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("$FreeBSD$"); 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/rwlock.h> 73 #include <sys/stat.h> 74 #include <sys/sysctl.h> 75 #include <sys/syslog.h> 76 #include <sys/vnode.h> 77 #include <sys/conf.h> 78 79 #include <ufs/ufs/dir.h> 80 #include <ufs/ufs/extattr.h> 81 #include <ufs/ufs/quota.h> 82 #include <ufs/ufs/inode.h> 83 #include <ufs/ufs/ufsmount.h> 84 #include <ufs/ffs/fs.h> 85 #include <ufs/ffs/softdep.h> 86 #include <ufs/ffs/ffs_extern.h> 87 #include <ufs/ufs/ufs_extern.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_extern.h> 91 #include <vm/vm_object.h> 92 93 #include <geom/geom.h> 94 95 #include <ddb/ddb.h> 96 97 #define KTR_SUJ 0 /* Define to KTR_SPARE. */ 98 99 #ifndef SOFTUPDATES 100 101 int 102 softdep_flushfiles(oldmnt, flags, td) 103 struct mount *oldmnt; 104 int flags; 105 struct thread *td; 106 { 107 108 panic("softdep_flushfiles called"); 109 } 110 111 int 112 softdep_mount(devvp, mp, fs, cred) 113 struct vnode *devvp; 114 struct mount *mp; 115 struct fs *fs; 116 struct ucred *cred; 117 { 118 119 return (0); 120 } 121 122 void 123 softdep_initialize() 124 { 125 126 return; 127 } 128 129 void 130 softdep_uninitialize() 131 { 132 133 return; 134 } 135 136 void 137 softdep_unmount(mp) 138 struct mount *mp; 139 { 140 141 panic("softdep_unmount called"); 142 } 143 144 void 145 softdep_setup_sbupdate(ump, fs, bp) 146 struct ufsmount *ump; 147 struct fs *fs; 148 struct buf *bp; 149 { 150 151 panic("softdep_setup_sbupdate called"); 152 } 153 154 void 155 softdep_setup_inomapdep(bp, ip, newinum, mode) 156 struct buf *bp; 157 struct inode *ip; 158 ino_t newinum; 159 int mode; 160 { 161 162 panic("softdep_setup_inomapdep called"); 163 } 164 165 void 166 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 167 struct buf *bp; 168 struct mount *mp; 169 ufs2_daddr_t newblkno; 170 int frags; 171 int oldfrags; 172 { 173 174 panic("softdep_setup_blkmapdep called"); 175 } 176 177 void 178 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 179 struct inode *ip; 180 ufs_lbn_t lbn; 181 ufs2_daddr_t newblkno; 182 ufs2_daddr_t oldblkno; 183 long newsize; 184 long oldsize; 185 struct buf *bp; 186 { 187 188 panic("softdep_setup_allocdirect called"); 189 } 190 191 void 192 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 193 struct inode *ip; 194 ufs_lbn_t lbn; 195 ufs2_daddr_t newblkno; 196 ufs2_daddr_t oldblkno; 197 long newsize; 198 long oldsize; 199 struct buf *bp; 200 { 201 202 panic("softdep_setup_allocext called"); 203 } 204 205 void 206 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 207 struct inode *ip; 208 ufs_lbn_t lbn; 209 struct buf *bp; 210 int ptrno; 211 ufs2_daddr_t newblkno; 212 ufs2_daddr_t oldblkno; 213 struct buf *nbp; 214 { 215 216 panic("softdep_setup_allocindir_page called"); 217 } 218 219 void 220 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 221 struct buf *nbp; 222 struct inode *ip; 223 struct buf *bp; 224 int ptrno; 225 ufs2_daddr_t newblkno; 226 { 227 228 panic("softdep_setup_allocindir_meta called"); 229 } 230 231 void 232 softdep_journal_freeblocks(ip, cred, length, flags) 233 struct inode *ip; 234 struct ucred *cred; 235 off_t length; 236 int flags; 237 { 238 239 panic("softdep_journal_freeblocks called"); 240 } 241 242 void 243 softdep_journal_fsync(ip) 244 struct inode *ip; 245 { 246 247 panic("softdep_journal_fsync called"); 248 } 249 250 void 251 softdep_setup_freeblocks(ip, length, flags) 252 struct inode *ip; 253 off_t length; 254 int flags; 255 { 256 257 panic("softdep_setup_freeblocks called"); 258 } 259 260 void 261 softdep_freefile(pvp, ino, mode) 262 struct vnode *pvp; 263 ino_t ino; 264 int mode; 265 { 266 267 panic("softdep_freefile called"); 268 } 269 270 int 271 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 272 struct buf *bp; 273 struct inode *dp; 274 off_t diroffset; 275 ino_t newinum; 276 struct buf *newdirbp; 277 int isnewblk; 278 { 279 280 panic("softdep_setup_directory_add called"); 281 } 282 283 void 284 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 285 struct buf *bp; 286 struct inode *dp; 287 caddr_t base; 288 caddr_t oldloc; 289 caddr_t newloc; 290 int entrysize; 291 { 292 293 panic("softdep_change_directoryentry_offset called"); 294 } 295 296 void 297 softdep_setup_remove(bp, dp, ip, isrmdir) 298 struct buf *bp; 299 struct inode *dp; 300 struct inode *ip; 301 int isrmdir; 302 { 303 304 panic("softdep_setup_remove called"); 305 } 306 307 void 308 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 309 struct buf *bp; 310 struct inode *dp; 311 struct inode *ip; 312 ino_t newinum; 313 int isrmdir; 314 { 315 316 panic("softdep_setup_directory_change called"); 317 } 318 319 void 320 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 321 struct mount *mp; 322 struct buf *bp; 323 ufs2_daddr_t blkno; 324 int frags; 325 struct workhead *wkhd; 326 { 327 328 panic("%s called", __FUNCTION__); 329 } 330 331 void 332 softdep_setup_inofree(mp, bp, ino, wkhd) 333 struct mount *mp; 334 struct buf *bp; 335 ino_t ino; 336 struct workhead *wkhd; 337 { 338 339 panic("%s called", __FUNCTION__); 340 } 341 342 void 343 softdep_setup_unlink(dp, ip) 344 struct inode *dp; 345 struct inode *ip; 346 { 347 348 panic("%s called", __FUNCTION__); 349 } 350 351 void 352 softdep_setup_link(dp, ip) 353 struct inode *dp; 354 struct inode *ip; 355 { 356 357 panic("%s called", __FUNCTION__); 358 } 359 360 void 361 softdep_revert_link(dp, ip) 362 struct inode *dp; 363 struct inode *ip; 364 { 365 366 panic("%s called", __FUNCTION__); 367 } 368 369 void 370 softdep_setup_rmdir(dp, ip) 371 struct inode *dp; 372 struct inode *ip; 373 { 374 375 panic("%s called", __FUNCTION__); 376 } 377 378 void 379 softdep_revert_rmdir(dp, ip) 380 struct inode *dp; 381 struct inode *ip; 382 { 383 384 panic("%s called", __FUNCTION__); 385 } 386 387 void 388 softdep_setup_create(dp, ip) 389 struct inode *dp; 390 struct inode *ip; 391 { 392 393 panic("%s called", __FUNCTION__); 394 } 395 396 void 397 softdep_revert_create(dp, ip) 398 struct inode *dp; 399 struct inode *ip; 400 { 401 402 panic("%s called", __FUNCTION__); 403 } 404 405 void 406 softdep_setup_mkdir(dp, ip) 407 struct inode *dp; 408 struct inode *ip; 409 { 410 411 panic("%s called", __FUNCTION__); 412 } 413 414 void 415 softdep_revert_mkdir(dp, ip) 416 struct inode *dp; 417 struct inode *ip; 418 { 419 420 panic("%s called", __FUNCTION__); 421 } 422 423 void 424 softdep_setup_dotdot_link(dp, ip) 425 struct inode *dp; 426 struct inode *ip; 427 { 428 429 panic("%s called", __FUNCTION__); 430 } 431 432 int 433 softdep_prealloc(vp, waitok) 434 struct vnode *vp; 435 int waitok; 436 { 437 438 panic("%s called", __FUNCTION__); 439 } 440 441 int 442 softdep_journal_lookup(mp, vpp) 443 struct mount *mp; 444 struct vnode **vpp; 445 { 446 447 return (ENOENT); 448 } 449 450 void 451 softdep_change_linkcnt(ip) 452 struct inode *ip; 453 { 454 455 panic("softdep_change_linkcnt called"); 456 } 457 458 void 459 softdep_load_inodeblock(ip) 460 struct inode *ip; 461 { 462 463 panic("softdep_load_inodeblock called"); 464 } 465 466 void 467 softdep_update_inodeblock(ip, bp, waitfor) 468 struct inode *ip; 469 struct buf *bp; 470 int waitfor; 471 { 472 473 panic("softdep_update_inodeblock called"); 474 } 475 476 int 477 softdep_fsync(vp) 478 struct vnode *vp; /* the "in_core" copy of the inode */ 479 { 480 481 return (0); 482 } 483 484 void 485 softdep_fsync_mountdev(vp) 486 struct vnode *vp; 487 { 488 489 return; 490 } 491 492 int 493 softdep_flushworklist(oldmnt, countp, td) 494 struct mount *oldmnt; 495 int *countp; 496 struct thread *td; 497 { 498 499 *countp = 0; 500 return (0); 501 } 502 503 int 504 softdep_sync_metadata(struct vnode *vp) 505 { 506 507 panic("softdep_sync_metadata called"); 508 } 509 510 int 511 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 512 { 513 514 panic("softdep_sync_buf called"); 515 } 516 517 int 518 softdep_slowdown(vp) 519 struct vnode *vp; 520 { 521 522 panic("softdep_slowdown called"); 523 } 524 525 int 526 softdep_request_cleanup(fs, vp, cred, resource) 527 struct fs *fs; 528 struct vnode *vp; 529 struct ucred *cred; 530 int resource; 531 { 532 533 return (0); 534 } 535 536 int 537 softdep_check_suspend(struct mount *mp, 538 struct vnode *devvp, 539 int softdep_depcnt, 540 int softdep_accdepcnt, 541 int secondary_writes, 542 int secondary_accwrites) 543 { 544 struct bufobj *bo; 545 int error; 546 547 (void) softdep_depcnt, 548 (void) softdep_accdepcnt; 549 550 bo = &devvp->v_bufobj; 551 ASSERT_BO_WLOCKED(bo); 552 553 MNT_ILOCK(mp); 554 while (mp->mnt_secondary_writes != 0) { 555 BO_UNLOCK(bo); 556 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 557 (PUSER - 1) | PDROP, "secwr", 0); 558 BO_LOCK(bo); 559 MNT_ILOCK(mp); 560 } 561 562 /* 563 * Reasons for needing more work before suspend: 564 * - Dirty buffers on devvp. 565 * - Secondary writes occurred after start of vnode sync loop 566 */ 567 error = 0; 568 if (bo->bo_numoutput > 0 || 569 bo->bo_dirty.bv_cnt > 0 || 570 secondary_writes != 0 || 571 mp->mnt_secondary_writes != 0 || 572 secondary_accwrites != mp->mnt_secondary_accwrites) 573 error = EAGAIN; 574 BO_UNLOCK(bo); 575 return (error); 576 } 577 578 void 579 softdep_get_depcounts(struct mount *mp, 580 int *softdepactivep, 581 int *softdepactiveaccp) 582 { 583 (void) mp; 584 *softdepactivep = 0; 585 *softdepactiveaccp = 0; 586 } 587 588 void 589 softdep_buf_append(bp, wkhd) 590 struct buf *bp; 591 struct workhead *wkhd; 592 { 593 594 panic("softdep_buf_appendwork called"); 595 } 596 597 void 598 softdep_inode_append(ip, cred, wkhd) 599 struct inode *ip; 600 struct ucred *cred; 601 struct workhead *wkhd; 602 { 603 604 panic("softdep_inode_appendwork called"); 605 } 606 607 void 608 softdep_freework(wkhd) 609 struct workhead *wkhd; 610 { 611 612 panic("softdep_freework called"); 613 } 614 615 #else 616 617 FEATURE(softupdates, "FFS soft-updates support"); 618 619 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 620 "soft updates stats"); 621 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 622 "total dependencies allocated"); 623 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 624 "high use dependencies allocated"); 625 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 626 "current dependencies allocated"); 627 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 628 "current dependencies written"); 629 630 unsigned long dep_current[D_LAST + 1]; 631 unsigned long dep_highuse[D_LAST + 1]; 632 unsigned long dep_total[D_LAST + 1]; 633 unsigned long dep_write[D_LAST + 1]; 634 635 #define SOFTDEP_TYPE(type, str, long) \ 636 static MALLOC_DEFINE(M_ ## type, #str, long); \ 637 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 638 &dep_total[D_ ## type], 0, ""); \ 639 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 640 &dep_current[D_ ## type], 0, ""); \ 641 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 642 &dep_highuse[D_ ## type], 0, ""); \ 643 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 644 &dep_write[D_ ## type], 0, ""); 645 646 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 647 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 648 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 649 "Block or frag allocated from cyl group map"); 650 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 651 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 652 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 653 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 654 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 655 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 656 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 657 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 658 SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 659 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 660 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 661 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 662 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 663 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 664 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 665 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 666 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 667 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 668 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 669 SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 670 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 671 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 672 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 673 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 674 675 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 676 677 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 678 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 679 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 680 681 #define M_SOFTDEP_FLAGS (M_WAITOK) 682 683 /* 684 * translate from workitem type to memory type 685 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 686 */ 687 static struct malloc_type *memtype[] = { 688 M_PAGEDEP, 689 M_INODEDEP, 690 M_BMSAFEMAP, 691 M_NEWBLK, 692 M_ALLOCDIRECT, 693 M_INDIRDEP, 694 M_ALLOCINDIR, 695 M_FREEFRAG, 696 M_FREEBLKS, 697 M_FREEFILE, 698 M_DIRADD, 699 M_MKDIR, 700 M_DIRREM, 701 M_NEWDIRBLK, 702 M_FREEWORK, 703 M_FREEDEP, 704 M_JADDREF, 705 M_JREMREF, 706 M_JMVREF, 707 M_JNEWBLK, 708 M_JFREEBLK, 709 M_JFREEFRAG, 710 M_JSEG, 711 M_JSEGDEP, 712 M_SBDEP, 713 M_JTRUNC, 714 M_JFSYNC, 715 M_SENTINEL 716 }; 717 718 #define DtoM(type) (memtype[type]) 719 720 /* 721 * Names of malloc types. 722 */ 723 #define TYPENAME(type) \ 724 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 725 /* 726 * End system adaptation definitions. 727 */ 728 729 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 730 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 731 732 /* 733 * Internal function prototypes. 734 */ 735 static void check_clear_deps(struct mount *); 736 static void softdep_error(char *, int); 737 static int softdep_process_worklist(struct mount *, int); 738 static int softdep_waitidle(struct mount *, int); 739 static void drain_output(struct vnode *); 740 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 741 static int check_inodedep_free(struct inodedep *); 742 static void clear_remove(struct mount *); 743 static void clear_inodedeps(struct mount *); 744 static void unlinked_inodedep(struct mount *, struct inodedep *); 745 static void clear_unlinked_inodedep(struct inodedep *); 746 static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 747 static int flush_pagedep_deps(struct vnode *, struct mount *, 748 struct diraddhd *); 749 static int free_pagedep(struct pagedep *); 750 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 751 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 752 static int flush_deplist(struct allocdirectlst *, int, int *); 753 static int sync_cgs(struct mount *, int); 754 static int handle_written_filepage(struct pagedep *, struct buf *); 755 static int handle_written_sbdep(struct sbdep *, struct buf *); 756 static void initiate_write_sbdep(struct sbdep *); 757 static void diradd_inode_written(struct diradd *, struct inodedep *); 758 static int handle_written_indirdep(struct indirdep *, struct buf *, 759 struct buf**); 760 static int handle_written_inodeblock(struct inodedep *, struct buf *); 761 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 762 uint8_t *); 763 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *); 764 static void handle_written_jaddref(struct jaddref *); 765 static void handle_written_jremref(struct jremref *); 766 static void handle_written_jseg(struct jseg *, struct buf *); 767 static void handle_written_jnewblk(struct jnewblk *); 768 static void handle_written_jblkdep(struct jblkdep *); 769 static void handle_written_jfreefrag(struct jfreefrag *); 770 static void complete_jseg(struct jseg *); 771 static void complete_jsegs(struct jseg *); 772 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 773 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 774 static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 775 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 776 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 777 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 778 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 779 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 780 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 781 static inline void inoref_write(struct inoref *, struct jseg *, 782 struct jrefrec *); 783 static void handle_allocdirect_partdone(struct allocdirect *, 784 struct workhead *); 785 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 786 struct workhead *); 787 static void indirdep_complete(struct indirdep *); 788 static int indirblk_lookup(struct mount *, ufs2_daddr_t); 789 static void indirblk_insert(struct freework *); 790 static void indirblk_remove(struct freework *); 791 static void handle_allocindir_partdone(struct allocindir *); 792 static void initiate_write_filepage(struct pagedep *, struct buf *); 793 static void initiate_write_indirdep(struct indirdep*, struct buf *); 794 static void handle_written_mkdir(struct mkdir *, int); 795 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 796 uint8_t *); 797 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 798 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 799 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 800 static void handle_workitem_freefile(struct freefile *); 801 static int handle_workitem_remove(struct dirrem *, int); 802 static struct dirrem *newdirrem(struct buf *, struct inode *, 803 struct inode *, int, struct dirrem **); 804 static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 805 struct buf *); 806 static void cancel_indirdep(struct indirdep *, struct buf *, 807 struct freeblks *); 808 static void free_indirdep(struct indirdep *); 809 static void free_diradd(struct diradd *, struct workhead *); 810 static void merge_diradd(struct inodedep *, struct diradd *); 811 static void complete_diradd(struct diradd *); 812 static struct diradd *diradd_lookup(struct pagedep *, int); 813 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 814 struct jremref *); 815 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 816 struct jremref *); 817 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 818 struct jremref *, struct jremref *); 819 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 820 struct jremref *); 821 static void cancel_allocindir(struct allocindir *, struct buf *bp, 822 struct freeblks *, int); 823 static int setup_trunc_indir(struct freeblks *, struct inode *, 824 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 825 static void complete_trunc_indir(struct freework *); 826 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 827 int); 828 static void complete_mkdir(struct mkdir *); 829 static void free_newdirblk(struct newdirblk *); 830 static void free_jremref(struct jremref *); 831 static void free_jaddref(struct jaddref *); 832 static void free_jsegdep(struct jsegdep *); 833 static void free_jsegs(struct jblocks *); 834 static void rele_jseg(struct jseg *); 835 static void free_jseg(struct jseg *, struct jblocks *); 836 static void free_jnewblk(struct jnewblk *); 837 static void free_jblkdep(struct jblkdep *); 838 static void free_jfreefrag(struct jfreefrag *); 839 static void free_freedep(struct freedep *); 840 static void journal_jremref(struct dirrem *, struct jremref *, 841 struct inodedep *); 842 static void cancel_jnewblk(struct jnewblk *, struct workhead *); 843 static int cancel_jaddref(struct jaddref *, struct inodedep *, 844 struct workhead *); 845 static void cancel_jfreefrag(struct jfreefrag *); 846 static inline void setup_freedirect(struct freeblks *, struct inode *, 847 int, int); 848 static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 849 static inline void setup_freeindir(struct freeblks *, struct inode *, int, 850 ufs_lbn_t, int); 851 static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 852 static void freeblks_free(struct ufsmount *, struct freeblks *, int); 853 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 854 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 855 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 856 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 857 int, int); 858 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 859 static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 860 static int deallocate_dependencies(struct buf *, struct freeblks *, int); 861 static void newblk_freefrag(struct newblk*); 862 static void free_newblk(struct newblk *); 863 static void cancel_allocdirect(struct allocdirectlst *, 864 struct allocdirect *, struct freeblks *); 865 static int check_inode_unwritten(struct inodedep *); 866 static int free_inodedep(struct inodedep *); 867 static void freework_freeblock(struct freework *); 868 static void freework_enqueue(struct freework *); 869 static int handle_workitem_freeblocks(struct freeblks *, int); 870 static int handle_complete_freeblocks(struct freeblks *, int); 871 static void handle_workitem_indirblk(struct freework *); 872 static void handle_written_freework(struct freework *); 873 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 874 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 875 struct workhead *); 876 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 877 struct inodedep *, struct allocindir *, ufs_lbn_t); 878 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 879 ufs2_daddr_t, ufs_lbn_t); 880 static void handle_workitem_freefrag(struct freefrag *); 881 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 882 ufs_lbn_t); 883 static void allocdirect_merge(struct allocdirectlst *, 884 struct allocdirect *, struct allocdirect *); 885 static struct freefrag *allocindir_merge(struct allocindir *, 886 struct allocindir *); 887 static int bmsafemap_find(struct bmsafemap_hashhead *, int, 888 struct bmsafemap **); 889 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 890 int cg, struct bmsafemap *); 891 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 892 struct newblk **); 893 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 894 static int inodedep_find(struct inodedep_hashhead *, ino_t, 895 struct inodedep **); 896 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 897 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 898 int, struct pagedep **); 899 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 900 struct pagedep **); 901 static void pause_timer(void *); 902 static int request_cleanup(struct mount *, int); 903 static void schedule_cleanup(struct mount *); 904 static void softdep_ast_cleanup_proc(void); 905 static int process_worklist_item(struct mount *, int, int); 906 static void process_removes(struct vnode *); 907 static void process_truncates(struct vnode *); 908 static void jwork_move(struct workhead *, struct workhead *); 909 static void jwork_insert(struct workhead *, struct jsegdep *); 910 static void add_to_worklist(struct worklist *, int); 911 static void wake_worklist(struct worklist *); 912 static void wait_worklist(struct worklist *, char *); 913 static void remove_from_worklist(struct worklist *); 914 static void softdep_flush(void *); 915 static void softdep_flushjournal(struct mount *); 916 static int softdep_speedup(struct ufsmount *); 917 static void worklist_speedup(struct mount *); 918 static int journal_mount(struct mount *, struct fs *, struct ucred *); 919 static void journal_unmount(struct ufsmount *); 920 static int journal_space(struct ufsmount *, int); 921 static void journal_suspend(struct ufsmount *); 922 static int journal_unsuspend(struct ufsmount *ump); 923 static void softdep_prelink(struct vnode *, struct vnode *); 924 static void add_to_journal(struct worklist *); 925 static void remove_from_journal(struct worklist *); 926 static bool softdep_excess_items(struct ufsmount *, int); 927 static void softdep_process_journal(struct mount *, struct worklist *, int); 928 static struct jremref *newjremref(struct dirrem *, struct inode *, 929 struct inode *ip, off_t, nlink_t); 930 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 931 uint16_t); 932 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 933 uint16_t); 934 static inline struct jsegdep *inoref_jseg(struct inoref *); 935 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 936 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 937 ufs2_daddr_t, int); 938 static void adjust_newfreework(struct freeblks *, int); 939 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 940 static void move_newblock_dep(struct jaddref *, struct inodedep *); 941 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 942 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 943 ufs2_daddr_t, long, ufs_lbn_t); 944 static struct freework *newfreework(struct ufsmount *, struct freeblks *, 945 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 946 static int jwait(struct worklist *, int); 947 static struct inodedep *inodedep_lookup_ip(struct inode *); 948 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 949 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 950 static void handle_jwork(struct workhead *); 951 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 952 struct mkdir **); 953 static struct jblocks *jblocks_create(void); 954 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 955 static void jblocks_free(struct jblocks *, struct mount *, int); 956 static void jblocks_destroy(struct jblocks *); 957 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 958 959 /* 960 * Exported softdep operations. 961 */ 962 static void softdep_disk_io_initiation(struct buf *); 963 static void softdep_disk_write_complete(struct buf *); 964 static void softdep_deallocate_dependencies(struct buf *); 965 static int softdep_count_dependencies(struct buf *bp, int); 966 967 /* 968 * Global lock over all of soft updates. 969 */ 970 static struct mtx lk; 971 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF); 972 973 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 974 #define FREE_GBLLOCK(lk) mtx_unlock(lk) 975 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 976 977 /* 978 * Per-filesystem soft-updates locking. 979 */ 980 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 981 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 982 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 983 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 984 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 985 RA_WLOCKED) 986 987 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 988 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 989 990 /* 991 * Worklist queue management. 992 * These routines require that the lock be held. 993 */ 994 #ifndef /* NOT */ DEBUG 995 #define WORKLIST_INSERT(head, item) do { \ 996 (item)->wk_state |= ONWORKLIST; \ 997 LIST_INSERT_HEAD(head, item, wk_list); \ 998 } while (0) 999 #define WORKLIST_REMOVE(item) do { \ 1000 (item)->wk_state &= ~ONWORKLIST; \ 1001 LIST_REMOVE(item, wk_list); \ 1002 } while (0) 1003 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1004 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1005 1006 #else /* DEBUG */ 1007 static void worklist_insert(struct workhead *, struct worklist *, int); 1008 static void worklist_remove(struct worklist *, int); 1009 1010 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1011 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1012 #define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1013 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1014 1015 static void 1016 worklist_insert(head, item, locked) 1017 struct workhead *head; 1018 struct worklist *item; 1019 int locked; 1020 { 1021 1022 if (locked) 1023 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1024 if (item->wk_state & ONWORKLIST) 1025 panic("worklist_insert: %p %s(0x%X) already on list", 1026 item, TYPENAME(item->wk_type), item->wk_state); 1027 item->wk_state |= ONWORKLIST; 1028 LIST_INSERT_HEAD(head, item, wk_list); 1029 } 1030 1031 static void 1032 worklist_remove(item, locked) 1033 struct worklist *item; 1034 int locked; 1035 { 1036 1037 if (locked) 1038 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1039 if ((item->wk_state & ONWORKLIST) == 0) 1040 panic("worklist_remove: %p %s(0x%X) not on list", 1041 item, TYPENAME(item->wk_type), item->wk_state); 1042 item->wk_state &= ~ONWORKLIST; 1043 LIST_REMOVE(item, wk_list); 1044 } 1045 #endif /* DEBUG */ 1046 1047 /* 1048 * Merge two jsegdeps keeping only the oldest one as newer references 1049 * can't be discarded until after older references. 1050 */ 1051 static inline struct jsegdep * 1052 jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1053 { 1054 struct jsegdep *swp; 1055 1056 if (two == NULL) 1057 return (one); 1058 1059 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1060 swp = one; 1061 one = two; 1062 two = swp; 1063 } 1064 WORKLIST_REMOVE(&two->jd_list); 1065 free_jsegdep(two); 1066 1067 return (one); 1068 } 1069 1070 /* 1071 * If two freedeps are compatible free one to reduce list size. 1072 */ 1073 static inline struct freedep * 1074 freedep_merge(struct freedep *one, struct freedep *two) 1075 { 1076 if (two == NULL) 1077 return (one); 1078 1079 if (one->fd_freework == two->fd_freework) { 1080 WORKLIST_REMOVE(&two->fd_list); 1081 free_freedep(two); 1082 } 1083 return (one); 1084 } 1085 1086 /* 1087 * Move journal work from one list to another. Duplicate freedeps and 1088 * jsegdeps are coalesced to keep the lists as small as possible. 1089 */ 1090 static void 1091 jwork_move(dst, src) 1092 struct workhead *dst; 1093 struct workhead *src; 1094 { 1095 struct freedep *freedep; 1096 struct jsegdep *jsegdep; 1097 struct worklist *wkn; 1098 struct worklist *wk; 1099 1100 KASSERT(dst != src, 1101 ("jwork_move: dst == src")); 1102 freedep = NULL; 1103 jsegdep = NULL; 1104 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1105 if (wk->wk_type == D_JSEGDEP) 1106 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1107 else if (wk->wk_type == D_FREEDEP) 1108 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1109 } 1110 1111 while ((wk = LIST_FIRST(src)) != NULL) { 1112 WORKLIST_REMOVE(wk); 1113 WORKLIST_INSERT(dst, wk); 1114 if (wk->wk_type == D_JSEGDEP) { 1115 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1116 continue; 1117 } 1118 if (wk->wk_type == D_FREEDEP) 1119 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1120 } 1121 } 1122 1123 static void 1124 jwork_insert(dst, jsegdep) 1125 struct workhead *dst; 1126 struct jsegdep *jsegdep; 1127 { 1128 struct jsegdep *jsegdepn; 1129 struct worklist *wk; 1130 1131 LIST_FOREACH(wk, dst, wk_list) 1132 if (wk->wk_type == D_JSEGDEP) 1133 break; 1134 if (wk == NULL) { 1135 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1136 return; 1137 } 1138 jsegdepn = WK_JSEGDEP(wk); 1139 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1140 WORKLIST_REMOVE(wk); 1141 free_jsegdep(jsegdepn); 1142 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1143 } else 1144 free_jsegdep(jsegdep); 1145 } 1146 1147 /* 1148 * Routines for tracking and managing workitems. 1149 */ 1150 static void workitem_free(struct worklist *, int); 1151 static void workitem_alloc(struct worklist *, int, struct mount *); 1152 static void workitem_reassign(struct worklist *, int); 1153 1154 #define WORKITEM_FREE(item, type) \ 1155 workitem_free((struct worklist *)(item), (type)) 1156 #define WORKITEM_REASSIGN(item, type) \ 1157 workitem_reassign((struct worklist *)(item), (type)) 1158 1159 static void 1160 workitem_free(item, type) 1161 struct worklist *item; 1162 int type; 1163 { 1164 struct ufsmount *ump; 1165 1166 #ifdef DEBUG 1167 if (item->wk_state & ONWORKLIST) 1168 panic("workitem_free: %s(0x%X) still on list", 1169 TYPENAME(item->wk_type), item->wk_state); 1170 if (item->wk_type != type && type != D_NEWBLK) 1171 panic("workitem_free: type mismatch %s != %s", 1172 TYPENAME(item->wk_type), TYPENAME(type)); 1173 #endif 1174 if (item->wk_state & IOWAITING) 1175 wakeup(item); 1176 ump = VFSTOUFS(item->wk_mp); 1177 LOCK_OWNED(ump); 1178 KASSERT(ump->softdep_deps > 0, 1179 ("workitem_free: %s: softdep_deps going negative", 1180 ump->um_fs->fs_fsmnt)); 1181 if (--ump->softdep_deps == 0 && ump->softdep_req) 1182 wakeup(&ump->softdep_deps); 1183 KASSERT(dep_current[item->wk_type] > 0, 1184 ("workitem_free: %s: dep_current[%s] going negative", 1185 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1186 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1187 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1188 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1189 atomic_subtract_long(&dep_current[item->wk_type], 1); 1190 ump->softdep_curdeps[item->wk_type] -= 1; 1191 free(item, DtoM(type)); 1192 } 1193 1194 static void 1195 workitem_alloc(item, type, mp) 1196 struct worklist *item; 1197 int type; 1198 struct mount *mp; 1199 { 1200 struct ufsmount *ump; 1201 1202 item->wk_type = type; 1203 item->wk_mp = mp; 1204 item->wk_state = 0; 1205 1206 ump = VFSTOUFS(mp); 1207 ACQUIRE_GBLLOCK(&lk); 1208 dep_current[type]++; 1209 if (dep_current[type] > dep_highuse[type]) 1210 dep_highuse[type] = dep_current[type]; 1211 dep_total[type]++; 1212 FREE_GBLLOCK(&lk); 1213 ACQUIRE_LOCK(ump); 1214 ump->softdep_curdeps[type] += 1; 1215 ump->softdep_deps++; 1216 ump->softdep_accdeps++; 1217 FREE_LOCK(ump); 1218 } 1219 1220 static void 1221 workitem_reassign(item, newtype) 1222 struct worklist *item; 1223 int newtype; 1224 { 1225 struct ufsmount *ump; 1226 1227 ump = VFSTOUFS(item->wk_mp); 1228 LOCK_OWNED(ump); 1229 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1230 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1231 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1232 ump->softdep_curdeps[item->wk_type] -= 1; 1233 ump->softdep_curdeps[newtype] += 1; 1234 KASSERT(dep_current[item->wk_type] > 0, 1235 ("workitem_reassign: %s: dep_current[%s] going negative", 1236 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1237 ACQUIRE_GBLLOCK(&lk); 1238 dep_current[newtype]++; 1239 dep_current[item->wk_type]--; 1240 if (dep_current[newtype] > dep_highuse[newtype]) 1241 dep_highuse[newtype] = dep_current[newtype]; 1242 dep_total[newtype]++; 1243 FREE_GBLLOCK(&lk); 1244 item->wk_type = newtype; 1245 } 1246 1247 /* 1248 * Workitem queue management 1249 */ 1250 static int max_softdeps; /* maximum number of structs before slowdown */ 1251 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1252 static int proc_waiting; /* tracks whether we have a timeout posted */ 1253 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1254 static struct callout softdep_callout; 1255 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1256 static int req_clear_remove; /* syncer process flush some freeblks */ 1257 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1258 1259 /* 1260 * runtime statistics 1261 */ 1262 static int stat_flush_threads; /* number of softdep flushing threads */ 1263 static int stat_worklist_push; /* number of worklist cleanups */ 1264 static int stat_blk_limit_push; /* number of times block limit neared */ 1265 static int stat_ino_limit_push; /* number of times inode limit neared */ 1266 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1267 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1268 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1269 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1270 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1271 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1272 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1273 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1274 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1275 static int stat_journal_min; /* Times hit journal min threshold */ 1276 static int stat_journal_low; /* Times hit journal low threshold */ 1277 static int stat_journal_wait; /* Times blocked in jwait(). */ 1278 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1279 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1280 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1281 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1282 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1283 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1284 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1285 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1286 static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1287 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1288 1289 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1290 &max_softdeps, 0, ""); 1291 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1292 &tickdelay, 0, ""); 1293 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1294 &stat_flush_threads, 0, ""); 1295 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1296 &stat_worklist_push, 0,""); 1297 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1298 &stat_blk_limit_push, 0,""); 1299 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1300 &stat_ino_limit_push, 0,""); 1301 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1302 &stat_blk_limit_hit, 0, ""); 1303 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1304 &stat_ino_limit_hit, 0, ""); 1305 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1306 &stat_sync_limit_hit, 0, ""); 1307 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1308 &stat_indir_blk_ptrs, 0, ""); 1309 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1310 &stat_inode_bitmap, 0, ""); 1311 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1312 &stat_direct_blk_ptrs, 0, ""); 1313 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1314 &stat_dir_entry, 0, ""); 1315 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1316 &stat_jaddref, 0, ""); 1317 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1318 &stat_jnewblk, 0, ""); 1319 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1320 &stat_journal_low, 0, ""); 1321 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1322 &stat_journal_min, 0, ""); 1323 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1324 &stat_journal_wait, 0, ""); 1325 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1326 &stat_jwait_filepage, 0, ""); 1327 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1328 &stat_jwait_freeblks, 0, ""); 1329 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1330 &stat_jwait_inode, 0, ""); 1331 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1332 &stat_jwait_newblk, 0, ""); 1333 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1334 &stat_cleanup_blkrequests, 0, ""); 1335 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1336 &stat_cleanup_inorequests, 0, ""); 1337 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1338 &stat_cleanup_high_delay, 0, ""); 1339 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1340 &stat_cleanup_retries, 0, ""); 1341 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1342 &stat_cleanup_failures, 0, ""); 1343 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1344 &softdep_flushcache, 0, ""); 1345 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1346 &stat_emptyjblocks, 0, ""); 1347 1348 SYSCTL_DECL(_vfs_ffs); 1349 1350 /* Whether to recompute the summary at mount time */ 1351 static int compute_summary_at_mount = 0; 1352 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1353 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1354 static int print_threads = 0; 1355 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1356 &print_threads, 0, "Notify flusher thread start/stop"); 1357 1358 /* List of all filesystems mounted with soft updates */ 1359 static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1360 1361 /* 1362 * This function cleans the worklist for a filesystem. 1363 * Each filesystem running with soft dependencies gets its own 1364 * thread to run in this function. The thread is started up in 1365 * softdep_mount and shutdown in softdep_unmount. They show up 1366 * as part of the kernel "bufdaemon" process whose process 1367 * entry is available in bufdaemonproc. 1368 */ 1369 static int searchfailed; 1370 extern struct proc *bufdaemonproc; 1371 static void 1372 softdep_flush(addr) 1373 void *addr; 1374 { 1375 struct mount *mp; 1376 struct thread *td; 1377 struct ufsmount *ump; 1378 1379 td = curthread; 1380 td->td_pflags |= TDP_NORUNNINGBUF; 1381 mp = (struct mount *)addr; 1382 ump = VFSTOUFS(mp); 1383 atomic_add_int(&stat_flush_threads, 1); 1384 ACQUIRE_LOCK(ump); 1385 ump->softdep_flags &= ~FLUSH_STARTING; 1386 wakeup(&ump->softdep_flushtd); 1387 FREE_LOCK(ump); 1388 if (print_threads) { 1389 if (stat_flush_threads == 1) 1390 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1391 bufdaemonproc->p_pid); 1392 printf("Start thread %s\n", td->td_name); 1393 } 1394 for (;;) { 1395 while (softdep_process_worklist(mp, 0) > 0 || 1396 (MOUNTEDSUJ(mp) && 1397 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1398 kthread_suspend_check(); 1399 ACQUIRE_LOCK(ump); 1400 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1401 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1402 "sdflush", hz / 2); 1403 ump->softdep_flags &= ~FLUSH_CLEANUP; 1404 /* 1405 * Check to see if we are done and need to exit. 1406 */ 1407 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1408 FREE_LOCK(ump); 1409 continue; 1410 } 1411 ump->softdep_flags &= ~FLUSH_EXIT; 1412 FREE_LOCK(ump); 1413 wakeup(&ump->softdep_flags); 1414 if (print_threads) 1415 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1416 atomic_subtract_int(&stat_flush_threads, 1); 1417 kthread_exit(); 1418 panic("kthread_exit failed\n"); 1419 } 1420 } 1421 1422 static void 1423 worklist_speedup(mp) 1424 struct mount *mp; 1425 { 1426 struct ufsmount *ump; 1427 1428 ump = VFSTOUFS(mp); 1429 LOCK_OWNED(ump); 1430 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1431 ump->softdep_flags |= FLUSH_CLEANUP; 1432 wakeup(&ump->softdep_flushtd); 1433 } 1434 1435 static int 1436 softdep_speedup(ump) 1437 struct ufsmount *ump; 1438 { 1439 struct ufsmount *altump; 1440 struct mount_softdeps *sdp; 1441 1442 LOCK_OWNED(ump); 1443 worklist_speedup(ump->um_mountp); 1444 bd_speedup(); 1445 /* 1446 * If we have global shortages, then we need other 1447 * filesystems to help with the cleanup. Here we wakeup a 1448 * flusher thread for a filesystem that is over its fair 1449 * share of resources. 1450 */ 1451 if (req_clear_inodedeps || req_clear_remove) { 1452 ACQUIRE_GBLLOCK(&lk); 1453 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1454 if ((altump = sdp->sd_ump) == ump) 1455 continue; 1456 if (((req_clear_inodedeps && 1457 altump->softdep_curdeps[D_INODEDEP] > 1458 max_softdeps / stat_flush_threads) || 1459 (req_clear_remove && 1460 altump->softdep_curdeps[D_DIRREM] > 1461 (max_softdeps / 2) / stat_flush_threads)) && 1462 TRY_ACQUIRE_LOCK(altump)) 1463 break; 1464 } 1465 if (sdp == NULL) { 1466 searchfailed++; 1467 FREE_GBLLOCK(&lk); 1468 } else { 1469 /* 1470 * Move to the end of the list so we pick a 1471 * different one on out next try. 1472 */ 1473 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1474 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1475 FREE_GBLLOCK(&lk); 1476 if ((altump->softdep_flags & 1477 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1478 altump->softdep_flags |= FLUSH_CLEANUP; 1479 altump->um_softdep->sd_cleanups++; 1480 wakeup(&altump->softdep_flushtd); 1481 FREE_LOCK(altump); 1482 } 1483 } 1484 return (speedup_syncer()); 1485 } 1486 1487 /* 1488 * Add an item to the end of the work queue. 1489 * This routine requires that the lock be held. 1490 * This is the only routine that adds items to the list. 1491 * The following routine is the only one that removes items 1492 * and does so in order from first to last. 1493 */ 1494 1495 #define WK_HEAD 0x0001 /* Add to HEAD. */ 1496 #define WK_NODELAY 0x0002 /* Process immediately. */ 1497 1498 static void 1499 add_to_worklist(wk, flags) 1500 struct worklist *wk; 1501 int flags; 1502 { 1503 struct ufsmount *ump; 1504 1505 ump = VFSTOUFS(wk->wk_mp); 1506 LOCK_OWNED(ump); 1507 if (wk->wk_state & ONWORKLIST) 1508 panic("add_to_worklist: %s(0x%X) already on list", 1509 TYPENAME(wk->wk_type), wk->wk_state); 1510 wk->wk_state |= ONWORKLIST; 1511 if (ump->softdep_on_worklist == 0) { 1512 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1513 ump->softdep_worklist_tail = wk; 1514 } else if (flags & WK_HEAD) { 1515 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1516 } else { 1517 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1518 ump->softdep_worklist_tail = wk; 1519 } 1520 ump->softdep_on_worklist += 1; 1521 if (flags & WK_NODELAY) 1522 worklist_speedup(wk->wk_mp); 1523 } 1524 1525 /* 1526 * Remove the item to be processed. If we are removing the last 1527 * item on the list, we need to recalculate the tail pointer. 1528 */ 1529 static void 1530 remove_from_worklist(wk) 1531 struct worklist *wk; 1532 { 1533 struct ufsmount *ump; 1534 1535 ump = VFSTOUFS(wk->wk_mp); 1536 WORKLIST_REMOVE(wk); 1537 if (ump->softdep_worklist_tail == wk) 1538 ump->softdep_worklist_tail = 1539 (struct worklist *)wk->wk_list.le_prev; 1540 ump->softdep_on_worklist -= 1; 1541 } 1542 1543 static void 1544 wake_worklist(wk) 1545 struct worklist *wk; 1546 { 1547 if (wk->wk_state & IOWAITING) { 1548 wk->wk_state &= ~IOWAITING; 1549 wakeup(wk); 1550 } 1551 } 1552 1553 static void 1554 wait_worklist(wk, wmesg) 1555 struct worklist *wk; 1556 char *wmesg; 1557 { 1558 struct ufsmount *ump; 1559 1560 ump = VFSTOUFS(wk->wk_mp); 1561 wk->wk_state |= IOWAITING; 1562 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1563 } 1564 1565 /* 1566 * Process that runs once per second to handle items in the background queue. 1567 * 1568 * Note that we ensure that everything is done in the order in which they 1569 * appear in the queue. The code below depends on this property to ensure 1570 * that blocks of a file are freed before the inode itself is freed. This 1571 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1572 * until all the old ones have been purged from the dependency lists. 1573 */ 1574 static int 1575 softdep_process_worklist(mp, full) 1576 struct mount *mp; 1577 int full; 1578 { 1579 int cnt, matchcnt; 1580 struct ufsmount *ump; 1581 long starttime; 1582 1583 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1584 if (MOUNTEDSOFTDEP(mp) == 0) 1585 return (0); 1586 matchcnt = 0; 1587 ump = VFSTOUFS(mp); 1588 ACQUIRE_LOCK(ump); 1589 starttime = time_second; 1590 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1591 check_clear_deps(mp); 1592 while (ump->softdep_on_worklist > 0) { 1593 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1594 break; 1595 else 1596 matchcnt += cnt; 1597 check_clear_deps(mp); 1598 /* 1599 * We do not generally want to stop for buffer space, but if 1600 * we are really being a buffer hog, we will stop and wait. 1601 */ 1602 if (should_yield()) { 1603 FREE_LOCK(ump); 1604 kern_yield(PRI_USER); 1605 bwillwrite(); 1606 ACQUIRE_LOCK(ump); 1607 } 1608 /* 1609 * Never allow processing to run for more than one 1610 * second. This gives the syncer thread the opportunity 1611 * to pause if appropriate. 1612 */ 1613 if (!full && starttime != time_second) 1614 break; 1615 } 1616 if (full == 0) 1617 journal_unsuspend(ump); 1618 FREE_LOCK(ump); 1619 return (matchcnt); 1620 } 1621 1622 /* 1623 * Process all removes associated with a vnode if we are running out of 1624 * journal space. Any other process which attempts to flush these will 1625 * be unable as we have the vnodes locked. 1626 */ 1627 static void 1628 process_removes(vp) 1629 struct vnode *vp; 1630 { 1631 struct inodedep *inodedep; 1632 struct dirrem *dirrem; 1633 struct ufsmount *ump; 1634 struct mount *mp; 1635 ino_t inum; 1636 1637 mp = vp->v_mount; 1638 ump = VFSTOUFS(mp); 1639 LOCK_OWNED(ump); 1640 inum = VTOI(vp)->i_number; 1641 for (;;) { 1642 top: 1643 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1644 return; 1645 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1646 /* 1647 * If another thread is trying to lock this vnode 1648 * it will fail but we must wait for it to do so 1649 * before we can proceed. 1650 */ 1651 if (dirrem->dm_state & INPROGRESS) { 1652 wait_worklist(&dirrem->dm_list, "pwrwait"); 1653 goto top; 1654 } 1655 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1656 (COMPLETE | ONWORKLIST)) 1657 break; 1658 } 1659 if (dirrem == NULL) 1660 return; 1661 remove_from_worklist(&dirrem->dm_list); 1662 FREE_LOCK(ump); 1663 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1664 panic("process_removes: suspended filesystem"); 1665 handle_workitem_remove(dirrem, 0); 1666 vn_finished_secondary_write(mp); 1667 ACQUIRE_LOCK(ump); 1668 } 1669 } 1670 1671 /* 1672 * Process all truncations associated with a vnode if we are running out 1673 * of journal space. This is called when the vnode lock is already held 1674 * and no other process can clear the truncation. This function returns 1675 * a value greater than zero if it did any work. 1676 */ 1677 static void 1678 process_truncates(vp) 1679 struct vnode *vp; 1680 { 1681 struct inodedep *inodedep; 1682 struct freeblks *freeblks; 1683 struct ufsmount *ump; 1684 struct mount *mp; 1685 ino_t inum; 1686 int cgwait; 1687 1688 mp = vp->v_mount; 1689 ump = VFSTOUFS(mp); 1690 LOCK_OWNED(ump); 1691 inum = VTOI(vp)->i_number; 1692 for (;;) { 1693 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1694 return; 1695 cgwait = 0; 1696 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1697 /* Journal entries not yet written. */ 1698 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1699 jwait(&LIST_FIRST( 1700 &freeblks->fb_jblkdephd)->jb_list, 1701 MNT_WAIT); 1702 break; 1703 } 1704 /* Another thread is executing this item. */ 1705 if (freeblks->fb_state & INPROGRESS) { 1706 wait_worklist(&freeblks->fb_list, "ptrwait"); 1707 break; 1708 } 1709 /* Freeblks is waiting on a inode write. */ 1710 if ((freeblks->fb_state & COMPLETE) == 0) { 1711 FREE_LOCK(ump); 1712 ffs_update(vp, 1); 1713 ACQUIRE_LOCK(ump); 1714 break; 1715 } 1716 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1717 (ALLCOMPLETE | ONWORKLIST)) { 1718 remove_from_worklist(&freeblks->fb_list); 1719 freeblks->fb_state |= INPROGRESS; 1720 FREE_LOCK(ump); 1721 if (vn_start_secondary_write(NULL, &mp, 1722 V_NOWAIT)) 1723 panic("process_truncates: " 1724 "suspended filesystem"); 1725 handle_workitem_freeblocks(freeblks, 0); 1726 vn_finished_secondary_write(mp); 1727 ACQUIRE_LOCK(ump); 1728 break; 1729 } 1730 if (freeblks->fb_cgwait) 1731 cgwait++; 1732 } 1733 if (cgwait) { 1734 FREE_LOCK(ump); 1735 sync_cgs(mp, MNT_WAIT); 1736 ffs_sync_snap(mp, MNT_WAIT); 1737 ACQUIRE_LOCK(ump); 1738 continue; 1739 } 1740 if (freeblks == NULL) 1741 break; 1742 } 1743 return; 1744 } 1745 1746 /* 1747 * Process one item on the worklist. 1748 */ 1749 static int 1750 process_worklist_item(mp, target, flags) 1751 struct mount *mp; 1752 int target; 1753 int flags; 1754 { 1755 struct worklist sentinel; 1756 struct worklist *wk; 1757 struct ufsmount *ump; 1758 int matchcnt; 1759 int error; 1760 1761 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1762 /* 1763 * If we are being called because of a process doing a 1764 * copy-on-write, then it is not safe to write as we may 1765 * recurse into the copy-on-write routine. 1766 */ 1767 if (curthread->td_pflags & TDP_COWINPROGRESS) 1768 return (-1); 1769 PHOLD(curproc); /* Don't let the stack go away. */ 1770 ump = VFSTOUFS(mp); 1771 LOCK_OWNED(ump); 1772 matchcnt = 0; 1773 sentinel.wk_mp = NULL; 1774 sentinel.wk_type = D_SENTINEL; 1775 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1776 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1777 wk = LIST_NEXT(&sentinel, wk_list)) { 1778 if (wk->wk_type == D_SENTINEL) { 1779 LIST_REMOVE(&sentinel, wk_list); 1780 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1781 continue; 1782 } 1783 if (wk->wk_state & INPROGRESS) 1784 panic("process_worklist_item: %p already in progress.", 1785 wk); 1786 wk->wk_state |= INPROGRESS; 1787 remove_from_worklist(wk); 1788 FREE_LOCK(ump); 1789 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1790 panic("process_worklist_item: suspended filesystem"); 1791 switch (wk->wk_type) { 1792 case D_DIRREM: 1793 /* removal of a directory entry */ 1794 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1795 break; 1796 1797 case D_FREEBLKS: 1798 /* releasing blocks and/or fragments from a file */ 1799 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1800 flags); 1801 break; 1802 1803 case D_FREEFRAG: 1804 /* releasing a fragment when replaced as a file grows */ 1805 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1806 error = 0; 1807 break; 1808 1809 case D_FREEFILE: 1810 /* releasing an inode when its link count drops to 0 */ 1811 handle_workitem_freefile(WK_FREEFILE(wk)); 1812 error = 0; 1813 break; 1814 1815 default: 1816 panic("%s_process_worklist: Unknown type %s", 1817 "softdep", TYPENAME(wk->wk_type)); 1818 /* NOTREACHED */ 1819 } 1820 vn_finished_secondary_write(mp); 1821 ACQUIRE_LOCK(ump); 1822 if (error == 0) { 1823 if (++matchcnt == target) 1824 break; 1825 continue; 1826 } 1827 /* 1828 * We have to retry the worklist item later. Wake up any 1829 * waiters who may be able to complete it immediately and 1830 * add the item back to the head so we don't try to execute 1831 * it again. 1832 */ 1833 wk->wk_state &= ~INPROGRESS; 1834 wake_worklist(wk); 1835 add_to_worklist(wk, WK_HEAD); 1836 } 1837 LIST_REMOVE(&sentinel, wk_list); 1838 /* Sentinal could've become the tail from remove_from_worklist. */ 1839 if (ump->softdep_worklist_tail == &sentinel) 1840 ump->softdep_worklist_tail = 1841 (struct worklist *)sentinel.wk_list.le_prev; 1842 PRELE(curproc); 1843 return (matchcnt); 1844 } 1845 1846 /* 1847 * Move dependencies from one buffer to another. 1848 */ 1849 int 1850 softdep_move_dependencies(oldbp, newbp) 1851 struct buf *oldbp; 1852 struct buf *newbp; 1853 { 1854 struct worklist *wk, *wktail; 1855 struct ufsmount *ump; 1856 int dirty; 1857 1858 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1859 return (0); 1860 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1861 ("softdep_move_dependencies called on non-softdep filesystem")); 1862 dirty = 0; 1863 wktail = NULL; 1864 ump = VFSTOUFS(wk->wk_mp); 1865 ACQUIRE_LOCK(ump); 1866 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1867 LIST_REMOVE(wk, wk_list); 1868 if (wk->wk_type == D_BMSAFEMAP && 1869 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1870 dirty = 1; 1871 if (wktail == 0) 1872 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1873 else 1874 LIST_INSERT_AFTER(wktail, wk, wk_list); 1875 wktail = wk; 1876 } 1877 FREE_LOCK(ump); 1878 1879 return (dirty); 1880 } 1881 1882 /* 1883 * Purge the work list of all items associated with a particular mount point. 1884 */ 1885 int 1886 softdep_flushworklist(oldmnt, countp, td) 1887 struct mount *oldmnt; 1888 int *countp; 1889 struct thread *td; 1890 { 1891 struct vnode *devvp; 1892 struct ufsmount *ump; 1893 int count, error; 1894 1895 /* 1896 * Alternately flush the block device associated with the mount 1897 * point and process any dependencies that the flushing 1898 * creates. We continue until no more worklist dependencies 1899 * are found. 1900 */ 1901 *countp = 0; 1902 error = 0; 1903 ump = VFSTOUFS(oldmnt); 1904 devvp = ump->um_devvp; 1905 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1906 *countp += count; 1907 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1908 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1909 VOP_UNLOCK(devvp, 0); 1910 if (error != 0) 1911 break; 1912 } 1913 return (error); 1914 } 1915 1916 #define SU_WAITIDLE_RETRIES 20 1917 static int 1918 softdep_waitidle(struct mount *mp, int flags __unused) 1919 { 1920 struct ufsmount *ump; 1921 struct vnode *devvp; 1922 struct thread *td; 1923 int error, i; 1924 1925 ump = VFSTOUFS(mp); 1926 devvp = ump->um_devvp; 1927 td = curthread; 1928 error = 0; 1929 ACQUIRE_LOCK(ump); 1930 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 1931 ump->softdep_req = 1; 1932 KASSERT((flags & FORCECLOSE) == 0 || 1933 ump->softdep_on_worklist == 0, 1934 ("softdep_waitidle: work added after flush")); 1935 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 1936 "softdeps", 10 * hz); 1937 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1938 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1939 VOP_UNLOCK(devvp, 0); 1940 if (error != 0) 1941 break; 1942 ACQUIRE_LOCK(ump); 1943 } 1944 ump->softdep_req = 0; 1945 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 1946 error = EBUSY; 1947 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1948 mp); 1949 } 1950 FREE_LOCK(ump); 1951 return (error); 1952 } 1953 1954 /* 1955 * Flush all vnodes and worklist items associated with a specified mount point. 1956 */ 1957 int 1958 softdep_flushfiles(oldmnt, flags, td) 1959 struct mount *oldmnt; 1960 int flags; 1961 struct thread *td; 1962 { 1963 #ifdef QUOTA 1964 struct ufsmount *ump; 1965 int i; 1966 #endif 1967 int error, early, depcount, loopcnt, retry_flush_count, retry; 1968 int morework; 1969 1970 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1971 ("softdep_flushfiles called on non-softdep filesystem")); 1972 loopcnt = 10; 1973 retry_flush_count = 3; 1974 retry_flush: 1975 error = 0; 1976 1977 /* 1978 * Alternately flush the vnodes associated with the mount 1979 * point and process any dependencies that the flushing 1980 * creates. In theory, this loop can happen at most twice, 1981 * but we give it a few extra just to be sure. 1982 */ 1983 for (; loopcnt > 0; loopcnt--) { 1984 /* 1985 * Do another flush in case any vnodes were brought in 1986 * as part of the cleanup operations. 1987 */ 1988 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1989 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1990 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1991 break; 1992 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1993 depcount == 0) 1994 break; 1995 } 1996 /* 1997 * If we are unmounting then it is an error to fail. If we 1998 * are simply trying to downgrade to read-only, then filesystem 1999 * activity can keep us busy forever, so we just fail with EBUSY. 2000 */ 2001 if (loopcnt == 0) { 2002 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2003 panic("softdep_flushfiles: looping"); 2004 error = EBUSY; 2005 } 2006 if (!error) 2007 error = softdep_waitidle(oldmnt, flags); 2008 if (!error) { 2009 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2010 retry = 0; 2011 MNT_ILOCK(oldmnt); 2012 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 2013 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 2014 morework = oldmnt->mnt_nvnodelistsize > 0; 2015 #ifdef QUOTA 2016 ump = VFSTOUFS(oldmnt); 2017 UFS_LOCK(ump); 2018 for (i = 0; i < MAXQUOTAS; i++) { 2019 if (ump->um_quotas[i] != NULLVP) 2020 morework = 1; 2021 } 2022 UFS_UNLOCK(ump); 2023 #endif 2024 if (morework) { 2025 if (--retry_flush_count > 0) { 2026 retry = 1; 2027 loopcnt = 3; 2028 } else 2029 error = EBUSY; 2030 } 2031 MNT_IUNLOCK(oldmnt); 2032 if (retry) 2033 goto retry_flush; 2034 } 2035 } 2036 return (error); 2037 } 2038 2039 /* 2040 * Structure hashing. 2041 * 2042 * There are four types of structures that can be looked up: 2043 * 1) pagedep structures identified by mount point, inode number, 2044 * and logical block. 2045 * 2) inodedep structures identified by mount point and inode number. 2046 * 3) newblk structures identified by mount point and 2047 * physical block number. 2048 * 4) bmsafemap structures identified by mount point and 2049 * cylinder group number. 2050 * 2051 * The "pagedep" and "inodedep" dependency structures are hashed 2052 * separately from the file blocks and inodes to which they correspond. 2053 * This separation helps when the in-memory copy of an inode or 2054 * file block must be replaced. It also obviates the need to access 2055 * an inode or file page when simply updating (or de-allocating) 2056 * dependency structures. Lookup of newblk structures is needed to 2057 * find newly allocated blocks when trying to associate them with 2058 * their allocdirect or allocindir structure. 2059 * 2060 * The lookup routines optionally create and hash a new instance when 2061 * an existing entry is not found. The bmsafemap lookup routine always 2062 * allocates a new structure if an existing one is not found. 2063 */ 2064 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2065 2066 /* 2067 * Structures and routines associated with pagedep caching. 2068 */ 2069 #define PAGEDEP_HASH(ump, inum, lbn) \ 2070 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2071 2072 static int 2073 pagedep_find(pagedephd, ino, lbn, pagedeppp) 2074 struct pagedep_hashhead *pagedephd; 2075 ino_t ino; 2076 ufs_lbn_t lbn; 2077 struct pagedep **pagedeppp; 2078 { 2079 struct pagedep *pagedep; 2080 2081 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2082 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2083 *pagedeppp = pagedep; 2084 return (1); 2085 } 2086 } 2087 *pagedeppp = NULL; 2088 return (0); 2089 } 2090 /* 2091 * Look up a pagedep. Return 1 if found, 0 otherwise. 2092 * If not found, allocate if DEPALLOC flag is passed. 2093 * Found or allocated entry is returned in pagedeppp. 2094 * This routine must be called with splbio interrupts blocked. 2095 */ 2096 static int 2097 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2098 struct mount *mp; 2099 struct buf *bp; 2100 ino_t ino; 2101 ufs_lbn_t lbn; 2102 int flags; 2103 struct pagedep **pagedeppp; 2104 { 2105 struct pagedep *pagedep; 2106 struct pagedep_hashhead *pagedephd; 2107 struct worklist *wk; 2108 struct ufsmount *ump; 2109 int ret; 2110 int i; 2111 2112 ump = VFSTOUFS(mp); 2113 LOCK_OWNED(ump); 2114 if (bp) { 2115 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2116 if (wk->wk_type == D_PAGEDEP) { 2117 *pagedeppp = WK_PAGEDEP(wk); 2118 return (1); 2119 } 2120 } 2121 } 2122 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2123 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2124 if (ret) { 2125 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2126 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2127 return (1); 2128 } 2129 if ((flags & DEPALLOC) == 0) 2130 return (0); 2131 FREE_LOCK(ump); 2132 pagedep = malloc(sizeof(struct pagedep), 2133 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2134 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2135 ACQUIRE_LOCK(ump); 2136 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2137 if (*pagedeppp) { 2138 /* 2139 * This should never happen since we only create pagedeps 2140 * with the vnode lock held. Could be an assert. 2141 */ 2142 WORKITEM_FREE(pagedep, D_PAGEDEP); 2143 return (ret); 2144 } 2145 pagedep->pd_ino = ino; 2146 pagedep->pd_lbn = lbn; 2147 LIST_INIT(&pagedep->pd_dirremhd); 2148 LIST_INIT(&pagedep->pd_pendinghd); 2149 for (i = 0; i < DAHASHSZ; i++) 2150 LIST_INIT(&pagedep->pd_diraddhd[i]); 2151 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2152 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2153 *pagedeppp = pagedep; 2154 return (0); 2155 } 2156 2157 /* 2158 * Structures and routines associated with inodedep caching. 2159 */ 2160 #define INODEDEP_HASH(ump, inum) \ 2161 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2162 2163 static int 2164 inodedep_find(inodedephd, inum, inodedeppp) 2165 struct inodedep_hashhead *inodedephd; 2166 ino_t inum; 2167 struct inodedep **inodedeppp; 2168 { 2169 struct inodedep *inodedep; 2170 2171 LIST_FOREACH(inodedep, inodedephd, id_hash) 2172 if (inum == inodedep->id_ino) 2173 break; 2174 if (inodedep) { 2175 *inodedeppp = inodedep; 2176 return (1); 2177 } 2178 *inodedeppp = NULL; 2179 2180 return (0); 2181 } 2182 /* 2183 * Look up an inodedep. Return 1 if found, 0 if not found. 2184 * If not found, allocate if DEPALLOC flag is passed. 2185 * Found or allocated entry is returned in inodedeppp. 2186 * This routine must be called with splbio interrupts blocked. 2187 */ 2188 static int 2189 inodedep_lookup(mp, inum, flags, inodedeppp) 2190 struct mount *mp; 2191 ino_t inum; 2192 int flags; 2193 struct inodedep **inodedeppp; 2194 { 2195 struct inodedep *inodedep; 2196 struct inodedep_hashhead *inodedephd; 2197 struct ufsmount *ump; 2198 struct fs *fs; 2199 2200 ump = VFSTOUFS(mp); 2201 LOCK_OWNED(ump); 2202 fs = ump->um_fs; 2203 inodedephd = INODEDEP_HASH(ump, inum); 2204 2205 if (inodedep_find(inodedephd, inum, inodedeppp)) 2206 return (1); 2207 if ((flags & DEPALLOC) == 0) 2208 return (0); 2209 /* 2210 * If the system is over its limit and our filesystem is 2211 * responsible for more than our share of that usage and 2212 * we are not in a rush, request some inodedep cleanup. 2213 */ 2214 if (softdep_excess_items(ump, D_INODEDEP)) 2215 schedule_cleanup(mp); 2216 else 2217 FREE_LOCK(ump); 2218 inodedep = malloc(sizeof(struct inodedep), 2219 M_INODEDEP, M_SOFTDEP_FLAGS); 2220 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2221 ACQUIRE_LOCK(ump); 2222 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2223 WORKITEM_FREE(inodedep, D_INODEDEP); 2224 return (1); 2225 } 2226 inodedep->id_fs = fs; 2227 inodedep->id_ino = inum; 2228 inodedep->id_state = ALLCOMPLETE; 2229 inodedep->id_nlinkdelta = 0; 2230 inodedep->id_savedino1 = NULL; 2231 inodedep->id_savedsize = -1; 2232 inodedep->id_savedextsize = -1; 2233 inodedep->id_savednlink = -1; 2234 inodedep->id_bmsafemap = NULL; 2235 inodedep->id_mkdiradd = NULL; 2236 LIST_INIT(&inodedep->id_dirremhd); 2237 LIST_INIT(&inodedep->id_pendinghd); 2238 LIST_INIT(&inodedep->id_inowait); 2239 LIST_INIT(&inodedep->id_bufwait); 2240 TAILQ_INIT(&inodedep->id_inoreflst); 2241 TAILQ_INIT(&inodedep->id_inoupdt); 2242 TAILQ_INIT(&inodedep->id_newinoupdt); 2243 TAILQ_INIT(&inodedep->id_extupdt); 2244 TAILQ_INIT(&inodedep->id_newextupdt); 2245 TAILQ_INIT(&inodedep->id_freeblklst); 2246 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2247 *inodedeppp = inodedep; 2248 return (0); 2249 } 2250 2251 /* 2252 * Structures and routines associated with newblk caching. 2253 */ 2254 #define NEWBLK_HASH(ump, inum) \ 2255 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2256 2257 static int 2258 newblk_find(newblkhd, newblkno, flags, newblkpp) 2259 struct newblk_hashhead *newblkhd; 2260 ufs2_daddr_t newblkno; 2261 int flags; 2262 struct newblk **newblkpp; 2263 { 2264 struct newblk *newblk; 2265 2266 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2267 if (newblkno != newblk->nb_newblkno) 2268 continue; 2269 /* 2270 * If we're creating a new dependency don't match those that 2271 * have already been converted to allocdirects. This is for 2272 * a frag extend. 2273 */ 2274 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2275 continue; 2276 break; 2277 } 2278 if (newblk) { 2279 *newblkpp = newblk; 2280 return (1); 2281 } 2282 *newblkpp = NULL; 2283 return (0); 2284 } 2285 2286 /* 2287 * Look up a newblk. Return 1 if found, 0 if not found. 2288 * If not found, allocate if DEPALLOC flag is passed. 2289 * Found or allocated entry is returned in newblkpp. 2290 */ 2291 static int 2292 newblk_lookup(mp, newblkno, flags, newblkpp) 2293 struct mount *mp; 2294 ufs2_daddr_t newblkno; 2295 int flags; 2296 struct newblk **newblkpp; 2297 { 2298 struct newblk *newblk; 2299 struct newblk_hashhead *newblkhd; 2300 struct ufsmount *ump; 2301 2302 ump = VFSTOUFS(mp); 2303 LOCK_OWNED(ump); 2304 newblkhd = NEWBLK_HASH(ump, newblkno); 2305 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2306 return (1); 2307 if ((flags & DEPALLOC) == 0) 2308 return (0); 2309 if (softdep_excess_items(ump, D_NEWBLK) || 2310 softdep_excess_items(ump, D_ALLOCDIRECT) || 2311 softdep_excess_items(ump, D_ALLOCINDIR)) 2312 schedule_cleanup(mp); 2313 else 2314 FREE_LOCK(ump); 2315 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2316 M_SOFTDEP_FLAGS | M_ZERO); 2317 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2318 ACQUIRE_LOCK(ump); 2319 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2320 WORKITEM_FREE(newblk, D_NEWBLK); 2321 return (1); 2322 } 2323 newblk->nb_freefrag = NULL; 2324 LIST_INIT(&newblk->nb_indirdeps); 2325 LIST_INIT(&newblk->nb_newdirblk); 2326 LIST_INIT(&newblk->nb_jwork); 2327 newblk->nb_state = ATTACHED; 2328 newblk->nb_newblkno = newblkno; 2329 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2330 *newblkpp = newblk; 2331 return (0); 2332 } 2333 2334 /* 2335 * Structures and routines associated with freed indirect block caching. 2336 */ 2337 #define INDIR_HASH(ump, blkno) \ 2338 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2339 2340 /* 2341 * Lookup an indirect block in the indir hash table. The freework is 2342 * removed and potentially freed. The caller must do a blocking journal 2343 * write before writing to the blkno. 2344 */ 2345 static int 2346 indirblk_lookup(mp, blkno) 2347 struct mount *mp; 2348 ufs2_daddr_t blkno; 2349 { 2350 struct freework *freework; 2351 struct indir_hashhead *wkhd; 2352 struct ufsmount *ump; 2353 2354 ump = VFSTOUFS(mp); 2355 wkhd = INDIR_HASH(ump, blkno); 2356 TAILQ_FOREACH(freework, wkhd, fw_next) { 2357 if (freework->fw_blkno != blkno) 2358 continue; 2359 indirblk_remove(freework); 2360 return (1); 2361 } 2362 return (0); 2363 } 2364 2365 /* 2366 * Insert an indirect block represented by freework into the indirblk 2367 * hash table so that it may prevent the block from being re-used prior 2368 * to the journal being written. 2369 */ 2370 static void 2371 indirblk_insert(freework) 2372 struct freework *freework; 2373 { 2374 struct jblocks *jblocks; 2375 struct jseg *jseg; 2376 struct ufsmount *ump; 2377 2378 ump = VFSTOUFS(freework->fw_list.wk_mp); 2379 jblocks = ump->softdep_jblocks; 2380 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2381 if (jseg == NULL) 2382 return; 2383 2384 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2385 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2386 fw_next); 2387 freework->fw_state &= ~DEPCOMPLETE; 2388 } 2389 2390 static void 2391 indirblk_remove(freework) 2392 struct freework *freework; 2393 { 2394 struct ufsmount *ump; 2395 2396 ump = VFSTOUFS(freework->fw_list.wk_mp); 2397 LIST_REMOVE(freework, fw_segs); 2398 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2399 freework->fw_state |= DEPCOMPLETE; 2400 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2401 WORKITEM_FREE(freework, D_FREEWORK); 2402 } 2403 2404 /* 2405 * Executed during filesystem system initialization before 2406 * mounting any filesystems. 2407 */ 2408 void 2409 softdep_initialize() 2410 { 2411 2412 TAILQ_INIT(&softdepmounts); 2413 #ifdef __LP64__ 2414 max_softdeps = desiredvnodes * 4; 2415 #else 2416 max_softdeps = desiredvnodes * 2; 2417 #endif 2418 2419 /* initialise bioops hack */ 2420 bioops.io_start = softdep_disk_io_initiation; 2421 bioops.io_complete = softdep_disk_write_complete; 2422 bioops.io_deallocate = softdep_deallocate_dependencies; 2423 bioops.io_countdeps = softdep_count_dependencies; 2424 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2425 2426 /* Initialize the callout with an mtx. */ 2427 callout_init_mtx(&softdep_callout, &lk, 0); 2428 } 2429 2430 /* 2431 * Executed after all filesystems have been unmounted during 2432 * filesystem module unload. 2433 */ 2434 void 2435 softdep_uninitialize() 2436 { 2437 2438 /* clear bioops hack */ 2439 bioops.io_start = NULL; 2440 bioops.io_complete = NULL; 2441 bioops.io_deallocate = NULL; 2442 bioops.io_countdeps = NULL; 2443 softdep_ast_cleanup = NULL; 2444 2445 callout_drain(&softdep_callout); 2446 } 2447 2448 /* 2449 * Called at mount time to notify the dependency code that a 2450 * filesystem wishes to use it. 2451 */ 2452 int 2453 softdep_mount(devvp, mp, fs, cred) 2454 struct vnode *devvp; 2455 struct mount *mp; 2456 struct fs *fs; 2457 struct ucred *cred; 2458 { 2459 struct csum_total cstotal; 2460 struct mount_softdeps *sdp; 2461 struct ufsmount *ump; 2462 struct cg *cgp; 2463 struct buf *bp; 2464 int i, error, cyl; 2465 2466 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2467 M_WAITOK | M_ZERO); 2468 MNT_ILOCK(mp); 2469 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2470 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2471 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2472 MNTK_SOFTDEP | MNTK_NOASYNC; 2473 } 2474 ump = VFSTOUFS(mp); 2475 ump->um_softdep = sdp; 2476 MNT_IUNLOCK(mp); 2477 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock"); 2478 sdp->sd_ump = ump; 2479 LIST_INIT(&ump->softdep_workitem_pending); 2480 LIST_INIT(&ump->softdep_journal_pending); 2481 TAILQ_INIT(&ump->softdep_unlinked); 2482 LIST_INIT(&ump->softdep_dirtycg); 2483 ump->softdep_worklist_tail = NULL; 2484 ump->softdep_on_worklist = 0; 2485 ump->softdep_deps = 0; 2486 LIST_INIT(&ump->softdep_mkdirlisthd); 2487 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2488 &ump->pagedep_hash_size); 2489 ump->pagedep_nextclean = 0; 2490 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2491 &ump->inodedep_hash_size); 2492 ump->inodedep_nextclean = 0; 2493 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2494 &ump->newblk_hash_size); 2495 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2496 &ump->bmsafemap_hash_size); 2497 i = 1 << (ffs(desiredvnodes / 10) - 1); 2498 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2499 M_FREEWORK, M_WAITOK); 2500 ump->indir_hash_size = i - 1; 2501 for (i = 0; i <= ump->indir_hash_size; i++) 2502 TAILQ_INIT(&ump->indir_hashtbl[i]); 2503 ACQUIRE_GBLLOCK(&lk); 2504 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2505 FREE_GBLLOCK(&lk); 2506 if ((fs->fs_flags & FS_SUJ) && 2507 (error = journal_mount(mp, fs, cred)) != 0) { 2508 printf("Failed to start journal: %d\n", error); 2509 softdep_unmount(mp); 2510 return (error); 2511 } 2512 /* 2513 * Start our flushing thread in the bufdaemon process. 2514 */ 2515 ACQUIRE_LOCK(ump); 2516 ump->softdep_flags |= FLUSH_STARTING; 2517 FREE_LOCK(ump); 2518 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2519 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2520 mp->mnt_stat.f_mntonname); 2521 ACQUIRE_LOCK(ump); 2522 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2523 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2524 hz / 2); 2525 } 2526 FREE_LOCK(ump); 2527 /* 2528 * When doing soft updates, the counters in the 2529 * superblock may have gotten out of sync. Recomputation 2530 * can take a long time and can be deferred for background 2531 * fsck. However, the old behavior of scanning the cylinder 2532 * groups and recalculating them at mount time is available 2533 * by setting vfs.ffs.compute_summary_at_mount to one. 2534 */ 2535 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2536 return (0); 2537 bzero(&cstotal, sizeof cstotal); 2538 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2539 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2540 fs->fs_cgsize, cred, &bp)) != 0) { 2541 brelse(bp); 2542 softdep_unmount(mp); 2543 return (error); 2544 } 2545 cgp = (struct cg *)bp->b_data; 2546 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2547 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2548 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2549 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2550 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2551 brelse(bp); 2552 } 2553 #ifdef DEBUG 2554 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2555 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2556 #endif 2557 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2558 return (0); 2559 } 2560 2561 void 2562 softdep_unmount(mp) 2563 struct mount *mp; 2564 { 2565 struct ufsmount *ump; 2566 #ifdef INVARIANTS 2567 int i; 2568 #endif 2569 2570 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2571 ("softdep_unmount called on non-softdep filesystem")); 2572 ump = VFSTOUFS(mp); 2573 MNT_ILOCK(mp); 2574 mp->mnt_flag &= ~MNT_SOFTDEP; 2575 if (MOUNTEDSUJ(mp) == 0) { 2576 MNT_IUNLOCK(mp); 2577 } else { 2578 mp->mnt_flag &= ~MNT_SUJ; 2579 MNT_IUNLOCK(mp); 2580 journal_unmount(ump); 2581 } 2582 /* 2583 * Shut down our flushing thread. Check for NULL is if 2584 * softdep_mount errors out before the thread has been created. 2585 */ 2586 if (ump->softdep_flushtd != NULL) { 2587 ACQUIRE_LOCK(ump); 2588 ump->softdep_flags |= FLUSH_EXIT; 2589 wakeup(&ump->softdep_flushtd); 2590 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2591 "sdwait", 0); 2592 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2593 ("Thread shutdown failed")); 2594 } 2595 /* 2596 * Free up our resources. 2597 */ 2598 ACQUIRE_GBLLOCK(&lk); 2599 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2600 FREE_GBLLOCK(&lk); 2601 rw_destroy(LOCK_PTR(ump)); 2602 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2603 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2604 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2605 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2606 ump->bmsafemap_hash_size); 2607 free(ump->indir_hashtbl, M_FREEWORK); 2608 #ifdef INVARIANTS 2609 for (i = 0; i <= D_LAST; i++) 2610 KASSERT(ump->softdep_curdeps[i] == 0, 2611 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2612 TYPENAME(i), ump->softdep_curdeps[i])); 2613 #endif 2614 free(ump->um_softdep, M_MOUNTDATA); 2615 } 2616 2617 static struct jblocks * 2618 jblocks_create(void) 2619 { 2620 struct jblocks *jblocks; 2621 2622 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2623 TAILQ_INIT(&jblocks->jb_segs); 2624 jblocks->jb_avail = 10; 2625 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2626 M_JBLOCKS, M_WAITOK | M_ZERO); 2627 2628 return (jblocks); 2629 } 2630 2631 static ufs2_daddr_t 2632 jblocks_alloc(jblocks, bytes, actual) 2633 struct jblocks *jblocks; 2634 int bytes; 2635 int *actual; 2636 { 2637 ufs2_daddr_t daddr; 2638 struct jextent *jext; 2639 int freecnt; 2640 int blocks; 2641 2642 blocks = bytes / DEV_BSIZE; 2643 jext = &jblocks->jb_extent[jblocks->jb_head]; 2644 freecnt = jext->je_blocks - jblocks->jb_off; 2645 if (freecnt == 0) { 2646 jblocks->jb_off = 0; 2647 if (++jblocks->jb_head > jblocks->jb_used) 2648 jblocks->jb_head = 0; 2649 jext = &jblocks->jb_extent[jblocks->jb_head]; 2650 freecnt = jext->je_blocks; 2651 } 2652 if (freecnt > blocks) 2653 freecnt = blocks; 2654 *actual = freecnt * DEV_BSIZE; 2655 daddr = jext->je_daddr + jblocks->jb_off; 2656 jblocks->jb_off += freecnt; 2657 jblocks->jb_free -= freecnt; 2658 2659 return (daddr); 2660 } 2661 2662 static void 2663 jblocks_free(jblocks, mp, bytes) 2664 struct jblocks *jblocks; 2665 struct mount *mp; 2666 int bytes; 2667 { 2668 2669 LOCK_OWNED(VFSTOUFS(mp)); 2670 jblocks->jb_free += bytes / DEV_BSIZE; 2671 if (jblocks->jb_suspended) 2672 worklist_speedup(mp); 2673 wakeup(jblocks); 2674 } 2675 2676 static void 2677 jblocks_destroy(jblocks) 2678 struct jblocks *jblocks; 2679 { 2680 2681 if (jblocks->jb_extent) 2682 free(jblocks->jb_extent, M_JBLOCKS); 2683 free(jblocks, M_JBLOCKS); 2684 } 2685 2686 static void 2687 jblocks_add(jblocks, daddr, blocks) 2688 struct jblocks *jblocks; 2689 ufs2_daddr_t daddr; 2690 int blocks; 2691 { 2692 struct jextent *jext; 2693 2694 jblocks->jb_blocks += blocks; 2695 jblocks->jb_free += blocks; 2696 jext = &jblocks->jb_extent[jblocks->jb_used]; 2697 /* Adding the first block. */ 2698 if (jext->je_daddr == 0) { 2699 jext->je_daddr = daddr; 2700 jext->je_blocks = blocks; 2701 return; 2702 } 2703 /* Extending the last extent. */ 2704 if (jext->je_daddr + jext->je_blocks == daddr) { 2705 jext->je_blocks += blocks; 2706 return; 2707 } 2708 /* Adding a new extent. */ 2709 if (++jblocks->jb_used == jblocks->jb_avail) { 2710 jblocks->jb_avail *= 2; 2711 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2712 M_JBLOCKS, M_WAITOK | M_ZERO); 2713 memcpy(jext, jblocks->jb_extent, 2714 sizeof(struct jextent) * jblocks->jb_used); 2715 free(jblocks->jb_extent, M_JBLOCKS); 2716 jblocks->jb_extent = jext; 2717 } 2718 jext = &jblocks->jb_extent[jblocks->jb_used]; 2719 jext->je_daddr = daddr; 2720 jext->je_blocks = blocks; 2721 return; 2722 } 2723 2724 int 2725 softdep_journal_lookup(mp, vpp) 2726 struct mount *mp; 2727 struct vnode **vpp; 2728 { 2729 struct componentname cnp; 2730 struct vnode *dvp; 2731 ino_t sujournal; 2732 int error; 2733 2734 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2735 if (error) 2736 return (error); 2737 bzero(&cnp, sizeof(cnp)); 2738 cnp.cn_nameiop = LOOKUP; 2739 cnp.cn_flags = ISLASTCN; 2740 cnp.cn_thread = curthread; 2741 cnp.cn_cred = curthread->td_ucred; 2742 cnp.cn_pnbuf = SUJ_FILE; 2743 cnp.cn_nameptr = SUJ_FILE; 2744 cnp.cn_namelen = strlen(SUJ_FILE); 2745 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2746 vput(dvp); 2747 if (error != 0) 2748 return (error); 2749 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2750 return (error); 2751 } 2752 2753 /* 2754 * Open and verify the journal file. 2755 */ 2756 static int 2757 journal_mount(mp, fs, cred) 2758 struct mount *mp; 2759 struct fs *fs; 2760 struct ucred *cred; 2761 { 2762 struct jblocks *jblocks; 2763 struct ufsmount *ump; 2764 struct vnode *vp; 2765 struct inode *ip; 2766 ufs2_daddr_t blkno; 2767 int bcount; 2768 int error; 2769 int i; 2770 2771 ump = VFSTOUFS(mp); 2772 ump->softdep_journal_tail = NULL; 2773 ump->softdep_on_journal = 0; 2774 ump->softdep_accdeps = 0; 2775 ump->softdep_req = 0; 2776 ump->softdep_jblocks = NULL; 2777 error = softdep_journal_lookup(mp, &vp); 2778 if (error != 0) { 2779 printf("Failed to find journal. Use tunefs to create one\n"); 2780 return (error); 2781 } 2782 ip = VTOI(vp); 2783 if (ip->i_size < SUJ_MIN) { 2784 error = ENOSPC; 2785 goto out; 2786 } 2787 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2788 jblocks = jblocks_create(); 2789 for (i = 0; i < bcount; i++) { 2790 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2791 if (error) 2792 break; 2793 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2794 } 2795 if (error) { 2796 jblocks_destroy(jblocks); 2797 goto out; 2798 } 2799 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2800 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2801 ump->softdep_jblocks = jblocks; 2802 out: 2803 if (error == 0) { 2804 MNT_ILOCK(mp); 2805 mp->mnt_flag |= MNT_SUJ; 2806 mp->mnt_flag &= ~MNT_SOFTDEP; 2807 MNT_IUNLOCK(mp); 2808 /* 2809 * Only validate the journal contents if the 2810 * filesystem is clean, otherwise we write the logs 2811 * but they'll never be used. If the filesystem was 2812 * still dirty when we mounted it the journal is 2813 * invalid and a new journal can only be valid if it 2814 * starts from a clean mount. 2815 */ 2816 if (fs->fs_clean) { 2817 DIP_SET(ip, i_modrev, fs->fs_mtime); 2818 ip->i_flags |= IN_MODIFIED; 2819 ffs_update(vp, 1); 2820 } 2821 } 2822 vput(vp); 2823 return (error); 2824 } 2825 2826 static void 2827 journal_unmount(ump) 2828 struct ufsmount *ump; 2829 { 2830 2831 if (ump->softdep_jblocks) 2832 jblocks_destroy(ump->softdep_jblocks); 2833 ump->softdep_jblocks = NULL; 2834 } 2835 2836 /* 2837 * Called when a journal record is ready to be written. Space is allocated 2838 * and the journal entry is created when the journal is flushed to stable 2839 * store. 2840 */ 2841 static void 2842 add_to_journal(wk) 2843 struct worklist *wk; 2844 { 2845 struct ufsmount *ump; 2846 2847 ump = VFSTOUFS(wk->wk_mp); 2848 LOCK_OWNED(ump); 2849 if (wk->wk_state & ONWORKLIST) 2850 panic("add_to_journal: %s(0x%X) already on list", 2851 TYPENAME(wk->wk_type), wk->wk_state); 2852 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2853 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2854 ump->softdep_jblocks->jb_age = ticks; 2855 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2856 } else 2857 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2858 ump->softdep_journal_tail = wk; 2859 ump->softdep_on_journal += 1; 2860 } 2861 2862 /* 2863 * Remove an arbitrary item for the journal worklist maintain the tail 2864 * pointer. This happens when a new operation obviates the need to 2865 * journal an old operation. 2866 */ 2867 static void 2868 remove_from_journal(wk) 2869 struct worklist *wk; 2870 { 2871 struct ufsmount *ump; 2872 2873 ump = VFSTOUFS(wk->wk_mp); 2874 LOCK_OWNED(ump); 2875 #ifdef SUJ_DEBUG 2876 { 2877 struct worklist *wkn; 2878 2879 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2880 if (wkn == wk) 2881 break; 2882 if (wkn == NULL) 2883 panic("remove_from_journal: %p is not in journal", wk); 2884 } 2885 #endif 2886 /* 2887 * We emulate a TAILQ to save space in most structures which do not 2888 * require TAILQ semantics. Here we must update the tail position 2889 * when removing the tail which is not the final entry. This works 2890 * only if the worklist linkage are at the beginning of the structure. 2891 */ 2892 if (ump->softdep_journal_tail == wk) 2893 ump->softdep_journal_tail = 2894 (struct worklist *)wk->wk_list.le_prev; 2895 2896 WORKLIST_REMOVE(wk); 2897 ump->softdep_on_journal -= 1; 2898 } 2899 2900 /* 2901 * Check for journal space as well as dependency limits so the prelink 2902 * code can throttle both journaled and non-journaled filesystems. 2903 * Threshold is 0 for low and 1 for min. 2904 */ 2905 static int 2906 journal_space(ump, thresh) 2907 struct ufsmount *ump; 2908 int thresh; 2909 { 2910 struct jblocks *jblocks; 2911 int limit, avail; 2912 2913 jblocks = ump->softdep_jblocks; 2914 if (jblocks == NULL) 2915 return (1); 2916 /* 2917 * We use a tighter restriction here to prevent request_cleanup() 2918 * running in threads from running into locks we currently hold. 2919 * We have to be over the limit and our filesystem has to be 2920 * responsible for more than our share of that usage. 2921 */ 2922 limit = (max_softdeps / 10) * 9; 2923 if (dep_current[D_INODEDEP] > limit && 2924 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2925 return (0); 2926 if (thresh) 2927 thresh = jblocks->jb_min; 2928 else 2929 thresh = jblocks->jb_low; 2930 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2931 avail = jblocks->jb_free - avail; 2932 2933 return (avail > thresh); 2934 } 2935 2936 static void 2937 journal_suspend(ump) 2938 struct ufsmount *ump; 2939 { 2940 struct jblocks *jblocks; 2941 struct mount *mp; 2942 2943 mp = UFSTOVFS(ump); 2944 jblocks = ump->softdep_jblocks; 2945 MNT_ILOCK(mp); 2946 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2947 stat_journal_min++; 2948 mp->mnt_kern_flag |= MNTK_SUSPEND; 2949 mp->mnt_susp_owner = ump->softdep_flushtd; 2950 } 2951 jblocks->jb_suspended = 1; 2952 MNT_IUNLOCK(mp); 2953 } 2954 2955 static int 2956 journal_unsuspend(struct ufsmount *ump) 2957 { 2958 struct jblocks *jblocks; 2959 struct mount *mp; 2960 2961 mp = UFSTOVFS(ump); 2962 jblocks = ump->softdep_jblocks; 2963 2964 if (jblocks != NULL && jblocks->jb_suspended && 2965 journal_space(ump, jblocks->jb_min)) { 2966 jblocks->jb_suspended = 0; 2967 FREE_LOCK(ump); 2968 mp->mnt_susp_owner = curthread; 2969 vfs_write_resume(mp, 0); 2970 ACQUIRE_LOCK(ump); 2971 return (1); 2972 } 2973 return (0); 2974 } 2975 2976 /* 2977 * Called before any allocation function to be certain that there is 2978 * sufficient space in the journal prior to creating any new records. 2979 * Since in the case of block allocation we may have multiple locked 2980 * buffers at the time of the actual allocation we can not block 2981 * when the journal records are created. Doing so would create a deadlock 2982 * if any of these buffers needed to be flushed to reclaim space. Instead 2983 * we require a sufficiently large amount of available space such that 2984 * each thread in the system could have passed this allocation check and 2985 * still have sufficient free space. With 20% of a minimum journal size 2986 * of 1MB we have 6553 records available. 2987 */ 2988 int 2989 softdep_prealloc(vp, waitok) 2990 struct vnode *vp; 2991 int waitok; 2992 { 2993 struct ufsmount *ump; 2994 2995 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 2996 ("softdep_prealloc called on non-softdep filesystem")); 2997 /* 2998 * Nothing to do if we are not running journaled soft updates. 2999 * If we currently hold the snapshot lock, we must avoid handling 3000 * other resources that could cause deadlock. 3001 */ 3002 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp))) 3003 return (0); 3004 ump = VFSTOUFS(vp->v_mount); 3005 ACQUIRE_LOCK(ump); 3006 if (journal_space(ump, 0)) { 3007 FREE_LOCK(ump); 3008 return (0); 3009 } 3010 stat_journal_low++; 3011 FREE_LOCK(ump); 3012 if (waitok == MNT_NOWAIT) 3013 return (ENOSPC); 3014 /* 3015 * Attempt to sync this vnode once to flush any journal 3016 * work attached to it. 3017 */ 3018 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3019 ffs_syncvnode(vp, waitok, 0); 3020 ACQUIRE_LOCK(ump); 3021 process_removes(vp); 3022 process_truncates(vp); 3023 if (journal_space(ump, 0) == 0) { 3024 softdep_speedup(ump); 3025 if (journal_space(ump, 1) == 0) 3026 journal_suspend(ump); 3027 } 3028 FREE_LOCK(ump); 3029 3030 return (0); 3031 } 3032 3033 /* 3034 * Before adjusting a link count on a vnode verify that we have sufficient 3035 * journal space. If not, process operations that depend on the currently 3036 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3037 * and softdep flush threads can not acquire these locks to reclaim space. 3038 */ 3039 static void 3040 softdep_prelink(dvp, vp) 3041 struct vnode *dvp; 3042 struct vnode *vp; 3043 { 3044 struct ufsmount *ump; 3045 3046 ump = VFSTOUFS(dvp->v_mount); 3047 LOCK_OWNED(ump); 3048 /* 3049 * Nothing to do if we have sufficient journal space. 3050 * If we currently hold the snapshot lock, we must avoid 3051 * handling other resources that could cause deadlock. 3052 */ 3053 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3054 return; 3055 stat_journal_low++; 3056 FREE_LOCK(ump); 3057 if (vp) 3058 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3059 ffs_syncvnode(dvp, MNT_WAIT, 0); 3060 ACQUIRE_LOCK(ump); 3061 /* Process vp before dvp as it may create .. removes. */ 3062 if (vp) { 3063 process_removes(vp); 3064 process_truncates(vp); 3065 } 3066 process_removes(dvp); 3067 process_truncates(dvp); 3068 softdep_speedup(ump); 3069 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3070 if (journal_space(ump, 0) == 0) { 3071 softdep_speedup(ump); 3072 if (journal_space(ump, 1) == 0) 3073 journal_suspend(ump); 3074 } 3075 } 3076 3077 static void 3078 jseg_write(ump, jseg, data) 3079 struct ufsmount *ump; 3080 struct jseg *jseg; 3081 uint8_t *data; 3082 { 3083 struct jsegrec *rec; 3084 3085 rec = (struct jsegrec *)data; 3086 rec->jsr_seq = jseg->js_seq; 3087 rec->jsr_oldest = jseg->js_oldseq; 3088 rec->jsr_cnt = jseg->js_cnt; 3089 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3090 rec->jsr_crc = 0; 3091 rec->jsr_time = ump->um_fs->fs_mtime; 3092 } 3093 3094 static inline void 3095 inoref_write(inoref, jseg, rec) 3096 struct inoref *inoref; 3097 struct jseg *jseg; 3098 struct jrefrec *rec; 3099 { 3100 3101 inoref->if_jsegdep->jd_seg = jseg; 3102 rec->jr_ino = inoref->if_ino; 3103 rec->jr_parent = inoref->if_parent; 3104 rec->jr_nlink = inoref->if_nlink; 3105 rec->jr_mode = inoref->if_mode; 3106 rec->jr_diroff = inoref->if_diroff; 3107 } 3108 3109 static void 3110 jaddref_write(jaddref, jseg, data) 3111 struct jaddref *jaddref; 3112 struct jseg *jseg; 3113 uint8_t *data; 3114 { 3115 struct jrefrec *rec; 3116 3117 rec = (struct jrefrec *)data; 3118 rec->jr_op = JOP_ADDREF; 3119 inoref_write(&jaddref->ja_ref, jseg, rec); 3120 } 3121 3122 static void 3123 jremref_write(jremref, jseg, data) 3124 struct jremref *jremref; 3125 struct jseg *jseg; 3126 uint8_t *data; 3127 { 3128 struct jrefrec *rec; 3129 3130 rec = (struct jrefrec *)data; 3131 rec->jr_op = JOP_REMREF; 3132 inoref_write(&jremref->jr_ref, jseg, rec); 3133 } 3134 3135 static void 3136 jmvref_write(jmvref, jseg, data) 3137 struct jmvref *jmvref; 3138 struct jseg *jseg; 3139 uint8_t *data; 3140 { 3141 struct jmvrec *rec; 3142 3143 rec = (struct jmvrec *)data; 3144 rec->jm_op = JOP_MVREF; 3145 rec->jm_ino = jmvref->jm_ino; 3146 rec->jm_parent = jmvref->jm_parent; 3147 rec->jm_oldoff = jmvref->jm_oldoff; 3148 rec->jm_newoff = jmvref->jm_newoff; 3149 } 3150 3151 static void 3152 jnewblk_write(jnewblk, jseg, data) 3153 struct jnewblk *jnewblk; 3154 struct jseg *jseg; 3155 uint8_t *data; 3156 { 3157 struct jblkrec *rec; 3158 3159 jnewblk->jn_jsegdep->jd_seg = jseg; 3160 rec = (struct jblkrec *)data; 3161 rec->jb_op = JOP_NEWBLK; 3162 rec->jb_ino = jnewblk->jn_ino; 3163 rec->jb_blkno = jnewblk->jn_blkno; 3164 rec->jb_lbn = jnewblk->jn_lbn; 3165 rec->jb_frags = jnewblk->jn_frags; 3166 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3167 } 3168 3169 static void 3170 jfreeblk_write(jfreeblk, jseg, data) 3171 struct jfreeblk *jfreeblk; 3172 struct jseg *jseg; 3173 uint8_t *data; 3174 { 3175 struct jblkrec *rec; 3176 3177 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3178 rec = (struct jblkrec *)data; 3179 rec->jb_op = JOP_FREEBLK; 3180 rec->jb_ino = jfreeblk->jf_ino; 3181 rec->jb_blkno = jfreeblk->jf_blkno; 3182 rec->jb_lbn = jfreeblk->jf_lbn; 3183 rec->jb_frags = jfreeblk->jf_frags; 3184 rec->jb_oldfrags = 0; 3185 } 3186 3187 static void 3188 jfreefrag_write(jfreefrag, jseg, data) 3189 struct jfreefrag *jfreefrag; 3190 struct jseg *jseg; 3191 uint8_t *data; 3192 { 3193 struct jblkrec *rec; 3194 3195 jfreefrag->fr_jsegdep->jd_seg = jseg; 3196 rec = (struct jblkrec *)data; 3197 rec->jb_op = JOP_FREEBLK; 3198 rec->jb_ino = jfreefrag->fr_ino; 3199 rec->jb_blkno = jfreefrag->fr_blkno; 3200 rec->jb_lbn = jfreefrag->fr_lbn; 3201 rec->jb_frags = jfreefrag->fr_frags; 3202 rec->jb_oldfrags = 0; 3203 } 3204 3205 static void 3206 jtrunc_write(jtrunc, jseg, data) 3207 struct jtrunc *jtrunc; 3208 struct jseg *jseg; 3209 uint8_t *data; 3210 { 3211 struct jtrncrec *rec; 3212 3213 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3214 rec = (struct jtrncrec *)data; 3215 rec->jt_op = JOP_TRUNC; 3216 rec->jt_ino = jtrunc->jt_ino; 3217 rec->jt_size = jtrunc->jt_size; 3218 rec->jt_extsize = jtrunc->jt_extsize; 3219 } 3220 3221 static void 3222 jfsync_write(jfsync, jseg, data) 3223 struct jfsync *jfsync; 3224 struct jseg *jseg; 3225 uint8_t *data; 3226 { 3227 struct jtrncrec *rec; 3228 3229 rec = (struct jtrncrec *)data; 3230 rec->jt_op = JOP_SYNC; 3231 rec->jt_ino = jfsync->jfs_ino; 3232 rec->jt_size = jfsync->jfs_size; 3233 rec->jt_extsize = jfsync->jfs_extsize; 3234 } 3235 3236 static void 3237 softdep_flushjournal(mp) 3238 struct mount *mp; 3239 { 3240 struct jblocks *jblocks; 3241 struct ufsmount *ump; 3242 3243 if (MOUNTEDSUJ(mp) == 0) 3244 return; 3245 ump = VFSTOUFS(mp); 3246 jblocks = ump->softdep_jblocks; 3247 ACQUIRE_LOCK(ump); 3248 while (ump->softdep_on_journal) { 3249 jblocks->jb_needseg = 1; 3250 softdep_process_journal(mp, NULL, MNT_WAIT); 3251 } 3252 FREE_LOCK(ump); 3253 } 3254 3255 static void softdep_synchronize_completed(struct bio *); 3256 static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3257 3258 static void 3259 softdep_synchronize_completed(bp) 3260 struct bio *bp; 3261 { 3262 struct jseg *oldest; 3263 struct jseg *jseg; 3264 struct ufsmount *ump; 3265 3266 /* 3267 * caller1 marks the last segment written before we issued the 3268 * synchronize cache. 3269 */ 3270 jseg = bp->bio_caller1; 3271 if (jseg == NULL) { 3272 g_destroy_bio(bp); 3273 return; 3274 } 3275 ump = VFSTOUFS(jseg->js_list.wk_mp); 3276 ACQUIRE_LOCK(ump); 3277 oldest = NULL; 3278 /* 3279 * Mark all the journal entries waiting on the synchronize cache 3280 * as completed so they may continue on. 3281 */ 3282 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3283 jseg->js_state |= COMPLETE; 3284 oldest = jseg; 3285 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3286 } 3287 /* 3288 * Restart deferred journal entry processing from the oldest 3289 * completed jseg. 3290 */ 3291 if (oldest) 3292 complete_jsegs(oldest); 3293 3294 FREE_LOCK(ump); 3295 g_destroy_bio(bp); 3296 } 3297 3298 /* 3299 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3300 * barriers. The journal must be written prior to any blocks that depend 3301 * on it and the journal can not be released until the blocks have be 3302 * written. This code handles both barriers simultaneously. 3303 */ 3304 static void 3305 softdep_synchronize(bp, ump, caller1) 3306 struct bio *bp; 3307 struct ufsmount *ump; 3308 void *caller1; 3309 { 3310 3311 bp->bio_cmd = BIO_FLUSH; 3312 bp->bio_flags |= BIO_ORDERED; 3313 bp->bio_data = NULL; 3314 bp->bio_offset = ump->um_cp->provider->mediasize; 3315 bp->bio_length = 0; 3316 bp->bio_done = softdep_synchronize_completed; 3317 bp->bio_caller1 = caller1; 3318 g_io_request(bp, 3319 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3320 } 3321 3322 /* 3323 * Flush some journal records to disk. 3324 */ 3325 static void 3326 softdep_process_journal(mp, needwk, flags) 3327 struct mount *mp; 3328 struct worklist *needwk; 3329 int flags; 3330 { 3331 struct jblocks *jblocks; 3332 struct ufsmount *ump; 3333 struct worklist *wk; 3334 struct jseg *jseg; 3335 struct buf *bp; 3336 struct bio *bio; 3337 uint8_t *data; 3338 struct fs *fs; 3339 int shouldflush; 3340 int segwritten; 3341 int jrecmin; /* Minimum records per block. */ 3342 int jrecmax; /* Maximum records per block. */ 3343 int size; 3344 int cnt; 3345 int off; 3346 int devbsize; 3347 3348 if (MOUNTEDSUJ(mp) == 0) 3349 return; 3350 shouldflush = softdep_flushcache; 3351 bio = NULL; 3352 jseg = NULL; 3353 ump = VFSTOUFS(mp); 3354 LOCK_OWNED(ump); 3355 fs = ump->um_fs; 3356 jblocks = ump->softdep_jblocks; 3357 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3358 /* 3359 * We write anywhere between a disk block and fs block. The upper 3360 * bound is picked to prevent buffer cache fragmentation and limit 3361 * processing time per I/O. 3362 */ 3363 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3364 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3365 segwritten = 0; 3366 for (;;) { 3367 cnt = ump->softdep_on_journal; 3368 /* 3369 * Criteria for writing a segment: 3370 * 1) We have a full block. 3371 * 2) We're called from jwait() and haven't found the 3372 * journal item yet. 3373 * 3) Always write if needseg is set. 3374 * 4) If we are called from process_worklist and have 3375 * not yet written anything we write a partial block 3376 * to enforce a 1 second maximum latency on journal 3377 * entries. 3378 */ 3379 if (cnt < (jrecmax - 1) && needwk == NULL && 3380 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3381 break; 3382 cnt++; 3383 /* 3384 * Verify some free journal space. softdep_prealloc() should 3385 * guarantee that we don't run out so this is indicative of 3386 * a problem with the flow control. Try to recover 3387 * gracefully in any event. 3388 */ 3389 while (jblocks->jb_free == 0) { 3390 if (flags != MNT_WAIT) 3391 break; 3392 printf("softdep: Out of journal space!\n"); 3393 softdep_speedup(ump); 3394 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3395 } 3396 FREE_LOCK(ump); 3397 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3398 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3399 LIST_INIT(&jseg->js_entries); 3400 LIST_INIT(&jseg->js_indirs); 3401 jseg->js_state = ATTACHED; 3402 if (shouldflush == 0) 3403 jseg->js_state |= COMPLETE; 3404 else if (bio == NULL) 3405 bio = g_alloc_bio(); 3406 jseg->js_jblocks = jblocks; 3407 bp = geteblk(fs->fs_bsize, 0); 3408 ACQUIRE_LOCK(ump); 3409 /* 3410 * If there was a race while we were allocating the block 3411 * and jseg the entry we care about was likely written. 3412 * We bail out in both the WAIT and NOWAIT case and assume 3413 * the caller will loop if the entry it cares about is 3414 * not written. 3415 */ 3416 cnt = ump->softdep_on_journal; 3417 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3418 bp->b_flags |= B_INVAL | B_NOCACHE; 3419 WORKITEM_FREE(jseg, D_JSEG); 3420 FREE_LOCK(ump); 3421 brelse(bp); 3422 ACQUIRE_LOCK(ump); 3423 break; 3424 } 3425 /* 3426 * Calculate the disk block size required for the available 3427 * records rounded to the min size. 3428 */ 3429 if (cnt == 0) 3430 size = devbsize; 3431 else if (cnt < jrecmax) 3432 size = howmany(cnt, jrecmin) * devbsize; 3433 else 3434 size = fs->fs_bsize; 3435 /* 3436 * Allocate a disk block for this journal data and account 3437 * for truncation of the requested size if enough contiguous 3438 * space was not available. 3439 */ 3440 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3441 bp->b_lblkno = bp->b_blkno; 3442 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3443 bp->b_bcount = size; 3444 bp->b_flags &= ~B_INVAL; 3445 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3446 /* 3447 * Initialize our jseg with cnt records. Assign the next 3448 * sequence number to it and link it in-order. 3449 */ 3450 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3451 jseg->js_buf = bp; 3452 jseg->js_cnt = cnt; 3453 jseg->js_refs = cnt + 1; /* Self ref. */ 3454 jseg->js_size = size; 3455 jseg->js_seq = jblocks->jb_nextseq++; 3456 if (jblocks->jb_oldestseg == NULL) 3457 jblocks->jb_oldestseg = jseg; 3458 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3459 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3460 if (jblocks->jb_writeseg == NULL) 3461 jblocks->jb_writeseg = jseg; 3462 /* 3463 * Start filling in records from the pending list. 3464 */ 3465 data = bp->b_data; 3466 off = 0; 3467 3468 /* 3469 * Always put a header on the first block. 3470 * XXX As with below, there might not be a chance to get 3471 * into the loop. Ensure that something valid is written. 3472 */ 3473 jseg_write(ump, jseg, data); 3474 off += JREC_SIZE; 3475 data = bp->b_data + off; 3476 3477 /* 3478 * XXX Something is wrong here. There's no work to do, 3479 * but we need to perform and I/O and allow it to complete 3480 * anyways. 3481 */ 3482 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3483 stat_emptyjblocks++; 3484 3485 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3486 != NULL) { 3487 if (cnt == 0) 3488 break; 3489 /* Place a segment header on every device block. */ 3490 if ((off % devbsize) == 0) { 3491 jseg_write(ump, jseg, data); 3492 off += JREC_SIZE; 3493 data = bp->b_data + off; 3494 } 3495 if (wk == needwk) 3496 needwk = NULL; 3497 remove_from_journal(wk); 3498 wk->wk_state |= INPROGRESS; 3499 WORKLIST_INSERT(&jseg->js_entries, wk); 3500 switch (wk->wk_type) { 3501 case D_JADDREF: 3502 jaddref_write(WK_JADDREF(wk), jseg, data); 3503 break; 3504 case D_JREMREF: 3505 jremref_write(WK_JREMREF(wk), jseg, data); 3506 break; 3507 case D_JMVREF: 3508 jmvref_write(WK_JMVREF(wk), jseg, data); 3509 break; 3510 case D_JNEWBLK: 3511 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3512 break; 3513 case D_JFREEBLK: 3514 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3515 break; 3516 case D_JFREEFRAG: 3517 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3518 break; 3519 case D_JTRUNC: 3520 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3521 break; 3522 case D_JFSYNC: 3523 jfsync_write(WK_JFSYNC(wk), jseg, data); 3524 break; 3525 default: 3526 panic("process_journal: Unknown type %s", 3527 TYPENAME(wk->wk_type)); 3528 /* NOTREACHED */ 3529 } 3530 off += JREC_SIZE; 3531 data = bp->b_data + off; 3532 cnt--; 3533 } 3534 3535 /* Clear any remaining space so we don't leak kernel data */ 3536 if (size > off) 3537 bzero(data, size - off); 3538 3539 /* 3540 * Write this one buffer and continue. 3541 */ 3542 segwritten = 1; 3543 jblocks->jb_needseg = 0; 3544 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3545 FREE_LOCK(ump); 3546 pbgetvp(ump->um_devvp, bp); 3547 /* 3548 * We only do the blocking wait once we find the journal 3549 * entry we're looking for. 3550 */ 3551 if (needwk == NULL && flags == MNT_WAIT) 3552 bwrite(bp); 3553 else 3554 bawrite(bp); 3555 ACQUIRE_LOCK(ump); 3556 } 3557 /* 3558 * If we wrote a segment issue a synchronize cache so the journal 3559 * is reflected on disk before the data is written. Since reclaiming 3560 * journal space also requires writing a journal record this 3561 * process also enforces a barrier before reclamation. 3562 */ 3563 if (segwritten && shouldflush) { 3564 softdep_synchronize(bio, ump, 3565 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3566 } else if (bio) 3567 g_destroy_bio(bio); 3568 /* 3569 * If we've suspended the filesystem because we ran out of journal 3570 * space either try to sync it here to make some progress or 3571 * unsuspend it if we already have. 3572 */ 3573 if (flags == 0 && jblocks->jb_suspended) { 3574 if (journal_unsuspend(ump)) 3575 return; 3576 FREE_LOCK(ump); 3577 VFS_SYNC(mp, MNT_NOWAIT); 3578 ffs_sbupdate(ump, MNT_WAIT, 0); 3579 ACQUIRE_LOCK(ump); 3580 } 3581 } 3582 3583 /* 3584 * Complete a jseg, allowing all dependencies awaiting journal writes 3585 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3586 * structures so that the journal segment can be freed to reclaim space. 3587 */ 3588 static void 3589 complete_jseg(jseg) 3590 struct jseg *jseg; 3591 { 3592 struct worklist *wk; 3593 struct jmvref *jmvref; 3594 int waiting; 3595 #ifdef INVARIANTS 3596 int i = 0; 3597 #endif 3598 3599 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3600 WORKLIST_REMOVE(wk); 3601 waiting = wk->wk_state & IOWAITING; 3602 wk->wk_state &= ~(INPROGRESS | IOWAITING); 3603 wk->wk_state |= COMPLETE; 3604 KASSERT(i++ < jseg->js_cnt, 3605 ("handle_written_jseg: overflow %d >= %d", 3606 i - 1, jseg->js_cnt)); 3607 switch (wk->wk_type) { 3608 case D_JADDREF: 3609 handle_written_jaddref(WK_JADDREF(wk)); 3610 break; 3611 case D_JREMREF: 3612 handle_written_jremref(WK_JREMREF(wk)); 3613 break; 3614 case D_JMVREF: 3615 rele_jseg(jseg); /* No jsegdep. */ 3616 jmvref = WK_JMVREF(wk); 3617 LIST_REMOVE(jmvref, jm_deps); 3618 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3619 free_pagedep(jmvref->jm_pagedep); 3620 WORKITEM_FREE(jmvref, D_JMVREF); 3621 break; 3622 case D_JNEWBLK: 3623 handle_written_jnewblk(WK_JNEWBLK(wk)); 3624 break; 3625 case D_JFREEBLK: 3626 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3627 break; 3628 case D_JTRUNC: 3629 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3630 break; 3631 case D_JFSYNC: 3632 rele_jseg(jseg); /* No jsegdep. */ 3633 WORKITEM_FREE(wk, D_JFSYNC); 3634 break; 3635 case D_JFREEFRAG: 3636 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3637 break; 3638 default: 3639 panic("handle_written_jseg: Unknown type %s", 3640 TYPENAME(wk->wk_type)); 3641 /* NOTREACHED */ 3642 } 3643 if (waiting) 3644 wakeup(wk); 3645 } 3646 /* Release the self reference so the structure may be freed. */ 3647 rele_jseg(jseg); 3648 } 3649 3650 /* 3651 * Determine which jsegs are ready for completion processing. Waits for 3652 * synchronize cache to complete as well as forcing in-order completion 3653 * of journal entries. 3654 */ 3655 static void 3656 complete_jsegs(jseg) 3657 struct jseg *jseg; 3658 { 3659 struct jblocks *jblocks; 3660 struct jseg *jsegn; 3661 3662 jblocks = jseg->js_jblocks; 3663 /* 3664 * Don't allow out of order completions. If this isn't the first 3665 * block wait for it to write before we're done. 3666 */ 3667 if (jseg != jblocks->jb_writeseg) 3668 return; 3669 /* Iterate through available jsegs processing their entries. */ 3670 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3671 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3672 jsegn = TAILQ_NEXT(jseg, js_next); 3673 complete_jseg(jseg); 3674 jseg = jsegn; 3675 } 3676 jblocks->jb_writeseg = jseg; 3677 /* 3678 * Attempt to free jsegs now that oldestwrseq may have advanced. 3679 */ 3680 free_jsegs(jblocks); 3681 } 3682 3683 /* 3684 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3685 * the final completions. 3686 */ 3687 static void 3688 handle_written_jseg(jseg, bp) 3689 struct jseg *jseg; 3690 struct buf *bp; 3691 { 3692 3693 if (jseg->js_refs == 0) 3694 panic("handle_written_jseg: No self-reference on %p", jseg); 3695 jseg->js_state |= DEPCOMPLETE; 3696 /* 3697 * We'll never need this buffer again, set flags so it will be 3698 * discarded. 3699 */ 3700 bp->b_flags |= B_INVAL | B_NOCACHE; 3701 pbrelvp(bp); 3702 complete_jsegs(jseg); 3703 } 3704 3705 static inline struct jsegdep * 3706 inoref_jseg(inoref) 3707 struct inoref *inoref; 3708 { 3709 struct jsegdep *jsegdep; 3710 3711 jsegdep = inoref->if_jsegdep; 3712 inoref->if_jsegdep = NULL; 3713 3714 return (jsegdep); 3715 } 3716 3717 /* 3718 * Called once a jremref has made it to stable store. The jremref is marked 3719 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3720 * for the jremref to complete will be awoken by free_jremref. 3721 */ 3722 static void 3723 handle_written_jremref(jremref) 3724 struct jremref *jremref; 3725 { 3726 struct inodedep *inodedep; 3727 struct jsegdep *jsegdep; 3728 struct dirrem *dirrem; 3729 3730 /* Grab the jsegdep. */ 3731 jsegdep = inoref_jseg(&jremref->jr_ref); 3732 /* 3733 * Remove us from the inoref list. 3734 */ 3735 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3736 0, &inodedep) == 0) 3737 panic("handle_written_jremref: Lost inodedep"); 3738 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3739 /* 3740 * Complete the dirrem. 3741 */ 3742 dirrem = jremref->jr_dirrem; 3743 jremref->jr_dirrem = NULL; 3744 LIST_REMOVE(jremref, jr_deps); 3745 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3746 jwork_insert(&dirrem->dm_jwork, jsegdep); 3747 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3748 (dirrem->dm_state & COMPLETE) != 0) 3749 add_to_worklist(&dirrem->dm_list, 0); 3750 free_jremref(jremref); 3751 } 3752 3753 /* 3754 * Called once a jaddref has made it to stable store. The dependency is 3755 * marked complete and any dependent structures are added to the inode 3756 * bufwait list to be completed as soon as it is written. If a bitmap write 3757 * depends on this entry we move the inode into the inodedephd of the 3758 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3759 */ 3760 static void 3761 handle_written_jaddref(jaddref) 3762 struct jaddref *jaddref; 3763 { 3764 struct jsegdep *jsegdep; 3765 struct inodedep *inodedep; 3766 struct diradd *diradd; 3767 struct mkdir *mkdir; 3768 3769 /* Grab the jsegdep. */ 3770 jsegdep = inoref_jseg(&jaddref->ja_ref); 3771 mkdir = NULL; 3772 diradd = NULL; 3773 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3774 0, &inodedep) == 0) 3775 panic("handle_written_jaddref: Lost inodedep."); 3776 if (jaddref->ja_diradd == NULL) 3777 panic("handle_written_jaddref: No dependency"); 3778 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3779 diradd = jaddref->ja_diradd; 3780 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3781 } else if (jaddref->ja_state & MKDIR_PARENT) { 3782 mkdir = jaddref->ja_mkdir; 3783 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3784 } else if (jaddref->ja_state & MKDIR_BODY) 3785 mkdir = jaddref->ja_mkdir; 3786 else 3787 panic("handle_written_jaddref: Unknown dependency %p", 3788 jaddref->ja_diradd); 3789 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3790 /* 3791 * Remove us from the inode list. 3792 */ 3793 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3794 /* 3795 * The mkdir may be waiting on the jaddref to clear before freeing. 3796 */ 3797 if (mkdir) { 3798 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3799 ("handle_written_jaddref: Incorrect type for mkdir %s", 3800 TYPENAME(mkdir->md_list.wk_type))); 3801 mkdir->md_jaddref = NULL; 3802 diradd = mkdir->md_diradd; 3803 mkdir->md_state |= DEPCOMPLETE; 3804 complete_mkdir(mkdir); 3805 } 3806 jwork_insert(&diradd->da_jwork, jsegdep); 3807 if (jaddref->ja_state & NEWBLOCK) { 3808 inodedep->id_state |= ONDEPLIST; 3809 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3810 inodedep, id_deps); 3811 } 3812 free_jaddref(jaddref); 3813 } 3814 3815 /* 3816 * Called once a jnewblk journal is written. The allocdirect or allocindir 3817 * is placed in the bmsafemap to await notification of a written bitmap. If 3818 * the operation was canceled we add the segdep to the appropriate 3819 * dependency to free the journal space once the canceling operation 3820 * completes. 3821 */ 3822 static void 3823 handle_written_jnewblk(jnewblk) 3824 struct jnewblk *jnewblk; 3825 { 3826 struct bmsafemap *bmsafemap; 3827 struct freefrag *freefrag; 3828 struct freework *freework; 3829 struct jsegdep *jsegdep; 3830 struct newblk *newblk; 3831 3832 /* Grab the jsegdep. */ 3833 jsegdep = jnewblk->jn_jsegdep; 3834 jnewblk->jn_jsegdep = NULL; 3835 if (jnewblk->jn_dep == NULL) 3836 panic("handle_written_jnewblk: No dependency for the segdep."); 3837 switch (jnewblk->jn_dep->wk_type) { 3838 case D_NEWBLK: 3839 case D_ALLOCDIRECT: 3840 case D_ALLOCINDIR: 3841 /* 3842 * Add the written block to the bmsafemap so it can 3843 * be notified when the bitmap is on disk. 3844 */ 3845 newblk = WK_NEWBLK(jnewblk->jn_dep); 3846 newblk->nb_jnewblk = NULL; 3847 if ((newblk->nb_state & GOINGAWAY) == 0) { 3848 bmsafemap = newblk->nb_bmsafemap; 3849 newblk->nb_state |= ONDEPLIST; 3850 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3851 nb_deps); 3852 } 3853 jwork_insert(&newblk->nb_jwork, jsegdep); 3854 break; 3855 case D_FREEFRAG: 3856 /* 3857 * A newblock being removed by a freefrag when replaced by 3858 * frag extension. 3859 */ 3860 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3861 freefrag->ff_jdep = NULL; 3862 jwork_insert(&freefrag->ff_jwork, jsegdep); 3863 break; 3864 case D_FREEWORK: 3865 /* 3866 * A direct block was removed by truncate. 3867 */ 3868 freework = WK_FREEWORK(jnewblk->jn_dep); 3869 freework->fw_jnewblk = NULL; 3870 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3871 break; 3872 default: 3873 panic("handle_written_jnewblk: Unknown type %d.", 3874 jnewblk->jn_dep->wk_type); 3875 } 3876 jnewblk->jn_dep = NULL; 3877 free_jnewblk(jnewblk); 3878 } 3879 3880 /* 3881 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3882 * an in-flight allocation that has not yet been committed. Divorce us 3883 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3884 * to the worklist. 3885 */ 3886 static void 3887 cancel_jfreefrag(jfreefrag) 3888 struct jfreefrag *jfreefrag; 3889 { 3890 struct freefrag *freefrag; 3891 3892 if (jfreefrag->fr_jsegdep) { 3893 free_jsegdep(jfreefrag->fr_jsegdep); 3894 jfreefrag->fr_jsegdep = NULL; 3895 } 3896 freefrag = jfreefrag->fr_freefrag; 3897 jfreefrag->fr_freefrag = NULL; 3898 free_jfreefrag(jfreefrag); 3899 freefrag->ff_state |= DEPCOMPLETE; 3900 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3901 } 3902 3903 /* 3904 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3905 */ 3906 static void 3907 free_jfreefrag(jfreefrag) 3908 struct jfreefrag *jfreefrag; 3909 { 3910 3911 if (jfreefrag->fr_state & INPROGRESS) 3912 WORKLIST_REMOVE(&jfreefrag->fr_list); 3913 else if (jfreefrag->fr_state & ONWORKLIST) 3914 remove_from_journal(&jfreefrag->fr_list); 3915 if (jfreefrag->fr_freefrag != NULL) 3916 panic("free_jfreefrag: Still attached to a freefrag."); 3917 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3918 } 3919 3920 /* 3921 * Called when the journal write for a jfreefrag completes. The parent 3922 * freefrag is added to the worklist if this completes its dependencies. 3923 */ 3924 static void 3925 handle_written_jfreefrag(jfreefrag) 3926 struct jfreefrag *jfreefrag; 3927 { 3928 struct jsegdep *jsegdep; 3929 struct freefrag *freefrag; 3930 3931 /* Grab the jsegdep. */ 3932 jsegdep = jfreefrag->fr_jsegdep; 3933 jfreefrag->fr_jsegdep = NULL; 3934 freefrag = jfreefrag->fr_freefrag; 3935 if (freefrag == NULL) 3936 panic("handle_written_jfreefrag: No freefrag."); 3937 freefrag->ff_state |= DEPCOMPLETE; 3938 freefrag->ff_jdep = NULL; 3939 jwork_insert(&freefrag->ff_jwork, jsegdep); 3940 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3941 add_to_worklist(&freefrag->ff_list, 0); 3942 jfreefrag->fr_freefrag = NULL; 3943 free_jfreefrag(jfreefrag); 3944 } 3945 3946 /* 3947 * Called when the journal write for a jfreeblk completes. The jfreeblk 3948 * is removed from the freeblks list of pending journal writes and the 3949 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3950 * have been reclaimed. 3951 */ 3952 static void 3953 handle_written_jblkdep(jblkdep) 3954 struct jblkdep *jblkdep; 3955 { 3956 struct freeblks *freeblks; 3957 struct jsegdep *jsegdep; 3958 3959 /* Grab the jsegdep. */ 3960 jsegdep = jblkdep->jb_jsegdep; 3961 jblkdep->jb_jsegdep = NULL; 3962 freeblks = jblkdep->jb_freeblks; 3963 LIST_REMOVE(jblkdep, jb_deps); 3964 jwork_insert(&freeblks->fb_jwork, jsegdep); 3965 /* 3966 * If the freeblks is all journaled, we can add it to the worklist. 3967 */ 3968 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3969 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3970 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3971 3972 free_jblkdep(jblkdep); 3973 } 3974 3975 static struct jsegdep * 3976 newjsegdep(struct worklist *wk) 3977 { 3978 struct jsegdep *jsegdep; 3979 3980 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3981 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3982 jsegdep->jd_seg = NULL; 3983 3984 return (jsegdep); 3985 } 3986 3987 static struct jmvref * 3988 newjmvref(dp, ino, oldoff, newoff) 3989 struct inode *dp; 3990 ino_t ino; 3991 off_t oldoff; 3992 off_t newoff; 3993 { 3994 struct jmvref *jmvref; 3995 3996 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3997 workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump)); 3998 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 3999 jmvref->jm_parent = dp->i_number; 4000 jmvref->jm_ino = ino; 4001 jmvref->jm_oldoff = oldoff; 4002 jmvref->jm_newoff = newoff; 4003 4004 return (jmvref); 4005 } 4006 4007 /* 4008 * Allocate a new jremref that tracks the removal of ip from dp with the 4009 * directory entry offset of diroff. Mark the entry as ATTACHED and 4010 * DEPCOMPLETE as we have all the information required for the journal write 4011 * and the directory has already been removed from the buffer. The caller 4012 * is responsible for linking the jremref into the pagedep and adding it 4013 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4014 * a DOTDOT addition so handle_workitem_remove() can properly assign 4015 * the jsegdep when we're done. 4016 */ 4017 static struct jremref * 4018 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4019 off_t diroff, nlink_t nlink) 4020 { 4021 struct jremref *jremref; 4022 4023 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4024 workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump)); 4025 jremref->jr_state = ATTACHED; 4026 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4027 nlink, ip->i_mode); 4028 jremref->jr_dirrem = dirrem; 4029 4030 return (jremref); 4031 } 4032 4033 static inline void 4034 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4035 nlink_t nlink, uint16_t mode) 4036 { 4037 4038 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4039 inoref->if_diroff = diroff; 4040 inoref->if_ino = ino; 4041 inoref->if_parent = parent; 4042 inoref->if_nlink = nlink; 4043 inoref->if_mode = mode; 4044 } 4045 4046 /* 4047 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4048 * directory offset may not be known until later. The caller is responsible 4049 * adding the entry to the journal when this information is available. nlink 4050 * should be the link count prior to the addition and mode is only required 4051 * to have the correct FMT. 4052 */ 4053 static struct jaddref * 4054 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4055 uint16_t mode) 4056 { 4057 struct jaddref *jaddref; 4058 4059 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4060 workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump)); 4061 jaddref->ja_state = ATTACHED; 4062 jaddref->ja_mkdir = NULL; 4063 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4064 4065 return (jaddref); 4066 } 4067 4068 /* 4069 * Create a new free dependency for a freework. The caller is responsible 4070 * for adjusting the reference count when it has the lock held. The freedep 4071 * will track an outstanding bitmap write that will ultimately clear the 4072 * freework to continue. 4073 */ 4074 static struct freedep * 4075 newfreedep(struct freework *freework) 4076 { 4077 struct freedep *freedep; 4078 4079 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4080 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4081 freedep->fd_freework = freework; 4082 4083 return (freedep); 4084 } 4085 4086 /* 4087 * Free a freedep structure once the buffer it is linked to is written. If 4088 * this is the last reference to the freework schedule it for completion. 4089 */ 4090 static void 4091 free_freedep(freedep) 4092 struct freedep *freedep; 4093 { 4094 struct freework *freework; 4095 4096 freework = freedep->fd_freework; 4097 freework->fw_freeblks->fb_cgwait--; 4098 if (--freework->fw_ref == 0) 4099 freework_enqueue(freework); 4100 WORKITEM_FREE(freedep, D_FREEDEP); 4101 } 4102 4103 /* 4104 * Allocate a new freework structure that may be a level in an indirect 4105 * when parent is not NULL or a top level block when it is. The top level 4106 * freework structures are allocated without the per-filesystem lock held 4107 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4108 */ 4109 static struct freework * 4110 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4111 struct ufsmount *ump; 4112 struct freeblks *freeblks; 4113 struct freework *parent; 4114 ufs_lbn_t lbn; 4115 ufs2_daddr_t nb; 4116 int frags; 4117 int off; 4118 int journal; 4119 { 4120 struct freework *freework; 4121 4122 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4123 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4124 freework->fw_state = ATTACHED; 4125 freework->fw_jnewblk = NULL; 4126 freework->fw_freeblks = freeblks; 4127 freework->fw_parent = parent; 4128 freework->fw_lbn = lbn; 4129 freework->fw_blkno = nb; 4130 freework->fw_frags = frags; 4131 freework->fw_indir = NULL; 4132 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR) 4133 ? 0 : NINDIR(ump->um_fs) + 1; 4134 freework->fw_start = freework->fw_off = off; 4135 if (journal) 4136 newjfreeblk(freeblks, lbn, nb, frags); 4137 if (parent == NULL) { 4138 ACQUIRE_LOCK(ump); 4139 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4140 freeblks->fb_ref++; 4141 FREE_LOCK(ump); 4142 } 4143 4144 return (freework); 4145 } 4146 4147 /* 4148 * Eliminate a jfreeblk for a block that does not need journaling. 4149 */ 4150 static void 4151 cancel_jfreeblk(freeblks, blkno) 4152 struct freeblks *freeblks; 4153 ufs2_daddr_t blkno; 4154 { 4155 struct jfreeblk *jfreeblk; 4156 struct jblkdep *jblkdep; 4157 4158 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4159 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4160 continue; 4161 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4162 if (jfreeblk->jf_blkno == blkno) 4163 break; 4164 } 4165 if (jblkdep == NULL) 4166 return; 4167 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4168 free_jsegdep(jblkdep->jb_jsegdep); 4169 LIST_REMOVE(jblkdep, jb_deps); 4170 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4171 } 4172 4173 /* 4174 * Allocate a new jfreeblk to journal top level block pointer when truncating 4175 * a file. The caller must add this to the worklist when the per-filesystem 4176 * lock is held. 4177 */ 4178 static struct jfreeblk * 4179 newjfreeblk(freeblks, lbn, blkno, frags) 4180 struct freeblks *freeblks; 4181 ufs_lbn_t lbn; 4182 ufs2_daddr_t blkno; 4183 int frags; 4184 { 4185 struct jfreeblk *jfreeblk; 4186 4187 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4188 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4189 freeblks->fb_list.wk_mp); 4190 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4191 jfreeblk->jf_dep.jb_freeblks = freeblks; 4192 jfreeblk->jf_ino = freeblks->fb_inum; 4193 jfreeblk->jf_lbn = lbn; 4194 jfreeblk->jf_blkno = blkno; 4195 jfreeblk->jf_frags = frags; 4196 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4197 4198 return (jfreeblk); 4199 } 4200 4201 /* 4202 * The journal is only prepared to handle full-size block numbers, so we 4203 * have to adjust the record to reflect the change to a full-size block. 4204 * For example, suppose we have a block made up of fragments 8-15 and 4205 * want to free its last two fragments. We are given a request that says: 4206 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4207 * where frags are the number of fragments to free and oldfrags are the 4208 * number of fragments to keep. To block align it, we have to change it to 4209 * have a valid full-size blkno, so it becomes: 4210 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4211 */ 4212 static void 4213 adjust_newfreework(freeblks, frag_offset) 4214 struct freeblks *freeblks; 4215 int frag_offset; 4216 { 4217 struct jfreeblk *jfreeblk; 4218 4219 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4220 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4221 ("adjust_newfreework: Missing freeblks dependency")); 4222 4223 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4224 jfreeblk->jf_blkno -= frag_offset; 4225 jfreeblk->jf_frags += frag_offset; 4226 } 4227 4228 /* 4229 * Allocate a new jtrunc to track a partial truncation. 4230 */ 4231 static struct jtrunc * 4232 newjtrunc(freeblks, size, extsize) 4233 struct freeblks *freeblks; 4234 off_t size; 4235 int extsize; 4236 { 4237 struct jtrunc *jtrunc; 4238 4239 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4240 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4241 freeblks->fb_list.wk_mp); 4242 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4243 jtrunc->jt_dep.jb_freeblks = freeblks; 4244 jtrunc->jt_ino = freeblks->fb_inum; 4245 jtrunc->jt_size = size; 4246 jtrunc->jt_extsize = extsize; 4247 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4248 4249 return (jtrunc); 4250 } 4251 4252 /* 4253 * If we're canceling a new bitmap we have to search for another ref 4254 * to move into the bmsafemap dep. This might be better expressed 4255 * with another structure. 4256 */ 4257 static void 4258 move_newblock_dep(jaddref, inodedep) 4259 struct jaddref *jaddref; 4260 struct inodedep *inodedep; 4261 { 4262 struct inoref *inoref; 4263 struct jaddref *jaddrefn; 4264 4265 jaddrefn = NULL; 4266 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4267 inoref = TAILQ_NEXT(inoref, if_deps)) { 4268 if ((jaddref->ja_state & NEWBLOCK) && 4269 inoref->if_list.wk_type == D_JADDREF) { 4270 jaddrefn = (struct jaddref *)inoref; 4271 break; 4272 } 4273 } 4274 if (jaddrefn == NULL) 4275 return; 4276 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4277 jaddrefn->ja_state |= jaddref->ja_state & 4278 (ATTACHED | UNDONE | NEWBLOCK); 4279 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4280 jaddref->ja_state |= ATTACHED; 4281 LIST_REMOVE(jaddref, ja_bmdeps); 4282 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4283 ja_bmdeps); 4284 } 4285 4286 /* 4287 * Cancel a jaddref either before it has been written or while it is being 4288 * written. This happens when a link is removed before the add reaches 4289 * the disk. The jaddref dependency is kept linked into the bmsafemap 4290 * and inode to prevent the link count or bitmap from reaching the disk 4291 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4292 * required. 4293 * 4294 * Returns 1 if the canceled addref requires journaling of the remove and 4295 * 0 otherwise. 4296 */ 4297 static int 4298 cancel_jaddref(jaddref, inodedep, wkhd) 4299 struct jaddref *jaddref; 4300 struct inodedep *inodedep; 4301 struct workhead *wkhd; 4302 { 4303 struct inoref *inoref; 4304 struct jsegdep *jsegdep; 4305 int needsj; 4306 4307 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4308 ("cancel_jaddref: Canceling complete jaddref")); 4309 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4310 needsj = 1; 4311 else 4312 needsj = 0; 4313 if (inodedep == NULL) 4314 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4315 0, &inodedep) == 0) 4316 panic("cancel_jaddref: Lost inodedep"); 4317 /* 4318 * We must adjust the nlink of any reference operation that follows 4319 * us so that it is consistent with the in-memory reference. This 4320 * ensures that inode nlink rollbacks always have the correct link. 4321 */ 4322 if (needsj == 0) { 4323 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4324 inoref = TAILQ_NEXT(inoref, if_deps)) { 4325 if (inoref->if_state & GOINGAWAY) 4326 break; 4327 inoref->if_nlink--; 4328 } 4329 } 4330 jsegdep = inoref_jseg(&jaddref->ja_ref); 4331 if (jaddref->ja_state & NEWBLOCK) 4332 move_newblock_dep(jaddref, inodedep); 4333 wake_worklist(&jaddref->ja_list); 4334 jaddref->ja_mkdir = NULL; 4335 if (jaddref->ja_state & INPROGRESS) { 4336 jaddref->ja_state &= ~INPROGRESS; 4337 WORKLIST_REMOVE(&jaddref->ja_list); 4338 jwork_insert(wkhd, jsegdep); 4339 } else { 4340 free_jsegdep(jsegdep); 4341 if (jaddref->ja_state & DEPCOMPLETE) 4342 remove_from_journal(&jaddref->ja_list); 4343 } 4344 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4345 /* 4346 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4347 * can arrange for them to be freed with the bitmap. Otherwise we 4348 * no longer need this addref attached to the inoreflst and it 4349 * will incorrectly adjust nlink if we leave it. 4350 */ 4351 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4352 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4353 if_deps); 4354 jaddref->ja_state |= COMPLETE; 4355 free_jaddref(jaddref); 4356 return (needsj); 4357 } 4358 /* 4359 * Leave the head of the list for jsegdeps for fast merging. 4360 */ 4361 if (LIST_FIRST(wkhd) != NULL) { 4362 jaddref->ja_state |= ONWORKLIST; 4363 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4364 } else 4365 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4366 4367 return (needsj); 4368 } 4369 4370 /* 4371 * Attempt to free a jaddref structure when some work completes. This 4372 * should only succeed once the entry is written and all dependencies have 4373 * been notified. 4374 */ 4375 static void 4376 free_jaddref(jaddref) 4377 struct jaddref *jaddref; 4378 { 4379 4380 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4381 return; 4382 if (jaddref->ja_ref.if_jsegdep) 4383 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4384 jaddref, jaddref->ja_state); 4385 if (jaddref->ja_state & NEWBLOCK) 4386 LIST_REMOVE(jaddref, ja_bmdeps); 4387 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4388 panic("free_jaddref: Bad state %p(0x%X)", 4389 jaddref, jaddref->ja_state); 4390 if (jaddref->ja_mkdir != NULL) 4391 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4392 WORKITEM_FREE(jaddref, D_JADDREF); 4393 } 4394 4395 /* 4396 * Free a jremref structure once it has been written or discarded. 4397 */ 4398 static void 4399 free_jremref(jremref) 4400 struct jremref *jremref; 4401 { 4402 4403 if (jremref->jr_ref.if_jsegdep) 4404 free_jsegdep(jremref->jr_ref.if_jsegdep); 4405 if (jremref->jr_state & INPROGRESS) 4406 panic("free_jremref: IO still pending"); 4407 WORKITEM_FREE(jremref, D_JREMREF); 4408 } 4409 4410 /* 4411 * Free a jnewblk structure. 4412 */ 4413 static void 4414 free_jnewblk(jnewblk) 4415 struct jnewblk *jnewblk; 4416 { 4417 4418 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4419 return; 4420 LIST_REMOVE(jnewblk, jn_deps); 4421 if (jnewblk->jn_dep != NULL) 4422 panic("free_jnewblk: Dependency still attached."); 4423 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4424 } 4425 4426 /* 4427 * Cancel a jnewblk which has been been made redundant by frag extension. 4428 */ 4429 static void 4430 cancel_jnewblk(jnewblk, wkhd) 4431 struct jnewblk *jnewblk; 4432 struct workhead *wkhd; 4433 { 4434 struct jsegdep *jsegdep; 4435 4436 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4437 jsegdep = jnewblk->jn_jsegdep; 4438 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4439 panic("cancel_jnewblk: Invalid state"); 4440 jnewblk->jn_jsegdep = NULL; 4441 jnewblk->jn_dep = NULL; 4442 jnewblk->jn_state |= GOINGAWAY; 4443 if (jnewblk->jn_state & INPROGRESS) { 4444 jnewblk->jn_state &= ~INPROGRESS; 4445 WORKLIST_REMOVE(&jnewblk->jn_list); 4446 jwork_insert(wkhd, jsegdep); 4447 } else { 4448 free_jsegdep(jsegdep); 4449 remove_from_journal(&jnewblk->jn_list); 4450 } 4451 wake_worklist(&jnewblk->jn_list); 4452 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4453 } 4454 4455 static void 4456 free_jblkdep(jblkdep) 4457 struct jblkdep *jblkdep; 4458 { 4459 4460 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4461 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4462 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4463 WORKITEM_FREE(jblkdep, D_JTRUNC); 4464 else 4465 panic("free_jblkdep: Unexpected type %s", 4466 TYPENAME(jblkdep->jb_list.wk_type)); 4467 } 4468 4469 /* 4470 * Free a single jseg once it is no longer referenced in memory or on 4471 * disk. Reclaim journal blocks and dependencies waiting for the segment 4472 * to disappear. 4473 */ 4474 static void 4475 free_jseg(jseg, jblocks) 4476 struct jseg *jseg; 4477 struct jblocks *jblocks; 4478 { 4479 struct freework *freework; 4480 4481 /* 4482 * Free freework structures that were lingering to indicate freed 4483 * indirect blocks that forced journal write ordering on reallocate. 4484 */ 4485 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4486 indirblk_remove(freework); 4487 if (jblocks->jb_oldestseg == jseg) 4488 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4489 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4490 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4491 KASSERT(LIST_EMPTY(&jseg->js_entries), 4492 ("free_jseg: Freed jseg has valid entries.")); 4493 WORKITEM_FREE(jseg, D_JSEG); 4494 } 4495 4496 /* 4497 * Free all jsegs that meet the criteria for being reclaimed and update 4498 * oldestseg. 4499 */ 4500 static void 4501 free_jsegs(jblocks) 4502 struct jblocks *jblocks; 4503 { 4504 struct jseg *jseg; 4505 4506 /* 4507 * Free only those jsegs which have none allocated before them to 4508 * preserve the journal space ordering. 4509 */ 4510 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4511 /* 4512 * Only reclaim space when nothing depends on this journal 4513 * set and another set has written that it is no longer 4514 * valid. 4515 */ 4516 if (jseg->js_refs != 0) { 4517 jblocks->jb_oldestseg = jseg; 4518 return; 4519 } 4520 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4521 break; 4522 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4523 break; 4524 /* 4525 * We can free jsegs that didn't write entries when 4526 * oldestwrseq == js_seq. 4527 */ 4528 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4529 jseg->js_cnt != 0) 4530 break; 4531 free_jseg(jseg, jblocks); 4532 } 4533 /* 4534 * If we exited the loop above we still must discover the 4535 * oldest valid segment. 4536 */ 4537 if (jseg) 4538 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4539 jseg = TAILQ_NEXT(jseg, js_next)) 4540 if (jseg->js_refs != 0) 4541 break; 4542 jblocks->jb_oldestseg = jseg; 4543 /* 4544 * The journal has no valid records but some jsegs may still be 4545 * waiting on oldestwrseq to advance. We force a small record 4546 * out to permit these lingering records to be reclaimed. 4547 */ 4548 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4549 jblocks->jb_needseg = 1; 4550 } 4551 4552 /* 4553 * Release one reference to a jseg and free it if the count reaches 0. This 4554 * should eventually reclaim journal space as well. 4555 */ 4556 static void 4557 rele_jseg(jseg) 4558 struct jseg *jseg; 4559 { 4560 4561 KASSERT(jseg->js_refs > 0, 4562 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4563 if (--jseg->js_refs != 0) 4564 return; 4565 free_jsegs(jseg->js_jblocks); 4566 } 4567 4568 /* 4569 * Release a jsegdep and decrement the jseg count. 4570 */ 4571 static void 4572 free_jsegdep(jsegdep) 4573 struct jsegdep *jsegdep; 4574 { 4575 4576 if (jsegdep->jd_seg) 4577 rele_jseg(jsegdep->jd_seg); 4578 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4579 } 4580 4581 /* 4582 * Wait for a journal item to make it to disk. Initiate journal processing 4583 * if required. 4584 */ 4585 static int 4586 jwait(wk, waitfor) 4587 struct worklist *wk; 4588 int waitfor; 4589 { 4590 4591 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4592 /* 4593 * Blocking journal waits cause slow synchronous behavior. Record 4594 * stats on the frequency of these blocking operations. 4595 */ 4596 if (waitfor == MNT_WAIT) { 4597 stat_journal_wait++; 4598 switch (wk->wk_type) { 4599 case D_JREMREF: 4600 case D_JMVREF: 4601 stat_jwait_filepage++; 4602 break; 4603 case D_JTRUNC: 4604 case D_JFREEBLK: 4605 stat_jwait_freeblks++; 4606 break; 4607 case D_JNEWBLK: 4608 stat_jwait_newblk++; 4609 break; 4610 case D_JADDREF: 4611 stat_jwait_inode++; 4612 break; 4613 default: 4614 break; 4615 } 4616 } 4617 /* 4618 * If IO has not started we process the journal. We can't mark the 4619 * worklist item as IOWAITING because we drop the lock while 4620 * processing the journal and the worklist entry may be freed after 4621 * this point. The caller may call back in and re-issue the request. 4622 */ 4623 if ((wk->wk_state & INPROGRESS) == 0) { 4624 softdep_process_journal(wk->wk_mp, wk, waitfor); 4625 if (waitfor != MNT_WAIT) 4626 return (EBUSY); 4627 return (0); 4628 } 4629 if (waitfor != MNT_WAIT) 4630 return (EBUSY); 4631 wait_worklist(wk, "jwait"); 4632 return (0); 4633 } 4634 4635 /* 4636 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4637 * appropriate. This is a convenience function to reduce duplicate code 4638 * for the setup and revert functions below. 4639 */ 4640 static struct inodedep * 4641 inodedep_lookup_ip(ip) 4642 struct inode *ip; 4643 { 4644 struct inodedep *inodedep; 4645 4646 KASSERT(ip->i_nlink >= ip->i_effnlink, 4647 ("inodedep_lookup_ip: bad delta")); 4648 (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, 4649 &inodedep); 4650 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4651 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4652 4653 return (inodedep); 4654 } 4655 4656 /* 4657 * Called prior to creating a new inode and linking it to a directory. The 4658 * jaddref structure must already be allocated by softdep_setup_inomapdep 4659 * and it is discovered here so we can initialize the mode and update 4660 * nlinkdelta. 4661 */ 4662 void 4663 softdep_setup_create(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 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4672 ("softdep_setup_create called on non-softdep filesystem")); 4673 KASSERT(ip->i_nlink == 1, 4674 ("softdep_setup_create: Invalid link count.")); 4675 dvp = ITOV(dp); 4676 ACQUIRE_LOCK(dp->i_ump); 4677 inodedep = inodedep_lookup_ip(ip); 4678 if (DOINGSUJ(dvp)) { 4679 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4680 inoreflst); 4681 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4682 ("softdep_setup_create: No addref structure present.")); 4683 } 4684 softdep_prelink(dvp, NULL); 4685 FREE_LOCK(dp->i_ump); 4686 } 4687 4688 /* 4689 * Create a jaddref structure to track the addition of a DOTDOT link when 4690 * we are reparenting an inode as part of a rename. This jaddref will be 4691 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4692 * non-journaling softdep. 4693 */ 4694 void 4695 softdep_setup_dotdot_link(dp, ip) 4696 struct inode *dp; 4697 struct inode *ip; 4698 { 4699 struct inodedep *inodedep; 4700 struct jaddref *jaddref; 4701 struct vnode *dvp; 4702 4703 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4704 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4705 dvp = ITOV(dp); 4706 jaddref = NULL; 4707 /* 4708 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4709 * is used as a normal link would be. 4710 */ 4711 if (DOINGSUJ(dvp)) 4712 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4713 dp->i_effnlink - 1, dp->i_mode); 4714 ACQUIRE_LOCK(dp->i_ump); 4715 inodedep = inodedep_lookup_ip(dp); 4716 if (jaddref) 4717 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4718 if_deps); 4719 softdep_prelink(dvp, ITOV(ip)); 4720 FREE_LOCK(dp->i_ump); 4721 } 4722 4723 /* 4724 * Create a jaddref structure to track a new link to an inode. The directory 4725 * offset is not known until softdep_setup_directory_add or 4726 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4727 * softdep. 4728 */ 4729 void 4730 softdep_setup_link(dp, ip) 4731 struct inode *dp; 4732 struct inode *ip; 4733 { 4734 struct inodedep *inodedep; 4735 struct jaddref *jaddref; 4736 struct vnode *dvp; 4737 4738 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4739 ("softdep_setup_link called on non-softdep filesystem")); 4740 dvp = ITOV(dp); 4741 jaddref = NULL; 4742 if (DOINGSUJ(dvp)) 4743 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4744 ip->i_mode); 4745 ACQUIRE_LOCK(dp->i_ump); 4746 inodedep = inodedep_lookup_ip(ip); 4747 if (jaddref) 4748 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4749 if_deps); 4750 softdep_prelink(dvp, ITOV(ip)); 4751 FREE_LOCK(dp->i_ump); 4752 } 4753 4754 /* 4755 * Called to create the jaddref structures to track . and .. references as 4756 * well as lookup and further initialize the incomplete jaddref created 4757 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4758 * nlinkdelta for non-journaling softdep. 4759 */ 4760 void 4761 softdep_setup_mkdir(dp, ip) 4762 struct inode *dp; 4763 struct inode *ip; 4764 { 4765 struct inodedep *inodedep; 4766 struct jaddref *dotdotaddref; 4767 struct jaddref *dotaddref; 4768 struct jaddref *jaddref; 4769 struct vnode *dvp; 4770 4771 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4772 ("softdep_setup_mkdir called on non-softdep filesystem")); 4773 dvp = ITOV(dp); 4774 dotaddref = dotdotaddref = NULL; 4775 if (DOINGSUJ(dvp)) { 4776 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4777 ip->i_mode); 4778 dotaddref->ja_state |= MKDIR_BODY; 4779 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4780 dp->i_effnlink - 1, dp->i_mode); 4781 dotdotaddref->ja_state |= MKDIR_PARENT; 4782 } 4783 ACQUIRE_LOCK(dp->i_ump); 4784 inodedep = inodedep_lookup_ip(ip); 4785 if (DOINGSUJ(dvp)) { 4786 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4787 inoreflst); 4788 KASSERT(jaddref != NULL, 4789 ("softdep_setup_mkdir: No addref structure present.")); 4790 KASSERT(jaddref->ja_parent == dp->i_number, 4791 ("softdep_setup_mkdir: bad parent %ju", 4792 (uintmax_t)jaddref->ja_parent)); 4793 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4794 if_deps); 4795 } 4796 inodedep = inodedep_lookup_ip(dp); 4797 if (DOINGSUJ(dvp)) 4798 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4799 &dotdotaddref->ja_ref, if_deps); 4800 softdep_prelink(ITOV(dp), NULL); 4801 FREE_LOCK(dp->i_ump); 4802 } 4803 4804 /* 4805 * Called to track nlinkdelta of the inode and parent directories prior to 4806 * unlinking a directory. 4807 */ 4808 void 4809 softdep_setup_rmdir(dp, ip) 4810 struct inode *dp; 4811 struct inode *ip; 4812 { 4813 struct vnode *dvp; 4814 4815 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4816 ("softdep_setup_rmdir called on non-softdep filesystem")); 4817 dvp = ITOV(dp); 4818 ACQUIRE_LOCK(dp->i_ump); 4819 (void) inodedep_lookup_ip(ip); 4820 (void) inodedep_lookup_ip(dp); 4821 softdep_prelink(dvp, ITOV(ip)); 4822 FREE_LOCK(dp->i_ump); 4823 } 4824 4825 /* 4826 * Called to track nlinkdelta of the inode and parent directories prior to 4827 * unlink. 4828 */ 4829 void 4830 softdep_setup_unlink(dp, ip) 4831 struct inode *dp; 4832 struct inode *ip; 4833 { 4834 struct vnode *dvp; 4835 4836 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4837 ("softdep_setup_unlink called on non-softdep filesystem")); 4838 dvp = ITOV(dp); 4839 ACQUIRE_LOCK(dp->i_ump); 4840 (void) inodedep_lookup_ip(ip); 4841 (void) inodedep_lookup_ip(dp); 4842 softdep_prelink(dvp, ITOV(ip)); 4843 FREE_LOCK(dp->i_ump); 4844 } 4845 4846 /* 4847 * Called to release the journal structures created by a failed non-directory 4848 * creation. Adjusts nlinkdelta for non-journaling softdep. 4849 */ 4850 void 4851 softdep_revert_create(dp, ip) 4852 struct inode *dp; 4853 struct inode *ip; 4854 { 4855 struct inodedep *inodedep; 4856 struct jaddref *jaddref; 4857 struct vnode *dvp; 4858 4859 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4860 ("softdep_revert_create called on non-softdep filesystem")); 4861 dvp = ITOV(dp); 4862 ACQUIRE_LOCK(dp->i_ump); 4863 inodedep = inodedep_lookup_ip(ip); 4864 if (DOINGSUJ(dvp)) { 4865 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4866 inoreflst); 4867 KASSERT(jaddref->ja_parent == dp->i_number, 4868 ("softdep_revert_create: addref parent mismatch")); 4869 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4870 } 4871 FREE_LOCK(dp->i_ump); 4872 } 4873 4874 /* 4875 * Called to release the journal structures created by a failed link 4876 * addition. Adjusts nlinkdelta for non-journaling softdep. 4877 */ 4878 void 4879 softdep_revert_link(dp, ip) 4880 struct inode *dp; 4881 struct inode *ip; 4882 { 4883 struct inodedep *inodedep; 4884 struct jaddref *jaddref; 4885 struct vnode *dvp; 4886 4887 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4888 ("softdep_revert_link called on non-softdep filesystem")); 4889 dvp = ITOV(dp); 4890 ACQUIRE_LOCK(dp->i_ump); 4891 inodedep = inodedep_lookup_ip(ip); 4892 if (DOINGSUJ(dvp)) { 4893 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4894 inoreflst); 4895 KASSERT(jaddref->ja_parent == dp->i_number, 4896 ("softdep_revert_link: addref parent mismatch")); 4897 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4898 } 4899 FREE_LOCK(dp->i_ump); 4900 } 4901 4902 /* 4903 * Called to release the journal structures created by a failed mkdir 4904 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4905 */ 4906 void 4907 softdep_revert_mkdir(dp, ip) 4908 struct inode *dp; 4909 struct inode *ip; 4910 { 4911 struct inodedep *inodedep; 4912 struct jaddref *jaddref; 4913 struct jaddref *dotaddref; 4914 struct vnode *dvp; 4915 4916 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4917 ("softdep_revert_mkdir called on non-softdep filesystem")); 4918 dvp = ITOV(dp); 4919 4920 ACQUIRE_LOCK(dp->i_ump); 4921 inodedep = inodedep_lookup_ip(dp); 4922 if (DOINGSUJ(dvp)) { 4923 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4924 inoreflst); 4925 KASSERT(jaddref->ja_parent == ip->i_number, 4926 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4927 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4928 } 4929 inodedep = inodedep_lookup_ip(ip); 4930 if (DOINGSUJ(dvp)) { 4931 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4932 inoreflst); 4933 KASSERT(jaddref->ja_parent == dp->i_number, 4934 ("softdep_revert_mkdir: addref parent mismatch")); 4935 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4936 inoreflst, if_deps); 4937 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4938 KASSERT(dotaddref->ja_parent == ip->i_number, 4939 ("softdep_revert_mkdir: dot addref parent mismatch")); 4940 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4941 } 4942 FREE_LOCK(dp->i_ump); 4943 } 4944 4945 /* 4946 * Called to correct nlinkdelta after a failed rmdir. 4947 */ 4948 void 4949 softdep_revert_rmdir(dp, ip) 4950 struct inode *dp; 4951 struct inode *ip; 4952 { 4953 4954 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(dp->i_ump)) != 0, 4955 ("softdep_revert_rmdir called on non-softdep filesystem")); 4956 ACQUIRE_LOCK(dp->i_ump); 4957 (void) inodedep_lookup_ip(ip); 4958 (void) inodedep_lookup_ip(dp); 4959 FREE_LOCK(dp->i_ump); 4960 } 4961 4962 /* 4963 * Protecting the freemaps (or bitmaps). 4964 * 4965 * To eliminate the need to execute fsck before mounting a filesystem 4966 * after a power failure, one must (conservatively) guarantee that the 4967 * on-disk copy of the bitmaps never indicate that a live inode or block is 4968 * free. So, when a block or inode is allocated, the bitmap should be 4969 * updated (on disk) before any new pointers. When a block or inode is 4970 * freed, the bitmap should not be updated until all pointers have been 4971 * reset. The latter dependency is handled by the delayed de-allocation 4972 * approach described below for block and inode de-allocation. The former 4973 * dependency is handled by calling the following procedure when a block or 4974 * inode is allocated. When an inode is allocated an "inodedep" is created 4975 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4976 * Each "inodedep" is also inserted into the hash indexing structure so 4977 * that any additional link additions can be made dependent on the inode 4978 * allocation. 4979 * 4980 * The ufs filesystem maintains a number of free block counts (e.g., per 4981 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4982 * in addition to the bitmaps. These counts are used to improve efficiency 4983 * during allocation and therefore must be consistent with the bitmaps. 4984 * There is no convenient way to guarantee post-crash consistency of these 4985 * counts with simple update ordering, for two main reasons: (1) The counts 4986 * and bitmaps for a single cylinder group block are not in the same disk 4987 * sector. If a disk write is interrupted (e.g., by power failure), one may 4988 * be written and the other not. (2) Some of the counts are located in the 4989 * superblock rather than the cylinder group block. So, we focus our soft 4990 * updates implementation on protecting the bitmaps. When mounting a 4991 * filesystem, we recompute the auxiliary counts from the bitmaps. 4992 */ 4993 4994 /* 4995 * Called just after updating the cylinder group block to allocate an inode. 4996 */ 4997 void 4998 softdep_setup_inomapdep(bp, ip, newinum, mode) 4999 struct buf *bp; /* buffer for cylgroup block with inode map */ 5000 struct inode *ip; /* inode related to allocation */ 5001 ino_t newinum; /* new inode number being allocated */ 5002 int mode; 5003 { 5004 struct inodedep *inodedep; 5005 struct bmsafemap *bmsafemap; 5006 struct jaddref *jaddref; 5007 struct mount *mp; 5008 struct fs *fs; 5009 5010 mp = UFSTOVFS(ip->i_ump); 5011 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5012 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5013 fs = ip->i_ump->um_fs; 5014 jaddref = NULL; 5015 5016 /* 5017 * Allocate the journal reference add structure so that the bitmap 5018 * can be dependent on it. 5019 */ 5020 if (MOUNTEDSUJ(mp)) { 5021 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5022 jaddref->ja_state |= NEWBLOCK; 5023 } 5024 5025 /* 5026 * Create a dependency for the newly allocated inode. 5027 * Panic if it already exists as something is seriously wrong. 5028 * Otherwise add it to the dependency list for the buffer holding 5029 * the cylinder group map from which it was allocated. 5030 * 5031 * We have to preallocate a bmsafemap entry in case it is needed 5032 * in bmsafemap_lookup since once we allocate the inodedep, we 5033 * have to finish initializing it before we can FREE_LOCK(). 5034 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5035 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5036 * creating the inodedep as it can be freed during the time 5037 * that we FREE_LOCK() while allocating the inodedep. We must 5038 * call workitem_alloc() before entering the locked section as 5039 * it also acquires the lock and we must avoid trying doing so 5040 * recursively. 5041 */ 5042 bmsafemap = malloc(sizeof(struct bmsafemap), 5043 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5044 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5045 ACQUIRE_LOCK(ip->i_ump); 5046 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5047 panic("softdep_setup_inomapdep: dependency %p for new" 5048 "inode already exists", inodedep); 5049 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5050 if (jaddref) { 5051 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5052 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5053 if_deps); 5054 } else { 5055 inodedep->id_state |= ONDEPLIST; 5056 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5057 } 5058 inodedep->id_bmsafemap = bmsafemap; 5059 inodedep->id_state &= ~DEPCOMPLETE; 5060 FREE_LOCK(ip->i_ump); 5061 } 5062 5063 /* 5064 * Called just after updating the cylinder group block to 5065 * allocate block or fragment. 5066 */ 5067 void 5068 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5069 struct buf *bp; /* buffer for cylgroup block with block map */ 5070 struct mount *mp; /* filesystem doing allocation */ 5071 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5072 int frags; /* Number of fragments. */ 5073 int oldfrags; /* Previous number of fragments for extend. */ 5074 { 5075 struct newblk *newblk; 5076 struct bmsafemap *bmsafemap; 5077 struct jnewblk *jnewblk; 5078 struct ufsmount *ump; 5079 struct fs *fs; 5080 5081 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5082 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5083 ump = VFSTOUFS(mp); 5084 fs = ump->um_fs; 5085 jnewblk = NULL; 5086 /* 5087 * Create a dependency for the newly allocated block. 5088 * Add it to the dependency list for the buffer holding 5089 * the cylinder group map from which it was allocated. 5090 */ 5091 if (MOUNTEDSUJ(mp)) { 5092 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5093 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5094 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5095 jnewblk->jn_state = ATTACHED; 5096 jnewblk->jn_blkno = newblkno; 5097 jnewblk->jn_frags = frags; 5098 jnewblk->jn_oldfrags = oldfrags; 5099 #ifdef SUJ_DEBUG 5100 { 5101 struct cg *cgp; 5102 uint8_t *blksfree; 5103 long bno; 5104 int i; 5105 5106 cgp = (struct cg *)bp->b_data; 5107 blksfree = cg_blksfree(cgp); 5108 bno = dtogd(fs, jnewblk->jn_blkno); 5109 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5110 i++) { 5111 if (isset(blksfree, bno + i)) 5112 panic("softdep_setup_blkmapdep: " 5113 "free fragment %d from %d-%d " 5114 "state 0x%X dep %p", i, 5115 jnewblk->jn_oldfrags, 5116 jnewblk->jn_frags, 5117 jnewblk->jn_state, 5118 jnewblk->jn_dep); 5119 } 5120 } 5121 #endif 5122 } 5123 5124 CTR3(KTR_SUJ, 5125 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5126 newblkno, frags, oldfrags); 5127 ACQUIRE_LOCK(ump); 5128 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5129 panic("softdep_setup_blkmapdep: found block"); 5130 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5131 dtog(fs, newblkno), NULL); 5132 if (jnewblk) { 5133 jnewblk->jn_dep = (struct worklist *)newblk; 5134 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5135 } else { 5136 newblk->nb_state |= ONDEPLIST; 5137 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5138 } 5139 newblk->nb_bmsafemap = bmsafemap; 5140 newblk->nb_jnewblk = jnewblk; 5141 FREE_LOCK(ump); 5142 } 5143 5144 #define BMSAFEMAP_HASH(ump, cg) \ 5145 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5146 5147 static int 5148 bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5149 struct bmsafemap_hashhead *bmsafemaphd; 5150 int cg; 5151 struct bmsafemap **bmsafemapp; 5152 { 5153 struct bmsafemap *bmsafemap; 5154 5155 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5156 if (bmsafemap->sm_cg == cg) 5157 break; 5158 if (bmsafemap) { 5159 *bmsafemapp = bmsafemap; 5160 return (1); 5161 } 5162 *bmsafemapp = NULL; 5163 5164 return (0); 5165 } 5166 5167 /* 5168 * Find the bmsafemap associated with a cylinder group buffer. 5169 * If none exists, create one. The buffer must be locked when 5170 * this routine is called and this routine must be called with 5171 * the softdep lock held. To avoid giving up the lock while 5172 * allocating a new bmsafemap, a preallocated bmsafemap may be 5173 * provided. If it is provided but not needed, it is freed. 5174 */ 5175 static struct bmsafemap * 5176 bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5177 struct mount *mp; 5178 struct buf *bp; 5179 int cg; 5180 struct bmsafemap *newbmsafemap; 5181 { 5182 struct bmsafemap_hashhead *bmsafemaphd; 5183 struct bmsafemap *bmsafemap, *collision; 5184 struct worklist *wk; 5185 struct ufsmount *ump; 5186 5187 ump = VFSTOUFS(mp); 5188 LOCK_OWNED(ump); 5189 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5190 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5191 if (wk->wk_type == D_BMSAFEMAP) { 5192 if (newbmsafemap) 5193 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5194 return (WK_BMSAFEMAP(wk)); 5195 } 5196 } 5197 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5198 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5199 if (newbmsafemap) 5200 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5201 return (bmsafemap); 5202 } 5203 if (newbmsafemap) { 5204 bmsafemap = newbmsafemap; 5205 } else { 5206 FREE_LOCK(ump); 5207 bmsafemap = malloc(sizeof(struct bmsafemap), 5208 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5209 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5210 ACQUIRE_LOCK(ump); 5211 } 5212 bmsafemap->sm_buf = bp; 5213 LIST_INIT(&bmsafemap->sm_inodedephd); 5214 LIST_INIT(&bmsafemap->sm_inodedepwr); 5215 LIST_INIT(&bmsafemap->sm_newblkhd); 5216 LIST_INIT(&bmsafemap->sm_newblkwr); 5217 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5218 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5219 LIST_INIT(&bmsafemap->sm_freehd); 5220 LIST_INIT(&bmsafemap->sm_freewr); 5221 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5222 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5223 return (collision); 5224 } 5225 bmsafemap->sm_cg = cg; 5226 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5227 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5228 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5229 return (bmsafemap); 5230 } 5231 5232 /* 5233 * Direct block allocation dependencies. 5234 * 5235 * When a new block is allocated, the corresponding disk locations must be 5236 * initialized (with zeros or new data) before the on-disk inode points to 5237 * them. Also, the freemap from which the block was allocated must be 5238 * updated (on disk) before the inode's pointer. These two dependencies are 5239 * independent of each other and are needed for all file blocks and indirect 5240 * blocks that are pointed to directly by the inode. Just before the 5241 * "in-core" version of the inode is updated with a newly allocated block 5242 * number, a procedure (below) is called to setup allocation dependency 5243 * structures. These structures are removed when the corresponding 5244 * dependencies are satisfied or when the block allocation becomes obsolete 5245 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5246 * fragment that gets upgraded). All of these cases are handled in 5247 * procedures described later. 5248 * 5249 * When a file extension causes a fragment to be upgraded, either to a larger 5250 * fragment or to a full block, the on-disk location may change (if the 5251 * previous fragment could not simply be extended). In this case, the old 5252 * fragment must be de-allocated, but not until after the inode's pointer has 5253 * been updated. In most cases, this is handled by later procedures, which 5254 * will construct a "freefrag" structure to be added to the workitem queue 5255 * when the inode update is complete (or obsolete). The main exception to 5256 * this is when an allocation occurs while a pending allocation dependency 5257 * (for the same block pointer) remains. This case is handled in the main 5258 * allocation dependency setup procedure by immediately freeing the 5259 * unreferenced fragments. 5260 */ 5261 void 5262 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5263 struct inode *ip; /* inode to which block is being added */ 5264 ufs_lbn_t off; /* block pointer within inode */ 5265 ufs2_daddr_t newblkno; /* disk block number being added */ 5266 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5267 long newsize; /* size of new block */ 5268 long oldsize; /* size of new block */ 5269 struct buf *bp; /* bp for allocated block */ 5270 { 5271 struct allocdirect *adp, *oldadp; 5272 struct allocdirectlst *adphead; 5273 struct freefrag *freefrag; 5274 struct inodedep *inodedep; 5275 struct pagedep *pagedep; 5276 struct jnewblk *jnewblk; 5277 struct newblk *newblk; 5278 struct mount *mp; 5279 ufs_lbn_t lbn; 5280 5281 lbn = bp->b_lblkno; 5282 mp = UFSTOVFS(ip->i_ump); 5283 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5284 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5285 if (oldblkno && oldblkno != newblkno) 5286 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5287 else 5288 freefrag = NULL; 5289 5290 CTR6(KTR_SUJ, 5291 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5292 "off %jd newsize %ld oldsize %d", 5293 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5294 ACQUIRE_LOCK(ip->i_ump); 5295 if (off >= NDADDR) { 5296 if (lbn > 0) 5297 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5298 lbn, off); 5299 /* allocating an indirect block */ 5300 if (oldblkno != 0) 5301 panic("softdep_setup_allocdirect: non-zero indir"); 5302 } else { 5303 if (off != lbn) 5304 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5305 lbn, off); 5306 /* 5307 * Allocating a direct block. 5308 * 5309 * If we are allocating a directory block, then we must 5310 * allocate an associated pagedep to track additions and 5311 * deletions. 5312 */ 5313 if ((ip->i_mode & IFMT) == IFDIR) 5314 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5315 &pagedep); 5316 } 5317 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5318 panic("softdep_setup_allocdirect: lost block"); 5319 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5320 ("softdep_setup_allocdirect: newblk already initialized")); 5321 /* 5322 * Convert the newblk to an allocdirect. 5323 */ 5324 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5325 adp = (struct allocdirect *)newblk; 5326 newblk->nb_freefrag = freefrag; 5327 adp->ad_offset = off; 5328 adp->ad_oldblkno = oldblkno; 5329 adp->ad_newsize = newsize; 5330 adp->ad_oldsize = oldsize; 5331 5332 /* 5333 * Finish initializing the journal. 5334 */ 5335 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5336 jnewblk->jn_ino = ip->i_number; 5337 jnewblk->jn_lbn = lbn; 5338 add_to_journal(&jnewblk->jn_list); 5339 } 5340 if (freefrag && freefrag->ff_jdep != NULL && 5341 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5342 add_to_journal(freefrag->ff_jdep); 5343 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5344 adp->ad_inodedep = inodedep; 5345 5346 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5347 /* 5348 * The list of allocdirects must be kept in sorted and ascending 5349 * order so that the rollback routines can quickly determine the 5350 * first uncommitted block (the size of the file stored on disk 5351 * ends at the end of the lowest committed fragment, or if there 5352 * are no fragments, at the end of the highest committed block). 5353 * Since files generally grow, the typical case is that the new 5354 * block is to be added at the end of the list. We speed this 5355 * special case by checking against the last allocdirect in the 5356 * list before laboriously traversing the list looking for the 5357 * insertion point. 5358 */ 5359 adphead = &inodedep->id_newinoupdt; 5360 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5361 if (oldadp == NULL || oldadp->ad_offset <= off) { 5362 /* insert at end of list */ 5363 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5364 if (oldadp != NULL && oldadp->ad_offset == off) 5365 allocdirect_merge(adphead, adp, oldadp); 5366 FREE_LOCK(ip->i_ump); 5367 return; 5368 } 5369 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5370 if (oldadp->ad_offset >= off) 5371 break; 5372 } 5373 if (oldadp == NULL) 5374 panic("softdep_setup_allocdirect: lost entry"); 5375 /* insert in middle of list */ 5376 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5377 if (oldadp->ad_offset == off) 5378 allocdirect_merge(adphead, adp, oldadp); 5379 5380 FREE_LOCK(ip->i_ump); 5381 } 5382 5383 /* 5384 * Merge a newer and older journal record to be stored either in a 5385 * newblock or freefrag. This handles aggregating journal records for 5386 * fragment allocation into a second record as well as replacing a 5387 * journal free with an aborted journal allocation. A segment for the 5388 * oldest record will be placed on wkhd if it has been written. If not 5389 * the segment for the newer record will suffice. 5390 */ 5391 static struct worklist * 5392 jnewblk_merge(new, old, wkhd) 5393 struct worklist *new; 5394 struct worklist *old; 5395 struct workhead *wkhd; 5396 { 5397 struct jnewblk *njnewblk; 5398 struct jnewblk *jnewblk; 5399 5400 /* Handle NULLs to simplify callers. */ 5401 if (new == NULL) 5402 return (old); 5403 if (old == NULL) 5404 return (new); 5405 /* Replace a jfreefrag with a jnewblk. */ 5406 if (new->wk_type == D_JFREEFRAG) { 5407 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5408 panic("jnewblk_merge: blkno mismatch: %p, %p", 5409 old, new); 5410 cancel_jfreefrag(WK_JFREEFRAG(new)); 5411 return (old); 5412 } 5413 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5414 panic("jnewblk_merge: Bad type: old %d new %d\n", 5415 old->wk_type, new->wk_type); 5416 /* 5417 * Handle merging of two jnewblk records that describe 5418 * different sets of fragments in the same block. 5419 */ 5420 jnewblk = WK_JNEWBLK(old); 5421 njnewblk = WK_JNEWBLK(new); 5422 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5423 panic("jnewblk_merge: Merging disparate blocks."); 5424 /* 5425 * The record may be rolled back in the cg. 5426 */ 5427 if (jnewblk->jn_state & UNDONE) { 5428 jnewblk->jn_state &= ~UNDONE; 5429 njnewblk->jn_state |= UNDONE; 5430 njnewblk->jn_state &= ~ATTACHED; 5431 } 5432 /* 5433 * We modify the newer addref and free the older so that if neither 5434 * has been written the most up-to-date copy will be on disk. If 5435 * both have been written but rolled back we only temporarily need 5436 * one of them to fix the bits when the cg write completes. 5437 */ 5438 jnewblk->jn_state |= ATTACHED | COMPLETE; 5439 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5440 cancel_jnewblk(jnewblk, wkhd); 5441 WORKLIST_REMOVE(&jnewblk->jn_list); 5442 free_jnewblk(jnewblk); 5443 return (new); 5444 } 5445 5446 /* 5447 * Replace an old allocdirect dependency with a newer one. 5448 * This routine must be called with splbio interrupts blocked. 5449 */ 5450 static void 5451 allocdirect_merge(adphead, newadp, oldadp) 5452 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5453 struct allocdirect *newadp; /* allocdirect being added */ 5454 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5455 { 5456 struct worklist *wk; 5457 struct freefrag *freefrag; 5458 5459 freefrag = NULL; 5460 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5461 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5462 newadp->ad_oldsize != oldadp->ad_newsize || 5463 newadp->ad_offset >= NDADDR) 5464 panic("%s %jd != new %jd || old size %ld != new %ld", 5465 "allocdirect_merge: old blkno", 5466 (intmax_t)newadp->ad_oldblkno, 5467 (intmax_t)oldadp->ad_newblkno, 5468 newadp->ad_oldsize, oldadp->ad_newsize); 5469 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5470 newadp->ad_oldsize = oldadp->ad_oldsize; 5471 /* 5472 * If the old dependency had a fragment to free or had never 5473 * previously had a block allocated, then the new dependency 5474 * can immediately post its freefrag and adopt the old freefrag. 5475 * This action is done by swapping the freefrag dependencies. 5476 * The new dependency gains the old one's freefrag, and the 5477 * old one gets the new one and then immediately puts it on 5478 * the worklist when it is freed by free_newblk. It is 5479 * not possible to do this swap when the old dependency had a 5480 * non-zero size but no previous fragment to free. This condition 5481 * arises when the new block is an extension of the old block. 5482 * Here, the first part of the fragment allocated to the new 5483 * dependency is part of the block currently claimed on disk by 5484 * the old dependency, so cannot legitimately be freed until the 5485 * conditions for the new dependency are fulfilled. 5486 */ 5487 freefrag = newadp->ad_freefrag; 5488 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5489 newadp->ad_freefrag = oldadp->ad_freefrag; 5490 oldadp->ad_freefrag = freefrag; 5491 } 5492 /* 5493 * If we are tracking a new directory-block allocation, 5494 * move it from the old allocdirect to the new allocdirect. 5495 */ 5496 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5497 WORKLIST_REMOVE(wk); 5498 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5499 panic("allocdirect_merge: extra newdirblk"); 5500 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5501 } 5502 TAILQ_REMOVE(adphead, oldadp, ad_next); 5503 /* 5504 * We need to move any journal dependencies over to the freefrag 5505 * that releases this block if it exists. Otherwise we are 5506 * extending an existing block and we'll wait until that is 5507 * complete to release the journal space and extend the 5508 * new journal to cover this old space as well. 5509 */ 5510 if (freefrag == NULL) { 5511 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5512 panic("allocdirect_merge: %jd != %jd", 5513 oldadp->ad_newblkno, newadp->ad_newblkno); 5514 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5515 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5516 &oldadp->ad_block.nb_jnewblk->jn_list, 5517 &newadp->ad_block.nb_jwork); 5518 oldadp->ad_block.nb_jnewblk = NULL; 5519 cancel_newblk(&oldadp->ad_block, NULL, 5520 &newadp->ad_block.nb_jwork); 5521 } else { 5522 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5523 &freefrag->ff_list, &freefrag->ff_jwork); 5524 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5525 &freefrag->ff_jwork); 5526 } 5527 free_newblk(&oldadp->ad_block); 5528 } 5529 5530 /* 5531 * Allocate a jfreefrag structure to journal a single block free. 5532 */ 5533 static struct jfreefrag * 5534 newjfreefrag(freefrag, ip, blkno, size, lbn) 5535 struct freefrag *freefrag; 5536 struct inode *ip; 5537 ufs2_daddr_t blkno; 5538 long size; 5539 ufs_lbn_t lbn; 5540 { 5541 struct jfreefrag *jfreefrag; 5542 struct fs *fs; 5543 5544 fs = ip->i_fs; 5545 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5546 M_SOFTDEP_FLAGS); 5547 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump)); 5548 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5549 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5550 jfreefrag->fr_ino = ip->i_number; 5551 jfreefrag->fr_lbn = lbn; 5552 jfreefrag->fr_blkno = blkno; 5553 jfreefrag->fr_frags = numfrags(fs, size); 5554 jfreefrag->fr_freefrag = freefrag; 5555 5556 return (jfreefrag); 5557 } 5558 5559 /* 5560 * Allocate a new freefrag structure. 5561 */ 5562 static struct freefrag * 5563 newfreefrag(ip, blkno, size, lbn) 5564 struct inode *ip; 5565 ufs2_daddr_t blkno; 5566 long size; 5567 ufs_lbn_t lbn; 5568 { 5569 struct freefrag *freefrag; 5570 struct fs *fs; 5571 5572 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5573 ip->i_number, blkno, size, lbn); 5574 fs = ip->i_fs; 5575 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5576 panic("newfreefrag: frag size"); 5577 freefrag = malloc(sizeof(struct freefrag), 5578 M_FREEFRAG, M_SOFTDEP_FLAGS); 5579 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump)); 5580 freefrag->ff_state = ATTACHED; 5581 LIST_INIT(&freefrag->ff_jwork); 5582 freefrag->ff_inum = ip->i_number; 5583 freefrag->ff_vtype = ITOV(ip)->v_type; 5584 freefrag->ff_blkno = blkno; 5585 freefrag->ff_fragsize = size; 5586 5587 if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) { 5588 freefrag->ff_jdep = (struct worklist *) 5589 newjfreefrag(freefrag, ip, blkno, size, lbn); 5590 } else { 5591 freefrag->ff_state |= DEPCOMPLETE; 5592 freefrag->ff_jdep = NULL; 5593 } 5594 5595 return (freefrag); 5596 } 5597 5598 /* 5599 * This workitem de-allocates fragments that were replaced during 5600 * file block allocation. 5601 */ 5602 static void 5603 handle_workitem_freefrag(freefrag) 5604 struct freefrag *freefrag; 5605 { 5606 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5607 struct workhead wkhd; 5608 5609 CTR3(KTR_SUJ, 5610 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5611 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5612 /* 5613 * It would be illegal to add new completion items to the 5614 * freefrag after it was schedule to be done so it must be 5615 * safe to modify the list head here. 5616 */ 5617 LIST_INIT(&wkhd); 5618 ACQUIRE_LOCK(ump); 5619 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5620 /* 5621 * If the journal has not been written we must cancel it here. 5622 */ 5623 if (freefrag->ff_jdep) { 5624 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5625 panic("handle_workitem_freefrag: Unexpected type %d\n", 5626 freefrag->ff_jdep->wk_type); 5627 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5628 } 5629 FREE_LOCK(ump); 5630 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5631 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5632 ACQUIRE_LOCK(ump); 5633 WORKITEM_FREE(freefrag, D_FREEFRAG); 5634 FREE_LOCK(ump); 5635 } 5636 5637 /* 5638 * Set up a dependency structure for an external attributes data block. 5639 * This routine follows much of the structure of softdep_setup_allocdirect. 5640 * See the description of softdep_setup_allocdirect above for details. 5641 */ 5642 void 5643 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5644 struct inode *ip; 5645 ufs_lbn_t off; 5646 ufs2_daddr_t newblkno; 5647 ufs2_daddr_t oldblkno; 5648 long newsize; 5649 long oldsize; 5650 struct buf *bp; 5651 { 5652 struct allocdirect *adp, *oldadp; 5653 struct allocdirectlst *adphead; 5654 struct freefrag *freefrag; 5655 struct inodedep *inodedep; 5656 struct jnewblk *jnewblk; 5657 struct newblk *newblk; 5658 struct mount *mp; 5659 ufs_lbn_t lbn; 5660 5661 mp = UFSTOVFS(ip->i_ump); 5662 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5663 ("softdep_setup_allocext called on non-softdep filesystem")); 5664 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR", 5665 (long long)off)); 5666 5667 lbn = bp->b_lblkno; 5668 if (oldblkno && oldblkno != newblkno) 5669 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5670 else 5671 freefrag = NULL; 5672 5673 ACQUIRE_LOCK(ip->i_ump); 5674 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5675 panic("softdep_setup_allocext: lost block"); 5676 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5677 ("softdep_setup_allocext: newblk already initialized")); 5678 /* 5679 * Convert the newblk to an allocdirect. 5680 */ 5681 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5682 adp = (struct allocdirect *)newblk; 5683 newblk->nb_freefrag = freefrag; 5684 adp->ad_offset = off; 5685 adp->ad_oldblkno = oldblkno; 5686 adp->ad_newsize = newsize; 5687 adp->ad_oldsize = oldsize; 5688 adp->ad_state |= EXTDATA; 5689 5690 /* 5691 * Finish initializing the journal. 5692 */ 5693 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5694 jnewblk->jn_ino = ip->i_number; 5695 jnewblk->jn_lbn = lbn; 5696 add_to_journal(&jnewblk->jn_list); 5697 } 5698 if (freefrag && freefrag->ff_jdep != NULL && 5699 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5700 add_to_journal(freefrag->ff_jdep); 5701 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5702 adp->ad_inodedep = inodedep; 5703 5704 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5705 /* 5706 * The list of allocdirects must be kept in sorted and ascending 5707 * order so that the rollback routines can quickly determine the 5708 * first uncommitted block (the size of the file stored on disk 5709 * ends at the end of the lowest committed fragment, or if there 5710 * are no fragments, at the end of the highest committed block). 5711 * Since files generally grow, the typical case is that the new 5712 * block is to be added at the end of the list. We speed this 5713 * special case by checking against the last allocdirect in the 5714 * list before laboriously traversing the list looking for the 5715 * insertion point. 5716 */ 5717 adphead = &inodedep->id_newextupdt; 5718 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5719 if (oldadp == NULL || oldadp->ad_offset <= off) { 5720 /* insert at end of list */ 5721 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5722 if (oldadp != NULL && oldadp->ad_offset == off) 5723 allocdirect_merge(adphead, adp, oldadp); 5724 FREE_LOCK(ip->i_ump); 5725 return; 5726 } 5727 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5728 if (oldadp->ad_offset >= off) 5729 break; 5730 } 5731 if (oldadp == NULL) 5732 panic("softdep_setup_allocext: lost entry"); 5733 /* insert in middle of list */ 5734 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5735 if (oldadp->ad_offset == off) 5736 allocdirect_merge(adphead, adp, oldadp); 5737 FREE_LOCK(ip->i_ump); 5738 } 5739 5740 /* 5741 * Indirect block allocation dependencies. 5742 * 5743 * The same dependencies that exist for a direct block also exist when 5744 * a new block is allocated and pointed to by an entry in a block of 5745 * indirect pointers. The undo/redo states described above are also 5746 * used here. Because an indirect block contains many pointers that 5747 * may have dependencies, a second copy of the entire in-memory indirect 5748 * block is kept. The buffer cache copy is always completely up-to-date. 5749 * The second copy, which is used only as a source for disk writes, 5750 * contains only the safe pointers (i.e., those that have no remaining 5751 * update dependencies). The second copy is freed when all pointers 5752 * are safe. The cache is not allowed to replace indirect blocks with 5753 * pending update dependencies. If a buffer containing an indirect 5754 * block with dependencies is written, these routines will mark it 5755 * dirty again. It can only be successfully written once all the 5756 * dependencies are removed. The ffs_fsync routine in conjunction with 5757 * softdep_sync_metadata work together to get all the dependencies 5758 * removed so that a file can be successfully written to disk. Three 5759 * procedures are used when setting up indirect block pointer 5760 * dependencies. The division is necessary because of the organization 5761 * of the "balloc" routine and because of the distinction between file 5762 * pages and file metadata blocks. 5763 */ 5764 5765 /* 5766 * Allocate a new allocindir structure. 5767 */ 5768 static struct allocindir * 5769 newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5770 struct inode *ip; /* inode for file being extended */ 5771 int ptrno; /* offset of pointer in indirect block */ 5772 ufs2_daddr_t newblkno; /* disk block number being added */ 5773 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5774 ufs_lbn_t lbn; 5775 { 5776 struct newblk *newblk; 5777 struct allocindir *aip; 5778 struct freefrag *freefrag; 5779 struct jnewblk *jnewblk; 5780 5781 if (oldblkno) 5782 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn); 5783 else 5784 freefrag = NULL; 5785 ACQUIRE_LOCK(ip->i_ump); 5786 if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0) 5787 panic("new_allocindir: lost block"); 5788 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5789 ("newallocindir: newblk already initialized")); 5790 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5791 newblk->nb_freefrag = freefrag; 5792 aip = (struct allocindir *)newblk; 5793 aip->ai_offset = ptrno; 5794 aip->ai_oldblkno = oldblkno; 5795 aip->ai_lbn = lbn; 5796 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5797 jnewblk->jn_ino = ip->i_number; 5798 jnewblk->jn_lbn = lbn; 5799 add_to_journal(&jnewblk->jn_list); 5800 } 5801 if (freefrag && freefrag->ff_jdep != NULL && 5802 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5803 add_to_journal(freefrag->ff_jdep); 5804 return (aip); 5805 } 5806 5807 /* 5808 * Called just before setting an indirect block pointer 5809 * to a newly allocated file page. 5810 */ 5811 void 5812 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5813 struct inode *ip; /* inode for file being extended */ 5814 ufs_lbn_t lbn; /* allocated block number within file */ 5815 struct buf *bp; /* buffer with indirect blk referencing page */ 5816 int ptrno; /* offset of pointer in indirect block */ 5817 ufs2_daddr_t newblkno; /* disk block number being added */ 5818 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5819 struct buf *nbp; /* buffer holding allocated page */ 5820 { 5821 struct inodedep *inodedep; 5822 struct freefrag *freefrag; 5823 struct allocindir *aip; 5824 struct pagedep *pagedep; 5825 struct mount *mp; 5826 5827 mp = UFSTOVFS(ip->i_ump); 5828 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5829 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5830 KASSERT(lbn == nbp->b_lblkno, 5831 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5832 lbn, bp->b_lblkno)); 5833 CTR4(KTR_SUJ, 5834 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5835 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5836 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5837 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5838 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5839 /* 5840 * If we are allocating a directory page, then we must 5841 * allocate an associated pagedep to track additions and 5842 * deletions. 5843 */ 5844 if ((ip->i_mode & IFMT) == IFDIR) 5845 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5846 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5847 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5848 FREE_LOCK(ip->i_ump); 5849 if (freefrag) 5850 handle_workitem_freefrag(freefrag); 5851 } 5852 5853 /* 5854 * Called just before setting an indirect block pointer to a 5855 * newly allocated indirect block. 5856 */ 5857 void 5858 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5859 struct buf *nbp; /* newly allocated indirect block */ 5860 struct inode *ip; /* inode for file being extended */ 5861 struct buf *bp; /* indirect block referencing allocated block */ 5862 int ptrno; /* offset of pointer in indirect block */ 5863 ufs2_daddr_t newblkno; /* disk block number being added */ 5864 { 5865 struct inodedep *inodedep; 5866 struct allocindir *aip; 5867 ufs_lbn_t lbn; 5868 5869 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 5870 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5871 CTR3(KTR_SUJ, 5872 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5873 ip->i_number, newblkno, ptrno); 5874 lbn = nbp->b_lblkno; 5875 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5876 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5877 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, 5878 &inodedep); 5879 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5880 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5881 panic("softdep_setup_allocindir_meta: Block already existed"); 5882 FREE_LOCK(ip->i_ump); 5883 } 5884 5885 static void 5886 indirdep_complete(indirdep) 5887 struct indirdep *indirdep; 5888 { 5889 struct allocindir *aip; 5890 5891 LIST_REMOVE(indirdep, ir_next); 5892 indirdep->ir_state |= DEPCOMPLETE; 5893 5894 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5895 LIST_REMOVE(aip, ai_next); 5896 free_newblk(&aip->ai_block); 5897 } 5898 /* 5899 * If this indirdep is not attached to a buf it was simply waiting 5900 * on completion to clear completehd. free_indirdep() asserts 5901 * that nothing is dangling. 5902 */ 5903 if ((indirdep->ir_state & ONWORKLIST) == 0) 5904 free_indirdep(indirdep); 5905 } 5906 5907 static struct indirdep * 5908 indirdep_lookup(mp, ip, bp) 5909 struct mount *mp; 5910 struct inode *ip; 5911 struct buf *bp; 5912 { 5913 struct indirdep *indirdep, *newindirdep; 5914 struct newblk *newblk; 5915 struct ufsmount *ump; 5916 struct worklist *wk; 5917 struct fs *fs; 5918 ufs2_daddr_t blkno; 5919 5920 ump = VFSTOUFS(mp); 5921 LOCK_OWNED(ump); 5922 indirdep = NULL; 5923 newindirdep = NULL; 5924 fs = ip->i_fs; 5925 for (;;) { 5926 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5927 if (wk->wk_type != D_INDIRDEP) 5928 continue; 5929 indirdep = WK_INDIRDEP(wk); 5930 break; 5931 } 5932 /* Found on the buffer worklist, no new structure to free. */ 5933 if (indirdep != NULL && newindirdep == NULL) 5934 return (indirdep); 5935 if (indirdep != NULL && newindirdep != NULL) 5936 panic("indirdep_lookup: simultaneous create"); 5937 /* None found on the buffer and a new structure is ready. */ 5938 if (indirdep == NULL && newindirdep != NULL) 5939 break; 5940 /* None found and no new structure available. */ 5941 FREE_LOCK(ump); 5942 newindirdep = malloc(sizeof(struct indirdep), 5943 M_INDIRDEP, M_SOFTDEP_FLAGS); 5944 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5945 newindirdep->ir_state = ATTACHED; 5946 if (ip->i_ump->um_fstype == UFS1) 5947 newindirdep->ir_state |= UFS1FMT; 5948 TAILQ_INIT(&newindirdep->ir_trunc); 5949 newindirdep->ir_saveddata = NULL; 5950 LIST_INIT(&newindirdep->ir_deplisthd); 5951 LIST_INIT(&newindirdep->ir_donehd); 5952 LIST_INIT(&newindirdep->ir_writehd); 5953 LIST_INIT(&newindirdep->ir_completehd); 5954 if (bp->b_blkno == bp->b_lblkno) { 5955 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5956 NULL, NULL); 5957 bp->b_blkno = blkno; 5958 } 5959 newindirdep->ir_freeblks = NULL; 5960 newindirdep->ir_savebp = 5961 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5962 newindirdep->ir_bp = bp; 5963 BUF_KERNPROC(newindirdep->ir_savebp); 5964 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5965 ACQUIRE_LOCK(ump); 5966 } 5967 indirdep = newindirdep; 5968 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5969 /* 5970 * If the block is not yet allocated we don't set DEPCOMPLETE so 5971 * that we don't free dependencies until the pointers are valid. 5972 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5973 * than using the hash. 5974 */ 5975 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5976 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5977 else 5978 indirdep->ir_state |= DEPCOMPLETE; 5979 return (indirdep); 5980 } 5981 5982 /* 5983 * Called to finish the allocation of the "aip" allocated 5984 * by one of the two routines above. 5985 */ 5986 static struct freefrag * 5987 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 5988 struct buf *bp; /* in-memory copy of the indirect block */ 5989 struct inode *ip; /* inode for file being extended */ 5990 struct inodedep *inodedep; /* Inodedep for ip */ 5991 struct allocindir *aip; /* allocindir allocated by the above routines */ 5992 ufs_lbn_t lbn; /* Logical block number for this block. */ 5993 { 5994 struct fs *fs; 5995 struct indirdep *indirdep; 5996 struct allocindir *oldaip; 5997 struct freefrag *freefrag; 5998 struct mount *mp; 5999 6000 LOCK_OWNED(ip->i_ump); 6001 mp = UFSTOVFS(ip->i_ump); 6002 fs = ip->i_fs; 6003 if (bp->b_lblkno >= 0) 6004 panic("setup_allocindir_phase2: not indir blk"); 6005 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6006 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6007 indirdep = indirdep_lookup(mp, ip, bp); 6008 KASSERT(indirdep->ir_savebp != NULL, 6009 ("setup_allocindir_phase2 NULL ir_savebp")); 6010 aip->ai_indirdep = indirdep; 6011 /* 6012 * Check for an unwritten dependency for this indirect offset. If 6013 * there is, merge the old dependency into the new one. This happens 6014 * as a result of reallocblk only. 6015 */ 6016 freefrag = NULL; 6017 if (aip->ai_oldblkno != 0) { 6018 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6019 if (oldaip->ai_offset == aip->ai_offset) { 6020 freefrag = allocindir_merge(aip, oldaip); 6021 goto done; 6022 } 6023 } 6024 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6025 if (oldaip->ai_offset == aip->ai_offset) { 6026 freefrag = allocindir_merge(aip, oldaip); 6027 goto done; 6028 } 6029 } 6030 } 6031 done: 6032 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6033 return (freefrag); 6034 } 6035 6036 /* 6037 * Merge two allocindirs which refer to the same block. Move newblock 6038 * dependencies and setup the freefrags appropriately. 6039 */ 6040 static struct freefrag * 6041 allocindir_merge(aip, oldaip) 6042 struct allocindir *aip; 6043 struct allocindir *oldaip; 6044 { 6045 struct freefrag *freefrag; 6046 struct worklist *wk; 6047 6048 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6049 panic("allocindir_merge: blkno"); 6050 aip->ai_oldblkno = oldaip->ai_oldblkno; 6051 freefrag = aip->ai_freefrag; 6052 aip->ai_freefrag = oldaip->ai_freefrag; 6053 oldaip->ai_freefrag = NULL; 6054 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6055 /* 6056 * If we are tracking a new directory-block allocation, 6057 * move it from the old allocindir to the new allocindir. 6058 */ 6059 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6060 WORKLIST_REMOVE(wk); 6061 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6062 panic("allocindir_merge: extra newdirblk"); 6063 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6064 } 6065 /* 6066 * We can skip journaling for this freefrag and just complete 6067 * any pending journal work for the allocindir that is being 6068 * removed after the freefrag completes. 6069 */ 6070 if (freefrag->ff_jdep) 6071 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6072 LIST_REMOVE(oldaip, ai_next); 6073 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6074 &freefrag->ff_list, &freefrag->ff_jwork); 6075 free_newblk(&oldaip->ai_block); 6076 6077 return (freefrag); 6078 } 6079 6080 static inline void 6081 setup_freedirect(freeblks, ip, i, needj) 6082 struct freeblks *freeblks; 6083 struct inode *ip; 6084 int i; 6085 int needj; 6086 { 6087 ufs2_daddr_t blkno; 6088 int frags; 6089 6090 blkno = DIP(ip, i_db[i]); 6091 if (blkno == 0) 6092 return; 6093 DIP_SET(ip, i_db[i], 0); 6094 frags = sblksize(ip->i_fs, ip->i_size, i); 6095 frags = numfrags(ip->i_fs, frags); 6096 newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj); 6097 } 6098 6099 static inline void 6100 setup_freeext(freeblks, ip, i, needj) 6101 struct freeblks *freeblks; 6102 struct inode *ip; 6103 int i; 6104 int needj; 6105 { 6106 ufs2_daddr_t blkno; 6107 int frags; 6108 6109 blkno = ip->i_din2->di_extb[i]; 6110 if (blkno == 0) 6111 return; 6112 ip->i_din2->di_extb[i] = 0; 6113 frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i); 6114 frags = numfrags(ip->i_fs, frags); 6115 newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6116 } 6117 6118 static inline void 6119 setup_freeindir(freeblks, ip, i, lbn, needj) 6120 struct freeblks *freeblks; 6121 struct inode *ip; 6122 int i; 6123 ufs_lbn_t lbn; 6124 int needj; 6125 { 6126 ufs2_daddr_t blkno; 6127 6128 blkno = DIP(ip, i_ib[i]); 6129 if (blkno == 0) 6130 return; 6131 DIP_SET(ip, i_ib[i], 0); 6132 newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag, 6133 0, needj); 6134 } 6135 6136 static inline struct freeblks * 6137 newfreeblks(mp, ip) 6138 struct mount *mp; 6139 struct inode *ip; 6140 { 6141 struct freeblks *freeblks; 6142 6143 freeblks = malloc(sizeof(struct freeblks), 6144 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6145 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6146 LIST_INIT(&freeblks->fb_jblkdephd); 6147 LIST_INIT(&freeblks->fb_jwork); 6148 freeblks->fb_ref = 0; 6149 freeblks->fb_cgwait = 0; 6150 freeblks->fb_state = ATTACHED; 6151 freeblks->fb_uid = ip->i_uid; 6152 freeblks->fb_inum = ip->i_number; 6153 freeblks->fb_vtype = ITOV(ip)->v_type; 6154 freeblks->fb_modrev = DIP(ip, i_modrev); 6155 freeblks->fb_devvp = ip->i_devvp; 6156 freeblks->fb_chkcnt = 0; 6157 freeblks->fb_len = 0; 6158 6159 return (freeblks); 6160 } 6161 6162 static void 6163 trunc_indirdep(indirdep, freeblks, bp, off) 6164 struct indirdep *indirdep; 6165 struct freeblks *freeblks; 6166 struct buf *bp; 6167 int off; 6168 { 6169 struct allocindir *aip, *aipn; 6170 6171 /* 6172 * The first set of allocindirs won't be in savedbp. 6173 */ 6174 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6175 if (aip->ai_offset > off) 6176 cancel_allocindir(aip, bp, freeblks, 1); 6177 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6178 if (aip->ai_offset > off) 6179 cancel_allocindir(aip, bp, freeblks, 1); 6180 /* 6181 * These will exist in savedbp. 6182 */ 6183 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6184 if (aip->ai_offset > off) 6185 cancel_allocindir(aip, NULL, freeblks, 0); 6186 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6187 if (aip->ai_offset > off) 6188 cancel_allocindir(aip, NULL, freeblks, 0); 6189 } 6190 6191 /* 6192 * Follow the chain of indirects down to lastlbn creating a freework 6193 * structure for each. This will be used to start indir_trunc() at 6194 * the right offset and create the journal records for the parrtial 6195 * truncation. A second step will handle the truncated dependencies. 6196 */ 6197 static int 6198 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6199 struct freeblks *freeblks; 6200 struct inode *ip; 6201 ufs_lbn_t lbn; 6202 ufs_lbn_t lastlbn; 6203 ufs2_daddr_t blkno; 6204 { 6205 struct indirdep *indirdep; 6206 struct indirdep *indirn; 6207 struct freework *freework; 6208 struct newblk *newblk; 6209 struct mount *mp; 6210 struct buf *bp; 6211 uint8_t *start; 6212 uint8_t *end; 6213 ufs_lbn_t lbnadd; 6214 int level; 6215 int error; 6216 int off; 6217 6218 6219 freework = NULL; 6220 if (blkno == 0) 6221 return (0); 6222 mp = freeblks->fb_list.wk_mp; 6223 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6224 if ((bp->b_flags & B_CACHE) == 0) { 6225 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6226 bp->b_iocmd = BIO_READ; 6227 bp->b_flags &= ~B_INVAL; 6228 bp->b_ioflags &= ~BIO_ERROR; 6229 vfs_busy_pages(bp, 0); 6230 bp->b_iooffset = dbtob(bp->b_blkno); 6231 bstrategy(bp); 6232 curthread->td_ru.ru_inblock++; 6233 error = bufwait(bp); 6234 if (error) { 6235 brelse(bp); 6236 return (error); 6237 } 6238 } 6239 level = lbn_level(lbn); 6240 lbnadd = lbn_offset(ip->i_fs, level); 6241 /* 6242 * Compute the offset of the last block we want to keep. Store 6243 * in the freework the first block we want to completely free. 6244 */ 6245 off = (lastlbn - -(lbn + level)) / lbnadd; 6246 if (off + 1 == NINDIR(ip->i_fs)) 6247 goto nowork; 6248 freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1, 6249 0); 6250 /* 6251 * Link the freework into the indirdep. This will prevent any new 6252 * allocations from proceeding until we are finished with the 6253 * truncate and the block is written. 6254 */ 6255 ACQUIRE_LOCK(ip->i_ump); 6256 indirdep = indirdep_lookup(mp, ip, bp); 6257 if (indirdep->ir_freeblks) 6258 panic("setup_trunc_indir: indirdep already truncated."); 6259 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6260 freework->fw_indir = indirdep; 6261 /* 6262 * Cancel any allocindirs that will not make it to disk. 6263 * We have to do this for all copies of the indirdep that 6264 * live on this newblk. 6265 */ 6266 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6267 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk); 6268 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6269 trunc_indirdep(indirn, freeblks, bp, off); 6270 } else 6271 trunc_indirdep(indirdep, freeblks, bp, off); 6272 FREE_LOCK(ip->i_ump); 6273 /* 6274 * Creation is protected by the buf lock. The saveddata is only 6275 * needed if a full truncation follows a partial truncation but it 6276 * is difficult to allocate in that case so we fetch it anyway. 6277 */ 6278 if (indirdep->ir_saveddata == NULL) 6279 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6280 M_SOFTDEP_FLAGS); 6281 nowork: 6282 /* Fetch the blkno of the child and the zero start offset. */ 6283 if (ip->i_ump->um_fstype == UFS1) { 6284 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6285 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6286 } else { 6287 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6288 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6289 } 6290 if (freework) { 6291 /* Zero the truncated pointers. */ 6292 end = bp->b_data + bp->b_bcount; 6293 bzero(start, end - start); 6294 bdwrite(bp); 6295 } else 6296 bqrelse(bp); 6297 if (level == 0) 6298 return (0); 6299 lbn++; /* adjust level */ 6300 lbn -= (off * lbnadd); 6301 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6302 } 6303 6304 /* 6305 * Complete the partial truncation of an indirect block setup by 6306 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6307 * copy and writes them to disk before the freeblks is allowed to complete. 6308 */ 6309 static void 6310 complete_trunc_indir(freework) 6311 struct freework *freework; 6312 { 6313 struct freework *fwn; 6314 struct indirdep *indirdep; 6315 struct ufsmount *ump; 6316 struct buf *bp; 6317 uintptr_t start; 6318 int count; 6319 6320 ump = VFSTOUFS(freework->fw_list.wk_mp); 6321 LOCK_OWNED(ump); 6322 indirdep = freework->fw_indir; 6323 for (;;) { 6324 bp = indirdep->ir_bp; 6325 /* See if the block was discarded. */ 6326 if (bp == NULL) 6327 break; 6328 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6329 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6330 break; 6331 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6332 LOCK_PTR(ump)) == 0) 6333 BUF_UNLOCK(bp); 6334 ACQUIRE_LOCK(ump); 6335 } 6336 freework->fw_state |= DEPCOMPLETE; 6337 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6338 /* 6339 * Zero the pointers in the saved copy. 6340 */ 6341 if (indirdep->ir_state & UFS1FMT) 6342 start = sizeof(ufs1_daddr_t); 6343 else 6344 start = sizeof(ufs2_daddr_t); 6345 start *= freework->fw_start; 6346 count = indirdep->ir_savebp->b_bcount - start; 6347 start += (uintptr_t)indirdep->ir_savebp->b_data; 6348 bzero((char *)start, count); 6349 /* 6350 * We need to start the next truncation in the list if it has not 6351 * been started yet. 6352 */ 6353 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6354 if (fwn != NULL) { 6355 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6356 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6357 if ((fwn->fw_state & ONWORKLIST) == 0) 6358 freework_enqueue(fwn); 6359 } 6360 /* 6361 * If bp is NULL the block was fully truncated, restore 6362 * the saved block list otherwise free it if it is no 6363 * longer needed. 6364 */ 6365 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6366 if (bp == NULL) 6367 bcopy(indirdep->ir_saveddata, 6368 indirdep->ir_savebp->b_data, 6369 indirdep->ir_savebp->b_bcount); 6370 free(indirdep->ir_saveddata, M_INDIRDEP); 6371 indirdep->ir_saveddata = NULL; 6372 } 6373 /* 6374 * When bp is NULL there is a full truncation pending. We 6375 * must wait for this full truncation to be journaled before 6376 * we can release this freework because the disk pointers will 6377 * never be written as zero. 6378 */ 6379 if (bp == NULL) { 6380 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6381 handle_written_freework(freework); 6382 else 6383 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6384 &freework->fw_list); 6385 } else { 6386 /* Complete when the real copy is written. */ 6387 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6388 BUF_UNLOCK(bp); 6389 } 6390 } 6391 6392 /* 6393 * Calculate the number of blocks we are going to release where datablocks 6394 * is the current total and length is the new file size. 6395 */ 6396 static ufs2_daddr_t 6397 blkcount(fs, datablocks, length) 6398 struct fs *fs; 6399 ufs2_daddr_t datablocks; 6400 off_t length; 6401 { 6402 off_t totblks, numblks; 6403 6404 totblks = 0; 6405 numblks = howmany(length, fs->fs_bsize); 6406 if (numblks <= NDADDR) { 6407 totblks = howmany(length, fs->fs_fsize); 6408 goto out; 6409 } 6410 totblks = blkstofrags(fs, numblks); 6411 numblks -= NDADDR; 6412 /* 6413 * Count all single, then double, then triple indirects required. 6414 * Subtracting one indirects worth of blocks for each pass 6415 * acknowledges one of each pointed to by the inode. 6416 */ 6417 for (;;) { 6418 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6419 numblks -= NINDIR(fs); 6420 if (numblks <= 0) 6421 break; 6422 numblks = howmany(numblks, NINDIR(fs)); 6423 } 6424 out: 6425 totblks = fsbtodb(fs, totblks); 6426 /* 6427 * Handle sparse files. We can't reclaim more blocks than the inode 6428 * references. We will correct it later in handle_complete_freeblks() 6429 * when we know the real count. 6430 */ 6431 if (totblks > datablocks) 6432 return (0); 6433 return (datablocks - totblks); 6434 } 6435 6436 /* 6437 * Handle freeblocks for journaled softupdate filesystems. 6438 * 6439 * Contrary to normal softupdates, we must preserve the block pointers in 6440 * indirects until their subordinates are free. This is to avoid journaling 6441 * every block that is freed which may consume more space than the journal 6442 * itself. The recovery program will see the free block journals at the 6443 * base of the truncated area and traverse them to reclaim space. The 6444 * pointers in the inode may be cleared immediately after the journal 6445 * records are written because each direct and indirect pointer in the 6446 * inode is recorded in a journal. This permits full truncation to proceed 6447 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6448 * 6449 * The algorithm is as follows: 6450 * 1) Traverse the in-memory state and create journal entries to release 6451 * the relevant blocks and full indirect trees. 6452 * 2) Traverse the indirect block chain adding partial truncation freework 6453 * records to indirects in the path to lastlbn. The freework will 6454 * prevent new allocation dependencies from being satisfied in this 6455 * indirect until the truncation completes. 6456 * 3) Read and lock the inode block, performing an update with the new size 6457 * and pointers. This prevents truncated data from becoming valid on 6458 * disk through step 4. 6459 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6460 * eliminate journal work for those records that do not require it. 6461 * 5) Schedule the journal records to be written followed by the inode block. 6462 * 6) Allocate any necessary frags for the end of file. 6463 * 7) Zero any partially truncated blocks. 6464 * 6465 * From this truncation proceeds asynchronously using the freework and 6466 * indir_trunc machinery. The file will not be extended again into a 6467 * partially truncated indirect block until all work is completed but 6468 * the normal dependency mechanism ensures that it is rolled back/forward 6469 * as appropriate. Further truncation may occur without delay and is 6470 * serialized in indir_trunc(). 6471 */ 6472 void 6473 softdep_journal_freeblocks(ip, cred, length, flags) 6474 struct inode *ip; /* The inode whose length is to be reduced */ 6475 struct ucred *cred; 6476 off_t length; /* The new length for the file */ 6477 int flags; /* IO_EXT and/or IO_NORMAL */ 6478 { 6479 struct freeblks *freeblks, *fbn; 6480 struct worklist *wk, *wkn; 6481 struct inodedep *inodedep; 6482 struct jblkdep *jblkdep; 6483 struct allocdirect *adp, *adpn; 6484 struct ufsmount *ump; 6485 struct fs *fs; 6486 struct buf *bp; 6487 struct vnode *vp; 6488 struct mount *mp; 6489 ufs2_daddr_t extblocks, datablocks; 6490 ufs_lbn_t tmpval, lbn, lastlbn; 6491 int frags, lastoff, iboff, allocblock, needj, error, i; 6492 6493 fs = ip->i_fs; 6494 ump = ip->i_ump; 6495 mp = UFSTOVFS(ump); 6496 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6497 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6498 vp = ITOV(ip); 6499 needj = 1; 6500 iboff = -1; 6501 allocblock = 0; 6502 extblocks = 0; 6503 datablocks = 0; 6504 frags = 0; 6505 freeblks = newfreeblks(mp, ip); 6506 ACQUIRE_LOCK(ump); 6507 /* 6508 * If we're truncating a removed file that will never be written 6509 * we don't need to journal the block frees. The canceled journals 6510 * for the allocations will suffice. 6511 */ 6512 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6513 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6514 length == 0) 6515 needj = 0; 6516 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6517 ip->i_number, length, needj); 6518 FREE_LOCK(ump); 6519 /* 6520 * Calculate the lbn that we are truncating to. This results in -1 6521 * if we're truncating the 0 bytes. So it is the last lbn we want 6522 * to keep, not the first lbn we want to truncate. 6523 */ 6524 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6525 lastoff = blkoff(fs, length); 6526 /* 6527 * Compute frags we are keeping in lastlbn. 0 means all. 6528 */ 6529 if (lastlbn >= 0 && lastlbn < NDADDR) { 6530 frags = fragroundup(fs, lastoff); 6531 /* adp offset of last valid allocdirect. */ 6532 iboff = lastlbn; 6533 } else if (lastlbn > 0) 6534 iboff = NDADDR; 6535 if (fs->fs_magic == FS_UFS2_MAGIC) 6536 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6537 /* 6538 * Handle normal data blocks and indirects. This section saves 6539 * values used after the inode update to complete frag and indirect 6540 * truncation. 6541 */ 6542 if ((flags & IO_NORMAL) != 0) { 6543 /* 6544 * Handle truncation of whole direct and indirect blocks. 6545 */ 6546 for (i = iboff + 1; i < NDADDR; i++) 6547 setup_freedirect(freeblks, ip, i, needj); 6548 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6549 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6550 /* Release a whole indirect tree. */ 6551 if (lbn > lastlbn) { 6552 setup_freeindir(freeblks, ip, i, -lbn -i, 6553 needj); 6554 continue; 6555 } 6556 iboff = i + NDADDR; 6557 /* 6558 * Traverse partially truncated indirect tree. 6559 */ 6560 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6561 setup_trunc_indir(freeblks, ip, -lbn - i, 6562 lastlbn, DIP(ip, i_ib[i])); 6563 } 6564 /* 6565 * Handle partial truncation to a frag boundary. 6566 */ 6567 if (frags) { 6568 ufs2_daddr_t blkno; 6569 long oldfrags; 6570 6571 oldfrags = blksize(fs, ip, lastlbn); 6572 blkno = DIP(ip, i_db[lastlbn]); 6573 if (blkno && oldfrags != frags) { 6574 oldfrags -= frags; 6575 oldfrags = numfrags(ip->i_fs, oldfrags); 6576 blkno += numfrags(ip->i_fs, frags); 6577 newfreework(ump, freeblks, NULL, lastlbn, 6578 blkno, oldfrags, 0, needj); 6579 if (needj) 6580 adjust_newfreework(freeblks, 6581 numfrags(ip->i_fs, frags)); 6582 } else if (blkno == 0) 6583 allocblock = 1; 6584 } 6585 /* 6586 * Add a journal record for partial truncate if we are 6587 * handling indirect blocks. Non-indirects need no extra 6588 * journaling. 6589 */ 6590 if (length != 0 && lastlbn >= NDADDR) { 6591 ip->i_flag |= IN_TRUNCATED; 6592 newjtrunc(freeblks, length, 0); 6593 } 6594 ip->i_size = length; 6595 DIP_SET(ip, i_size, ip->i_size); 6596 datablocks = DIP(ip, i_blocks) - extblocks; 6597 if (length != 0) 6598 datablocks = blkcount(ip->i_fs, datablocks, length); 6599 freeblks->fb_len = length; 6600 } 6601 if ((flags & IO_EXT) != 0) { 6602 for (i = 0; i < NXADDR; i++) 6603 setup_freeext(freeblks, ip, i, needj); 6604 ip->i_din2->di_extsize = 0; 6605 datablocks += extblocks; 6606 } 6607 #ifdef QUOTA 6608 /* Reference the quotas in case the block count is wrong in the end. */ 6609 quotaref(vp, freeblks->fb_quota); 6610 (void) chkdq(ip, -datablocks, NOCRED, 0); 6611 #endif 6612 freeblks->fb_chkcnt = -datablocks; 6613 UFS_LOCK(ump); 6614 fs->fs_pendingblocks += datablocks; 6615 UFS_UNLOCK(ump); 6616 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6617 /* 6618 * Handle truncation of incomplete alloc direct dependencies. We 6619 * hold the inode block locked to prevent incomplete dependencies 6620 * from reaching the disk while we are eliminating those that 6621 * have been truncated. This is a partially inlined ffs_update(). 6622 */ 6623 ufs_itimes(vp); 6624 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6625 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6626 (int)fs->fs_bsize, cred, &bp); 6627 if (error) { 6628 brelse(bp); 6629 softdep_error("softdep_journal_freeblocks", error); 6630 return; 6631 } 6632 if (bp->b_bufsize == fs->fs_bsize) 6633 bp->b_flags |= B_CLUSTEROK; 6634 softdep_update_inodeblock(ip, bp, 0); 6635 if (ump->um_fstype == UFS1) 6636 *((struct ufs1_dinode *)bp->b_data + 6637 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6638 else 6639 *((struct ufs2_dinode *)bp->b_data + 6640 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6641 ACQUIRE_LOCK(ump); 6642 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6643 if ((inodedep->id_state & IOSTARTED) != 0) 6644 panic("softdep_setup_freeblocks: inode busy"); 6645 /* 6646 * Add the freeblks structure to the list of operations that 6647 * must await the zero'ed inode being written to disk. If we 6648 * still have a bitmap dependency (needj), then the inode 6649 * has never been written to disk, so we can process the 6650 * freeblks below once we have deleted the dependencies. 6651 */ 6652 if (needj) 6653 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6654 else 6655 freeblks->fb_state |= COMPLETE; 6656 if ((flags & IO_NORMAL) != 0) { 6657 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6658 if (adp->ad_offset > iboff) 6659 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6660 freeblks); 6661 /* 6662 * Truncate the allocdirect. We could eliminate 6663 * or modify journal records as well. 6664 */ 6665 else if (adp->ad_offset == iboff && frags) 6666 adp->ad_newsize = frags; 6667 } 6668 } 6669 if ((flags & IO_EXT) != 0) 6670 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6671 cancel_allocdirect(&inodedep->id_extupdt, adp, 6672 freeblks); 6673 /* 6674 * Scan the bufwait list for newblock dependencies that will never 6675 * make it to disk. 6676 */ 6677 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6678 if (wk->wk_type != D_ALLOCDIRECT) 6679 continue; 6680 adp = WK_ALLOCDIRECT(wk); 6681 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6682 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6683 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6684 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6685 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6686 } 6687 } 6688 /* 6689 * Add journal work. 6690 */ 6691 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6692 add_to_journal(&jblkdep->jb_list); 6693 FREE_LOCK(ump); 6694 bdwrite(bp); 6695 /* 6696 * Truncate dependency structures beyond length. 6697 */ 6698 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6699 /* 6700 * This is only set when we need to allocate a fragment because 6701 * none existed at the end of a frag-sized file. It handles only 6702 * allocating a new, zero filled block. 6703 */ 6704 if (allocblock) { 6705 ip->i_size = length - lastoff; 6706 DIP_SET(ip, i_size, ip->i_size); 6707 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6708 if (error != 0) { 6709 softdep_error("softdep_journal_freeblks", error); 6710 return; 6711 } 6712 ip->i_size = length; 6713 DIP_SET(ip, i_size, length); 6714 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6715 allocbuf(bp, frags); 6716 ffs_update(vp, 0); 6717 bawrite(bp); 6718 } else if (lastoff != 0 && vp->v_type != VDIR) { 6719 int size; 6720 6721 /* 6722 * Zero the end of a truncated frag or block. 6723 */ 6724 size = sblksize(fs, length, lastlbn); 6725 error = bread(vp, lastlbn, size, cred, &bp); 6726 if (error) { 6727 softdep_error("softdep_journal_freeblks", error); 6728 return; 6729 } 6730 bzero((char *)bp->b_data + lastoff, size - lastoff); 6731 bawrite(bp); 6732 6733 } 6734 ACQUIRE_LOCK(ump); 6735 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6736 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6737 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6738 /* 6739 * We zero earlier truncations so they don't erroneously 6740 * update i_blocks. 6741 */ 6742 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6743 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6744 fbn->fb_len = 0; 6745 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6746 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6747 freeblks->fb_state |= INPROGRESS; 6748 else 6749 freeblks = NULL; 6750 FREE_LOCK(ump); 6751 if (freeblks) 6752 handle_workitem_freeblocks(freeblks, 0); 6753 trunc_pages(ip, length, extblocks, flags); 6754 6755 } 6756 6757 /* 6758 * Flush a JOP_SYNC to the journal. 6759 */ 6760 void 6761 softdep_journal_fsync(ip) 6762 struct inode *ip; 6763 { 6764 struct jfsync *jfsync; 6765 6766 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 6767 ("softdep_journal_fsync called on non-softdep filesystem")); 6768 if ((ip->i_flag & IN_TRUNCATED) == 0) 6769 return; 6770 ip->i_flag &= ~IN_TRUNCATED; 6771 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6772 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump)); 6773 jfsync->jfs_size = ip->i_size; 6774 jfsync->jfs_ino = ip->i_number; 6775 ACQUIRE_LOCK(ip->i_ump); 6776 add_to_journal(&jfsync->jfs_list); 6777 jwait(&jfsync->jfs_list, MNT_WAIT); 6778 FREE_LOCK(ip->i_ump); 6779 } 6780 6781 /* 6782 * Block de-allocation dependencies. 6783 * 6784 * When blocks are de-allocated, the on-disk pointers must be nullified before 6785 * the blocks are made available for use by other files. (The true 6786 * requirement is that old pointers must be nullified before new on-disk 6787 * pointers are set. We chose this slightly more stringent requirement to 6788 * reduce complexity.) Our implementation handles this dependency by updating 6789 * the inode (or indirect block) appropriately but delaying the actual block 6790 * de-allocation (i.e., freemap and free space count manipulation) until 6791 * after the updated versions reach stable storage. After the disk is 6792 * updated, the blocks can be safely de-allocated whenever it is convenient. 6793 * This implementation handles only the common case of reducing a file's 6794 * length to zero. Other cases are handled by the conventional synchronous 6795 * write approach. 6796 * 6797 * The ffs implementation with which we worked double-checks 6798 * the state of the block pointers and file size as it reduces 6799 * a file's length. Some of this code is replicated here in our 6800 * soft updates implementation. The freeblks->fb_chkcnt field is 6801 * used to transfer a part of this information to the procedure 6802 * that eventually de-allocates the blocks. 6803 * 6804 * This routine should be called from the routine that shortens 6805 * a file's length, before the inode's size or block pointers 6806 * are modified. It will save the block pointer information for 6807 * later release and zero the inode so that the calling routine 6808 * can release it. 6809 */ 6810 void 6811 softdep_setup_freeblocks(ip, length, flags) 6812 struct inode *ip; /* The inode whose length is to be reduced */ 6813 off_t length; /* The new length for the file */ 6814 int flags; /* IO_EXT and/or IO_NORMAL */ 6815 { 6816 struct ufs1_dinode *dp1; 6817 struct ufs2_dinode *dp2; 6818 struct freeblks *freeblks; 6819 struct inodedep *inodedep; 6820 struct allocdirect *adp; 6821 struct ufsmount *ump; 6822 struct buf *bp; 6823 struct fs *fs; 6824 ufs2_daddr_t extblocks, datablocks; 6825 struct mount *mp; 6826 int i, delay, error; 6827 ufs_lbn_t tmpval; 6828 ufs_lbn_t lbn; 6829 6830 ump = ip->i_ump; 6831 mp = UFSTOVFS(ump); 6832 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6833 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6834 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6835 ip->i_number, length); 6836 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6837 fs = ip->i_fs; 6838 if ((error = bread(ip->i_devvp, 6839 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6840 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6841 brelse(bp); 6842 softdep_error("softdep_setup_freeblocks", error); 6843 return; 6844 } 6845 freeblks = newfreeblks(mp, ip); 6846 extblocks = 0; 6847 datablocks = 0; 6848 if (fs->fs_magic == FS_UFS2_MAGIC) 6849 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6850 if ((flags & IO_NORMAL) != 0) { 6851 for (i = 0; i < NDADDR; i++) 6852 setup_freedirect(freeblks, ip, i, 0); 6853 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6854 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6855 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6856 ip->i_size = 0; 6857 DIP_SET(ip, i_size, 0); 6858 datablocks = DIP(ip, i_blocks) - extblocks; 6859 } 6860 if ((flags & IO_EXT) != 0) { 6861 for (i = 0; i < NXADDR; i++) 6862 setup_freeext(freeblks, ip, i, 0); 6863 ip->i_din2->di_extsize = 0; 6864 datablocks += extblocks; 6865 } 6866 #ifdef QUOTA 6867 /* Reference the quotas in case the block count is wrong in the end. */ 6868 quotaref(ITOV(ip), freeblks->fb_quota); 6869 (void) chkdq(ip, -datablocks, NOCRED, 0); 6870 #endif 6871 freeblks->fb_chkcnt = -datablocks; 6872 UFS_LOCK(ump); 6873 fs->fs_pendingblocks += datablocks; 6874 UFS_UNLOCK(ump); 6875 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6876 /* 6877 * Push the zero'ed inode to to its disk buffer so that we are free 6878 * to delete its dependencies below. Once the dependencies are gone 6879 * the buffer can be safely released. 6880 */ 6881 if (ump->um_fstype == UFS1) { 6882 dp1 = ((struct ufs1_dinode *)bp->b_data + 6883 ino_to_fsbo(fs, ip->i_number)); 6884 ip->i_din1->di_freelink = dp1->di_freelink; 6885 *dp1 = *ip->i_din1; 6886 } else { 6887 dp2 = ((struct ufs2_dinode *)bp->b_data + 6888 ino_to_fsbo(fs, ip->i_number)); 6889 ip->i_din2->di_freelink = dp2->di_freelink; 6890 *dp2 = *ip->i_din2; 6891 } 6892 /* 6893 * Find and eliminate any inode dependencies. 6894 */ 6895 ACQUIRE_LOCK(ump); 6896 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6897 if ((inodedep->id_state & IOSTARTED) != 0) 6898 panic("softdep_setup_freeblocks: inode busy"); 6899 /* 6900 * Add the freeblks structure to the list of operations that 6901 * must await the zero'ed inode being written to disk. If we 6902 * still have a bitmap dependency (delay == 0), then the inode 6903 * has never been written to disk, so we can process the 6904 * freeblks below once we have deleted the dependencies. 6905 */ 6906 delay = (inodedep->id_state & DEPCOMPLETE); 6907 if (delay) 6908 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6909 else 6910 freeblks->fb_state |= COMPLETE; 6911 /* 6912 * Because the file length has been truncated to zero, any 6913 * pending block allocation dependency structures associated 6914 * with this inode are obsolete and can simply be de-allocated. 6915 * We must first merge the two dependency lists to get rid of 6916 * any duplicate freefrag structures, then purge the merged list. 6917 * If we still have a bitmap dependency, then the inode has never 6918 * been written to disk, so we can free any fragments without delay. 6919 */ 6920 if (flags & IO_NORMAL) { 6921 merge_inode_lists(&inodedep->id_newinoupdt, 6922 &inodedep->id_inoupdt); 6923 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0) 6924 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6925 freeblks); 6926 } 6927 if (flags & IO_EXT) { 6928 merge_inode_lists(&inodedep->id_newextupdt, 6929 &inodedep->id_extupdt); 6930 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0) 6931 cancel_allocdirect(&inodedep->id_extupdt, adp, 6932 freeblks); 6933 } 6934 FREE_LOCK(ump); 6935 bdwrite(bp); 6936 trunc_dependencies(ip, freeblks, -1, 0, flags); 6937 ACQUIRE_LOCK(ump); 6938 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6939 (void) free_inodedep(inodedep); 6940 freeblks->fb_state |= DEPCOMPLETE; 6941 /* 6942 * If the inode with zeroed block pointers is now on disk 6943 * we can start freeing blocks. 6944 */ 6945 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6946 freeblks->fb_state |= INPROGRESS; 6947 else 6948 freeblks = NULL; 6949 FREE_LOCK(ump); 6950 if (freeblks) 6951 handle_workitem_freeblocks(freeblks, 0); 6952 trunc_pages(ip, length, extblocks, flags); 6953 } 6954 6955 /* 6956 * Eliminate pages from the page cache that back parts of this inode and 6957 * adjust the vnode pager's idea of our size. This prevents stale data 6958 * from hanging around in the page cache. 6959 */ 6960 static void 6961 trunc_pages(ip, length, extblocks, flags) 6962 struct inode *ip; 6963 off_t length; 6964 ufs2_daddr_t extblocks; 6965 int flags; 6966 { 6967 struct vnode *vp; 6968 struct fs *fs; 6969 ufs_lbn_t lbn; 6970 off_t end, extend; 6971 6972 vp = ITOV(ip); 6973 fs = ip->i_fs; 6974 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 6975 if ((flags & IO_EXT) != 0) 6976 vn_pages_remove(vp, extend, 0); 6977 if ((flags & IO_NORMAL) == 0) 6978 return; 6979 BO_LOCK(&vp->v_bufobj); 6980 drain_output(vp); 6981 BO_UNLOCK(&vp->v_bufobj); 6982 /* 6983 * The vnode pager eliminates file pages we eliminate indirects 6984 * below. 6985 */ 6986 vnode_pager_setsize(vp, length); 6987 /* 6988 * Calculate the end based on the last indirect we want to keep. If 6989 * the block extends into indirects we can just use the negative of 6990 * its lbn. Doubles and triples exist at lower numbers so we must 6991 * be careful not to remove those, if they exist. double and triple 6992 * indirect lbns do not overlap with others so it is not important 6993 * to verify how many levels are required. 6994 */ 6995 lbn = lblkno(fs, length); 6996 if (lbn >= NDADDR) { 6997 /* Calculate the virtual lbn of the triple indirect. */ 6998 lbn = -lbn - (NIADDR - 1); 6999 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7000 } else 7001 end = extend; 7002 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7003 } 7004 7005 /* 7006 * See if the buf bp is in the range eliminated by truncation. 7007 */ 7008 static int 7009 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7010 struct buf *bp; 7011 int *blkoffp; 7012 ufs_lbn_t lastlbn; 7013 int lastoff; 7014 int flags; 7015 { 7016 ufs_lbn_t lbn; 7017 7018 *blkoffp = 0; 7019 /* Only match ext/normal blocks as appropriate. */ 7020 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7021 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7022 return (0); 7023 /* ALTDATA is always a full truncation. */ 7024 if ((bp->b_xflags & BX_ALTDATA) != 0) 7025 return (1); 7026 /* -1 is full truncation. */ 7027 if (lastlbn == -1) 7028 return (1); 7029 /* 7030 * If this is a partial truncate we only want those 7031 * blocks and indirect blocks that cover the range 7032 * we're after. 7033 */ 7034 lbn = bp->b_lblkno; 7035 if (lbn < 0) 7036 lbn = -(lbn + lbn_level(lbn)); 7037 if (lbn < lastlbn) 7038 return (0); 7039 /* Here we only truncate lblkno if it's partial. */ 7040 if (lbn == lastlbn) { 7041 if (lastoff == 0) 7042 return (0); 7043 *blkoffp = lastoff; 7044 } 7045 return (1); 7046 } 7047 7048 /* 7049 * Eliminate any dependencies that exist in memory beyond lblkno:off 7050 */ 7051 static void 7052 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7053 struct inode *ip; 7054 struct freeblks *freeblks; 7055 ufs_lbn_t lastlbn; 7056 int lastoff; 7057 int flags; 7058 { 7059 struct bufobj *bo; 7060 struct vnode *vp; 7061 struct buf *bp; 7062 int blkoff; 7063 7064 /* 7065 * We must wait for any I/O in progress to finish so that 7066 * all potential buffers on the dirty list will be visible. 7067 * Once they are all there, walk the list and get rid of 7068 * any dependencies. 7069 */ 7070 vp = ITOV(ip); 7071 bo = &vp->v_bufobj; 7072 BO_LOCK(bo); 7073 drain_output(vp); 7074 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7075 bp->b_vflags &= ~BV_SCANNED; 7076 restart: 7077 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7078 if (bp->b_vflags & BV_SCANNED) 7079 continue; 7080 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7081 bp->b_vflags |= BV_SCANNED; 7082 continue; 7083 } 7084 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7085 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7086 goto restart; 7087 BO_UNLOCK(bo); 7088 if (deallocate_dependencies(bp, freeblks, blkoff)) 7089 bqrelse(bp); 7090 else 7091 brelse(bp); 7092 BO_LOCK(bo); 7093 goto restart; 7094 } 7095 /* 7096 * Now do the work of vtruncbuf while also matching indirect blocks. 7097 */ 7098 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7099 bp->b_vflags &= ~BV_SCANNED; 7100 cleanrestart: 7101 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7102 if (bp->b_vflags & BV_SCANNED) 7103 continue; 7104 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7105 bp->b_vflags |= BV_SCANNED; 7106 continue; 7107 } 7108 if (BUF_LOCK(bp, 7109 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7110 BO_LOCKPTR(bo)) == ENOLCK) { 7111 BO_LOCK(bo); 7112 goto cleanrestart; 7113 } 7114 bp->b_vflags |= BV_SCANNED; 7115 bremfree(bp); 7116 if (blkoff != 0) { 7117 allocbuf(bp, blkoff); 7118 bqrelse(bp); 7119 } else { 7120 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7121 brelse(bp); 7122 } 7123 BO_LOCK(bo); 7124 goto cleanrestart; 7125 } 7126 drain_output(vp); 7127 BO_UNLOCK(bo); 7128 } 7129 7130 static int 7131 cancel_pagedep(pagedep, freeblks, blkoff) 7132 struct pagedep *pagedep; 7133 struct freeblks *freeblks; 7134 int blkoff; 7135 { 7136 struct jremref *jremref; 7137 struct jmvref *jmvref; 7138 struct dirrem *dirrem, *tmp; 7139 int i; 7140 7141 /* 7142 * Copy any directory remove dependencies to the list 7143 * to be processed after the freeblks proceeds. If 7144 * directory entry never made it to disk they 7145 * can be dumped directly onto the work list. 7146 */ 7147 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7148 /* Skip this directory removal if it is intended to remain. */ 7149 if (dirrem->dm_offset < blkoff) 7150 continue; 7151 /* 7152 * If there are any dirrems we wait for the journal write 7153 * to complete and then restart the buf scan as the lock 7154 * has been dropped. 7155 */ 7156 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7157 jwait(&jremref->jr_list, MNT_WAIT); 7158 return (ERESTART); 7159 } 7160 LIST_REMOVE(dirrem, dm_next); 7161 dirrem->dm_dirinum = pagedep->pd_ino; 7162 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7163 } 7164 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7165 jwait(&jmvref->jm_list, MNT_WAIT); 7166 return (ERESTART); 7167 } 7168 /* 7169 * When we're partially truncating a pagedep we just want to flush 7170 * journal entries and return. There can not be any adds in the 7171 * truncated portion of the directory and newblk must remain if 7172 * part of the block remains. 7173 */ 7174 if (blkoff != 0) { 7175 struct diradd *dap; 7176 7177 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7178 if (dap->da_offset > blkoff) 7179 panic("cancel_pagedep: diradd %p off %d > %d", 7180 dap, dap->da_offset, blkoff); 7181 for (i = 0; i < DAHASHSZ; i++) 7182 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7183 if (dap->da_offset > blkoff) 7184 panic("cancel_pagedep: diradd %p off %d > %d", 7185 dap, dap->da_offset, blkoff); 7186 return (0); 7187 } 7188 /* 7189 * There should be no directory add dependencies present 7190 * as the directory could not be truncated until all 7191 * children were removed. 7192 */ 7193 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7194 ("deallocate_dependencies: pendinghd != NULL")); 7195 for (i = 0; i < DAHASHSZ; i++) 7196 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7197 ("deallocate_dependencies: diraddhd != NULL")); 7198 if ((pagedep->pd_state & NEWBLOCK) != 0) 7199 free_newdirblk(pagedep->pd_newdirblk); 7200 if (free_pagedep(pagedep) == 0) 7201 panic("Failed to free pagedep %p", pagedep); 7202 return (0); 7203 } 7204 7205 /* 7206 * Reclaim any dependency structures from a buffer that is about to 7207 * be reallocated to a new vnode. The buffer must be locked, thus, 7208 * no I/O completion operations can occur while we are manipulating 7209 * its associated dependencies. The mutex is held so that other I/O's 7210 * associated with related dependencies do not occur. 7211 */ 7212 static int 7213 deallocate_dependencies(bp, freeblks, off) 7214 struct buf *bp; 7215 struct freeblks *freeblks; 7216 int off; 7217 { 7218 struct indirdep *indirdep; 7219 struct pagedep *pagedep; 7220 struct worklist *wk, *wkn; 7221 struct ufsmount *ump; 7222 7223 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 7224 goto done; 7225 ump = VFSTOUFS(wk->wk_mp); 7226 ACQUIRE_LOCK(ump); 7227 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7228 switch (wk->wk_type) { 7229 case D_INDIRDEP: 7230 indirdep = WK_INDIRDEP(wk); 7231 if (bp->b_lblkno >= 0 || 7232 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7233 panic("deallocate_dependencies: not indir"); 7234 cancel_indirdep(indirdep, bp, freeblks); 7235 continue; 7236 7237 case D_PAGEDEP: 7238 pagedep = WK_PAGEDEP(wk); 7239 if (cancel_pagedep(pagedep, freeblks, off)) { 7240 FREE_LOCK(ump); 7241 return (ERESTART); 7242 } 7243 continue; 7244 7245 case D_ALLOCINDIR: 7246 /* 7247 * Simply remove the allocindir, we'll find it via 7248 * the indirdep where we can clear pointers if 7249 * needed. 7250 */ 7251 WORKLIST_REMOVE(wk); 7252 continue; 7253 7254 case D_FREEWORK: 7255 /* 7256 * A truncation is waiting for the zero'd pointers 7257 * to be written. It can be freed when the freeblks 7258 * is journaled. 7259 */ 7260 WORKLIST_REMOVE(wk); 7261 wk->wk_state |= ONDEPLIST; 7262 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7263 break; 7264 7265 case D_ALLOCDIRECT: 7266 if (off != 0) 7267 continue; 7268 /* FALLTHROUGH */ 7269 default: 7270 panic("deallocate_dependencies: Unexpected type %s", 7271 TYPENAME(wk->wk_type)); 7272 /* NOTREACHED */ 7273 } 7274 } 7275 FREE_LOCK(ump); 7276 done: 7277 /* 7278 * Don't throw away this buf, we were partially truncating and 7279 * some deps may always remain. 7280 */ 7281 if (off) { 7282 allocbuf(bp, off); 7283 bp->b_vflags |= BV_SCANNED; 7284 return (EBUSY); 7285 } 7286 bp->b_flags |= B_INVAL | B_NOCACHE; 7287 7288 return (0); 7289 } 7290 7291 /* 7292 * An allocdirect is being canceled due to a truncate. We must make sure 7293 * the journal entry is released in concert with the blkfree that releases 7294 * the storage. Completed journal entries must not be released until the 7295 * space is no longer pointed to by the inode or in the bitmap. 7296 */ 7297 static void 7298 cancel_allocdirect(adphead, adp, freeblks) 7299 struct allocdirectlst *adphead; 7300 struct allocdirect *adp; 7301 struct freeblks *freeblks; 7302 { 7303 struct freework *freework; 7304 struct newblk *newblk; 7305 struct worklist *wk; 7306 7307 TAILQ_REMOVE(adphead, adp, ad_next); 7308 newblk = (struct newblk *)adp; 7309 freework = NULL; 7310 /* 7311 * Find the correct freework structure. 7312 */ 7313 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7314 if (wk->wk_type != D_FREEWORK) 7315 continue; 7316 freework = WK_FREEWORK(wk); 7317 if (freework->fw_blkno == newblk->nb_newblkno) 7318 break; 7319 } 7320 if (freework == NULL) 7321 panic("cancel_allocdirect: Freework not found"); 7322 /* 7323 * If a newblk exists at all we still have the journal entry that 7324 * initiated the allocation so we do not need to journal the free. 7325 */ 7326 cancel_jfreeblk(freeblks, freework->fw_blkno); 7327 /* 7328 * If the journal hasn't been written the jnewblk must be passed 7329 * to the call to ffs_blkfree that reclaims the space. We accomplish 7330 * this by linking the journal dependency into the freework to be 7331 * freed when freework_freeblock() is called. If the journal has 7332 * been written we can simply reclaim the journal space when the 7333 * freeblks work is complete. 7334 */ 7335 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7336 &freeblks->fb_jwork); 7337 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7338 } 7339 7340 7341 /* 7342 * Cancel a new block allocation. May be an indirect or direct block. We 7343 * remove it from various lists and return any journal record that needs to 7344 * be resolved by the caller. 7345 * 7346 * A special consideration is made for indirects which were never pointed 7347 * at on disk and will never be found once this block is released. 7348 */ 7349 static struct jnewblk * 7350 cancel_newblk(newblk, wk, wkhd) 7351 struct newblk *newblk; 7352 struct worklist *wk; 7353 struct workhead *wkhd; 7354 { 7355 struct jnewblk *jnewblk; 7356 7357 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7358 7359 newblk->nb_state |= GOINGAWAY; 7360 /* 7361 * Previously we traversed the completedhd on each indirdep 7362 * attached to this newblk to cancel them and gather journal 7363 * work. Since we need only the oldest journal segment and 7364 * the lowest point on the tree will always have the oldest 7365 * journal segment we are free to release the segments 7366 * of any subordinates and may leave the indirdep list to 7367 * indirdep_complete() when this newblk is freed. 7368 */ 7369 if (newblk->nb_state & ONDEPLIST) { 7370 newblk->nb_state &= ~ONDEPLIST; 7371 LIST_REMOVE(newblk, nb_deps); 7372 } 7373 if (newblk->nb_state & ONWORKLIST) 7374 WORKLIST_REMOVE(&newblk->nb_list); 7375 /* 7376 * If the journal entry hasn't been written we save a pointer to 7377 * the dependency that frees it until it is written or the 7378 * superseding operation completes. 7379 */ 7380 jnewblk = newblk->nb_jnewblk; 7381 if (jnewblk != NULL && wk != NULL) { 7382 newblk->nb_jnewblk = NULL; 7383 jnewblk->jn_dep = wk; 7384 } 7385 if (!LIST_EMPTY(&newblk->nb_jwork)) 7386 jwork_move(wkhd, &newblk->nb_jwork); 7387 /* 7388 * When truncating we must free the newdirblk early to remove 7389 * the pagedep from the hash before returning. 7390 */ 7391 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7392 free_newdirblk(WK_NEWDIRBLK(wk)); 7393 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7394 panic("cancel_newblk: extra newdirblk"); 7395 7396 return (jnewblk); 7397 } 7398 7399 /* 7400 * Schedule the freefrag associated with a newblk to be released once 7401 * the pointers are written and the previous block is no longer needed. 7402 */ 7403 static void 7404 newblk_freefrag(newblk) 7405 struct newblk *newblk; 7406 { 7407 struct freefrag *freefrag; 7408 7409 if (newblk->nb_freefrag == NULL) 7410 return; 7411 freefrag = newblk->nb_freefrag; 7412 newblk->nb_freefrag = NULL; 7413 freefrag->ff_state |= COMPLETE; 7414 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7415 add_to_worklist(&freefrag->ff_list, 0); 7416 } 7417 7418 /* 7419 * Free a newblk. Generate a new freefrag work request if appropriate. 7420 * This must be called after the inode pointer and any direct block pointers 7421 * are valid or fully removed via truncate or frag extension. 7422 */ 7423 static void 7424 free_newblk(newblk) 7425 struct newblk *newblk; 7426 { 7427 struct indirdep *indirdep; 7428 struct worklist *wk; 7429 7430 KASSERT(newblk->nb_jnewblk == NULL, 7431 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7432 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7433 ("free_newblk: unclaimed newblk")); 7434 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7435 newblk_freefrag(newblk); 7436 if (newblk->nb_state & ONDEPLIST) 7437 LIST_REMOVE(newblk, nb_deps); 7438 if (newblk->nb_state & ONWORKLIST) 7439 WORKLIST_REMOVE(&newblk->nb_list); 7440 LIST_REMOVE(newblk, nb_hash); 7441 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7442 free_newdirblk(WK_NEWDIRBLK(wk)); 7443 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7444 panic("free_newblk: extra newdirblk"); 7445 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7446 indirdep_complete(indirdep); 7447 handle_jwork(&newblk->nb_jwork); 7448 WORKITEM_FREE(newblk, D_NEWBLK); 7449 } 7450 7451 /* 7452 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7453 * This routine must be called with splbio interrupts blocked. 7454 */ 7455 static void 7456 free_newdirblk(newdirblk) 7457 struct newdirblk *newdirblk; 7458 { 7459 struct pagedep *pagedep; 7460 struct diradd *dap; 7461 struct worklist *wk; 7462 7463 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7464 WORKLIST_REMOVE(&newdirblk->db_list); 7465 /* 7466 * If the pagedep is still linked onto the directory buffer 7467 * dependency chain, then some of the entries on the 7468 * pd_pendinghd list may not be committed to disk yet. In 7469 * this case, we will simply clear the NEWBLOCK flag and 7470 * let the pd_pendinghd list be processed when the pagedep 7471 * is next written. If the pagedep is no longer on the buffer 7472 * dependency chain, then all the entries on the pd_pending 7473 * list are committed to disk and we can free them here. 7474 */ 7475 pagedep = newdirblk->db_pagedep; 7476 pagedep->pd_state &= ~NEWBLOCK; 7477 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7478 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7479 free_diradd(dap, NULL); 7480 /* 7481 * If no dependencies remain, the pagedep will be freed. 7482 */ 7483 free_pagedep(pagedep); 7484 } 7485 /* Should only ever be one item in the list. */ 7486 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7487 WORKLIST_REMOVE(wk); 7488 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7489 } 7490 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7491 } 7492 7493 /* 7494 * Prepare an inode to be freed. The actual free operation is not 7495 * done until the zero'ed inode has been written to disk. 7496 */ 7497 void 7498 softdep_freefile(pvp, ino, mode) 7499 struct vnode *pvp; 7500 ino_t ino; 7501 int mode; 7502 { 7503 struct inode *ip = VTOI(pvp); 7504 struct inodedep *inodedep; 7505 struct freefile *freefile; 7506 struct freeblks *freeblks; 7507 struct ufsmount *ump; 7508 7509 ump = ip->i_ump; 7510 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7511 ("softdep_freefile called on non-softdep filesystem")); 7512 /* 7513 * This sets up the inode de-allocation dependency. 7514 */ 7515 freefile = malloc(sizeof(struct freefile), 7516 M_FREEFILE, M_SOFTDEP_FLAGS); 7517 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7518 freefile->fx_mode = mode; 7519 freefile->fx_oldinum = ino; 7520 freefile->fx_devvp = ip->i_devvp; 7521 LIST_INIT(&freefile->fx_jwork); 7522 UFS_LOCK(ump); 7523 ip->i_fs->fs_pendinginodes += 1; 7524 UFS_UNLOCK(ump); 7525 7526 /* 7527 * If the inodedep does not exist, then the zero'ed inode has 7528 * been written to disk. If the allocated inode has never been 7529 * written to disk, then the on-disk inode is zero'ed. In either 7530 * case we can free the file immediately. If the journal was 7531 * canceled before being written the inode will never make it to 7532 * disk and we must send the canceled journal entrys to 7533 * ffs_freefile() to be cleared in conjunction with the bitmap. 7534 * Any blocks waiting on the inode to write can be safely freed 7535 * here as it will never been written. 7536 */ 7537 ACQUIRE_LOCK(ump); 7538 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7539 if (inodedep) { 7540 /* 7541 * Clear out freeblks that no longer need to reference 7542 * this inode. 7543 */ 7544 while ((freeblks = 7545 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7546 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7547 fb_next); 7548 freeblks->fb_state &= ~ONDEPLIST; 7549 } 7550 /* 7551 * Remove this inode from the unlinked list. 7552 */ 7553 if (inodedep->id_state & UNLINKED) { 7554 /* 7555 * Save the journal work to be freed with the bitmap 7556 * before we clear UNLINKED. Otherwise it can be lost 7557 * if the inode block is written. 7558 */ 7559 handle_bufwait(inodedep, &freefile->fx_jwork); 7560 clear_unlinked_inodedep(inodedep); 7561 /* 7562 * Re-acquire inodedep as we've dropped the 7563 * per-filesystem lock in clear_unlinked_inodedep(). 7564 */ 7565 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7566 } 7567 } 7568 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7569 FREE_LOCK(ump); 7570 handle_workitem_freefile(freefile); 7571 return; 7572 } 7573 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7574 inodedep->id_state |= GOINGAWAY; 7575 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7576 FREE_LOCK(ump); 7577 if (ip->i_number == ino) 7578 ip->i_flag |= IN_MODIFIED; 7579 } 7580 7581 /* 7582 * Check to see if an inode has never been written to disk. If 7583 * so free the inodedep and return success, otherwise return failure. 7584 * This routine must be called with splbio interrupts blocked. 7585 * 7586 * If we still have a bitmap dependency, then the inode has never 7587 * been written to disk. Drop the dependency as it is no longer 7588 * necessary since the inode is being deallocated. We set the 7589 * ALLCOMPLETE flags since the bitmap now properly shows that the 7590 * inode is not allocated. Even if the inode is actively being 7591 * written, it has been rolled back to its zero'ed state, so we 7592 * are ensured that a zero inode is what is on the disk. For short 7593 * lived files, this change will usually result in removing all the 7594 * dependencies from the inode so that it can be freed immediately. 7595 */ 7596 static int 7597 check_inode_unwritten(inodedep) 7598 struct inodedep *inodedep; 7599 { 7600 7601 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7602 7603 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7604 !LIST_EMPTY(&inodedep->id_dirremhd) || 7605 !LIST_EMPTY(&inodedep->id_pendinghd) || 7606 !LIST_EMPTY(&inodedep->id_bufwait) || 7607 !LIST_EMPTY(&inodedep->id_inowait) || 7608 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7609 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7610 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7611 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7612 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7613 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7614 inodedep->id_mkdiradd != NULL || 7615 inodedep->id_nlinkdelta != 0) 7616 return (0); 7617 /* 7618 * Another process might be in initiate_write_inodeblock_ufs[12] 7619 * trying to allocate memory without holding "Softdep Lock". 7620 */ 7621 if ((inodedep->id_state & IOSTARTED) != 0 && 7622 inodedep->id_savedino1 == NULL) 7623 return (0); 7624 7625 if (inodedep->id_state & ONDEPLIST) 7626 LIST_REMOVE(inodedep, id_deps); 7627 inodedep->id_state &= ~ONDEPLIST; 7628 inodedep->id_state |= ALLCOMPLETE; 7629 inodedep->id_bmsafemap = NULL; 7630 if (inodedep->id_state & ONWORKLIST) 7631 WORKLIST_REMOVE(&inodedep->id_list); 7632 if (inodedep->id_savedino1 != NULL) { 7633 free(inodedep->id_savedino1, M_SAVEDINO); 7634 inodedep->id_savedino1 = NULL; 7635 } 7636 if (free_inodedep(inodedep) == 0) 7637 panic("check_inode_unwritten: busy inode"); 7638 return (1); 7639 } 7640 7641 static int 7642 check_inodedep_free(inodedep) 7643 struct inodedep *inodedep; 7644 { 7645 7646 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7647 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7648 !LIST_EMPTY(&inodedep->id_dirremhd) || 7649 !LIST_EMPTY(&inodedep->id_pendinghd) || 7650 !LIST_EMPTY(&inodedep->id_bufwait) || 7651 !LIST_EMPTY(&inodedep->id_inowait) || 7652 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7653 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7654 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7655 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7656 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7657 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7658 inodedep->id_mkdiradd != NULL || 7659 inodedep->id_nlinkdelta != 0 || 7660 inodedep->id_savedino1 != NULL) 7661 return (0); 7662 return (1); 7663 } 7664 7665 /* 7666 * Try to free an inodedep structure. Return 1 if it could be freed. 7667 */ 7668 static int 7669 free_inodedep(inodedep) 7670 struct inodedep *inodedep; 7671 { 7672 7673 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7674 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7675 !check_inodedep_free(inodedep)) 7676 return (0); 7677 if (inodedep->id_state & ONDEPLIST) 7678 LIST_REMOVE(inodedep, id_deps); 7679 LIST_REMOVE(inodedep, id_hash); 7680 WORKITEM_FREE(inodedep, D_INODEDEP); 7681 return (1); 7682 } 7683 7684 /* 7685 * Free the block referenced by a freework structure. The parent freeblks 7686 * structure is released and completed when the final cg bitmap reaches 7687 * the disk. This routine may be freeing a jnewblk which never made it to 7688 * disk in which case we do not have to wait as the operation is undone 7689 * in memory immediately. 7690 */ 7691 static void 7692 freework_freeblock(freework) 7693 struct freework *freework; 7694 { 7695 struct freeblks *freeblks; 7696 struct jnewblk *jnewblk; 7697 struct ufsmount *ump; 7698 struct workhead wkhd; 7699 struct fs *fs; 7700 int bsize; 7701 int needj; 7702 7703 ump = VFSTOUFS(freework->fw_list.wk_mp); 7704 LOCK_OWNED(ump); 7705 /* 7706 * Handle partial truncate separately. 7707 */ 7708 if (freework->fw_indir) { 7709 complete_trunc_indir(freework); 7710 return; 7711 } 7712 freeblks = freework->fw_freeblks; 7713 fs = ump->um_fs; 7714 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7715 bsize = lfragtosize(fs, freework->fw_frags); 7716 LIST_INIT(&wkhd); 7717 /* 7718 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7719 * on the indirblk hashtable and prevents premature freeing. 7720 */ 7721 freework->fw_state |= DEPCOMPLETE; 7722 /* 7723 * SUJ needs to wait for the segment referencing freed indirect 7724 * blocks to expire so that we know the checker will not confuse 7725 * a re-allocated indirect block with its old contents. 7726 */ 7727 if (needj && freework->fw_lbn <= -NDADDR) 7728 indirblk_insert(freework); 7729 /* 7730 * If we are canceling an existing jnewblk pass it to the free 7731 * routine, otherwise pass the freeblk which will ultimately 7732 * release the freeblks. If we're not journaling, we can just 7733 * free the freeblks immediately. 7734 */ 7735 jnewblk = freework->fw_jnewblk; 7736 if (jnewblk != NULL) { 7737 cancel_jnewblk(jnewblk, &wkhd); 7738 needj = 0; 7739 } else if (needj) { 7740 freework->fw_state |= DELAYEDFREE; 7741 freeblks->fb_cgwait++; 7742 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7743 } 7744 FREE_LOCK(ump); 7745 freeblks_free(ump, freeblks, btodb(bsize)); 7746 CTR4(KTR_SUJ, 7747 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld", 7748 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7749 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7750 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7751 ACQUIRE_LOCK(ump); 7752 /* 7753 * The jnewblk will be discarded and the bits in the map never 7754 * made it to disk. We can immediately free the freeblk. 7755 */ 7756 if (needj == 0) 7757 handle_written_freework(freework); 7758 } 7759 7760 /* 7761 * We enqueue freework items that need processing back on the freeblks and 7762 * add the freeblks to the worklist. This makes it easier to find all work 7763 * required to flush a truncation in process_truncates(). 7764 */ 7765 static void 7766 freework_enqueue(freework) 7767 struct freework *freework; 7768 { 7769 struct freeblks *freeblks; 7770 7771 freeblks = freework->fw_freeblks; 7772 if ((freework->fw_state & INPROGRESS) == 0) 7773 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7774 if ((freeblks->fb_state & 7775 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7776 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7777 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7778 } 7779 7780 /* 7781 * Start, continue, or finish the process of freeing an indirect block tree. 7782 * The free operation may be paused at any point with fw_off containing the 7783 * offset to restart from. This enables us to implement some flow control 7784 * for large truncates which may fan out and generate a huge number of 7785 * dependencies. 7786 */ 7787 static void 7788 handle_workitem_indirblk(freework) 7789 struct freework *freework; 7790 { 7791 struct freeblks *freeblks; 7792 struct ufsmount *ump; 7793 struct fs *fs; 7794 7795 freeblks = freework->fw_freeblks; 7796 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7797 fs = ump->um_fs; 7798 if (freework->fw_state & DEPCOMPLETE) { 7799 handle_written_freework(freework); 7800 return; 7801 } 7802 if (freework->fw_off == NINDIR(fs)) { 7803 freework_freeblock(freework); 7804 return; 7805 } 7806 freework->fw_state |= INPROGRESS; 7807 FREE_LOCK(ump); 7808 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7809 freework->fw_lbn); 7810 ACQUIRE_LOCK(ump); 7811 } 7812 7813 /* 7814 * Called when a freework structure attached to a cg buf is written. The 7815 * ref on either the parent or the freeblks structure is released and 7816 * the freeblks is added back to the worklist if there is more work to do. 7817 */ 7818 static void 7819 handle_written_freework(freework) 7820 struct freework *freework; 7821 { 7822 struct freeblks *freeblks; 7823 struct freework *parent; 7824 7825 freeblks = freework->fw_freeblks; 7826 parent = freework->fw_parent; 7827 if (freework->fw_state & DELAYEDFREE) 7828 freeblks->fb_cgwait--; 7829 freework->fw_state |= COMPLETE; 7830 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7831 WORKITEM_FREE(freework, D_FREEWORK); 7832 if (parent) { 7833 if (--parent->fw_ref == 0) 7834 freework_enqueue(parent); 7835 return; 7836 } 7837 if (--freeblks->fb_ref != 0) 7838 return; 7839 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7840 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7841 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7842 } 7843 7844 /* 7845 * This workitem routine performs the block de-allocation. 7846 * The workitem is added to the pending list after the updated 7847 * inode block has been written to disk. As mentioned above, 7848 * checks regarding the number of blocks de-allocated (compared 7849 * to the number of blocks allocated for the file) are also 7850 * performed in this function. 7851 */ 7852 static int 7853 handle_workitem_freeblocks(freeblks, flags) 7854 struct freeblks *freeblks; 7855 int flags; 7856 { 7857 struct freework *freework; 7858 struct newblk *newblk; 7859 struct allocindir *aip; 7860 struct ufsmount *ump; 7861 struct worklist *wk; 7862 7863 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7864 ("handle_workitem_freeblocks: Journal entries not written.")); 7865 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7866 ACQUIRE_LOCK(ump); 7867 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7868 WORKLIST_REMOVE(wk); 7869 switch (wk->wk_type) { 7870 case D_DIRREM: 7871 wk->wk_state |= COMPLETE; 7872 add_to_worklist(wk, 0); 7873 continue; 7874 7875 case D_ALLOCDIRECT: 7876 free_newblk(WK_NEWBLK(wk)); 7877 continue; 7878 7879 case D_ALLOCINDIR: 7880 aip = WK_ALLOCINDIR(wk); 7881 freework = NULL; 7882 if (aip->ai_state & DELAYEDFREE) { 7883 FREE_LOCK(ump); 7884 freework = newfreework(ump, freeblks, NULL, 7885 aip->ai_lbn, aip->ai_newblkno, 7886 ump->um_fs->fs_frag, 0, 0); 7887 ACQUIRE_LOCK(ump); 7888 } 7889 newblk = WK_NEWBLK(wk); 7890 if (newblk->nb_jnewblk) { 7891 freework->fw_jnewblk = newblk->nb_jnewblk; 7892 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7893 newblk->nb_jnewblk = NULL; 7894 } 7895 free_newblk(newblk); 7896 continue; 7897 7898 case D_FREEWORK: 7899 freework = WK_FREEWORK(wk); 7900 if (freework->fw_lbn <= -NDADDR) 7901 handle_workitem_indirblk(freework); 7902 else 7903 freework_freeblock(freework); 7904 continue; 7905 default: 7906 panic("handle_workitem_freeblocks: Unknown type %s", 7907 TYPENAME(wk->wk_type)); 7908 } 7909 } 7910 if (freeblks->fb_ref != 0) { 7911 freeblks->fb_state &= ~INPROGRESS; 7912 wake_worklist(&freeblks->fb_list); 7913 freeblks = NULL; 7914 } 7915 FREE_LOCK(ump); 7916 if (freeblks) 7917 return handle_complete_freeblocks(freeblks, flags); 7918 return (0); 7919 } 7920 7921 /* 7922 * Handle completion of block free via truncate. This allows fs_pending 7923 * to track the actual free block count more closely than if we only updated 7924 * it at the end. We must be careful to handle cases where the block count 7925 * on free was incorrect. 7926 */ 7927 static void 7928 freeblks_free(ump, freeblks, blocks) 7929 struct ufsmount *ump; 7930 struct freeblks *freeblks; 7931 int blocks; 7932 { 7933 struct fs *fs; 7934 ufs2_daddr_t remain; 7935 7936 UFS_LOCK(ump); 7937 remain = -freeblks->fb_chkcnt; 7938 freeblks->fb_chkcnt += blocks; 7939 if (remain > 0) { 7940 if (remain < blocks) 7941 blocks = remain; 7942 fs = ump->um_fs; 7943 fs->fs_pendingblocks -= blocks; 7944 } 7945 UFS_UNLOCK(ump); 7946 } 7947 7948 /* 7949 * Once all of the freework workitems are complete we can retire the 7950 * freeblocks dependency and any journal work awaiting completion. This 7951 * can not be called until all other dependencies are stable on disk. 7952 */ 7953 static int 7954 handle_complete_freeblocks(freeblks, flags) 7955 struct freeblks *freeblks; 7956 int flags; 7957 { 7958 struct inodedep *inodedep; 7959 struct inode *ip; 7960 struct vnode *vp; 7961 struct fs *fs; 7962 struct ufsmount *ump; 7963 ufs2_daddr_t spare; 7964 7965 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7966 fs = ump->um_fs; 7967 flags = LK_EXCLUSIVE | flags; 7968 spare = freeblks->fb_chkcnt; 7969 7970 /* 7971 * If we did not release the expected number of blocks we may have 7972 * to adjust the inode block count here. Only do so if it wasn't 7973 * a truncation to zero and the modrev still matches. 7974 */ 7975 if (spare && freeblks->fb_len != 0) { 7976 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 7977 flags, &vp, FFSV_FORCEINSMQ) != 0) 7978 return (EBUSY); 7979 ip = VTOI(vp); 7980 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 7981 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 7982 ip->i_flag |= IN_CHANGE; 7983 /* 7984 * We must wait so this happens before the 7985 * journal is reclaimed. 7986 */ 7987 ffs_update(vp, 1); 7988 } 7989 vput(vp); 7990 } 7991 if (spare < 0) { 7992 UFS_LOCK(ump); 7993 fs->fs_pendingblocks += spare; 7994 UFS_UNLOCK(ump); 7995 } 7996 #ifdef QUOTA 7997 /* Handle spare. */ 7998 if (spare) 7999 quotaadj(freeblks->fb_quota, ump, -spare); 8000 quotarele(freeblks->fb_quota); 8001 #endif 8002 ACQUIRE_LOCK(ump); 8003 if (freeblks->fb_state & ONDEPLIST) { 8004 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8005 0, &inodedep); 8006 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8007 freeblks->fb_state &= ~ONDEPLIST; 8008 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8009 free_inodedep(inodedep); 8010 } 8011 /* 8012 * All of the freeblock deps must be complete prior to this call 8013 * so it's now safe to complete earlier outstanding journal entries. 8014 */ 8015 handle_jwork(&freeblks->fb_jwork); 8016 WORKITEM_FREE(freeblks, D_FREEBLKS); 8017 FREE_LOCK(ump); 8018 return (0); 8019 } 8020 8021 /* 8022 * Release blocks associated with the freeblks and stored in the indirect 8023 * block dbn. If level is greater than SINGLE, the block is an indirect block 8024 * and recursive calls to indirtrunc must be used to cleanse other indirect 8025 * blocks. 8026 * 8027 * This handles partial and complete truncation of blocks. Partial is noted 8028 * with goingaway == 0. In this case the freework is completed after the 8029 * zero'd indirects are written to disk. For full truncation the freework 8030 * is completed after the block is freed. 8031 */ 8032 static void 8033 indir_trunc(freework, dbn, lbn) 8034 struct freework *freework; 8035 ufs2_daddr_t dbn; 8036 ufs_lbn_t lbn; 8037 { 8038 struct freework *nfreework; 8039 struct workhead wkhd; 8040 struct freeblks *freeblks; 8041 struct buf *bp; 8042 struct fs *fs; 8043 struct indirdep *indirdep; 8044 struct ufsmount *ump; 8045 ufs1_daddr_t *bap1 = 0; 8046 ufs2_daddr_t nb, nnb, *bap2 = 0; 8047 ufs_lbn_t lbnadd, nlbn; 8048 int i, nblocks, ufs1fmt; 8049 int freedblocks; 8050 int goingaway; 8051 int freedeps; 8052 int needj; 8053 int level; 8054 int cnt; 8055 8056 freeblks = freework->fw_freeblks; 8057 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8058 fs = ump->um_fs; 8059 /* 8060 * Get buffer of block pointers to be freed. There are three cases: 8061 * 8062 * 1) Partial truncate caches the indirdep pointer in the freework 8063 * which provides us a back copy to the save bp which holds the 8064 * pointers we want to clear. When this completes the zero 8065 * pointers are written to the real copy. 8066 * 2) The indirect is being completely truncated, cancel_indirdep() 8067 * eliminated the real copy and placed the indirdep on the saved 8068 * copy. The indirdep and buf are discarded when this completes. 8069 * 3) The indirect was not in memory, we read a copy off of the disk 8070 * using the devvp and drop and invalidate the buffer when we're 8071 * done. 8072 */ 8073 goingaway = 1; 8074 indirdep = NULL; 8075 if (freework->fw_indir != NULL) { 8076 goingaway = 0; 8077 indirdep = freework->fw_indir; 8078 bp = indirdep->ir_savebp; 8079 if (bp == NULL || bp->b_blkno != dbn) 8080 panic("indir_trunc: Bad saved buf %p blkno %jd", 8081 bp, (intmax_t)dbn); 8082 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8083 /* 8084 * The lock prevents the buf dep list from changing and 8085 * indirects on devvp should only ever have one dependency. 8086 */ 8087 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8088 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8089 panic("indir_trunc: Bad indirdep %p from buf %p", 8090 indirdep, bp); 8091 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8092 NOCRED, &bp) != 0) { 8093 brelse(bp); 8094 return; 8095 } 8096 ACQUIRE_LOCK(ump); 8097 /* Protects against a race with complete_trunc_indir(). */ 8098 freework->fw_state &= ~INPROGRESS; 8099 /* 8100 * If we have an indirdep we need to enforce the truncation order 8101 * and discard it when it is complete. 8102 */ 8103 if (indirdep) { 8104 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8105 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8106 /* 8107 * Add the complete truncate to the list on the 8108 * indirdep to enforce in-order processing. 8109 */ 8110 if (freework->fw_indir == NULL) 8111 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8112 freework, fw_next); 8113 FREE_LOCK(ump); 8114 return; 8115 } 8116 /* 8117 * If we're goingaway, free the indirdep. Otherwise it will 8118 * linger until the write completes. 8119 */ 8120 if (goingaway) 8121 free_indirdep(indirdep); 8122 } 8123 FREE_LOCK(ump); 8124 /* Initialize pointers depending on block size. */ 8125 if (ump->um_fstype == UFS1) { 8126 bap1 = (ufs1_daddr_t *)bp->b_data; 8127 nb = bap1[freework->fw_off]; 8128 ufs1fmt = 1; 8129 } else { 8130 bap2 = (ufs2_daddr_t *)bp->b_data; 8131 nb = bap2[freework->fw_off]; 8132 ufs1fmt = 0; 8133 } 8134 level = lbn_level(lbn); 8135 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8136 lbnadd = lbn_offset(fs, level); 8137 nblocks = btodb(fs->fs_bsize); 8138 nfreework = freework; 8139 freedeps = 0; 8140 cnt = 0; 8141 /* 8142 * Reclaim blocks. Traverses into nested indirect levels and 8143 * arranges for the current level to be freed when subordinates 8144 * are free when journaling. 8145 */ 8146 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8147 if (i != NINDIR(fs) - 1) { 8148 if (ufs1fmt) 8149 nnb = bap1[i+1]; 8150 else 8151 nnb = bap2[i+1]; 8152 } else 8153 nnb = 0; 8154 if (nb == 0) 8155 continue; 8156 cnt++; 8157 if (level != 0) { 8158 nlbn = (lbn + 1) - (i * lbnadd); 8159 if (needj != 0) { 8160 nfreework = newfreework(ump, freeblks, freework, 8161 nlbn, nb, fs->fs_frag, 0, 0); 8162 freedeps++; 8163 } 8164 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8165 } else { 8166 struct freedep *freedep; 8167 8168 /* 8169 * Attempt to aggregate freedep dependencies for 8170 * all blocks being released to the same CG. 8171 */ 8172 LIST_INIT(&wkhd); 8173 if (needj != 0 && 8174 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8175 freedep = newfreedep(freework); 8176 WORKLIST_INSERT_UNLOCKED(&wkhd, 8177 &freedep->fd_list); 8178 freedeps++; 8179 } 8180 CTR3(KTR_SUJ, 8181 "indir_trunc: ino %d blkno %jd size %ld", 8182 freeblks->fb_inum, nb, fs->fs_bsize); 8183 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8184 fs->fs_bsize, freeblks->fb_inum, 8185 freeblks->fb_vtype, &wkhd); 8186 } 8187 } 8188 if (goingaway) { 8189 bp->b_flags |= B_INVAL | B_NOCACHE; 8190 brelse(bp); 8191 } 8192 freedblocks = 0; 8193 if (level == 0) 8194 freedblocks = (nblocks * cnt); 8195 if (needj == 0) 8196 freedblocks += nblocks; 8197 freeblks_free(ump, freeblks, freedblocks); 8198 /* 8199 * If we are journaling set up the ref counts and offset so this 8200 * indirect can be completed when its children are free. 8201 */ 8202 if (needj) { 8203 ACQUIRE_LOCK(ump); 8204 freework->fw_off = i; 8205 freework->fw_ref += freedeps; 8206 freework->fw_ref -= NINDIR(fs) + 1; 8207 if (level == 0) 8208 freeblks->fb_cgwait += freedeps; 8209 if (freework->fw_ref == 0) 8210 freework_freeblock(freework); 8211 FREE_LOCK(ump); 8212 return; 8213 } 8214 /* 8215 * If we're not journaling we can free the indirect now. 8216 */ 8217 dbn = dbtofsb(fs, dbn); 8218 CTR3(KTR_SUJ, 8219 "indir_trunc 2: ino %d blkno %jd size %ld", 8220 freeblks->fb_inum, dbn, fs->fs_bsize); 8221 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8222 freeblks->fb_inum, freeblks->fb_vtype, NULL); 8223 /* Non SUJ softdep does single-threaded truncations. */ 8224 if (freework->fw_blkno == dbn) { 8225 freework->fw_state |= ALLCOMPLETE; 8226 ACQUIRE_LOCK(ump); 8227 handle_written_freework(freework); 8228 FREE_LOCK(ump); 8229 } 8230 return; 8231 } 8232 8233 /* 8234 * Cancel an allocindir when it is removed via truncation. When bp is not 8235 * NULL the indirect never appeared on disk and is scheduled to be freed 8236 * independently of the indir so we can more easily track journal work. 8237 */ 8238 static void 8239 cancel_allocindir(aip, bp, freeblks, trunc) 8240 struct allocindir *aip; 8241 struct buf *bp; 8242 struct freeblks *freeblks; 8243 int trunc; 8244 { 8245 struct indirdep *indirdep; 8246 struct freefrag *freefrag; 8247 struct newblk *newblk; 8248 8249 newblk = (struct newblk *)aip; 8250 LIST_REMOVE(aip, ai_next); 8251 /* 8252 * We must eliminate the pointer in bp if it must be freed on its 8253 * own due to partial truncate or pending journal work. 8254 */ 8255 if (bp && (trunc || newblk->nb_jnewblk)) { 8256 /* 8257 * Clear the pointer and mark the aip to be freed 8258 * directly if it never existed on disk. 8259 */ 8260 aip->ai_state |= DELAYEDFREE; 8261 indirdep = aip->ai_indirdep; 8262 if (indirdep->ir_state & UFS1FMT) 8263 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8264 else 8265 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8266 } 8267 /* 8268 * When truncating the previous pointer will be freed via 8269 * savedbp. Eliminate the freefrag which would dup free. 8270 */ 8271 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8272 newblk->nb_freefrag = NULL; 8273 if (freefrag->ff_jdep) 8274 cancel_jfreefrag( 8275 WK_JFREEFRAG(freefrag->ff_jdep)); 8276 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8277 WORKITEM_FREE(freefrag, D_FREEFRAG); 8278 } 8279 /* 8280 * If the journal hasn't been written the jnewblk must be passed 8281 * to the call to ffs_blkfree that reclaims the space. We accomplish 8282 * this by leaving the journal dependency on the newblk to be freed 8283 * when a freework is created in handle_workitem_freeblocks(). 8284 */ 8285 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8286 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8287 } 8288 8289 /* 8290 * Create the mkdir dependencies for . and .. in a new directory. Link them 8291 * in to a newdirblk so any subsequent additions are tracked properly. The 8292 * caller is responsible for adding the mkdir1 dependency to the journal 8293 * and updating id_mkdiradd. This function returns with the per-filesystem 8294 * lock held. 8295 */ 8296 static struct mkdir * 8297 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8298 struct diradd *dap; 8299 ino_t newinum; 8300 ino_t dinum; 8301 struct buf *newdirbp; 8302 struct mkdir **mkdirp; 8303 { 8304 struct newblk *newblk; 8305 struct pagedep *pagedep; 8306 struct inodedep *inodedep; 8307 struct newdirblk *newdirblk = 0; 8308 struct mkdir *mkdir1, *mkdir2; 8309 struct worklist *wk; 8310 struct jaddref *jaddref; 8311 struct ufsmount *ump; 8312 struct mount *mp; 8313 8314 mp = dap->da_list.wk_mp; 8315 ump = VFSTOUFS(mp); 8316 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8317 M_SOFTDEP_FLAGS); 8318 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8319 LIST_INIT(&newdirblk->db_mkdir); 8320 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8321 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8322 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8323 mkdir1->md_diradd = dap; 8324 mkdir1->md_jaddref = NULL; 8325 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8326 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8327 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8328 mkdir2->md_diradd = dap; 8329 mkdir2->md_jaddref = NULL; 8330 if (MOUNTEDSUJ(mp) == 0) { 8331 mkdir1->md_state |= DEPCOMPLETE; 8332 mkdir2->md_state |= DEPCOMPLETE; 8333 } 8334 /* 8335 * Dependency on "." and ".." being written to disk. 8336 */ 8337 mkdir1->md_buf = newdirbp; 8338 ACQUIRE_LOCK(VFSTOUFS(mp)); 8339 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8340 /* 8341 * We must link the pagedep, allocdirect, and newdirblk for 8342 * the initial file page so the pointer to the new directory 8343 * is not written until the directory contents are live and 8344 * any subsequent additions are not marked live until the 8345 * block is reachable via the inode. 8346 */ 8347 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8348 panic("setup_newdir: lost pagedep"); 8349 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8350 if (wk->wk_type == D_ALLOCDIRECT) 8351 break; 8352 if (wk == NULL) 8353 panic("setup_newdir: lost allocdirect"); 8354 if (pagedep->pd_state & NEWBLOCK) 8355 panic("setup_newdir: NEWBLOCK already set"); 8356 newblk = WK_NEWBLK(wk); 8357 pagedep->pd_state |= NEWBLOCK; 8358 pagedep->pd_newdirblk = newdirblk; 8359 newdirblk->db_pagedep = pagedep; 8360 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8361 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8362 /* 8363 * Look up the inodedep for the parent directory so that we 8364 * can link mkdir2 into the pending dotdot jaddref or 8365 * the inode write if there is none. If the inode is 8366 * ALLCOMPLETE and no jaddref is present all dependencies have 8367 * been satisfied and mkdir2 can be freed. 8368 */ 8369 inodedep_lookup(mp, dinum, 0, &inodedep); 8370 if (MOUNTEDSUJ(mp)) { 8371 if (inodedep == NULL) 8372 panic("setup_newdir: Lost parent."); 8373 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8374 inoreflst); 8375 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8376 (jaddref->ja_state & MKDIR_PARENT), 8377 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8378 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8379 mkdir2->md_jaddref = jaddref; 8380 jaddref->ja_mkdir = mkdir2; 8381 } else if (inodedep == NULL || 8382 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8383 dap->da_state &= ~MKDIR_PARENT; 8384 WORKITEM_FREE(mkdir2, D_MKDIR); 8385 mkdir2 = NULL; 8386 } else { 8387 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8388 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8389 } 8390 *mkdirp = mkdir2; 8391 8392 return (mkdir1); 8393 } 8394 8395 /* 8396 * Directory entry addition dependencies. 8397 * 8398 * When adding a new directory entry, the inode (with its incremented link 8399 * count) must be written to disk before the directory entry's pointer to it. 8400 * Also, if the inode is newly allocated, the corresponding freemap must be 8401 * updated (on disk) before the directory entry's pointer. These requirements 8402 * are met via undo/redo on the directory entry's pointer, which consists 8403 * simply of the inode number. 8404 * 8405 * As directory entries are added and deleted, the free space within a 8406 * directory block can become fragmented. The ufs filesystem will compact 8407 * a fragmented directory block to make space for a new entry. When this 8408 * occurs, the offsets of previously added entries change. Any "diradd" 8409 * dependency structures corresponding to these entries must be updated with 8410 * the new offsets. 8411 */ 8412 8413 /* 8414 * This routine is called after the in-memory inode's link 8415 * count has been incremented, but before the directory entry's 8416 * pointer to the inode has been set. 8417 */ 8418 int 8419 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8420 struct buf *bp; /* buffer containing directory block */ 8421 struct inode *dp; /* inode for directory */ 8422 off_t diroffset; /* offset of new entry in directory */ 8423 ino_t newinum; /* inode referenced by new directory entry */ 8424 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8425 int isnewblk; /* entry is in a newly allocated block */ 8426 { 8427 int offset; /* offset of new entry within directory block */ 8428 ufs_lbn_t lbn; /* block in directory containing new entry */ 8429 struct fs *fs; 8430 struct diradd *dap; 8431 struct newblk *newblk; 8432 struct pagedep *pagedep; 8433 struct inodedep *inodedep; 8434 struct newdirblk *newdirblk = 0; 8435 struct mkdir *mkdir1, *mkdir2; 8436 struct jaddref *jaddref; 8437 struct ufsmount *ump; 8438 struct mount *mp; 8439 int isindir; 8440 8441 ump = dp->i_ump; 8442 mp = UFSTOVFS(ump); 8443 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8444 ("softdep_setup_directory_add called on non-softdep filesystem")); 8445 /* 8446 * Whiteouts have no dependencies. 8447 */ 8448 if (newinum == WINO) { 8449 if (newdirbp != NULL) 8450 bdwrite(newdirbp); 8451 return (0); 8452 } 8453 jaddref = NULL; 8454 mkdir1 = mkdir2 = NULL; 8455 fs = dp->i_fs; 8456 lbn = lblkno(fs, diroffset); 8457 offset = blkoff(fs, diroffset); 8458 dap = malloc(sizeof(struct diradd), M_DIRADD, 8459 M_SOFTDEP_FLAGS|M_ZERO); 8460 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8461 dap->da_offset = offset; 8462 dap->da_newinum = newinum; 8463 dap->da_state = ATTACHED; 8464 LIST_INIT(&dap->da_jwork); 8465 isindir = bp->b_lblkno >= NDADDR; 8466 if (isnewblk && 8467 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8468 newdirblk = malloc(sizeof(struct newdirblk), 8469 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8470 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8471 LIST_INIT(&newdirblk->db_mkdir); 8472 } 8473 /* 8474 * If we're creating a new directory setup the dependencies and set 8475 * the dap state to wait for them. Otherwise it's COMPLETE and 8476 * we can move on. 8477 */ 8478 if (newdirbp == NULL) { 8479 dap->da_state |= DEPCOMPLETE; 8480 ACQUIRE_LOCK(ump); 8481 } else { 8482 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8483 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8484 &mkdir2); 8485 } 8486 /* 8487 * Link into parent directory pagedep to await its being written. 8488 */ 8489 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8490 #ifdef DEBUG 8491 if (diradd_lookup(pagedep, offset) != NULL) 8492 panic("softdep_setup_directory_add: %p already at off %d\n", 8493 diradd_lookup(pagedep, offset), offset); 8494 #endif 8495 dap->da_pagedep = pagedep; 8496 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8497 da_pdlist); 8498 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8499 /* 8500 * If we're journaling, link the diradd into the jaddref so it 8501 * may be completed after the journal entry is written. Otherwise, 8502 * link the diradd into its inodedep. If the inode is not yet 8503 * written place it on the bufwait list, otherwise do the post-inode 8504 * write processing to put it on the id_pendinghd list. 8505 */ 8506 if (MOUNTEDSUJ(mp)) { 8507 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8508 inoreflst); 8509 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8510 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8511 jaddref->ja_diroff = diroffset; 8512 jaddref->ja_diradd = dap; 8513 add_to_journal(&jaddref->ja_list); 8514 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8515 diradd_inode_written(dap, inodedep); 8516 else 8517 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8518 /* 8519 * Add the journal entries for . and .. links now that the primary 8520 * link is written. 8521 */ 8522 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8523 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8524 inoreflst, if_deps); 8525 KASSERT(jaddref != NULL && 8526 jaddref->ja_ino == jaddref->ja_parent && 8527 (jaddref->ja_state & MKDIR_BODY), 8528 ("softdep_setup_directory_add: bad dot jaddref %p", 8529 jaddref)); 8530 mkdir1->md_jaddref = jaddref; 8531 jaddref->ja_mkdir = mkdir1; 8532 /* 8533 * It is important that the dotdot journal entry 8534 * is added prior to the dot entry since dot writes 8535 * both the dot and dotdot links. These both must 8536 * be added after the primary link for the journal 8537 * to remain consistent. 8538 */ 8539 add_to_journal(&mkdir2->md_jaddref->ja_list); 8540 add_to_journal(&jaddref->ja_list); 8541 } 8542 /* 8543 * If we are adding a new directory remember this diradd so that if 8544 * we rename it we can keep the dot and dotdot dependencies. If 8545 * we are adding a new name for an inode that has a mkdiradd we 8546 * must be in rename and we have to move the dot and dotdot 8547 * dependencies to this new name. The old name is being orphaned 8548 * soon. 8549 */ 8550 if (mkdir1 != NULL) { 8551 if (inodedep->id_mkdiradd != NULL) 8552 panic("softdep_setup_directory_add: Existing mkdir"); 8553 inodedep->id_mkdiradd = dap; 8554 } else if (inodedep->id_mkdiradd) 8555 merge_diradd(inodedep, dap); 8556 if (newdirblk) { 8557 /* 8558 * There is nothing to do if we are already tracking 8559 * this block. 8560 */ 8561 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8562 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8563 FREE_LOCK(ump); 8564 return (0); 8565 } 8566 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8567 == 0) 8568 panic("softdep_setup_directory_add: lost entry"); 8569 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8570 pagedep->pd_state |= NEWBLOCK; 8571 pagedep->pd_newdirblk = newdirblk; 8572 newdirblk->db_pagedep = pagedep; 8573 FREE_LOCK(ump); 8574 /* 8575 * If we extended into an indirect signal direnter to sync. 8576 */ 8577 if (isindir) 8578 return (1); 8579 return (0); 8580 } 8581 FREE_LOCK(ump); 8582 return (0); 8583 } 8584 8585 /* 8586 * This procedure is called to change the offset of a directory 8587 * entry when compacting a directory block which must be owned 8588 * exclusively by the caller. Note that the actual entry movement 8589 * must be done in this procedure to ensure that no I/O completions 8590 * occur while the move is in progress. 8591 */ 8592 void 8593 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8594 struct buf *bp; /* Buffer holding directory block. */ 8595 struct inode *dp; /* inode for directory */ 8596 caddr_t base; /* address of dp->i_offset */ 8597 caddr_t oldloc; /* address of old directory location */ 8598 caddr_t newloc; /* address of new directory location */ 8599 int entrysize; /* size of directory entry */ 8600 { 8601 int offset, oldoffset, newoffset; 8602 struct pagedep *pagedep; 8603 struct jmvref *jmvref; 8604 struct diradd *dap; 8605 struct direct *de; 8606 struct mount *mp; 8607 ufs_lbn_t lbn; 8608 int flags; 8609 8610 mp = UFSTOVFS(dp->i_ump); 8611 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8612 ("softdep_change_directoryentry_offset called on " 8613 "non-softdep filesystem")); 8614 de = (struct direct *)oldloc; 8615 jmvref = NULL; 8616 flags = 0; 8617 /* 8618 * Moves are always journaled as it would be too complex to 8619 * determine if any affected adds or removes are present in the 8620 * journal. 8621 */ 8622 if (MOUNTEDSUJ(mp)) { 8623 flags = DEPALLOC; 8624 jmvref = newjmvref(dp, de->d_ino, 8625 dp->i_offset + (oldloc - base), 8626 dp->i_offset + (newloc - base)); 8627 } 8628 lbn = lblkno(dp->i_fs, dp->i_offset); 8629 offset = blkoff(dp->i_fs, dp->i_offset); 8630 oldoffset = offset + (oldloc - base); 8631 newoffset = offset + (newloc - base); 8632 ACQUIRE_LOCK(dp->i_ump); 8633 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8634 goto done; 8635 dap = diradd_lookup(pagedep, oldoffset); 8636 if (dap) { 8637 dap->da_offset = newoffset; 8638 newoffset = DIRADDHASH(newoffset); 8639 oldoffset = DIRADDHASH(oldoffset); 8640 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8641 newoffset != oldoffset) { 8642 LIST_REMOVE(dap, da_pdlist); 8643 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8644 dap, da_pdlist); 8645 } 8646 } 8647 done: 8648 if (jmvref) { 8649 jmvref->jm_pagedep = pagedep; 8650 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8651 add_to_journal(&jmvref->jm_list); 8652 } 8653 bcopy(oldloc, newloc, entrysize); 8654 FREE_LOCK(dp->i_ump); 8655 } 8656 8657 /* 8658 * Move the mkdir dependencies and journal work from one diradd to another 8659 * when renaming a directory. The new name must depend on the mkdir deps 8660 * completing as the old name did. Directories can only have one valid link 8661 * at a time so one must be canonical. 8662 */ 8663 static void 8664 merge_diradd(inodedep, newdap) 8665 struct inodedep *inodedep; 8666 struct diradd *newdap; 8667 { 8668 struct diradd *olddap; 8669 struct mkdir *mkdir, *nextmd; 8670 struct ufsmount *ump; 8671 short state; 8672 8673 olddap = inodedep->id_mkdiradd; 8674 inodedep->id_mkdiradd = newdap; 8675 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8676 newdap->da_state &= ~DEPCOMPLETE; 8677 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8678 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8679 mkdir = nextmd) { 8680 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8681 if (mkdir->md_diradd != olddap) 8682 continue; 8683 mkdir->md_diradd = newdap; 8684 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8685 newdap->da_state |= state; 8686 olddap->da_state &= ~state; 8687 if ((olddap->da_state & 8688 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8689 break; 8690 } 8691 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8692 panic("merge_diradd: unfound ref"); 8693 } 8694 /* 8695 * Any mkdir related journal items are not safe to be freed until 8696 * the new name is stable. 8697 */ 8698 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8699 olddap->da_state |= DEPCOMPLETE; 8700 complete_diradd(olddap); 8701 } 8702 8703 /* 8704 * Move the diradd to the pending list when all diradd dependencies are 8705 * complete. 8706 */ 8707 static void 8708 complete_diradd(dap) 8709 struct diradd *dap; 8710 { 8711 struct pagedep *pagedep; 8712 8713 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8714 if (dap->da_state & DIRCHG) 8715 pagedep = dap->da_previous->dm_pagedep; 8716 else 8717 pagedep = dap->da_pagedep; 8718 LIST_REMOVE(dap, da_pdlist); 8719 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8720 } 8721 } 8722 8723 /* 8724 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8725 * add entries and conditonally journal the remove. 8726 */ 8727 static void 8728 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8729 struct diradd *dap; 8730 struct dirrem *dirrem; 8731 struct jremref *jremref; 8732 struct jremref *dotremref; 8733 struct jremref *dotdotremref; 8734 { 8735 struct inodedep *inodedep; 8736 struct jaddref *jaddref; 8737 struct inoref *inoref; 8738 struct ufsmount *ump; 8739 struct mkdir *mkdir; 8740 8741 /* 8742 * If no remove references were allocated we're on a non-journaled 8743 * filesystem and can skip the cancel step. 8744 */ 8745 if (jremref == NULL) { 8746 free_diradd(dap, NULL); 8747 return; 8748 } 8749 /* 8750 * Cancel the primary name an free it if it does not require 8751 * journaling. 8752 */ 8753 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8754 0, &inodedep) != 0) { 8755 /* Abort the addref that reference this diradd. */ 8756 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8757 if (inoref->if_list.wk_type != D_JADDREF) 8758 continue; 8759 jaddref = (struct jaddref *)inoref; 8760 if (jaddref->ja_diradd != dap) 8761 continue; 8762 if (cancel_jaddref(jaddref, inodedep, 8763 &dirrem->dm_jwork) == 0) { 8764 free_jremref(jremref); 8765 jremref = NULL; 8766 } 8767 break; 8768 } 8769 } 8770 /* 8771 * Cancel subordinate names and free them if they do not require 8772 * journaling. 8773 */ 8774 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8775 ump = VFSTOUFS(dap->da_list.wk_mp); 8776 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8777 if (mkdir->md_diradd != dap) 8778 continue; 8779 if ((jaddref = mkdir->md_jaddref) == NULL) 8780 continue; 8781 mkdir->md_jaddref = NULL; 8782 if (mkdir->md_state & MKDIR_PARENT) { 8783 if (cancel_jaddref(jaddref, NULL, 8784 &dirrem->dm_jwork) == 0) { 8785 free_jremref(dotdotremref); 8786 dotdotremref = NULL; 8787 } 8788 } else { 8789 if (cancel_jaddref(jaddref, inodedep, 8790 &dirrem->dm_jwork) == 0) { 8791 free_jremref(dotremref); 8792 dotremref = NULL; 8793 } 8794 } 8795 } 8796 } 8797 8798 if (jremref) 8799 journal_jremref(dirrem, jremref, inodedep); 8800 if (dotremref) 8801 journal_jremref(dirrem, dotremref, inodedep); 8802 if (dotdotremref) 8803 journal_jremref(dirrem, dotdotremref, NULL); 8804 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8805 free_diradd(dap, &dirrem->dm_jwork); 8806 } 8807 8808 /* 8809 * Free a diradd dependency structure. This routine must be called 8810 * with splbio interrupts blocked. 8811 */ 8812 static void 8813 free_diradd(dap, wkhd) 8814 struct diradd *dap; 8815 struct workhead *wkhd; 8816 { 8817 struct dirrem *dirrem; 8818 struct pagedep *pagedep; 8819 struct inodedep *inodedep; 8820 struct mkdir *mkdir, *nextmd; 8821 struct ufsmount *ump; 8822 8823 ump = VFSTOUFS(dap->da_list.wk_mp); 8824 LOCK_OWNED(ump); 8825 LIST_REMOVE(dap, da_pdlist); 8826 if (dap->da_state & ONWORKLIST) 8827 WORKLIST_REMOVE(&dap->da_list); 8828 if ((dap->da_state & DIRCHG) == 0) { 8829 pagedep = dap->da_pagedep; 8830 } else { 8831 dirrem = dap->da_previous; 8832 pagedep = dirrem->dm_pagedep; 8833 dirrem->dm_dirinum = pagedep->pd_ino; 8834 dirrem->dm_state |= COMPLETE; 8835 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8836 add_to_worklist(&dirrem->dm_list, 0); 8837 } 8838 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8839 0, &inodedep) != 0) 8840 if (inodedep->id_mkdiradd == dap) 8841 inodedep->id_mkdiradd = NULL; 8842 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8843 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8844 mkdir = nextmd) { 8845 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8846 if (mkdir->md_diradd != dap) 8847 continue; 8848 dap->da_state &= 8849 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8850 LIST_REMOVE(mkdir, md_mkdirs); 8851 if (mkdir->md_state & ONWORKLIST) 8852 WORKLIST_REMOVE(&mkdir->md_list); 8853 if (mkdir->md_jaddref != NULL) 8854 panic("free_diradd: Unexpected jaddref"); 8855 WORKITEM_FREE(mkdir, D_MKDIR); 8856 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8857 break; 8858 } 8859 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8860 panic("free_diradd: unfound ref"); 8861 } 8862 if (inodedep) 8863 free_inodedep(inodedep); 8864 /* 8865 * Free any journal segments waiting for the directory write. 8866 */ 8867 handle_jwork(&dap->da_jwork); 8868 WORKITEM_FREE(dap, D_DIRADD); 8869 } 8870 8871 /* 8872 * Directory entry removal dependencies. 8873 * 8874 * When removing a directory entry, the entry's inode pointer must be 8875 * zero'ed on disk before the corresponding inode's link count is decremented 8876 * (possibly freeing the inode for re-use). This dependency is handled by 8877 * updating the directory entry but delaying the inode count reduction until 8878 * after the directory block has been written to disk. After this point, the 8879 * inode count can be decremented whenever it is convenient. 8880 */ 8881 8882 /* 8883 * This routine should be called immediately after removing 8884 * a directory entry. The inode's link count should not be 8885 * decremented by the calling procedure -- the soft updates 8886 * code will do this task when it is safe. 8887 */ 8888 void 8889 softdep_setup_remove(bp, dp, ip, isrmdir) 8890 struct buf *bp; /* buffer containing directory block */ 8891 struct inode *dp; /* inode for the directory being modified */ 8892 struct inode *ip; /* inode for directory entry being removed */ 8893 int isrmdir; /* indicates if doing RMDIR */ 8894 { 8895 struct dirrem *dirrem, *prevdirrem; 8896 struct inodedep *inodedep; 8897 int direct; 8898 8899 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 8900 ("softdep_setup_remove called on non-softdep filesystem")); 8901 /* 8902 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8903 * newdirrem() to setup the full directory remove which requires 8904 * isrmdir > 1. 8905 */ 8906 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8907 /* 8908 * Add the dirrem to the inodedep's pending remove list for quick 8909 * discovery later. 8910 */ 8911 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 8912 &inodedep) == 0) 8913 panic("softdep_setup_remove: Lost inodedep."); 8914 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8915 dirrem->dm_state |= ONDEPLIST; 8916 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8917 8918 /* 8919 * If the COMPLETE flag is clear, then there were no active 8920 * entries and we want to roll back to a zeroed entry until 8921 * the new inode is committed to disk. If the COMPLETE flag is 8922 * set then we have deleted an entry that never made it to 8923 * disk. If the entry we deleted resulted from a name change, 8924 * then the old name still resides on disk. We cannot delete 8925 * its inode (returned to us in prevdirrem) until the zeroed 8926 * directory entry gets to disk. The new inode has never been 8927 * referenced on the disk, so can be deleted immediately. 8928 */ 8929 if ((dirrem->dm_state & COMPLETE) == 0) { 8930 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8931 dm_next); 8932 FREE_LOCK(ip->i_ump); 8933 } else { 8934 if (prevdirrem != NULL) 8935 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8936 prevdirrem, dm_next); 8937 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8938 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8939 FREE_LOCK(ip->i_ump); 8940 if (direct) 8941 handle_workitem_remove(dirrem, 0); 8942 } 8943 } 8944 8945 /* 8946 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8947 * pd_pendinghd list of a pagedep. 8948 */ 8949 static struct diradd * 8950 diradd_lookup(pagedep, offset) 8951 struct pagedep *pagedep; 8952 int offset; 8953 { 8954 struct diradd *dap; 8955 8956 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8957 if (dap->da_offset == offset) 8958 return (dap); 8959 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 8960 if (dap->da_offset == offset) 8961 return (dap); 8962 return (NULL); 8963 } 8964 8965 /* 8966 * Search for a .. diradd dependency in a directory that is being removed. 8967 * If the directory was renamed to a new parent we have a diradd rather 8968 * than a mkdir for the .. entry. We need to cancel it now before 8969 * it is found in truncate(). 8970 */ 8971 static struct jremref * 8972 cancel_diradd_dotdot(ip, dirrem, jremref) 8973 struct inode *ip; 8974 struct dirrem *dirrem; 8975 struct jremref *jremref; 8976 { 8977 struct pagedep *pagedep; 8978 struct diradd *dap; 8979 struct worklist *wk; 8980 8981 if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0, 8982 &pagedep) == 0) 8983 return (jremref); 8984 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 8985 if (dap == NULL) 8986 return (jremref); 8987 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 8988 /* 8989 * Mark any journal work as belonging to the parent so it is freed 8990 * with the .. reference. 8991 */ 8992 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 8993 wk->wk_state |= MKDIR_PARENT; 8994 return (NULL); 8995 } 8996 8997 /* 8998 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 8999 * replace it with a dirrem/diradd pair as a result of re-parenting a 9000 * directory. This ensures that we don't simultaneously have a mkdir and 9001 * a diradd for the same .. entry. 9002 */ 9003 static struct jremref * 9004 cancel_mkdir_dotdot(ip, dirrem, jremref) 9005 struct inode *ip; 9006 struct dirrem *dirrem; 9007 struct jremref *jremref; 9008 { 9009 struct inodedep *inodedep; 9010 struct jaddref *jaddref; 9011 struct ufsmount *ump; 9012 struct mkdir *mkdir; 9013 struct diradd *dap; 9014 9015 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 9016 &inodedep) == 0) 9017 return (jremref); 9018 dap = inodedep->id_mkdiradd; 9019 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9020 return (jremref); 9021 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9022 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9023 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9024 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9025 break; 9026 if (mkdir == NULL) 9027 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9028 if ((jaddref = mkdir->md_jaddref) != NULL) { 9029 mkdir->md_jaddref = NULL; 9030 jaddref->ja_state &= ~MKDIR_PARENT; 9031 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0, 9032 &inodedep) == 0) 9033 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9034 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9035 journal_jremref(dirrem, jremref, inodedep); 9036 jremref = NULL; 9037 } 9038 } 9039 if (mkdir->md_state & ONWORKLIST) 9040 WORKLIST_REMOVE(&mkdir->md_list); 9041 mkdir->md_state |= ALLCOMPLETE; 9042 complete_mkdir(mkdir); 9043 return (jremref); 9044 } 9045 9046 static void 9047 journal_jremref(dirrem, jremref, inodedep) 9048 struct dirrem *dirrem; 9049 struct jremref *jremref; 9050 struct inodedep *inodedep; 9051 { 9052 9053 if (inodedep == NULL) 9054 if (inodedep_lookup(jremref->jr_list.wk_mp, 9055 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9056 panic("journal_jremref: Lost inodedep"); 9057 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9058 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9059 add_to_journal(&jremref->jr_list); 9060 } 9061 9062 static void 9063 dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9064 struct dirrem *dirrem; 9065 struct jremref *jremref; 9066 struct jremref *dotremref; 9067 struct jremref *dotdotremref; 9068 { 9069 struct inodedep *inodedep; 9070 9071 9072 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9073 &inodedep) == 0) 9074 panic("dirrem_journal: Lost inodedep"); 9075 journal_jremref(dirrem, jremref, inodedep); 9076 if (dotremref) 9077 journal_jremref(dirrem, dotremref, inodedep); 9078 if (dotdotremref) 9079 journal_jremref(dirrem, dotdotremref, NULL); 9080 } 9081 9082 /* 9083 * Allocate a new dirrem if appropriate and return it along with 9084 * its associated pagedep. Called without a lock, returns with lock. 9085 */ 9086 static struct dirrem * 9087 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9088 struct buf *bp; /* buffer containing directory block */ 9089 struct inode *dp; /* inode for the directory being modified */ 9090 struct inode *ip; /* inode for directory entry being removed */ 9091 int isrmdir; /* indicates if doing RMDIR */ 9092 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9093 { 9094 int offset; 9095 ufs_lbn_t lbn; 9096 struct diradd *dap; 9097 struct dirrem *dirrem; 9098 struct pagedep *pagedep; 9099 struct jremref *jremref; 9100 struct jremref *dotremref; 9101 struct jremref *dotdotremref; 9102 struct vnode *dvp; 9103 9104 /* 9105 * Whiteouts have no deletion dependencies. 9106 */ 9107 if (ip == NULL) 9108 panic("newdirrem: whiteout"); 9109 dvp = ITOV(dp); 9110 /* 9111 * If the system is over its limit and our filesystem is 9112 * responsible for more than our share of that usage and 9113 * we are not a snapshot, request some inodedep cleanup. 9114 * Limiting the number of dirrem structures will also limit 9115 * the number of freefile and freeblks structures. 9116 */ 9117 ACQUIRE_LOCK(ip->i_ump); 9118 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ip->i_ump, D_DIRREM)) 9119 schedule_cleanup(ITOV(dp)->v_mount); 9120 else 9121 FREE_LOCK(ip->i_ump); 9122 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9123 M_ZERO); 9124 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9125 LIST_INIT(&dirrem->dm_jremrefhd); 9126 LIST_INIT(&dirrem->dm_jwork); 9127 dirrem->dm_state = isrmdir ? RMDIR : 0; 9128 dirrem->dm_oldinum = ip->i_number; 9129 *prevdirremp = NULL; 9130 /* 9131 * Allocate remove reference structures to track journal write 9132 * dependencies. We will always have one for the link and 9133 * when doing directories we will always have one more for dot. 9134 * When renaming a directory we skip the dotdot link change so 9135 * this is not needed. 9136 */ 9137 jremref = dotremref = dotdotremref = NULL; 9138 if (DOINGSUJ(dvp)) { 9139 if (isrmdir) { 9140 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9141 ip->i_effnlink + 2); 9142 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9143 ip->i_effnlink + 1); 9144 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9145 dp->i_effnlink + 1); 9146 dotdotremref->jr_state |= MKDIR_PARENT; 9147 } else 9148 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9149 ip->i_effnlink + 1); 9150 } 9151 ACQUIRE_LOCK(ip->i_ump); 9152 lbn = lblkno(dp->i_fs, dp->i_offset); 9153 offset = blkoff(dp->i_fs, dp->i_offset); 9154 pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC, 9155 &pagedep); 9156 dirrem->dm_pagedep = pagedep; 9157 dirrem->dm_offset = offset; 9158 /* 9159 * If we're renaming a .. link to a new directory, cancel any 9160 * existing MKDIR_PARENT mkdir. If it has already been canceled 9161 * the jremref is preserved for any potential diradd in this 9162 * location. This can not coincide with a rmdir. 9163 */ 9164 if (dp->i_offset == DOTDOT_OFFSET) { 9165 if (isrmdir) 9166 panic("newdirrem: .. directory change during remove?"); 9167 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9168 } 9169 /* 9170 * If we're removing a directory search for the .. dependency now and 9171 * cancel it. Any pending journal work will be added to the dirrem 9172 * to be completed when the workitem remove completes. 9173 */ 9174 if (isrmdir) 9175 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9176 /* 9177 * Check for a diradd dependency for the same directory entry. 9178 * If present, then both dependencies become obsolete and can 9179 * be de-allocated. 9180 */ 9181 dap = diradd_lookup(pagedep, offset); 9182 if (dap == NULL) { 9183 /* 9184 * Link the jremref structures into the dirrem so they are 9185 * written prior to the pagedep. 9186 */ 9187 if (jremref) 9188 dirrem_journal(dirrem, jremref, dotremref, 9189 dotdotremref); 9190 return (dirrem); 9191 } 9192 /* 9193 * Must be ATTACHED at this point. 9194 */ 9195 if ((dap->da_state & ATTACHED) == 0) 9196 panic("newdirrem: not ATTACHED"); 9197 if (dap->da_newinum != ip->i_number) 9198 panic("newdirrem: inum %ju should be %ju", 9199 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9200 /* 9201 * If we are deleting a changed name that never made it to disk, 9202 * then return the dirrem describing the previous inode (which 9203 * represents the inode currently referenced from this entry on disk). 9204 */ 9205 if ((dap->da_state & DIRCHG) != 0) { 9206 *prevdirremp = dap->da_previous; 9207 dap->da_state &= ~DIRCHG; 9208 dap->da_pagedep = pagedep; 9209 } 9210 /* 9211 * We are deleting an entry that never made it to disk. 9212 * Mark it COMPLETE so we can delete its inode immediately. 9213 */ 9214 dirrem->dm_state |= COMPLETE; 9215 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9216 #ifdef SUJ_DEBUG 9217 if (isrmdir == 0) { 9218 struct worklist *wk; 9219 9220 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9221 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9222 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9223 } 9224 #endif 9225 9226 return (dirrem); 9227 } 9228 9229 /* 9230 * Directory entry change dependencies. 9231 * 9232 * Changing an existing directory entry requires that an add operation 9233 * be completed first followed by a deletion. The semantics for the addition 9234 * are identical to the description of adding a new entry above except 9235 * that the rollback is to the old inode number rather than zero. Once 9236 * the addition dependency is completed, the removal is done as described 9237 * in the removal routine above. 9238 */ 9239 9240 /* 9241 * This routine should be called immediately after changing 9242 * a directory entry. The inode's link count should not be 9243 * decremented by the calling procedure -- the soft updates 9244 * code will perform this task when it is safe. 9245 */ 9246 void 9247 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9248 struct buf *bp; /* buffer containing directory block */ 9249 struct inode *dp; /* inode for the directory being modified */ 9250 struct inode *ip; /* inode for directory entry being removed */ 9251 ino_t newinum; /* new inode number for changed entry */ 9252 int isrmdir; /* indicates if doing RMDIR */ 9253 { 9254 int offset; 9255 struct diradd *dap = NULL; 9256 struct dirrem *dirrem, *prevdirrem; 9257 struct pagedep *pagedep; 9258 struct inodedep *inodedep; 9259 struct jaddref *jaddref; 9260 struct mount *mp; 9261 9262 offset = blkoff(dp->i_fs, dp->i_offset); 9263 mp = UFSTOVFS(dp->i_ump); 9264 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9265 ("softdep_setup_directory_change called on non-softdep filesystem")); 9266 9267 /* 9268 * Whiteouts do not need diradd dependencies. 9269 */ 9270 if (newinum != WINO) { 9271 dap = malloc(sizeof(struct diradd), 9272 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9273 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9274 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9275 dap->da_offset = offset; 9276 dap->da_newinum = newinum; 9277 LIST_INIT(&dap->da_jwork); 9278 } 9279 9280 /* 9281 * Allocate a new dirrem and ACQUIRE_LOCK. 9282 */ 9283 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9284 pagedep = dirrem->dm_pagedep; 9285 /* 9286 * The possible values for isrmdir: 9287 * 0 - non-directory file rename 9288 * 1 - directory rename within same directory 9289 * inum - directory rename to new directory of given inode number 9290 * When renaming to a new directory, we are both deleting and 9291 * creating a new directory entry, so the link count on the new 9292 * directory should not change. Thus we do not need the followup 9293 * dirrem which is usually done in handle_workitem_remove. We set 9294 * the DIRCHG flag to tell handle_workitem_remove to skip the 9295 * followup dirrem. 9296 */ 9297 if (isrmdir > 1) 9298 dirrem->dm_state |= DIRCHG; 9299 9300 /* 9301 * Whiteouts have no additional dependencies, 9302 * so just put the dirrem on the correct list. 9303 */ 9304 if (newinum == WINO) { 9305 if ((dirrem->dm_state & COMPLETE) == 0) { 9306 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9307 dm_next); 9308 } else { 9309 dirrem->dm_dirinum = pagedep->pd_ino; 9310 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9311 add_to_worklist(&dirrem->dm_list, 0); 9312 } 9313 FREE_LOCK(dp->i_ump); 9314 return; 9315 } 9316 /* 9317 * Add the dirrem to the inodedep's pending remove list for quick 9318 * discovery later. A valid nlinkdelta ensures that this lookup 9319 * will not fail. 9320 */ 9321 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9322 panic("softdep_setup_directory_change: Lost inodedep."); 9323 dirrem->dm_state |= ONDEPLIST; 9324 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9325 9326 /* 9327 * If the COMPLETE flag is clear, then there were no active 9328 * entries and we want to roll back to the previous inode until 9329 * the new inode is committed to disk. If the COMPLETE flag is 9330 * set, then we have deleted an entry that never made it to disk. 9331 * If the entry we deleted resulted from a name change, then the old 9332 * inode reference still resides on disk. Any rollback that we do 9333 * needs to be to that old inode (returned to us in prevdirrem). If 9334 * the entry we deleted resulted from a create, then there is 9335 * no entry on the disk, so we want to roll back to zero rather 9336 * than the uncommitted inode. In either of the COMPLETE cases we 9337 * want to immediately free the unwritten and unreferenced inode. 9338 */ 9339 if ((dirrem->dm_state & COMPLETE) == 0) { 9340 dap->da_previous = dirrem; 9341 } else { 9342 if (prevdirrem != NULL) { 9343 dap->da_previous = prevdirrem; 9344 } else { 9345 dap->da_state &= ~DIRCHG; 9346 dap->da_pagedep = pagedep; 9347 } 9348 dirrem->dm_dirinum = pagedep->pd_ino; 9349 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9350 add_to_worklist(&dirrem->dm_list, 0); 9351 } 9352 /* 9353 * Lookup the jaddref for this journal entry. We must finish 9354 * initializing it and make the diradd write dependent on it. 9355 * If we're not journaling, put it on the id_bufwait list if the 9356 * inode is not yet written. If it is written, do the post-inode 9357 * write processing to put it on the id_pendinghd list. 9358 */ 9359 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9360 if (MOUNTEDSUJ(mp)) { 9361 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9362 inoreflst); 9363 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9364 ("softdep_setup_directory_change: bad jaddref %p", 9365 jaddref)); 9366 jaddref->ja_diroff = dp->i_offset; 9367 jaddref->ja_diradd = dap; 9368 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9369 dap, da_pdlist); 9370 add_to_journal(&jaddref->ja_list); 9371 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9372 dap->da_state |= COMPLETE; 9373 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9374 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9375 } else { 9376 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9377 dap, da_pdlist); 9378 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9379 } 9380 /* 9381 * If we're making a new name for a directory that has not been 9382 * committed when need to move the dot and dotdot references to 9383 * this new name. 9384 */ 9385 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9386 merge_diradd(inodedep, dap); 9387 FREE_LOCK(dp->i_ump); 9388 } 9389 9390 /* 9391 * Called whenever the link count on an inode is changed. 9392 * It creates an inode dependency so that the new reference(s) 9393 * to the inode cannot be committed to disk until the updated 9394 * inode has been written. 9395 */ 9396 void 9397 softdep_change_linkcnt(ip) 9398 struct inode *ip; /* the inode with the increased link count */ 9399 { 9400 struct inodedep *inodedep; 9401 9402 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 9403 ("softdep_change_linkcnt called on non-softdep filesystem")); 9404 ACQUIRE_LOCK(ip->i_ump); 9405 inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, 9406 &inodedep); 9407 if (ip->i_nlink < ip->i_effnlink) 9408 panic("softdep_change_linkcnt: bad delta"); 9409 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9410 FREE_LOCK(ip->i_ump); 9411 } 9412 9413 /* 9414 * Attach a sbdep dependency to the superblock buf so that we can keep 9415 * track of the head of the linked list of referenced but unlinked inodes. 9416 */ 9417 void 9418 softdep_setup_sbupdate(ump, fs, bp) 9419 struct ufsmount *ump; 9420 struct fs *fs; 9421 struct buf *bp; 9422 { 9423 struct sbdep *sbdep; 9424 struct worklist *wk; 9425 9426 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9427 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9428 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9429 if (wk->wk_type == D_SBDEP) 9430 break; 9431 if (wk != NULL) 9432 return; 9433 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9434 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9435 sbdep->sb_fs = fs; 9436 sbdep->sb_ump = ump; 9437 ACQUIRE_LOCK(ump); 9438 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9439 FREE_LOCK(ump); 9440 } 9441 9442 /* 9443 * Return the first unlinked inodedep which is ready to be the head of the 9444 * list. The inodedep and all those after it must have valid next pointers. 9445 */ 9446 static struct inodedep * 9447 first_unlinked_inodedep(ump) 9448 struct ufsmount *ump; 9449 { 9450 struct inodedep *inodedep; 9451 struct inodedep *idp; 9452 9453 LOCK_OWNED(ump); 9454 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9455 inodedep; inodedep = idp) { 9456 if ((inodedep->id_state & UNLINKNEXT) == 0) 9457 return (NULL); 9458 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9459 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9460 break; 9461 if ((inodedep->id_state & UNLINKPREV) == 0) 9462 break; 9463 } 9464 return (inodedep); 9465 } 9466 9467 /* 9468 * Set the sujfree unlinked head pointer prior to writing a superblock. 9469 */ 9470 static void 9471 initiate_write_sbdep(sbdep) 9472 struct sbdep *sbdep; 9473 { 9474 struct inodedep *inodedep; 9475 struct fs *bpfs; 9476 struct fs *fs; 9477 9478 bpfs = sbdep->sb_fs; 9479 fs = sbdep->sb_ump->um_fs; 9480 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9481 if (inodedep) { 9482 fs->fs_sujfree = inodedep->id_ino; 9483 inodedep->id_state |= UNLINKPREV; 9484 } else 9485 fs->fs_sujfree = 0; 9486 bpfs->fs_sujfree = fs->fs_sujfree; 9487 } 9488 9489 /* 9490 * After a superblock is written determine whether it must be written again 9491 * due to a changing unlinked list head. 9492 */ 9493 static int 9494 handle_written_sbdep(sbdep, bp) 9495 struct sbdep *sbdep; 9496 struct buf *bp; 9497 { 9498 struct inodedep *inodedep; 9499 struct fs *fs; 9500 9501 LOCK_OWNED(sbdep->sb_ump); 9502 fs = sbdep->sb_fs; 9503 /* 9504 * If the superblock doesn't match the in-memory list start over. 9505 */ 9506 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9507 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9508 (inodedep == NULL && fs->fs_sujfree != 0)) { 9509 bdirty(bp); 9510 return (1); 9511 } 9512 WORKITEM_FREE(sbdep, D_SBDEP); 9513 if (fs->fs_sujfree == 0) 9514 return (0); 9515 /* 9516 * Now that we have a record of this inode in stable store allow it 9517 * to be written to free up pending work. Inodes may see a lot of 9518 * write activity after they are unlinked which we must not hold up. 9519 */ 9520 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9521 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9522 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9523 inodedep, inodedep->id_state); 9524 if (inodedep->id_state & UNLINKONLIST) 9525 break; 9526 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9527 } 9528 9529 return (0); 9530 } 9531 9532 /* 9533 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9534 */ 9535 static void 9536 unlinked_inodedep(mp, inodedep) 9537 struct mount *mp; 9538 struct inodedep *inodedep; 9539 { 9540 struct ufsmount *ump; 9541 9542 ump = VFSTOUFS(mp); 9543 LOCK_OWNED(ump); 9544 if (MOUNTEDSUJ(mp) == 0) 9545 return; 9546 ump->um_fs->fs_fmod = 1; 9547 if (inodedep->id_state & UNLINKED) 9548 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9549 inodedep->id_state |= UNLINKED; 9550 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9551 } 9552 9553 /* 9554 * Remove an inodedep from the unlinked inodedep list. This may require 9555 * disk writes if the inode has made it that far. 9556 */ 9557 static void 9558 clear_unlinked_inodedep(inodedep) 9559 struct inodedep *inodedep; 9560 { 9561 struct ufsmount *ump; 9562 struct inodedep *idp; 9563 struct inodedep *idn; 9564 struct fs *fs; 9565 struct buf *bp; 9566 ino_t ino; 9567 ino_t nino; 9568 ino_t pino; 9569 int error; 9570 9571 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9572 fs = ump->um_fs; 9573 ino = inodedep->id_ino; 9574 error = 0; 9575 for (;;) { 9576 LOCK_OWNED(ump); 9577 KASSERT((inodedep->id_state & UNLINKED) != 0, 9578 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9579 inodedep)); 9580 /* 9581 * If nothing has yet been written simply remove us from 9582 * the in memory list and return. This is the most common 9583 * case where handle_workitem_remove() loses the final 9584 * reference. 9585 */ 9586 if ((inodedep->id_state & UNLINKLINKS) == 0) 9587 break; 9588 /* 9589 * If we have a NEXT pointer and no PREV pointer we can simply 9590 * clear NEXT's PREV and remove ourselves from the list. Be 9591 * careful not to clear PREV if the superblock points at 9592 * next as well. 9593 */ 9594 idn = TAILQ_NEXT(inodedep, id_unlinked); 9595 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9596 if (idn && fs->fs_sujfree != idn->id_ino) 9597 idn->id_state &= ~UNLINKPREV; 9598 break; 9599 } 9600 /* 9601 * Here we have an inodedep which is actually linked into 9602 * the list. We must remove it by forcing a write to the 9603 * link before us, whether it be the superblock or an inode. 9604 * Unfortunately the list may change while we're waiting 9605 * on the buf lock for either resource so we must loop until 9606 * we lock the right one. If both the superblock and an 9607 * inode point to this inode we must clear the inode first 9608 * followed by the superblock. 9609 */ 9610 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9611 pino = 0; 9612 if (idp && (idp->id_state & UNLINKNEXT)) 9613 pino = idp->id_ino; 9614 FREE_LOCK(ump); 9615 if (pino == 0) { 9616 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9617 (int)fs->fs_sbsize, 0, 0, 0); 9618 } else { 9619 error = bread(ump->um_devvp, 9620 fsbtodb(fs, ino_to_fsba(fs, pino)), 9621 (int)fs->fs_bsize, NOCRED, &bp); 9622 if (error) 9623 brelse(bp); 9624 } 9625 ACQUIRE_LOCK(ump); 9626 if (error) 9627 break; 9628 /* If the list has changed restart the loop. */ 9629 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9630 nino = 0; 9631 if (idp && (idp->id_state & UNLINKNEXT)) 9632 nino = idp->id_ino; 9633 if (nino != pino || 9634 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9635 FREE_LOCK(ump); 9636 brelse(bp); 9637 ACQUIRE_LOCK(ump); 9638 continue; 9639 } 9640 nino = 0; 9641 idn = TAILQ_NEXT(inodedep, id_unlinked); 9642 if (idn) 9643 nino = idn->id_ino; 9644 /* 9645 * Remove us from the in memory list. After this we cannot 9646 * access the inodedep. 9647 */ 9648 KASSERT((inodedep->id_state & UNLINKED) != 0, 9649 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9650 inodedep)); 9651 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9652 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9653 FREE_LOCK(ump); 9654 /* 9655 * The predecessor's next pointer is manually updated here 9656 * so that the NEXT flag is never cleared for an element 9657 * that is in the list. 9658 */ 9659 if (pino == 0) { 9660 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9661 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9662 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9663 bp); 9664 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9665 ((struct ufs1_dinode *)bp->b_data + 9666 ino_to_fsbo(fs, pino))->di_freelink = nino; 9667 else 9668 ((struct ufs2_dinode *)bp->b_data + 9669 ino_to_fsbo(fs, pino))->di_freelink = nino; 9670 /* 9671 * If the bwrite fails we have no recourse to recover. The 9672 * filesystem is corrupted already. 9673 */ 9674 bwrite(bp); 9675 ACQUIRE_LOCK(ump); 9676 /* 9677 * If the superblock pointer still needs to be cleared force 9678 * a write here. 9679 */ 9680 if (fs->fs_sujfree == ino) { 9681 FREE_LOCK(ump); 9682 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9683 (int)fs->fs_sbsize, 0, 0, 0); 9684 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9685 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9686 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9687 bp); 9688 bwrite(bp); 9689 ACQUIRE_LOCK(ump); 9690 } 9691 9692 if (fs->fs_sujfree != ino) 9693 return; 9694 panic("clear_unlinked_inodedep: Failed to clear free head"); 9695 } 9696 if (inodedep->id_ino == fs->fs_sujfree) 9697 panic("clear_unlinked_inodedep: Freeing head of free list"); 9698 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9699 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9700 return; 9701 } 9702 9703 /* 9704 * This workitem decrements the inode's link count. 9705 * If the link count reaches zero, the file is removed. 9706 */ 9707 static int 9708 handle_workitem_remove(dirrem, flags) 9709 struct dirrem *dirrem; 9710 int flags; 9711 { 9712 struct inodedep *inodedep; 9713 struct workhead dotdotwk; 9714 struct worklist *wk; 9715 struct ufsmount *ump; 9716 struct mount *mp; 9717 struct vnode *vp; 9718 struct inode *ip; 9719 ino_t oldinum; 9720 9721 if (dirrem->dm_state & ONWORKLIST) 9722 panic("handle_workitem_remove: dirrem %p still on worklist", 9723 dirrem); 9724 oldinum = dirrem->dm_oldinum; 9725 mp = dirrem->dm_list.wk_mp; 9726 ump = VFSTOUFS(mp); 9727 flags |= LK_EXCLUSIVE; 9728 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9729 return (EBUSY); 9730 ip = VTOI(vp); 9731 ACQUIRE_LOCK(ump); 9732 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9733 panic("handle_workitem_remove: lost inodedep"); 9734 if (dirrem->dm_state & ONDEPLIST) 9735 LIST_REMOVE(dirrem, dm_inonext); 9736 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9737 ("handle_workitem_remove: Journal entries not written.")); 9738 9739 /* 9740 * Move all dependencies waiting on the remove to complete 9741 * from the dirrem to the inode inowait list to be completed 9742 * after the inode has been updated and written to disk. Any 9743 * marked MKDIR_PARENT are saved to be completed when the .. ref 9744 * is removed. 9745 */ 9746 LIST_INIT(&dotdotwk); 9747 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9748 WORKLIST_REMOVE(wk); 9749 if (wk->wk_state & MKDIR_PARENT) { 9750 wk->wk_state &= ~MKDIR_PARENT; 9751 WORKLIST_INSERT(&dotdotwk, wk); 9752 continue; 9753 } 9754 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9755 } 9756 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9757 /* 9758 * Normal file deletion. 9759 */ 9760 if ((dirrem->dm_state & RMDIR) == 0) { 9761 ip->i_nlink--; 9762 DIP_SET(ip, i_nlink, ip->i_nlink); 9763 ip->i_flag |= IN_CHANGE; 9764 if (ip->i_nlink < ip->i_effnlink) 9765 panic("handle_workitem_remove: bad file delta"); 9766 if (ip->i_nlink == 0) 9767 unlinked_inodedep(mp, inodedep); 9768 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9769 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9770 ("handle_workitem_remove: worklist not empty. %s", 9771 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9772 WORKITEM_FREE(dirrem, D_DIRREM); 9773 FREE_LOCK(ump); 9774 goto out; 9775 } 9776 /* 9777 * Directory deletion. Decrement reference count for both the 9778 * just deleted parent directory entry and the reference for ".". 9779 * Arrange to have the reference count on the parent decremented 9780 * to account for the loss of "..". 9781 */ 9782 ip->i_nlink -= 2; 9783 DIP_SET(ip, i_nlink, ip->i_nlink); 9784 ip->i_flag |= IN_CHANGE; 9785 if (ip->i_nlink < ip->i_effnlink) 9786 panic("handle_workitem_remove: bad dir delta"); 9787 if (ip->i_nlink == 0) 9788 unlinked_inodedep(mp, inodedep); 9789 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9790 /* 9791 * Rename a directory to a new parent. Since, we are both deleting 9792 * and creating a new directory entry, the link count on the new 9793 * directory should not change. Thus we skip the followup dirrem. 9794 */ 9795 if (dirrem->dm_state & DIRCHG) { 9796 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9797 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9798 WORKITEM_FREE(dirrem, D_DIRREM); 9799 FREE_LOCK(ump); 9800 goto out; 9801 } 9802 dirrem->dm_state = ONDEPLIST; 9803 dirrem->dm_oldinum = dirrem->dm_dirinum; 9804 /* 9805 * Place the dirrem on the parent's diremhd list. 9806 */ 9807 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9808 panic("handle_workitem_remove: lost dir inodedep"); 9809 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9810 /* 9811 * If the allocated inode has never been written to disk, then 9812 * the on-disk inode is zero'ed and we can remove the file 9813 * immediately. When journaling if the inode has been marked 9814 * unlinked and not DEPCOMPLETE we know it can never be written. 9815 */ 9816 inodedep_lookup(mp, oldinum, 0, &inodedep); 9817 if (inodedep == NULL || 9818 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9819 check_inode_unwritten(inodedep)) { 9820 FREE_LOCK(ump); 9821 vput(vp); 9822 return handle_workitem_remove(dirrem, flags); 9823 } 9824 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9825 FREE_LOCK(ump); 9826 ip->i_flag |= IN_CHANGE; 9827 out: 9828 ffs_update(vp, 0); 9829 vput(vp); 9830 return (0); 9831 } 9832 9833 /* 9834 * Inode de-allocation dependencies. 9835 * 9836 * When an inode's link count is reduced to zero, it can be de-allocated. We 9837 * found it convenient to postpone de-allocation until after the inode is 9838 * written to disk with its new link count (zero). At this point, all of the 9839 * on-disk inode's block pointers are nullified and, with careful dependency 9840 * list ordering, all dependencies related to the inode will be satisfied and 9841 * the corresponding dependency structures de-allocated. So, if/when the 9842 * inode is reused, there will be no mixing of old dependencies with new 9843 * ones. This artificial dependency is set up by the block de-allocation 9844 * procedure above (softdep_setup_freeblocks) and completed by the 9845 * following procedure. 9846 */ 9847 static void 9848 handle_workitem_freefile(freefile) 9849 struct freefile *freefile; 9850 { 9851 struct workhead wkhd; 9852 struct fs *fs; 9853 struct inodedep *idp; 9854 struct ufsmount *ump; 9855 int error; 9856 9857 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9858 fs = ump->um_fs; 9859 #ifdef DEBUG 9860 ACQUIRE_LOCK(ump); 9861 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9862 FREE_LOCK(ump); 9863 if (error) 9864 panic("handle_workitem_freefile: inodedep %p survived", idp); 9865 #endif 9866 UFS_LOCK(ump); 9867 fs->fs_pendinginodes -= 1; 9868 UFS_UNLOCK(ump); 9869 LIST_INIT(&wkhd); 9870 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9871 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9872 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9873 softdep_error("handle_workitem_freefile", error); 9874 ACQUIRE_LOCK(ump); 9875 WORKITEM_FREE(freefile, D_FREEFILE); 9876 FREE_LOCK(ump); 9877 } 9878 9879 9880 /* 9881 * Helper function which unlinks marker element from work list and returns 9882 * the next element on the list. 9883 */ 9884 static __inline struct worklist * 9885 markernext(struct worklist *marker) 9886 { 9887 struct worklist *next; 9888 9889 next = LIST_NEXT(marker, wk_list); 9890 LIST_REMOVE(marker, wk_list); 9891 return next; 9892 } 9893 9894 /* 9895 * Disk writes. 9896 * 9897 * The dependency structures constructed above are most actively used when file 9898 * system blocks are written to disk. No constraints are placed on when a 9899 * block can be written, but unsatisfied update dependencies are made safe by 9900 * modifying (or replacing) the source memory for the duration of the disk 9901 * write. When the disk write completes, the memory block is again brought 9902 * up-to-date. 9903 * 9904 * In-core inode structure reclamation. 9905 * 9906 * Because there are a finite number of "in-core" inode structures, they are 9907 * reused regularly. By transferring all inode-related dependencies to the 9908 * in-memory inode block and indexing them separately (via "inodedep"s), we 9909 * can allow "in-core" inode structures to be reused at any time and avoid 9910 * any increase in contention. 9911 * 9912 * Called just before entering the device driver to initiate a new disk I/O. 9913 * The buffer must be locked, thus, no I/O completion operations can occur 9914 * while we are manipulating its associated dependencies. 9915 */ 9916 static void 9917 softdep_disk_io_initiation(bp) 9918 struct buf *bp; /* structure describing disk write to occur */ 9919 { 9920 struct worklist *wk; 9921 struct worklist marker; 9922 struct inodedep *inodedep; 9923 struct freeblks *freeblks; 9924 struct jblkdep *jblkdep; 9925 struct newblk *newblk; 9926 struct ufsmount *ump; 9927 9928 /* 9929 * We only care about write operations. There should never 9930 * be dependencies for reads. 9931 */ 9932 if (bp->b_iocmd != BIO_WRITE) 9933 panic("softdep_disk_io_initiation: not write"); 9934 9935 if (bp->b_vflags & BV_BKGRDINPROG) 9936 panic("softdep_disk_io_initiation: Writing buffer with " 9937 "background write in progress: %p", bp); 9938 9939 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 9940 return; 9941 ump = VFSTOUFS(wk->wk_mp); 9942 9943 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9944 PHOLD(curproc); /* Don't swap out kernel stack */ 9945 ACQUIRE_LOCK(ump); 9946 /* 9947 * Do any necessary pre-I/O processing. 9948 */ 9949 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9950 wk = markernext(&marker)) { 9951 LIST_INSERT_AFTER(wk, &marker, wk_list); 9952 switch (wk->wk_type) { 9953 9954 case D_PAGEDEP: 9955 initiate_write_filepage(WK_PAGEDEP(wk), bp); 9956 continue; 9957 9958 case D_INODEDEP: 9959 inodedep = WK_INODEDEP(wk); 9960 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 9961 initiate_write_inodeblock_ufs1(inodedep, bp); 9962 else 9963 initiate_write_inodeblock_ufs2(inodedep, bp); 9964 continue; 9965 9966 case D_INDIRDEP: 9967 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 9968 continue; 9969 9970 case D_BMSAFEMAP: 9971 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 9972 continue; 9973 9974 case D_JSEG: 9975 WK_JSEG(wk)->js_buf = NULL; 9976 continue; 9977 9978 case D_FREEBLKS: 9979 freeblks = WK_FREEBLKS(wk); 9980 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 9981 /* 9982 * We have to wait for the freeblks to be journaled 9983 * before we can write an inodeblock with updated 9984 * pointers. Be careful to arrange the marker so 9985 * we revisit the freeblks if it's not removed by 9986 * the first jwait(). 9987 */ 9988 if (jblkdep != NULL) { 9989 LIST_REMOVE(&marker, wk_list); 9990 LIST_INSERT_BEFORE(wk, &marker, wk_list); 9991 jwait(&jblkdep->jb_list, MNT_WAIT); 9992 } 9993 continue; 9994 case D_ALLOCDIRECT: 9995 case D_ALLOCINDIR: 9996 /* 9997 * We have to wait for the jnewblk to be journaled 9998 * before we can write to a block if the contents 9999 * may be confused with an earlier file's indirect 10000 * at recovery time. Handle the marker as described 10001 * above. 10002 */ 10003 newblk = WK_NEWBLK(wk); 10004 if (newblk->nb_jnewblk != NULL && 10005 indirblk_lookup(newblk->nb_list.wk_mp, 10006 newblk->nb_newblkno)) { 10007 LIST_REMOVE(&marker, wk_list); 10008 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10009 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10010 } 10011 continue; 10012 10013 case D_SBDEP: 10014 initiate_write_sbdep(WK_SBDEP(wk)); 10015 continue; 10016 10017 case D_MKDIR: 10018 case D_FREEWORK: 10019 case D_FREEDEP: 10020 case D_JSEGDEP: 10021 continue; 10022 10023 default: 10024 panic("handle_disk_io_initiation: Unexpected type %s", 10025 TYPENAME(wk->wk_type)); 10026 /* NOTREACHED */ 10027 } 10028 } 10029 FREE_LOCK(ump); 10030 PRELE(curproc); /* Allow swapout of kernel stack */ 10031 } 10032 10033 /* 10034 * Called from within the procedure above to deal with unsatisfied 10035 * allocation dependencies in a directory. The buffer must be locked, 10036 * thus, no I/O completion operations can occur while we are 10037 * manipulating its associated dependencies. 10038 */ 10039 static void 10040 initiate_write_filepage(pagedep, bp) 10041 struct pagedep *pagedep; 10042 struct buf *bp; 10043 { 10044 struct jremref *jremref; 10045 struct jmvref *jmvref; 10046 struct dirrem *dirrem; 10047 struct diradd *dap; 10048 struct direct *ep; 10049 int i; 10050 10051 if (pagedep->pd_state & IOSTARTED) { 10052 /* 10053 * This can only happen if there is a driver that does not 10054 * understand chaining. Here biodone will reissue the call 10055 * to strategy for the incomplete buffers. 10056 */ 10057 printf("initiate_write_filepage: already started\n"); 10058 return; 10059 } 10060 pagedep->pd_state |= IOSTARTED; 10061 /* 10062 * Wait for all journal remove dependencies to hit the disk. 10063 * We can not allow any potentially conflicting directory adds 10064 * to be visible before removes and rollback is too difficult. 10065 * The per-filesystem lock may be dropped and re-acquired, however 10066 * we hold the buf locked so the dependency can not go away. 10067 */ 10068 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10069 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10070 jwait(&jremref->jr_list, MNT_WAIT); 10071 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10072 jwait(&jmvref->jm_list, MNT_WAIT); 10073 for (i = 0; i < DAHASHSZ; i++) { 10074 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10075 ep = (struct direct *) 10076 ((char *)bp->b_data + dap->da_offset); 10077 if (ep->d_ino != dap->da_newinum) 10078 panic("%s: dir inum %ju != new %ju", 10079 "initiate_write_filepage", 10080 (uintmax_t)ep->d_ino, 10081 (uintmax_t)dap->da_newinum); 10082 if (dap->da_state & DIRCHG) 10083 ep->d_ino = dap->da_previous->dm_oldinum; 10084 else 10085 ep->d_ino = 0; 10086 dap->da_state &= ~ATTACHED; 10087 dap->da_state |= UNDONE; 10088 } 10089 } 10090 } 10091 10092 /* 10093 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10094 * Note that any bug fixes made to this routine must be done in the 10095 * version found below. 10096 * 10097 * Called from within the procedure above to deal with unsatisfied 10098 * allocation dependencies in an inodeblock. The buffer must be 10099 * locked, thus, no I/O completion operations can occur while we 10100 * are manipulating its associated dependencies. 10101 */ 10102 static void 10103 initiate_write_inodeblock_ufs1(inodedep, bp) 10104 struct inodedep *inodedep; 10105 struct buf *bp; /* The inode block */ 10106 { 10107 struct allocdirect *adp, *lastadp; 10108 struct ufs1_dinode *dp; 10109 struct ufs1_dinode *sip; 10110 struct inoref *inoref; 10111 struct ufsmount *ump; 10112 struct fs *fs; 10113 ufs_lbn_t i; 10114 #ifdef INVARIANTS 10115 ufs_lbn_t prevlbn = 0; 10116 #endif 10117 int deplist; 10118 10119 if (inodedep->id_state & IOSTARTED) 10120 panic("initiate_write_inodeblock_ufs1: already started"); 10121 inodedep->id_state |= IOSTARTED; 10122 fs = inodedep->id_fs; 10123 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10124 LOCK_OWNED(ump); 10125 dp = (struct ufs1_dinode *)bp->b_data + 10126 ino_to_fsbo(fs, inodedep->id_ino); 10127 10128 /* 10129 * If we're on the unlinked list but have not yet written our 10130 * next pointer initialize it here. 10131 */ 10132 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10133 struct inodedep *inon; 10134 10135 inon = TAILQ_NEXT(inodedep, id_unlinked); 10136 dp->di_freelink = inon ? inon->id_ino : 0; 10137 } 10138 /* 10139 * If the bitmap is not yet written, then the allocated 10140 * inode cannot be written to disk. 10141 */ 10142 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10143 if (inodedep->id_savedino1 != NULL) 10144 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10145 FREE_LOCK(ump); 10146 sip = malloc(sizeof(struct ufs1_dinode), 10147 M_SAVEDINO, M_SOFTDEP_FLAGS); 10148 ACQUIRE_LOCK(ump); 10149 inodedep->id_savedino1 = sip; 10150 *inodedep->id_savedino1 = *dp; 10151 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10152 dp->di_gen = inodedep->id_savedino1->di_gen; 10153 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10154 return; 10155 } 10156 /* 10157 * If no dependencies, then there is nothing to roll back. 10158 */ 10159 inodedep->id_savedsize = dp->di_size; 10160 inodedep->id_savedextsize = 0; 10161 inodedep->id_savednlink = dp->di_nlink; 10162 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10163 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10164 return; 10165 /* 10166 * Revert the link count to that of the first unwritten journal entry. 10167 */ 10168 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10169 if (inoref) 10170 dp->di_nlink = inoref->if_nlink; 10171 /* 10172 * Set the dependencies to busy. 10173 */ 10174 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10175 adp = TAILQ_NEXT(adp, ad_next)) { 10176 #ifdef INVARIANTS 10177 if (deplist != 0 && prevlbn >= adp->ad_offset) 10178 panic("softdep_write_inodeblock: lbn order"); 10179 prevlbn = adp->ad_offset; 10180 if (adp->ad_offset < NDADDR && 10181 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10182 panic("%s: direct pointer #%jd mismatch %d != %jd", 10183 "softdep_write_inodeblock", 10184 (intmax_t)adp->ad_offset, 10185 dp->di_db[adp->ad_offset], 10186 (intmax_t)adp->ad_newblkno); 10187 if (adp->ad_offset >= NDADDR && 10188 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10189 panic("%s: indirect pointer #%jd mismatch %d != %jd", 10190 "softdep_write_inodeblock", 10191 (intmax_t)adp->ad_offset - NDADDR, 10192 dp->di_ib[adp->ad_offset - NDADDR], 10193 (intmax_t)adp->ad_newblkno); 10194 deplist |= 1 << adp->ad_offset; 10195 if ((adp->ad_state & ATTACHED) == 0) 10196 panic("softdep_write_inodeblock: Unknown state 0x%x", 10197 adp->ad_state); 10198 #endif /* INVARIANTS */ 10199 adp->ad_state &= ~ATTACHED; 10200 adp->ad_state |= UNDONE; 10201 } 10202 /* 10203 * The on-disk inode cannot claim to be any larger than the last 10204 * fragment that has been written. Otherwise, the on-disk inode 10205 * might have fragments that were not the last block in the file 10206 * which would corrupt the filesystem. 10207 */ 10208 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10209 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10210 if (adp->ad_offset >= NDADDR) 10211 break; 10212 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10213 /* keep going until hitting a rollback to a frag */ 10214 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10215 continue; 10216 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10217 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10218 #ifdef INVARIANTS 10219 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10220 panic("softdep_write_inodeblock: lost dep1"); 10221 #endif /* INVARIANTS */ 10222 dp->di_db[i] = 0; 10223 } 10224 for (i = 0; i < NIADDR; i++) { 10225 #ifdef INVARIANTS 10226 if (dp->di_ib[i] != 0 && 10227 (deplist & ((1 << NDADDR) << i)) == 0) 10228 panic("softdep_write_inodeblock: lost dep2"); 10229 #endif /* INVARIANTS */ 10230 dp->di_ib[i] = 0; 10231 } 10232 return; 10233 } 10234 /* 10235 * If we have zero'ed out the last allocated block of the file, 10236 * roll back the size to the last currently allocated block. 10237 * We know that this last allocated block is a full-sized as 10238 * we already checked for fragments in the loop above. 10239 */ 10240 if (lastadp != NULL && 10241 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10242 for (i = lastadp->ad_offset; i >= 0; i--) 10243 if (dp->di_db[i] != 0) 10244 break; 10245 dp->di_size = (i + 1) * fs->fs_bsize; 10246 } 10247 /* 10248 * The only dependencies are for indirect blocks. 10249 * 10250 * The file size for indirect block additions is not guaranteed. 10251 * Such a guarantee would be non-trivial to achieve. The conventional 10252 * synchronous write implementation also does not make this guarantee. 10253 * Fsck should catch and fix discrepancies. Arguably, the file size 10254 * can be over-estimated without destroying integrity when the file 10255 * moves into the indirect blocks (i.e., is large). If we want to 10256 * postpone fsck, we are stuck with this argument. 10257 */ 10258 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10259 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10260 } 10261 10262 /* 10263 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10264 * Note that any bug fixes made to this routine must be done in the 10265 * version found above. 10266 * 10267 * Called from within the procedure above to deal with unsatisfied 10268 * allocation dependencies in an inodeblock. The buffer must be 10269 * locked, thus, no I/O completion operations can occur while we 10270 * are manipulating its associated dependencies. 10271 */ 10272 static void 10273 initiate_write_inodeblock_ufs2(inodedep, bp) 10274 struct inodedep *inodedep; 10275 struct buf *bp; /* The inode block */ 10276 { 10277 struct allocdirect *adp, *lastadp; 10278 struct ufs2_dinode *dp; 10279 struct ufs2_dinode *sip; 10280 struct inoref *inoref; 10281 struct ufsmount *ump; 10282 struct fs *fs; 10283 ufs_lbn_t i; 10284 #ifdef INVARIANTS 10285 ufs_lbn_t prevlbn = 0; 10286 #endif 10287 int deplist; 10288 10289 if (inodedep->id_state & IOSTARTED) 10290 panic("initiate_write_inodeblock_ufs2: already started"); 10291 inodedep->id_state |= IOSTARTED; 10292 fs = inodedep->id_fs; 10293 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10294 LOCK_OWNED(ump); 10295 dp = (struct ufs2_dinode *)bp->b_data + 10296 ino_to_fsbo(fs, inodedep->id_ino); 10297 10298 /* 10299 * If we're on the unlinked list but have not yet written our 10300 * next pointer initialize it here. 10301 */ 10302 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10303 struct inodedep *inon; 10304 10305 inon = TAILQ_NEXT(inodedep, id_unlinked); 10306 dp->di_freelink = inon ? inon->id_ino : 0; 10307 } 10308 /* 10309 * If the bitmap is not yet written, then the allocated 10310 * inode cannot be written to disk. 10311 */ 10312 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10313 if (inodedep->id_savedino2 != NULL) 10314 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10315 FREE_LOCK(ump); 10316 sip = malloc(sizeof(struct ufs2_dinode), 10317 M_SAVEDINO, M_SOFTDEP_FLAGS); 10318 ACQUIRE_LOCK(ump); 10319 inodedep->id_savedino2 = sip; 10320 *inodedep->id_savedino2 = *dp; 10321 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10322 dp->di_gen = inodedep->id_savedino2->di_gen; 10323 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10324 return; 10325 } 10326 /* 10327 * If no dependencies, then there is nothing to roll back. 10328 */ 10329 inodedep->id_savedsize = dp->di_size; 10330 inodedep->id_savedextsize = dp->di_extsize; 10331 inodedep->id_savednlink = dp->di_nlink; 10332 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10333 TAILQ_EMPTY(&inodedep->id_extupdt) && 10334 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10335 return; 10336 /* 10337 * Revert the link count to that of the first unwritten journal entry. 10338 */ 10339 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10340 if (inoref) 10341 dp->di_nlink = inoref->if_nlink; 10342 10343 /* 10344 * Set the ext data dependencies to busy. 10345 */ 10346 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10347 adp = TAILQ_NEXT(adp, ad_next)) { 10348 #ifdef INVARIANTS 10349 if (deplist != 0 && prevlbn >= adp->ad_offset) 10350 panic("softdep_write_inodeblock: lbn order"); 10351 prevlbn = adp->ad_offset; 10352 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10353 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10354 "softdep_write_inodeblock", 10355 (intmax_t)adp->ad_offset, 10356 (intmax_t)dp->di_extb[adp->ad_offset], 10357 (intmax_t)adp->ad_newblkno); 10358 deplist |= 1 << adp->ad_offset; 10359 if ((adp->ad_state & ATTACHED) == 0) 10360 panic("softdep_write_inodeblock: Unknown state 0x%x", 10361 adp->ad_state); 10362 #endif /* INVARIANTS */ 10363 adp->ad_state &= ~ATTACHED; 10364 adp->ad_state |= UNDONE; 10365 } 10366 /* 10367 * The on-disk inode cannot claim to be any larger than the last 10368 * fragment that has been written. Otherwise, the on-disk inode 10369 * might have fragments that were not the last block in the ext 10370 * data which would corrupt the filesystem. 10371 */ 10372 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10373 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10374 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10375 /* keep going until hitting a rollback to a frag */ 10376 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10377 continue; 10378 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10379 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 10380 #ifdef INVARIANTS 10381 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10382 panic("softdep_write_inodeblock: lost dep1"); 10383 #endif /* INVARIANTS */ 10384 dp->di_extb[i] = 0; 10385 } 10386 lastadp = NULL; 10387 break; 10388 } 10389 /* 10390 * If we have zero'ed out the last allocated block of the ext 10391 * data, roll back the size to the last currently allocated block. 10392 * We know that this last allocated block is a full-sized as 10393 * we already checked for fragments in the loop above. 10394 */ 10395 if (lastadp != NULL && 10396 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10397 for (i = lastadp->ad_offset; i >= 0; i--) 10398 if (dp->di_extb[i] != 0) 10399 break; 10400 dp->di_extsize = (i + 1) * fs->fs_bsize; 10401 } 10402 /* 10403 * Set the file data dependencies to busy. 10404 */ 10405 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10406 adp = TAILQ_NEXT(adp, ad_next)) { 10407 #ifdef INVARIANTS 10408 if (deplist != 0 && prevlbn >= adp->ad_offset) 10409 panic("softdep_write_inodeblock: lbn order"); 10410 if ((adp->ad_state & ATTACHED) == 0) 10411 panic("inodedep %p and adp %p not attached", inodedep, adp); 10412 prevlbn = adp->ad_offset; 10413 if (adp->ad_offset < NDADDR && 10414 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10415 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10416 "softdep_write_inodeblock", 10417 (intmax_t)adp->ad_offset, 10418 (intmax_t)dp->di_db[adp->ad_offset], 10419 (intmax_t)adp->ad_newblkno); 10420 if (adp->ad_offset >= NDADDR && 10421 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10422 panic("%s indirect pointer #%jd mismatch %jd != %jd", 10423 "softdep_write_inodeblock:", 10424 (intmax_t)adp->ad_offset - NDADDR, 10425 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 10426 (intmax_t)adp->ad_newblkno); 10427 deplist |= 1 << adp->ad_offset; 10428 if ((adp->ad_state & ATTACHED) == 0) 10429 panic("softdep_write_inodeblock: Unknown state 0x%x", 10430 adp->ad_state); 10431 #endif /* INVARIANTS */ 10432 adp->ad_state &= ~ATTACHED; 10433 adp->ad_state |= UNDONE; 10434 } 10435 /* 10436 * The on-disk inode cannot claim to be any larger than the last 10437 * fragment that has been written. Otherwise, the on-disk inode 10438 * might have fragments that were not the last block in the file 10439 * which would corrupt the filesystem. 10440 */ 10441 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10442 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10443 if (adp->ad_offset >= NDADDR) 10444 break; 10445 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10446 /* keep going until hitting a rollback to a frag */ 10447 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10448 continue; 10449 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10450 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10451 #ifdef INVARIANTS 10452 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10453 panic("softdep_write_inodeblock: lost dep2"); 10454 #endif /* INVARIANTS */ 10455 dp->di_db[i] = 0; 10456 } 10457 for (i = 0; i < NIADDR; i++) { 10458 #ifdef INVARIANTS 10459 if (dp->di_ib[i] != 0 && 10460 (deplist & ((1 << NDADDR) << i)) == 0) 10461 panic("softdep_write_inodeblock: lost dep3"); 10462 #endif /* INVARIANTS */ 10463 dp->di_ib[i] = 0; 10464 } 10465 return; 10466 } 10467 /* 10468 * If we have zero'ed out the last allocated block of the file, 10469 * roll back the size to the last currently allocated block. 10470 * We know that this last allocated block is a full-sized as 10471 * we already checked for fragments in the loop above. 10472 */ 10473 if (lastadp != NULL && 10474 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10475 for (i = lastadp->ad_offset; i >= 0; i--) 10476 if (dp->di_db[i] != 0) 10477 break; 10478 dp->di_size = (i + 1) * fs->fs_bsize; 10479 } 10480 /* 10481 * The only dependencies are for indirect blocks. 10482 * 10483 * The file size for indirect block additions is not guaranteed. 10484 * Such a guarantee would be non-trivial to achieve. The conventional 10485 * synchronous write implementation also does not make this guarantee. 10486 * Fsck should catch and fix discrepancies. Arguably, the file size 10487 * can be over-estimated without destroying integrity when the file 10488 * moves into the indirect blocks (i.e., is large). If we want to 10489 * postpone fsck, we are stuck with this argument. 10490 */ 10491 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10492 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10493 } 10494 10495 /* 10496 * Cancel an indirdep as a result of truncation. Release all of the 10497 * children allocindirs and place their journal work on the appropriate 10498 * list. 10499 */ 10500 static void 10501 cancel_indirdep(indirdep, bp, freeblks) 10502 struct indirdep *indirdep; 10503 struct buf *bp; 10504 struct freeblks *freeblks; 10505 { 10506 struct allocindir *aip; 10507 10508 /* 10509 * None of the indirect pointers will ever be visible, 10510 * so they can simply be tossed. GOINGAWAY ensures 10511 * that allocated pointers will be saved in the buffer 10512 * cache until they are freed. Note that they will 10513 * only be able to be found by their physical address 10514 * since the inode mapping the logical address will 10515 * be gone. The save buffer used for the safe copy 10516 * was allocated in setup_allocindir_phase2 using 10517 * the physical address so it could be used for this 10518 * purpose. Hence we swap the safe copy with the real 10519 * copy, allowing the safe copy to be freed and holding 10520 * on to the real copy for later use in indir_trunc. 10521 */ 10522 if (indirdep->ir_state & GOINGAWAY) 10523 panic("cancel_indirdep: already gone"); 10524 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10525 indirdep->ir_state |= DEPCOMPLETE; 10526 LIST_REMOVE(indirdep, ir_next); 10527 } 10528 indirdep->ir_state |= GOINGAWAY; 10529 /* 10530 * Pass in bp for blocks still have journal writes 10531 * pending so we can cancel them on their own. 10532 */ 10533 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0) 10534 cancel_allocindir(aip, bp, freeblks, 0); 10535 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) 10536 cancel_allocindir(aip, NULL, freeblks, 0); 10537 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) 10538 cancel_allocindir(aip, NULL, freeblks, 0); 10539 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0) 10540 cancel_allocindir(aip, NULL, freeblks, 0); 10541 /* 10542 * If there are pending partial truncations we need to keep the 10543 * old block copy around until they complete. This is because 10544 * the current b_data is not a perfect superset of the available 10545 * blocks. 10546 */ 10547 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10548 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10549 else 10550 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10551 WORKLIST_REMOVE(&indirdep->ir_list); 10552 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10553 indirdep->ir_bp = NULL; 10554 indirdep->ir_freeblks = freeblks; 10555 } 10556 10557 /* 10558 * Free an indirdep once it no longer has new pointers to track. 10559 */ 10560 static void 10561 free_indirdep(indirdep) 10562 struct indirdep *indirdep; 10563 { 10564 10565 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10566 ("free_indirdep: Indir trunc list not empty.")); 10567 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10568 ("free_indirdep: Complete head not empty.")); 10569 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10570 ("free_indirdep: write head not empty.")); 10571 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10572 ("free_indirdep: done head not empty.")); 10573 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10574 ("free_indirdep: deplist head not empty.")); 10575 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10576 ("free_indirdep: %p still on newblk list.", indirdep)); 10577 KASSERT(indirdep->ir_saveddata == NULL, 10578 ("free_indirdep: %p still has saved data.", indirdep)); 10579 if (indirdep->ir_state & ONWORKLIST) 10580 WORKLIST_REMOVE(&indirdep->ir_list); 10581 WORKITEM_FREE(indirdep, D_INDIRDEP); 10582 } 10583 10584 /* 10585 * Called before a write to an indirdep. This routine is responsible for 10586 * rolling back pointers to a safe state which includes only those 10587 * allocindirs which have been completed. 10588 */ 10589 static void 10590 initiate_write_indirdep(indirdep, bp) 10591 struct indirdep *indirdep; 10592 struct buf *bp; 10593 { 10594 struct ufsmount *ump; 10595 10596 indirdep->ir_state |= IOSTARTED; 10597 if (indirdep->ir_state & GOINGAWAY) 10598 panic("disk_io_initiation: indirdep gone"); 10599 /* 10600 * If there are no remaining dependencies, this will be writing 10601 * the real pointers. 10602 */ 10603 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10604 TAILQ_EMPTY(&indirdep->ir_trunc)) 10605 return; 10606 /* 10607 * Replace up-to-date version with safe version. 10608 */ 10609 if (indirdep->ir_saveddata == NULL) { 10610 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10611 LOCK_OWNED(ump); 10612 FREE_LOCK(ump); 10613 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10614 M_SOFTDEP_FLAGS); 10615 ACQUIRE_LOCK(ump); 10616 } 10617 indirdep->ir_state &= ~ATTACHED; 10618 indirdep->ir_state |= UNDONE; 10619 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10620 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10621 bp->b_bcount); 10622 } 10623 10624 /* 10625 * Called when an inode has been cleared in a cg bitmap. This finally 10626 * eliminates any canceled jaddrefs 10627 */ 10628 void 10629 softdep_setup_inofree(mp, bp, ino, wkhd) 10630 struct mount *mp; 10631 struct buf *bp; 10632 ino_t ino; 10633 struct workhead *wkhd; 10634 { 10635 struct worklist *wk, *wkn; 10636 struct inodedep *inodedep; 10637 struct ufsmount *ump; 10638 uint8_t *inosused; 10639 struct cg *cgp; 10640 struct fs *fs; 10641 10642 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10643 ("softdep_setup_inofree called on non-softdep filesystem")); 10644 ump = VFSTOUFS(mp); 10645 ACQUIRE_LOCK(ump); 10646 fs = ump->um_fs; 10647 cgp = (struct cg *)bp->b_data; 10648 inosused = cg_inosused(cgp); 10649 if (isset(inosused, ino % fs->fs_ipg)) 10650 panic("softdep_setup_inofree: inode %ju not freed.", 10651 (uintmax_t)ino); 10652 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10653 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10654 (uintmax_t)ino, inodedep); 10655 if (wkhd) { 10656 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10657 if (wk->wk_type != D_JADDREF) 10658 continue; 10659 WORKLIST_REMOVE(wk); 10660 /* 10661 * We can free immediately even if the jaddref 10662 * isn't attached in a background write as now 10663 * the bitmaps are reconciled. 10664 */ 10665 wk->wk_state |= COMPLETE | ATTACHED; 10666 free_jaddref(WK_JADDREF(wk)); 10667 } 10668 jwork_move(&bp->b_dep, wkhd); 10669 } 10670 FREE_LOCK(ump); 10671 } 10672 10673 10674 /* 10675 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10676 * map. Any dependencies waiting for the write to clear are added to the 10677 * buf's list and any jnewblks that are being canceled are discarded 10678 * immediately. 10679 */ 10680 void 10681 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10682 struct mount *mp; 10683 struct buf *bp; 10684 ufs2_daddr_t blkno; 10685 int frags; 10686 struct workhead *wkhd; 10687 { 10688 struct bmsafemap *bmsafemap; 10689 struct jnewblk *jnewblk; 10690 struct ufsmount *ump; 10691 struct worklist *wk; 10692 struct fs *fs; 10693 #ifdef SUJ_DEBUG 10694 uint8_t *blksfree; 10695 struct cg *cgp; 10696 ufs2_daddr_t jstart; 10697 ufs2_daddr_t jend; 10698 ufs2_daddr_t end; 10699 long bno; 10700 int i; 10701 #endif 10702 10703 CTR3(KTR_SUJ, 10704 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10705 blkno, frags, wkhd); 10706 10707 ump = VFSTOUFS(mp); 10708 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10709 ("softdep_setup_blkfree called on non-softdep filesystem")); 10710 ACQUIRE_LOCK(ump); 10711 /* Lookup the bmsafemap so we track when it is dirty. */ 10712 fs = ump->um_fs; 10713 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10714 /* 10715 * Detach any jnewblks which have been canceled. They must linger 10716 * until the bitmap is cleared again by ffs_blkfree() to prevent 10717 * an unjournaled allocation from hitting the disk. 10718 */ 10719 if (wkhd) { 10720 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10721 CTR2(KTR_SUJ, 10722 "softdep_setup_blkfree: blkno %jd wk type %d", 10723 blkno, wk->wk_type); 10724 WORKLIST_REMOVE(wk); 10725 if (wk->wk_type != D_JNEWBLK) { 10726 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10727 continue; 10728 } 10729 jnewblk = WK_JNEWBLK(wk); 10730 KASSERT(jnewblk->jn_state & GOINGAWAY, 10731 ("softdep_setup_blkfree: jnewblk not canceled.")); 10732 #ifdef SUJ_DEBUG 10733 /* 10734 * Assert that this block is free in the bitmap 10735 * before we discard the jnewblk. 10736 */ 10737 cgp = (struct cg *)bp->b_data; 10738 blksfree = cg_blksfree(cgp); 10739 bno = dtogd(fs, jnewblk->jn_blkno); 10740 for (i = jnewblk->jn_oldfrags; 10741 i < jnewblk->jn_frags; i++) { 10742 if (isset(blksfree, bno + i)) 10743 continue; 10744 panic("softdep_setup_blkfree: not free"); 10745 } 10746 #endif 10747 /* 10748 * Even if it's not attached we can free immediately 10749 * as the new bitmap is correct. 10750 */ 10751 wk->wk_state |= COMPLETE | ATTACHED; 10752 free_jnewblk(jnewblk); 10753 } 10754 } 10755 10756 #ifdef SUJ_DEBUG 10757 /* 10758 * Assert that we are not freeing a block which has an outstanding 10759 * allocation dependency. 10760 */ 10761 fs = VFSTOUFS(mp)->um_fs; 10762 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10763 end = blkno + frags; 10764 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10765 /* 10766 * Don't match against blocks that will be freed when the 10767 * background write is done. 10768 */ 10769 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10770 (COMPLETE | DEPCOMPLETE)) 10771 continue; 10772 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10773 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10774 if ((blkno >= jstart && blkno < jend) || 10775 (end > jstart && end <= jend)) { 10776 printf("state 0x%X %jd - %d %d dep %p\n", 10777 jnewblk->jn_state, jnewblk->jn_blkno, 10778 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10779 jnewblk->jn_dep); 10780 panic("softdep_setup_blkfree: " 10781 "%jd-%jd(%d) overlaps with %jd-%jd", 10782 blkno, end, frags, jstart, jend); 10783 } 10784 } 10785 #endif 10786 FREE_LOCK(ump); 10787 } 10788 10789 /* 10790 * Revert a block allocation when the journal record that describes it 10791 * is not yet written. 10792 */ 10793 static int 10794 jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10795 struct jnewblk *jnewblk; 10796 struct fs *fs; 10797 struct cg *cgp; 10798 uint8_t *blksfree; 10799 { 10800 ufs1_daddr_t fragno; 10801 long cgbno, bbase; 10802 int frags, blk; 10803 int i; 10804 10805 frags = 0; 10806 cgbno = dtogd(fs, jnewblk->jn_blkno); 10807 /* 10808 * We have to test which frags need to be rolled back. We may 10809 * be operating on a stale copy when doing background writes. 10810 */ 10811 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10812 if (isclr(blksfree, cgbno + i)) 10813 frags++; 10814 if (frags == 0) 10815 return (0); 10816 /* 10817 * This is mostly ffs_blkfree() sans some validation and 10818 * superblock updates. 10819 */ 10820 if (frags == fs->fs_frag) { 10821 fragno = fragstoblks(fs, cgbno); 10822 ffs_setblock(fs, blksfree, fragno); 10823 ffs_clusteracct(fs, cgp, fragno, 1); 10824 cgp->cg_cs.cs_nbfree++; 10825 } else { 10826 cgbno += jnewblk->jn_oldfrags; 10827 bbase = cgbno - fragnum(fs, cgbno); 10828 /* Decrement the old frags. */ 10829 blk = blkmap(fs, blksfree, bbase); 10830 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10831 /* Deallocate the fragment */ 10832 for (i = 0; i < frags; i++) 10833 setbit(blksfree, cgbno + i); 10834 cgp->cg_cs.cs_nffree += frags; 10835 /* Add back in counts associated with the new frags */ 10836 blk = blkmap(fs, blksfree, bbase); 10837 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10838 /* If a complete block has been reassembled, account for it. */ 10839 fragno = fragstoblks(fs, bbase); 10840 if (ffs_isblock(fs, blksfree, fragno)) { 10841 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10842 ffs_clusteracct(fs, cgp, fragno, 1); 10843 cgp->cg_cs.cs_nbfree++; 10844 } 10845 } 10846 stat_jnewblk++; 10847 jnewblk->jn_state &= ~ATTACHED; 10848 jnewblk->jn_state |= UNDONE; 10849 10850 return (frags); 10851 } 10852 10853 static void 10854 initiate_write_bmsafemap(bmsafemap, bp) 10855 struct bmsafemap *bmsafemap; 10856 struct buf *bp; /* The cg block. */ 10857 { 10858 struct jaddref *jaddref; 10859 struct jnewblk *jnewblk; 10860 uint8_t *inosused; 10861 uint8_t *blksfree; 10862 struct cg *cgp; 10863 struct fs *fs; 10864 ino_t ino; 10865 10866 if (bmsafemap->sm_state & IOSTARTED) 10867 return; 10868 bmsafemap->sm_state |= IOSTARTED; 10869 /* 10870 * Clear any inode allocations which are pending journal writes. 10871 */ 10872 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10873 cgp = (struct cg *)bp->b_data; 10874 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10875 inosused = cg_inosused(cgp); 10876 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10877 ino = jaddref->ja_ino % fs->fs_ipg; 10878 if (isset(inosused, ino)) { 10879 if ((jaddref->ja_mode & IFMT) == IFDIR) 10880 cgp->cg_cs.cs_ndir--; 10881 cgp->cg_cs.cs_nifree++; 10882 clrbit(inosused, ino); 10883 jaddref->ja_state &= ~ATTACHED; 10884 jaddref->ja_state |= UNDONE; 10885 stat_jaddref++; 10886 } else 10887 panic("initiate_write_bmsafemap: inode %ju " 10888 "marked free", (uintmax_t)jaddref->ja_ino); 10889 } 10890 } 10891 /* 10892 * Clear any block allocations which are pending journal writes. 10893 */ 10894 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10895 cgp = (struct cg *)bp->b_data; 10896 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10897 blksfree = cg_blksfree(cgp); 10898 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10899 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10900 continue; 10901 panic("initiate_write_bmsafemap: block %jd " 10902 "marked free", jnewblk->jn_blkno); 10903 } 10904 } 10905 /* 10906 * Move allocation lists to the written lists so they can be 10907 * cleared once the block write is complete. 10908 */ 10909 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10910 inodedep, id_deps); 10911 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10912 newblk, nb_deps); 10913 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10914 wk_list); 10915 } 10916 10917 /* 10918 * This routine is called during the completion interrupt 10919 * service routine for a disk write (from the procedure called 10920 * by the device driver to inform the filesystem caches of 10921 * a request completion). It should be called early in this 10922 * procedure, before the block is made available to other 10923 * processes or other routines are called. 10924 * 10925 */ 10926 static void 10927 softdep_disk_write_complete(bp) 10928 struct buf *bp; /* describes the completed disk write */ 10929 { 10930 struct worklist *wk; 10931 struct worklist *owk; 10932 struct ufsmount *ump; 10933 struct workhead reattach; 10934 struct freeblks *freeblks; 10935 struct buf *sbp; 10936 10937 /* 10938 * If an error occurred while doing the write, then the data 10939 * has not hit the disk and the dependencies cannot be unrolled. 10940 */ 10941 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) 10942 return; 10943 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 10944 return; 10945 ump = VFSTOUFS(wk->wk_mp); 10946 LIST_INIT(&reattach); 10947 /* 10948 * This lock must not be released anywhere in this code segment. 10949 */ 10950 sbp = NULL; 10951 owk = NULL; 10952 ACQUIRE_LOCK(ump); 10953 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 10954 WORKLIST_REMOVE(wk); 10955 atomic_add_long(&dep_write[wk->wk_type], 1); 10956 if (wk == owk) 10957 panic("duplicate worklist: %p\n", wk); 10958 owk = wk; 10959 switch (wk->wk_type) { 10960 10961 case D_PAGEDEP: 10962 if (handle_written_filepage(WK_PAGEDEP(wk), bp)) 10963 WORKLIST_INSERT(&reattach, wk); 10964 continue; 10965 10966 case D_INODEDEP: 10967 if (handle_written_inodeblock(WK_INODEDEP(wk), bp)) 10968 WORKLIST_INSERT(&reattach, wk); 10969 continue; 10970 10971 case D_BMSAFEMAP: 10972 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp)) 10973 WORKLIST_INSERT(&reattach, wk); 10974 continue; 10975 10976 case D_MKDIR: 10977 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 10978 continue; 10979 10980 case D_ALLOCDIRECT: 10981 wk->wk_state |= COMPLETE; 10982 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 10983 continue; 10984 10985 case D_ALLOCINDIR: 10986 wk->wk_state |= COMPLETE; 10987 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 10988 continue; 10989 10990 case D_INDIRDEP: 10991 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp)) 10992 WORKLIST_INSERT(&reattach, wk); 10993 continue; 10994 10995 case D_FREEBLKS: 10996 wk->wk_state |= COMPLETE; 10997 freeblks = WK_FREEBLKS(wk); 10998 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 10999 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11000 add_to_worklist(wk, WK_NODELAY); 11001 continue; 11002 11003 case D_FREEWORK: 11004 handle_written_freework(WK_FREEWORK(wk)); 11005 break; 11006 11007 case D_JSEGDEP: 11008 free_jsegdep(WK_JSEGDEP(wk)); 11009 continue; 11010 11011 case D_JSEG: 11012 handle_written_jseg(WK_JSEG(wk), bp); 11013 continue; 11014 11015 case D_SBDEP: 11016 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11017 WORKLIST_INSERT(&reattach, wk); 11018 continue; 11019 11020 case D_FREEDEP: 11021 free_freedep(WK_FREEDEP(wk)); 11022 continue; 11023 11024 default: 11025 panic("handle_disk_write_complete: Unknown type %s", 11026 TYPENAME(wk->wk_type)); 11027 /* NOTREACHED */ 11028 } 11029 } 11030 /* 11031 * Reattach any requests that must be redone. 11032 */ 11033 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11034 WORKLIST_REMOVE(wk); 11035 WORKLIST_INSERT(&bp->b_dep, wk); 11036 } 11037 FREE_LOCK(ump); 11038 if (sbp) 11039 brelse(sbp); 11040 } 11041 11042 /* 11043 * Called from within softdep_disk_write_complete above. Note that 11044 * this routine is always called from interrupt level with further 11045 * splbio interrupts blocked. 11046 */ 11047 static void 11048 handle_allocdirect_partdone(adp, wkhd) 11049 struct allocdirect *adp; /* the completed allocdirect */ 11050 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11051 { 11052 struct allocdirectlst *listhead; 11053 struct allocdirect *listadp; 11054 struct inodedep *inodedep; 11055 long bsize; 11056 11057 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11058 return; 11059 /* 11060 * The on-disk inode cannot claim to be any larger than the last 11061 * fragment that has been written. Otherwise, the on-disk inode 11062 * might have fragments that were not the last block in the file 11063 * which would corrupt the filesystem. Thus, we cannot free any 11064 * allocdirects after one whose ad_oldblkno claims a fragment as 11065 * these blocks must be rolled back to zero before writing the inode. 11066 * We check the currently active set of allocdirects in id_inoupdt 11067 * or id_extupdt as appropriate. 11068 */ 11069 inodedep = adp->ad_inodedep; 11070 bsize = inodedep->id_fs->fs_bsize; 11071 if (adp->ad_state & EXTDATA) 11072 listhead = &inodedep->id_extupdt; 11073 else 11074 listhead = &inodedep->id_inoupdt; 11075 TAILQ_FOREACH(listadp, listhead, ad_next) { 11076 /* found our block */ 11077 if (listadp == adp) 11078 break; 11079 /* continue if ad_oldlbn is not a fragment */ 11080 if (listadp->ad_oldsize == 0 || 11081 listadp->ad_oldsize == bsize) 11082 continue; 11083 /* hit a fragment */ 11084 return; 11085 } 11086 /* 11087 * If we have reached the end of the current list without 11088 * finding the just finished dependency, then it must be 11089 * on the future dependency list. Future dependencies cannot 11090 * be freed until they are moved to the current list. 11091 */ 11092 if (listadp == NULL) { 11093 #ifdef DEBUG 11094 if (adp->ad_state & EXTDATA) 11095 listhead = &inodedep->id_newextupdt; 11096 else 11097 listhead = &inodedep->id_newinoupdt; 11098 TAILQ_FOREACH(listadp, listhead, ad_next) 11099 /* found our block */ 11100 if (listadp == adp) 11101 break; 11102 if (listadp == NULL) 11103 panic("handle_allocdirect_partdone: lost dep"); 11104 #endif /* DEBUG */ 11105 return; 11106 } 11107 /* 11108 * If we have found the just finished dependency, then queue 11109 * it along with anything that follows it that is complete. 11110 * Since the pointer has not yet been written in the inode 11111 * as the dependency prevents it, place the allocdirect on the 11112 * bufwait list where it will be freed once the pointer is 11113 * valid. 11114 */ 11115 if (wkhd == NULL) 11116 wkhd = &inodedep->id_bufwait; 11117 for (; adp; adp = listadp) { 11118 listadp = TAILQ_NEXT(adp, ad_next); 11119 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11120 return; 11121 TAILQ_REMOVE(listhead, adp, ad_next); 11122 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11123 } 11124 } 11125 11126 /* 11127 * Called from within softdep_disk_write_complete above. This routine 11128 * completes successfully written allocindirs. 11129 */ 11130 static void 11131 handle_allocindir_partdone(aip) 11132 struct allocindir *aip; /* the completed allocindir */ 11133 { 11134 struct indirdep *indirdep; 11135 11136 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11137 return; 11138 indirdep = aip->ai_indirdep; 11139 LIST_REMOVE(aip, ai_next); 11140 /* 11141 * Don't set a pointer while the buffer is undergoing IO or while 11142 * we have active truncations. 11143 */ 11144 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11145 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11146 return; 11147 } 11148 if (indirdep->ir_state & UFS1FMT) 11149 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11150 aip->ai_newblkno; 11151 else 11152 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11153 aip->ai_newblkno; 11154 /* 11155 * Await the pointer write before freeing the allocindir. 11156 */ 11157 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11158 } 11159 11160 /* 11161 * Release segments held on a jwork list. 11162 */ 11163 static void 11164 handle_jwork(wkhd) 11165 struct workhead *wkhd; 11166 { 11167 struct worklist *wk; 11168 11169 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11170 WORKLIST_REMOVE(wk); 11171 switch (wk->wk_type) { 11172 case D_JSEGDEP: 11173 free_jsegdep(WK_JSEGDEP(wk)); 11174 continue; 11175 case D_FREEDEP: 11176 free_freedep(WK_FREEDEP(wk)); 11177 continue; 11178 case D_FREEFRAG: 11179 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11180 WORKITEM_FREE(wk, D_FREEFRAG); 11181 continue; 11182 case D_FREEWORK: 11183 handle_written_freework(WK_FREEWORK(wk)); 11184 continue; 11185 default: 11186 panic("handle_jwork: Unknown type %s\n", 11187 TYPENAME(wk->wk_type)); 11188 } 11189 } 11190 } 11191 11192 /* 11193 * Handle the bufwait list on an inode when it is safe to release items 11194 * held there. This normally happens after an inode block is written but 11195 * may be delayed and handled later if there are pending journal items that 11196 * are not yet safe to be released. 11197 */ 11198 static struct freefile * 11199 handle_bufwait(inodedep, refhd) 11200 struct inodedep *inodedep; 11201 struct workhead *refhd; 11202 { 11203 struct jaddref *jaddref; 11204 struct freefile *freefile; 11205 struct worklist *wk; 11206 11207 freefile = NULL; 11208 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11209 WORKLIST_REMOVE(wk); 11210 switch (wk->wk_type) { 11211 case D_FREEFILE: 11212 /* 11213 * We defer adding freefile to the worklist 11214 * until all other additions have been made to 11215 * ensure that it will be done after all the 11216 * old blocks have been freed. 11217 */ 11218 if (freefile != NULL) 11219 panic("handle_bufwait: freefile"); 11220 freefile = WK_FREEFILE(wk); 11221 continue; 11222 11223 case D_MKDIR: 11224 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11225 continue; 11226 11227 case D_DIRADD: 11228 diradd_inode_written(WK_DIRADD(wk), inodedep); 11229 continue; 11230 11231 case D_FREEFRAG: 11232 wk->wk_state |= COMPLETE; 11233 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11234 add_to_worklist(wk, 0); 11235 continue; 11236 11237 case D_DIRREM: 11238 wk->wk_state |= COMPLETE; 11239 add_to_worklist(wk, 0); 11240 continue; 11241 11242 case D_ALLOCDIRECT: 11243 case D_ALLOCINDIR: 11244 free_newblk(WK_NEWBLK(wk)); 11245 continue; 11246 11247 case D_JNEWBLK: 11248 wk->wk_state |= COMPLETE; 11249 free_jnewblk(WK_JNEWBLK(wk)); 11250 continue; 11251 11252 /* 11253 * Save freed journal segments and add references on 11254 * the supplied list which will delay their release 11255 * until the cg bitmap is cleared on disk. 11256 */ 11257 case D_JSEGDEP: 11258 if (refhd == NULL) 11259 free_jsegdep(WK_JSEGDEP(wk)); 11260 else 11261 WORKLIST_INSERT(refhd, wk); 11262 continue; 11263 11264 case D_JADDREF: 11265 jaddref = WK_JADDREF(wk); 11266 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11267 if_deps); 11268 /* 11269 * Transfer any jaddrefs to the list to be freed with 11270 * the bitmap if we're handling a removed file. 11271 */ 11272 if (refhd == NULL) { 11273 wk->wk_state |= COMPLETE; 11274 free_jaddref(jaddref); 11275 } else 11276 WORKLIST_INSERT(refhd, wk); 11277 continue; 11278 11279 default: 11280 panic("handle_bufwait: Unknown type %p(%s)", 11281 wk, TYPENAME(wk->wk_type)); 11282 /* NOTREACHED */ 11283 } 11284 } 11285 return (freefile); 11286 } 11287 /* 11288 * Called from within softdep_disk_write_complete above to restore 11289 * in-memory inode block contents to their most up-to-date state. Note 11290 * that this routine is always called from interrupt level with further 11291 * splbio interrupts blocked. 11292 */ 11293 static int 11294 handle_written_inodeblock(inodedep, bp) 11295 struct inodedep *inodedep; 11296 struct buf *bp; /* buffer containing the inode block */ 11297 { 11298 struct freefile *freefile; 11299 struct allocdirect *adp, *nextadp; 11300 struct ufs1_dinode *dp1 = NULL; 11301 struct ufs2_dinode *dp2 = NULL; 11302 struct workhead wkhd; 11303 int hadchanges, fstype; 11304 ino_t freelink; 11305 11306 LIST_INIT(&wkhd); 11307 hadchanges = 0; 11308 freefile = NULL; 11309 if ((inodedep->id_state & IOSTARTED) == 0) 11310 panic("handle_written_inodeblock: not started"); 11311 inodedep->id_state &= ~IOSTARTED; 11312 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11313 fstype = UFS1; 11314 dp1 = (struct ufs1_dinode *)bp->b_data + 11315 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11316 freelink = dp1->di_freelink; 11317 } else { 11318 fstype = UFS2; 11319 dp2 = (struct ufs2_dinode *)bp->b_data + 11320 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11321 freelink = dp2->di_freelink; 11322 } 11323 /* 11324 * Leave this inodeblock dirty until it's in the list. 11325 */ 11326 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) { 11327 struct inodedep *inon; 11328 11329 inon = TAILQ_NEXT(inodedep, id_unlinked); 11330 if ((inon == NULL && freelink == 0) || 11331 (inon && inon->id_ino == freelink)) { 11332 if (inon) 11333 inon->id_state |= UNLINKPREV; 11334 inodedep->id_state |= UNLINKNEXT; 11335 } 11336 hadchanges = 1; 11337 } 11338 /* 11339 * If we had to rollback the inode allocation because of 11340 * bitmaps being incomplete, then simply restore it. 11341 * Keep the block dirty so that it will not be reclaimed until 11342 * all associated dependencies have been cleared and the 11343 * corresponding updates written to disk. 11344 */ 11345 if (inodedep->id_savedino1 != NULL) { 11346 hadchanges = 1; 11347 if (fstype == UFS1) 11348 *dp1 = *inodedep->id_savedino1; 11349 else 11350 *dp2 = *inodedep->id_savedino2; 11351 free(inodedep->id_savedino1, M_SAVEDINO); 11352 inodedep->id_savedino1 = NULL; 11353 if ((bp->b_flags & B_DELWRI) == 0) 11354 stat_inode_bitmap++; 11355 bdirty(bp); 11356 /* 11357 * If the inode is clear here and GOINGAWAY it will never 11358 * be written. Process the bufwait and clear any pending 11359 * work which may include the freefile. 11360 */ 11361 if (inodedep->id_state & GOINGAWAY) 11362 goto bufwait; 11363 return (1); 11364 } 11365 inodedep->id_state |= COMPLETE; 11366 /* 11367 * Roll forward anything that had to be rolled back before 11368 * the inode could be updated. 11369 */ 11370 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11371 nextadp = TAILQ_NEXT(adp, ad_next); 11372 if (adp->ad_state & ATTACHED) 11373 panic("handle_written_inodeblock: new entry"); 11374 if (fstype == UFS1) { 11375 if (adp->ad_offset < NDADDR) { 11376 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11377 panic("%s %s #%jd mismatch %d != %jd", 11378 "handle_written_inodeblock:", 11379 "direct pointer", 11380 (intmax_t)adp->ad_offset, 11381 dp1->di_db[adp->ad_offset], 11382 (intmax_t)adp->ad_oldblkno); 11383 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11384 } else { 11385 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 11386 panic("%s: %s #%jd allocated as %d", 11387 "handle_written_inodeblock", 11388 "indirect pointer", 11389 (intmax_t)adp->ad_offset - NDADDR, 11390 dp1->di_ib[adp->ad_offset - NDADDR]); 11391 dp1->di_ib[adp->ad_offset - NDADDR] = 11392 adp->ad_newblkno; 11393 } 11394 } else { 11395 if (adp->ad_offset < NDADDR) { 11396 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11397 panic("%s: %s #%jd %s %jd != %jd", 11398 "handle_written_inodeblock", 11399 "direct pointer", 11400 (intmax_t)adp->ad_offset, "mismatch", 11401 (intmax_t)dp2->di_db[adp->ad_offset], 11402 (intmax_t)adp->ad_oldblkno); 11403 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11404 } else { 11405 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 11406 panic("%s: %s #%jd allocated as %jd", 11407 "handle_written_inodeblock", 11408 "indirect pointer", 11409 (intmax_t)adp->ad_offset - NDADDR, 11410 (intmax_t) 11411 dp2->di_ib[adp->ad_offset - NDADDR]); 11412 dp2->di_ib[adp->ad_offset - NDADDR] = 11413 adp->ad_newblkno; 11414 } 11415 } 11416 adp->ad_state &= ~UNDONE; 11417 adp->ad_state |= ATTACHED; 11418 hadchanges = 1; 11419 } 11420 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11421 nextadp = TAILQ_NEXT(adp, ad_next); 11422 if (adp->ad_state & ATTACHED) 11423 panic("handle_written_inodeblock: new entry"); 11424 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11425 panic("%s: direct pointers #%jd %s %jd != %jd", 11426 "handle_written_inodeblock", 11427 (intmax_t)adp->ad_offset, "mismatch", 11428 (intmax_t)dp2->di_extb[adp->ad_offset], 11429 (intmax_t)adp->ad_oldblkno); 11430 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11431 adp->ad_state &= ~UNDONE; 11432 adp->ad_state |= ATTACHED; 11433 hadchanges = 1; 11434 } 11435 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11436 stat_direct_blk_ptrs++; 11437 /* 11438 * Reset the file size to its most up-to-date value. 11439 */ 11440 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11441 panic("handle_written_inodeblock: bad size"); 11442 if (inodedep->id_savednlink > LINK_MAX) 11443 panic("handle_written_inodeblock: Invalid link count " 11444 "%d for inodedep %p", inodedep->id_savednlink, inodedep); 11445 if (fstype == UFS1) { 11446 if (dp1->di_nlink != inodedep->id_savednlink) { 11447 dp1->di_nlink = inodedep->id_savednlink; 11448 hadchanges = 1; 11449 } 11450 if (dp1->di_size != inodedep->id_savedsize) { 11451 dp1->di_size = inodedep->id_savedsize; 11452 hadchanges = 1; 11453 } 11454 } else { 11455 if (dp2->di_nlink != inodedep->id_savednlink) { 11456 dp2->di_nlink = inodedep->id_savednlink; 11457 hadchanges = 1; 11458 } 11459 if (dp2->di_size != inodedep->id_savedsize) { 11460 dp2->di_size = inodedep->id_savedsize; 11461 hadchanges = 1; 11462 } 11463 if (dp2->di_extsize != inodedep->id_savedextsize) { 11464 dp2->di_extsize = inodedep->id_savedextsize; 11465 hadchanges = 1; 11466 } 11467 } 11468 inodedep->id_savedsize = -1; 11469 inodedep->id_savedextsize = -1; 11470 inodedep->id_savednlink = -1; 11471 /* 11472 * If there were any rollbacks in the inode block, then it must be 11473 * marked dirty so that its will eventually get written back in 11474 * its correct form. 11475 */ 11476 if (hadchanges) 11477 bdirty(bp); 11478 bufwait: 11479 /* 11480 * Process any allocdirects that completed during the update. 11481 */ 11482 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11483 handle_allocdirect_partdone(adp, &wkhd); 11484 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11485 handle_allocdirect_partdone(adp, &wkhd); 11486 /* 11487 * Process deallocations that were held pending until the 11488 * inode had been written to disk. Freeing of the inode 11489 * is delayed until after all blocks have been freed to 11490 * avoid creation of new <vfsid, inum, lbn> triples 11491 * before the old ones have been deleted. Completely 11492 * unlinked inodes are not processed until the unlinked 11493 * inode list is written or the last reference is removed. 11494 */ 11495 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11496 freefile = handle_bufwait(inodedep, NULL); 11497 if (freefile && !LIST_EMPTY(&wkhd)) { 11498 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11499 freefile = NULL; 11500 } 11501 } 11502 /* 11503 * Move rolled forward dependency completions to the bufwait list 11504 * now that those that were already written have been processed. 11505 */ 11506 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11507 panic("handle_written_inodeblock: bufwait but no changes"); 11508 jwork_move(&inodedep->id_bufwait, &wkhd); 11509 11510 if (freefile != NULL) { 11511 /* 11512 * If the inode is goingaway it was never written. Fake up 11513 * the state here so free_inodedep() can succeed. 11514 */ 11515 if (inodedep->id_state & GOINGAWAY) 11516 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11517 if (free_inodedep(inodedep) == 0) 11518 panic("handle_written_inodeblock: live inodedep %p", 11519 inodedep); 11520 add_to_worklist(&freefile->fx_list, 0); 11521 return (0); 11522 } 11523 11524 /* 11525 * If no outstanding dependencies, free it. 11526 */ 11527 if (free_inodedep(inodedep) || 11528 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11529 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11530 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11531 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11532 return (0); 11533 return (hadchanges); 11534 } 11535 11536 static int 11537 handle_written_indirdep(indirdep, bp, bpp) 11538 struct indirdep *indirdep; 11539 struct buf *bp; 11540 struct buf **bpp; 11541 { 11542 struct allocindir *aip; 11543 struct buf *sbp; 11544 int chgs; 11545 11546 if (indirdep->ir_state & GOINGAWAY) 11547 panic("handle_written_indirdep: indirdep gone"); 11548 if ((indirdep->ir_state & IOSTARTED) == 0) 11549 panic("handle_written_indirdep: IO not started"); 11550 chgs = 0; 11551 /* 11552 * If there were rollbacks revert them here. 11553 */ 11554 if (indirdep->ir_saveddata) { 11555 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11556 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11557 free(indirdep->ir_saveddata, M_INDIRDEP); 11558 indirdep->ir_saveddata = NULL; 11559 } 11560 chgs = 1; 11561 } 11562 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11563 indirdep->ir_state |= ATTACHED; 11564 /* 11565 * Move allocindirs with written pointers to the completehd if 11566 * the indirdep's pointer is not yet written. Otherwise 11567 * free them here. 11568 */ 11569 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) { 11570 LIST_REMOVE(aip, ai_next); 11571 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11572 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11573 ai_next); 11574 newblk_freefrag(&aip->ai_block); 11575 continue; 11576 } 11577 free_newblk(&aip->ai_block); 11578 } 11579 /* 11580 * Move allocindirs that have finished dependency processing from 11581 * the done list to the write list after updating the pointers. 11582 */ 11583 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11584 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) { 11585 handle_allocindir_partdone(aip); 11586 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11587 panic("disk_write_complete: not gone"); 11588 chgs = 1; 11589 } 11590 } 11591 /* 11592 * Preserve the indirdep if there were any changes or if it is not 11593 * yet valid on disk. 11594 */ 11595 if (chgs) { 11596 stat_indir_blk_ptrs++; 11597 bdirty(bp); 11598 return (1); 11599 } 11600 /* 11601 * If there were no changes we can discard the savedbp and detach 11602 * ourselves from the buf. We are only carrying completed pointers 11603 * in this case. 11604 */ 11605 sbp = indirdep->ir_savebp; 11606 sbp->b_flags |= B_INVAL | B_NOCACHE; 11607 indirdep->ir_savebp = NULL; 11608 indirdep->ir_bp = NULL; 11609 if (*bpp != NULL) 11610 panic("handle_written_indirdep: bp already exists."); 11611 *bpp = sbp; 11612 /* 11613 * The indirdep may not be freed until its parent points at it. 11614 */ 11615 if (indirdep->ir_state & DEPCOMPLETE) 11616 free_indirdep(indirdep); 11617 11618 return (0); 11619 } 11620 11621 /* 11622 * Process a diradd entry after its dependent inode has been written. 11623 * This routine must be called with splbio interrupts blocked. 11624 */ 11625 static void 11626 diradd_inode_written(dap, inodedep) 11627 struct diradd *dap; 11628 struct inodedep *inodedep; 11629 { 11630 11631 dap->da_state |= COMPLETE; 11632 complete_diradd(dap); 11633 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11634 } 11635 11636 /* 11637 * Returns true if the bmsafemap will have rollbacks when written. Must only 11638 * be called with the per-filesystem lock and the buf lock on the cg held. 11639 */ 11640 static int 11641 bmsafemap_backgroundwrite(bmsafemap, bp) 11642 struct bmsafemap *bmsafemap; 11643 struct buf *bp; 11644 { 11645 int dirty; 11646 11647 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11648 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11649 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11650 /* 11651 * If we're initiating a background write we need to process the 11652 * rollbacks as they exist now, not as they exist when IO starts. 11653 * No other consumers will look at the contents of the shadowed 11654 * buf so this is safe to do here. 11655 */ 11656 if (bp->b_xflags & BX_BKGRDMARKER) 11657 initiate_write_bmsafemap(bmsafemap, bp); 11658 11659 return (dirty); 11660 } 11661 11662 /* 11663 * Re-apply an allocation when a cg write is complete. 11664 */ 11665 static int 11666 jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11667 struct jnewblk *jnewblk; 11668 struct fs *fs; 11669 struct cg *cgp; 11670 uint8_t *blksfree; 11671 { 11672 ufs1_daddr_t fragno; 11673 ufs2_daddr_t blkno; 11674 long cgbno, bbase; 11675 int frags, blk; 11676 int i; 11677 11678 frags = 0; 11679 cgbno = dtogd(fs, jnewblk->jn_blkno); 11680 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11681 if (isclr(blksfree, cgbno + i)) 11682 panic("jnewblk_rollforward: re-allocated fragment"); 11683 frags++; 11684 } 11685 if (frags == fs->fs_frag) { 11686 blkno = fragstoblks(fs, cgbno); 11687 ffs_clrblock(fs, blksfree, (long)blkno); 11688 ffs_clusteracct(fs, cgp, blkno, -1); 11689 cgp->cg_cs.cs_nbfree--; 11690 } else { 11691 bbase = cgbno - fragnum(fs, cgbno); 11692 cgbno += jnewblk->jn_oldfrags; 11693 /* If a complete block had been reassembled, account for it. */ 11694 fragno = fragstoblks(fs, bbase); 11695 if (ffs_isblock(fs, blksfree, fragno)) { 11696 cgp->cg_cs.cs_nffree += fs->fs_frag; 11697 ffs_clusteracct(fs, cgp, fragno, -1); 11698 cgp->cg_cs.cs_nbfree--; 11699 } 11700 /* Decrement the old frags. */ 11701 blk = blkmap(fs, blksfree, bbase); 11702 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11703 /* Allocate the fragment */ 11704 for (i = 0; i < frags; i++) 11705 clrbit(blksfree, cgbno + i); 11706 cgp->cg_cs.cs_nffree -= frags; 11707 /* Add back in counts associated with the new frags */ 11708 blk = blkmap(fs, blksfree, bbase); 11709 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11710 } 11711 return (frags); 11712 } 11713 11714 /* 11715 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11716 * changes if it's not a background write. Set all written dependencies 11717 * to DEPCOMPLETE and free the structure if possible. 11718 */ 11719 static int 11720 handle_written_bmsafemap(bmsafemap, bp) 11721 struct bmsafemap *bmsafemap; 11722 struct buf *bp; 11723 { 11724 struct newblk *newblk; 11725 struct inodedep *inodedep; 11726 struct jaddref *jaddref, *jatmp; 11727 struct jnewblk *jnewblk, *jntmp; 11728 struct ufsmount *ump; 11729 uint8_t *inosused; 11730 uint8_t *blksfree; 11731 struct cg *cgp; 11732 struct fs *fs; 11733 ino_t ino; 11734 int foreground; 11735 int chgs; 11736 11737 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11738 panic("initiate_write_bmsafemap: Not started\n"); 11739 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11740 chgs = 0; 11741 bmsafemap->sm_state &= ~IOSTARTED; 11742 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11743 /* 11744 * Release journal work that was waiting on the write. 11745 */ 11746 handle_jwork(&bmsafemap->sm_freewr); 11747 11748 /* 11749 * Restore unwritten inode allocation pending jaddref writes. 11750 */ 11751 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11752 cgp = (struct cg *)bp->b_data; 11753 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11754 inosused = cg_inosused(cgp); 11755 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11756 ja_bmdeps, jatmp) { 11757 if ((jaddref->ja_state & UNDONE) == 0) 11758 continue; 11759 ino = jaddref->ja_ino % fs->fs_ipg; 11760 if (isset(inosused, ino)) 11761 panic("handle_written_bmsafemap: " 11762 "re-allocated inode"); 11763 /* Do the roll-forward only if it's a real copy. */ 11764 if (foreground) { 11765 if ((jaddref->ja_mode & IFMT) == IFDIR) 11766 cgp->cg_cs.cs_ndir++; 11767 cgp->cg_cs.cs_nifree--; 11768 setbit(inosused, ino); 11769 chgs = 1; 11770 } 11771 jaddref->ja_state &= ~UNDONE; 11772 jaddref->ja_state |= ATTACHED; 11773 free_jaddref(jaddref); 11774 } 11775 } 11776 /* 11777 * Restore any block allocations which are pending journal writes. 11778 */ 11779 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11780 cgp = (struct cg *)bp->b_data; 11781 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11782 blksfree = cg_blksfree(cgp); 11783 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11784 jntmp) { 11785 if ((jnewblk->jn_state & UNDONE) == 0) 11786 continue; 11787 /* Do the roll-forward only if it's a real copy. */ 11788 if (foreground && 11789 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11790 chgs = 1; 11791 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11792 jnewblk->jn_state |= ATTACHED; 11793 free_jnewblk(jnewblk); 11794 } 11795 } 11796 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11797 newblk->nb_state |= DEPCOMPLETE; 11798 newblk->nb_state &= ~ONDEPLIST; 11799 newblk->nb_bmsafemap = NULL; 11800 LIST_REMOVE(newblk, nb_deps); 11801 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11802 handle_allocdirect_partdone( 11803 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11804 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11805 handle_allocindir_partdone( 11806 WK_ALLOCINDIR(&newblk->nb_list)); 11807 else if (newblk->nb_list.wk_type != D_NEWBLK) 11808 panic("handle_written_bmsafemap: Unexpected type: %s", 11809 TYPENAME(newblk->nb_list.wk_type)); 11810 } 11811 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11812 inodedep->id_state |= DEPCOMPLETE; 11813 inodedep->id_state &= ~ONDEPLIST; 11814 LIST_REMOVE(inodedep, id_deps); 11815 inodedep->id_bmsafemap = NULL; 11816 } 11817 LIST_REMOVE(bmsafemap, sm_next); 11818 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11819 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11820 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11821 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11822 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11823 LIST_REMOVE(bmsafemap, sm_hash); 11824 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11825 return (0); 11826 } 11827 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11828 if (foreground) 11829 bdirty(bp); 11830 return (1); 11831 } 11832 11833 /* 11834 * Try to free a mkdir dependency. 11835 */ 11836 static void 11837 complete_mkdir(mkdir) 11838 struct mkdir *mkdir; 11839 { 11840 struct diradd *dap; 11841 11842 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11843 return; 11844 LIST_REMOVE(mkdir, md_mkdirs); 11845 dap = mkdir->md_diradd; 11846 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11847 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11848 dap->da_state |= DEPCOMPLETE; 11849 complete_diradd(dap); 11850 } 11851 WORKITEM_FREE(mkdir, D_MKDIR); 11852 } 11853 11854 /* 11855 * Handle the completion of a mkdir dependency. 11856 */ 11857 static void 11858 handle_written_mkdir(mkdir, type) 11859 struct mkdir *mkdir; 11860 int type; 11861 { 11862 11863 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 11864 panic("handle_written_mkdir: bad type"); 11865 mkdir->md_state |= COMPLETE; 11866 complete_mkdir(mkdir); 11867 } 11868 11869 static int 11870 free_pagedep(pagedep) 11871 struct pagedep *pagedep; 11872 { 11873 int i; 11874 11875 if (pagedep->pd_state & NEWBLOCK) 11876 return (0); 11877 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 11878 return (0); 11879 for (i = 0; i < DAHASHSZ; i++) 11880 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 11881 return (0); 11882 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 11883 return (0); 11884 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 11885 return (0); 11886 if (pagedep->pd_state & ONWORKLIST) 11887 WORKLIST_REMOVE(&pagedep->pd_list); 11888 LIST_REMOVE(pagedep, pd_hash); 11889 WORKITEM_FREE(pagedep, D_PAGEDEP); 11890 11891 return (1); 11892 } 11893 11894 /* 11895 * Called from within softdep_disk_write_complete above. 11896 * A write operation was just completed. Removed inodes can 11897 * now be freed and associated block pointers may be committed. 11898 * Note that this routine is always called from interrupt level 11899 * with further splbio interrupts blocked. 11900 */ 11901 static int 11902 handle_written_filepage(pagedep, bp) 11903 struct pagedep *pagedep; 11904 struct buf *bp; /* buffer containing the written page */ 11905 { 11906 struct dirrem *dirrem; 11907 struct diradd *dap, *nextdap; 11908 struct direct *ep; 11909 int i, chgs; 11910 11911 if ((pagedep->pd_state & IOSTARTED) == 0) 11912 panic("handle_written_filepage: not started"); 11913 pagedep->pd_state &= ~IOSTARTED; 11914 /* 11915 * Process any directory removals that have been committed. 11916 */ 11917 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 11918 LIST_REMOVE(dirrem, dm_next); 11919 dirrem->dm_state |= COMPLETE; 11920 dirrem->dm_dirinum = pagedep->pd_ino; 11921 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 11922 ("handle_written_filepage: Journal entries not written.")); 11923 add_to_worklist(&dirrem->dm_list, 0); 11924 } 11925 /* 11926 * Free any directory additions that have been committed. 11927 * If it is a newly allocated block, we have to wait until 11928 * the on-disk directory inode claims the new block. 11929 */ 11930 if ((pagedep->pd_state & NEWBLOCK) == 0) 11931 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 11932 free_diradd(dap, NULL); 11933 /* 11934 * Uncommitted directory entries must be restored. 11935 */ 11936 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 11937 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 11938 dap = nextdap) { 11939 nextdap = LIST_NEXT(dap, da_pdlist); 11940 if (dap->da_state & ATTACHED) 11941 panic("handle_written_filepage: attached"); 11942 ep = (struct direct *) 11943 ((char *)bp->b_data + dap->da_offset); 11944 ep->d_ino = dap->da_newinum; 11945 dap->da_state &= ~UNDONE; 11946 dap->da_state |= ATTACHED; 11947 chgs = 1; 11948 /* 11949 * If the inode referenced by the directory has 11950 * been written out, then the dependency can be 11951 * moved to the pending list. 11952 */ 11953 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 11954 LIST_REMOVE(dap, da_pdlist); 11955 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 11956 da_pdlist); 11957 } 11958 } 11959 } 11960 /* 11961 * If there were any rollbacks in the directory, then it must be 11962 * marked dirty so that its will eventually get written back in 11963 * its correct form. 11964 */ 11965 if (chgs) { 11966 if ((bp->b_flags & B_DELWRI) == 0) 11967 stat_dir_entry++; 11968 bdirty(bp); 11969 return (1); 11970 } 11971 /* 11972 * If we are not waiting for a new directory block to be 11973 * claimed by its inode, then the pagedep will be freed. 11974 * Otherwise it will remain to track any new entries on 11975 * the page in case they are fsync'ed. 11976 */ 11977 free_pagedep(pagedep); 11978 return (0); 11979 } 11980 11981 /* 11982 * Writing back in-core inode structures. 11983 * 11984 * The filesystem only accesses an inode's contents when it occupies an 11985 * "in-core" inode structure. These "in-core" structures are separate from 11986 * the page frames used to cache inode blocks. Only the latter are 11987 * transferred to/from the disk. So, when the updated contents of the 11988 * "in-core" inode structure are copied to the corresponding in-memory inode 11989 * block, the dependencies are also transferred. The following procedure is 11990 * called when copying a dirty "in-core" inode to a cached inode block. 11991 */ 11992 11993 /* 11994 * Called when an inode is loaded from disk. If the effective link count 11995 * differed from the actual link count when it was last flushed, then we 11996 * need to ensure that the correct effective link count is put back. 11997 */ 11998 void 11999 softdep_load_inodeblock(ip) 12000 struct inode *ip; /* the "in_core" copy of the inode */ 12001 { 12002 struct inodedep *inodedep; 12003 12004 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 12005 ("softdep_load_inodeblock called on non-softdep filesystem")); 12006 /* 12007 * Check for alternate nlink count. 12008 */ 12009 ip->i_effnlink = ip->i_nlink; 12010 ACQUIRE_LOCK(ip->i_ump); 12011 if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 12012 &inodedep) == 0) { 12013 FREE_LOCK(ip->i_ump); 12014 return; 12015 } 12016 ip->i_effnlink -= inodedep->id_nlinkdelta; 12017 FREE_LOCK(ip->i_ump); 12018 } 12019 12020 /* 12021 * This routine is called just before the "in-core" inode 12022 * information is to be copied to the in-memory inode block. 12023 * Recall that an inode block contains several inodes. If 12024 * the force flag is set, then the dependencies will be 12025 * cleared so that the update can always be made. Note that 12026 * the buffer is locked when this routine is called, so we 12027 * will never be in the middle of writing the inode block 12028 * to disk. 12029 */ 12030 void 12031 softdep_update_inodeblock(ip, bp, waitfor) 12032 struct inode *ip; /* the "in_core" copy of the inode */ 12033 struct buf *bp; /* the buffer containing the inode block */ 12034 int waitfor; /* nonzero => update must be allowed */ 12035 { 12036 struct inodedep *inodedep; 12037 struct inoref *inoref; 12038 struct ufsmount *ump; 12039 struct worklist *wk; 12040 struct mount *mp; 12041 struct buf *ibp; 12042 struct fs *fs; 12043 int error; 12044 12045 ump = ip->i_ump; 12046 mp = UFSTOVFS(ump); 12047 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12048 ("softdep_update_inodeblock called on non-softdep filesystem")); 12049 fs = ip->i_fs; 12050 /* 12051 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12052 * does not have access to the in-core ip so must write directly into 12053 * the inode block buffer when setting freelink. 12054 */ 12055 if (fs->fs_magic == FS_UFS1_MAGIC) 12056 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12057 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12058 else 12059 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12060 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12061 /* 12062 * If the effective link count is not equal to the actual link 12063 * count, then we must track the difference in an inodedep while 12064 * the inode is (potentially) tossed out of the cache. Otherwise, 12065 * if there is no existing inodedep, then there are no dependencies 12066 * to track. 12067 */ 12068 ACQUIRE_LOCK(ump); 12069 again: 12070 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12071 FREE_LOCK(ump); 12072 if (ip->i_effnlink != ip->i_nlink) 12073 panic("softdep_update_inodeblock: bad link count"); 12074 return; 12075 } 12076 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12077 panic("softdep_update_inodeblock: bad delta"); 12078 /* 12079 * If we're flushing all dependencies we must also move any waiting 12080 * for journal writes onto the bufwait list prior to I/O. 12081 */ 12082 if (waitfor) { 12083 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12084 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12085 == DEPCOMPLETE) { 12086 jwait(&inoref->if_list, MNT_WAIT); 12087 goto again; 12088 } 12089 } 12090 } 12091 /* 12092 * Changes have been initiated. Anything depending on these 12093 * changes cannot occur until this inode has been written. 12094 */ 12095 inodedep->id_state &= ~COMPLETE; 12096 if ((inodedep->id_state & ONWORKLIST) == 0) 12097 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12098 /* 12099 * Any new dependencies associated with the incore inode must 12100 * now be moved to the list associated with the buffer holding 12101 * the in-memory copy of the inode. Once merged process any 12102 * allocdirects that are completed by the merger. 12103 */ 12104 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12105 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12106 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12107 NULL); 12108 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12109 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12110 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12111 NULL); 12112 /* 12113 * Now that the inode has been pushed into the buffer, the 12114 * operations dependent on the inode being written to disk 12115 * can be moved to the id_bufwait so that they will be 12116 * processed when the buffer I/O completes. 12117 */ 12118 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12119 WORKLIST_REMOVE(wk); 12120 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12121 } 12122 /* 12123 * Newly allocated inodes cannot be written until the bitmap 12124 * that allocates them have been written (indicated by 12125 * DEPCOMPLETE being set in id_state). If we are doing a 12126 * forced sync (e.g., an fsync on a file), we force the bitmap 12127 * to be written so that the update can be done. 12128 */ 12129 if (waitfor == 0) { 12130 FREE_LOCK(ump); 12131 return; 12132 } 12133 retry: 12134 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12135 FREE_LOCK(ump); 12136 return; 12137 } 12138 ibp = inodedep->id_bmsafemap->sm_buf; 12139 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12140 if (ibp == NULL) { 12141 /* 12142 * If ibp came back as NULL, the dependency could have been 12143 * freed while we slept. Look it up again, and check to see 12144 * that it has completed. 12145 */ 12146 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12147 goto retry; 12148 FREE_LOCK(ump); 12149 return; 12150 } 12151 FREE_LOCK(ump); 12152 if ((error = bwrite(ibp)) != 0) 12153 softdep_error("softdep_update_inodeblock: bwrite", error); 12154 } 12155 12156 /* 12157 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12158 * old inode dependency list (such as id_inoupdt). This routine must be 12159 * called with splbio interrupts blocked. 12160 */ 12161 static void 12162 merge_inode_lists(newlisthead, oldlisthead) 12163 struct allocdirectlst *newlisthead; 12164 struct allocdirectlst *oldlisthead; 12165 { 12166 struct allocdirect *listadp, *newadp; 12167 12168 newadp = TAILQ_FIRST(newlisthead); 12169 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12170 if (listadp->ad_offset < newadp->ad_offset) { 12171 listadp = TAILQ_NEXT(listadp, ad_next); 12172 continue; 12173 } 12174 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12175 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12176 if (listadp->ad_offset == newadp->ad_offset) { 12177 allocdirect_merge(oldlisthead, newadp, 12178 listadp); 12179 listadp = newadp; 12180 } 12181 newadp = TAILQ_FIRST(newlisthead); 12182 } 12183 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12184 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12185 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12186 } 12187 } 12188 12189 /* 12190 * If we are doing an fsync, then we must ensure that any directory 12191 * entries for the inode have been written after the inode gets to disk. 12192 */ 12193 int 12194 softdep_fsync(vp) 12195 struct vnode *vp; /* the "in_core" copy of the inode */ 12196 { 12197 struct inodedep *inodedep; 12198 struct pagedep *pagedep; 12199 struct inoref *inoref; 12200 struct ufsmount *ump; 12201 struct worklist *wk; 12202 struct diradd *dap; 12203 struct mount *mp; 12204 struct vnode *pvp; 12205 struct inode *ip; 12206 struct buf *bp; 12207 struct fs *fs; 12208 struct thread *td = curthread; 12209 int error, flushparent, pagedep_new_block; 12210 ino_t parentino; 12211 ufs_lbn_t lbn; 12212 12213 ip = VTOI(vp); 12214 fs = ip->i_fs; 12215 ump = ip->i_ump; 12216 mp = vp->v_mount; 12217 if (MOUNTEDSOFTDEP(mp) == 0) 12218 return (0); 12219 ACQUIRE_LOCK(ump); 12220 restart: 12221 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12222 FREE_LOCK(ump); 12223 return (0); 12224 } 12225 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12226 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12227 == DEPCOMPLETE) { 12228 jwait(&inoref->if_list, MNT_WAIT); 12229 goto restart; 12230 } 12231 } 12232 if (!LIST_EMPTY(&inodedep->id_inowait) || 12233 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12234 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12235 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12236 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12237 panic("softdep_fsync: pending ops %p", inodedep); 12238 for (error = 0, flushparent = 0; ; ) { 12239 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12240 break; 12241 if (wk->wk_type != D_DIRADD) 12242 panic("softdep_fsync: Unexpected type %s", 12243 TYPENAME(wk->wk_type)); 12244 dap = WK_DIRADD(wk); 12245 /* 12246 * Flush our parent if this directory entry has a MKDIR_PARENT 12247 * dependency or is contained in a newly allocated block. 12248 */ 12249 if (dap->da_state & DIRCHG) 12250 pagedep = dap->da_previous->dm_pagedep; 12251 else 12252 pagedep = dap->da_pagedep; 12253 parentino = pagedep->pd_ino; 12254 lbn = pagedep->pd_lbn; 12255 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12256 panic("softdep_fsync: dirty"); 12257 if ((dap->da_state & MKDIR_PARENT) || 12258 (pagedep->pd_state & NEWBLOCK)) 12259 flushparent = 1; 12260 else 12261 flushparent = 0; 12262 /* 12263 * If we are being fsync'ed as part of vgone'ing this vnode, 12264 * then we will not be able to release and recover the 12265 * vnode below, so we just have to give up on writing its 12266 * directory entry out. It will eventually be written, just 12267 * not now, but then the user was not asking to have it 12268 * written, so we are not breaking any promises. 12269 */ 12270 if (vp->v_iflag & VI_DOOMED) 12271 break; 12272 /* 12273 * We prevent deadlock by always fetching inodes from the 12274 * root, moving down the directory tree. Thus, when fetching 12275 * our parent directory, we first try to get the lock. If 12276 * that fails, we must unlock ourselves before requesting 12277 * the lock on our parent. See the comment in ufs_lookup 12278 * for details on possible races. 12279 */ 12280 FREE_LOCK(ump); 12281 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12282 FFSV_FORCEINSMQ)) { 12283 error = vfs_busy(mp, MBF_NOWAIT); 12284 if (error != 0) { 12285 vfs_ref(mp); 12286 VOP_UNLOCK(vp, 0); 12287 error = vfs_busy(mp, 0); 12288 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12289 vfs_rel(mp); 12290 if (error != 0) 12291 return (ENOENT); 12292 if (vp->v_iflag & VI_DOOMED) { 12293 vfs_unbusy(mp); 12294 return (ENOENT); 12295 } 12296 } 12297 VOP_UNLOCK(vp, 0); 12298 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12299 &pvp, FFSV_FORCEINSMQ); 12300 vfs_unbusy(mp); 12301 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12302 if (vp->v_iflag & VI_DOOMED) { 12303 if (error == 0) 12304 vput(pvp); 12305 error = ENOENT; 12306 } 12307 if (error != 0) 12308 return (error); 12309 } 12310 /* 12311 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12312 * that are contained in direct blocks will be resolved by 12313 * doing a ffs_update. Pagedeps contained in indirect blocks 12314 * may require a complete sync'ing of the directory. So, we 12315 * try the cheap and fast ffs_update first, and if that fails, 12316 * then we do the slower ffs_syncvnode of the directory. 12317 */ 12318 if (flushparent) { 12319 int locked; 12320 12321 if ((error = ffs_update(pvp, 1)) != 0) { 12322 vput(pvp); 12323 return (error); 12324 } 12325 ACQUIRE_LOCK(ump); 12326 locked = 1; 12327 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12328 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12329 if (wk->wk_type != D_DIRADD) 12330 panic("softdep_fsync: Unexpected type %s", 12331 TYPENAME(wk->wk_type)); 12332 dap = WK_DIRADD(wk); 12333 if (dap->da_state & DIRCHG) 12334 pagedep = dap->da_previous->dm_pagedep; 12335 else 12336 pagedep = dap->da_pagedep; 12337 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12338 FREE_LOCK(ump); 12339 locked = 0; 12340 if (pagedep_new_block && (error = 12341 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12342 vput(pvp); 12343 return (error); 12344 } 12345 } 12346 } 12347 if (locked) 12348 FREE_LOCK(ump); 12349 } 12350 /* 12351 * Flush directory page containing the inode's name. 12352 */ 12353 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12354 &bp); 12355 if (error == 0) 12356 error = bwrite(bp); 12357 else 12358 brelse(bp); 12359 vput(pvp); 12360 if (error != 0) 12361 return (error); 12362 ACQUIRE_LOCK(ump); 12363 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12364 break; 12365 } 12366 FREE_LOCK(ump); 12367 return (0); 12368 } 12369 12370 /* 12371 * Flush all the dirty bitmaps associated with the block device 12372 * before flushing the rest of the dirty blocks so as to reduce 12373 * the number of dependencies that will have to be rolled back. 12374 * 12375 * XXX Unused? 12376 */ 12377 void 12378 softdep_fsync_mountdev(vp) 12379 struct vnode *vp; 12380 { 12381 struct buf *bp, *nbp; 12382 struct worklist *wk; 12383 struct bufobj *bo; 12384 12385 if (!vn_isdisk(vp, NULL)) 12386 panic("softdep_fsync_mountdev: vnode not a disk"); 12387 bo = &vp->v_bufobj; 12388 restart: 12389 BO_LOCK(bo); 12390 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12391 /* 12392 * If it is already scheduled, skip to the next buffer. 12393 */ 12394 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12395 continue; 12396 12397 if ((bp->b_flags & B_DELWRI) == 0) 12398 panic("softdep_fsync_mountdev: not dirty"); 12399 /* 12400 * We are only interested in bitmaps with outstanding 12401 * dependencies. 12402 */ 12403 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12404 wk->wk_type != D_BMSAFEMAP || 12405 (bp->b_vflags & BV_BKGRDINPROG)) { 12406 BUF_UNLOCK(bp); 12407 continue; 12408 } 12409 BO_UNLOCK(bo); 12410 bremfree(bp); 12411 (void) bawrite(bp); 12412 goto restart; 12413 } 12414 drain_output(vp); 12415 BO_UNLOCK(bo); 12416 } 12417 12418 /* 12419 * Sync all cylinder groups that were dirty at the time this function is 12420 * called. Newly dirtied cgs will be inserted before the sentinel. This 12421 * is used to flush freedep activity that may be holding up writes to a 12422 * indirect block. 12423 */ 12424 static int 12425 sync_cgs(mp, waitfor) 12426 struct mount *mp; 12427 int waitfor; 12428 { 12429 struct bmsafemap *bmsafemap; 12430 struct bmsafemap *sentinel; 12431 struct ufsmount *ump; 12432 struct buf *bp; 12433 int error; 12434 12435 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12436 sentinel->sm_cg = -1; 12437 ump = VFSTOUFS(mp); 12438 error = 0; 12439 ACQUIRE_LOCK(ump); 12440 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12441 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12442 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12443 /* Skip sentinels and cgs with no work to release. */ 12444 if (bmsafemap->sm_cg == -1 || 12445 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12446 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12447 LIST_REMOVE(sentinel, sm_next); 12448 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12449 continue; 12450 } 12451 /* 12452 * If we don't get the lock and we're waiting try again, if 12453 * not move on to the next buf and try to sync it. 12454 */ 12455 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12456 if (bp == NULL && waitfor == MNT_WAIT) 12457 continue; 12458 LIST_REMOVE(sentinel, sm_next); 12459 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12460 if (bp == NULL) 12461 continue; 12462 FREE_LOCK(ump); 12463 if (waitfor == MNT_NOWAIT) 12464 bawrite(bp); 12465 else 12466 error = bwrite(bp); 12467 ACQUIRE_LOCK(ump); 12468 if (error) 12469 break; 12470 } 12471 LIST_REMOVE(sentinel, sm_next); 12472 FREE_LOCK(ump); 12473 free(sentinel, M_BMSAFEMAP); 12474 return (error); 12475 } 12476 12477 /* 12478 * This routine is called when we are trying to synchronously flush a 12479 * file. This routine must eliminate any filesystem metadata dependencies 12480 * so that the syncing routine can succeed. 12481 */ 12482 int 12483 softdep_sync_metadata(struct vnode *vp) 12484 { 12485 struct inode *ip; 12486 int error; 12487 12488 ip = VTOI(vp); 12489 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 12490 ("softdep_sync_metadata called on non-softdep filesystem")); 12491 /* 12492 * Ensure that any direct block dependencies have been cleared, 12493 * truncations are started, and inode references are journaled. 12494 */ 12495 ACQUIRE_LOCK(ip->i_ump); 12496 /* 12497 * Write all journal records to prevent rollbacks on devvp. 12498 */ 12499 if (vp->v_type == VCHR) 12500 softdep_flushjournal(vp->v_mount); 12501 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12502 /* 12503 * Ensure that all truncates are written so we won't find deps on 12504 * indirect blocks. 12505 */ 12506 process_truncates(vp); 12507 FREE_LOCK(ip->i_ump); 12508 12509 return (error); 12510 } 12511 12512 /* 12513 * This routine is called when we are attempting to sync a buf with 12514 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12515 * other IO it can but returns EBUSY if the buffer is not yet able to 12516 * be written. Dependencies which will not cause rollbacks will always 12517 * return 0. 12518 */ 12519 int 12520 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12521 { 12522 struct indirdep *indirdep; 12523 struct pagedep *pagedep; 12524 struct allocindir *aip; 12525 struct newblk *newblk; 12526 struct ufsmount *ump; 12527 struct buf *nbp; 12528 struct worklist *wk; 12529 int i, error; 12530 12531 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12532 ("softdep_sync_buf called on non-softdep filesystem")); 12533 /* 12534 * For VCHR we just don't want to force flush any dependencies that 12535 * will cause rollbacks. 12536 */ 12537 if (vp->v_type == VCHR) { 12538 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12539 return (EBUSY); 12540 return (0); 12541 } 12542 ump = VTOI(vp)->i_ump; 12543 ACQUIRE_LOCK(ump); 12544 /* 12545 * As we hold the buffer locked, none of its dependencies 12546 * will disappear. 12547 */ 12548 error = 0; 12549 top: 12550 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12551 switch (wk->wk_type) { 12552 12553 case D_ALLOCDIRECT: 12554 case D_ALLOCINDIR: 12555 newblk = WK_NEWBLK(wk); 12556 if (newblk->nb_jnewblk != NULL) { 12557 if (waitfor == MNT_NOWAIT) { 12558 error = EBUSY; 12559 goto out_unlock; 12560 } 12561 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12562 goto top; 12563 } 12564 if (newblk->nb_state & DEPCOMPLETE || 12565 waitfor == MNT_NOWAIT) 12566 continue; 12567 nbp = newblk->nb_bmsafemap->sm_buf; 12568 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12569 if (nbp == NULL) 12570 goto top; 12571 FREE_LOCK(ump); 12572 if ((error = bwrite(nbp)) != 0) 12573 goto out; 12574 ACQUIRE_LOCK(ump); 12575 continue; 12576 12577 case D_INDIRDEP: 12578 indirdep = WK_INDIRDEP(wk); 12579 if (waitfor == MNT_NOWAIT) { 12580 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12581 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12582 error = EBUSY; 12583 goto out_unlock; 12584 } 12585 } 12586 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12587 panic("softdep_sync_buf: truncation pending."); 12588 restart: 12589 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12590 newblk = (struct newblk *)aip; 12591 if (newblk->nb_jnewblk != NULL) { 12592 jwait(&newblk->nb_jnewblk->jn_list, 12593 waitfor); 12594 goto restart; 12595 } 12596 if (newblk->nb_state & DEPCOMPLETE) 12597 continue; 12598 nbp = newblk->nb_bmsafemap->sm_buf; 12599 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12600 if (nbp == NULL) 12601 goto restart; 12602 FREE_LOCK(ump); 12603 if ((error = bwrite(nbp)) != 0) 12604 goto out; 12605 ACQUIRE_LOCK(ump); 12606 goto restart; 12607 } 12608 continue; 12609 12610 case D_PAGEDEP: 12611 /* 12612 * Only flush directory entries in synchronous passes. 12613 */ 12614 if (waitfor != MNT_WAIT) { 12615 error = EBUSY; 12616 goto out_unlock; 12617 } 12618 /* 12619 * While syncing snapshots, we must allow recursive 12620 * lookups. 12621 */ 12622 BUF_AREC(bp); 12623 /* 12624 * We are trying to sync a directory that may 12625 * have dependencies on both its own metadata 12626 * and/or dependencies on the inodes of any 12627 * recently allocated files. We walk its diradd 12628 * lists pushing out the associated inode. 12629 */ 12630 pagedep = WK_PAGEDEP(wk); 12631 for (i = 0; i < DAHASHSZ; i++) { 12632 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12633 continue; 12634 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12635 &pagedep->pd_diraddhd[i]))) { 12636 BUF_NOREC(bp); 12637 goto out_unlock; 12638 } 12639 } 12640 BUF_NOREC(bp); 12641 continue; 12642 12643 case D_FREEWORK: 12644 case D_FREEDEP: 12645 case D_JSEGDEP: 12646 case D_JNEWBLK: 12647 continue; 12648 12649 default: 12650 panic("softdep_sync_buf: Unknown type %s", 12651 TYPENAME(wk->wk_type)); 12652 /* NOTREACHED */ 12653 } 12654 } 12655 out_unlock: 12656 FREE_LOCK(ump); 12657 out: 12658 return (error); 12659 } 12660 12661 /* 12662 * Flush the dependencies associated with an inodedep. 12663 * Called with splbio blocked. 12664 */ 12665 static int 12666 flush_inodedep_deps(vp, mp, ino) 12667 struct vnode *vp; 12668 struct mount *mp; 12669 ino_t ino; 12670 { 12671 struct inodedep *inodedep; 12672 struct inoref *inoref; 12673 struct ufsmount *ump; 12674 int error, waitfor; 12675 12676 /* 12677 * This work is done in two passes. The first pass grabs most 12678 * of the buffers and begins asynchronously writing them. The 12679 * only way to wait for these asynchronous writes is to sleep 12680 * on the filesystem vnode which may stay busy for a long time 12681 * if the filesystem is active. So, instead, we make a second 12682 * pass over the dependencies blocking on each write. In the 12683 * usual case we will be blocking against a write that we 12684 * initiated, so when it is done the dependency will have been 12685 * resolved. Thus the second pass is expected to end quickly. 12686 * We give a brief window at the top of the loop to allow 12687 * any pending I/O to complete. 12688 */ 12689 ump = VFSTOUFS(mp); 12690 LOCK_OWNED(ump); 12691 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12692 if (error) 12693 return (error); 12694 FREE_LOCK(ump); 12695 ACQUIRE_LOCK(ump); 12696 restart: 12697 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12698 return (0); 12699 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12700 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12701 == DEPCOMPLETE) { 12702 jwait(&inoref->if_list, MNT_WAIT); 12703 goto restart; 12704 } 12705 } 12706 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12707 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12708 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12709 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12710 continue; 12711 /* 12712 * If pass2, we are done, otherwise do pass 2. 12713 */ 12714 if (waitfor == MNT_WAIT) 12715 break; 12716 waitfor = MNT_WAIT; 12717 } 12718 /* 12719 * Try freeing inodedep in case all dependencies have been removed. 12720 */ 12721 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12722 (void) free_inodedep(inodedep); 12723 return (0); 12724 } 12725 12726 /* 12727 * Flush an inode dependency list. 12728 * Called with splbio blocked. 12729 */ 12730 static int 12731 flush_deplist(listhead, waitfor, errorp) 12732 struct allocdirectlst *listhead; 12733 int waitfor; 12734 int *errorp; 12735 { 12736 struct allocdirect *adp; 12737 struct newblk *newblk; 12738 struct ufsmount *ump; 12739 struct buf *bp; 12740 12741 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12742 return (0); 12743 ump = VFSTOUFS(adp->ad_list.wk_mp); 12744 LOCK_OWNED(ump); 12745 TAILQ_FOREACH(adp, listhead, ad_next) { 12746 newblk = (struct newblk *)adp; 12747 if (newblk->nb_jnewblk != NULL) { 12748 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12749 return (1); 12750 } 12751 if (newblk->nb_state & DEPCOMPLETE) 12752 continue; 12753 bp = newblk->nb_bmsafemap->sm_buf; 12754 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12755 if (bp == NULL) { 12756 if (waitfor == MNT_NOWAIT) 12757 continue; 12758 return (1); 12759 } 12760 FREE_LOCK(ump); 12761 if (waitfor == MNT_NOWAIT) 12762 bawrite(bp); 12763 else 12764 *errorp = bwrite(bp); 12765 ACQUIRE_LOCK(ump); 12766 return (1); 12767 } 12768 return (0); 12769 } 12770 12771 /* 12772 * Flush dependencies associated with an allocdirect block. 12773 */ 12774 static int 12775 flush_newblk_dep(vp, mp, lbn) 12776 struct vnode *vp; 12777 struct mount *mp; 12778 ufs_lbn_t lbn; 12779 { 12780 struct newblk *newblk; 12781 struct ufsmount *ump; 12782 struct bufobj *bo; 12783 struct inode *ip; 12784 struct buf *bp; 12785 ufs2_daddr_t blkno; 12786 int error; 12787 12788 error = 0; 12789 bo = &vp->v_bufobj; 12790 ip = VTOI(vp); 12791 blkno = DIP(ip, i_db[lbn]); 12792 if (blkno == 0) 12793 panic("flush_newblk_dep: Missing block"); 12794 ump = VFSTOUFS(mp); 12795 ACQUIRE_LOCK(ump); 12796 /* 12797 * Loop until all dependencies related to this block are satisfied. 12798 * We must be careful to restart after each sleep in case a write 12799 * completes some part of this process for us. 12800 */ 12801 for (;;) { 12802 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12803 FREE_LOCK(ump); 12804 break; 12805 } 12806 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12807 panic("flush_newblk_deps: Bad newblk %p", newblk); 12808 /* 12809 * Flush the journal. 12810 */ 12811 if (newblk->nb_jnewblk != NULL) { 12812 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12813 continue; 12814 } 12815 /* 12816 * Write the bitmap dependency. 12817 */ 12818 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12819 bp = newblk->nb_bmsafemap->sm_buf; 12820 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12821 if (bp == NULL) 12822 continue; 12823 FREE_LOCK(ump); 12824 error = bwrite(bp); 12825 if (error) 12826 break; 12827 ACQUIRE_LOCK(ump); 12828 continue; 12829 } 12830 /* 12831 * Write the buffer. 12832 */ 12833 FREE_LOCK(ump); 12834 BO_LOCK(bo); 12835 bp = gbincore(bo, lbn); 12836 if (bp != NULL) { 12837 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12838 LK_INTERLOCK, BO_LOCKPTR(bo)); 12839 if (error == ENOLCK) { 12840 ACQUIRE_LOCK(ump); 12841 continue; /* Slept, retry */ 12842 } 12843 if (error != 0) 12844 break; /* Failed */ 12845 if (bp->b_flags & B_DELWRI) { 12846 bremfree(bp); 12847 error = bwrite(bp); 12848 if (error) 12849 break; 12850 } else 12851 BUF_UNLOCK(bp); 12852 } else 12853 BO_UNLOCK(bo); 12854 /* 12855 * We have to wait for the direct pointers to 12856 * point at the newdirblk before the dependency 12857 * will go away. 12858 */ 12859 error = ffs_update(vp, 1); 12860 if (error) 12861 break; 12862 ACQUIRE_LOCK(ump); 12863 } 12864 return (error); 12865 } 12866 12867 /* 12868 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 12869 * Called with splbio blocked. 12870 */ 12871 static int 12872 flush_pagedep_deps(pvp, mp, diraddhdp) 12873 struct vnode *pvp; 12874 struct mount *mp; 12875 struct diraddhd *diraddhdp; 12876 { 12877 struct inodedep *inodedep; 12878 struct inoref *inoref; 12879 struct ufsmount *ump; 12880 struct diradd *dap; 12881 struct vnode *vp; 12882 int error = 0; 12883 struct buf *bp; 12884 ino_t inum; 12885 struct diraddhd unfinished; 12886 12887 LIST_INIT(&unfinished); 12888 ump = VFSTOUFS(mp); 12889 LOCK_OWNED(ump); 12890 restart: 12891 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 12892 /* 12893 * Flush ourselves if this directory entry 12894 * has a MKDIR_PARENT dependency. 12895 */ 12896 if (dap->da_state & MKDIR_PARENT) { 12897 FREE_LOCK(ump); 12898 if ((error = ffs_update(pvp, 1)) != 0) 12899 break; 12900 ACQUIRE_LOCK(ump); 12901 /* 12902 * If that cleared dependencies, go on to next. 12903 */ 12904 if (dap != LIST_FIRST(diraddhdp)) 12905 continue; 12906 /* 12907 * All MKDIR_PARENT dependencies and all the 12908 * NEWBLOCK pagedeps that are contained in direct 12909 * blocks were resolved by doing above ffs_update. 12910 * Pagedeps contained in indirect blocks may 12911 * require a complete sync'ing of the directory. 12912 * We are in the midst of doing a complete sync, 12913 * so if they are not resolved in this pass we 12914 * defer them for now as they will be sync'ed by 12915 * our caller shortly. 12916 */ 12917 LIST_REMOVE(dap, da_pdlist); 12918 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 12919 continue; 12920 } 12921 /* 12922 * A newly allocated directory must have its "." and 12923 * ".." entries written out before its name can be 12924 * committed in its parent. 12925 */ 12926 inum = dap->da_newinum; 12927 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12928 panic("flush_pagedep_deps: lost inode1"); 12929 /* 12930 * Wait for any pending journal adds to complete so we don't 12931 * cause rollbacks while syncing. 12932 */ 12933 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12934 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12935 == DEPCOMPLETE) { 12936 jwait(&inoref->if_list, MNT_WAIT); 12937 goto restart; 12938 } 12939 } 12940 if (dap->da_state & MKDIR_BODY) { 12941 FREE_LOCK(ump); 12942 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 12943 FFSV_FORCEINSMQ))) 12944 break; 12945 error = flush_newblk_dep(vp, mp, 0); 12946 /* 12947 * If we still have the dependency we might need to 12948 * update the vnode to sync the new link count to 12949 * disk. 12950 */ 12951 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 12952 error = ffs_update(vp, 1); 12953 vput(vp); 12954 if (error != 0) 12955 break; 12956 ACQUIRE_LOCK(ump); 12957 /* 12958 * If that cleared dependencies, go on to next. 12959 */ 12960 if (dap != LIST_FIRST(diraddhdp)) 12961 continue; 12962 if (dap->da_state & MKDIR_BODY) { 12963 inodedep_lookup(UFSTOVFS(ump), inum, 0, 12964 &inodedep); 12965 panic("flush_pagedep_deps: MKDIR_BODY " 12966 "inodedep %p dap %p vp %p", 12967 inodedep, dap, vp); 12968 } 12969 } 12970 /* 12971 * Flush the inode on which the directory entry depends. 12972 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 12973 * the only remaining dependency is that the updated inode 12974 * count must get pushed to disk. The inode has already 12975 * been pushed into its inode buffer (via VOP_UPDATE) at 12976 * the time of the reference count change. So we need only 12977 * locate that buffer, ensure that there will be no rollback 12978 * caused by a bitmap dependency, then write the inode buffer. 12979 */ 12980 retry: 12981 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 12982 panic("flush_pagedep_deps: lost inode"); 12983 /* 12984 * If the inode still has bitmap dependencies, 12985 * push them to disk. 12986 */ 12987 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 12988 bp = inodedep->id_bmsafemap->sm_buf; 12989 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12990 if (bp == NULL) 12991 goto retry; 12992 FREE_LOCK(ump); 12993 if ((error = bwrite(bp)) != 0) 12994 break; 12995 ACQUIRE_LOCK(ump); 12996 if (dap != LIST_FIRST(diraddhdp)) 12997 continue; 12998 } 12999 /* 13000 * If the inode is still sitting in a buffer waiting 13001 * to be written or waiting for the link count to be 13002 * adjusted update it here to flush it to disk. 13003 */ 13004 if (dap == LIST_FIRST(diraddhdp)) { 13005 FREE_LOCK(ump); 13006 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13007 FFSV_FORCEINSMQ))) 13008 break; 13009 error = ffs_update(vp, 1); 13010 vput(vp); 13011 if (error) 13012 break; 13013 ACQUIRE_LOCK(ump); 13014 } 13015 /* 13016 * If we have failed to get rid of all the dependencies 13017 * then something is seriously wrong. 13018 */ 13019 if (dap == LIST_FIRST(diraddhdp)) { 13020 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13021 panic("flush_pagedep_deps: failed to flush " 13022 "inodedep %p ino %ju dap %p", 13023 inodedep, (uintmax_t)inum, dap); 13024 } 13025 } 13026 if (error) 13027 ACQUIRE_LOCK(ump); 13028 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13029 LIST_REMOVE(dap, da_pdlist); 13030 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13031 } 13032 return (error); 13033 } 13034 13035 /* 13036 * A large burst of file addition or deletion activity can drive the 13037 * memory load excessively high. First attempt to slow things down 13038 * using the techniques below. If that fails, this routine requests 13039 * the offending operations to fall back to running synchronously 13040 * until the memory load returns to a reasonable level. 13041 */ 13042 int 13043 softdep_slowdown(vp) 13044 struct vnode *vp; 13045 { 13046 struct ufsmount *ump; 13047 int jlow; 13048 int max_softdeps_hard; 13049 13050 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13051 ("softdep_slowdown called on non-softdep filesystem")); 13052 ump = VFSTOUFS(vp->v_mount); 13053 ACQUIRE_LOCK(ump); 13054 jlow = 0; 13055 /* 13056 * Check for journal space if needed. 13057 */ 13058 if (DOINGSUJ(vp)) { 13059 if (journal_space(ump, 0) == 0) 13060 jlow = 1; 13061 } 13062 /* 13063 * If the system is under its limits and our filesystem is 13064 * not responsible for more than our share of the usage and 13065 * we are not low on journal space, then no need to slow down. 13066 */ 13067 max_softdeps_hard = max_softdeps * 11 / 10; 13068 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13069 dep_current[D_INODEDEP] < max_softdeps_hard && 13070 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13071 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13072 ump->softdep_curdeps[D_DIRREM] < 13073 (max_softdeps_hard / 2) / stat_flush_threads && 13074 ump->softdep_curdeps[D_INODEDEP] < 13075 max_softdeps_hard / stat_flush_threads && 13076 ump->softdep_curdeps[D_INDIRDEP] < 13077 (max_softdeps_hard / 1000) / stat_flush_threads && 13078 ump->softdep_curdeps[D_FREEBLKS] < 13079 max_softdeps_hard / stat_flush_threads) { 13080 FREE_LOCK(ump); 13081 return (0); 13082 } 13083 /* 13084 * If the journal is low or our filesystem is over its limit 13085 * then speedup the cleanup. 13086 */ 13087 if (ump->softdep_curdeps[D_INDIRDEP] < 13088 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13089 softdep_speedup(ump); 13090 stat_sync_limit_hit += 1; 13091 FREE_LOCK(ump); 13092 /* 13093 * We only slow down the rate at which new dependencies are 13094 * generated if we are not using journaling. With journaling, 13095 * the cleanup should always be sufficient to keep things 13096 * under control. 13097 */ 13098 if (DOINGSUJ(vp)) 13099 return (0); 13100 return (1); 13101 } 13102 13103 /* 13104 * Called by the allocation routines when they are about to fail 13105 * in the hope that we can free up the requested resource (inodes 13106 * or disk space). 13107 * 13108 * First check to see if the work list has anything on it. If it has, 13109 * clean up entries until we successfully free the requested resource. 13110 * Because this process holds inodes locked, we cannot handle any remove 13111 * requests that might block on a locked inode as that could lead to 13112 * deadlock. If the worklist yields none of the requested resource, 13113 * start syncing out vnodes to free up the needed space. 13114 */ 13115 int 13116 softdep_request_cleanup(fs, vp, cred, resource) 13117 struct fs *fs; 13118 struct vnode *vp; 13119 struct ucred *cred; 13120 int resource; 13121 { 13122 struct ufsmount *ump; 13123 struct mount *mp; 13124 struct vnode *lvp, *mvp; 13125 long starttime; 13126 ufs2_daddr_t needed; 13127 int error; 13128 13129 /* 13130 * If we are being called because of a process doing a 13131 * copy-on-write, then it is not safe to process any 13132 * worklist items as we will recurse into the copyonwrite 13133 * routine. This will result in an incoherent snapshot. 13134 * If the vnode that we hold is a snapshot, we must avoid 13135 * handling other resources that could cause deadlock. 13136 */ 13137 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13138 return (0); 13139 13140 if (resource == FLUSH_BLOCKS_WAIT) 13141 stat_cleanup_blkrequests += 1; 13142 else 13143 stat_cleanup_inorequests += 1; 13144 13145 mp = vp->v_mount; 13146 ump = VFSTOUFS(mp); 13147 mtx_assert(UFS_MTX(ump), MA_OWNED); 13148 UFS_UNLOCK(ump); 13149 error = ffs_update(vp, 1); 13150 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13151 UFS_LOCK(ump); 13152 return (0); 13153 } 13154 /* 13155 * If we are in need of resources, start by cleaning up 13156 * any block removals associated with our inode. 13157 */ 13158 ACQUIRE_LOCK(ump); 13159 process_removes(vp); 13160 process_truncates(vp); 13161 FREE_LOCK(ump); 13162 /* 13163 * Now clean up at least as many resources as we will need. 13164 * 13165 * When requested to clean up inodes, the number that are needed 13166 * is set by the number of simultaneous writers (mnt_writeopcount) 13167 * plus a bit of slop (2) in case some more writers show up while 13168 * we are cleaning. 13169 * 13170 * When requested to free up space, the amount of space that 13171 * we need is enough blocks to allocate a full-sized segment 13172 * (fs_contigsumsize). The number of such segments that will 13173 * be needed is set by the number of simultaneous writers 13174 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13175 * writers show up while we are cleaning. 13176 * 13177 * Additionally, if we are unpriviledged and allocating space, 13178 * we need to ensure that we clean up enough blocks to get the 13179 * needed number of blocks over the threshhold of the minimum 13180 * number of blocks required to be kept free by the filesystem 13181 * (fs_minfree). 13182 */ 13183 if (resource == FLUSH_INODES_WAIT) { 13184 needed = vp->v_mount->mnt_writeopcount + 2; 13185 } else if (resource == FLUSH_BLOCKS_WAIT) { 13186 needed = (vp->v_mount->mnt_writeopcount + 2) * 13187 fs->fs_contigsumsize; 13188 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 13189 needed += fragstoblks(fs, 13190 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13191 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13192 } else { 13193 UFS_LOCK(ump); 13194 printf("softdep_request_cleanup: Unknown resource type %d\n", 13195 resource); 13196 return (0); 13197 } 13198 starttime = time_second; 13199 retry: 13200 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13201 fs->fs_cstotal.cs_nbfree <= needed) || 13202 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13203 fs->fs_cstotal.cs_nifree <= needed)) { 13204 ACQUIRE_LOCK(ump); 13205 if (ump->softdep_on_worklist > 0 && 13206 process_worklist_item(UFSTOVFS(ump), 13207 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13208 stat_worklist_push += 1; 13209 FREE_LOCK(ump); 13210 } 13211 /* 13212 * If we still need resources and there are no more worklist 13213 * entries to process to obtain them, we have to start flushing 13214 * the dirty vnodes to force the release of additional requests 13215 * to the worklist that we can then process to reap addition 13216 * resources. We walk the vnodes associated with the mount point 13217 * until we get the needed worklist requests that we can reap. 13218 */ 13219 if ((resource == FLUSH_BLOCKS_WAIT && 13220 fs->fs_cstotal.cs_nbfree <= needed) || 13221 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13222 fs->fs_cstotal.cs_nifree <= needed)) { 13223 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13224 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13225 VI_UNLOCK(lvp); 13226 continue; 13227 } 13228 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13229 curthread)) 13230 continue; 13231 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13232 vput(lvp); 13233 continue; 13234 } 13235 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13236 vput(lvp); 13237 } 13238 lvp = ump->um_devvp; 13239 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13240 VOP_FSYNC(lvp, MNT_NOWAIT, curthread); 13241 VOP_UNLOCK(lvp, 0); 13242 } 13243 if (ump->softdep_on_worklist > 0) { 13244 stat_cleanup_retries += 1; 13245 goto retry; 13246 } 13247 stat_cleanup_failures += 1; 13248 } 13249 if (time_second - starttime > stat_cleanup_high_delay) 13250 stat_cleanup_high_delay = time_second - starttime; 13251 UFS_LOCK(ump); 13252 return (1); 13253 } 13254 13255 static bool 13256 softdep_excess_items(struct ufsmount *ump, int item) 13257 { 13258 13259 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13260 return (dep_current[item] > max_softdeps && 13261 ump->softdep_curdeps[item] > max_softdeps / 13262 stat_flush_threads); 13263 } 13264 13265 static void 13266 schedule_cleanup(struct mount *mp) 13267 { 13268 struct ufsmount *ump; 13269 struct thread *td; 13270 13271 ump = VFSTOUFS(mp); 13272 LOCK_OWNED(ump); 13273 FREE_LOCK(ump); 13274 td = curthread; 13275 if ((td->td_pflags & TDP_KTHREAD) != 0 && 13276 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 13277 /* 13278 * No ast is delivered to kernel threads, so nobody 13279 * would deref the mp. Some kernel threads 13280 * explicitely check for AST, e.g. NFS daemon does 13281 * this in the serving loop. 13282 */ 13283 return; 13284 } 13285 if (td->td_su != NULL) 13286 vfs_rel(td->td_su); 13287 vfs_ref(mp); 13288 td->td_su = mp; 13289 thread_lock(td); 13290 td->td_flags |= TDF_ASTPENDING; 13291 thread_unlock(td); 13292 } 13293 13294 static void 13295 softdep_ast_cleanup_proc(void) 13296 { 13297 struct thread *td; 13298 struct mount *mp; 13299 struct ufsmount *ump; 13300 int error; 13301 bool req; 13302 13303 td = curthread; 13304 mp = td->td_su; 13305 if (mp == NULL) 13306 return; 13307 td->td_su = NULL; 13308 error = vfs_busy(mp, MBF_NOWAIT); 13309 vfs_rel(mp); 13310 if (error != 0) 13311 return; 13312 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 13313 ump = VFSTOUFS(mp); 13314 for (;;) { 13315 req = false; 13316 ACQUIRE_LOCK(ump); 13317 if (softdep_excess_items(ump, D_INODEDEP)) { 13318 req = true; 13319 request_cleanup(mp, FLUSH_INODES); 13320 } 13321 if (softdep_excess_items(ump, D_DIRREM)) { 13322 req = true; 13323 request_cleanup(mp, FLUSH_BLOCKS); 13324 } 13325 FREE_LOCK(ump); 13326 if (softdep_excess_items(ump, D_NEWBLK) || 13327 softdep_excess_items(ump, D_ALLOCDIRECT) || 13328 softdep_excess_items(ump, D_ALLOCINDIR)) { 13329 error = vn_start_write(NULL, &mp, V_WAIT); 13330 if (error == 0) { 13331 req = true; 13332 VFS_SYNC(mp, MNT_WAIT); 13333 vn_finished_write(mp); 13334 } 13335 } 13336 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 13337 break; 13338 } 13339 } 13340 vfs_unbusy(mp); 13341 } 13342 13343 /* 13344 * If memory utilization has gotten too high, deliberately slow things 13345 * down and speed up the I/O processing. 13346 */ 13347 static int 13348 request_cleanup(mp, resource) 13349 struct mount *mp; 13350 int resource; 13351 { 13352 struct thread *td = curthread; 13353 struct ufsmount *ump; 13354 13355 ump = VFSTOUFS(mp); 13356 LOCK_OWNED(ump); 13357 /* 13358 * We never hold up the filesystem syncer or buf daemon. 13359 */ 13360 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13361 return (0); 13362 /* 13363 * First check to see if the work list has gotten backlogged. 13364 * If it has, co-opt this process to help clean up two entries. 13365 * Because this process may hold inodes locked, we cannot 13366 * handle any remove requests that might block on a locked 13367 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13368 * to avoid recursively processing the worklist. 13369 */ 13370 if (ump->softdep_on_worklist > max_softdeps / 10) { 13371 td->td_pflags |= TDP_SOFTDEP; 13372 process_worklist_item(mp, 2, LK_NOWAIT); 13373 td->td_pflags &= ~TDP_SOFTDEP; 13374 stat_worklist_push += 2; 13375 return(1); 13376 } 13377 /* 13378 * Next, we attempt to speed up the syncer process. If that 13379 * is successful, then we allow the process to continue. 13380 */ 13381 if (softdep_speedup(ump) && 13382 resource != FLUSH_BLOCKS_WAIT && 13383 resource != FLUSH_INODES_WAIT) 13384 return(0); 13385 /* 13386 * If we are resource constrained on inode dependencies, try 13387 * flushing some dirty inodes. Otherwise, we are constrained 13388 * by file deletions, so try accelerating flushes of directories 13389 * with removal dependencies. We would like to do the cleanup 13390 * here, but we probably hold an inode locked at this point and 13391 * that might deadlock against one that we try to clean. So, 13392 * the best that we can do is request the syncer daemon to do 13393 * the cleanup for us. 13394 */ 13395 switch (resource) { 13396 13397 case FLUSH_INODES: 13398 case FLUSH_INODES_WAIT: 13399 ACQUIRE_GBLLOCK(&lk); 13400 stat_ino_limit_push += 1; 13401 req_clear_inodedeps += 1; 13402 FREE_GBLLOCK(&lk); 13403 stat_countp = &stat_ino_limit_hit; 13404 break; 13405 13406 case FLUSH_BLOCKS: 13407 case FLUSH_BLOCKS_WAIT: 13408 ACQUIRE_GBLLOCK(&lk); 13409 stat_blk_limit_push += 1; 13410 req_clear_remove += 1; 13411 FREE_GBLLOCK(&lk); 13412 stat_countp = &stat_blk_limit_hit; 13413 break; 13414 13415 default: 13416 panic("request_cleanup: unknown type"); 13417 } 13418 /* 13419 * Hopefully the syncer daemon will catch up and awaken us. 13420 * We wait at most tickdelay before proceeding in any case. 13421 */ 13422 ACQUIRE_GBLLOCK(&lk); 13423 FREE_LOCK(ump); 13424 proc_waiting += 1; 13425 if (callout_pending(&softdep_callout) == FALSE) 13426 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13427 pause_timer, 0); 13428 13429 if ((td->td_pflags & TDP_KTHREAD) == 0) 13430 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13431 proc_waiting -= 1; 13432 FREE_GBLLOCK(&lk); 13433 ACQUIRE_LOCK(ump); 13434 return (1); 13435 } 13436 13437 /* 13438 * Awaken processes pausing in request_cleanup and clear proc_waiting 13439 * to indicate that there is no longer a timer running. Pause_timer 13440 * will be called with the global softdep mutex (&lk) locked. 13441 */ 13442 static void 13443 pause_timer(arg) 13444 void *arg; 13445 { 13446 13447 GBLLOCK_OWNED(&lk); 13448 /* 13449 * The callout_ API has acquired mtx and will hold it around this 13450 * function call. 13451 */ 13452 *stat_countp += proc_waiting; 13453 wakeup(&proc_waiting); 13454 } 13455 13456 /* 13457 * If requested, try removing inode or removal dependencies. 13458 */ 13459 static void 13460 check_clear_deps(mp) 13461 struct mount *mp; 13462 { 13463 13464 /* 13465 * If we are suspended, it may be because of our using 13466 * too many inodedeps, so help clear them out. 13467 */ 13468 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13469 clear_inodedeps(mp); 13470 /* 13471 * General requests for cleanup of backed up dependencies 13472 */ 13473 ACQUIRE_GBLLOCK(&lk); 13474 if (req_clear_inodedeps) { 13475 req_clear_inodedeps -= 1; 13476 FREE_GBLLOCK(&lk); 13477 clear_inodedeps(mp); 13478 ACQUIRE_GBLLOCK(&lk); 13479 wakeup(&proc_waiting); 13480 } 13481 if (req_clear_remove) { 13482 req_clear_remove -= 1; 13483 FREE_GBLLOCK(&lk); 13484 clear_remove(mp); 13485 ACQUIRE_GBLLOCK(&lk); 13486 wakeup(&proc_waiting); 13487 } 13488 FREE_GBLLOCK(&lk); 13489 } 13490 13491 /* 13492 * Flush out a directory with at least one removal dependency in an effort to 13493 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13494 */ 13495 static void 13496 clear_remove(mp) 13497 struct mount *mp; 13498 { 13499 struct pagedep_hashhead *pagedephd; 13500 struct pagedep *pagedep; 13501 struct ufsmount *ump; 13502 struct vnode *vp; 13503 struct bufobj *bo; 13504 int error, cnt; 13505 ino_t ino; 13506 13507 ump = VFSTOUFS(mp); 13508 LOCK_OWNED(ump); 13509 13510 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13511 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13512 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13513 ump->pagedep_nextclean = 0; 13514 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13515 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13516 continue; 13517 ino = pagedep->pd_ino; 13518 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13519 continue; 13520 FREE_LOCK(ump); 13521 13522 /* 13523 * Let unmount clear deps 13524 */ 13525 error = vfs_busy(mp, MBF_NOWAIT); 13526 if (error != 0) 13527 goto finish_write; 13528 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13529 FFSV_FORCEINSMQ); 13530 vfs_unbusy(mp); 13531 if (error != 0) { 13532 softdep_error("clear_remove: vget", error); 13533 goto finish_write; 13534 } 13535 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13536 softdep_error("clear_remove: fsync", error); 13537 bo = &vp->v_bufobj; 13538 BO_LOCK(bo); 13539 drain_output(vp); 13540 BO_UNLOCK(bo); 13541 vput(vp); 13542 finish_write: 13543 vn_finished_write(mp); 13544 ACQUIRE_LOCK(ump); 13545 return; 13546 } 13547 } 13548 } 13549 13550 /* 13551 * Clear out a block of dirty inodes in an effort to reduce 13552 * the number of inodedep dependency structures. 13553 */ 13554 static void 13555 clear_inodedeps(mp) 13556 struct mount *mp; 13557 { 13558 struct inodedep_hashhead *inodedephd; 13559 struct inodedep *inodedep; 13560 struct ufsmount *ump; 13561 struct vnode *vp; 13562 struct fs *fs; 13563 int error, cnt; 13564 ino_t firstino, lastino, ino; 13565 13566 ump = VFSTOUFS(mp); 13567 fs = ump->um_fs; 13568 LOCK_OWNED(ump); 13569 /* 13570 * Pick a random inode dependency to be cleared. 13571 * We will then gather up all the inodes in its block 13572 * that have dependencies and flush them out. 13573 */ 13574 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13575 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13576 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13577 ump->inodedep_nextclean = 0; 13578 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13579 break; 13580 } 13581 if (inodedep == NULL) 13582 return; 13583 /* 13584 * Find the last inode in the block with dependencies. 13585 */ 13586 firstino = inodedep->id_ino & ~(INOPB(fs) - 1); 13587 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13588 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13589 break; 13590 /* 13591 * Asynchronously push all but the last inode with dependencies. 13592 * Synchronously push the last inode with dependencies to ensure 13593 * that the inode block gets written to free up the inodedeps. 13594 */ 13595 for (ino = firstino; ino <= lastino; ino++) { 13596 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13597 continue; 13598 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13599 continue; 13600 FREE_LOCK(ump); 13601 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13602 if (error != 0) { 13603 vn_finished_write(mp); 13604 ACQUIRE_LOCK(ump); 13605 return; 13606 } 13607 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13608 FFSV_FORCEINSMQ)) != 0) { 13609 softdep_error("clear_inodedeps: vget", error); 13610 vfs_unbusy(mp); 13611 vn_finished_write(mp); 13612 ACQUIRE_LOCK(ump); 13613 return; 13614 } 13615 vfs_unbusy(mp); 13616 if (ino == lastino) { 13617 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13618 softdep_error("clear_inodedeps: fsync1", error); 13619 } else { 13620 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13621 softdep_error("clear_inodedeps: fsync2", error); 13622 BO_LOCK(&vp->v_bufobj); 13623 drain_output(vp); 13624 BO_UNLOCK(&vp->v_bufobj); 13625 } 13626 vput(vp); 13627 vn_finished_write(mp); 13628 ACQUIRE_LOCK(ump); 13629 } 13630 } 13631 13632 void 13633 softdep_buf_append(bp, wkhd) 13634 struct buf *bp; 13635 struct workhead *wkhd; 13636 { 13637 struct worklist *wk; 13638 struct ufsmount *ump; 13639 13640 if ((wk = LIST_FIRST(wkhd)) == NULL) 13641 return; 13642 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13643 ("softdep_buf_append called on non-softdep filesystem")); 13644 ump = VFSTOUFS(wk->wk_mp); 13645 ACQUIRE_LOCK(ump); 13646 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13647 WORKLIST_REMOVE(wk); 13648 WORKLIST_INSERT(&bp->b_dep, wk); 13649 } 13650 FREE_LOCK(ump); 13651 13652 } 13653 13654 void 13655 softdep_inode_append(ip, cred, wkhd) 13656 struct inode *ip; 13657 struct ucred *cred; 13658 struct workhead *wkhd; 13659 { 13660 struct buf *bp; 13661 struct fs *fs; 13662 int error; 13663 13664 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ip->i_ump)) != 0, 13665 ("softdep_inode_append called on non-softdep filesystem")); 13666 fs = ip->i_fs; 13667 error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13668 (int)fs->fs_bsize, cred, &bp); 13669 if (error) { 13670 bqrelse(bp); 13671 softdep_freework(wkhd); 13672 return; 13673 } 13674 softdep_buf_append(bp, wkhd); 13675 bqrelse(bp); 13676 } 13677 13678 void 13679 softdep_freework(wkhd) 13680 struct workhead *wkhd; 13681 { 13682 struct worklist *wk; 13683 struct ufsmount *ump; 13684 13685 if ((wk = LIST_FIRST(wkhd)) == NULL) 13686 return; 13687 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13688 ("softdep_freework called on non-softdep filesystem")); 13689 ump = VFSTOUFS(wk->wk_mp); 13690 ACQUIRE_LOCK(ump); 13691 handle_jwork(wkhd); 13692 FREE_LOCK(ump); 13693 } 13694 13695 /* 13696 * Function to determine if the buffer has outstanding dependencies 13697 * that will cause a roll-back if the buffer is written. If wantcount 13698 * is set, return number of dependencies, otherwise just yes or no. 13699 */ 13700 static int 13701 softdep_count_dependencies(bp, wantcount) 13702 struct buf *bp; 13703 int wantcount; 13704 { 13705 struct worklist *wk; 13706 struct ufsmount *ump; 13707 struct bmsafemap *bmsafemap; 13708 struct freework *freework; 13709 struct inodedep *inodedep; 13710 struct indirdep *indirdep; 13711 struct freeblks *freeblks; 13712 struct allocindir *aip; 13713 struct pagedep *pagedep; 13714 struct dirrem *dirrem; 13715 struct newblk *newblk; 13716 struct mkdir *mkdir; 13717 struct diradd *dap; 13718 int i, retval; 13719 13720 retval = 0; 13721 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL) 13722 return (0); 13723 ump = VFSTOUFS(wk->wk_mp); 13724 ACQUIRE_LOCK(ump); 13725 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13726 switch (wk->wk_type) { 13727 13728 case D_INODEDEP: 13729 inodedep = WK_INODEDEP(wk); 13730 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13731 /* bitmap allocation dependency */ 13732 retval += 1; 13733 if (!wantcount) 13734 goto out; 13735 } 13736 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13737 /* direct block pointer dependency */ 13738 retval += 1; 13739 if (!wantcount) 13740 goto out; 13741 } 13742 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13743 /* direct block pointer dependency */ 13744 retval += 1; 13745 if (!wantcount) 13746 goto out; 13747 } 13748 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 13749 /* Add reference dependency. */ 13750 retval += 1; 13751 if (!wantcount) 13752 goto out; 13753 } 13754 continue; 13755 13756 case D_INDIRDEP: 13757 indirdep = WK_INDIRDEP(wk); 13758 13759 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 13760 /* indirect truncation dependency */ 13761 retval += 1; 13762 if (!wantcount) 13763 goto out; 13764 } 13765 13766 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 13767 /* indirect block pointer dependency */ 13768 retval += 1; 13769 if (!wantcount) 13770 goto out; 13771 } 13772 continue; 13773 13774 case D_PAGEDEP: 13775 pagedep = WK_PAGEDEP(wk); 13776 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 13777 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 13778 /* Journal remove ref dependency. */ 13779 retval += 1; 13780 if (!wantcount) 13781 goto out; 13782 } 13783 } 13784 for (i = 0; i < DAHASHSZ; i++) { 13785 13786 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 13787 /* directory entry dependency */ 13788 retval += 1; 13789 if (!wantcount) 13790 goto out; 13791 } 13792 } 13793 continue; 13794 13795 case D_BMSAFEMAP: 13796 bmsafemap = WK_BMSAFEMAP(wk); 13797 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 13798 /* Add reference dependency. */ 13799 retval += 1; 13800 if (!wantcount) 13801 goto out; 13802 } 13803 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 13804 /* Allocate block dependency. */ 13805 retval += 1; 13806 if (!wantcount) 13807 goto out; 13808 } 13809 continue; 13810 13811 case D_FREEBLKS: 13812 freeblks = WK_FREEBLKS(wk); 13813 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 13814 /* Freeblk journal dependency. */ 13815 retval += 1; 13816 if (!wantcount) 13817 goto out; 13818 } 13819 continue; 13820 13821 case D_ALLOCDIRECT: 13822 case D_ALLOCINDIR: 13823 newblk = WK_NEWBLK(wk); 13824 if (newblk->nb_jnewblk) { 13825 /* Journal allocate dependency. */ 13826 retval += 1; 13827 if (!wantcount) 13828 goto out; 13829 } 13830 continue; 13831 13832 case D_MKDIR: 13833 mkdir = WK_MKDIR(wk); 13834 if (mkdir->md_jaddref) { 13835 /* Journal reference dependency. */ 13836 retval += 1; 13837 if (!wantcount) 13838 goto out; 13839 } 13840 continue; 13841 13842 case D_FREEWORK: 13843 case D_FREEDEP: 13844 case D_JSEGDEP: 13845 case D_JSEG: 13846 case D_SBDEP: 13847 /* never a dependency on these blocks */ 13848 continue; 13849 13850 default: 13851 panic("softdep_count_dependencies: Unexpected type %s", 13852 TYPENAME(wk->wk_type)); 13853 /* NOTREACHED */ 13854 } 13855 } 13856 out: 13857 FREE_LOCK(ump); 13858 return retval; 13859 } 13860 13861 /* 13862 * Acquire exclusive access to a buffer. 13863 * Must be called with a locked mtx parameter. 13864 * Return acquired buffer or NULL on failure. 13865 */ 13866 static struct buf * 13867 getdirtybuf(bp, lock, waitfor) 13868 struct buf *bp; 13869 struct rwlock *lock; 13870 int waitfor; 13871 { 13872 int error; 13873 13874 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 13875 if (waitfor != MNT_WAIT) 13876 return (NULL); 13877 error = BUF_LOCK(bp, 13878 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 13879 /* 13880 * Even if we sucessfully acquire bp here, we have dropped 13881 * lock, which may violates our guarantee. 13882 */ 13883 if (error == 0) 13884 BUF_UNLOCK(bp); 13885 else if (error != ENOLCK) 13886 panic("getdirtybuf: inconsistent lock: %d", error); 13887 rw_wlock(lock); 13888 return (NULL); 13889 } 13890 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13891 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 13892 rw_wunlock(lock); 13893 BO_LOCK(bp->b_bufobj); 13894 BUF_UNLOCK(bp); 13895 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 13896 bp->b_vflags |= BV_BKGRDWAIT; 13897 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 13898 PRIBIO | PDROP, "getbuf", 0); 13899 } else 13900 BO_UNLOCK(bp->b_bufobj); 13901 rw_wlock(lock); 13902 return (NULL); 13903 } 13904 BUF_UNLOCK(bp); 13905 if (waitfor != MNT_WAIT) 13906 return (NULL); 13907 /* 13908 * The lock argument must be bp->b_vp's mutex in 13909 * this case. 13910 */ 13911 #ifdef DEBUG_VFS_LOCKS 13912 if (bp->b_vp->v_type != VCHR) 13913 ASSERT_BO_WLOCKED(bp->b_bufobj); 13914 #endif 13915 bp->b_vflags |= BV_BKGRDWAIT; 13916 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 13917 return (NULL); 13918 } 13919 if ((bp->b_flags & B_DELWRI) == 0) { 13920 BUF_UNLOCK(bp); 13921 return (NULL); 13922 } 13923 bremfree(bp); 13924 return (bp); 13925 } 13926 13927 13928 /* 13929 * Check if it is safe to suspend the file system now. On entry, 13930 * the vnode interlock for devvp should be held. Return 0 with 13931 * the mount interlock held if the file system can be suspended now, 13932 * otherwise return EAGAIN with the mount interlock held. 13933 */ 13934 int 13935 softdep_check_suspend(struct mount *mp, 13936 struct vnode *devvp, 13937 int softdep_depcnt, 13938 int softdep_accdepcnt, 13939 int secondary_writes, 13940 int secondary_accwrites) 13941 { 13942 struct bufobj *bo; 13943 struct ufsmount *ump; 13944 struct inodedep *inodedep; 13945 int error, unlinked; 13946 13947 bo = &devvp->v_bufobj; 13948 ASSERT_BO_WLOCKED(bo); 13949 13950 /* 13951 * If we are not running with soft updates, then we need only 13952 * deal with secondary writes as we try to suspend. 13953 */ 13954 if (MOUNTEDSOFTDEP(mp) == 0) { 13955 MNT_ILOCK(mp); 13956 while (mp->mnt_secondary_writes != 0) { 13957 BO_UNLOCK(bo); 13958 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 13959 (PUSER - 1) | PDROP, "secwr", 0); 13960 BO_LOCK(bo); 13961 MNT_ILOCK(mp); 13962 } 13963 13964 /* 13965 * Reasons for needing more work before suspend: 13966 * - Dirty buffers on devvp. 13967 * - Secondary writes occurred after start of vnode sync loop 13968 */ 13969 error = 0; 13970 if (bo->bo_numoutput > 0 || 13971 bo->bo_dirty.bv_cnt > 0 || 13972 secondary_writes != 0 || 13973 mp->mnt_secondary_writes != 0 || 13974 secondary_accwrites != mp->mnt_secondary_accwrites) 13975 error = EAGAIN; 13976 BO_UNLOCK(bo); 13977 return (error); 13978 } 13979 13980 /* 13981 * If we are running with soft updates, then we need to coordinate 13982 * with them as we try to suspend. 13983 */ 13984 ump = VFSTOUFS(mp); 13985 for (;;) { 13986 if (!TRY_ACQUIRE_LOCK(ump)) { 13987 BO_UNLOCK(bo); 13988 ACQUIRE_LOCK(ump); 13989 FREE_LOCK(ump); 13990 BO_LOCK(bo); 13991 continue; 13992 } 13993 MNT_ILOCK(mp); 13994 if (mp->mnt_secondary_writes != 0) { 13995 FREE_LOCK(ump); 13996 BO_UNLOCK(bo); 13997 msleep(&mp->mnt_secondary_writes, 13998 MNT_MTX(mp), 13999 (PUSER - 1) | PDROP, "secwr", 0); 14000 BO_LOCK(bo); 14001 continue; 14002 } 14003 break; 14004 } 14005 14006 unlinked = 0; 14007 if (MOUNTEDSUJ(mp)) { 14008 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14009 inodedep != NULL; 14010 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14011 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14012 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14013 UNLINKONLIST) || 14014 !check_inodedep_free(inodedep)) 14015 continue; 14016 unlinked++; 14017 } 14018 } 14019 14020 /* 14021 * Reasons for needing more work before suspend: 14022 * - Dirty buffers on devvp. 14023 * - Softdep activity occurred after start of vnode sync loop 14024 * - Secondary writes occurred after start of vnode sync loop 14025 */ 14026 error = 0; 14027 if (bo->bo_numoutput > 0 || 14028 bo->bo_dirty.bv_cnt > 0 || 14029 softdep_depcnt != unlinked || 14030 ump->softdep_deps != unlinked || 14031 softdep_accdepcnt != ump->softdep_accdeps || 14032 secondary_writes != 0 || 14033 mp->mnt_secondary_writes != 0 || 14034 secondary_accwrites != mp->mnt_secondary_accwrites) 14035 error = EAGAIN; 14036 FREE_LOCK(ump); 14037 BO_UNLOCK(bo); 14038 return (error); 14039 } 14040 14041 14042 /* 14043 * Get the number of dependency structures for the file system, both 14044 * the current number and the total number allocated. These will 14045 * later be used to detect that softdep processing has occurred. 14046 */ 14047 void 14048 softdep_get_depcounts(struct mount *mp, 14049 int *softdep_depsp, 14050 int *softdep_accdepsp) 14051 { 14052 struct ufsmount *ump; 14053 14054 if (MOUNTEDSOFTDEP(mp) == 0) { 14055 *softdep_depsp = 0; 14056 *softdep_accdepsp = 0; 14057 return; 14058 } 14059 ump = VFSTOUFS(mp); 14060 ACQUIRE_LOCK(ump); 14061 *softdep_depsp = ump->softdep_deps; 14062 *softdep_accdepsp = ump->softdep_accdeps; 14063 FREE_LOCK(ump); 14064 } 14065 14066 /* 14067 * Wait for pending output on a vnode to complete. 14068 * Must be called with vnode lock and interlock locked. 14069 * 14070 * XXX: Should just be a call to bufobj_wwait(). 14071 */ 14072 static void 14073 drain_output(vp) 14074 struct vnode *vp; 14075 { 14076 struct bufobj *bo; 14077 14078 bo = &vp->v_bufobj; 14079 ASSERT_VOP_LOCKED(vp, "drain_output"); 14080 ASSERT_BO_WLOCKED(bo); 14081 14082 while (bo->bo_numoutput) { 14083 bo->bo_flag |= BO_WWAIT; 14084 msleep((caddr_t)&bo->bo_numoutput, 14085 BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0); 14086 } 14087 } 14088 14089 /* 14090 * Called whenever a buffer that is being invalidated or reallocated 14091 * contains dependencies. This should only happen if an I/O error has 14092 * occurred. The routine is called with the buffer locked. 14093 */ 14094 static void 14095 softdep_deallocate_dependencies(bp) 14096 struct buf *bp; 14097 { 14098 14099 if ((bp->b_ioflags & BIO_ERROR) == 0) 14100 panic("softdep_deallocate_dependencies: dangling deps"); 14101 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14102 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14103 else 14104 printf("softdep_deallocate_dependencies: " 14105 "got error %d while accessing filesystem\n", bp->b_error); 14106 if (bp->b_error != ENXIO) 14107 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14108 } 14109 14110 /* 14111 * Function to handle asynchronous write errors in the filesystem. 14112 */ 14113 static void 14114 softdep_error(func, error) 14115 char *func; 14116 int error; 14117 { 14118 14119 /* XXX should do something better! */ 14120 printf("%s: got error %d while accessing filesystem\n", func, error); 14121 } 14122 14123 #ifdef DDB 14124 14125 static void 14126 inodedep_print(struct inodedep *inodedep, int verbose) 14127 { 14128 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d" 14129 " saveino %p\n", 14130 inodedep, inodedep->id_fs, inodedep->id_state, 14131 (intmax_t)inodedep->id_ino, 14132 (intmax_t)fsbtodb(inodedep->id_fs, 14133 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14134 inodedep->id_nlinkdelta, inodedep->id_savednlink, 14135 inodedep->id_savedino1); 14136 14137 if (verbose == 0) 14138 return; 14139 14140 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 14141 "mkdiradd %p\n", 14142 LIST_FIRST(&inodedep->id_pendinghd), 14143 LIST_FIRST(&inodedep->id_bufwait), 14144 LIST_FIRST(&inodedep->id_inowait), 14145 TAILQ_FIRST(&inodedep->id_inoreflst), 14146 inodedep->id_mkdiradd); 14147 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 14148 TAILQ_FIRST(&inodedep->id_inoupdt), 14149 TAILQ_FIRST(&inodedep->id_newinoupdt), 14150 TAILQ_FIRST(&inodedep->id_extupdt), 14151 TAILQ_FIRST(&inodedep->id_newextupdt)); 14152 } 14153 14154 DB_SHOW_COMMAND(inodedep, db_show_inodedep) 14155 { 14156 14157 if (have_addr == 0) { 14158 db_printf("Address required\n"); 14159 return; 14160 } 14161 inodedep_print((struct inodedep*)addr, 1); 14162 } 14163 14164 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 14165 { 14166 struct inodedep_hashhead *inodedephd; 14167 struct inodedep *inodedep; 14168 struct ufsmount *ump; 14169 int cnt; 14170 14171 if (have_addr == 0) { 14172 db_printf("Address required\n"); 14173 return; 14174 } 14175 ump = (struct ufsmount *)addr; 14176 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14177 inodedephd = &ump->inodedep_hashtbl[cnt]; 14178 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14179 inodedep_print(inodedep, 0); 14180 } 14181 } 14182 } 14183 14184 DB_SHOW_COMMAND(worklist, db_show_worklist) 14185 { 14186 struct worklist *wk; 14187 14188 if (have_addr == 0) { 14189 db_printf("Address required\n"); 14190 return; 14191 } 14192 wk = (struct worklist *)addr; 14193 printf("worklist: %p type %s state 0x%X\n", 14194 wk, TYPENAME(wk->wk_type), wk->wk_state); 14195 } 14196 14197 DB_SHOW_COMMAND(workhead, db_show_workhead) 14198 { 14199 struct workhead *wkhd; 14200 struct worklist *wk; 14201 int i; 14202 14203 if (have_addr == 0) { 14204 db_printf("Address required\n"); 14205 return; 14206 } 14207 wkhd = (struct workhead *)addr; 14208 wk = LIST_FIRST(wkhd); 14209 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 14210 db_printf("worklist: %p type %s state 0x%X", 14211 wk, TYPENAME(wk->wk_type), wk->wk_state); 14212 if (i == 100) 14213 db_printf("workhead overflow"); 14214 printf("\n"); 14215 } 14216 14217 14218 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 14219 { 14220 struct mkdirlist *mkdirlisthd; 14221 struct jaddref *jaddref; 14222 struct diradd *diradd; 14223 struct mkdir *mkdir; 14224 14225 if (have_addr == 0) { 14226 db_printf("Address required\n"); 14227 return; 14228 } 14229 mkdirlisthd = (struct mkdirlist *)addr; 14230 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14231 diradd = mkdir->md_diradd; 14232 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 14233 mkdir, mkdir->md_state, diradd, diradd->da_state); 14234 if ((jaddref = mkdir->md_jaddref) != NULL) 14235 db_printf(" jaddref %p jaddref state 0x%X", 14236 jaddref, jaddref->ja_state); 14237 db_printf("\n"); 14238 } 14239 } 14240 14241 /* exported to ffs_vfsops.c */ 14242 extern void db_print_ffs(struct ufsmount *ump); 14243 void 14244 db_print_ffs(struct ufsmount *ump) 14245 { 14246 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n", 14247 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname, 14248 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist, 14249 ump->softdep_deps, ump->softdep_req); 14250 } 14251 14252 #endif /* DDB */ 14253 14254 #endif /* SOFTUPDATES */ 14255