8318 if (newdirblk) { 8319 /* 8320 * There is nothing to do if we are already tracking 8321 * this block. 8322 */ 8323 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8324 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8325 FREE_LOCK(&lk); 8326 return (0); 8327 } 8328 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8329 == 0) 8330 panic("softdep_setup_directory_add: lost entry"); 8331 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8332 pagedep->pd_state |= NEWBLOCK; 8333 pagedep->pd_newdirblk = newdirblk; 8334 newdirblk->db_pagedep = pagedep; 8335 FREE_LOCK(&lk); 8336 /* 8337 * If we extended into an indirect signal direnter to sync. 8338 */ 8339 if (isindir) 8340 return (1); 8341 return (0); 8342 } 8343 FREE_LOCK(&lk); 8344 return (0); 8345} 8346 8347/* 8348 * This procedure is called to change the offset of a directory 8349 * entry when compacting a directory block which must be owned 8350 * exclusively by the caller. Note that the actual entry movement 8351 * must be done in this procedure to ensure that no I/O completions 8352 * occur while the move is in progress. 8353 */ 8354void 8355softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8356 struct buf *bp; /* Buffer holding directory block. */ 8357 struct inode *dp; /* inode for directory */ 8358 caddr_t base; /* address of dp->i_offset */ 8359 caddr_t oldloc; /* address of old directory location */ 8360 caddr_t newloc; /* address of new directory location */ 8361 int entrysize; /* size of directory entry */ 8362{ 8363 int offset, oldoffset, newoffset; 8364 struct pagedep *pagedep; 8365 struct jmvref *jmvref; 8366 struct diradd *dap; 8367 struct direct *de; 8368 struct mount *mp; 8369 ufs_lbn_t lbn; 8370 int flags; 8371 8372 mp = UFSTOVFS(dp->i_ump); 8373 de = (struct direct *)oldloc; 8374 jmvref = NULL; 8375 flags = 0; 8376 /* 8377 * Moves are always journaled as it would be too complex to 8378 * determine if any affected adds or removes are present in the 8379 * journal. 8380 */ 8381 if (MOUNTEDSUJ(mp)) { 8382 flags = DEPALLOC; 8383 jmvref = newjmvref(dp, de->d_ino, 8384 dp->i_offset + (oldloc - base), 8385 dp->i_offset + (newloc - base)); 8386 } 8387 lbn = lblkno(dp->i_fs, dp->i_offset); 8388 offset = blkoff(dp->i_fs, dp->i_offset); 8389 oldoffset = offset + (oldloc - base); 8390 newoffset = offset + (newloc - base); 8391 ACQUIRE_LOCK(&lk); 8392 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8393 goto done; 8394 dap = diradd_lookup(pagedep, oldoffset); 8395 if (dap) { 8396 dap->da_offset = newoffset; 8397 newoffset = DIRADDHASH(newoffset); 8398 oldoffset = DIRADDHASH(oldoffset); 8399 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8400 newoffset != oldoffset) { 8401 LIST_REMOVE(dap, da_pdlist); 8402 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8403 dap, da_pdlist); 8404 } 8405 } 8406done: 8407 if (jmvref) { 8408 jmvref->jm_pagedep = pagedep; 8409 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8410 add_to_journal(&jmvref->jm_list); 8411 } 8412 bcopy(oldloc, newloc, entrysize); 8413 FREE_LOCK(&lk); 8414} 8415 8416/* 8417 * Move the mkdir dependencies and journal work from one diradd to another 8418 * when renaming a directory. The new name must depend on the mkdir deps 8419 * completing as the old name did. Directories can only have one valid link 8420 * at a time so one must be canonical. 8421 */ 8422static void 8423merge_diradd(inodedep, newdap) 8424 struct inodedep *inodedep; 8425 struct diradd *newdap; 8426{ 8427 struct diradd *olddap; 8428 struct mkdir *mkdir, *nextmd; 8429 short state; 8430 8431 olddap = inodedep->id_mkdiradd; 8432 inodedep->id_mkdiradd = newdap; 8433 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8434 newdap->da_state &= ~DEPCOMPLETE; 8435 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 8436 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8437 if (mkdir->md_diradd != olddap) 8438 continue; 8439 mkdir->md_diradd = newdap; 8440 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8441 newdap->da_state |= state; 8442 olddap->da_state &= ~state; 8443 if ((olddap->da_state & 8444 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8445 break; 8446 } 8447 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8448 panic("merge_diradd: unfound ref"); 8449 } 8450 /* 8451 * Any mkdir related journal items are not safe to be freed until 8452 * the new name is stable. 8453 */ 8454 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8455 olddap->da_state |= DEPCOMPLETE; 8456 complete_diradd(olddap); 8457} 8458 8459/* 8460 * Move the diradd to the pending list when all diradd dependencies are 8461 * complete. 8462 */ 8463static void 8464complete_diradd(dap) 8465 struct diradd *dap; 8466{ 8467 struct pagedep *pagedep; 8468 8469 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8470 if (dap->da_state & DIRCHG) 8471 pagedep = dap->da_previous->dm_pagedep; 8472 else 8473 pagedep = dap->da_pagedep; 8474 LIST_REMOVE(dap, da_pdlist); 8475 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8476 } 8477} 8478 8479/* 8480 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8481 * add entries and conditonally journal the remove. 8482 */ 8483static void 8484cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8485 struct diradd *dap; 8486 struct dirrem *dirrem; 8487 struct jremref *jremref; 8488 struct jremref *dotremref; 8489 struct jremref *dotdotremref; 8490{ 8491 struct inodedep *inodedep; 8492 struct jaddref *jaddref; 8493 struct inoref *inoref; 8494 struct mkdir *mkdir; 8495 8496 /* 8497 * If no remove references were allocated we're on a non-journaled 8498 * filesystem and can skip the cancel step. 8499 */ 8500 if (jremref == NULL) { 8501 free_diradd(dap, NULL); 8502 return; 8503 } 8504 /* 8505 * Cancel the primary name an free it if it does not require 8506 * journaling. 8507 */ 8508 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8509 0, &inodedep) != 0) { 8510 /* Abort the addref that reference this diradd. */ 8511 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8512 if (inoref->if_list.wk_type != D_JADDREF) 8513 continue; 8514 jaddref = (struct jaddref *)inoref; 8515 if (jaddref->ja_diradd != dap) 8516 continue; 8517 if (cancel_jaddref(jaddref, inodedep, 8518 &dirrem->dm_jwork) == 0) { 8519 free_jremref(jremref); 8520 jremref = NULL; 8521 } 8522 break; 8523 } 8524 } 8525 /* 8526 * Cancel subordinate names and free them if they do not require 8527 * journaling. 8528 */ 8529 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8530 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 8531 if (mkdir->md_diradd != dap) 8532 continue; 8533 if ((jaddref = mkdir->md_jaddref) == NULL) 8534 continue; 8535 mkdir->md_jaddref = NULL; 8536 if (mkdir->md_state & MKDIR_PARENT) { 8537 if (cancel_jaddref(jaddref, NULL, 8538 &dirrem->dm_jwork) == 0) { 8539 free_jremref(dotdotremref); 8540 dotdotremref = NULL; 8541 } 8542 } else { 8543 if (cancel_jaddref(jaddref, inodedep, 8544 &dirrem->dm_jwork) == 0) { 8545 free_jremref(dotremref); 8546 dotremref = NULL; 8547 } 8548 } 8549 } 8550 } 8551 8552 if (jremref) 8553 journal_jremref(dirrem, jremref, inodedep); 8554 if (dotremref) 8555 journal_jremref(dirrem, dotremref, inodedep); 8556 if (dotdotremref) 8557 journal_jremref(dirrem, dotdotremref, NULL); 8558 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8559 free_diradd(dap, &dirrem->dm_jwork); 8560} 8561 8562/* 8563 * Free a diradd dependency structure. This routine must be called 8564 * with splbio interrupts blocked. 8565 */ 8566static void 8567free_diradd(dap, wkhd) 8568 struct diradd *dap; 8569 struct workhead *wkhd; 8570{ 8571 struct dirrem *dirrem; 8572 struct pagedep *pagedep; 8573 struct inodedep *inodedep; 8574 struct mkdir *mkdir, *nextmd; 8575 8576 mtx_assert(&lk, MA_OWNED); 8577 LIST_REMOVE(dap, da_pdlist); 8578 if (dap->da_state & ONWORKLIST) 8579 WORKLIST_REMOVE(&dap->da_list); 8580 if ((dap->da_state & DIRCHG) == 0) { 8581 pagedep = dap->da_pagedep; 8582 } else { 8583 dirrem = dap->da_previous; 8584 pagedep = dirrem->dm_pagedep; 8585 dirrem->dm_dirinum = pagedep->pd_ino; 8586 dirrem->dm_state |= COMPLETE; 8587 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8588 add_to_worklist(&dirrem->dm_list, 0); 8589 } 8590 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8591 0, &inodedep) != 0) 8592 if (inodedep->id_mkdiradd == dap) 8593 inodedep->id_mkdiradd = NULL; 8594 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8595 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) { 8596 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8597 if (mkdir->md_diradd != dap) 8598 continue; 8599 dap->da_state &= 8600 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8601 LIST_REMOVE(mkdir, md_mkdirs); 8602 if (mkdir->md_state & ONWORKLIST) 8603 WORKLIST_REMOVE(&mkdir->md_list); 8604 if (mkdir->md_jaddref != NULL) 8605 panic("free_diradd: Unexpected jaddref"); 8606 WORKITEM_FREE(mkdir, D_MKDIR); 8607 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8608 break; 8609 } 8610 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8611 panic("free_diradd: unfound ref"); 8612 } 8613 if (inodedep) 8614 free_inodedep(inodedep); 8615 /* 8616 * Free any journal segments waiting for the directory write. 8617 */ 8618 handle_jwork(&dap->da_jwork); 8619 WORKITEM_FREE(dap, D_DIRADD); 8620} 8621 8622/* 8623 * Directory entry removal dependencies. 8624 * 8625 * When removing a directory entry, the entry's inode pointer must be 8626 * zero'ed on disk before the corresponding inode's link count is decremented 8627 * (possibly freeing the inode for re-use). This dependency is handled by 8628 * updating the directory entry but delaying the inode count reduction until 8629 * after the directory block has been written to disk. After this point, the 8630 * inode count can be decremented whenever it is convenient. 8631 */ 8632 8633/* 8634 * This routine should be called immediately after removing 8635 * a directory entry. The inode's link count should not be 8636 * decremented by the calling procedure -- the soft updates 8637 * code will do this task when it is safe. 8638 */ 8639void 8640softdep_setup_remove(bp, dp, ip, isrmdir) 8641 struct buf *bp; /* buffer containing directory block */ 8642 struct inode *dp; /* inode for the directory being modified */ 8643 struct inode *ip; /* inode for directory entry being removed */ 8644 int isrmdir; /* indicates if doing RMDIR */ 8645{ 8646 struct dirrem *dirrem, *prevdirrem; 8647 struct inodedep *inodedep; 8648 int direct; 8649 8650 /* 8651 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8652 * newdirrem() to setup the full directory remove which requires 8653 * isrmdir > 1. 8654 */ 8655 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8656 /* 8657 * Add the dirrem to the inodedep's pending remove list for quick 8658 * discovery later. 8659 */ 8660 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8661 &inodedep) == 0) 8662 panic("softdep_setup_remove: Lost inodedep."); 8663 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8664 dirrem->dm_state |= ONDEPLIST; 8665 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8666 8667 /* 8668 * If the COMPLETE flag is clear, then there were no active 8669 * entries and we want to roll back to a zeroed entry until 8670 * the new inode is committed to disk. If the COMPLETE flag is 8671 * set then we have deleted an entry that never made it to 8672 * disk. If the entry we deleted resulted from a name change, 8673 * then the old name still resides on disk. We cannot delete 8674 * its inode (returned to us in prevdirrem) until the zeroed 8675 * directory entry gets to disk. The new inode has never been 8676 * referenced on the disk, so can be deleted immediately. 8677 */ 8678 if ((dirrem->dm_state & COMPLETE) == 0) { 8679 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8680 dm_next); 8681 FREE_LOCK(&lk); 8682 } else { 8683 if (prevdirrem != NULL) 8684 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8685 prevdirrem, dm_next); 8686 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8687 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8688 FREE_LOCK(&lk); 8689 if (direct) 8690 handle_workitem_remove(dirrem, 0); 8691 } 8692} 8693 8694/* 8695 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8696 * pd_pendinghd list of a pagedep. 8697 */ 8698static struct diradd * 8699diradd_lookup(pagedep, offset) 8700 struct pagedep *pagedep; 8701 int offset; 8702{ 8703 struct diradd *dap; 8704 8705 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8706 if (dap->da_offset == offset) 8707 return (dap); 8708 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 8709 if (dap->da_offset == offset) 8710 return (dap); 8711 return (NULL); 8712} 8713 8714/* 8715 * Search for a .. diradd dependency in a directory that is being removed. 8716 * If the directory was renamed to a new parent we have a diradd rather 8717 * than a mkdir for the .. entry. We need to cancel it now before 8718 * it is found in truncate(). 8719 */ 8720static struct jremref * 8721cancel_diradd_dotdot(ip, dirrem, jremref) 8722 struct inode *ip; 8723 struct dirrem *dirrem; 8724 struct jremref *jremref; 8725{ 8726 struct pagedep *pagedep; 8727 struct diradd *dap; 8728 struct worklist *wk; 8729 8730 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0, 8731 &pagedep) == 0) 8732 return (jremref); 8733 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 8734 if (dap == NULL) 8735 return (jremref); 8736 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 8737 /* 8738 * Mark any journal work as belonging to the parent so it is freed 8739 * with the .. reference. 8740 */ 8741 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8742 wk->wk_state |= MKDIR_PARENT; 8743 return (NULL); 8744} 8745 8746/* 8747 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 8748 * replace it with a dirrem/diradd pair as a result of re-parenting a 8749 * directory. This ensures that we don't simultaneously have a mkdir and 8750 * a diradd for the same .. entry. 8751 */ 8752static struct jremref * 8753cancel_mkdir_dotdot(ip, dirrem, jremref) 8754 struct inode *ip; 8755 struct dirrem *dirrem; 8756 struct jremref *jremref; 8757{ 8758 struct inodedep *inodedep; 8759 struct jaddref *jaddref; 8760 struct mkdir *mkdir; 8761 struct diradd *dap; 8762 8763 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8764 &inodedep) == 0) 8765 return (jremref); 8766 dap = inodedep->id_mkdiradd; 8767 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 8768 return (jremref); 8769 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; 8770 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 8771 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 8772 break; 8773 if (mkdir == NULL) 8774 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 8775 if ((jaddref = mkdir->md_jaddref) != NULL) { 8776 mkdir->md_jaddref = NULL; 8777 jaddref->ja_state &= ~MKDIR_PARENT; 8778 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0, 8779 &inodedep) == 0) 8780 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 8781 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 8782 journal_jremref(dirrem, jremref, inodedep); 8783 jremref = NULL; 8784 } 8785 } 8786 if (mkdir->md_state & ONWORKLIST) 8787 WORKLIST_REMOVE(&mkdir->md_list); 8788 mkdir->md_state |= ALLCOMPLETE; 8789 complete_mkdir(mkdir); 8790 return (jremref); 8791} 8792 8793static void 8794journal_jremref(dirrem, jremref, inodedep) 8795 struct dirrem *dirrem; 8796 struct jremref *jremref; 8797 struct inodedep *inodedep; 8798{ 8799 8800 if (inodedep == NULL) 8801 if (inodedep_lookup(jremref->jr_list.wk_mp, 8802 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 8803 panic("journal_jremref: Lost inodedep"); 8804 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 8805 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 8806 add_to_journal(&jremref->jr_list); 8807} 8808 8809static void 8810dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 8811 struct dirrem *dirrem; 8812 struct jremref *jremref; 8813 struct jremref *dotremref; 8814 struct jremref *dotdotremref; 8815{ 8816 struct inodedep *inodedep; 8817 8818 8819 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 8820 &inodedep) == 0) 8821 panic("dirrem_journal: Lost inodedep"); 8822 journal_jremref(dirrem, jremref, inodedep); 8823 if (dotremref) 8824 journal_jremref(dirrem, dotremref, inodedep); 8825 if (dotdotremref) 8826 journal_jremref(dirrem, dotdotremref, NULL); 8827} 8828 8829/* 8830 * Allocate a new dirrem if appropriate and return it along with 8831 * its associated pagedep. Called without a lock, returns with lock. 8832 */ 8833static struct dirrem * 8834newdirrem(bp, dp, ip, isrmdir, prevdirremp) 8835 struct buf *bp; /* buffer containing directory block */ 8836 struct inode *dp; /* inode for the directory being modified */ 8837 struct inode *ip; /* inode for directory entry being removed */ 8838 int isrmdir; /* indicates if doing RMDIR */ 8839 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 8840{ 8841 int offset; 8842 ufs_lbn_t lbn; 8843 struct diradd *dap; 8844 struct dirrem *dirrem; 8845 struct pagedep *pagedep; 8846 struct jremref *jremref; 8847 struct jremref *dotremref; 8848 struct jremref *dotdotremref; 8849 struct vnode *dvp; 8850 8851 /* 8852 * Whiteouts have no deletion dependencies. 8853 */ 8854 if (ip == NULL) 8855 panic("newdirrem: whiteout"); 8856 dvp = ITOV(dp); 8857 /* 8858 * If we are over our limit, try to improve the situation. 8859 * Limiting the number of dirrem structures will also limit 8860 * the number of freefile and freeblks structures. 8861 */ 8862 ACQUIRE_LOCK(&lk); 8863 if (!IS_SNAPSHOT(ip) && dep_current[D_DIRREM] > max_softdeps / 2) 8864 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_BLOCKS); 8865 FREE_LOCK(&lk); 8866 dirrem = malloc(sizeof(struct dirrem), 8867 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO); 8868 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 8869 LIST_INIT(&dirrem->dm_jremrefhd); 8870 LIST_INIT(&dirrem->dm_jwork); 8871 dirrem->dm_state = isrmdir ? RMDIR : 0; 8872 dirrem->dm_oldinum = ip->i_number; 8873 *prevdirremp = NULL; 8874 /* 8875 * Allocate remove reference structures to track journal write 8876 * dependencies. We will always have one for the link and 8877 * when doing directories we will always have one more for dot. 8878 * When renaming a directory we skip the dotdot link change so 8879 * this is not needed. 8880 */ 8881 jremref = dotremref = dotdotremref = NULL; 8882 if (DOINGSUJ(dvp)) { 8883 if (isrmdir) { 8884 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 8885 ip->i_effnlink + 2); 8886 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 8887 ip->i_effnlink + 1); 8888 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 8889 dp->i_effnlink + 1); 8890 dotdotremref->jr_state |= MKDIR_PARENT; 8891 } else 8892 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 8893 ip->i_effnlink + 1); 8894 } 8895 ACQUIRE_LOCK(&lk); 8896 lbn = lblkno(dp->i_fs, dp->i_offset); 8897 offset = blkoff(dp->i_fs, dp->i_offset); 8898 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC, 8899 &pagedep); 8900 dirrem->dm_pagedep = pagedep; 8901 dirrem->dm_offset = offset; 8902 /* 8903 * If we're renaming a .. link to a new directory, cancel any 8904 * existing MKDIR_PARENT mkdir. If it has already been canceled 8905 * the jremref is preserved for any potential diradd in this 8906 * location. This can not coincide with a rmdir. 8907 */ 8908 if (dp->i_offset == DOTDOT_OFFSET) { 8909 if (isrmdir) 8910 panic("newdirrem: .. directory change during remove?"); 8911 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 8912 } 8913 /* 8914 * If we're removing a directory search for the .. dependency now and 8915 * cancel it. Any pending journal work will be added to the dirrem 8916 * to be completed when the workitem remove completes. 8917 */ 8918 if (isrmdir) 8919 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 8920 /* 8921 * Check for a diradd dependency for the same directory entry. 8922 * If present, then both dependencies become obsolete and can 8923 * be de-allocated. 8924 */ 8925 dap = diradd_lookup(pagedep, offset); 8926 if (dap == NULL) { 8927 /* 8928 * Link the jremref structures into the dirrem so they are 8929 * written prior to the pagedep. 8930 */ 8931 if (jremref) 8932 dirrem_journal(dirrem, jremref, dotremref, 8933 dotdotremref); 8934 return (dirrem); 8935 } 8936 /* 8937 * Must be ATTACHED at this point. 8938 */ 8939 if ((dap->da_state & ATTACHED) == 0) 8940 panic("newdirrem: not ATTACHED"); 8941 if (dap->da_newinum != ip->i_number) 8942 panic("newdirrem: inum %ju should be %ju", 8943 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 8944 /* 8945 * If we are deleting a changed name that never made it to disk, 8946 * then return the dirrem describing the previous inode (which 8947 * represents the inode currently referenced from this entry on disk). 8948 */ 8949 if ((dap->da_state & DIRCHG) != 0) { 8950 *prevdirremp = dap->da_previous; 8951 dap->da_state &= ~DIRCHG; 8952 dap->da_pagedep = pagedep; 8953 } 8954 /* 8955 * We are deleting an entry that never made it to disk. 8956 * Mark it COMPLETE so we can delete its inode immediately. 8957 */ 8958 dirrem->dm_state |= COMPLETE; 8959 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 8960#ifdef SUJ_DEBUG 8961 if (isrmdir == 0) { 8962 struct worklist *wk; 8963 8964 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8965 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 8966 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 8967 } 8968#endif 8969 8970 return (dirrem); 8971} 8972 8973/* 8974 * Directory entry change dependencies. 8975 * 8976 * Changing an existing directory entry requires that an add operation 8977 * be completed first followed by a deletion. The semantics for the addition 8978 * are identical to the description of adding a new entry above except 8979 * that the rollback is to the old inode number rather than zero. Once 8980 * the addition dependency is completed, the removal is done as described 8981 * in the removal routine above. 8982 */ 8983 8984/* 8985 * This routine should be called immediately after changing 8986 * a directory entry. The inode's link count should not be 8987 * decremented by the calling procedure -- the soft updates 8988 * code will perform this task when it is safe. 8989 */ 8990void 8991softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 8992 struct buf *bp; /* buffer containing directory block */ 8993 struct inode *dp; /* inode for the directory being modified */ 8994 struct inode *ip; /* inode for directory entry being removed */ 8995 ino_t newinum; /* new inode number for changed entry */ 8996 int isrmdir; /* indicates if doing RMDIR */ 8997{ 8998 int offset; 8999 struct diradd *dap = NULL; 9000 struct dirrem *dirrem, *prevdirrem; 9001 struct pagedep *pagedep; 9002 struct inodedep *inodedep; 9003 struct jaddref *jaddref; 9004 struct mount *mp; 9005 9006 offset = blkoff(dp->i_fs, dp->i_offset); 9007 mp = UFSTOVFS(dp->i_ump); 9008 9009 /* 9010 * Whiteouts do not need diradd dependencies. 9011 */ 9012 if (newinum != WINO) { 9013 dap = malloc(sizeof(struct diradd), 9014 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9015 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9016 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9017 dap->da_offset = offset; 9018 dap->da_newinum = newinum; 9019 LIST_INIT(&dap->da_jwork); 9020 } 9021 9022 /* 9023 * Allocate a new dirrem and ACQUIRE_LOCK. 9024 */ 9025 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9026 pagedep = dirrem->dm_pagedep; 9027 /* 9028 * The possible values for isrmdir: 9029 * 0 - non-directory file rename 9030 * 1 - directory rename within same directory 9031 * inum - directory rename to new directory of given inode number 9032 * When renaming to a new directory, we are both deleting and 9033 * creating a new directory entry, so the link count on the new 9034 * directory should not change. Thus we do not need the followup 9035 * dirrem which is usually done in handle_workitem_remove. We set 9036 * the DIRCHG flag to tell handle_workitem_remove to skip the 9037 * followup dirrem. 9038 */ 9039 if (isrmdir > 1) 9040 dirrem->dm_state |= DIRCHG; 9041 9042 /* 9043 * Whiteouts have no additional dependencies, 9044 * so just put the dirrem on the correct list. 9045 */ 9046 if (newinum == WINO) { 9047 if ((dirrem->dm_state & COMPLETE) == 0) { 9048 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9049 dm_next); 9050 } else { 9051 dirrem->dm_dirinum = pagedep->pd_ino; 9052 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9053 add_to_worklist(&dirrem->dm_list, 0); 9054 } 9055 FREE_LOCK(&lk); 9056 return; 9057 } 9058 /* 9059 * Add the dirrem to the inodedep's pending remove list for quick 9060 * discovery later. A valid nlinkdelta ensures that this lookup 9061 * will not fail. 9062 */ 9063 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9064 panic("softdep_setup_directory_change: Lost inodedep."); 9065 dirrem->dm_state |= ONDEPLIST; 9066 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9067 9068 /* 9069 * If the COMPLETE flag is clear, then there were no active 9070 * entries and we want to roll back to the previous inode until 9071 * the new inode is committed to disk. If the COMPLETE flag is 9072 * set, then we have deleted an entry that never made it to disk. 9073 * If the entry we deleted resulted from a name change, then the old 9074 * inode reference still resides on disk. Any rollback that we do 9075 * needs to be to that old inode (returned to us in prevdirrem). If 9076 * the entry we deleted resulted from a create, then there is 9077 * no entry on the disk, so we want to roll back to zero rather 9078 * than the uncommitted inode. In either of the COMPLETE cases we 9079 * want to immediately free the unwritten and unreferenced inode. 9080 */ 9081 if ((dirrem->dm_state & COMPLETE) == 0) { 9082 dap->da_previous = dirrem; 9083 } else { 9084 if (prevdirrem != NULL) { 9085 dap->da_previous = prevdirrem; 9086 } else { 9087 dap->da_state &= ~DIRCHG; 9088 dap->da_pagedep = pagedep; 9089 } 9090 dirrem->dm_dirinum = pagedep->pd_ino; 9091 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9092 add_to_worklist(&dirrem->dm_list, 0); 9093 } 9094 /* 9095 * Lookup the jaddref for this journal entry. We must finish 9096 * initializing it and make the diradd write dependent on it. 9097 * If we're not journaling, put it on the id_bufwait list if the 9098 * inode is not yet written. If it is written, do the post-inode 9099 * write processing to put it on the id_pendinghd list. 9100 */ 9101 inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep); 9102 if (MOUNTEDSUJ(mp)) { 9103 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9104 inoreflst); 9105 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9106 ("softdep_setup_directory_change: bad jaddref %p", 9107 jaddref)); 9108 jaddref->ja_diroff = dp->i_offset; 9109 jaddref->ja_diradd = dap; 9110 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9111 dap, da_pdlist); 9112 add_to_journal(&jaddref->ja_list); 9113 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9114 dap->da_state |= COMPLETE; 9115 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9116 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9117 } else { 9118 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9119 dap, da_pdlist); 9120 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9121 } 9122 /* 9123 * If we're making a new name for a directory that has not been 9124 * committed when need to move the dot and dotdot references to 9125 * this new name. 9126 */ 9127 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9128 merge_diradd(inodedep, dap); 9129 FREE_LOCK(&lk); 9130} 9131 9132/* 9133 * Called whenever the link count on an inode is changed. 9134 * It creates an inode dependency so that the new reference(s) 9135 * to the inode cannot be committed to disk until the updated 9136 * inode has been written. 9137 */ 9138void 9139softdep_change_linkcnt(ip) 9140 struct inode *ip; /* the inode with the increased link count */ 9141{ 9142 struct inodedep *inodedep; 9143 int dflags; 9144 9145 ACQUIRE_LOCK(&lk); 9146 dflags = DEPALLOC; 9147 if (IS_SNAPSHOT(ip)) 9148 dflags |= NODELAY; 9149 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep); 9150 if (ip->i_nlink < ip->i_effnlink) 9151 panic("softdep_change_linkcnt: bad delta"); 9152 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9153 FREE_LOCK(&lk); 9154} 9155 9156/* 9157 * Attach a sbdep dependency to the superblock buf so that we can keep 9158 * track of the head of the linked list of referenced but unlinked inodes. 9159 */ 9160void 9161softdep_setup_sbupdate(ump, fs, bp) 9162 struct ufsmount *ump; 9163 struct fs *fs; 9164 struct buf *bp; 9165{ 9166 struct sbdep *sbdep; 9167 struct worklist *wk; 9168 9169 if (MOUNTEDSUJ(UFSTOVFS(ump)) == 0) 9170 return; 9171 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9172 if (wk->wk_type == D_SBDEP) 9173 break; 9174 if (wk != NULL) 9175 return; 9176 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9177 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9178 sbdep->sb_fs = fs; 9179 sbdep->sb_ump = ump; 9180 ACQUIRE_LOCK(&lk); 9181 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9182 FREE_LOCK(&lk); 9183} 9184 9185/* 9186 * Return the first unlinked inodedep which is ready to be the head of the 9187 * list. The inodedep and all those after it must have valid next pointers. 9188 */ 9189static struct inodedep * 9190first_unlinked_inodedep(ump) 9191 struct ufsmount *ump; 9192{ 9193 struct inodedep *inodedep; 9194 struct inodedep *idp; 9195 9196 mtx_assert(&lk, MA_OWNED); 9197 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9198 inodedep; inodedep = idp) { 9199 if ((inodedep->id_state & UNLINKNEXT) == 0) 9200 return (NULL); 9201 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9202 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9203 break; 9204 if ((inodedep->id_state & UNLINKPREV) == 0) 9205 break; 9206 } 9207 return (inodedep); 9208} 9209 9210/* 9211 * Set the sujfree unlinked head pointer prior to writing a superblock. 9212 */ 9213static void 9214initiate_write_sbdep(sbdep) 9215 struct sbdep *sbdep; 9216{ 9217 struct inodedep *inodedep; 9218 struct fs *bpfs; 9219 struct fs *fs; 9220 9221 bpfs = sbdep->sb_fs; 9222 fs = sbdep->sb_ump->um_fs; 9223 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9224 if (inodedep) { 9225 fs->fs_sujfree = inodedep->id_ino; 9226 inodedep->id_state |= UNLINKPREV; 9227 } else 9228 fs->fs_sujfree = 0; 9229 bpfs->fs_sujfree = fs->fs_sujfree; 9230} 9231 9232/* 9233 * After a superblock is written determine whether it must be written again 9234 * due to a changing unlinked list head. 9235 */ 9236static int 9237handle_written_sbdep(sbdep, bp) 9238 struct sbdep *sbdep; 9239 struct buf *bp; 9240{ 9241 struct inodedep *inodedep; 9242 struct mount *mp; 9243 struct fs *fs; 9244 9245 mtx_assert(&lk, MA_OWNED); 9246 fs = sbdep->sb_fs; 9247 mp = UFSTOVFS(sbdep->sb_ump); 9248 /* 9249 * If the superblock doesn't match the in-memory list start over. 9250 */ 9251 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9252 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9253 (inodedep == NULL && fs->fs_sujfree != 0)) { 9254 bdirty(bp); 9255 return (1); 9256 } 9257 WORKITEM_FREE(sbdep, D_SBDEP); 9258 if (fs->fs_sujfree == 0) 9259 return (0); 9260 /* 9261 * Now that we have a record of this inode in stable store allow it 9262 * to be written to free up pending work. Inodes may see a lot of 9263 * write activity after they are unlinked which we must not hold up. 9264 */ 9265 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9266 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9267 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9268 inodedep, inodedep->id_state); 9269 if (inodedep->id_state & UNLINKONLIST) 9270 break; 9271 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9272 } 9273 9274 return (0); 9275} 9276 9277/* 9278 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9279 */ 9280static void 9281unlinked_inodedep(mp, inodedep) 9282 struct mount *mp; 9283 struct inodedep *inodedep; 9284{ 9285 struct ufsmount *ump; 9286 9287 mtx_assert(&lk, MA_OWNED); 9288 if (MOUNTEDSUJ(mp) == 0) 9289 return; 9290 ump = VFSTOUFS(mp); 9291 ump->um_fs->fs_fmod = 1; 9292 if (inodedep->id_state & UNLINKED) 9293 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9294 inodedep->id_state |= UNLINKED; 9295 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9296} 9297 9298/* 9299 * Remove an inodedep from the unlinked inodedep list. This may require 9300 * disk writes if the inode has made it that far. 9301 */ 9302static void 9303clear_unlinked_inodedep(inodedep) 9304 struct inodedep *inodedep; 9305{ 9306 struct ufsmount *ump; 9307 struct inodedep *idp; 9308 struct inodedep *idn; 9309 struct fs *fs; 9310 struct buf *bp; 9311 ino_t ino; 9312 ino_t nino; 9313 ino_t pino; 9314 int error; 9315 9316 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9317 fs = ump->um_fs; 9318 ino = inodedep->id_ino; 9319 error = 0; 9320 for (;;) { 9321 mtx_assert(&lk, MA_OWNED); 9322 KASSERT((inodedep->id_state & UNLINKED) != 0, 9323 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9324 inodedep)); 9325 /* 9326 * If nothing has yet been written simply remove us from 9327 * the in memory list and return. This is the most common 9328 * case where handle_workitem_remove() loses the final 9329 * reference. 9330 */ 9331 if ((inodedep->id_state & UNLINKLINKS) == 0) 9332 break; 9333 /* 9334 * If we have a NEXT pointer and no PREV pointer we can simply 9335 * clear NEXT's PREV and remove ourselves from the list. Be 9336 * careful not to clear PREV if the superblock points at 9337 * next as well. 9338 */ 9339 idn = TAILQ_NEXT(inodedep, id_unlinked); 9340 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9341 if (idn && fs->fs_sujfree != idn->id_ino) 9342 idn->id_state &= ~UNLINKPREV; 9343 break; 9344 } 9345 /* 9346 * Here we have an inodedep which is actually linked into 9347 * the list. We must remove it by forcing a write to the 9348 * link before us, whether it be the superblock or an inode. 9349 * Unfortunately the list may change while we're waiting 9350 * on the buf lock for either resource so we must loop until 9351 * we lock the right one. If both the superblock and an 9352 * inode point to this inode we must clear the inode first 9353 * followed by the superblock. 9354 */ 9355 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9356 pino = 0; 9357 if (idp && (idp->id_state & UNLINKNEXT)) 9358 pino = idp->id_ino; 9359 FREE_LOCK(&lk); 9360 if (pino == 0) 9361 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9362 (int)fs->fs_sbsize, 0, 0, 0); 9363 else 9364 error = bread(ump->um_devvp, 9365 fsbtodb(fs, ino_to_fsba(fs, pino)), 9366 (int)fs->fs_bsize, NOCRED, &bp); 9367 ACQUIRE_LOCK(&lk); 9368 if (error) 9369 break; 9370 /* If the list has changed restart the loop. */ 9371 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9372 nino = 0; 9373 if (idp && (idp->id_state & UNLINKNEXT)) 9374 nino = idp->id_ino; 9375 if (nino != pino || 9376 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9377 FREE_LOCK(&lk); 9378 brelse(bp); 9379 ACQUIRE_LOCK(&lk); 9380 continue; 9381 } 9382 nino = 0; 9383 idn = TAILQ_NEXT(inodedep, id_unlinked); 9384 if (idn) 9385 nino = idn->id_ino; 9386 /* 9387 * Remove us from the in memory list. After this we cannot 9388 * access the inodedep. 9389 */ 9390 KASSERT((inodedep->id_state & UNLINKED) != 0, 9391 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9392 inodedep)); 9393 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9394 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9395 FREE_LOCK(&lk); 9396 /* 9397 * The predecessor's next pointer is manually updated here 9398 * so that the NEXT flag is never cleared for an element 9399 * that is in the list. 9400 */ 9401 if (pino == 0) { 9402 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9403 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9404 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9405 bp); 9406 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9407 ((struct ufs1_dinode *)bp->b_data + 9408 ino_to_fsbo(fs, pino))->di_freelink = nino; 9409 else 9410 ((struct ufs2_dinode *)bp->b_data + 9411 ino_to_fsbo(fs, pino))->di_freelink = nino; 9412 /* 9413 * If the bwrite fails we have no recourse to recover. The 9414 * filesystem is corrupted already. 9415 */ 9416 bwrite(bp); 9417 ACQUIRE_LOCK(&lk); 9418 /* 9419 * If the superblock pointer still needs to be cleared force 9420 * a write here. 9421 */ 9422 if (fs->fs_sujfree == ino) { 9423 FREE_LOCK(&lk); 9424 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9425 (int)fs->fs_sbsize, 0, 0, 0); 9426 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9427 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9428 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9429 bp); 9430 bwrite(bp); 9431 ACQUIRE_LOCK(&lk); 9432 } 9433 9434 if (fs->fs_sujfree != ino) 9435 return; 9436 panic("clear_unlinked_inodedep: Failed to clear free head"); 9437 } 9438 if (inodedep->id_ino == fs->fs_sujfree) 9439 panic("clear_unlinked_inodedep: Freeing head of free list"); 9440 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9441 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9442 return; 9443} 9444 9445/* 9446 * This workitem decrements the inode's link count. 9447 * If the link count reaches zero, the file is removed. 9448 */ 9449static int 9450handle_workitem_remove(dirrem, flags) 9451 struct dirrem *dirrem; 9452 int flags; 9453{ 9454 struct inodedep *inodedep; 9455 struct workhead dotdotwk; 9456 struct worklist *wk; 9457 struct ufsmount *ump; 9458 struct mount *mp; 9459 struct vnode *vp; 9460 struct inode *ip; 9461 ino_t oldinum; 9462 9463 if (dirrem->dm_state & ONWORKLIST) 9464 panic("handle_workitem_remove: dirrem %p still on worklist", 9465 dirrem); 9466 oldinum = dirrem->dm_oldinum; 9467 mp = dirrem->dm_list.wk_mp; 9468 ump = VFSTOUFS(mp); 9469 flags |= LK_EXCLUSIVE; 9470 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9471 return (EBUSY); 9472 ip = VTOI(vp); 9473 ACQUIRE_LOCK(&lk); 9474 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9475 panic("handle_workitem_remove: lost inodedep"); 9476 if (dirrem->dm_state & ONDEPLIST) 9477 LIST_REMOVE(dirrem, dm_inonext); 9478 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9479 ("handle_workitem_remove: Journal entries not written.")); 9480 9481 /* 9482 * Move all dependencies waiting on the remove to complete 9483 * from the dirrem to the inode inowait list to be completed 9484 * after the inode has been updated and written to disk. Any 9485 * marked MKDIR_PARENT are saved to be completed when the .. ref 9486 * is removed. 9487 */ 9488 LIST_INIT(&dotdotwk); 9489 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9490 WORKLIST_REMOVE(wk); 9491 if (wk->wk_state & MKDIR_PARENT) { 9492 wk->wk_state &= ~MKDIR_PARENT; 9493 WORKLIST_INSERT(&dotdotwk, wk); 9494 continue; 9495 } 9496 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9497 } 9498 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9499 /* 9500 * Normal file deletion. 9501 */ 9502 if ((dirrem->dm_state & RMDIR) == 0) { 9503 ip->i_nlink--; 9504 DIP_SET(ip, i_nlink, ip->i_nlink); 9505 ip->i_flag |= IN_CHANGE; 9506 if (ip->i_nlink < ip->i_effnlink) 9507 panic("handle_workitem_remove: bad file delta"); 9508 if (ip->i_nlink == 0) 9509 unlinked_inodedep(mp, inodedep); 9510 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9511 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9512 ("handle_workitem_remove: worklist not empty. %s", 9513 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9514 WORKITEM_FREE(dirrem, D_DIRREM); 9515 FREE_LOCK(&lk); 9516 goto out; 9517 } 9518 /* 9519 * Directory deletion. Decrement reference count for both the 9520 * just deleted parent directory entry and the reference for ".". 9521 * Arrange to have the reference count on the parent decremented 9522 * to account for the loss of "..". 9523 */ 9524 ip->i_nlink -= 2; 9525 DIP_SET(ip, i_nlink, ip->i_nlink); 9526 ip->i_flag |= IN_CHANGE; 9527 if (ip->i_nlink < ip->i_effnlink) 9528 panic("handle_workitem_remove: bad dir delta"); 9529 if (ip->i_nlink == 0) 9530 unlinked_inodedep(mp, inodedep); 9531 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9532 /* 9533 * Rename a directory to a new parent. Since, we are both deleting 9534 * and creating a new directory entry, the link count on the new 9535 * directory should not change. Thus we skip the followup dirrem. 9536 */ 9537 if (dirrem->dm_state & DIRCHG) { 9538 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9539 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9540 WORKITEM_FREE(dirrem, D_DIRREM); 9541 FREE_LOCK(&lk); 9542 goto out; 9543 } 9544 dirrem->dm_state = ONDEPLIST; 9545 dirrem->dm_oldinum = dirrem->dm_dirinum; 9546 /* 9547 * Place the dirrem on the parent's diremhd list. 9548 */ 9549 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9550 panic("handle_workitem_remove: lost dir inodedep"); 9551 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9552 /* 9553 * If the allocated inode has never been written to disk, then 9554 * the on-disk inode is zero'ed and we can remove the file 9555 * immediately. When journaling if the inode has been marked 9556 * unlinked and not DEPCOMPLETE we know it can never be written. 9557 */ 9558 inodedep_lookup(mp, oldinum, 0, &inodedep); 9559 if (inodedep == NULL || 9560 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9561 check_inode_unwritten(inodedep)) { 9562 FREE_LOCK(&lk); 9563 vput(vp); 9564 return handle_workitem_remove(dirrem, flags); 9565 } 9566 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9567 FREE_LOCK(&lk); 9568 ip->i_flag |= IN_CHANGE; 9569out: 9570 ffs_update(vp, 0); 9571 vput(vp); 9572 return (0); 9573} 9574 9575/* 9576 * Inode de-allocation dependencies. 9577 * 9578 * When an inode's link count is reduced to zero, it can be de-allocated. We 9579 * found it convenient to postpone de-allocation until after the inode is 9580 * written to disk with its new link count (zero). At this point, all of the 9581 * on-disk inode's block pointers are nullified and, with careful dependency 9582 * list ordering, all dependencies related to the inode will be satisfied and 9583 * the corresponding dependency structures de-allocated. So, if/when the 9584 * inode is reused, there will be no mixing of old dependencies with new 9585 * ones. This artificial dependency is set up by the block de-allocation 9586 * procedure above (softdep_setup_freeblocks) and completed by the 9587 * following procedure. 9588 */ 9589static void 9590handle_workitem_freefile(freefile) 9591 struct freefile *freefile; 9592{ 9593 struct workhead wkhd; 9594 struct fs *fs; 9595 struct inodedep *idp; 9596 struct ufsmount *ump; 9597 int error; 9598 9599 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9600 fs = ump->um_fs; 9601#ifdef DEBUG 9602 ACQUIRE_LOCK(&lk); 9603 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9604 FREE_LOCK(&lk); 9605 if (error) 9606 panic("handle_workitem_freefile: inodedep %p survived", idp); 9607#endif 9608 UFS_LOCK(ump); 9609 fs->fs_pendinginodes -= 1; 9610 UFS_UNLOCK(ump); 9611 LIST_INIT(&wkhd); 9612 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9613 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9614 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9615 softdep_error("handle_workitem_freefile", error); 9616 ACQUIRE_LOCK(&lk); 9617 WORKITEM_FREE(freefile, D_FREEFILE); 9618 FREE_LOCK(&lk); 9619} 9620 9621 9622/* 9623 * Helper function which unlinks marker element from work list and returns 9624 * the next element on the list. 9625 */ 9626static __inline struct worklist * 9627markernext(struct worklist *marker) 9628{ 9629 struct worklist *next; 9630 9631 next = LIST_NEXT(marker, wk_list); 9632 LIST_REMOVE(marker, wk_list); 9633 return next; 9634} 9635 9636/* 9637 * Disk writes. 9638 * 9639 * The dependency structures constructed above are most actively used when file 9640 * system blocks are written to disk. No constraints are placed on when a 9641 * block can be written, but unsatisfied update dependencies are made safe by 9642 * modifying (or replacing) the source memory for the duration of the disk 9643 * write. When the disk write completes, the memory block is again brought 9644 * up-to-date. 9645 * 9646 * In-core inode structure reclamation. 9647 * 9648 * Because there are a finite number of "in-core" inode structures, they are 9649 * reused regularly. By transferring all inode-related dependencies to the 9650 * in-memory inode block and indexing them separately (via "inodedep"s), we 9651 * can allow "in-core" inode structures to be reused at any time and avoid 9652 * any increase in contention. 9653 * 9654 * Called just before entering the device driver to initiate a new disk I/O. 9655 * The buffer must be locked, thus, no I/O completion operations can occur 9656 * while we are manipulating its associated dependencies. 9657 */ 9658static void 9659softdep_disk_io_initiation(bp) 9660 struct buf *bp; /* structure describing disk write to occur */ 9661{ 9662 struct worklist *wk; 9663 struct worklist marker; 9664 struct inodedep *inodedep; 9665 struct freeblks *freeblks; 9666 struct jblkdep *jblkdep; 9667 struct newblk *newblk; 9668 9669 /* 9670 * We only care about write operations. There should never 9671 * be dependencies for reads. 9672 */ 9673 if (bp->b_iocmd != BIO_WRITE) 9674 panic("softdep_disk_io_initiation: not write"); 9675 9676 if (bp->b_vflags & BV_BKGRDINPROG) 9677 panic("softdep_disk_io_initiation: Writing buffer with " 9678 "background write in progress: %p", bp); 9679 9680 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9681 PHOLD(curproc); /* Don't swap out kernel stack */ 9682 9683 ACQUIRE_LOCK(&lk); 9684 /* 9685 * Do any necessary pre-I/O processing. 9686 */ 9687 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9688 wk = markernext(&marker)) { 9689 LIST_INSERT_AFTER(wk, &marker, wk_list); 9690 switch (wk->wk_type) { 9691 9692 case D_PAGEDEP: 9693 initiate_write_filepage(WK_PAGEDEP(wk), bp); 9694 continue; 9695 9696 case D_INODEDEP: 9697 inodedep = WK_INODEDEP(wk); 9698 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 9699 initiate_write_inodeblock_ufs1(inodedep, bp); 9700 else 9701 initiate_write_inodeblock_ufs2(inodedep, bp); 9702 continue; 9703 9704 case D_INDIRDEP: 9705 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 9706 continue; 9707 9708 case D_BMSAFEMAP: 9709 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 9710 continue; 9711 9712 case D_JSEG: 9713 WK_JSEG(wk)->js_buf = NULL; 9714 continue; 9715 9716 case D_FREEBLKS: 9717 freeblks = WK_FREEBLKS(wk); 9718 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 9719 /* 9720 * We have to wait for the freeblks to be journaled 9721 * before we can write an inodeblock with updated 9722 * pointers. Be careful to arrange the marker so 9723 * we revisit the freeblks if it's not removed by 9724 * the first jwait(). 9725 */ 9726 if (jblkdep != NULL) { 9727 LIST_REMOVE(&marker, wk_list); 9728 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9729 jwait(&jblkdep->jb_list, MNT_WAIT); 9730 } 9731 continue; 9732 case D_ALLOCDIRECT: 9733 case D_ALLOCINDIR: 9734 /* 9735 * We have to wait for the jnewblk to be journaled 9736 * before we can write to a block if the contents 9737 * may be confused with an earlier file's indirect 9738 * at recovery time. Handle the marker as described 9739 * above. 9740 */ 9741 newblk = WK_NEWBLK(wk); 9742 if (newblk->nb_jnewblk != NULL && 9743 indirblk_lookup(newblk->nb_list.wk_mp, 9744 newblk->nb_newblkno)) { 9745 LIST_REMOVE(&marker, wk_list); 9746 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9747 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 9748 } 9749 continue; 9750 9751 case D_SBDEP: 9752 initiate_write_sbdep(WK_SBDEP(wk)); 9753 continue; 9754 9755 case D_MKDIR: 9756 case D_FREEWORK: 9757 case D_FREEDEP: 9758 case D_JSEGDEP: 9759 continue; 9760 9761 default: 9762 panic("handle_disk_io_initiation: Unexpected type %s", 9763 TYPENAME(wk->wk_type)); 9764 /* NOTREACHED */ 9765 } 9766 } 9767 FREE_LOCK(&lk); 9768 PRELE(curproc); /* Allow swapout of kernel stack */ 9769} 9770 9771/* 9772 * Called from within the procedure above to deal with unsatisfied 9773 * allocation dependencies in a directory. The buffer must be locked, 9774 * thus, no I/O completion operations can occur while we are 9775 * manipulating its associated dependencies. 9776 */ 9777static void 9778initiate_write_filepage(pagedep, bp) 9779 struct pagedep *pagedep; 9780 struct buf *bp; 9781{ 9782 struct jremref *jremref; 9783 struct jmvref *jmvref; 9784 struct dirrem *dirrem; 9785 struct diradd *dap; 9786 struct direct *ep; 9787 int i; 9788 9789 if (pagedep->pd_state & IOSTARTED) { 9790 /* 9791 * This can only happen if there is a driver that does not 9792 * understand chaining. Here biodone will reissue the call 9793 * to strategy for the incomplete buffers. 9794 */ 9795 printf("initiate_write_filepage: already started\n"); 9796 return; 9797 } 9798 pagedep->pd_state |= IOSTARTED; 9799 /* 9800 * Wait for all journal remove dependencies to hit the disk. 9801 * We can not allow any potentially conflicting directory adds 9802 * to be visible before removes and rollback is too difficult. 9803 * lk may be dropped and re-acquired, however we hold the buf 9804 * locked so the dependency can not go away. 9805 */ 9806 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 9807 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 9808 jwait(&jremref->jr_list, MNT_WAIT); 9809 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 9810 jwait(&jmvref->jm_list, MNT_WAIT); 9811 for (i = 0; i < DAHASHSZ; i++) { 9812 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 9813 ep = (struct direct *) 9814 ((char *)bp->b_data + dap->da_offset); 9815 if (ep->d_ino != dap->da_newinum) 9816 panic("%s: dir inum %ju != new %ju", 9817 "initiate_write_filepage", 9818 (uintmax_t)ep->d_ino, 9819 (uintmax_t)dap->da_newinum); 9820 if (dap->da_state & DIRCHG) 9821 ep->d_ino = dap->da_previous->dm_oldinum; 9822 else 9823 ep->d_ino = 0; 9824 dap->da_state &= ~ATTACHED; 9825 dap->da_state |= UNDONE; 9826 } 9827 } 9828} 9829 9830/* 9831 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 9832 * Note that any bug fixes made to this routine must be done in the 9833 * version found below. 9834 * 9835 * Called from within the procedure above to deal with unsatisfied 9836 * allocation dependencies in an inodeblock. The buffer must be 9837 * locked, thus, no I/O completion operations can occur while we 9838 * are manipulating its associated dependencies. 9839 */ 9840static void 9841initiate_write_inodeblock_ufs1(inodedep, bp) 9842 struct inodedep *inodedep; 9843 struct buf *bp; /* The inode block */ 9844{ 9845 struct allocdirect *adp, *lastadp; 9846 struct ufs1_dinode *dp; 9847 struct ufs1_dinode *sip; 9848 struct inoref *inoref; 9849 struct fs *fs; 9850 ufs_lbn_t i; 9851#ifdef INVARIANTS 9852 ufs_lbn_t prevlbn = 0; 9853#endif 9854 int deplist; 9855 9856 if (inodedep->id_state & IOSTARTED) 9857 panic("initiate_write_inodeblock_ufs1: already started"); 9858 inodedep->id_state |= IOSTARTED; 9859 fs = inodedep->id_fs; 9860 dp = (struct ufs1_dinode *)bp->b_data + 9861 ino_to_fsbo(fs, inodedep->id_ino); 9862 9863 /* 9864 * If we're on the unlinked list but have not yet written our 9865 * next pointer initialize it here. 9866 */ 9867 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 9868 struct inodedep *inon; 9869 9870 inon = TAILQ_NEXT(inodedep, id_unlinked); 9871 dp->di_freelink = inon ? inon->id_ino : 0; 9872 } 9873 /* 9874 * If the bitmap is not yet written, then the allocated 9875 * inode cannot be written to disk. 9876 */ 9877 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 9878 if (inodedep->id_savedino1 != NULL) 9879 panic("initiate_write_inodeblock_ufs1: I/O underway"); 9880 FREE_LOCK(&lk); 9881 sip = malloc(sizeof(struct ufs1_dinode), 9882 M_SAVEDINO, M_SOFTDEP_FLAGS); 9883 ACQUIRE_LOCK(&lk); 9884 inodedep->id_savedino1 = sip; 9885 *inodedep->id_savedino1 = *dp; 9886 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 9887 dp->di_gen = inodedep->id_savedino1->di_gen; 9888 dp->di_freelink = inodedep->id_savedino1->di_freelink; 9889 return; 9890 } 9891 /* 9892 * If no dependencies, then there is nothing to roll back. 9893 */ 9894 inodedep->id_savedsize = dp->di_size; 9895 inodedep->id_savedextsize = 0; 9896 inodedep->id_savednlink = dp->di_nlink; 9897 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 9898 TAILQ_EMPTY(&inodedep->id_inoreflst)) 9899 return; 9900 /* 9901 * Revert the link count to that of the first unwritten journal entry. 9902 */ 9903 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 9904 if (inoref) 9905 dp->di_nlink = inoref->if_nlink; 9906 /* 9907 * Set the dependencies to busy. 9908 */ 9909 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9910 adp = TAILQ_NEXT(adp, ad_next)) { 9911#ifdef INVARIANTS 9912 if (deplist != 0 && prevlbn >= adp->ad_offset) 9913 panic("softdep_write_inodeblock: lbn order"); 9914 prevlbn = adp->ad_offset; 9915 if (adp->ad_offset < NDADDR && 9916 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 9917 panic("%s: direct pointer #%jd mismatch %d != %jd", 9918 "softdep_write_inodeblock", 9919 (intmax_t)adp->ad_offset, 9920 dp->di_db[adp->ad_offset], 9921 (intmax_t)adp->ad_newblkno); 9922 if (adp->ad_offset >= NDADDR && 9923 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 9924 panic("%s: indirect pointer #%jd mismatch %d != %jd", 9925 "softdep_write_inodeblock", 9926 (intmax_t)adp->ad_offset - NDADDR, 9927 dp->di_ib[adp->ad_offset - NDADDR], 9928 (intmax_t)adp->ad_newblkno); 9929 deplist |= 1 << adp->ad_offset; 9930 if ((adp->ad_state & ATTACHED) == 0) 9931 panic("softdep_write_inodeblock: Unknown state 0x%x", 9932 adp->ad_state); 9933#endif /* INVARIANTS */ 9934 adp->ad_state &= ~ATTACHED; 9935 adp->ad_state |= UNDONE; 9936 } 9937 /* 9938 * The on-disk inode cannot claim to be any larger than the last 9939 * fragment that has been written. Otherwise, the on-disk inode 9940 * might have fragments that were not the last block in the file 9941 * which would corrupt the filesystem. 9942 */ 9943 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 9944 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 9945 if (adp->ad_offset >= NDADDR) 9946 break; 9947 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 9948 /* keep going until hitting a rollback to a frag */ 9949 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 9950 continue; 9951 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 9952 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 9953#ifdef INVARIANTS 9954 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 9955 panic("softdep_write_inodeblock: lost dep1"); 9956#endif /* INVARIANTS */ 9957 dp->di_db[i] = 0; 9958 } 9959 for (i = 0; i < NIADDR; i++) { 9960#ifdef INVARIANTS 9961 if (dp->di_ib[i] != 0 && 9962 (deplist & ((1 << NDADDR) << i)) == 0) 9963 panic("softdep_write_inodeblock: lost dep2"); 9964#endif /* INVARIANTS */ 9965 dp->di_ib[i] = 0; 9966 } 9967 return; 9968 } 9969 /* 9970 * If we have zero'ed out the last allocated block of the file, 9971 * roll back the size to the last currently allocated block. 9972 * We know that this last allocated block is a full-sized as 9973 * we already checked for fragments in the loop above. 9974 */ 9975 if (lastadp != NULL && 9976 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 9977 for (i = lastadp->ad_offset; i >= 0; i--) 9978 if (dp->di_db[i] != 0) 9979 break; 9980 dp->di_size = (i + 1) * fs->fs_bsize; 9981 } 9982 /* 9983 * The only dependencies are for indirect blocks. 9984 * 9985 * The file size for indirect block additions is not guaranteed. 9986 * Such a guarantee would be non-trivial to achieve. The conventional 9987 * synchronous write implementation also does not make this guarantee. 9988 * Fsck should catch and fix discrepancies. Arguably, the file size 9989 * can be over-estimated without destroying integrity when the file 9990 * moves into the indirect blocks (i.e., is large). If we want to 9991 * postpone fsck, we are stuck with this argument. 9992 */ 9993 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 9994 dp->di_ib[adp->ad_offset - NDADDR] = 0; 9995} 9996 9997/* 9998 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 9999 * Note that any bug fixes made to this routine must be done in the 10000 * version found above. 10001 * 10002 * Called from within the procedure above to deal with unsatisfied 10003 * allocation dependencies in an inodeblock. The buffer must be 10004 * locked, thus, no I/O completion operations can occur while we 10005 * are manipulating its associated dependencies. 10006 */ 10007static void 10008initiate_write_inodeblock_ufs2(inodedep, bp) 10009 struct inodedep *inodedep; 10010 struct buf *bp; /* The inode block */ 10011{ 10012 struct allocdirect *adp, *lastadp; 10013 struct ufs2_dinode *dp; 10014 struct ufs2_dinode *sip; 10015 struct inoref *inoref; 10016 struct fs *fs; 10017 ufs_lbn_t i; 10018#ifdef INVARIANTS 10019 ufs_lbn_t prevlbn = 0; 10020#endif 10021 int deplist; 10022 10023 if (inodedep->id_state & IOSTARTED) 10024 panic("initiate_write_inodeblock_ufs2: already started"); 10025 inodedep->id_state |= IOSTARTED; 10026 fs = inodedep->id_fs; 10027 dp = (struct ufs2_dinode *)bp->b_data + 10028 ino_to_fsbo(fs, inodedep->id_ino); 10029 10030 /* 10031 * If we're on the unlinked list but have not yet written our 10032 * next pointer initialize it here. 10033 */ 10034 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10035 struct inodedep *inon; 10036 10037 inon = TAILQ_NEXT(inodedep, id_unlinked); 10038 dp->di_freelink = inon ? inon->id_ino : 0; 10039 } 10040 /* 10041 * If the bitmap is not yet written, then the allocated 10042 * inode cannot be written to disk. 10043 */ 10044 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10045 if (inodedep->id_savedino2 != NULL) 10046 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10047 FREE_LOCK(&lk); 10048 sip = malloc(sizeof(struct ufs2_dinode), 10049 M_SAVEDINO, M_SOFTDEP_FLAGS); 10050 ACQUIRE_LOCK(&lk); 10051 inodedep->id_savedino2 = sip; 10052 *inodedep->id_savedino2 = *dp; 10053 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10054 dp->di_gen = inodedep->id_savedino2->di_gen; 10055 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10056 return; 10057 } 10058 /* 10059 * If no dependencies, then there is nothing to roll back. 10060 */ 10061 inodedep->id_savedsize = dp->di_size; 10062 inodedep->id_savedextsize = dp->di_extsize; 10063 inodedep->id_savednlink = dp->di_nlink; 10064 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10065 TAILQ_EMPTY(&inodedep->id_extupdt) && 10066 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10067 return; 10068 /* 10069 * Revert the link count to that of the first unwritten journal entry. 10070 */ 10071 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10072 if (inoref) 10073 dp->di_nlink = inoref->if_nlink; 10074 10075 /* 10076 * Set the ext data dependencies to busy. 10077 */ 10078 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10079 adp = TAILQ_NEXT(adp, ad_next)) { 10080#ifdef INVARIANTS 10081 if (deplist != 0 && prevlbn >= adp->ad_offset) 10082 panic("softdep_write_inodeblock: lbn order"); 10083 prevlbn = adp->ad_offset; 10084 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10085 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10086 "softdep_write_inodeblock", 10087 (intmax_t)adp->ad_offset, 10088 (intmax_t)dp->di_extb[adp->ad_offset], 10089 (intmax_t)adp->ad_newblkno); 10090 deplist |= 1 << adp->ad_offset; 10091 if ((adp->ad_state & ATTACHED) == 0) 10092 panic("softdep_write_inodeblock: Unknown state 0x%x", 10093 adp->ad_state); 10094#endif /* INVARIANTS */ 10095 adp->ad_state &= ~ATTACHED; 10096 adp->ad_state |= UNDONE; 10097 } 10098 /* 10099 * The on-disk inode cannot claim to be any larger than the last 10100 * fragment that has been written. Otherwise, the on-disk inode 10101 * might have fragments that were not the last block in the ext 10102 * data which would corrupt the filesystem. 10103 */ 10104 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10105 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10106 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10107 /* keep going until hitting a rollback to a frag */ 10108 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10109 continue; 10110 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10111 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 10112#ifdef INVARIANTS 10113 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10114 panic("softdep_write_inodeblock: lost dep1"); 10115#endif /* INVARIANTS */ 10116 dp->di_extb[i] = 0; 10117 } 10118 lastadp = NULL; 10119 break; 10120 } 10121 /* 10122 * If we have zero'ed out the last allocated block of the ext 10123 * data, roll back the size to the last currently allocated block. 10124 * We know that this last allocated block is a full-sized as 10125 * we already checked for fragments in the loop above. 10126 */ 10127 if (lastadp != NULL && 10128 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10129 for (i = lastadp->ad_offset; i >= 0; i--) 10130 if (dp->di_extb[i] != 0) 10131 break; 10132 dp->di_extsize = (i + 1) * fs->fs_bsize; 10133 } 10134 /* 10135 * Set the file data dependencies to busy. 10136 */ 10137 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10138 adp = TAILQ_NEXT(adp, ad_next)) { 10139#ifdef INVARIANTS 10140 if (deplist != 0 && prevlbn >= adp->ad_offset) 10141 panic("softdep_write_inodeblock: lbn order"); 10142 if ((adp->ad_state & ATTACHED) == 0) 10143 panic("inodedep %p and adp %p not attached", inodedep, adp); 10144 prevlbn = adp->ad_offset; 10145 if (adp->ad_offset < NDADDR && 10146 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10147 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10148 "softdep_write_inodeblock", 10149 (intmax_t)adp->ad_offset, 10150 (intmax_t)dp->di_db[adp->ad_offset], 10151 (intmax_t)adp->ad_newblkno); 10152 if (adp->ad_offset >= NDADDR && 10153 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10154 panic("%s indirect pointer #%jd mismatch %jd != %jd", 10155 "softdep_write_inodeblock:", 10156 (intmax_t)adp->ad_offset - NDADDR, 10157 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 10158 (intmax_t)adp->ad_newblkno); 10159 deplist |= 1 << adp->ad_offset; 10160 if ((adp->ad_state & ATTACHED) == 0) 10161 panic("softdep_write_inodeblock: Unknown state 0x%x", 10162 adp->ad_state); 10163#endif /* INVARIANTS */ 10164 adp->ad_state &= ~ATTACHED; 10165 adp->ad_state |= UNDONE; 10166 } 10167 /* 10168 * The on-disk inode cannot claim to be any larger than the last 10169 * fragment that has been written. Otherwise, the on-disk inode 10170 * might have fragments that were not the last block in the file 10171 * which would corrupt the filesystem. 10172 */ 10173 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10174 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10175 if (adp->ad_offset >= NDADDR) 10176 break; 10177 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10178 /* keep going until hitting a rollback to a frag */ 10179 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10180 continue; 10181 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10182 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10183#ifdef INVARIANTS 10184 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10185 panic("softdep_write_inodeblock: lost dep2"); 10186#endif /* INVARIANTS */ 10187 dp->di_db[i] = 0; 10188 } 10189 for (i = 0; i < NIADDR; i++) { 10190#ifdef INVARIANTS 10191 if (dp->di_ib[i] != 0 && 10192 (deplist & ((1 << NDADDR) << i)) == 0) 10193 panic("softdep_write_inodeblock: lost dep3"); 10194#endif /* INVARIANTS */ 10195 dp->di_ib[i] = 0; 10196 } 10197 return; 10198 } 10199 /* 10200 * If we have zero'ed out the last allocated block of the file, 10201 * roll back the size to the last currently allocated block. 10202 * We know that this last allocated block is a full-sized as 10203 * we already checked for fragments in the loop above. 10204 */ 10205 if (lastadp != NULL && 10206 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10207 for (i = lastadp->ad_offset; i >= 0; i--) 10208 if (dp->di_db[i] != 0) 10209 break; 10210 dp->di_size = (i + 1) * fs->fs_bsize; 10211 } 10212 /* 10213 * The only dependencies are for indirect blocks. 10214 * 10215 * The file size for indirect block additions is not guaranteed. 10216 * Such a guarantee would be non-trivial to achieve. The conventional 10217 * synchronous write implementation also does not make this guarantee. 10218 * Fsck should catch and fix discrepancies. Arguably, the file size 10219 * can be over-estimated without destroying integrity when the file 10220 * moves into the indirect blocks (i.e., is large). If we want to 10221 * postpone fsck, we are stuck with this argument. 10222 */ 10223 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10224 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10225} 10226 10227/* 10228 * Cancel an indirdep as a result of truncation. Release all of the 10229 * children allocindirs and place their journal work on the appropriate 10230 * list. 10231 */ 10232static void 10233cancel_indirdep(indirdep, bp, freeblks) 10234 struct indirdep *indirdep; 10235 struct buf *bp; 10236 struct freeblks *freeblks; 10237{ 10238 struct allocindir *aip; 10239 10240 /* 10241 * None of the indirect pointers will ever be visible, 10242 * so they can simply be tossed. GOINGAWAY ensures 10243 * that allocated pointers will be saved in the buffer 10244 * cache until they are freed. Note that they will 10245 * only be able to be found by their physical address 10246 * since the inode mapping the logical address will 10247 * be gone. The save buffer used for the safe copy 10248 * was allocated in setup_allocindir_phase2 using 10249 * the physical address so it could be used for this 10250 * purpose. Hence we swap the safe copy with the real 10251 * copy, allowing the safe copy to be freed and holding 10252 * on to the real copy for later use in indir_trunc. 10253 */ 10254 if (indirdep->ir_state & GOINGAWAY) 10255 panic("cancel_indirdep: already gone"); 10256 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10257 indirdep->ir_state |= DEPCOMPLETE; 10258 LIST_REMOVE(indirdep, ir_next); 10259 } 10260 indirdep->ir_state |= GOINGAWAY; 10261 VFSTOUFS(indirdep->ir_list.wk_mp)->um_numindirdeps += 1; 10262 /* 10263 * Pass in bp for blocks still have journal writes 10264 * pending so we can cancel them on their own. 10265 */ 10266 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 10267 cancel_allocindir(aip, bp, freeblks, 0); 10268 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) 10269 cancel_allocindir(aip, NULL, freeblks, 0); 10270 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) 10271 cancel_allocindir(aip, NULL, freeblks, 0); 10272 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0) 10273 cancel_allocindir(aip, NULL, freeblks, 0); 10274 /* 10275 * If there are pending partial truncations we need to keep the 10276 * old block copy around until they complete. This is because 10277 * the current b_data is not a perfect superset of the available 10278 * blocks. 10279 */ 10280 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10281 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10282 else 10283 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10284 WORKLIST_REMOVE(&indirdep->ir_list); 10285 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10286 indirdep->ir_bp = NULL; 10287 indirdep->ir_freeblks = freeblks; 10288} 10289 10290/* 10291 * Free an indirdep once it no longer has new pointers to track. 10292 */ 10293static void 10294free_indirdep(indirdep) 10295 struct indirdep *indirdep; 10296{ 10297 10298 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10299 ("free_indirdep: Indir trunc list not empty.")); 10300 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10301 ("free_indirdep: Complete head not empty.")); 10302 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10303 ("free_indirdep: write head not empty.")); 10304 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10305 ("free_indirdep: done head not empty.")); 10306 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10307 ("free_indirdep: deplist head not empty.")); 10308 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10309 ("free_indirdep: %p still on newblk list.", indirdep)); 10310 KASSERT(indirdep->ir_saveddata == NULL, 10311 ("free_indirdep: %p still has saved data.", indirdep)); 10312 if (indirdep->ir_state & ONWORKLIST) 10313 WORKLIST_REMOVE(&indirdep->ir_list); 10314 WORKITEM_FREE(indirdep, D_INDIRDEP); 10315} 10316 10317/* 10318 * Called before a write to an indirdep. This routine is responsible for 10319 * rolling back pointers to a safe state which includes only those 10320 * allocindirs which have been completed. 10321 */ 10322static void 10323initiate_write_indirdep(indirdep, bp) 10324 struct indirdep *indirdep; 10325 struct buf *bp; 10326{ 10327 10328 indirdep->ir_state |= IOSTARTED; 10329 if (indirdep->ir_state & GOINGAWAY) 10330 panic("disk_io_initiation: indirdep gone"); 10331 /* 10332 * If there are no remaining dependencies, this will be writing 10333 * the real pointers. 10334 */ 10335 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10336 TAILQ_EMPTY(&indirdep->ir_trunc)) 10337 return; 10338 /* 10339 * Replace up-to-date version with safe version. 10340 */ 10341 if (indirdep->ir_saveddata == NULL) { 10342 FREE_LOCK(&lk); 10343 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10344 M_SOFTDEP_FLAGS); 10345 ACQUIRE_LOCK(&lk); 10346 } 10347 indirdep->ir_state &= ~ATTACHED; 10348 indirdep->ir_state |= UNDONE; 10349 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10350 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10351 bp->b_bcount); 10352} 10353 10354/* 10355 * Called when an inode has been cleared in a cg bitmap. This finally 10356 * eliminates any canceled jaddrefs 10357 */ 10358void 10359softdep_setup_inofree(mp, bp, ino, wkhd) 10360 struct mount *mp; 10361 struct buf *bp; 10362 ino_t ino; 10363 struct workhead *wkhd; 10364{ 10365 struct worklist *wk, *wkn; 10366 struct inodedep *inodedep; 10367 uint8_t *inosused; 10368 struct cg *cgp; 10369 struct fs *fs; 10370 10371 ACQUIRE_LOCK(&lk); 10372 fs = VFSTOUFS(mp)->um_fs; 10373 cgp = (struct cg *)bp->b_data; 10374 inosused = cg_inosused(cgp); 10375 if (isset(inosused, ino % fs->fs_ipg)) 10376 panic("softdep_setup_inofree: inode %ju not freed.", 10377 (uintmax_t)ino); 10378 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10379 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10380 (uintmax_t)ino, inodedep); 10381 if (wkhd) { 10382 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10383 if (wk->wk_type != D_JADDREF) 10384 continue; 10385 WORKLIST_REMOVE(wk); 10386 /* 10387 * We can free immediately even if the jaddref 10388 * isn't attached in a background write as now 10389 * the bitmaps are reconciled. 10390 */ 10391 wk->wk_state |= COMPLETE | ATTACHED; 10392 free_jaddref(WK_JADDREF(wk)); 10393 } 10394 jwork_move(&bp->b_dep, wkhd); 10395 } 10396 FREE_LOCK(&lk); 10397} 10398 10399 10400/* 10401 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10402 * map. Any dependencies waiting for the write to clear are added to the 10403 * buf's list and any jnewblks that are being canceled are discarded 10404 * immediately. 10405 */ 10406void 10407softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10408 struct mount *mp; 10409 struct buf *bp; 10410 ufs2_daddr_t blkno; 10411 int frags; 10412 struct workhead *wkhd; 10413{ 10414 struct bmsafemap *bmsafemap; 10415 struct jnewblk *jnewblk; 10416 struct worklist *wk; 10417 struct fs *fs; 10418#ifdef SUJ_DEBUG 10419 uint8_t *blksfree; 10420 struct cg *cgp; 10421 ufs2_daddr_t jstart; 10422 ufs2_daddr_t jend; 10423 ufs2_daddr_t end; 10424 long bno; 10425 int i; 10426#endif 10427 10428 CTR3(KTR_SUJ, 10429 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10430 blkno, frags, wkhd); 10431 10432 ACQUIRE_LOCK(&lk); 10433 /* Lookup the bmsafemap so we track when it is dirty. */ 10434 fs = VFSTOUFS(mp)->um_fs; 10435 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10436 /* 10437 * Detach any jnewblks which have been canceled. They must linger 10438 * until the bitmap is cleared again by ffs_blkfree() to prevent 10439 * an unjournaled allocation from hitting the disk. 10440 */ 10441 if (wkhd) { 10442 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10443 CTR2(KTR_SUJ, 10444 "softdep_setup_blkfree: blkno %jd wk type %d", 10445 blkno, wk->wk_type); 10446 WORKLIST_REMOVE(wk); 10447 if (wk->wk_type != D_JNEWBLK) { 10448 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10449 continue; 10450 } 10451 jnewblk = WK_JNEWBLK(wk); 10452 KASSERT(jnewblk->jn_state & GOINGAWAY, 10453 ("softdep_setup_blkfree: jnewblk not canceled.")); 10454#ifdef SUJ_DEBUG 10455 /* 10456 * Assert that this block is free in the bitmap 10457 * before we discard the jnewblk. 10458 */ 10459 cgp = (struct cg *)bp->b_data; 10460 blksfree = cg_blksfree(cgp); 10461 bno = dtogd(fs, jnewblk->jn_blkno); 10462 for (i = jnewblk->jn_oldfrags; 10463 i < jnewblk->jn_frags; i++) { 10464 if (isset(blksfree, bno + i)) 10465 continue; 10466 panic("softdep_setup_blkfree: not free"); 10467 } 10468#endif 10469 /* 10470 * Even if it's not attached we can free immediately 10471 * as the new bitmap is correct. 10472 */ 10473 wk->wk_state |= COMPLETE | ATTACHED; 10474 free_jnewblk(jnewblk); 10475 } 10476 } 10477 10478#ifdef SUJ_DEBUG 10479 /* 10480 * Assert that we are not freeing a block which has an outstanding 10481 * allocation dependency. 10482 */ 10483 fs = VFSTOUFS(mp)->um_fs; 10484 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10485 end = blkno + frags; 10486 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10487 /* 10488 * Don't match against blocks that will be freed when the 10489 * background write is done. 10490 */ 10491 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10492 (COMPLETE | DEPCOMPLETE)) 10493 continue; 10494 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10495 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10496 if ((blkno >= jstart && blkno < jend) || 10497 (end > jstart && end <= jend)) { 10498 printf("state 0x%X %jd - %d %d dep %p\n", 10499 jnewblk->jn_state, jnewblk->jn_blkno, 10500 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10501 jnewblk->jn_dep); 10502 panic("softdep_setup_blkfree: " 10503 "%jd-%jd(%d) overlaps with %jd-%jd", 10504 blkno, end, frags, jstart, jend); 10505 } 10506 } 10507#endif 10508 FREE_LOCK(&lk); 10509} 10510 10511/* 10512 * Revert a block allocation when the journal record that describes it 10513 * is not yet written. 10514 */ 10515int 10516jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10517 struct jnewblk *jnewblk; 10518 struct fs *fs; 10519 struct cg *cgp; 10520 uint8_t *blksfree; 10521{ 10522 ufs1_daddr_t fragno; 10523 long cgbno, bbase; 10524 int frags, blk; 10525 int i; 10526 10527 frags = 0; 10528 cgbno = dtogd(fs, jnewblk->jn_blkno); 10529 /* 10530 * We have to test which frags need to be rolled back. We may 10531 * be operating on a stale copy when doing background writes. 10532 */ 10533 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10534 if (isclr(blksfree, cgbno + i)) 10535 frags++; 10536 if (frags == 0) 10537 return (0); 10538 /* 10539 * This is mostly ffs_blkfree() sans some validation and 10540 * superblock updates. 10541 */ 10542 if (frags == fs->fs_frag) { 10543 fragno = fragstoblks(fs, cgbno); 10544 ffs_setblock(fs, blksfree, fragno); 10545 ffs_clusteracct(fs, cgp, fragno, 1); 10546 cgp->cg_cs.cs_nbfree++; 10547 } else { 10548 cgbno += jnewblk->jn_oldfrags; 10549 bbase = cgbno - fragnum(fs, cgbno); 10550 /* Decrement the old frags. */ 10551 blk = blkmap(fs, blksfree, bbase); 10552 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10553 /* Deallocate the fragment */ 10554 for (i = 0; i < frags; i++) 10555 setbit(blksfree, cgbno + i); 10556 cgp->cg_cs.cs_nffree += frags; 10557 /* Add back in counts associated with the new frags */ 10558 blk = blkmap(fs, blksfree, bbase); 10559 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10560 /* If a complete block has been reassembled, account for it. */ 10561 fragno = fragstoblks(fs, bbase); 10562 if (ffs_isblock(fs, blksfree, fragno)) { 10563 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10564 ffs_clusteracct(fs, cgp, fragno, 1); 10565 cgp->cg_cs.cs_nbfree++; 10566 } 10567 } 10568 stat_jnewblk++; 10569 jnewblk->jn_state &= ~ATTACHED; 10570 jnewblk->jn_state |= UNDONE; 10571 10572 return (frags); 10573} 10574 10575static void 10576initiate_write_bmsafemap(bmsafemap, bp) 10577 struct bmsafemap *bmsafemap; 10578 struct buf *bp; /* The cg block. */ 10579{ 10580 struct jaddref *jaddref; 10581 struct jnewblk *jnewblk; 10582 uint8_t *inosused; 10583 uint8_t *blksfree; 10584 struct cg *cgp; 10585 struct fs *fs; 10586 ino_t ino; 10587 10588 if (bmsafemap->sm_state & IOSTARTED) 10589 return; 10590 bmsafemap->sm_state |= IOSTARTED; 10591 /* 10592 * Clear any inode allocations which are pending journal writes. 10593 */ 10594 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10595 cgp = (struct cg *)bp->b_data; 10596 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10597 inosused = cg_inosused(cgp); 10598 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10599 ino = jaddref->ja_ino % fs->fs_ipg; 10600 if (isset(inosused, ino)) { 10601 if ((jaddref->ja_mode & IFMT) == IFDIR) 10602 cgp->cg_cs.cs_ndir--; 10603 cgp->cg_cs.cs_nifree++; 10604 clrbit(inosused, ino); 10605 jaddref->ja_state &= ~ATTACHED; 10606 jaddref->ja_state |= UNDONE; 10607 stat_jaddref++; 10608 } else 10609 panic("initiate_write_bmsafemap: inode %ju " 10610 "marked free", (uintmax_t)jaddref->ja_ino); 10611 } 10612 } 10613 /* 10614 * Clear any block allocations which are pending journal writes. 10615 */ 10616 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10617 cgp = (struct cg *)bp->b_data; 10618 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10619 blksfree = cg_blksfree(cgp); 10620 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10621 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10622 continue; 10623 panic("initiate_write_bmsafemap: block %jd " 10624 "marked free", jnewblk->jn_blkno); 10625 } 10626 } 10627 /* 10628 * Move allocation lists to the written lists so they can be 10629 * cleared once the block write is complete. 10630 */ 10631 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10632 inodedep, id_deps); 10633 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10634 newblk, nb_deps); 10635 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10636 wk_list); 10637} 10638 10639/* 10640 * This routine is called during the completion interrupt 10641 * service routine for a disk write (from the procedure called 10642 * by the device driver to inform the filesystem caches of 10643 * a request completion). It should be called early in this 10644 * procedure, before the block is made available to other 10645 * processes or other routines are called. 10646 * 10647 */ 10648static void 10649softdep_disk_write_complete(bp) 10650 struct buf *bp; /* describes the completed disk write */ 10651{ 10652 struct worklist *wk; 10653 struct worklist *owk; 10654 struct workhead reattach; 10655 struct freeblks *freeblks; 10656 struct buf *sbp; 10657 10658 /* 10659 * If an error occurred while doing the write, then the data 10660 * has not hit the disk and the dependencies cannot be unrolled. 10661 */ 10662 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 10663 return; 10664 LIST_INIT(&reattach); 10665 /* 10666 * This lock must not be released anywhere in this code segment. 10667 */ 10668 sbp = NULL; 10669 owk = NULL; 10670 ACQUIRE_LOCK(&lk); 10671 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 10672 WORKLIST_REMOVE(wk); 10673 dep_write[wk->wk_type]++; 10674 if (wk == owk) 10675 panic("duplicate worklist: %p\n", wk); 10676 owk = wk; 10677 switch (wk->wk_type) { 10678 10679 case D_PAGEDEP: 10680 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 10681 WORKLIST_INSERT(&reattach, wk); 10682 continue; 10683 10684 case D_INODEDEP: 10685 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 10686 WORKLIST_INSERT(&reattach, wk); 10687 continue; 10688 10689 case D_BMSAFEMAP: 10690 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp)) 10691 WORKLIST_INSERT(&reattach, wk); 10692 continue; 10693 10694 case D_MKDIR: 10695 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 10696 continue; 10697 10698 case D_ALLOCDIRECT: 10699 wk->wk_state |= COMPLETE; 10700 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 10701 continue; 10702 10703 case D_ALLOCINDIR: 10704 wk->wk_state |= COMPLETE; 10705 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 10706 continue; 10707 10708 case D_INDIRDEP: 10709 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp)) 10710 WORKLIST_INSERT(&reattach, wk); 10711 continue; 10712 10713 case D_FREEBLKS: 10714 wk->wk_state |= COMPLETE; 10715 freeblks = WK_FREEBLKS(wk); 10716 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 10717 LIST_EMPTY(&freeblks->fb_jblkdephd)) 10718 add_to_worklist(wk, WK_NODELAY); 10719 continue; 10720 10721 case D_FREEWORK: 10722 handle_written_freework(WK_FREEWORK(wk)); 10723 break; 10724 10725 case D_JSEGDEP: 10726 free_jsegdep(WK_JSEGDEP(wk)); 10727 continue; 10728 10729 case D_JSEG: 10730 handle_written_jseg(WK_JSEG(wk), bp); 10731 continue; 10732 10733 case D_SBDEP: 10734 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 10735 WORKLIST_INSERT(&reattach, wk); 10736 continue; 10737 10738 case D_FREEDEP: 10739 free_freedep(WK_FREEDEP(wk)); 10740 continue; 10741 10742 default: 10743 panic("handle_disk_write_complete: Unknown type %s", 10744 TYPENAME(wk->wk_type)); 10745 /* NOTREACHED */ 10746 } 10747 } 10748 /* 10749 * Reattach any requests that must be redone. 10750 */ 10751 while ((wk = LIST_FIRST(&reattach)) != NULL) { 10752 WORKLIST_REMOVE(wk); 10753 WORKLIST_INSERT(&bp->b_dep, wk); 10754 } 10755 FREE_LOCK(&lk); 10756 if (sbp) 10757 brelse(sbp); 10758} 10759 10760/* 10761 * Called from within softdep_disk_write_complete above. Note that 10762 * this routine is always called from interrupt level with further 10763 * splbio interrupts blocked. 10764 */ 10765static void 10766handle_allocdirect_partdone(adp, wkhd) 10767 struct allocdirect *adp; /* the completed allocdirect */ 10768 struct workhead *wkhd; /* Work to do when inode is writtne. */ 10769{ 10770 struct allocdirectlst *listhead; 10771 struct allocdirect *listadp; 10772 struct inodedep *inodedep; 10773 long bsize; 10774 10775 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 10776 return; 10777 /* 10778 * The on-disk inode cannot claim to be any larger than the last 10779 * fragment that has been written. Otherwise, the on-disk inode 10780 * might have fragments that were not the last block in the file 10781 * which would corrupt the filesystem. Thus, we cannot free any 10782 * allocdirects after one whose ad_oldblkno claims a fragment as 10783 * these blocks must be rolled back to zero before writing the inode. 10784 * We check the currently active set of allocdirects in id_inoupdt 10785 * or id_extupdt as appropriate. 10786 */ 10787 inodedep = adp->ad_inodedep; 10788 bsize = inodedep->id_fs->fs_bsize; 10789 if (adp->ad_state & EXTDATA) 10790 listhead = &inodedep->id_extupdt; 10791 else 10792 listhead = &inodedep->id_inoupdt; 10793 TAILQ_FOREACH(listadp, listhead, ad_next) { 10794 /* found our block */ 10795 if (listadp == adp) 10796 break; 10797 /* continue if ad_oldlbn is not a fragment */ 10798 if (listadp->ad_oldsize == 0 || 10799 listadp->ad_oldsize == bsize) 10800 continue; 10801 /* hit a fragment */ 10802 return; 10803 } 10804 /* 10805 * If we have reached the end of the current list without 10806 * finding the just finished dependency, then it must be 10807 * on the future dependency list. Future dependencies cannot 10808 * be freed until they are moved to the current list. 10809 */ 10810 if (listadp == NULL) { 10811#ifdef DEBUG 10812 if (adp->ad_state & EXTDATA) 10813 listhead = &inodedep->id_newextupdt; 10814 else 10815 listhead = &inodedep->id_newinoupdt; 10816 TAILQ_FOREACH(listadp, listhead, ad_next) 10817 /* found our block */ 10818 if (listadp == adp) 10819 break; 10820 if (listadp == NULL) 10821 panic("handle_allocdirect_partdone: lost dep"); 10822#endif /* DEBUG */ 10823 return; 10824 } 10825 /* 10826 * If we have found the just finished dependency, then queue 10827 * it along with anything that follows it that is complete. 10828 * Since the pointer has not yet been written in the inode 10829 * as the dependency prevents it, place the allocdirect on the 10830 * bufwait list where it will be freed once the pointer is 10831 * valid. 10832 */ 10833 if (wkhd == NULL) 10834 wkhd = &inodedep->id_bufwait; 10835 for (; adp; adp = listadp) { 10836 listadp = TAILQ_NEXT(adp, ad_next); 10837 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 10838 return; 10839 TAILQ_REMOVE(listhead, adp, ad_next); 10840 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 10841 } 10842} 10843 10844/* 10845 * Called from within softdep_disk_write_complete above. This routine 10846 * completes successfully written allocindirs. 10847 */ 10848static void 10849handle_allocindir_partdone(aip) 10850 struct allocindir *aip; /* the completed allocindir */ 10851{ 10852 struct indirdep *indirdep; 10853 10854 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 10855 return; 10856 indirdep = aip->ai_indirdep; 10857 LIST_REMOVE(aip, ai_next); 10858 /* 10859 * Don't set a pointer while the buffer is undergoing IO or while 10860 * we have active truncations. 10861 */ 10862 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 10863 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 10864 return; 10865 } 10866 if (indirdep->ir_state & UFS1FMT) 10867 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 10868 aip->ai_newblkno; 10869 else 10870 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 10871 aip->ai_newblkno; 10872 /* 10873 * Await the pointer write before freeing the allocindir. 10874 */ 10875 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 10876} 10877 10878/* 10879 * Release segments held on a jwork list. 10880 */ 10881static void 10882handle_jwork(wkhd) 10883 struct workhead *wkhd; 10884{ 10885 struct worklist *wk; 10886 10887 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10888 WORKLIST_REMOVE(wk); 10889 switch (wk->wk_type) { 10890 case D_JSEGDEP: 10891 free_jsegdep(WK_JSEGDEP(wk)); 10892 continue; 10893 case D_FREEDEP: 10894 free_freedep(WK_FREEDEP(wk)); 10895 continue; 10896 case D_FREEFRAG: 10897 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 10898 WORKITEM_FREE(wk, D_FREEFRAG); 10899 continue; 10900 case D_FREEWORK: 10901 handle_written_freework(WK_FREEWORK(wk)); 10902 continue; 10903 default: 10904 panic("handle_jwork: Unknown type %s\n", 10905 TYPENAME(wk->wk_type)); 10906 } 10907 } 10908} 10909 10910/* 10911 * Handle the bufwait list on an inode when it is safe to release items 10912 * held there. This normally happens after an inode block is written but 10913 * may be delayed and handled later if there are pending journal items that 10914 * are not yet safe to be released. 10915 */ 10916static struct freefile * 10917handle_bufwait(inodedep, refhd) 10918 struct inodedep *inodedep; 10919 struct workhead *refhd; 10920{ 10921 struct jaddref *jaddref; 10922 struct freefile *freefile; 10923 struct worklist *wk; 10924 10925 freefile = NULL; 10926 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 10927 WORKLIST_REMOVE(wk); 10928 switch (wk->wk_type) { 10929 case D_FREEFILE: 10930 /* 10931 * We defer adding freefile to the worklist 10932 * until all other additions have been made to 10933 * ensure that it will be done after all the 10934 * old blocks have been freed. 10935 */ 10936 if (freefile != NULL) 10937 panic("handle_bufwait: freefile"); 10938 freefile = WK_FREEFILE(wk); 10939 continue; 10940 10941 case D_MKDIR: 10942 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 10943 continue; 10944 10945 case D_DIRADD: 10946 diradd_inode_written(WK_DIRADD(wk), inodedep); 10947 continue; 10948 10949 case D_FREEFRAG: 10950 wk->wk_state |= COMPLETE; 10951 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 10952 add_to_worklist(wk, 0); 10953 continue; 10954 10955 case D_DIRREM: 10956 wk->wk_state |= COMPLETE; 10957 add_to_worklist(wk, 0); 10958 continue; 10959 10960 case D_ALLOCDIRECT: 10961 case D_ALLOCINDIR: 10962 free_newblk(WK_NEWBLK(wk)); 10963 continue; 10964 10965 case D_JNEWBLK: 10966 wk->wk_state |= COMPLETE; 10967 free_jnewblk(WK_JNEWBLK(wk)); 10968 continue; 10969 10970 /* 10971 * Save freed journal segments and add references on 10972 * the supplied list which will delay their release 10973 * until the cg bitmap is cleared on disk. 10974 */ 10975 case D_JSEGDEP: 10976 if (refhd == NULL) 10977 free_jsegdep(WK_JSEGDEP(wk)); 10978 else 10979 WORKLIST_INSERT(refhd, wk); 10980 continue; 10981 10982 case D_JADDREF: 10983 jaddref = WK_JADDREF(wk); 10984 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 10985 if_deps); 10986 /* 10987 * Transfer any jaddrefs to the list to be freed with 10988 * the bitmap if we're handling a removed file. 10989 */ 10990 if (refhd == NULL) { 10991 wk->wk_state |= COMPLETE; 10992 free_jaddref(jaddref); 10993 } else 10994 WORKLIST_INSERT(refhd, wk); 10995 continue; 10996 10997 default: 10998 panic("handle_bufwait: Unknown type %p(%s)", 10999 wk, TYPENAME(wk->wk_type)); 11000 /* NOTREACHED */ 11001 } 11002 } 11003 return (freefile); 11004} 11005/* 11006 * Called from within softdep_disk_write_complete above to restore 11007 * in-memory inode block contents to their most up-to-date state. Note 11008 * that this routine is always called from interrupt level with further 11009 * splbio interrupts blocked. 11010 */ 11011static int 11012handle_written_inodeblock(inodedep, bp) 11013 struct inodedep *inodedep; 11014 struct buf *bp; /* buffer containing the inode block */ 11015{ 11016 struct freefile *freefile; 11017 struct allocdirect *adp, *nextadp; 11018 struct ufs1_dinode *dp1 = NULL; 11019 struct ufs2_dinode *dp2 = NULL; 11020 struct workhead wkhd; 11021 int hadchanges, fstype; 11022 ino_t freelink; 11023 11024 LIST_INIT(&wkhd); 11025 hadchanges = 0; 11026 freefile = NULL; 11027 if ((inodedep->id_state & IOSTARTED) == 0) 11028 panic("handle_written_inodeblock: not started"); 11029 inodedep->id_state &= ~IOSTARTED; 11030 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11031 fstype = UFS1; 11032 dp1 = (struct ufs1_dinode *)bp->b_data + 11033 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11034 freelink = dp1->di_freelink; 11035 } else { 11036 fstype = UFS2; 11037 dp2 = (struct ufs2_dinode *)bp->b_data + 11038 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11039 freelink = dp2->di_freelink; 11040 } 11041 /* 11042 * Leave this inodeblock dirty until it's in the list. 11043 */ 11044 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) { 11045 struct inodedep *inon; 11046 11047 inon = TAILQ_NEXT(inodedep, id_unlinked); 11048 if ((inon == NULL && freelink == 0) || 11049 (inon && inon->id_ino == freelink)) { 11050 if (inon) 11051 inon->id_state |= UNLINKPREV; 11052 inodedep->id_state |= UNLINKNEXT; 11053 } 11054 hadchanges = 1; 11055 } 11056 /* 11057 * If we had to rollback the inode allocation because of 11058 * bitmaps being incomplete, then simply restore it. 11059 * Keep the block dirty so that it will not be reclaimed until 11060 * all associated dependencies have been cleared and the 11061 * corresponding updates written to disk. 11062 */ 11063 if (inodedep->id_savedino1 != NULL) { 11064 hadchanges = 1; 11065 if (fstype == UFS1) 11066 *dp1 = *inodedep->id_savedino1; 11067 else 11068 *dp2 = *inodedep->id_savedino2; 11069 free(inodedep->id_savedino1, M_SAVEDINO); 11070 inodedep->id_savedino1 = NULL; 11071 if ((bp->b_flags & B_DELWRI) == 0) 11072 stat_inode_bitmap++; 11073 bdirty(bp); 11074 /* 11075 * If the inode is clear here and GOINGAWAY it will never 11076 * be written. Process the bufwait and clear any pending 11077 * work which may include the freefile. 11078 */ 11079 if (inodedep->id_state & GOINGAWAY) 11080 goto bufwait; 11081 return (1); 11082 } 11083 inodedep->id_state |= COMPLETE; 11084 /* 11085 * Roll forward anything that had to be rolled back before 11086 * the inode could be updated. 11087 */ 11088 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11089 nextadp = TAILQ_NEXT(adp, ad_next); 11090 if (adp->ad_state & ATTACHED) 11091 panic("handle_written_inodeblock: new entry"); 11092 if (fstype == UFS1) { 11093 if (adp->ad_offset < NDADDR) { 11094 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11095 panic("%s %s #%jd mismatch %d != %jd", 11096 "handle_written_inodeblock:", 11097 "direct pointer", 11098 (intmax_t)adp->ad_offset, 11099 dp1->di_db[adp->ad_offset], 11100 (intmax_t)adp->ad_oldblkno); 11101 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11102 } else { 11103 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 11104 panic("%s: %s #%jd allocated as %d", 11105 "handle_written_inodeblock", 11106 "indirect pointer", 11107 (intmax_t)adp->ad_offset - NDADDR, 11108 dp1->di_ib[adp->ad_offset - NDADDR]); 11109 dp1->di_ib[adp->ad_offset - NDADDR] = 11110 adp->ad_newblkno; 11111 } 11112 } else { 11113 if (adp->ad_offset < NDADDR) { 11114 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11115 panic("%s: %s #%jd %s %jd != %jd", 11116 "handle_written_inodeblock", 11117 "direct pointer", 11118 (intmax_t)adp->ad_offset, "mismatch", 11119 (intmax_t)dp2->di_db[adp->ad_offset], 11120 (intmax_t)adp->ad_oldblkno); 11121 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11122 } else { 11123 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 11124 panic("%s: %s #%jd allocated as %jd", 11125 "handle_written_inodeblock", 11126 "indirect pointer", 11127 (intmax_t)adp->ad_offset - NDADDR, 11128 (intmax_t) 11129 dp2->di_ib[adp->ad_offset - NDADDR]); 11130 dp2->di_ib[adp->ad_offset - NDADDR] = 11131 adp->ad_newblkno; 11132 } 11133 } 11134 adp->ad_state &= ~UNDONE; 11135 adp->ad_state |= ATTACHED; 11136 hadchanges = 1; 11137 } 11138 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11139 nextadp = TAILQ_NEXT(adp, ad_next); 11140 if (adp->ad_state & ATTACHED) 11141 panic("handle_written_inodeblock: new entry"); 11142 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11143 panic("%s: direct pointers #%jd %s %jd != %jd", 11144 "handle_written_inodeblock", 11145 (intmax_t)adp->ad_offset, "mismatch", 11146 (intmax_t)dp2->di_extb[adp->ad_offset], 11147 (intmax_t)adp->ad_oldblkno); 11148 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11149 adp->ad_state &= ~UNDONE; 11150 adp->ad_state |= ATTACHED; 11151 hadchanges = 1; 11152 } 11153 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11154 stat_direct_blk_ptrs++; 11155 /* 11156 * Reset the file size to its most up-to-date value. 11157 */ 11158 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11159 panic("handle_written_inodeblock: bad size"); 11160 if (inodedep->id_savednlink > LINK_MAX) 11161 panic("handle_written_inodeblock: Invalid link count " 11162 "%d for inodedep %p", inodedep->id_savednlink, inodedep); 11163 if (fstype == UFS1) { 11164 if (dp1->di_nlink != inodedep->id_savednlink) { 11165 dp1->di_nlink = inodedep->id_savednlink; 11166 hadchanges = 1; 11167 } 11168 if (dp1->di_size != inodedep->id_savedsize) { 11169 dp1->di_size = inodedep->id_savedsize; 11170 hadchanges = 1; 11171 } 11172 } else { 11173 if (dp2->di_nlink != inodedep->id_savednlink) { 11174 dp2->di_nlink = inodedep->id_savednlink; 11175 hadchanges = 1; 11176 } 11177 if (dp2->di_size != inodedep->id_savedsize) { 11178 dp2->di_size = inodedep->id_savedsize; 11179 hadchanges = 1; 11180 } 11181 if (dp2->di_extsize != inodedep->id_savedextsize) { 11182 dp2->di_extsize = inodedep->id_savedextsize; 11183 hadchanges = 1; 11184 } 11185 } 11186 inodedep->id_savedsize = -1; 11187 inodedep->id_savedextsize = -1; 11188 inodedep->id_savednlink = -1; 11189 /* 11190 * If there were any rollbacks in the inode block, then it must be 11191 * marked dirty so that its will eventually get written back in 11192 * its correct form. 11193 */ 11194 if (hadchanges) 11195 bdirty(bp); 11196bufwait: 11197 /* 11198 * Process any allocdirects that completed during the update. 11199 */ 11200 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11201 handle_allocdirect_partdone(adp, &wkhd); 11202 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11203 handle_allocdirect_partdone(adp, &wkhd); 11204 /* 11205 * Process deallocations that were held pending until the 11206 * inode had been written to disk. Freeing of the inode 11207 * is delayed until after all blocks have been freed to 11208 * avoid creation of new <vfsid, inum, lbn> triples 11209 * before the old ones have been deleted. Completely 11210 * unlinked inodes are not processed until the unlinked 11211 * inode list is written or the last reference is removed. 11212 */ 11213 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11214 freefile = handle_bufwait(inodedep, NULL); 11215 if (freefile && !LIST_EMPTY(&wkhd)) { 11216 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11217 freefile = NULL; 11218 } 11219 } 11220 /* 11221 * Move rolled forward dependency completions to the bufwait list 11222 * now that those that were already written have been processed. 11223 */ 11224 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11225 panic("handle_written_inodeblock: bufwait but no changes"); 11226 jwork_move(&inodedep->id_bufwait, &wkhd); 11227 11228 if (freefile != NULL) { 11229 /* 11230 * If the inode is goingaway it was never written. Fake up 11231 * the state here so free_inodedep() can succeed. 11232 */ 11233 if (inodedep->id_state & GOINGAWAY) 11234 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11235 if (free_inodedep(inodedep) == 0) 11236 panic("handle_written_inodeblock: live inodedep %p", 11237 inodedep); 11238 add_to_worklist(&freefile->fx_list, 0); 11239 return (0); 11240 } 11241 11242 /* 11243 * If no outstanding dependencies, free it. 11244 */ 11245 if (free_inodedep(inodedep) || 11246 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11247 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11248 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11249 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11250 return (0); 11251 return (hadchanges); 11252} 11253 11254static int 11255handle_written_indirdep(indirdep, bp, bpp) 11256 struct indirdep *indirdep; 11257 struct buf *bp; 11258 struct buf **bpp; 11259{ 11260 struct allocindir *aip; 11261 struct buf *sbp; 11262 int chgs; 11263 11264 if (indirdep->ir_state & GOINGAWAY) 11265 panic("handle_written_indirdep: indirdep gone"); 11266 if ((indirdep->ir_state & IOSTARTED) == 0) 11267 panic("handle_written_indirdep: IO not started"); 11268 chgs = 0; 11269 /* 11270 * If there were rollbacks revert them here. 11271 */ 11272 if (indirdep->ir_saveddata) { 11273 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11274 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11275 free(indirdep->ir_saveddata, M_INDIRDEP); 11276 indirdep->ir_saveddata = NULL; 11277 } 11278 chgs = 1; 11279 } 11280 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11281 indirdep->ir_state |= ATTACHED; 11282 /* 11283 * Move allocindirs with written pointers to the completehd if 11284 * the indirdep's pointer is not yet written. Otherwise 11285 * free them here. 11286 */ 11287 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) { 11288 LIST_REMOVE(aip, ai_next); 11289 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11290 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11291 ai_next); 11292 newblk_freefrag(&aip->ai_block); 11293 continue; 11294 } 11295 free_newblk(&aip->ai_block); 11296 } 11297 /* 11298 * Move allocindirs that have finished dependency processing from 11299 * the done list to the write list after updating the pointers. 11300 */ 11301 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11302 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 11303 handle_allocindir_partdone(aip); 11304 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11305 panic("disk_write_complete: not gone"); 11306 chgs = 1; 11307 } 11308 } 11309 /* 11310 * Preserve the indirdep if there were any changes or if it is not 11311 * yet valid on disk. 11312 */ 11313 if (chgs) { 11314 stat_indir_blk_ptrs++; 11315 bdirty(bp); 11316 return (1); 11317 } 11318 /* 11319 * If there were no changes we can discard the savedbp and detach 11320 * ourselves from the buf. We are only carrying completed pointers 11321 * in this case. 11322 */ 11323 sbp = indirdep->ir_savebp; 11324 sbp->b_flags |= B_INVAL | B_NOCACHE; 11325 indirdep->ir_savebp = NULL; 11326 indirdep->ir_bp = NULL; 11327 if (*bpp != NULL) 11328 panic("handle_written_indirdep: bp already exists."); 11329 *bpp = sbp; 11330 /* 11331 * The indirdep may not be freed until its parent points at it. 11332 */ 11333 if (indirdep->ir_state & DEPCOMPLETE) 11334 free_indirdep(indirdep); 11335 11336 return (0); 11337} 11338 11339/* 11340 * Process a diradd entry after its dependent inode has been written. 11341 * This routine must be called with splbio interrupts blocked. 11342 */ 11343static void 11344diradd_inode_written(dap, inodedep) 11345 struct diradd *dap; 11346 struct inodedep *inodedep; 11347{ 11348 11349 dap->da_state |= COMPLETE; 11350 complete_diradd(dap); 11351 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11352} 11353 11354/* 11355 * Returns true if the bmsafemap will have rollbacks when written. Must 11356 * only be called with lk and the buf lock on the cg held. 11357 */ 11358static int 11359bmsafemap_backgroundwrite(bmsafemap, bp) 11360 struct bmsafemap *bmsafemap; 11361 struct buf *bp; 11362{ 11363 int dirty; 11364 11365 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11366 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11367 /* 11368 * If we're initiating a background write we need to process the 11369 * rollbacks as they exist now, not as they exist when IO starts. 11370 * No other consumers will look at the contents of the shadowed 11371 * buf so this is safe to do here. 11372 */ 11373 if (bp->b_xflags & BX_BKGRDMARKER) 11374 initiate_write_bmsafemap(bmsafemap, bp); 11375 11376 return (dirty); 11377} 11378 11379/* 11380 * Re-apply an allocation when a cg write is complete. 11381 */ 11382static int 11383jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11384 struct jnewblk *jnewblk; 11385 struct fs *fs; 11386 struct cg *cgp; 11387 uint8_t *blksfree; 11388{ 11389 ufs1_daddr_t fragno; 11390 ufs2_daddr_t blkno; 11391 long cgbno, bbase; 11392 int frags, blk; 11393 int i; 11394 11395 frags = 0; 11396 cgbno = dtogd(fs, jnewblk->jn_blkno); 11397 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11398 if (isclr(blksfree, cgbno + i)) 11399 panic("jnewblk_rollforward: re-allocated fragment"); 11400 frags++; 11401 } 11402 if (frags == fs->fs_frag) { 11403 blkno = fragstoblks(fs, cgbno); 11404 ffs_clrblock(fs, blksfree, (long)blkno); 11405 ffs_clusteracct(fs, cgp, blkno, -1); 11406 cgp->cg_cs.cs_nbfree--; 11407 } else { 11408 bbase = cgbno - fragnum(fs, cgbno); 11409 cgbno += jnewblk->jn_oldfrags; 11410 /* If a complete block had been reassembled, account for it. */ 11411 fragno = fragstoblks(fs, bbase); 11412 if (ffs_isblock(fs, blksfree, fragno)) { 11413 cgp->cg_cs.cs_nffree += fs->fs_frag; 11414 ffs_clusteracct(fs, cgp, fragno, -1); 11415 cgp->cg_cs.cs_nbfree--; 11416 } 11417 /* Decrement the old frags. */ 11418 blk = blkmap(fs, blksfree, bbase); 11419 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11420 /* Allocate the fragment */ 11421 for (i = 0; i < frags; i++) 11422 clrbit(blksfree, cgbno + i); 11423 cgp->cg_cs.cs_nffree -= frags; 11424 /* Add back in counts associated with the new frags */ 11425 blk = blkmap(fs, blksfree, bbase); 11426 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11427 } 11428 return (frags); 11429} 11430 11431/* 11432 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11433 * changes if it's not a background write. Set all written dependencies 11434 * to DEPCOMPLETE and free the structure if possible. 11435 */ 11436static int 11437handle_written_bmsafemap(bmsafemap, bp) 11438 struct bmsafemap *bmsafemap; 11439 struct buf *bp; 11440{ 11441 struct newblk *newblk; 11442 struct inodedep *inodedep; 11443 struct jaddref *jaddref, *jatmp; 11444 struct jnewblk *jnewblk, *jntmp; 11445 struct ufsmount *ump; 11446 uint8_t *inosused; 11447 uint8_t *blksfree; 11448 struct cg *cgp; 11449 struct fs *fs; 11450 ino_t ino; 11451 int foreground; 11452 int chgs; 11453 11454 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11455 panic("initiate_write_bmsafemap: Not started\n"); 11456 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11457 chgs = 0; 11458 bmsafemap->sm_state &= ~IOSTARTED; 11459 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11460 /* 11461 * Release journal work that was waiting on the write. 11462 */ 11463 handle_jwork(&bmsafemap->sm_freewr); 11464 11465 /* 11466 * Restore unwritten inode allocation pending jaddref writes. 11467 */ 11468 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11469 cgp = (struct cg *)bp->b_data; 11470 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11471 inosused = cg_inosused(cgp); 11472 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11473 ja_bmdeps, jatmp) { 11474 if ((jaddref->ja_state & UNDONE) == 0) 11475 continue; 11476 ino = jaddref->ja_ino % fs->fs_ipg; 11477 if (isset(inosused, ino)) 11478 panic("handle_written_bmsafemap: " 11479 "re-allocated inode"); 11480 /* Do the roll-forward only if it's a real copy. */ 11481 if (foreground) { 11482 if ((jaddref->ja_mode & IFMT) == IFDIR) 11483 cgp->cg_cs.cs_ndir++; 11484 cgp->cg_cs.cs_nifree--; 11485 setbit(inosused, ino); 11486 chgs = 1; 11487 } 11488 jaddref->ja_state &= ~UNDONE; 11489 jaddref->ja_state |= ATTACHED; 11490 free_jaddref(jaddref); 11491 } 11492 } 11493 /* 11494 * Restore any block allocations which are pending journal writes. 11495 */ 11496 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11497 cgp = (struct cg *)bp->b_data; 11498 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11499 blksfree = cg_blksfree(cgp); 11500 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11501 jntmp) { 11502 if ((jnewblk->jn_state & UNDONE) == 0) 11503 continue; 11504 /* Do the roll-forward only if it's a real copy. */ 11505 if (foreground && 11506 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11507 chgs = 1; 11508 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11509 jnewblk->jn_state |= ATTACHED; 11510 free_jnewblk(jnewblk); 11511 } 11512 } 11513 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11514 newblk->nb_state |= DEPCOMPLETE; 11515 newblk->nb_state &= ~ONDEPLIST; 11516 newblk->nb_bmsafemap = NULL; 11517 LIST_REMOVE(newblk, nb_deps); 11518 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11519 handle_allocdirect_partdone( 11520 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11521 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11522 handle_allocindir_partdone( 11523 WK_ALLOCINDIR(&newblk->nb_list)); 11524 else if (newblk->nb_list.wk_type != D_NEWBLK) 11525 panic("handle_written_bmsafemap: Unexpected type: %s", 11526 TYPENAME(newblk->nb_list.wk_type)); 11527 } 11528 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11529 inodedep->id_state |= DEPCOMPLETE; 11530 inodedep->id_state &= ~ONDEPLIST; 11531 LIST_REMOVE(inodedep, id_deps); 11532 inodedep->id_bmsafemap = NULL; 11533 } 11534 LIST_REMOVE(bmsafemap, sm_next); 11535 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11536 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11537 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11538 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11539 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11540 LIST_REMOVE(bmsafemap, sm_hash); 11541 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11542 return (0); 11543 } 11544 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11545 if (foreground) 11546 bdirty(bp); 11547 return (1); 11548} 11549 11550/* 11551 * Try to free a mkdir dependency. 11552 */ 11553static void 11554complete_mkdir(mkdir) 11555 struct mkdir *mkdir; 11556{ 11557 struct diradd *dap; 11558 11559 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11560 return; 11561 LIST_REMOVE(mkdir, md_mkdirs); 11562 dap = mkdir->md_diradd; 11563 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11564 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11565 dap->da_state |= DEPCOMPLETE; 11566 complete_diradd(dap); 11567 } 11568 WORKITEM_FREE(mkdir, D_MKDIR); 11569} 11570 11571/* 11572 * Handle the completion of a mkdir dependency. 11573 */ 11574static void 11575handle_written_mkdir(mkdir, type) 11576 struct mkdir *mkdir; 11577 int type; 11578{ 11579 11580 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 11581 panic("handle_written_mkdir: bad type"); 11582 mkdir->md_state |= COMPLETE; 11583 complete_mkdir(mkdir); 11584} 11585 11586static int 11587free_pagedep(pagedep) 11588 struct pagedep *pagedep; 11589{ 11590 int i; 11591 11592 if (pagedep->pd_state & NEWBLOCK) 11593 return (0); 11594 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 11595 return (0); 11596 for (i = 0; i < DAHASHSZ; i++) 11597 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 11598 return (0); 11599 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 11600 return (0); 11601 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 11602 return (0); 11603 if (pagedep->pd_state & ONWORKLIST) 11604 WORKLIST_REMOVE(&pagedep->pd_list); 11605 LIST_REMOVE(pagedep, pd_hash); 11606 WORKITEM_FREE(pagedep, D_PAGEDEP); 11607 11608 return (1); 11609} 11610 11611/* 11612 * Called from within softdep_disk_write_complete above. 11613 * A write operation was just completed. Removed inodes can 11614 * now be freed and associated block pointers may be committed. 11615 * Note that this routine is always called from interrupt level 11616 * with further splbio interrupts blocked. 11617 */ 11618static int 11619handle_written_filepage(pagedep, bp) 11620 struct pagedep *pagedep; 11621 struct buf *bp; /* buffer containing the written page */ 11622{ 11623 struct dirrem *dirrem; 11624 struct diradd *dap, *nextdap; 11625 struct direct *ep; 11626 int i, chgs; 11627 11628 if ((pagedep->pd_state & IOSTARTED) == 0) 11629 panic("handle_written_filepage: not started"); 11630 pagedep->pd_state &= ~IOSTARTED; 11631 /* 11632 * Process any directory removals that have been committed. 11633 */ 11634 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 11635 LIST_REMOVE(dirrem, dm_next); 11636 dirrem->dm_state |= COMPLETE; 11637 dirrem->dm_dirinum = pagedep->pd_ino; 11638 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 11639 ("handle_written_filepage: Journal entries not written.")); 11640 add_to_worklist(&dirrem->dm_list, 0); 11641 } 11642 /* 11643 * Free any directory additions that have been committed. 11644 * If it is a newly allocated block, we have to wait until 11645 * the on-disk directory inode claims the new block. 11646 */ 11647 if ((pagedep->pd_state & NEWBLOCK) == 0) 11648 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 11649 free_diradd(dap, NULL); 11650 /* 11651 * Uncommitted directory entries must be restored. 11652 */ 11653 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 11654 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 11655 dap = nextdap) { 11656 nextdap = LIST_NEXT(dap, da_pdlist); 11657 if (dap->da_state & ATTACHED) 11658 panic("handle_written_filepage: attached"); 11659 ep = (struct direct *) 11660 ((char *)bp->b_data + dap->da_offset); 11661 ep->d_ino = dap->da_newinum; 11662 dap->da_state &= ~UNDONE; 11663 dap->da_state |= ATTACHED; 11664 chgs = 1; 11665 /* 11666 * If the inode referenced by the directory has 11667 * been written out, then the dependency can be 11668 * moved to the pending list. 11669 */ 11670 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 11671 LIST_REMOVE(dap, da_pdlist); 11672 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 11673 da_pdlist); 11674 } 11675 } 11676 } 11677 /* 11678 * If there were any rollbacks in the directory, then it must be 11679 * marked dirty so that its will eventually get written back in 11680 * its correct form. 11681 */ 11682 if (chgs) { 11683 if ((bp->b_flags & B_DELWRI) == 0) 11684 stat_dir_entry++; 11685 bdirty(bp); 11686 return (1); 11687 } 11688 /* 11689 * If we are not waiting for a new directory block to be 11690 * claimed by its inode, then the pagedep will be freed. 11691 * Otherwise it will remain to track any new entries on 11692 * the page in case they are fsync'ed. 11693 */ 11694 free_pagedep(pagedep); 11695 return (0); 11696} 11697 11698/* 11699 * Writing back in-core inode structures. 11700 * 11701 * The filesystem only accesses an inode's contents when it occupies an 11702 * "in-core" inode structure. These "in-core" structures are separate from 11703 * the page frames used to cache inode blocks. Only the latter are 11704 * transferred to/from the disk. So, when the updated contents of the 11705 * "in-core" inode structure are copied to the corresponding in-memory inode 11706 * block, the dependencies are also transferred. The following procedure is 11707 * called when copying a dirty "in-core" inode to a cached inode block. 11708 */ 11709 11710/* 11711 * Called when an inode is loaded from disk. If the effective link count 11712 * differed from the actual link count when it was last flushed, then we 11713 * need to ensure that the correct effective link count is put back. 11714 */ 11715void 11716softdep_load_inodeblock(ip) 11717 struct inode *ip; /* the "in_core" copy of the inode */ 11718{ 11719 struct inodedep *inodedep; 11720 11721 /* 11722 * Check for alternate nlink count. 11723 */ 11724 ip->i_effnlink = ip->i_nlink; 11725 ACQUIRE_LOCK(&lk); 11726 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 11727 &inodedep) == 0) { 11728 FREE_LOCK(&lk); 11729 return; 11730 } 11731 ip->i_effnlink -= inodedep->id_nlinkdelta; 11732 FREE_LOCK(&lk); 11733} 11734 11735/* 11736 * This routine is called just before the "in-core" inode 11737 * information is to be copied to the in-memory inode block. 11738 * Recall that an inode block contains several inodes. If 11739 * the force flag is set, then the dependencies will be 11740 * cleared so that the update can always be made. Note that 11741 * the buffer is locked when this routine is called, so we 11742 * will never be in the middle of writing the inode block 11743 * to disk. 11744 */ 11745void 11746softdep_update_inodeblock(ip, bp, waitfor) 11747 struct inode *ip; /* the "in_core" copy of the inode */ 11748 struct buf *bp; /* the buffer containing the inode block */ 11749 int waitfor; /* nonzero => update must be allowed */ 11750{ 11751 struct inodedep *inodedep; 11752 struct inoref *inoref; 11753 struct worklist *wk; 11754 struct mount *mp; 11755 struct buf *ibp; 11756 struct fs *fs; 11757 int error; 11758 11759 mp = UFSTOVFS(ip->i_ump); 11760 fs = ip->i_fs; 11761 /* 11762 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 11763 * does not have access to the in-core ip so must write directly into 11764 * the inode block buffer when setting freelink. 11765 */ 11766 if (fs->fs_magic == FS_UFS1_MAGIC) 11767 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 11768 ino_to_fsbo(fs, ip->i_number))->di_freelink); 11769 else 11770 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 11771 ino_to_fsbo(fs, ip->i_number))->di_freelink); 11772 /* 11773 * If the effective link count is not equal to the actual link 11774 * count, then we must track the difference in an inodedep while 11775 * the inode is (potentially) tossed out of the cache. Otherwise, 11776 * if there is no existing inodedep, then there are no dependencies 11777 * to track. 11778 */ 11779 ACQUIRE_LOCK(&lk); 11780again: 11781 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 11782 FREE_LOCK(&lk); 11783 if (ip->i_effnlink != ip->i_nlink) 11784 panic("softdep_update_inodeblock: bad link count"); 11785 return; 11786 } 11787 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 11788 panic("softdep_update_inodeblock: bad delta"); 11789 /* 11790 * If we're flushing all dependencies we must also move any waiting 11791 * for journal writes onto the bufwait list prior to I/O. 11792 */ 11793 if (waitfor) { 11794 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 11795 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 11796 == DEPCOMPLETE) { 11797 jwait(&inoref->if_list, MNT_WAIT); 11798 goto again; 11799 } 11800 } 11801 } 11802 /* 11803 * Changes have been initiated. Anything depending on these 11804 * changes cannot occur until this inode has been written. 11805 */ 11806 inodedep->id_state &= ~COMPLETE; 11807 if ((inodedep->id_state & ONWORKLIST) == 0) 11808 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 11809 /* 11810 * Any new dependencies associated with the incore inode must 11811 * now be moved to the list associated with the buffer holding 11812 * the in-memory copy of the inode. Once merged process any 11813 * allocdirects that are completed by the merger. 11814 */ 11815 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 11816 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 11817 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 11818 NULL); 11819 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 11820 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 11821 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 11822 NULL); 11823 /* 11824 * Now that the inode has been pushed into the buffer, the 11825 * operations dependent on the inode being written to disk 11826 * can be moved to the id_bufwait so that they will be 11827 * processed when the buffer I/O completes. 11828 */ 11829 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 11830 WORKLIST_REMOVE(wk); 11831 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 11832 } 11833 /* 11834 * Newly allocated inodes cannot be written until the bitmap 11835 * that allocates them have been written (indicated by 11836 * DEPCOMPLETE being set in id_state). If we are doing a 11837 * forced sync (e.g., an fsync on a file), we force the bitmap 11838 * to be written so that the update can be done. 11839 */ 11840 if (waitfor == 0) { 11841 FREE_LOCK(&lk); 11842 return; 11843 } 11844retry: 11845 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 11846 FREE_LOCK(&lk); 11847 return; 11848 } 11849 ibp = inodedep->id_bmsafemap->sm_buf; 11850 ibp = getdirtybuf(ibp, &lk, MNT_WAIT); 11851 if (ibp == NULL) { 11852 /* 11853 * If ibp came back as NULL, the dependency could have been 11854 * freed while we slept. Look it up again, and check to see 11855 * that it has completed. 11856 */ 11857 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 11858 goto retry; 11859 FREE_LOCK(&lk); 11860 return; 11861 } 11862 FREE_LOCK(&lk); 11863 if ((error = bwrite(ibp)) != 0) 11864 softdep_error("softdep_update_inodeblock: bwrite", error); 11865} 11866 11867/* 11868 * Merge the a new inode dependency list (such as id_newinoupdt) into an 11869 * old inode dependency list (such as id_inoupdt). This routine must be 11870 * called with splbio interrupts blocked. 11871 */ 11872static void 11873merge_inode_lists(newlisthead, oldlisthead) 11874 struct allocdirectlst *newlisthead; 11875 struct allocdirectlst *oldlisthead; 11876{ 11877 struct allocdirect *listadp, *newadp; 11878 11879 newadp = TAILQ_FIRST(newlisthead); 11880 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 11881 if (listadp->ad_offset < newadp->ad_offset) { 11882 listadp = TAILQ_NEXT(listadp, ad_next); 11883 continue; 11884 } 11885 TAILQ_REMOVE(newlisthead, newadp, ad_next); 11886 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 11887 if (listadp->ad_offset == newadp->ad_offset) { 11888 allocdirect_merge(oldlisthead, newadp, 11889 listadp); 11890 listadp = newadp; 11891 } 11892 newadp = TAILQ_FIRST(newlisthead); 11893 } 11894 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 11895 TAILQ_REMOVE(newlisthead, newadp, ad_next); 11896 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 11897 } 11898} 11899 11900/* 11901 * If we are doing an fsync, then we must ensure that any directory 11902 * entries for the inode have been written after the inode gets to disk. 11903 */ 11904int 11905softdep_fsync(vp) 11906 struct vnode *vp; /* the "in_core" copy of the inode */ 11907{ 11908 struct inodedep *inodedep; 11909 struct pagedep *pagedep; 11910 struct inoref *inoref; 11911 struct worklist *wk; 11912 struct diradd *dap; 11913 struct mount *mp; 11914 struct vnode *pvp; 11915 struct inode *ip; 11916 struct buf *bp; 11917 struct fs *fs; 11918 struct thread *td = curthread; 11919 int error, flushparent, pagedep_new_block; 11920 ino_t parentino; 11921 ufs_lbn_t lbn; 11922 11923 ip = VTOI(vp); 11924 fs = ip->i_fs; 11925 mp = vp->v_mount; 11926 ACQUIRE_LOCK(&lk); 11927restart: 11928 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 11929 FREE_LOCK(&lk); 11930 return (0); 11931 } 11932 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 11933 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 11934 == DEPCOMPLETE) { 11935 jwait(&inoref->if_list, MNT_WAIT); 11936 goto restart; 11937 } 11938 } 11939 if (!LIST_EMPTY(&inodedep->id_inowait) || 11940 !TAILQ_EMPTY(&inodedep->id_extupdt) || 11941 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 11942 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 11943 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 11944 panic("softdep_fsync: pending ops %p", inodedep); 11945 for (error = 0, flushparent = 0; ; ) { 11946 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 11947 break; 11948 if (wk->wk_type != D_DIRADD) 11949 panic("softdep_fsync: Unexpected type %s", 11950 TYPENAME(wk->wk_type)); 11951 dap = WK_DIRADD(wk); 11952 /* 11953 * Flush our parent if this directory entry has a MKDIR_PARENT 11954 * dependency or is contained in a newly allocated block. 11955 */ 11956 if (dap->da_state & DIRCHG) 11957 pagedep = dap->da_previous->dm_pagedep; 11958 else 11959 pagedep = dap->da_pagedep; 11960 parentino = pagedep->pd_ino; 11961 lbn = pagedep->pd_lbn; 11962 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 11963 panic("softdep_fsync: dirty"); 11964 if ((dap->da_state & MKDIR_PARENT) || 11965 (pagedep->pd_state & NEWBLOCK)) 11966 flushparent = 1; 11967 else 11968 flushparent = 0; 11969 /* 11970 * If we are being fsync'ed as part of vgone'ing this vnode, 11971 * then we will not be able to release and recover the 11972 * vnode below, so we just have to give up on writing its 11973 * directory entry out. It will eventually be written, just 11974 * not now, but then the user was not asking to have it 11975 * written, so we are not breaking any promises. 11976 */ 11977 if (vp->v_iflag & VI_DOOMED) 11978 break; 11979 /* 11980 * We prevent deadlock by always fetching inodes from the 11981 * root, moving down the directory tree. Thus, when fetching 11982 * our parent directory, we first try to get the lock. If 11983 * that fails, we must unlock ourselves before requesting 11984 * the lock on our parent. See the comment in ufs_lookup 11985 * for details on possible races. 11986 */ 11987 FREE_LOCK(&lk); 11988 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 11989 FFSV_FORCEINSMQ)) { 11990 error = vfs_busy(mp, MBF_NOWAIT); 11991 if (error != 0) { 11992 vfs_ref(mp); 11993 VOP_UNLOCK(vp, 0); 11994 error = vfs_busy(mp, 0); 11995 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 11996 vfs_rel(mp); 11997 if (error != 0) 11998 return (ENOENT); 11999 if (vp->v_iflag & VI_DOOMED) { 12000 vfs_unbusy(mp); 12001 return (ENOENT); 12002 } 12003 } 12004 VOP_UNLOCK(vp, 0); 12005 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12006 &pvp, FFSV_FORCEINSMQ); 12007 vfs_unbusy(mp); 12008 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12009 if (vp->v_iflag & VI_DOOMED) { 12010 if (error == 0) 12011 vput(pvp); 12012 error = ENOENT; 12013 } 12014 if (error != 0) 12015 return (error); 12016 } 12017 /* 12018 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12019 * that are contained in direct blocks will be resolved by 12020 * doing a ffs_update. Pagedeps contained in indirect blocks 12021 * may require a complete sync'ing of the directory. So, we 12022 * try the cheap and fast ffs_update first, and if that fails, 12023 * then we do the slower ffs_syncvnode of the directory. 12024 */ 12025 if (flushparent) { 12026 int locked; 12027 12028 if ((error = ffs_update(pvp, 1)) != 0) { 12029 vput(pvp); 12030 return (error); 12031 } 12032 ACQUIRE_LOCK(&lk); 12033 locked = 1; 12034 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12035 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12036 if (wk->wk_type != D_DIRADD) 12037 panic("softdep_fsync: Unexpected type %s", 12038 TYPENAME(wk->wk_type)); 12039 dap = WK_DIRADD(wk); 12040 if (dap->da_state & DIRCHG) 12041 pagedep = dap->da_previous->dm_pagedep; 12042 else 12043 pagedep = dap->da_pagedep; 12044 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12045 FREE_LOCK(&lk); 12046 locked = 0; 12047 if (pagedep_new_block && (error = 12048 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12049 vput(pvp); 12050 return (error); 12051 } 12052 } 12053 } 12054 if (locked) 12055 FREE_LOCK(&lk); 12056 } 12057 /* 12058 * Flush directory page containing the inode's name. 12059 */ 12060 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12061 &bp); 12062 if (error == 0) 12063 error = bwrite(bp); 12064 else 12065 brelse(bp); 12066 vput(pvp); 12067 if (error != 0) 12068 return (error); 12069 ACQUIRE_LOCK(&lk); 12070 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12071 break; 12072 } 12073 FREE_LOCK(&lk); 12074 return (0); 12075} 12076 12077/* 12078 * Flush all the dirty bitmaps associated with the block device 12079 * before flushing the rest of the dirty blocks so as to reduce 12080 * the number of dependencies that will have to be rolled back. 12081 * 12082 * XXX Unused? 12083 */ 12084void 12085softdep_fsync_mountdev(vp) 12086 struct vnode *vp; 12087{ 12088 struct buf *bp, *nbp; 12089 struct worklist *wk; 12090 struct bufobj *bo; 12091 12092 if (!vn_isdisk(vp, NULL)) 12093 panic("softdep_fsync_mountdev: vnode not a disk"); 12094 bo = &vp->v_bufobj; 12095restart: 12096 BO_LOCK(bo); 12097 ACQUIRE_LOCK(&lk); 12098 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12099 /* 12100 * If it is already scheduled, skip to the next buffer. 12101 */ 12102 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12103 continue; 12104 12105 if ((bp->b_flags & B_DELWRI) == 0) 12106 panic("softdep_fsync_mountdev: not dirty"); 12107 /* 12108 * We are only interested in bitmaps with outstanding 12109 * dependencies. 12110 */ 12111 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12112 wk->wk_type != D_BMSAFEMAP || 12113 (bp->b_vflags & BV_BKGRDINPROG)) { 12114 BUF_UNLOCK(bp); 12115 continue; 12116 } 12117 FREE_LOCK(&lk); 12118 BO_UNLOCK(bo); 12119 bremfree(bp); 12120 (void) bawrite(bp); 12121 goto restart; 12122 } 12123 FREE_LOCK(&lk); 12124 drain_output(vp); 12125 BO_UNLOCK(bo); 12126} 12127 12128/* 12129 * Sync all cylinder groups that were dirty at the time this function is 12130 * called. Newly dirtied cgs will be inserted before the sintenel. This 12131 * is used to flush freedep activity that may be holding up writes to a 12132 * indirect block. 12133 */ 12134static int 12135sync_cgs(mp, waitfor) 12136 struct mount *mp; 12137 int waitfor; 12138{ 12139 struct bmsafemap *bmsafemap; 12140 struct bmsafemap *sintenel; 12141 struct ufsmount *ump; 12142 struct buf *bp; 12143 int error; 12144 12145 sintenel = malloc(sizeof(*sintenel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12146 sintenel->sm_cg = -1; 12147 ump = VFSTOUFS(mp); 12148 error = 0; 12149 ACQUIRE_LOCK(&lk); 12150 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sintenel, sm_next); 12151 for (bmsafemap = LIST_NEXT(sintenel, sm_next); bmsafemap != NULL; 12152 bmsafemap = LIST_NEXT(sintenel, sm_next)) { 12153 /* Skip sintenels and cgs with no work to release. */ 12154 if (bmsafemap->sm_cg == -1 || 12155 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12156 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12157 LIST_REMOVE(sintenel, sm_next); 12158 LIST_INSERT_AFTER(bmsafemap, sintenel, sm_next); 12159 continue; 12160 } 12161 /* 12162 * If we don't get the lock and we're waiting try again, if 12163 * not move on to the next buf and try to sync it. 12164 */ 12165 bp = getdirtybuf(bmsafemap->sm_buf, &lk, waitfor); 12166 if (bp == NULL && waitfor == MNT_WAIT) 12167 continue; 12168 LIST_REMOVE(sintenel, sm_next); 12169 LIST_INSERT_AFTER(bmsafemap, sintenel, sm_next); 12170 if (bp == NULL) 12171 continue; 12172 FREE_LOCK(&lk); 12173 if (waitfor == MNT_NOWAIT) 12174 bawrite(bp); 12175 else 12176 error = bwrite(bp); 12177 ACQUIRE_LOCK(&lk); 12178 if (error) 12179 break; 12180 } 12181 LIST_REMOVE(sintenel, sm_next); 12182 FREE_LOCK(&lk); 12183 free(sintenel, M_BMSAFEMAP); 12184 return (error); 12185} 12186 12187/* 12188 * This routine is called when we are trying to synchronously flush a 12189 * file. This routine must eliminate any filesystem metadata dependencies 12190 * so that the syncing routine can succeed. 12191 */ 12192int 12193softdep_sync_metadata(struct vnode *vp) 12194{ 12195 int error; 12196 12197 /* 12198 * Ensure that any direct block dependencies have been cleared, 12199 * truncations are started, and inode references are journaled. 12200 */ 12201 ACQUIRE_LOCK(&lk); 12202 /* 12203 * Write all journal records to prevent rollbacks on devvp. 12204 */ 12205 if (vp->v_type == VCHR) 12206 softdep_flushjournal(vp->v_mount); 12207 error = flush_inodedep_deps(vp, vp->v_mount, VTOI(vp)->i_number); 12208 /* 12209 * Ensure that all truncates are written so we won't find deps on 12210 * indirect blocks. 12211 */ 12212 process_truncates(vp); 12213 FREE_LOCK(&lk); 12214 12215 return (error); 12216} 12217 12218/* 12219 * This routine is called when we are attempting to sync a buf with 12220 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12221 * other IO it can but returns EBUSY if the buffer is not yet able to 12222 * be written. Dependencies which will not cause rollbacks will always 12223 * return 0. 12224 */ 12225int 12226softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12227{ 12228 struct indirdep *indirdep; 12229 struct pagedep *pagedep; 12230 struct allocindir *aip; 12231 struct newblk *newblk; 12232 struct buf *nbp; 12233 struct worklist *wk; 12234 int i, error; 12235 12236 /* 12237 * For VCHR we just don't want to force flush any dependencies that 12238 * will cause rollbacks. 12239 */ 12240 if (vp->v_type == VCHR) { 12241 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12242 return (EBUSY); 12243 return (0); 12244 } 12245 ACQUIRE_LOCK(&lk); 12246 /* 12247 * As we hold the buffer locked, none of its dependencies 12248 * will disappear. 12249 */ 12250 error = 0; 12251top: 12252 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12253 switch (wk->wk_type) { 12254 12255 case D_ALLOCDIRECT: 12256 case D_ALLOCINDIR: 12257 newblk = WK_NEWBLK(wk); 12258 if (newblk->nb_jnewblk != NULL) { 12259 if (waitfor == MNT_NOWAIT) { 12260 error = EBUSY; 12261 goto out_unlock; 12262 } 12263 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12264 goto top; 12265 } 12266 if (newblk->nb_state & DEPCOMPLETE || 12267 waitfor == MNT_NOWAIT) 12268 continue; 12269 nbp = newblk->nb_bmsafemap->sm_buf; 12270 nbp = getdirtybuf(nbp, &lk, waitfor); 12271 if (nbp == NULL) 12272 goto top; 12273 FREE_LOCK(&lk); 12274 if ((error = bwrite(nbp)) != 0) 12275 goto out; 12276 ACQUIRE_LOCK(&lk); 12277 continue; 12278 12279 case D_INDIRDEP: 12280 indirdep = WK_INDIRDEP(wk); 12281 if (waitfor == MNT_NOWAIT) { 12282 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12283 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12284 error = EBUSY; 12285 goto out_unlock; 12286 } 12287 } 12288 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12289 panic("softdep_sync_buf: truncation pending."); 12290 restart: 12291 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12292 newblk = (struct newblk *)aip; 12293 if (newblk->nb_jnewblk != NULL) { 12294 jwait(&newblk->nb_jnewblk->jn_list, 12295 waitfor); 12296 goto restart; 12297 } 12298 if (newblk->nb_state & DEPCOMPLETE) 12299 continue; 12300 nbp = newblk->nb_bmsafemap->sm_buf; 12301 nbp = getdirtybuf(nbp, &lk, waitfor); 12302 if (nbp == NULL) 12303 goto restart; 12304 FREE_LOCK(&lk); 12305 if ((error = bwrite(nbp)) != 0) 12306 goto out; 12307 ACQUIRE_LOCK(&lk); 12308 goto restart; 12309 } 12310 continue; 12311 12312 case D_PAGEDEP: 12313 /* 12314 * Only flush directory entries in synchronous passes. 12315 */ 12316 if (waitfor != MNT_WAIT) { 12317 error = EBUSY; 12318 goto out_unlock; 12319 } 12320 /* 12321 * While syncing snapshots, we must allow recursive 12322 * lookups. 12323 */ 12324 BUF_AREC(bp); 12325 /* 12326 * We are trying to sync a directory that may 12327 * have dependencies on both its own metadata 12328 * and/or dependencies on the inodes of any 12329 * recently allocated files. We walk its diradd 12330 * lists pushing out the associated inode. 12331 */ 12332 pagedep = WK_PAGEDEP(wk); 12333 for (i = 0; i < DAHASHSZ; i++) { 12334 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12335 continue; 12336 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12337 &pagedep->pd_diraddhd[i]))) { 12338 BUF_NOREC(bp); 12339 goto out_unlock; 12340 } 12341 } 12342 BUF_NOREC(bp); 12343 continue; 12344 12345 case D_FREEWORK: 12346 case D_FREEDEP: 12347 case D_JSEGDEP: 12348 case D_JNEWBLK: 12349 continue; 12350 12351 default: 12352 panic("softdep_sync_buf: Unknown type %s", 12353 TYPENAME(wk->wk_type)); 12354 /* NOTREACHED */ 12355 } 12356 } 12357out_unlock: 12358 FREE_LOCK(&lk); 12359out: 12360 return (error); 12361} 12362 12363/* 12364 * Flush the dependencies associated with an inodedep. 12365 * Called with splbio blocked. 12366 */ 12367static int 12368flush_inodedep_deps(vp, mp, ino) 12369 struct vnode *vp; 12370 struct mount *mp; 12371 ino_t ino; 12372{ 12373 struct inodedep *inodedep; 12374 struct inoref *inoref; 12375 int error, waitfor; 12376 12377 /* 12378 * This work is done in two passes. The first pass grabs most 12379 * of the buffers and begins asynchronously writing them. The 12380 * only way to wait for these asynchronous writes is to sleep 12381 * on the filesystem vnode which may stay busy for a long time 12382 * if the filesystem is active. So, instead, we make a second 12383 * pass over the dependencies blocking on each write. In the 12384 * usual case we will be blocking against a write that we 12385 * initiated, so when it is done the dependency will have been 12386 * resolved. Thus the second pass is expected to end quickly. 12387 * We give a brief window at the top of the loop to allow 12388 * any pending I/O to complete. 12389 */ 12390 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12391 if (error) 12392 return (error); 12393 FREE_LOCK(&lk); 12394 ACQUIRE_LOCK(&lk); 12395restart: 12396 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12397 return (0); 12398 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12399 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12400 == DEPCOMPLETE) { 12401 jwait(&inoref->if_list, MNT_WAIT); 12402 goto restart; 12403 } 12404 } 12405 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12406 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12407 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12408 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12409 continue; 12410 /* 12411 * If pass2, we are done, otherwise do pass 2. 12412 */ 12413 if (waitfor == MNT_WAIT) 12414 break; 12415 waitfor = MNT_WAIT; 12416 } 12417 /* 12418 * Try freeing inodedep in case all dependencies have been removed. 12419 */ 12420 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12421 (void) free_inodedep(inodedep); 12422 return (0); 12423} 12424 12425/* 12426 * Flush an inode dependency list. 12427 * Called with splbio blocked. 12428 */ 12429static int 12430flush_deplist(listhead, waitfor, errorp) 12431 struct allocdirectlst *listhead; 12432 int waitfor; 12433 int *errorp; 12434{ 12435 struct allocdirect *adp; 12436 struct newblk *newblk; 12437 struct buf *bp; 12438 12439 mtx_assert(&lk, MA_OWNED); 12440 TAILQ_FOREACH(adp, listhead, ad_next) { 12441 newblk = (struct newblk *)adp; 12442 if (newblk->nb_jnewblk != NULL) { 12443 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12444 return (1); 12445 } 12446 if (newblk->nb_state & DEPCOMPLETE) 12447 continue; 12448 bp = newblk->nb_bmsafemap->sm_buf; 12449 bp = getdirtybuf(bp, &lk, waitfor); 12450 if (bp == NULL) { 12451 if (waitfor == MNT_NOWAIT) 12452 continue; 12453 return (1); 12454 } 12455 FREE_LOCK(&lk); 12456 if (waitfor == MNT_NOWAIT) 12457 bawrite(bp); 12458 else 12459 *errorp = bwrite(bp); 12460 ACQUIRE_LOCK(&lk); 12461 return (1); 12462 } 12463 return (0); 12464} 12465 12466/* 12467 * Flush dependencies associated with an allocdirect block. 12468 */ 12469static int 12470flush_newblk_dep(vp, mp, lbn) 12471 struct vnode *vp; 12472 struct mount *mp; 12473 ufs_lbn_t lbn; 12474{ 12475 struct newblk *newblk; 12476 struct bufobj *bo; 12477 struct inode *ip; 12478 struct buf *bp; 12479 ufs2_daddr_t blkno; 12480 int error; 12481 12482 error = 0; 12483 bo = &vp->v_bufobj; 12484 ip = VTOI(vp); 12485 blkno = DIP(ip, i_db[lbn]); 12486 if (blkno == 0) 12487 panic("flush_newblk_dep: Missing block"); 12488 ACQUIRE_LOCK(&lk); 12489 /* 12490 * Loop until all dependencies related to this block are satisfied. 12491 * We must be careful to restart after each sleep in case a write 12492 * completes some part of this process for us. 12493 */ 12494 for (;;) { 12495 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12496 FREE_LOCK(&lk); 12497 break; 12498 } 12499 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12500 panic("flush_newblk_deps: Bad newblk %p", newblk); 12501 /* 12502 * Flush the journal. 12503 */ 12504 if (newblk->nb_jnewblk != NULL) { 12505 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12506 continue; 12507 } 12508 /* 12509 * Write the bitmap dependency. 12510 */ 12511 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12512 bp = newblk->nb_bmsafemap->sm_buf; 12513 bp = getdirtybuf(bp, &lk, MNT_WAIT); 12514 if (bp == NULL) 12515 continue; 12516 FREE_LOCK(&lk); 12517 error = bwrite(bp); 12518 if (error) 12519 break; 12520 ACQUIRE_LOCK(&lk); 12521 continue; 12522 } 12523 /* 12524 * Write the buffer. 12525 */ 12526 FREE_LOCK(&lk); 12527 BO_LOCK(bo); 12528 bp = gbincore(bo, lbn); 12529 if (bp != NULL) { 12530 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12531 LK_INTERLOCK, BO_MTX(bo)); 12532 if (error == ENOLCK) { 12533 ACQUIRE_LOCK(&lk); 12534 continue; /* Slept, retry */ 12535 } 12536 if (error != 0) 12537 break; /* Failed */ 12538 if (bp->b_flags & B_DELWRI) { 12539 bremfree(bp); 12540 error = bwrite(bp); 12541 if (error) 12542 break; 12543 } else 12544 BUF_UNLOCK(bp); 12545 } else 12546 BO_UNLOCK(bo); 12547 /* 12548 * We have to wait for the direct pointers to 12549 * point at the newdirblk before the dependency 12550 * will go away. 12551 */ 12552 error = ffs_update(vp, 1); 12553 if (error) 12554 break; 12555 ACQUIRE_LOCK(&lk); 12556 } 12557 return (error); 12558} 12559 12560/* 12561 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 12562 * Called with splbio blocked. 12563 */ 12564static int 12565flush_pagedep_deps(pvp, mp, diraddhdp) 12566 struct vnode *pvp; 12567 struct mount *mp; 12568 struct diraddhd *diraddhdp; 12569{ 12570 struct inodedep *inodedep; 12571 struct inoref *inoref; 12572 struct ufsmount *ump; 12573 struct diradd *dap; 12574 struct vnode *vp; 12575 int error = 0; 12576 struct buf *bp; 12577 ino_t inum; 12578 12579 ump = VFSTOUFS(mp); 12580restart: 12581 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 12582 /* 12583 * Flush ourselves if this directory entry 12584 * has a MKDIR_PARENT dependency. 12585 */ 12586 if (dap->da_state & MKDIR_PARENT) { 12587 FREE_LOCK(&lk); 12588 if ((error = ffs_update(pvp, 1)) != 0) 12589 break; 12590 ACQUIRE_LOCK(&lk); 12591 /* 12592 * If that cleared dependencies, go on to next. 12593 */ 12594 if (dap != LIST_FIRST(diraddhdp)) 12595 continue; 12596 if (dap->da_state & MKDIR_PARENT) 12597 panic("flush_pagedep_deps: MKDIR_PARENT"); 12598 } 12599 /* 12600 * A newly allocated directory must have its "." and 12601 * ".." entries written out before its name can be 12602 * committed in its parent. 12603 */ 12604 inum = dap->da_newinum; 12605 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12606 panic("flush_pagedep_deps: lost inode1"); 12607 /* 12608 * Wait for any pending journal adds to complete so we don't 12609 * cause rollbacks while syncing. 12610 */ 12611 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12612 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12613 == DEPCOMPLETE) { 12614 jwait(&inoref->if_list, MNT_WAIT); 12615 goto restart; 12616 } 12617 } 12618 if (dap->da_state & MKDIR_BODY) { 12619 FREE_LOCK(&lk); 12620 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12621 FFSV_FORCEINSMQ))) 12622 break; 12623 error = flush_newblk_dep(vp, mp, 0); 12624 /* 12625 * If we still have the dependency we might need to 12626 * update the vnode to sync the new link count to 12627 * disk. 12628 */ 12629 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 12630 error = ffs_update(vp, 1); 12631 vput(vp); 12632 if (error != 0) 12633 break; 12634 ACQUIRE_LOCK(&lk); 12635 /* 12636 * If that cleared dependencies, go on to next. 12637 */ 12638 if (dap != LIST_FIRST(diraddhdp)) 12639 continue; 12640 if (dap->da_state & MKDIR_BODY) { 12641 inodedep_lookup(UFSTOVFS(ump), inum, 0, 12642 &inodedep); 12643 panic("flush_pagedep_deps: MKDIR_BODY " 12644 "inodedep %p dap %p vp %p", 12645 inodedep, dap, vp); 12646 } 12647 } 12648 /* 12649 * Flush the inode on which the directory entry depends. 12650 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 12651 * the only remaining dependency is that the updated inode 12652 * count must get pushed to disk. The inode has already 12653 * been pushed into its inode buffer (via VOP_UPDATE) at 12654 * the time of the reference count change. So we need only 12655 * locate that buffer, ensure that there will be no rollback 12656 * caused by a bitmap dependency, then write the inode buffer. 12657 */ 12658retry: 12659 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12660 panic("flush_pagedep_deps: lost inode"); 12661 /* 12662 * If the inode still has bitmap dependencies, 12663 * push them to disk. 12664 */ 12665 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 12666 bp = inodedep->id_bmsafemap->sm_buf; 12667 bp = getdirtybuf(bp, &lk, MNT_WAIT); 12668 if (bp == NULL) 12669 goto retry; 12670 FREE_LOCK(&lk); 12671 if ((error = bwrite(bp)) != 0) 12672 break; 12673 ACQUIRE_LOCK(&lk); 12674 if (dap != LIST_FIRST(diraddhdp)) 12675 continue; 12676 } 12677 /* 12678 * If the inode is still sitting in a buffer waiting 12679 * to be written or waiting for the link count to be 12680 * adjusted update it here to flush it to disk. 12681 */ 12682 if (dap == LIST_FIRST(diraddhdp)) { 12683 FREE_LOCK(&lk); 12684 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12685 FFSV_FORCEINSMQ))) 12686 break; 12687 error = ffs_update(vp, 1); 12688 vput(vp); 12689 if (error) 12690 break; 12691 ACQUIRE_LOCK(&lk); 12692 } 12693 /* 12694 * If we have failed to get rid of all the dependencies 12695 * then something is seriously wrong. 12696 */ 12697 if (dap == LIST_FIRST(diraddhdp)) { 12698 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 12699 panic("flush_pagedep_deps: failed to flush " 12700 "inodedep %p ino %ju dap %p", 12701 inodedep, (uintmax_t)inum, dap); 12702 } 12703 } 12704 if (error) 12705 ACQUIRE_LOCK(&lk); 12706 return (error); 12707} 12708 12709/* 12710 * A large burst of file addition or deletion activity can drive the 12711 * memory load excessively high. First attempt to slow things down 12712 * using the techniques below. If that fails, this routine requests 12713 * the offending operations to fall back to running synchronously 12714 * until the memory load returns to a reasonable level. 12715 */ 12716int 12717softdep_slowdown(vp) 12718 struct vnode *vp; 12719{ 12720 struct ufsmount *ump; 12721 int jlow; 12722 int max_softdeps_hard; 12723 12724 ACQUIRE_LOCK(&lk); 12725 jlow = 0; 12726 /* 12727 * Check for journal space if needed. 12728 */ 12729 if (DOINGSUJ(vp)) { 12730 ump = VFSTOUFS(vp->v_mount); 12731 if (journal_space(ump, 0) == 0) 12732 jlow = 1; 12733 } 12734 max_softdeps_hard = max_softdeps * 11 / 10; 12735 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 12736 dep_current[D_INODEDEP] < max_softdeps_hard && 12737 VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps && 12738 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0) { 12739 FREE_LOCK(&lk); 12740 return (0); 12741 } 12742 if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps || jlow) 12743 softdep_speedup(); 12744 stat_sync_limit_hit += 1; 12745 FREE_LOCK(&lk); 12746 if (DOINGSUJ(vp)) 12747 return (0); 12748 return (1); 12749} 12750 12751/* 12752 * Called by the allocation routines when they are about to fail 12753 * in the hope that we can free up the requested resource (inodes 12754 * or disk space). 12755 * 12756 * First check to see if the work list has anything on it. If it has, 12757 * clean up entries until we successfully free the requested resource. 12758 * Because this process holds inodes locked, we cannot handle any remove 12759 * requests that might block on a locked inode as that could lead to 12760 * deadlock. If the worklist yields none of the requested resource, 12761 * start syncing out vnodes to free up the needed space. 12762 */ 12763int 12764softdep_request_cleanup(fs, vp, cred, resource) 12765 struct fs *fs; 12766 struct vnode *vp; 12767 struct ucred *cred; 12768 int resource; 12769{ 12770 struct ufsmount *ump; 12771 struct mount *mp; 12772 struct vnode *lvp, *mvp; 12773 long starttime; 12774 ufs2_daddr_t needed; 12775 int error; 12776 12777 /* 12778 * If we are being called because of a process doing a 12779 * copy-on-write, then it is not safe to process any 12780 * worklist items as we will recurse into the copyonwrite 12781 * routine. This will result in an incoherent snapshot. 12782 * If the vnode that we hold is a snapshot, we must avoid 12783 * handling other resources that could cause deadlock. 12784 */ 12785 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 12786 return (0); 12787 12788 if (resource == FLUSH_BLOCKS_WAIT) 12789 stat_cleanup_blkrequests += 1; 12790 else 12791 stat_cleanup_inorequests += 1; 12792 12793 mp = vp->v_mount; 12794 ump = VFSTOUFS(mp); 12795 mtx_assert(UFS_MTX(ump), MA_OWNED); 12796 UFS_UNLOCK(ump); 12797 error = ffs_update(vp, 1); 12798 if (error != 0) { 12799 UFS_LOCK(ump); 12800 return (0); 12801 } 12802 /* 12803 * If we are in need of resources, consider pausing for 12804 * tickdelay to give ourselves some breathing room. 12805 */ 12806 ACQUIRE_LOCK(&lk); 12807 process_removes(vp); 12808 process_truncates(vp); 12809 request_cleanup(UFSTOVFS(ump), resource); 12810 FREE_LOCK(&lk); 12811 /* 12812 * Now clean up at least as many resources as we will need. 12813 * 12814 * When requested to clean up inodes, the number that are needed 12815 * is set by the number of simultaneous writers (mnt_writeopcount) 12816 * plus a bit of slop (2) in case some more writers show up while 12817 * we are cleaning. 12818 * 12819 * When requested to free up space, the amount of space that 12820 * we need is enough blocks to allocate a full-sized segment 12821 * (fs_contigsumsize). The number of such segments that will 12822 * be needed is set by the number of simultaneous writers 12823 * (mnt_writeopcount) plus a bit of slop (2) in case some more 12824 * writers show up while we are cleaning. 12825 * 12826 * Additionally, if we are unpriviledged and allocating space, 12827 * we need to ensure that we clean up enough blocks to get the 12828 * needed number of blocks over the threshhold of the minimum 12829 * number of blocks required to be kept free by the filesystem 12830 * (fs_minfree). 12831 */ 12832 if (resource == FLUSH_INODES_WAIT) { 12833 needed = vp->v_mount->mnt_writeopcount + 2; 12834 } else if (resource == FLUSH_BLOCKS_WAIT) { 12835 needed = (vp->v_mount->mnt_writeopcount + 2) * 12836 fs->fs_contigsumsize; 12837 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 12838 needed += fragstoblks(fs, 12839 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 12840 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 12841 } else { 12842 UFS_LOCK(ump); 12843 printf("softdep_request_cleanup: Unknown resource type %d\n", 12844 resource); 12845 return (0); 12846 } 12847 starttime = time_second; 12848retry: 12849 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 12850 fs->fs_cstotal.cs_nbfree <= needed) || 12851 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 12852 fs->fs_cstotal.cs_nifree <= needed)) { 12853 ACQUIRE_LOCK(&lk); 12854 if (ump->softdep_on_worklist > 0 && 12855 process_worklist_item(UFSTOVFS(ump), 12856 ump->softdep_on_worklist, LK_NOWAIT) != 0) 12857 stat_worklist_push += 1; 12858 FREE_LOCK(&lk); 12859 } 12860 /* 12861 * If we still need resources and there are no more worklist 12862 * entries to process to obtain them, we have to start flushing 12863 * the dirty vnodes to force the release of additional requests 12864 * to the worklist that we can then process to reap addition 12865 * resources. We walk the vnodes associated with the mount point 12866 * until we get the needed worklist requests that we can reap. 12867 */ 12868 if ((resource == FLUSH_BLOCKS_WAIT && 12869 fs->fs_cstotal.cs_nbfree <= needed) || 12870 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 12871 fs->fs_cstotal.cs_nifree <= needed)) { 12872 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 12873 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 12874 VI_UNLOCK(lvp); 12875 continue; 12876 } 12877 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 12878 curthread)) 12879 continue; 12880 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 12881 vput(lvp); 12882 continue; 12883 } 12884 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 12885 vput(lvp); 12886 } 12887 lvp = ump->um_devvp; 12888 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 12889 VOP_FSYNC(lvp, MNT_NOWAIT, curthread); 12890 VOP_UNLOCK(lvp, 0); 12891 } 12892 if (ump->softdep_on_worklist > 0) { 12893 stat_cleanup_retries += 1; 12894 goto retry; 12895 } 12896 stat_cleanup_failures += 1; 12897 } 12898 if (time_second - starttime > stat_cleanup_high_delay) 12899 stat_cleanup_high_delay = time_second - starttime; 12900 UFS_LOCK(ump); 12901 return (1); 12902} 12903 12904/* 12905 * If memory utilization has gotten too high, deliberately slow things 12906 * down and speed up the I/O processing. 12907 */ 12908extern struct thread *syncertd; 12909static int 12910request_cleanup(mp, resource) 12911 struct mount *mp; 12912 int resource; 12913{ 12914 struct thread *td = curthread; 12915 struct ufsmount *ump; 12916 12917 mtx_assert(&lk, MA_OWNED); 12918 /* 12919 * We never hold up the filesystem syncer or buf daemon. 12920 */ 12921 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 12922 return (0); 12923 ump = VFSTOUFS(mp); 12924 /* 12925 * First check to see if the work list has gotten backlogged. 12926 * If it has, co-opt this process to help clean up two entries. 12927 * Because this process may hold inodes locked, we cannot 12928 * handle any remove requests that might block on a locked 12929 * inode as that could lead to deadlock. We set TDP_SOFTDEP 12930 * to avoid recursively processing the worklist. 12931 */ 12932 if (ump->softdep_on_worklist > max_softdeps / 10) { 12933 td->td_pflags |= TDP_SOFTDEP; 12934 process_worklist_item(mp, 2, LK_NOWAIT); 12935 td->td_pflags &= ~TDP_SOFTDEP; 12936 stat_worklist_push += 2; 12937 return(1); 12938 } 12939 /* 12940 * Next, we attempt to speed up the syncer process. If that 12941 * is successful, then we allow the process to continue. 12942 */ 12943 if (softdep_speedup() && 12944 resource != FLUSH_BLOCKS_WAIT && 12945 resource != FLUSH_INODES_WAIT) 12946 return(0); 12947 /* 12948 * If we are resource constrained on inode dependencies, try 12949 * flushing some dirty inodes. Otherwise, we are constrained 12950 * by file deletions, so try accelerating flushes of directories 12951 * with removal dependencies. We would like to do the cleanup 12952 * here, but we probably hold an inode locked at this point and 12953 * that might deadlock against one that we try to clean. So, 12954 * the best that we can do is request the syncer daemon to do 12955 * the cleanup for us. 12956 */ 12957 switch (resource) { 12958 12959 case FLUSH_INODES: 12960 case FLUSH_INODES_WAIT: 12961 stat_ino_limit_push += 1; 12962 req_clear_inodedeps += 1; 12963 stat_countp = &stat_ino_limit_hit; 12964 break; 12965 12966 case FLUSH_BLOCKS: 12967 case FLUSH_BLOCKS_WAIT: 12968 stat_blk_limit_push += 1; 12969 req_clear_remove += 1; 12970 stat_countp = &stat_blk_limit_hit; 12971 break; 12972 12973 default: 12974 panic("request_cleanup: unknown type"); 12975 } 12976 /* 12977 * Hopefully the syncer daemon will catch up and awaken us. 12978 * We wait at most tickdelay before proceeding in any case. 12979 */ 12980 proc_waiting += 1; 12981 if (callout_pending(&softdep_callout) == FALSE) 12982 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 12983 pause_timer, 0); 12984 12985 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 12986 proc_waiting -= 1; 12987 return (1); 12988} 12989 12990/* 12991 * Awaken processes pausing in request_cleanup and clear proc_waiting 12992 * to indicate that there is no longer a timer running. 12993 */ 12994static void 12995pause_timer(arg) 12996 void *arg; 12997{ 12998 12999 /* 13000 * The callout_ API has acquired mtx and will hold it around this 13001 * function call. 13002 */ 13003 *stat_countp += 1; 13004 wakeup_one(&proc_waiting); 13005 if (proc_waiting > 0) 13006 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13007 pause_timer, 0); 13008} 13009 13010/* 13011 * Flush out a directory with at least one removal dependency in an effort to 13012 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13013 */ 13014static void 13015clear_remove(void) 13016{ 13017 struct pagedep_hashhead *pagedephd; 13018 struct pagedep *pagedep; 13019 static int next = 0; 13020 struct mount *mp; 13021 struct vnode *vp; 13022 struct bufobj *bo; 13023 int error, cnt; 13024 ino_t ino; 13025 13026 mtx_assert(&lk, MA_OWNED); 13027 13028 for (cnt = 0; cnt <= pagedep_hash; cnt++) { 13029 pagedephd = &pagedep_hashtbl[next++]; 13030 if (next > pagedep_hash) 13031 next = 0; 13032 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13033 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13034 continue; 13035 mp = pagedep->pd_list.wk_mp; 13036 ino = pagedep->pd_ino; 13037 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13038 continue; 13039 FREE_LOCK(&lk); 13040 13041 /* 13042 * Let unmount clear deps 13043 */ 13044 error = vfs_busy(mp, MBF_NOWAIT); 13045 if (error != 0) 13046 goto finish_write; 13047 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13048 FFSV_FORCEINSMQ); 13049 vfs_unbusy(mp); 13050 if (error != 0) { 13051 softdep_error("clear_remove: vget", error); 13052 goto finish_write; 13053 } 13054 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13055 softdep_error("clear_remove: fsync", error); 13056 bo = &vp->v_bufobj; 13057 BO_LOCK(bo); 13058 drain_output(vp); 13059 BO_UNLOCK(bo); 13060 vput(vp); 13061 finish_write: 13062 vn_finished_write(mp); 13063 ACQUIRE_LOCK(&lk); 13064 return; 13065 } 13066 } 13067} 13068 13069/* 13070 * Clear out a block of dirty inodes in an effort to reduce 13071 * the number of inodedep dependency structures. 13072 */ 13073static void 13074clear_inodedeps(void) 13075{ 13076 struct inodedep_hashhead *inodedephd; 13077 struct inodedep *inodedep; 13078 static int next = 0; 13079 struct mount *mp; 13080 struct vnode *vp; 13081 struct fs *fs; 13082 int error, cnt; 13083 ino_t firstino, lastino, ino; 13084 13085 mtx_assert(&lk, MA_OWNED); 13086 /* 13087 * Pick a random inode dependency to be cleared. 13088 * We will then gather up all the inodes in its block 13089 * that have dependencies and flush them out. 13090 */ 13091 for (cnt = 0; cnt <= inodedep_hash; cnt++) { 13092 inodedephd = &inodedep_hashtbl[next++]; 13093 if (next > inodedep_hash) 13094 next = 0; 13095 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13096 break; 13097 } 13098 if (inodedep == NULL) 13099 return; 13100 fs = inodedep->id_fs; 13101 mp = inodedep->id_list.wk_mp; 13102 /* 13103 * Find the last inode in the block with dependencies. 13104 */ 13105 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 13106 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13107 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13108 break; 13109 /* 13110 * Asynchronously push all but the last inode with dependencies. 13111 * Synchronously push the last inode with dependencies to ensure 13112 * that the inode block gets written to free up the inodedeps. 13113 */ 13114 for (ino = firstino; ino <= lastino; ino++) { 13115 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13116 continue; 13117 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13118 continue; 13119 FREE_LOCK(&lk); 13120 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13121 if (error != 0) { 13122 vn_finished_write(mp); 13123 ACQUIRE_LOCK(&lk); 13124 return; 13125 } 13126 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13127 FFSV_FORCEINSMQ)) != 0) { 13128 softdep_error("clear_inodedeps: vget", error); 13129 vfs_unbusy(mp); 13130 vn_finished_write(mp); 13131 ACQUIRE_LOCK(&lk); 13132 return; 13133 } 13134 vfs_unbusy(mp); 13135 if (ino == lastino) { 13136 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13137 softdep_error("clear_inodedeps: fsync1", error); 13138 } else { 13139 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13140 softdep_error("clear_inodedeps: fsync2", error); 13141 BO_LOCK(&vp->v_bufobj); 13142 drain_output(vp); 13143 BO_UNLOCK(&vp->v_bufobj); 13144 } 13145 vput(vp); 13146 vn_finished_write(mp); 13147 ACQUIRE_LOCK(&lk); 13148 } 13149} 13150 13151void 13152softdep_buf_append(bp, wkhd) 13153 struct buf *bp; 13154 struct workhead *wkhd; 13155{ 13156 struct worklist *wk; 13157 13158 ACQUIRE_LOCK(&lk); 13159 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13160 WORKLIST_REMOVE(wk); 13161 WORKLIST_INSERT(&bp->b_dep, wk); 13162 } 13163 FREE_LOCK(&lk); 13164 13165} 13166 13167void 13168softdep_inode_append(ip, cred, wkhd) 13169 struct inode *ip; 13170 struct ucred *cred; 13171 struct workhead *wkhd; 13172{ 13173 struct buf *bp; 13174 struct fs *fs; 13175 int error; 13176 13177 fs = ip->i_fs; 13178 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13179 (int)fs->fs_bsize, cred, &bp); 13180 if (error) { 13181 softdep_freework(wkhd); 13182 return; 13183 } 13184 softdep_buf_append(bp, wkhd); 13185 bqrelse(bp); 13186} 13187 13188void 13189softdep_freework(wkhd) 13190 struct workhead *wkhd; 13191{ 13192 13193 ACQUIRE_LOCK(&lk); 13194 handle_jwork(wkhd); 13195 FREE_LOCK(&lk); 13196} 13197 13198/* 13199 * Function to determine if the buffer has outstanding dependencies 13200 * that will cause a roll-back if the buffer is written. If wantcount 13201 * is set, return number of dependencies, otherwise just yes or no. 13202 */ 13203static int 13204softdep_count_dependencies(bp, wantcount) 13205 struct buf *bp; 13206 int wantcount; 13207{ 13208 struct worklist *wk; 13209 struct bmsafemap *bmsafemap; 13210 struct freework *freework; 13211 struct inodedep *inodedep; 13212 struct indirdep *indirdep; 13213 struct freeblks *freeblks; 13214 struct allocindir *aip; 13215 struct pagedep *pagedep; 13216 struct dirrem *dirrem; 13217 struct newblk *newblk; 13218 struct mkdir *mkdir; 13219 struct diradd *dap; 13220 int i, retval; 13221 13222 retval = 0; 13223 ACQUIRE_LOCK(&lk); 13224 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13225 switch (wk->wk_type) { 13226 13227 case D_INODEDEP: 13228 inodedep = WK_INODEDEP(wk); 13229 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13230 /* bitmap allocation dependency */ 13231 retval += 1; 13232 if (!wantcount) 13233 goto out; 13234 } 13235 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13236 /* direct block pointer dependency */ 13237 retval += 1; 13238 if (!wantcount) 13239 goto out; 13240 } 13241 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13242 /* direct block pointer dependency */ 13243 retval += 1; 13244 if (!wantcount) 13245 goto out; 13246 } 13247 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 13248 /* Add reference dependency. */ 13249 retval += 1; 13250 if (!wantcount) 13251 goto out; 13252 } 13253 continue; 13254 13255 case D_INDIRDEP: 13256 indirdep = WK_INDIRDEP(wk); 13257 13258 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 13259 /* indirect truncation dependency */ 13260 retval += 1; 13261 if (!wantcount) 13262 goto out; 13263 } 13264 13265 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 13266 /* indirect block pointer dependency */ 13267 retval += 1; 13268 if (!wantcount) 13269 goto out; 13270 } 13271 continue; 13272 13273 case D_PAGEDEP: 13274 pagedep = WK_PAGEDEP(wk); 13275 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 13276 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 13277 /* Journal remove ref dependency. */ 13278 retval += 1; 13279 if (!wantcount) 13280 goto out; 13281 } 13282 } 13283 for (i = 0; i < DAHASHSZ; i++) { 13284 13285 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 13286 /* directory entry dependency */ 13287 retval += 1; 13288 if (!wantcount) 13289 goto out; 13290 } 13291 } 13292 continue; 13293 13294 case D_BMSAFEMAP: 13295 bmsafemap = WK_BMSAFEMAP(wk); 13296 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 13297 /* Add reference dependency. */ 13298 retval += 1; 13299 if (!wantcount) 13300 goto out; 13301 } 13302 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 13303 /* Allocate block dependency. */ 13304 retval += 1; 13305 if (!wantcount) 13306 goto out; 13307 } 13308 continue; 13309 13310 case D_FREEBLKS: 13311 freeblks = WK_FREEBLKS(wk); 13312 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 13313 /* Freeblk journal dependency. */ 13314 retval += 1; 13315 if (!wantcount) 13316 goto out; 13317 } 13318 continue; 13319 13320 case D_ALLOCDIRECT: 13321 case D_ALLOCINDIR: 13322 newblk = WK_NEWBLK(wk); 13323 if (newblk->nb_jnewblk) { 13324 /* Journal allocate dependency. */ 13325 retval += 1; 13326 if (!wantcount) 13327 goto out; 13328 } 13329 continue; 13330 13331 case D_MKDIR: 13332 mkdir = WK_MKDIR(wk); 13333 if (mkdir->md_jaddref) { 13334 /* Journal reference dependency. */ 13335 retval += 1; 13336 if (!wantcount) 13337 goto out; 13338 } 13339 continue; 13340 13341 case D_FREEWORK: 13342 case D_FREEDEP: 13343 case D_JSEGDEP: 13344 case D_JSEG: 13345 case D_SBDEP: 13346 /* never a dependency on these blocks */ 13347 continue; 13348 13349 default: 13350 panic("softdep_count_dependencies: Unexpected type %s", 13351 TYPENAME(wk->wk_type)); 13352 /* NOTREACHED */ 13353 } 13354 } 13355out: 13356 FREE_LOCK(&lk); 13357 return retval; 13358} 13359 13360/* 13361 * Acquire exclusive access to a buffer. 13362 * Must be called with a locked mtx parameter. 13363 * Return acquired buffer or NULL on failure. 13364 */ 13365static struct buf * 13366getdirtybuf(bp, mtx, waitfor) 13367 struct buf *bp; 13368 struct mtx *mtx; 13369 int waitfor; 13370{ 13371 int error; 13372 13373 mtx_assert(mtx, MA_OWNED); 13374 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 13375 if (waitfor != MNT_WAIT) 13376 return (NULL); 13377 error = BUF_LOCK(bp, 13378 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx); 13379 /* 13380 * Even if we sucessfully acquire bp here, we have dropped 13381 * mtx, which may violates our guarantee. 13382 */ 13383 if (error == 0) 13384 BUF_UNLOCK(bp); 13385 else if (error != ENOLCK) 13386 panic("getdirtybuf: inconsistent lock: %d", error); 13387 mtx_lock(mtx); 13388 return (NULL); 13389 } 13390 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13391 if (mtx == &lk && waitfor == MNT_WAIT) { 13392 mtx_unlock(mtx); 13393 BO_LOCK(bp->b_bufobj); 13394 BUF_UNLOCK(bp); 13395 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13396 bp->b_vflags |= BV_BKGRDWAIT; 13397 msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj), 13398 PRIBIO | PDROP, "getbuf", 0); 13399 } else 13400 BO_UNLOCK(bp->b_bufobj); 13401 mtx_lock(mtx); 13402 return (NULL); 13403 } 13404 BUF_UNLOCK(bp); 13405 if (waitfor != MNT_WAIT) 13406 return (NULL); 13407 /* 13408 * The mtx argument must be bp->b_vp's mutex in 13409 * this case. 13410 */ 13411#ifdef DEBUG_VFS_LOCKS 13412 if (bp->b_vp->v_type != VCHR) 13413 ASSERT_BO_LOCKED(bp->b_bufobj); 13414#endif 13415 bp->b_vflags |= BV_BKGRDWAIT; 13416 msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0); 13417 return (NULL); 13418 } 13419 if ((bp->b_flags & B_DELWRI) == 0) { 13420 BUF_UNLOCK(bp); 13421 return (NULL); 13422 } 13423 bremfree(bp); 13424 return (bp); 13425} 13426 13427 13428/* 13429 * Check if it is safe to suspend the file system now. On entry, 13430 * the vnode interlock for devvp should be held. Return 0 with 13431 * the mount interlock held if the file system can be suspended now, 13432 * otherwise return EAGAIN with the mount interlock held. 13433 */ 13434int 13435softdep_check_suspend(struct mount *mp, 13436 struct vnode *devvp, 13437 int softdep_deps, 13438 int softdep_accdeps, 13439 int secondary_writes, 13440 int secondary_accwrites) 13441{ 13442 struct bufobj *bo; 13443 struct ufsmount *ump; 13444 int error; 13445 13446 ump = VFSTOUFS(mp); 13447 bo = &devvp->v_bufobj; 13448 ASSERT_BO_LOCKED(bo); 13449 13450 for (;;) { 13451 if (!TRY_ACQUIRE_LOCK(&lk)) { 13452 BO_UNLOCK(bo); 13453 ACQUIRE_LOCK(&lk); 13454 FREE_LOCK(&lk); 13455 BO_LOCK(bo); 13456 continue; 13457 } 13458 MNT_ILOCK(mp); 13459 if (mp->mnt_secondary_writes != 0) { 13460 FREE_LOCK(&lk); 13461 BO_UNLOCK(bo); 13462 msleep(&mp->mnt_secondary_writes, 13463 MNT_MTX(mp), 13464 (PUSER - 1) | PDROP, "secwr", 0); 13465 BO_LOCK(bo); 13466 continue; 13467 } 13468 break; 13469 } 13470 13471 /* 13472 * Reasons for needing more work before suspend: 13473 * - Dirty buffers on devvp. 13474 * - Softdep activity occurred after start of vnode sync loop 13475 * - Secondary writes occurred after start of vnode sync loop 13476 */ 13477 error = 0; 13478 if (bo->bo_numoutput > 0 || 13479 bo->bo_dirty.bv_cnt > 0 || 13480 softdep_deps != 0 || 13481 ump->softdep_deps != 0 || 13482 softdep_accdeps != ump->softdep_accdeps || 13483 secondary_writes != 0 || 13484 mp->mnt_secondary_writes != 0 || 13485 secondary_accwrites != mp->mnt_secondary_accwrites) 13486 error = EAGAIN; 13487 FREE_LOCK(&lk); 13488 BO_UNLOCK(bo); 13489 return (error); 13490} 13491 13492 13493/* 13494 * Get the number of dependency structures for the file system, both 13495 * the current number and the total number allocated. These will 13496 * later be used to detect that softdep processing has occurred. 13497 */ 13498void 13499softdep_get_depcounts(struct mount *mp, 13500 int *softdep_depsp, 13501 int *softdep_accdepsp) 13502{ 13503 struct ufsmount *ump; 13504 13505 ump = VFSTOUFS(mp); 13506 ACQUIRE_LOCK(&lk); 13507 *softdep_depsp = ump->softdep_deps; 13508 *softdep_accdepsp = ump->softdep_accdeps; 13509 FREE_LOCK(&lk); 13510} 13511 13512/* 13513 * Wait for pending output on a vnode to complete. 13514 * Must be called with vnode lock and interlock locked. 13515 * 13516 * XXX: Should just be a call to bufobj_wwait(). 13517 */ 13518static void 13519drain_output(vp) 13520 struct vnode *vp; 13521{ 13522 struct bufobj *bo; 13523 13524 bo = &vp->v_bufobj; 13525 ASSERT_VOP_LOCKED(vp, "drain_output"); 13526 ASSERT_BO_LOCKED(bo); 13527 13528 while (bo->bo_numoutput) { 13529 bo->bo_flag |= BO_WWAIT; 13530 msleep((caddr_t)&bo->bo_numoutput, 13531 BO_MTX(bo), PRIBIO + 1, "drainvp", 0); 13532 } 13533} 13534 13535/* 13536 * Called whenever a buffer that is being invalidated or reallocated 13537 * contains dependencies. This should only happen if an I/O error has 13538 * occurred. The routine is called with the buffer locked. 13539 */ 13540static void 13541softdep_deallocate_dependencies(bp) 13542 struct buf *bp; 13543{ 13544 13545 if ((bp->b_ioflags & BIO_ERROR) == 0) 13546 panic("softdep_deallocate_dependencies: dangling deps"); 13547 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 13548 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 13549 else 13550 printf("softdep_deallocate_dependencies: " 13551 "got error %d while accessing filesystem\n", bp->b_error); 13552 if (bp->b_error != ENXIO) 13553 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 13554} 13555 13556/* 13557 * Function to handle asynchronous write errors in the filesystem. 13558 */ 13559static void 13560softdep_error(func, error) 13561 char *func; 13562 int error; 13563{ 13564 13565 /* XXX should do something better! */ 13566 printf("%s: got error %d while accessing filesystem\n", func, error); 13567} 13568 13569#ifdef DDB 13570 13571static void 13572inodedep_print(struct inodedep *inodedep, int verbose) 13573{ 13574 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d" 13575 " saveino %p\n", 13576 inodedep, inodedep->id_fs, inodedep->id_state, 13577 (intmax_t)inodedep->id_ino, 13578 (intmax_t)fsbtodb(inodedep->id_fs, 13579 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 13580 inodedep->id_nlinkdelta, inodedep->id_savednlink, 13581 inodedep->id_savedino1); 13582 13583 if (verbose == 0) 13584 return; 13585 13586 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 13587 "mkdiradd %p\n", 13588 LIST_FIRST(&inodedep->id_pendinghd), 13589 LIST_FIRST(&inodedep->id_bufwait), 13590 LIST_FIRST(&inodedep->id_inowait), 13591 TAILQ_FIRST(&inodedep->id_inoreflst), 13592 inodedep->id_mkdiradd); 13593 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 13594 TAILQ_FIRST(&inodedep->id_inoupdt), 13595 TAILQ_FIRST(&inodedep->id_newinoupdt), 13596 TAILQ_FIRST(&inodedep->id_extupdt), 13597 TAILQ_FIRST(&inodedep->id_newextupdt)); 13598} 13599 13600DB_SHOW_COMMAND(inodedep, db_show_inodedep) 13601{ 13602 13603 if (have_addr == 0) { 13604 db_printf("Address required\n"); 13605 return; 13606 } 13607 inodedep_print((struct inodedep*)addr, 1); 13608} 13609 13610DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 13611{ 13612 struct inodedep_hashhead *inodedephd; 13613 struct inodedep *inodedep; 13614 struct fs *fs; 13615 int cnt; 13616 13617 fs = have_addr ? (struct fs *)addr : NULL; 13618 for (cnt = 0; cnt < inodedep_hash; cnt++) { 13619 inodedephd = &inodedep_hashtbl[cnt]; 13620 LIST_FOREACH(inodedep, inodedephd, id_hash) { 13621 if (fs != NULL && fs != inodedep->id_fs) 13622 continue; 13623 inodedep_print(inodedep, 0); 13624 } 13625 } 13626} 13627 13628DB_SHOW_COMMAND(worklist, db_show_worklist) 13629{ 13630 struct worklist *wk; 13631 13632 if (have_addr == 0) { 13633 db_printf("Address required\n"); 13634 return; 13635 } 13636 wk = (struct worklist *)addr; 13637 printf("worklist: %p type %s state 0x%X\n", 13638 wk, TYPENAME(wk->wk_type), wk->wk_state); 13639} 13640 13641DB_SHOW_COMMAND(workhead, db_show_workhead) 13642{ 13643 struct workhead *wkhd; 13644 struct worklist *wk; 13645 int i; 13646 13647 if (have_addr == 0) { 13648 db_printf("Address required\n"); 13649 return; 13650 } 13651 wkhd = (struct workhead *)addr; 13652 wk = LIST_FIRST(wkhd); 13653 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 13654 db_printf("worklist: %p type %s state 0x%X", 13655 wk, TYPENAME(wk->wk_type), wk->wk_state); 13656 if (i == 100) 13657 db_printf("workhead overflow"); 13658 printf("\n"); 13659} 13660 13661 13662DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 13663{ 13664 struct jaddref *jaddref; 13665 struct diradd *diradd; 13666 struct mkdir *mkdir; 13667 13668 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) { 13669 diradd = mkdir->md_diradd; 13670 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 13671 mkdir, mkdir->md_state, diradd, diradd->da_state); 13672 if ((jaddref = mkdir->md_jaddref) != NULL) 13673 db_printf(" jaddref %p jaddref state 0x%X", 13674 jaddref, jaddref->ja_state); 13675 db_printf("\n"); 13676 } 13677} 13678 13679#endif /* DDB */ 13680 13681#endif /* SOFTUPDATES */ 13682