1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright 1998, 2000 Marshall Kirk McKusick. 5 * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org> 6 * All rights reserved. 7 * 8 * The soft updates code is derived from the appendix of a University 9 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 10 * "Soft Updates: A Solution to the Metadata Update Problem in File 11 * Systems", CSE-TR-254-95, August 1995). 12 * 13 * Further information about soft updates can be obtained from: 14 * 15 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 16 * 1614 Oxford Street mckusick@mckusick.com 17 * Berkeley, CA 94709-1608 +1-510-843-9542 18 * USA 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * 1. Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * 2. Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in the 28 * documentation and/or other materials provided with the distribution. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 31 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 32 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 33 * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, 34 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 35 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 36 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 37 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR 38 * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE 39 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 * 41 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00 42 */ 43 44 #include <sys/cdefs.h> 45 __FBSDID("$FreeBSD$"); 46 47 #include "opt_ffs.h" 48 #include "opt_quota.h" 49 #include "opt_ddb.h" 50 51 /* 52 * For now we want the safety net that the DEBUG flag provides. 53 */ 54 #ifndef DEBUG 55 #define DEBUG 56 #endif 57 58 #include <sys/param.h> 59 #include <sys/kernel.h> 60 #include <sys/systm.h> 61 #include <sys/bio.h> 62 #include <sys/buf.h> 63 #include <sys/kdb.h> 64 #include <sys/kthread.h> 65 #include <sys/ktr.h> 66 #include <sys/limits.h> 67 #include <sys/lock.h> 68 #include <sys/malloc.h> 69 #include <sys/mount.h> 70 #include <sys/mutex.h> 71 #include <sys/namei.h> 72 #include <sys/priv.h> 73 #include <sys/proc.h> 74 #include <sys/racct.h> 75 #include <sys/rwlock.h> 76 #include <sys/stat.h> 77 #include <sys/sysctl.h> 78 #include <sys/syslog.h> 79 #include <sys/vnode.h> 80 #include <sys/conf.h> 81 82 #include <ufs/ufs/dir.h> 83 #include <ufs/ufs/extattr.h> 84 #include <ufs/ufs/quota.h> 85 #include <ufs/ufs/inode.h> 86 #include <ufs/ufs/ufsmount.h> 87 #include <ufs/ffs/fs.h> 88 #include <ufs/ffs/softdep.h> 89 #include <ufs/ffs/ffs_extern.h> 90 #include <ufs/ufs/ufs_extern.h> 91 92 #include <vm/vm.h> 93 #include <vm/vm_extern.h> 94 #include <vm/vm_object.h> 95 96 #include <geom/geom.h> 97 98 #include <ddb/ddb.h> 99 100 #define KTR_SUJ 0 /* Define to KTR_SPARE. */ 101 102 #ifndef SOFTUPDATES 103 104 int 105 softdep_flushfiles(oldmnt, flags, td) 106 struct mount *oldmnt; 107 int flags; 108 struct thread *td; 109 { 110 111 panic("softdep_flushfiles called"); 112 } 113 114 int 115 softdep_mount(devvp, mp, fs, cred) 116 struct vnode *devvp; 117 struct mount *mp; 118 struct fs *fs; 119 struct ucred *cred; 120 { 121 122 return (0); 123 } 124 125 void 126 softdep_initialize() 127 { 128 129 return; 130 } 131 132 void 133 softdep_uninitialize() 134 { 135 136 return; 137 } 138 139 void 140 softdep_unmount(mp) 141 struct mount *mp; 142 { 143 144 panic("softdep_unmount called"); 145 } 146 147 void 148 softdep_setup_sbupdate(ump, fs, bp) 149 struct ufsmount *ump; 150 struct fs *fs; 151 struct buf *bp; 152 { 153 154 panic("softdep_setup_sbupdate called"); 155 } 156 157 void 158 softdep_setup_inomapdep(bp, ip, newinum, mode) 159 struct buf *bp; 160 struct inode *ip; 161 ino_t newinum; 162 int mode; 163 { 164 165 panic("softdep_setup_inomapdep called"); 166 } 167 168 void 169 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 170 struct buf *bp; 171 struct mount *mp; 172 ufs2_daddr_t newblkno; 173 int frags; 174 int oldfrags; 175 { 176 177 panic("softdep_setup_blkmapdep called"); 178 } 179 180 void 181 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 182 struct inode *ip; 183 ufs_lbn_t lbn; 184 ufs2_daddr_t newblkno; 185 ufs2_daddr_t oldblkno; 186 long newsize; 187 long oldsize; 188 struct buf *bp; 189 { 190 191 panic("softdep_setup_allocdirect called"); 192 } 193 194 void 195 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 196 struct inode *ip; 197 ufs_lbn_t lbn; 198 ufs2_daddr_t newblkno; 199 ufs2_daddr_t oldblkno; 200 long newsize; 201 long oldsize; 202 struct buf *bp; 203 { 204 205 panic("softdep_setup_allocext called"); 206 } 207 208 void 209 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 210 struct inode *ip; 211 ufs_lbn_t lbn; 212 struct buf *bp; 213 int ptrno; 214 ufs2_daddr_t newblkno; 215 ufs2_daddr_t oldblkno; 216 struct buf *nbp; 217 { 218 219 panic("softdep_setup_allocindir_page called"); 220 } 221 222 void 223 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 224 struct buf *nbp; 225 struct inode *ip; 226 struct buf *bp; 227 int ptrno; 228 ufs2_daddr_t newblkno; 229 { 230 231 panic("softdep_setup_allocindir_meta called"); 232 } 233 234 void 235 softdep_journal_freeblocks(ip, cred, length, flags) 236 struct inode *ip; 237 struct ucred *cred; 238 off_t length; 239 int flags; 240 { 241 242 panic("softdep_journal_freeblocks called"); 243 } 244 245 void 246 softdep_journal_fsync(ip) 247 struct inode *ip; 248 { 249 250 panic("softdep_journal_fsync called"); 251 } 252 253 void 254 softdep_setup_freeblocks(ip, length, flags) 255 struct inode *ip; 256 off_t length; 257 int flags; 258 { 259 260 panic("softdep_setup_freeblocks called"); 261 } 262 263 void 264 softdep_freefile(pvp, ino, mode) 265 struct vnode *pvp; 266 ino_t ino; 267 int mode; 268 { 269 270 panic("softdep_freefile called"); 271 } 272 273 int 274 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 275 struct buf *bp; 276 struct inode *dp; 277 off_t diroffset; 278 ino_t newinum; 279 struct buf *newdirbp; 280 int isnewblk; 281 { 282 283 panic("softdep_setup_directory_add called"); 284 } 285 286 void 287 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 288 struct buf *bp; 289 struct inode *dp; 290 caddr_t base; 291 caddr_t oldloc; 292 caddr_t newloc; 293 int entrysize; 294 { 295 296 panic("softdep_change_directoryentry_offset called"); 297 } 298 299 void 300 softdep_setup_remove(bp, dp, ip, isrmdir) 301 struct buf *bp; 302 struct inode *dp; 303 struct inode *ip; 304 int isrmdir; 305 { 306 307 panic("softdep_setup_remove called"); 308 } 309 310 void 311 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 312 struct buf *bp; 313 struct inode *dp; 314 struct inode *ip; 315 ino_t newinum; 316 int isrmdir; 317 { 318 319 panic("softdep_setup_directory_change called"); 320 } 321 322 void 323 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 324 struct mount *mp; 325 struct buf *bp; 326 ufs2_daddr_t blkno; 327 int frags; 328 struct workhead *wkhd; 329 { 330 331 panic("%s called", __FUNCTION__); 332 } 333 334 void 335 softdep_setup_inofree(mp, bp, ino, wkhd) 336 struct mount *mp; 337 struct buf *bp; 338 ino_t ino; 339 struct workhead *wkhd; 340 { 341 342 panic("%s called", __FUNCTION__); 343 } 344 345 void 346 softdep_setup_unlink(dp, ip) 347 struct inode *dp; 348 struct inode *ip; 349 { 350 351 panic("%s called", __FUNCTION__); 352 } 353 354 void 355 softdep_setup_link(dp, ip) 356 struct inode *dp; 357 struct inode *ip; 358 { 359 360 panic("%s called", __FUNCTION__); 361 } 362 363 void 364 softdep_revert_link(dp, ip) 365 struct inode *dp; 366 struct inode *ip; 367 { 368 369 panic("%s called", __FUNCTION__); 370 } 371 372 void 373 softdep_setup_rmdir(dp, ip) 374 struct inode *dp; 375 struct inode *ip; 376 { 377 378 panic("%s called", __FUNCTION__); 379 } 380 381 void 382 softdep_revert_rmdir(dp, ip) 383 struct inode *dp; 384 struct inode *ip; 385 { 386 387 panic("%s called", __FUNCTION__); 388 } 389 390 void 391 softdep_setup_create(dp, ip) 392 struct inode *dp; 393 struct inode *ip; 394 { 395 396 panic("%s called", __FUNCTION__); 397 } 398 399 void 400 softdep_revert_create(dp, ip) 401 struct inode *dp; 402 struct inode *ip; 403 { 404 405 panic("%s called", __FUNCTION__); 406 } 407 408 void 409 softdep_setup_mkdir(dp, ip) 410 struct inode *dp; 411 struct inode *ip; 412 { 413 414 panic("%s called", __FUNCTION__); 415 } 416 417 void 418 softdep_revert_mkdir(dp, ip) 419 struct inode *dp; 420 struct inode *ip; 421 { 422 423 panic("%s called", __FUNCTION__); 424 } 425 426 void 427 softdep_setup_dotdot_link(dp, ip) 428 struct inode *dp; 429 struct inode *ip; 430 { 431 432 panic("%s called", __FUNCTION__); 433 } 434 435 int 436 softdep_prealloc(vp, waitok) 437 struct vnode *vp; 438 int waitok; 439 { 440 441 panic("%s called", __FUNCTION__); 442 } 443 444 int 445 softdep_journal_lookup(mp, vpp) 446 struct mount *mp; 447 struct vnode **vpp; 448 { 449 450 return (ENOENT); 451 } 452 453 void 454 softdep_change_linkcnt(ip) 455 struct inode *ip; 456 { 457 458 panic("softdep_change_linkcnt called"); 459 } 460 461 void 462 softdep_load_inodeblock(ip) 463 struct inode *ip; 464 { 465 466 panic("softdep_load_inodeblock called"); 467 } 468 469 void 470 softdep_update_inodeblock(ip, bp, waitfor) 471 struct inode *ip; 472 struct buf *bp; 473 int waitfor; 474 { 475 476 panic("softdep_update_inodeblock called"); 477 } 478 479 int 480 softdep_fsync(vp) 481 struct vnode *vp; /* the "in_core" copy of the inode */ 482 { 483 484 return (0); 485 } 486 487 void 488 softdep_fsync_mountdev(vp) 489 struct vnode *vp; 490 { 491 492 return; 493 } 494 495 int 496 softdep_flushworklist(oldmnt, countp, td) 497 struct mount *oldmnt; 498 int *countp; 499 struct thread *td; 500 { 501 502 *countp = 0; 503 return (0); 504 } 505 506 int 507 softdep_sync_metadata(struct vnode *vp) 508 { 509 510 panic("softdep_sync_metadata called"); 511 } 512 513 int 514 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 515 { 516 517 panic("softdep_sync_buf called"); 518 } 519 520 int 521 softdep_slowdown(vp) 522 struct vnode *vp; 523 { 524 525 panic("softdep_slowdown called"); 526 } 527 528 int 529 softdep_request_cleanup(fs, vp, cred, resource) 530 struct fs *fs; 531 struct vnode *vp; 532 struct ucred *cred; 533 int resource; 534 { 535 536 return (0); 537 } 538 539 int 540 softdep_check_suspend(struct mount *mp, 541 struct vnode *devvp, 542 int softdep_depcnt, 543 int softdep_accdepcnt, 544 int secondary_writes, 545 int secondary_accwrites) 546 { 547 struct bufobj *bo; 548 int error; 549 550 (void) softdep_depcnt, 551 (void) softdep_accdepcnt; 552 553 bo = &devvp->v_bufobj; 554 ASSERT_BO_WLOCKED(bo); 555 556 MNT_ILOCK(mp); 557 while (mp->mnt_secondary_writes != 0) { 558 BO_UNLOCK(bo); 559 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 560 (PUSER - 1) | PDROP, "secwr", 0); 561 BO_LOCK(bo); 562 MNT_ILOCK(mp); 563 } 564 565 /* 566 * Reasons for needing more work before suspend: 567 * - Dirty buffers on devvp. 568 * - Secondary writes occurred after start of vnode sync loop 569 */ 570 error = 0; 571 if (bo->bo_numoutput > 0 || 572 bo->bo_dirty.bv_cnt > 0 || 573 secondary_writes != 0 || 574 mp->mnt_secondary_writes != 0 || 575 secondary_accwrites != mp->mnt_secondary_accwrites) 576 error = EAGAIN; 577 BO_UNLOCK(bo); 578 return (error); 579 } 580 581 void 582 softdep_get_depcounts(struct mount *mp, 583 int *softdepactivep, 584 int *softdepactiveaccp) 585 { 586 (void) mp; 587 *softdepactivep = 0; 588 *softdepactiveaccp = 0; 589 } 590 591 void 592 softdep_buf_append(bp, wkhd) 593 struct buf *bp; 594 struct workhead *wkhd; 595 { 596 597 panic("softdep_buf_appendwork called"); 598 } 599 600 void 601 softdep_inode_append(ip, cred, wkhd) 602 struct inode *ip; 603 struct ucred *cred; 604 struct workhead *wkhd; 605 { 606 607 panic("softdep_inode_appendwork called"); 608 } 609 610 void 611 softdep_freework(wkhd) 612 struct workhead *wkhd; 613 { 614 615 panic("softdep_freework called"); 616 } 617 618 #else 619 620 FEATURE(softupdates, "FFS soft-updates support"); 621 622 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 623 "soft updates stats"); 624 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 625 "total dependencies allocated"); 626 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 627 "high use dependencies allocated"); 628 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 629 "current dependencies allocated"); 630 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 631 "current dependencies written"); 632 633 unsigned long dep_current[D_LAST + 1]; 634 unsigned long dep_highuse[D_LAST + 1]; 635 unsigned long dep_total[D_LAST + 1]; 636 unsigned long dep_write[D_LAST + 1]; 637 638 #define SOFTDEP_TYPE(type, str, long) \ 639 static MALLOC_DEFINE(M_ ## type, #str, long); \ 640 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 641 &dep_total[D_ ## type], 0, ""); \ 642 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 643 &dep_current[D_ ## type], 0, ""); \ 644 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 645 &dep_highuse[D_ ## type], 0, ""); \ 646 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 647 &dep_write[D_ ## type], 0, ""); 648 649 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 650 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 651 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 652 "Block or frag allocated from cyl group map"); 653 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 654 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 655 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 656 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 657 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 658 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 659 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 660 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 661 SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 662 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 663 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 664 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 665 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 666 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 667 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 668 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 669 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 670 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 671 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 672 SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 673 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 674 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 675 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 676 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 677 678 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 679 680 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 681 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 682 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 683 684 #define M_SOFTDEP_FLAGS (M_WAITOK) 685 686 /* 687 * translate from workitem type to memory type 688 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 689 */ 690 static struct malloc_type *memtype[] = { 691 NULL, 692 M_PAGEDEP, 693 M_INODEDEP, 694 M_BMSAFEMAP, 695 M_NEWBLK, 696 M_ALLOCDIRECT, 697 M_INDIRDEP, 698 M_ALLOCINDIR, 699 M_FREEFRAG, 700 M_FREEBLKS, 701 M_FREEFILE, 702 M_DIRADD, 703 M_MKDIR, 704 M_DIRREM, 705 M_NEWDIRBLK, 706 M_FREEWORK, 707 M_FREEDEP, 708 M_JADDREF, 709 M_JREMREF, 710 M_JMVREF, 711 M_JNEWBLK, 712 M_JFREEBLK, 713 M_JFREEFRAG, 714 M_JSEG, 715 M_JSEGDEP, 716 M_SBDEP, 717 M_JTRUNC, 718 M_JFSYNC, 719 M_SENTINEL 720 }; 721 722 #define DtoM(type) (memtype[type]) 723 724 /* 725 * Names of malloc types. 726 */ 727 #define TYPENAME(type) \ 728 ((unsigned)(type) <= D_LAST && (unsigned)(type) >= D_FIRST ? \ 729 memtype[type]->ks_shortdesc : "???") 730 /* 731 * End system adaptation definitions. 732 */ 733 734 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 735 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 736 737 /* 738 * Internal function prototypes. 739 */ 740 static void check_clear_deps(struct mount *); 741 static void softdep_error(char *, int); 742 static int softdep_process_worklist(struct mount *, int); 743 static int softdep_waitidle(struct mount *, int); 744 static void drain_output(struct vnode *); 745 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 746 static int check_inodedep_free(struct inodedep *); 747 static void clear_remove(struct mount *); 748 static void clear_inodedeps(struct mount *); 749 static void unlinked_inodedep(struct mount *, struct inodedep *); 750 static void clear_unlinked_inodedep(struct inodedep *); 751 static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 752 static int flush_pagedep_deps(struct vnode *, struct mount *, 753 struct diraddhd *); 754 static int free_pagedep(struct pagedep *); 755 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 756 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 757 static int flush_deplist(struct allocdirectlst *, int, int *); 758 static int sync_cgs(struct mount *, int); 759 static int handle_written_filepage(struct pagedep *, struct buf *, int); 760 static int handle_written_sbdep(struct sbdep *, struct buf *); 761 static void initiate_write_sbdep(struct sbdep *); 762 static void diradd_inode_written(struct diradd *, struct inodedep *); 763 static int handle_written_indirdep(struct indirdep *, struct buf *, 764 struct buf**, int); 765 static int handle_written_inodeblock(struct inodedep *, struct buf *, int); 766 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 767 uint8_t *); 768 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int); 769 static void handle_written_jaddref(struct jaddref *); 770 static void handle_written_jremref(struct jremref *); 771 static void handle_written_jseg(struct jseg *, struct buf *); 772 static void handle_written_jnewblk(struct jnewblk *); 773 static void handle_written_jblkdep(struct jblkdep *); 774 static void handle_written_jfreefrag(struct jfreefrag *); 775 static void complete_jseg(struct jseg *); 776 static void complete_jsegs(struct jseg *); 777 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 778 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 779 static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 780 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 781 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 782 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 783 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 784 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 785 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 786 static inline void inoref_write(struct inoref *, struct jseg *, 787 struct jrefrec *); 788 static void handle_allocdirect_partdone(struct allocdirect *, 789 struct workhead *); 790 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 791 struct workhead *); 792 static void indirdep_complete(struct indirdep *); 793 static int indirblk_lookup(struct mount *, ufs2_daddr_t); 794 static void indirblk_insert(struct freework *); 795 static void indirblk_remove(struct freework *); 796 static void handle_allocindir_partdone(struct allocindir *); 797 static void initiate_write_filepage(struct pagedep *, struct buf *); 798 static void initiate_write_indirdep(struct indirdep*, struct buf *); 799 static void handle_written_mkdir(struct mkdir *, int); 800 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 801 uint8_t *); 802 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 803 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 804 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 805 static void handle_workitem_freefile(struct freefile *); 806 static int handle_workitem_remove(struct dirrem *, int); 807 static struct dirrem *newdirrem(struct buf *, struct inode *, 808 struct inode *, int, struct dirrem **); 809 static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 810 struct buf *); 811 static void cancel_indirdep(struct indirdep *, struct buf *, 812 struct freeblks *); 813 static void free_indirdep(struct indirdep *); 814 static void free_diradd(struct diradd *, struct workhead *); 815 static void merge_diradd(struct inodedep *, struct diradd *); 816 static void complete_diradd(struct diradd *); 817 static struct diradd *diradd_lookup(struct pagedep *, int); 818 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 819 struct jremref *); 820 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 821 struct jremref *); 822 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 823 struct jremref *, struct jremref *); 824 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 825 struct jremref *); 826 static void cancel_allocindir(struct allocindir *, struct buf *bp, 827 struct freeblks *, int); 828 static int setup_trunc_indir(struct freeblks *, struct inode *, 829 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 830 static void complete_trunc_indir(struct freework *); 831 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 832 int); 833 static void complete_mkdir(struct mkdir *); 834 static void free_newdirblk(struct newdirblk *); 835 static void free_jremref(struct jremref *); 836 static void free_jaddref(struct jaddref *); 837 static void free_jsegdep(struct jsegdep *); 838 static void free_jsegs(struct jblocks *); 839 static void rele_jseg(struct jseg *); 840 static void free_jseg(struct jseg *, struct jblocks *); 841 static void free_jnewblk(struct jnewblk *); 842 static void free_jblkdep(struct jblkdep *); 843 static void free_jfreefrag(struct jfreefrag *); 844 static void free_freedep(struct freedep *); 845 static void journal_jremref(struct dirrem *, struct jremref *, 846 struct inodedep *); 847 static void cancel_jnewblk(struct jnewblk *, struct workhead *); 848 static int cancel_jaddref(struct jaddref *, struct inodedep *, 849 struct workhead *); 850 static void cancel_jfreefrag(struct jfreefrag *); 851 static inline void setup_freedirect(struct freeblks *, struct inode *, 852 int, int); 853 static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 854 static inline void setup_freeindir(struct freeblks *, struct inode *, int, 855 ufs_lbn_t, int); 856 static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 857 static void freeblks_free(struct ufsmount *, struct freeblks *, int); 858 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 859 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 860 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 861 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 862 int, int); 863 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 864 static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 865 static int deallocate_dependencies(struct buf *, struct freeblks *, int); 866 static void newblk_freefrag(struct newblk*); 867 static void free_newblk(struct newblk *); 868 static void cancel_allocdirect(struct allocdirectlst *, 869 struct allocdirect *, struct freeblks *); 870 static int check_inode_unwritten(struct inodedep *); 871 static int free_inodedep(struct inodedep *); 872 static void freework_freeblock(struct freework *); 873 static void freework_enqueue(struct freework *); 874 static int handle_workitem_freeblocks(struct freeblks *, int); 875 static int handle_complete_freeblocks(struct freeblks *, int); 876 static void handle_workitem_indirblk(struct freework *); 877 static void handle_written_freework(struct freework *); 878 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 879 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 880 struct workhead *); 881 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 882 struct inodedep *, struct allocindir *, ufs_lbn_t); 883 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 884 ufs2_daddr_t, ufs_lbn_t); 885 static void handle_workitem_freefrag(struct freefrag *); 886 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 887 ufs_lbn_t); 888 static void allocdirect_merge(struct allocdirectlst *, 889 struct allocdirect *, struct allocdirect *); 890 static struct freefrag *allocindir_merge(struct allocindir *, 891 struct allocindir *); 892 static int bmsafemap_find(struct bmsafemap_hashhead *, int, 893 struct bmsafemap **); 894 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 895 int cg, struct bmsafemap *); 896 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 897 struct newblk **); 898 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 899 static int inodedep_find(struct inodedep_hashhead *, ino_t, 900 struct inodedep **); 901 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 902 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 903 int, struct pagedep **); 904 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 905 struct pagedep **); 906 static void pause_timer(void *); 907 static int request_cleanup(struct mount *, int); 908 static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *); 909 static void schedule_cleanup(struct mount *); 910 static void softdep_ast_cleanup_proc(struct thread *); 911 static struct ufsmount *softdep_bp_to_mp(struct buf *bp); 912 static int process_worklist_item(struct mount *, int, int); 913 static void process_removes(struct vnode *); 914 static void process_truncates(struct vnode *); 915 static void jwork_move(struct workhead *, struct workhead *); 916 static void jwork_insert(struct workhead *, struct jsegdep *); 917 static void add_to_worklist(struct worklist *, int); 918 static void wake_worklist(struct worklist *); 919 static void wait_worklist(struct worklist *, char *); 920 static void remove_from_worklist(struct worklist *); 921 static void softdep_flush(void *); 922 static void softdep_flushjournal(struct mount *); 923 static int softdep_speedup(struct ufsmount *); 924 static void worklist_speedup(struct mount *); 925 static int journal_mount(struct mount *, struct fs *, struct ucred *); 926 static void journal_unmount(struct ufsmount *); 927 static int journal_space(struct ufsmount *, int); 928 static void journal_suspend(struct ufsmount *); 929 static int journal_unsuspend(struct ufsmount *ump); 930 static void softdep_prelink(struct vnode *, struct vnode *); 931 static void add_to_journal(struct worklist *); 932 static void remove_from_journal(struct worklist *); 933 static bool softdep_excess_items(struct ufsmount *, int); 934 static void softdep_process_journal(struct mount *, struct worklist *, int); 935 static struct jremref *newjremref(struct dirrem *, struct inode *, 936 struct inode *ip, off_t, nlink_t); 937 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 938 uint16_t); 939 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 940 uint16_t); 941 static inline struct jsegdep *inoref_jseg(struct inoref *); 942 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 943 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 944 ufs2_daddr_t, int); 945 static void adjust_newfreework(struct freeblks *, int); 946 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 947 static void move_newblock_dep(struct jaddref *, struct inodedep *); 948 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 949 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 950 ufs2_daddr_t, long, ufs_lbn_t); 951 static struct freework *newfreework(struct ufsmount *, struct freeblks *, 952 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 953 static int jwait(struct worklist *, int); 954 static struct inodedep *inodedep_lookup_ip(struct inode *); 955 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 956 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 957 static void handle_jwork(struct workhead *); 958 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 959 struct mkdir **); 960 static struct jblocks *jblocks_create(void); 961 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 962 static void jblocks_free(struct jblocks *, struct mount *, int); 963 static void jblocks_destroy(struct jblocks *); 964 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 965 966 /* 967 * Exported softdep operations. 968 */ 969 static void softdep_disk_io_initiation(struct buf *); 970 static void softdep_disk_write_complete(struct buf *); 971 static void softdep_deallocate_dependencies(struct buf *); 972 static int softdep_count_dependencies(struct buf *bp, int); 973 974 /* 975 * Global lock over all of soft updates. 976 */ 977 static struct mtx lk; 978 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF); 979 980 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 981 #define FREE_GBLLOCK(lk) mtx_unlock(lk) 982 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 983 984 /* 985 * Per-filesystem soft-updates locking. 986 */ 987 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 988 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 989 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 990 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 991 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 992 RA_WLOCKED) 993 994 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 995 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 996 997 /* 998 * Worklist queue management. 999 * These routines require that the lock be held. 1000 */ 1001 #ifndef /* NOT */ DEBUG 1002 #define WORKLIST_INSERT(head, item) do { \ 1003 (item)->wk_state |= ONWORKLIST; \ 1004 LIST_INSERT_HEAD(head, item, wk_list); \ 1005 } while (0) 1006 #define WORKLIST_REMOVE(item) do { \ 1007 (item)->wk_state &= ~ONWORKLIST; \ 1008 LIST_REMOVE(item, wk_list); \ 1009 } while (0) 1010 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1011 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1012 1013 #else /* DEBUG */ 1014 static void worklist_insert(struct workhead *, struct worklist *, int); 1015 static void worklist_remove(struct worklist *, int); 1016 1017 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1018 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1019 #define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1020 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1021 1022 static void 1023 worklist_insert(head, item, locked) 1024 struct workhead *head; 1025 struct worklist *item; 1026 int locked; 1027 { 1028 1029 if (locked) 1030 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1031 if (item->wk_state & ONWORKLIST) 1032 panic("worklist_insert: %p %s(0x%X) already on list", 1033 item, TYPENAME(item->wk_type), item->wk_state); 1034 item->wk_state |= ONWORKLIST; 1035 LIST_INSERT_HEAD(head, item, wk_list); 1036 } 1037 1038 static void 1039 worklist_remove(item, locked) 1040 struct worklist *item; 1041 int locked; 1042 { 1043 1044 if (locked) 1045 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1046 if ((item->wk_state & ONWORKLIST) == 0) 1047 panic("worklist_remove: %p %s(0x%X) not on list", 1048 item, TYPENAME(item->wk_type), item->wk_state); 1049 item->wk_state &= ~ONWORKLIST; 1050 LIST_REMOVE(item, wk_list); 1051 } 1052 #endif /* DEBUG */ 1053 1054 /* 1055 * Merge two jsegdeps keeping only the oldest one as newer references 1056 * can't be discarded until after older references. 1057 */ 1058 static inline struct jsegdep * 1059 jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1060 { 1061 struct jsegdep *swp; 1062 1063 if (two == NULL) 1064 return (one); 1065 1066 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1067 swp = one; 1068 one = two; 1069 two = swp; 1070 } 1071 WORKLIST_REMOVE(&two->jd_list); 1072 free_jsegdep(two); 1073 1074 return (one); 1075 } 1076 1077 /* 1078 * If two freedeps are compatible free one to reduce list size. 1079 */ 1080 static inline struct freedep * 1081 freedep_merge(struct freedep *one, struct freedep *two) 1082 { 1083 if (two == NULL) 1084 return (one); 1085 1086 if (one->fd_freework == two->fd_freework) { 1087 WORKLIST_REMOVE(&two->fd_list); 1088 free_freedep(two); 1089 } 1090 return (one); 1091 } 1092 1093 /* 1094 * Move journal work from one list to another. Duplicate freedeps and 1095 * jsegdeps are coalesced to keep the lists as small as possible. 1096 */ 1097 static void 1098 jwork_move(dst, src) 1099 struct workhead *dst; 1100 struct workhead *src; 1101 { 1102 struct freedep *freedep; 1103 struct jsegdep *jsegdep; 1104 struct worklist *wkn; 1105 struct worklist *wk; 1106 1107 KASSERT(dst != src, 1108 ("jwork_move: dst == src")); 1109 freedep = NULL; 1110 jsegdep = NULL; 1111 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1112 if (wk->wk_type == D_JSEGDEP) 1113 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1114 else if (wk->wk_type == D_FREEDEP) 1115 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1116 } 1117 1118 while ((wk = LIST_FIRST(src)) != NULL) { 1119 WORKLIST_REMOVE(wk); 1120 WORKLIST_INSERT(dst, wk); 1121 if (wk->wk_type == D_JSEGDEP) { 1122 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1123 continue; 1124 } 1125 if (wk->wk_type == D_FREEDEP) 1126 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1127 } 1128 } 1129 1130 static void 1131 jwork_insert(dst, jsegdep) 1132 struct workhead *dst; 1133 struct jsegdep *jsegdep; 1134 { 1135 struct jsegdep *jsegdepn; 1136 struct worklist *wk; 1137 1138 LIST_FOREACH(wk, dst, wk_list) 1139 if (wk->wk_type == D_JSEGDEP) 1140 break; 1141 if (wk == NULL) { 1142 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1143 return; 1144 } 1145 jsegdepn = WK_JSEGDEP(wk); 1146 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1147 WORKLIST_REMOVE(wk); 1148 free_jsegdep(jsegdepn); 1149 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1150 } else 1151 free_jsegdep(jsegdep); 1152 } 1153 1154 /* 1155 * Routines for tracking and managing workitems. 1156 */ 1157 static void workitem_free(struct worklist *, int); 1158 static void workitem_alloc(struct worklist *, int, struct mount *); 1159 static void workitem_reassign(struct worklist *, int); 1160 1161 #define WORKITEM_FREE(item, type) \ 1162 workitem_free((struct worklist *)(item), (type)) 1163 #define WORKITEM_REASSIGN(item, type) \ 1164 workitem_reassign((struct worklist *)(item), (type)) 1165 1166 static void 1167 workitem_free(item, type) 1168 struct worklist *item; 1169 int type; 1170 { 1171 struct ufsmount *ump; 1172 1173 #ifdef DEBUG 1174 if (item->wk_state & ONWORKLIST) 1175 panic("workitem_free: %s(0x%X) still on list", 1176 TYPENAME(item->wk_type), item->wk_state); 1177 if (item->wk_type != type && type != D_NEWBLK) 1178 panic("workitem_free: type mismatch %s != %s", 1179 TYPENAME(item->wk_type), TYPENAME(type)); 1180 #endif 1181 if (item->wk_state & IOWAITING) 1182 wakeup(item); 1183 ump = VFSTOUFS(item->wk_mp); 1184 LOCK_OWNED(ump); 1185 KASSERT(ump->softdep_deps > 0, 1186 ("workitem_free: %s: softdep_deps going negative", 1187 ump->um_fs->fs_fsmnt)); 1188 if (--ump->softdep_deps == 0 && ump->softdep_req) 1189 wakeup(&ump->softdep_deps); 1190 KASSERT(dep_current[item->wk_type] > 0, 1191 ("workitem_free: %s: dep_current[%s] going negative", 1192 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1193 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1194 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1195 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1196 atomic_subtract_long(&dep_current[item->wk_type], 1); 1197 ump->softdep_curdeps[item->wk_type] -= 1; 1198 free(item, DtoM(type)); 1199 } 1200 1201 static void 1202 workitem_alloc(item, type, mp) 1203 struct worklist *item; 1204 int type; 1205 struct mount *mp; 1206 { 1207 struct ufsmount *ump; 1208 1209 item->wk_type = type; 1210 item->wk_mp = mp; 1211 item->wk_state = 0; 1212 1213 ump = VFSTOUFS(mp); 1214 ACQUIRE_GBLLOCK(&lk); 1215 dep_current[type]++; 1216 if (dep_current[type] > dep_highuse[type]) 1217 dep_highuse[type] = dep_current[type]; 1218 dep_total[type]++; 1219 FREE_GBLLOCK(&lk); 1220 ACQUIRE_LOCK(ump); 1221 ump->softdep_curdeps[type] += 1; 1222 ump->softdep_deps++; 1223 ump->softdep_accdeps++; 1224 FREE_LOCK(ump); 1225 } 1226 1227 static void 1228 workitem_reassign(item, newtype) 1229 struct worklist *item; 1230 int newtype; 1231 { 1232 struct ufsmount *ump; 1233 1234 ump = VFSTOUFS(item->wk_mp); 1235 LOCK_OWNED(ump); 1236 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1237 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1238 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1239 ump->softdep_curdeps[item->wk_type] -= 1; 1240 ump->softdep_curdeps[newtype] += 1; 1241 KASSERT(dep_current[item->wk_type] > 0, 1242 ("workitem_reassign: %s: dep_current[%s] going negative", 1243 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1244 ACQUIRE_GBLLOCK(&lk); 1245 dep_current[newtype]++; 1246 dep_current[item->wk_type]--; 1247 if (dep_current[newtype] > dep_highuse[newtype]) 1248 dep_highuse[newtype] = dep_current[newtype]; 1249 dep_total[newtype]++; 1250 FREE_GBLLOCK(&lk); 1251 item->wk_type = newtype; 1252 } 1253 1254 /* 1255 * Workitem queue management 1256 */ 1257 static int max_softdeps; /* maximum number of structs before slowdown */ 1258 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1259 static int proc_waiting; /* tracks whether we have a timeout posted */ 1260 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1261 static struct callout softdep_callout; 1262 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1263 static int req_clear_remove; /* syncer process flush some freeblks */ 1264 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1265 1266 /* 1267 * runtime statistics 1268 */ 1269 static int stat_flush_threads; /* number of softdep flushing threads */ 1270 static int stat_worklist_push; /* number of worklist cleanups */ 1271 static int stat_blk_limit_push; /* number of times block limit neared */ 1272 static int stat_ino_limit_push; /* number of times inode limit neared */ 1273 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1274 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1275 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1276 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1277 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1278 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1279 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1280 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1281 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1282 static int stat_journal_min; /* Times hit journal min threshold */ 1283 static int stat_journal_low; /* Times hit journal low threshold */ 1284 static int stat_journal_wait; /* Times blocked in jwait(). */ 1285 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1286 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1287 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1288 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1289 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1290 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1291 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1292 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1293 static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1294 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1295 1296 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1297 &max_softdeps, 0, ""); 1298 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1299 &tickdelay, 0, ""); 1300 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1301 &stat_flush_threads, 0, ""); 1302 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1303 &stat_worklist_push, 0,""); 1304 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1305 &stat_blk_limit_push, 0,""); 1306 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1307 &stat_ino_limit_push, 0,""); 1308 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1309 &stat_blk_limit_hit, 0, ""); 1310 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1311 &stat_ino_limit_hit, 0, ""); 1312 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1313 &stat_sync_limit_hit, 0, ""); 1314 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1315 &stat_indir_blk_ptrs, 0, ""); 1316 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1317 &stat_inode_bitmap, 0, ""); 1318 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1319 &stat_direct_blk_ptrs, 0, ""); 1320 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1321 &stat_dir_entry, 0, ""); 1322 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1323 &stat_jaddref, 0, ""); 1324 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1325 &stat_jnewblk, 0, ""); 1326 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1327 &stat_journal_low, 0, ""); 1328 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1329 &stat_journal_min, 0, ""); 1330 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1331 &stat_journal_wait, 0, ""); 1332 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1333 &stat_jwait_filepage, 0, ""); 1334 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1335 &stat_jwait_freeblks, 0, ""); 1336 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1337 &stat_jwait_inode, 0, ""); 1338 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1339 &stat_jwait_newblk, 0, ""); 1340 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1341 &stat_cleanup_blkrequests, 0, ""); 1342 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1343 &stat_cleanup_inorequests, 0, ""); 1344 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1345 &stat_cleanup_high_delay, 0, ""); 1346 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1347 &stat_cleanup_retries, 0, ""); 1348 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1349 &stat_cleanup_failures, 0, ""); 1350 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1351 &softdep_flushcache, 0, ""); 1352 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1353 &stat_emptyjblocks, 0, ""); 1354 1355 SYSCTL_DECL(_vfs_ffs); 1356 1357 /* Whether to recompute the summary at mount time */ 1358 static int compute_summary_at_mount = 0; 1359 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1360 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1361 static int print_threads = 0; 1362 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1363 &print_threads, 0, "Notify flusher thread start/stop"); 1364 1365 /* List of all filesystems mounted with soft updates */ 1366 static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1367 1368 /* 1369 * This function cleans the worklist for a filesystem. 1370 * Each filesystem running with soft dependencies gets its own 1371 * thread to run in this function. The thread is started up in 1372 * softdep_mount and shutdown in softdep_unmount. They show up 1373 * as part of the kernel "bufdaemon" process whose process 1374 * entry is available in bufdaemonproc. 1375 */ 1376 static int searchfailed; 1377 extern struct proc *bufdaemonproc; 1378 static void 1379 softdep_flush(addr) 1380 void *addr; 1381 { 1382 struct mount *mp; 1383 struct thread *td; 1384 struct ufsmount *ump; 1385 1386 td = curthread; 1387 td->td_pflags |= TDP_NORUNNINGBUF; 1388 mp = (struct mount *)addr; 1389 ump = VFSTOUFS(mp); 1390 atomic_add_int(&stat_flush_threads, 1); 1391 ACQUIRE_LOCK(ump); 1392 ump->softdep_flags &= ~FLUSH_STARTING; 1393 wakeup(&ump->softdep_flushtd); 1394 FREE_LOCK(ump); 1395 if (print_threads) { 1396 if (stat_flush_threads == 1) 1397 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1398 bufdaemonproc->p_pid); 1399 printf("Start thread %s\n", td->td_name); 1400 } 1401 for (;;) { 1402 while (softdep_process_worklist(mp, 0) > 0 || 1403 (MOUNTEDSUJ(mp) && 1404 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1405 kthread_suspend_check(); 1406 ACQUIRE_LOCK(ump); 1407 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1408 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1409 "sdflush", hz / 2); 1410 ump->softdep_flags &= ~FLUSH_CLEANUP; 1411 /* 1412 * Check to see if we are done and need to exit. 1413 */ 1414 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1415 FREE_LOCK(ump); 1416 continue; 1417 } 1418 ump->softdep_flags &= ~FLUSH_EXIT; 1419 FREE_LOCK(ump); 1420 wakeup(&ump->softdep_flags); 1421 if (print_threads) 1422 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1423 atomic_subtract_int(&stat_flush_threads, 1); 1424 kthread_exit(); 1425 panic("kthread_exit failed\n"); 1426 } 1427 } 1428 1429 static void 1430 worklist_speedup(mp) 1431 struct mount *mp; 1432 { 1433 struct ufsmount *ump; 1434 1435 ump = VFSTOUFS(mp); 1436 LOCK_OWNED(ump); 1437 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1438 ump->softdep_flags |= FLUSH_CLEANUP; 1439 wakeup(&ump->softdep_flushtd); 1440 } 1441 1442 static int 1443 softdep_speedup(ump) 1444 struct ufsmount *ump; 1445 { 1446 struct ufsmount *altump; 1447 struct mount_softdeps *sdp; 1448 1449 LOCK_OWNED(ump); 1450 worklist_speedup(ump->um_mountp); 1451 bd_speedup(); 1452 /* 1453 * If we have global shortages, then we need other 1454 * filesystems to help with the cleanup. Here we wakeup a 1455 * flusher thread for a filesystem that is over its fair 1456 * share of resources. 1457 */ 1458 if (req_clear_inodedeps || req_clear_remove) { 1459 ACQUIRE_GBLLOCK(&lk); 1460 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1461 if ((altump = sdp->sd_ump) == ump) 1462 continue; 1463 if (((req_clear_inodedeps && 1464 altump->softdep_curdeps[D_INODEDEP] > 1465 max_softdeps / stat_flush_threads) || 1466 (req_clear_remove && 1467 altump->softdep_curdeps[D_DIRREM] > 1468 (max_softdeps / 2) / stat_flush_threads)) && 1469 TRY_ACQUIRE_LOCK(altump)) 1470 break; 1471 } 1472 if (sdp == NULL) { 1473 searchfailed++; 1474 FREE_GBLLOCK(&lk); 1475 } else { 1476 /* 1477 * Move to the end of the list so we pick a 1478 * different one on out next try. 1479 */ 1480 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1481 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1482 FREE_GBLLOCK(&lk); 1483 if ((altump->softdep_flags & 1484 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1485 altump->softdep_flags |= FLUSH_CLEANUP; 1486 altump->um_softdep->sd_cleanups++; 1487 wakeup(&altump->softdep_flushtd); 1488 FREE_LOCK(altump); 1489 } 1490 } 1491 return (speedup_syncer()); 1492 } 1493 1494 /* 1495 * Add an item to the end of the work queue. 1496 * This routine requires that the lock be held. 1497 * This is the only routine that adds items to the list. 1498 * The following routine is the only one that removes items 1499 * and does so in order from first to last. 1500 */ 1501 1502 #define WK_HEAD 0x0001 /* Add to HEAD. */ 1503 #define WK_NODELAY 0x0002 /* Process immediately. */ 1504 1505 static void 1506 add_to_worklist(wk, flags) 1507 struct worklist *wk; 1508 int flags; 1509 { 1510 struct ufsmount *ump; 1511 1512 ump = VFSTOUFS(wk->wk_mp); 1513 LOCK_OWNED(ump); 1514 if (wk->wk_state & ONWORKLIST) 1515 panic("add_to_worklist: %s(0x%X) already on list", 1516 TYPENAME(wk->wk_type), wk->wk_state); 1517 wk->wk_state |= ONWORKLIST; 1518 if (ump->softdep_on_worklist == 0) { 1519 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1520 ump->softdep_worklist_tail = wk; 1521 } else if (flags & WK_HEAD) { 1522 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1523 } else { 1524 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1525 ump->softdep_worklist_tail = wk; 1526 } 1527 ump->softdep_on_worklist += 1; 1528 if (flags & WK_NODELAY) 1529 worklist_speedup(wk->wk_mp); 1530 } 1531 1532 /* 1533 * Remove the item to be processed. If we are removing the last 1534 * item on the list, we need to recalculate the tail pointer. 1535 */ 1536 static void 1537 remove_from_worklist(wk) 1538 struct worklist *wk; 1539 { 1540 struct ufsmount *ump; 1541 1542 ump = VFSTOUFS(wk->wk_mp); 1543 if (ump->softdep_worklist_tail == wk) 1544 ump->softdep_worklist_tail = 1545 (struct worklist *)wk->wk_list.le_prev; 1546 WORKLIST_REMOVE(wk); 1547 ump->softdep_on_worklist -= 1; 1548 } 1549 1550 static void 1551 wake_worklist(wk) 1552 struct worklist *wk; 1553 { 1554 if (wk->wk_state & IOWAITING) { 1555 wk->wk_state &= ~IOWAITING; 1556 wakeup(wk); 1557 } 1558 } 1559 1560 static void 1561 wait_worklist(wk, wmesg) 1562 struct worklist *wk; 1563 char *wmesg; 1564 { 1565 struct ufsmount *ump; 1566 1567 ump = VFSTOUFS(wk->wk_mp); 1568 wk->wk_state |= IOWAITING; 1569 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1570 } 1571 1572 /* 1573 * Process that runs once per second to handle items in the background queue. 1574 * 1575 * Note that we ensure that everything is done in the order in which they 1576 * appear in the queue. The code below depends on this property to ensure 1577 * that blocks of a file are freed before the inode itself is freed. This 1578 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1579 * until all the old ones have been purged from the dependency lists. 1580 */ 1581 static int 1582 softdep_process_worklist(mp, full) 1583 struct mount *mp; 1584 int full; 1585 { 1586 int cnt, matchcnt; 1587 struct ufsmount *ump; 1588 long starttime; 1589 1590 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1591 if (MOUNTEDSOFTDEP(mp) == 0) 1592 return (0); 1593 matchcnt = 0; 1594 ump = VFSTOUFS(mp); 1595 ACQUIRE_LOCK(ump); 1596 starttime = time_second; 1597 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1598 check_clear_deps(mp); 1599 while (ump->softdep_on_worklist > 0) { 1600 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1601 break; 1602 else 1603 matchcnt += cnt; 1604 check_clear_deps(mp); 1605 /* 1606 * We do not generally want to stop for buffer space, but if 1607 * we are really being a buffer hog, we will stop and wait. 1608 */ 1609 if (should_yield()) { 1610 FREE_LOCK(ump); 1611 kern_yield(PRI_USER); 1612 bwillwrite(); 1613 ACQUIRE_LOCK(ump); 1614 } 1615 /* 1616 * Never allow processing to run for more than one 1617 * second. This gives the syncer thread the opportunity 1618 * to pause if appropriate. 1619 */ 1620 if (!full && starttime != time_second) 1621 break; 1622 } 1623 if (full == 0) 1624 journal_unsuspend(ump); 1625 FREE_LOCK(ump); 1626 return (matchcnt); 1627 } 1628 1629 /* 1630 * Process all removes associated with a vnode if we are running out of 1631 * journal space. Any other process which attempts to flush these will 1632 * be unable as we have the vnodes locked. 1633 */ 1634 static void 1635 process_removes(vp) 1636 struct vnode *vp; 1637 { 1638 struct inodedep *inodedep; 1639 struct dirrem *dirrem; 1640 struct ufsmount *ump; 1641 struct mount *mp; 1642 ino_t inum; 1643 1644 mp = vp->v_mount; 1645 ump = VFSTOUFS(mp); 1646 LOCK_OWNED(ump); 1647 inum = VTOI(vp)->i_number; 1648 for (;;) { 1649 top: 1650 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1651 return; 1652 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1653 /* 1654 * If another thread is trying to lock this vnode 1655 * it will fail but we must wait for it to do so 1656 * before we can proceed. 1657 */ 1658 if (dirrem->dm_state & INPROGRESS) { 1659 wait_worklist(&dirrem->dm_list, "pwrwait"); 1660 goto top; 1661 } 1662 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1663 (COMPLETE | ONWORKLIST)) 1664 break; 1665 } 1666 if (dirrem == NULL) 1667 return; 1668 remove_from_worklist(&dirrem->dm_list); 1669 FREE_LOCK(ump); 1670 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1671 panic("process_removes: suspended filesystem"); 1672 handle_workitem_remove(dirrem, 0); 1673 vn_finished_secondary_write(mp); 1674 ACQUIRE_LOCK(ump); 1675 } 1676 } 1677 1678 /* 1679 * Process all truncations associated with a vnode if we are running out 1680 * of journal space. This is called when the vnode lock is already held 1681 * and no other process can clear the truncation. This function returns 1682 * a value greater than zero if it did any work. 1683 */ 1684 static void 1685 process_truncates(vp) 1686 struct vnode *vp; 1687 { 1688 struct inodedep *inodedep; 1689 struct freeblks *freeblks; 1690 struct ufsmount *ump; 1691 struct mount *mp; 1692 ino_t inum; 1693 int cgwait; 1694 1695 mp = vp->v_mount; 1696 ump = VFSTOUFS(mp); 1697 LOCK_OWNED(ump); 1698 inum = VTOI(vp)->i_number; 1699 for (;;) { 1700 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1701 return; 1702 cgwait = 0; 1703 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1704 /* Journal entries not yet written. */ 1705 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1706 jwait(&LIST_FIRST( 1707 &freeblks->fb_jblkdephd)->jb_list, 1708 MNT_WAIT); 1709 break; 1710 } 1711 /* Another thread is executing this item. */ 1712 if (freeblks->fb_state & INPROGRESS) { 1713 wait_worklist(&freeblks->fb_list, "ptrwait"); 1714 break; 1715 } 1716 /* Freeblks is waiting on a inode write. */ 1717 if ((freeblks->fb_state & COMPLETE) == 0) { 1718 FREE_LOCK(ump); 1719 ffs_update(vp, 1); 1720 ACQUIRE_LOCK(ump); 1721 break; 1722 } 1723 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1724 (ALLCOMPLETE | ONWORKLIST)) { 1725 remove_from_worklist(&freeblks->fb_list); 1726 freeblks->fb_state |= INPROGRESS; 1727 FREE_LOCK(ump); 1728 if (vn_start_secondary_write(NULL, &mp, 1729 V_NOWAIT)) 1730 panic("process_truncates: " 1731 "suspended filesystem"); 1732 handle_workitem_freeblocks(freeblks, 0); 1733 vn_finished_secondary_write(mp); 1734 ACQUIRE_LOCK(ump); 1735 break; 1736 } 1737 if (freeblks->fb_cgwait) 1738 cgwait++; 1739 } 1740 if (cgwait) { 1741 FREE_LOCK(ump); 1742 sync_cgs(mp, MNT_WAIT); 1743 ffs_sync_snap(mp, MNT_WAIT); 1744 ACQUIRE_LOCK(ump); 1745 continue; 1746 } 1747 if (freeblks == NULL) 1748 break; 1749 } 1750 return; 1751 } 1752 1753 /* 1754 * Process one item on the worklist. 1755 */ 1756 static int 1757 process_worklist_item(mp, target, flags) 1758 struct mount *mp; 1759 int target; 1760 int flags; 1761 { 1762 struct worklist sentinel; 1763 struct worklist *wk; 1764 struct ufsmount *ump; 1765 int matchcnt; 1766 int error; 1767 1768 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1769 /* 1770 * If we are being called because of a process doing a 1771 * copy-on-write, then it is not safe to write as we may 1772 * recurse into the copy-on-write routine. 1773 */ 1774 if (curthread->td_pflags & TDP_COWINPROGRESS) 1775 return (-1); 1776 PHOLD(curproc); /* Don't let the stack go away. */ 1777 ump = VFSTOUFS(mp); 1778 LOCK_OWNED(ump); 1779 matchcnt = 0; 1780 sentinel.wk_mp = NULL; 1781 sentinel.wk_type = D_SENTINEL; 1782 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1783 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1784 wk = LIST_NEXT(&sentinel, wk_list)) { 1785 if (wk->wk_type == D_SENTINEL) { 1786 LIST_REMOVE(&sentinel, wk_list); 1787 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1788 continue; 1789 } 1790 if (wk->wk_state & INPROGRESS) 1791 panic("process_worklist_item: %p already in progress.", 1792 wk); 1793 wk->wk_state |= INPROGRESS; 1794 remove_from_worklist(wk); 1795 FREE_LOCK(ump); 1796 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1797 panic("process_worklist_item: suspended filesystem"); 1798 switch (wk->wk_type) { 1799 case D_DIRREM: 1800 /* removal of a directory entry */ 1801 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1802 break; 1803 1804 case D_FREEBLKS: 1805 /* releasing blocks and/or fragments from a file */ 1806 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1807 flags); 1808 break; 1809 1810 case D_FREEFRAG: 1811 /* releasing a fragment when replaced as a file grows */ 1812 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1813 error = 0; 1814 break; 1815 1816 case D_FREEFILE: 1817 /* releasing an inode when its link count drops to 0 */ 1818 handle_workitem_freefile(WK_FREEFILE(wk)); 1819 error = 0; 1820 break; 1821 1822 default: 1823 panic("%s_process_worklist: Unknown type %s", 1824 "softdep", TYPENAME(wk->wk_type)); 1825 /* NOTREACHED */ 1826 } 1827 vn_finished_secondary_write(mp); 1828 ACQUIRE_LOCK(ump); 1829 if (error == 0) { 1830 if (++matchcnt == target) 1831 break; 1832 continue; 1833 } 1834 /* 1835 * We have to retry the worklist item later. Wake up any 1836 * waiters who may be able to complete it immediately and 1837 * add the item back to the head so we don't try to execute 1838 * it again. 1839 */ 1840 wk->wk_state &= ~INPROGRESS; 1841 wake_worklist(wk); 1842 add_to_worklist(wk, WK_HEAD); 1843 } 1844 /* Sentinal could've become the tail from remove_from_worklist. */ 1845 if (ump->softdep_worklist_tail == &sentinel) 1846 ump->softdep_worklist_tail = 1847 (struct worklist *)sentinel.wk_list.le_prev; 1848 LIST_REMOVE(&sentinel, wk_list); 1849 PRELE(curproc); 1850 return (matchcnt); 1851 } 1852 1853 /* 1854 * Move dependencies from one buffer to another. 1855 */ 1856 int 1857 softdep_move_dependencies(oldbp, newbp) 1858 struct buf *oldbp; 1859 struct buf *newbp; 1860 { 1861 struct worklist *wk, *wktail; 1862 struct ufsmount *ump; 1863 int dirty; 1864 1865 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1866 return (0); 1867 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1868 ("softdep_move_dependencies called on non-softdep filesystem")); 1869 dirty = 0; 1870 wktail = NULL; 1871 ump = VFSTOUFS(wk->wk_mp); 1872 ACQUIRE_LOCK(ump); 1873 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1874 LIST_REMOVE(wk, wk_list); 1875 if (wk->wk_type == D_BMSAFEMAP && 1876 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1877 dirty = 1; 1878 if (wktail == NULL) 1879 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1880 else 1881 LIST_INSERT_AFTER(wktail, wk, wk_list); 1882 wktail = wk; 1883 } 1884 FREE_LOCK(ump); 1885 1886 return (dirty); 1887 } 1888 1889 /* 1890 * Purge the work list of all items associated with a particular mount point. 1891 */ 1892 int 1893 softdep_flushworklist(oldmnt, countp, td) 1894 struct mount *oldmnt; 1895 int *countp; 1896 struct thread *td; 1897 { 1898 struct vnode *devvp; 1899 struct ufsmount *ump; 1900 int count, error; 1901 1902 /* 1903 * Alternately flush the block device associated with the mount 1904 * point and process any dependencies that the flushing 1905 * creates. We continue until no more worklist dependencies 1906 * are found. 1907 */ 1908 *countp = 0; 1909 error = 0; 1910 ump = VFSTOUFS(oldmnt); 1911 devvp = ump->um_devvp; 1912 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1913 *countp += count; 1914 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1915 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1916 VOP_UNLOCK(devvp, 0); 1917 if (error != 0) 1918 break; 1919 } 1920 return (error); 1921 } 1922 1923 #define SU_WAITIDLE_RETRIES 20 1924 static int 1925 softdep_waitidle(struct mount *mp, int flags __unused) 1926 { 1927 struct ufsmount *ump; 1928 struct vnode *devvp; 1929 struct thread *td; 1930 int error, i; 1931 1932 ump = VFSTOUFS(mp); 1933 devvp = ump->um_devvp; 1934 td = curthread; 1935 error = 0; 1936 ACQUIRE_LOCK(ump); 1937 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 1938 ump->softdep_req = 1; 1939 KASSERT((flags & FORCECLOSE) == 0 || 1940 ump->softdep_on_worklist == 0, 1941 ("softdep_waitidle: work added after flush")); 1942 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 1943 "softdeps", 10 * hz); 1944 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1945 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1946 VOP_UNLOCK(devvp, 0); 1947 ACQUIRE_LOCK(ump); 1948 if (error != 0) 1949 break; 1950 } 1951 ump->softdep_req = 0; 1952 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 1953 error = EBUSY; 1954 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1955 mp); 1956 } 1957 FREE_LOCK(ump); 1958 return (error); 1959 } 1960 1961 /* 1962 * Flush all vnodes and worklist items associated with a specified mount point. 1963 */ 1964 int 1965 softdep_flushfiles(oldmnt, flags, td) 1966 struct mount *oldmnt; 1967 int flags; 1968 struct thread *td; 1969 { 1970 #ifdef QUOTA 1971 struct ufsmount *ump; 1972 int i; 1973 #endif 1974 int error, early, depcount, loopcnt, retry_flush_count, retry; 1975 int morework; 1976 1977 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1978 ("softdep_flushfiles called on non-softdep filesystem")); 1979 loopcnt = 10; 1980 retry_flush_count = 3; 1981 retry_flush: 1982 error = 0; 1983 1984 /* 1985 * Alternately flush the vnodes associated with the mount 1986 * point and process any dependencies that the flushing 1987 * creates. In theory, this loop can happen at most twice, 1988 * but we give it a few extra just to be sure. 1989 */ 1990 for (; loopcnt > 0; loopcnt--) { 1991 /* 1992 * Do another flush in case any vnodes were brought in 1993 * as part of the cleanup operations. 1994 */ 1995 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1996 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1997 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1998 break; 1999 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 2000 depcount == 0) 2001 break; 2002 } 2003 /* 2004 * If we are unmounting then it is an error to fail. If we 2005 * are simply trying to downgrade to read-only, then filesystem 2006 * activity can keep us busy forever, so we just fail with EBUSY. 2007 */ 2008 if (loopcnt == 0) { 2009 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2010 panic("softdep_flushfiles: looping"); 2011 error = EBUSY; 2012 } 2013 if (!error) 2014 error = softdep_waitidle(oldmnt, flags); 2015 if (!error) { 2016 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2017 retry = 0; 2018 MNT_ILOCK(oldmnt); 2019 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 2020 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 2021 morework = oldmnt->mnt_nvnodelistsize > 0; 2022 #ifdef QUOTA 2023 ump = VFSTOUFS(oldmnt); 2024 UFS_LOCK(ump); 2025 for (i = 0; i < MAXQUOTAS; i++) { 2026 if (ump->um_quotas[i] != NULLVP) 2027 morework = 1; 2028 } 2029 UFS_UNLOCK(ump); 2030 #endif 2031 if (morework) { 2032 if (--retry_flush_count > 0) { 2033 retry = 1; 2034 loopcnt = 3; 2035 } else 2036 error = EBUSY; 2037 } 2038 MNT_IUNLOCK(oldmnt); 2039 if (retry) 2040 goto retry_flush; 2041 } 2042 } 2043 return (error); 2044 } 2045 2046 /* 2047 * Structure hashing. 2048 * 2049 * There are four types of structures that can be looked up: 2050 * 1) pagedep structures identified by mount point, inode number, 2051 * and logical block. 2052 * 2) inodedep structures identified by mount point and inode number. 2053 * 3) newblk structures identified by mount point and 2054 * physical block number. 2055 * 4) bmsafemap structures identified by mount point and 2056 * cylinder group number. 2057 * 2058 * The "pagedep" and "inodedep" dependency structures are hashed 2059 * separately from the file blocks and inodes to which they correspond. 2060 * This separation helps when the in-memory copy of an inode or 2061 * file block must be replaced. It also obviates the need to access 2062 * an inode or file page when simply updating (or de-allocating) 2063 * dependency structures. Lookup of newblk structures is needed to 2064 * find newly allocated blocks when trying to associate them with 2065 * their allocdirect or allocindir structure. 2066 * 2067 * The lookup routines optionally create and hash a new instance when 2068 * an existing entry is not found. The bmsafemap lookup routine always 2069 * allocates a new structure if an existing one is not found. 2070 */ 2071 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2072 2073 /* 2074 * Structures and routines associated with pagedep caching. 2075 */ 2076 #define PAGEDEP_HASH(ump, inum, lbn) \ 2077 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2078 2079 static int 2080 pagedep_find(pagedephd, ino, lbn, pagedeppp) 2081 struct pagedep_hashhead *pagedephd; 2082 ino_t ino; 2083 ufs_lbn_t lbn; 2084 struct pagedep **pagedeppp; 2085 { 2086 struct pagedep *pagedep; 2087 2088 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2089 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2090 *pagedeppp = pagedep; 2091 return (1); 2092 } 2093 } 2094 *pagedeppp = NULL; 2095 return (0); 2096 } 2097 /* 2098 * Look up a pagedep. Return 1 if found, 0 otherwise. 2099 * If not found, allocate if DEPALLOC flag is passed. 2100 * Found or allocated entry is returned in pagedeppp. 2101 * This routine must be called with splbio interrupts blocked. 2102 */ 2103 static int 2104 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2105 struct mount *mp; 2106 struct buf *bp; 2107 ino_t ino; 2108 ufs_lbn_t lbn; 2109 int flags; 2110 struct pagedep **pagedeppp; 2111 { 2112 struct pagedep *pagedep; 2113 struct pagedep_hashhead *pagedephd; 2114 struct worklist *wk; 2115 struct ufsmount *ump; 2116 int ret; 2117 int i; 2118 2119 ump = VFSTOUFS(mp); 2120 LOCK_OWNED(ump); 2121 if (bp) { 2122 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2123 if (wk->wk_type == D_PAGEDEP) { 2124 *pagedeppp = WK_PAGEDEP(wk); 2125 return (1); 2126 } 2127 } 2128 } 2129 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2130 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2131 if (ret) { 2132 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2133 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2134 return (1); 2135 } 2136 if ((flags & DEPALLOC) == 0) 2137 return (0); 2138 FREE_LOCK(ump); 2139 pagedep = malloc(sizeof(struct pagedep), 2140 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2141 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2142 ACQUIRE_LOCK(ump); 2143 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2144 if (*pagedeppp) { 2145 /* 2146 * This should never happen since we only create pagedeps 2147 * with the vnode lock held. Could be an assert. 2148 */ 2149 WORKITEM_FREE(pagedep, D_PAGEDEP); 2150 return (ret); 2151 } 2152 pagedep->pd_ino = ino; 2153 pagedep->pd_lbn = lbn; 2154 LIST_INIT(&pagedep->pd_dirremhd); 2155 LIST_INIT(&pagedep->pd_pendinghd); 2156 for (i = 0; i < DAHASHSZ; i++) 2157 LIST_INIT(&pagedep->pd_diraddhd[i]); 2158 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2159 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2160 *pagedeppp = pagedep; 2161 return (0); 2162 } 2163 2164 /* 2165 * Structures and routines associated with inodedep caching. 2166 */ 2167 #define INODEDEP_HASH(ump, inum) \ 2168 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2169 2170 static int 2171 inodedep_find(inodedephd, inum, inodedeppp) 2172 struct inodedep_hashhead *inodedephd; 2173 ino_t inum; 2174 struct inodedep **inodedeppp; 2175 { 2176 struct inodedep *inodedep; 2177 2178 LIST_FOREACH(inodedep, inodedephd, id_hash) 2179 if (inum == inodedep->id_ino) 2180 break; 2181 if (inodedep) { 2182 *inodedeppp = inodedep; 2183 return (1); 2184 } 2185 *inodedeppp = NULL; 2186 2187 return (0); 2188 } 2189 /* 2190 * Look up an inodedep. Return 1 if found, 0 if not found. 2191 * If not found, allocate if DEPALLOC flag is passed. 2192 * Found or allocated entry is returned in inodedeppp. 2193 * This routine must be called with splbio interrupts blocked. 2194 */ 2195 static int 2196 inodedep_lookup(mp, inum, flags, inodedeppp) 2197 struct mount *mp; 2198 ino_t inum; 2199 int flags; 2200 struct inodedep **inodedeppp; 2201 { 2202 struct inodedep *inodedep; 2203 struct inodedep_hashhead *inodedephd; 2204 struct ufsmount *ump; 2205 struct fs *fs; 2206 2207 ump = VFSTOUFS(mp); 2208 LOCK_OWNED(ump); 2209 fs = ump->um_fs; 2210 inodedephd = INODEDEP_HASH(ump, inum); 2211 2212 if (inodedep_find(inodedephd, inum, inodedeppp)) 2213 return (1); 2214 if ((flags & DEPALLOC) == 0) 2215 return (0); 2216 /* 2217 * If the system is over its limit and our filesystem is 2218 * responsible for more than our share of that usage and 2219 * we are not in a rush, request some inodedep cleanup. 2220 */ 2221 if (softdep_excess_items(ump, D_INODEDEP)) 2222 schedule_cleanup(mp); 2223 else 2224 FREE_LOCK(ump); 2225 inodedep = malloc(sizeof(struct inodedep), 2226 M_INODEDEP, M_SOFTDEP_FLAGS); 2227 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2228 ACQUIRE_LOCK(ump); 2229 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2230 WORKITEM_FREE(inodedep, D_INODEDEP); 2231 return (1); 2232 } 2233 inodedep->id_fs = fs; 2234 inodedep->id_ino = inum; 2235 inodedep->id_state = ALLCOMPLETE; 2236 inodedep->id_nlinkdelta = 0; 2237 inodedep->id_savedino1 = NULL; 2238 inodedep->id_savedsize = -1; 2239 inodedep->id_savedextsize = -1; 2240 inodedep->id_savednlink = -1; 2241 inodedep->id_bmsafemap = NULL; 2242 inodedep->id_mkdiradd = NULL; 2243 LIST_INIT(&inodedep->id_dirremhd); 2244 LIST_INIT(&inodedep->id_pendinghd); 2245 LIST_INIT(&inodedep->id_inowait); 2246 LIST_INIT(&inodedep->id_bufwait); 2247 TAILQ_INIT(&inodedep->id_inoreflst); 2248 TAILQ_INIT(&inodedep->id_inoupdt); 2249 TAILQ_INIT(&inodedep->id_newinoupdt); 2250 TAILQ_INIT(&inodedep->id_extupdt); 2251 TAILQ_INIT(&inodedep->id_newextupdt); 2252 TAILQ_INIT(&inodedep->id_freeblklst); 2253 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2254 *inodedeppp = inodedep; 2255 return (0); 2256 } 2257 2258 /* 2259 * Structures and routines associated with newblk caching. 2260 */ 2261 #define NEWBLK_HASH(ump, inum) \ 2262 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2263 2264 static int 2265 newblk_find(newblkhd, newblkno, flags, newblkpp) 2266 struct newblk_hashhead *newblkhd; 2267 ufs2_daddr_t newblkno; 2268 int flags; 2269 struct newblk **newblkpp; 2270 { 2271 struct newblk *newblk; 2272 2273 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2274 if (newblkno != newblk->nb_newblkno) 2275 continue; 2276 /* 2277 * If we're creating a new dependency don't match those that 2278 * have already been converted to allocdirects. This is for 2279 * a frag extend. 2280 */ 2281 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2282 continue; 2283 break; 2284 } 2285 if (newblk) { 2286 *newblkpp = newblk; 2287 return (1); 2288 } 2289 *newblkpp = NULL; 2290 return (0); 2291 } 2292 2293 /* 2294 * Look up a newblk. Return 1 if found, 0 if not found. 2295 * If not found, allocate if DEPALLOC flag is passed. 2296 * Found or allocated entry is returned in newblkpp. 2297 */ 2298 static int 2299 newblk_lookup(mp, newblkno, flags, newblkpp) 2300 struct mount *mp; 2301 ufs2_daddr_t newblkno; 2302 int flags; 2303 struct newblk **newblkpp; 2304 { 2305 struct newblk *newblk; 2306 struct newblk_hashhead *newblkhd; 2307 struct ufsmount *ump; 2308 2309 ump = VFSTOUFS(mp); 2310 LOCK_OWNED(ump); 2311 newblkhd = NEWBLK_HASH(ump, newblkno); 2312 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2313 return (1); 2314 if ((flags & DEPALLOC) == 0) 2315 return (0); 2316 if (softdep_excess_items(ump, D_NEWBLK) || 2317 softdep_excess_items(ump, D_ALLOCDIRECT) || 2318 softdep_excess_items(ump, D_ALLOCINDIR)) 2319 schedule_cleanup(mp); 2320 else 2321 FREE_LOCK(ump); 2322 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2323 M_SOFTDEP_FLAGS | M_ZERO); 2324 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2325 ACQUIRE_LOCK(ump); 2326 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2327 WORKITEM_FREE(newblk, D_NEWBLK); 2328 return (1); 2329 } 2330 newblk->nb_freefrag = NULL; 2331 LIST_INIT(&newblk->nb_indirdeps); 2332 LIST_INIT(&newblk->nb_newdirblk); 2333 LIST_INIT(&newblk->nb_jwork); 2334 newblk->nb_state = ATTACHED; 2335 newblk->nb_newblkno = newblkno; 2336 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2337 *newblkpp = newblk; 2338 return (0); 2339 } 2340 2341 /* 2342 * Structures and routines associated with freed indirect block caching. 2343 */ 2344 #define INDIR_HASH(ump, blkno) \ 2345 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2346 2347 /* 2348 * Lookup an indirect block in the indir hash table. The freework is 2349 * removed and potentially freed. The caller must do a blocking journal 2350 * write before writing to the blkno. 2351 */ 2352 static int 2353 indirblk_lookup(mp, blkno) 2354 struct mount *mp; 2355 ufs2_daddr_t blkno; 2356 { 2357 struct freework *freework; 2358 struct indir_hashhead *wkhd; 2359 struct ufsmount *ump; 2360 2361 ump = VFSTOUFS(mp); 2362 wkhd = INDIR_HASH(ump, blkno); 2363 TAILQ_FOREACH(freework, wkhd, fw_next) { 2364 if (freework->fw_blkno != blkno) 2365 continue; 2366 indirblk_remove(freework); 2367 return (1); 2368 } 2369 return (0); 2370 } 2371 2372 /* 2373 * Insert an indirect block represented by freework into the indirblk 2374 * hash table so that it may prevent the block from being re-used prior 2375 * to the journal being written. 2376 */ 2377 static void 2378 indirblk_insert(freework) 2379 struct freework *freework; 2380 { 2381 struct jblocks *jblocks; 2382 struct jseg *jseg; 2383 struct ufsmount *ump; 2384 2385 ump = VFSTOUFS(freework->fw_list.wk_mp); 2386 jblocks = ump->softdep_jblocks; 2387 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2388 if (jseg == NULL) 2389 return; 2390 2391 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2392 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2393 fw_next); 2394 freework->fw_state &= ~DEPCOMPLETE; 2395 } 2396 2397 static void 2398 indirblk_remove(freework) 2399 struct freework *freework; 2400 { 2401 struct ufsmount *ump; 2402 2403 ump = VFSTOUFS(freework->fw_list.wk_mp); 2404 LIST_REMOVE(freework, fw_segs); 2405 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2406 freework->fw_state |= DEPCOMPLETE; 2407 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2408 WORKITEM_FREE(freework, D_FREEWORK); 2409 } 2410 2411 /* 2412 * Executed during filesystem system initialization before 2413 * mounting any filesystems. 2414 */ 2415 void 2416 softdep_initialize() 2417 { 2418 2419 TAILQ_INIT(&softdepmounts); 2420 #ifdef __LP64__ 2421 max_softdeps = desiredvnodes * 4; 2422 #else 2423 max_softdeps = desiredvnodes * 2; 2424 #endif 2425 2426 /* initialise bioops hack */ 2427 bioops.io_start = softdep_disk_io_initiation; 2428 bioops.io_complete = softdep_disk_write_complete; 2429 bioops.io_deallocate = softdep_deallocate_dependencies; 2430 bioops.io_countdeps = softdep_count_dependencies; 2431 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2432 2433 /* Initialize the callout with an mtx. */ 2434 callout_init_mtx(&softdep_callout, &lk, 0); 2435 } 2436 2437 /* 2438 * Executed after all filesystems have been unmounted during 2439 * filesystem module unload. 2440 */ 2441 void 2442 softdep_uninitialize() 2443 { 2444 2445 /* clear bioops hack */ 2446 bioops.io_start = NULL; 2447 bioops.io_complete = NULL; 2448 bioops.io_deallocate = NULL; 2449 bioops.io_countdeps = NULL; 2450 softdep_ast_cleanup = NULL; 2451 2452 callout_drain(&softdep_callout); 2453 } 2454 2455 /* 2456 * Called at mount time to notify the dependency code that a 2457 * filesystem wishes to use it. 2458 */ 2459 int 2460 softdep_mount(devvp, mp, fs, cred) 2461 struct vnode *devvp; 2462 struct mount *mp; 2463 struct fs *fs; 2464 struct ucred *cred; 2465 { 2466 struct csum_total cstotal; 2467 struct mount_softdeps *sdp; 2468 struct ufsmount *ump; 2469 struct cg *cgp; 2470 struct buf *bp; 2471 u_int cyl, i; 2472 int error; 2473 2474 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2475 M_WAITOK | M_ZERO); 2476 MNT_ILOCK(mp); 2477 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2478 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2479 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2480 MNTK_SOFTDEP | MNTK_NOASYNC; 2481 } 2482 ump = VFSTOUFS(mp); 2483 ump->um_softdep = sdp; 2484 MNT_IUNLOCK(mp); 2485 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock"); 2486 sdp->sd_ump = ump; 2487 LIST_INIT(&ump->softdep_workitem_pending); 2488 LIST_INIT(&ump->softdep_journal_pending); 2489 TAILQ_INIT(&ump->softdep_unlinked); 2490 LIST_INIT(&ump->softdep_dirtycg); 2491 ump->softdep_worklist_tail = NULL; 2492 ump->softdep_on_worklist = 0; 2493 ump->softdep_deps = 0; 2494 LIST_INIT(&ump->softdep_mkdirlisthd); 2495 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2496 &ump->pagedep_hash_size); 2497 ump->pagedep_nextclean = 0; 2498 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2499 &ump->inodedep_hash_size); 2500 ump->inodedep_nextclean = 0; 2501 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2502 &ump->newblk_hash_size); 2503 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2504 &ump->bmsafemap_hash_size); 2505 i = 1 << (ffs(desiredvnodes / 10) - 1); 2506 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2507 M_FREEWORK, M_WAITOK); 2508 ump->indir_hash_size = i - 1; 2509 for (i = 0; i <= ump->indir_hash_size; i++) 2510 TAILQ_INIT(&ump->indir_hashtbl[i]); 2511 ACQUIRE_GBLLOCK(&lk); 2512 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2513 FREE_GBLLOCK(&lk); 2514 if ((fs->fs_flags & FS_SUJ) && 2515 (error = journal_mount(mp, fs, cred)) != 0) { 2516 printf("Failed to start journal: %d\n", error); 2517 softdep_unmount(mp); 2518 return (error); 2519 } 2520 /* 2521 * Start our flushing thread in the bufdaemon process. 2522 */ 2523 ACQUIRE_LOCK(ump); 2524 ump->softdep_flags |= FLUSH_STARTING; 2525 FREE_LOCK(ump); 2526 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2527 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2528 mp->mnt_stat.f_mntonname); 2529 ACQUIRE_LOCK(ump); 2530 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2531 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2532 hz / 2); 2533 } 2534 FREE_LOCK(ump); 2535 /* 2536 * When doing soft updates, the counters in the 2537 * superblock may have gotten out of sync. Recomputation 2538 * can take a long time and can be deferred for background 2539 * fsck. However, the old behavior of scanning the cylinder 2540 * groups and recalculating them at mount time is available 2541 * by setting vfs.ffs.compute_summary_at_mount to one. 2542 */ 2543 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2544 return (0); 2545 bzero(&cstotal, sizeof cstotal); 2546 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2547 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2548 fs->fs_cgsize, cred, &bp)) != 0) { 2549 brelse(bp); 2550 softdep_unmount(mp); 2551 return (error); 2552 } 2553 cgp = (struct cg *)bp->b_data; 2554 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2555 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2556 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2557 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2558 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2559 brelse(bp); 2560 } 2561 #ifdef DEBUG 2562 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2563 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2564 #endif 2565 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2566 return (0); 2567 } 2568 2569 void 2570 softdep_unmount(mp) 2571 struct mount *mp; 2572 { 2573 struct ufsmount *ump; 2574 #ifdef INVARIANTS 2575 int i; 2576 #endif 2577 2578 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2579 ("softdep_unmount called on non-softdep filesystem")); 2580 ump = VFSTOUFS(mp); 2581 MNT_ILOCK(mp); 2582 mp->mnt_flag &= ~MNT_SOFTDEP; 2583 if (MOUNTEDSUJ(mp) == 0) { 2584 MNT_IUNLOCK(mp); 2585 } else { 2586 mp->mnt_flag &= ~MNT_SUJ; 2587 MNT_IUNLOCK(mp); 2588 journal_unmount(ump); 2589 } 2590 /* 2591 * Shut down our flushing thread. Check for NULL is if 2592 * softdep_mount errors out before the thread has been created. 2593 */ 2594 if (ump->softdep_flushtd != NULL) { 2595 ACQUIRE_LOCK(ump); 2596 ump->softdep_flags |= FLUSH_EXIT; 2597 wakeup(&ump->softdep_flushtd); 2598 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2599 "sdwait", 0); 2600 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2601 ("Thread shutdown failed")); 2602 } 2603 /* 2604 * Free up our resources. 2605 */ 2606 ACQUIRE_GBLLOCK(&lk); 2607 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2608 FREE_GBLLOCK(&lk); 2609 rw_destroy(LOCK_PTR(ump)); 2610 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2611 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2612 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2613 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2614 ump->bmsafemap_hash_size); 2615 free(ump->indir_hashtbl, M_FREEWORK); 2616 #ifdef INVARIANTS 2617 for (i = 0; i <= D_LAST; i++) 2618 KASSERT(ump->softdep_curdeps[i] == 0, 2619 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2620 TYPENAME(i), ump->softdep_curdeps[i])); 2621 #endif 2622 free(ump->um_softdep, M_MOUNTDATA); 2623 } 2624 2625 static struct jblocks * 2626 jblocks_create(void) 2627 { 2628 struct jblocks *jblocks; 2629 2630 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2631 TAILQ_INIT(&jblocks->jb_segs); 2632 jblocks->jb_avail = 10; 2633 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2634 M_JBLOCKS, M_WAITOK | M_ZERO); 2635 2636 return (jblocks); 2637 } 2638 2639 static ufs2_daddr_t 2640 jblocks_alloc(jblocks, bytes, actual) 2641 struct jblocks *jblocks; 2642 int bytes; 2643 int *actual; 2644 { 2645 ufs2_daddr_t daddr; 2646 struct jextent *jext; 2647 int freecnt; 2648 int blocks; 2649 2650 blocks = bytes / DEV_BSIZE; 2651 jext = &jblocks->jb_extent[jblocks->jb_head]; 2652 freecnt = jext->je_blocks - jblocks->jb_off; 2653 if (freecnt == 0) { 2654 jblocks->jb_off = 0; 2655 if (++jblocks->jb_head > jblocks->jb_used) 2656 jblocks->jb_head = 0; 2657 jext = &jblocks->jb_extent[jblocks->jb_head]; 2658 freecnt = jext->je_blocks; 2659 } 2660 if (freecnt > blocks) 2661 freecnt = blocks; 2662 *actual = freecnt * DEV_BSIZE; 2663 daddr = jext->je_daddr + jblocks->jb_off; 2664 jblocks->jb_off += freecnt; 2665 jblocks->jb_free -= freecnt; 2666 2667 return (daddr); 2668 } 2669 2670 static void 2671 jblocks_free(jblocks, mp, bytes) 2672 struct jblocks *jblocks; 2673 struct mount *mp; 2674 int bytes; 2675 { 2676 2677 LOCK_OWNED(VFSTOUFS(mp)); 2678 jblocks->jb_free += bytes / DEV_BSIZE; 2679 if (jblocks->jb_suspended) 2680 worklist_speedup(mp); 2681 wakeup(jblocks); 2682 } 2683 2684 static void 2685 jblocks_destroy(jblocks) 2686 struct jblocks *jblocks; 2687 { 2688 2689 if (jblocks->jb_extent) 2690 free(jblocks->jb_extent, M_JBLOCKS); 2691 free(jblocks, M_JBLOCKS); 2692 } 2693 2694 static void 2695 jblocks_add(jblocks, daddr, blocks) 2696 struct jblocks *jblocks; 2697 ufs2_daddr_t daddr; 2698 int blocks; 2699 { 2700 struct jextent *jext; 2701 2702 jblocks->jb_blocks += blocks; 2703 jblocks->jb_free += blocks; 2704 jext = &jblocks->jb_extent[jblocks->jb_used]; 2705 /* Adding the first block. */ 2706 if (jext->je_daddr == 0) { 2707 jext->je_daddr = daddr; 2708 jext->je_blocks = blocks; 2709 return; 2710 } 2711 /* Extending the last extent. */ 2712 if (jext->je_daddr + jext->je_blocks == daddr) { 2713 jext->je_blocks += blocks; 2714 return; 2715 } 2716 /* Adding a new extent. */ 2717 if (++jblocks->jb_used == jblocks->jb_avail) { 2718 jblocks->jb_avail *= 2; 2719 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2720 M_JBLOCKS, M_WAITOK | M_ZERO); 2721 memcpy(jext, jblocks->jb_extent, 2722 sizeof(struct jextent) * jblocks->jb_used); 2723 free(jblocks->jb_extent, M_JBLOCKS); 2724 jblocks->jb_extent = jext; 2725 } 2726 jext = &jblocks->jb_extent[jblocks->jb_used]; 2727 jext->je_daddr = daddr; 2728 jext->je_blocks = blocks; 2729 return; 2730 } 2731 2732 int 2733 softdep_journal_lookup(mp, vpp) 2734 struct mount *mp; 2735 struct vnode **vpp; 2736 { 2737 struct componentname cnp; 2738 struct vnode *dvp; 2739 ino_t sujournal; 2740 int error; 2741 2742 error = VFS_VGET(mp, UFS_ROOTINO, LK_EXCLUSIVE, &dvp); 2743 if (error) 2744 return (error); 2745 bzero(&cnp, sizeof(cnp)); 2746 cnp.cn_nameiop = LOOKUP; 2747 cnp.cn_flags = ISLASTCN; 2748 cnp.cn_thread = curthread; 2749 cnp.cn_cred = curthread->td_ucred; 2750 cnp.cn_pnbuf = SUJ_FILE; 2751 cnp.cn_nameptr = SUJ_FILE; 2752 cnp.cn_namelen = strlen(SUJ_FILE); 2753 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2754 vput(dvp); 2755 if (error != 0) 2756 return (error); 2757 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2758 return (error); 2759 } 2760 2761 /* 2762 * Open and verify the journal file. 2763 */ 2764 static int 2765 journal_mount(mp, fs, cred) 2766 struct mount *mp; 2767 struct fs *fs; 2768 struct ucred *cred; 2769 { 2770 struct jblocks *jblocks; 2771 struct ufsmount *ump; 2772 struct vnode *vp; 2773 struct inode *ip; 2774 ufs2_daddr_t blkno; 2775 int bcount; 2776 int error; 2777 int i; 2778 2779 ump = VFSTOUFS(mp); 2780 ump->softdep_journal_tail = NULL; 2781 ump->softdep_on_journal = 0; 2782 ump->softdep_accdeps = 0; 2783 ump->softdep_req = 0; 2784 ump->softdep_jblocks = NULL; 2785 error = softdep_journal_lookup(mp, &vp); 2786 if (error != 0) { 2787 printf("Failed to find journal. Use tunefs to create one\n"); 2788 return (error); 2789 } 2790 ip = VTOI(vp); 2791 if (ip->i_size < SUJ_MIN) { 2792 error = ENOSPC; 2793 goto out; 2794 } 2795 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2796 jblocks = jblocks_create(); 2797 for (i = 0; i < bcount; i++) { 2798 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2799 if (error) 2800 break; 2801 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2802 } 2803 if (error) { 2804 jblocks_destroy(jblocks); 2805 goto out; 2806 } 2807 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2808 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2809 ump->softdep_jblocks = jblocks; 2810 out: 2811 if (error == 0) { 2812 MNT_ILOCK(mp); 2813 mp->mnt_flag |= MNT_SUJ; 2814 mp->mnt_flag &= ~MNT_SOFTDEP; 2815 MNT_IUNLOCK(mp); 2816 /* 2817 * Only validate the journal contents if the 2818 * filesystem is clean, otherwise we write the logs 2819 * but they'll never be used. If the filesystem was 2820 * still dirty when we mounted it the journal is 2821 * invalid and a new journal can only be valid if it 2822 * starts from a clean mount. 2823 */ 2824 if (fs->fs_clean) { 2825 DIP_SET(ip, i_modrev, fs->fs_mtime); 2826 ip->i_flags |= IN_MODIFIED; 2827 ffs_update(vp, 1); 2828 } 2829 } 2830 vput(vp); 2831 return (error); 2832 } 2833 2834 static void 2835 journal_unmount(ump) 2836 struct ufsmount *ump; 2837 { 2838 2839 if (ump->softdep_jblocks) 2840 jblocks_destroy(ump->softdep_jblocks); 2841 ump->softdep_jblocks = NULL; 2842 } 2843 2844 /* 2845 * Called when a journal record is ready to be written. Space is allocated 2846 * and the journal entry is created when the journal is flushed to stable 2847 * store. 2848 */ 2849 static void 2850 add_to_journal(wk) 2851 struct worklist *wk; 2852 { 2853 struct ufsmount *ump; 2854 2855 ump = VFSTOUFS(wk->wk_mp); 2856 LOCK_OWNED(ump); 2857 if (wk->wk_state & ONWORKLIST) 2858 panic("add_to_journal: %s(0x%X) already on list", 2859 TYPENAME(wk->wk_type), wk->wk_state); 2860 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2861 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2862 ump->softdep_jblocks->jb_age = ticks; 2863 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2864 } else 2865 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2866 ump->softdep_journal_tail = wk; 2867 ump->softdep_on_journal += 1; 2868 } 2869 2870 /* 2871 * Remove an arbitrary item for the journal worklist maintain the tail 2872 * pointer. This happens when a new operation obviates the need to 2873 * journal an old operation. 2874 */ 2875 static void 2876 remove_from_journal(wk) 2877 struct worklist *wk; 2878 { 2879 struct ufsmount *ump; 2880 2881 ump = VFSTOUFS(wk->wk_mp); 2882 LOCK_OWNED(ump); 2883 #ifdef SUJ_DEBUG 2884 { 2885 struct worklist *wkn; 2886 2887 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2888 if (wkn == wk) 2889 break; 2890 if (wkn == NULL) 2891 panic("remove_from_journal: %p is not in journal", wk); 2892 } 2893 #endif 2894 /* 2895 * We emulate a TAILQ to save space in most structures which do not 2896 * require TAILQ semantics. Here we must update the tail position 2897 * when removing the tail which is not the final entry. This works 2898 * only if the worklist linkage are at the beginning of the structure. 2899 */ 2900 if (ump->softdep_journal_tail == wk) 2901 ump->softdep_journal_tail = 2902 (struct worklist *)wk->wk_list.le_prev; 2903 WORKLIST_REMOVE(wk); 2904 ump->softdep_on_journal -= 1; 2905 } 2906 2907 /* 2908 * Check for journal space as well as dependency limits so the prelink 2909 * code can throttle both journaled and non-journaled filesystems. 2910 * Threshold is 0 for low and 1 for min. 2911 */ 2912 static int 2913 journal_space(ump, thresh) 2914 struct ufsmount *ump; 2915 int thresh; 2916 { 2917 struct jblocks *jblocks; 2918 int limit, avail; 2919 2920 jblocks = ump->softdep_jblocks; 2921 if (jblocks == NULL) 2922 return (1); 2923 /* 2924 * We use a tighter restriction here to prevent request_cleanup() 2925 * running in threads from running into locks we currently hold. 2926 * We have to be over the limit and our filesystem has to be 2927 * responsible for more than our share of that usage. 2928 */ 2929 limit = (max_softdeps / 10) * 9; 2930 if (dep_current[D_INODEDEP] > limit && 2931 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2932 return (0); 2933 if (thresh) 2934 thresh = jblocks->jb_min; 2935 else 2936 thresh = jblocks->jb_low; 2937 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2938 avail = jblocks->jb_free - avail; 2939 2940 return (avail > thresh); 2941 } 2942 2943 static void 2944 journal_suspend(ump) 2945 struct ufsmount *ump; 2946 { 2947 struct jblocks *jblocks; 2948 struct mount *mp; 2949 2950 mp = UFSTOVFS(ump); 2951 jblocks = ump->softdep_jblocks; 2952 MNT_ILOCK(mp); 2953 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2954 stat_journal_min++; 2955 mp->mnt_kern_flag |= MNTK_SUSPEND; 2956 mp->mnt_susp_owner = ump->softdep_flushtd; 2957 } 2958 jblocks->jb_suspended = 1; 2959 MNT_IUNLOCK(mp); 2960 } 2961 2962 static int 2963 journal_unsuspend(struct ufsmount *ump) 2964 { 2965 struct jblocks *jblocks; 2966 struct mount *mp; 2967 2968 mp = UFSTOVFS(ump); 2969 jblocks = ump->softdep_jblocks; 2970 2971 if (jblocks != NULL && jblocks->jb_suspended && 2972 journal_space(ump, jblocks->jb_min)) { 2973 jblocks->jb_suspended = 0; 2974 FREE_LOCK(ump); 2975 mp->mnt_susp_owner = curthread; 2976 vfs_write_resume(mp, 0); 2977 ACQUIRE_LOCK(ump); 2978 return (1); 2979 } 2980 return (0); 2981 } 2982 2983 /* 2984 * Called before any allocation function to be certain that there is 2985 * sufficient space in the journal prior to creating any new records. 2986 * Since in the case of block allocation we may have multiple locked 2987 * buffers at the time of the actual allocation we can not block 2988 * when the journal records are created. Doing so would create a deadlock 2989 * if any of these buffers needed to be flushed to reclaim space. Instead 2990 * we require a sufficiently large amount of available space such that 2991 * each thread in the system could have passed this allocation check and 2992 * still have sufficient free space. With 20% of a minimum journal size 2993 * of 1MB we have 6553 records available. 2994 */ 2995 int 2996 softdep_prealloc(vp, waitok) 2997 struct vnode *vp; 2998 int waitok; 2999 { 3000 struct ufsmount *ump; 3001 3002 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 3003 ("softdep_prealloc called on non-softdep filesystem")); 3004 /* 3005 * Nothing to do if we are not running journaled soft updates. 3006 * If we currently hold the snapshot lock, we must avoid 3007 * handling other resources that could cause deadlock. Do not 3008 * touch quotas vnode since it is typically recursed with 3009 * other vnode locks held. 3010 */ 3011 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) || 3012 (vp->v_vflag & VV_SYSTEM) != 0) 3013 return (0); 3014 ump = VFSTOUFS(vp->v_mount); 3015 ACQUIRE_LOCK(ump); 3016 if (journal_space(ump, 0)) { 3017 FREE_LOCK(ump); 3018 return (0); 3019 } 3020 stat_journal_low++; 3021 FREE_LOCK(ump); 3022 if (waitok == MNT_NOWAIT) 3023 return (ENOSPC); 3024 /* 3025 * Attempt to sync this vnode once to flush any journal 3026 * work attached to it. 3027 */ 3028 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3029 ffs_syncvnode(vp, waitok, 0); 3030 ACQUIRE_LOCK(ump); 3031 process_removes(vp); 3032 process_truncates(vp); 3033 if (journal_space(ump, 0) == 0) { 3034 softdep_speedup(ump); 3035 if (journal_space(ump, 1) == 0) 3036 journal_suspend(ump); 3037 } 3038 FREE_LOCK(ump); 3039 3040 return (0); 3041 } 3042 3043 /* 3044 * Before adjusting a link count on a vnode verify that we have sufficient 3045 * journal space. If not, process operations that depend on the currently 3046 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3047 * and softdep flush threads can not acquire these locks to reclaim space. 3048 */ 3049 static void 3050 softdep_prelink(dvp, vp) 3051 struct vnode *dvp; 3052 struct vnode *vp; 3053 { 3054 struct ufsmount *ump; 3055 3056 ump = VFSTOUFS(dvp->v_mount); 3057 LOCK_OWNED(ump); 3058 /* 3059 * Nothing to do if we have sufficient journal space. 3060 * If we currently hold the snapshot lock, we must avoid 3061 * handling other resources that could cause deadlock. 3062 */ 3063 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3064 return; 3065 stat_journal_low++; 3066 FREE_LOCK(ump); 3067 if (vp) 3068 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3069 ffs_syncvnode(dvp, MNT_WAIT, 0); 3070 ACQUIRE_LOCK(ump); 3071 /* Process vp before dvp as it may create .. removes. */ 3072 if (vp) { 3073 process_removes(vp); 3074 process_truncates(vp); 3075 } 3076 process_removes(dvp); 3077 process_truncates(dvp); 3078 softdep_speedup(ump); 3079 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3080 if (journal_space(ump, 0) == 0) { 3081 softdep_speedup(ump); 3082 if (journal_space(ump, 1) == 0) 3083 journal_suspend(ump); 3084 } 3085 } 3086 3087 static void 3088 jseg_write(ump, jseg, data) 3089 struct ufsmount *ump; 3090 struct jseg *jseg; 3091 uint8_t *data; 3092 { 3093 struct jsegrec *rec; 3094 3095 rec = (struct jsegrec *)data; 3096 rec->jsr_seq = jseg->js_seq; 3097 rec->jsr_oldest = jseg->js_oldseq; 3098 rec->jsr_cnt = jseg->js_cnt; 3099 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3100 rec->jsr_crc = 0; 3101 rec->jsr_time = ump->um_fs->fs_mtime; 3102 } 3103 3104 static inline void 3105 inoref_write(inoref, jseg, rec) 3106 struct inoref *inoref; 3107 struct jseg *jseg; 3108 struct jrefrec *rec; 3109 { 3110 3111 inoref->if_jsegdep->jd_seg = jseg; 3112 rec->jr_ino = inoref->if_ino; 3113 rec->jr_parent = inoref->if_parent; 3114 rec->jr_nlink = inoref->if_nlink; 3115 rec->jr_mode = inoref->if_mode; 3116 rec->jr_diroff = inoref->if_diroff; 3117 } 3118 3119 static void 3120 jaddref_write(jaddref, jseg, data) 3121 struct jaddref *jaddref; 3122 struct jseg *jseg; 3123 uint8_t *data; 3124 { 3125 struct jrefrec *rec; 3126 3127 rec = (struct jrefrec *)data; 3128 rec->jr_op = JOP_ADDREF; 3129 inoref_write(&jaddref->ja_ref, jseg, rec); 3130 } 3131 3132 static void 3133 jremref_write(jremref, jseg, data) 3134 struct jremref *jremref; 3135 struct jseg *jseg; 3136 uint8_t *data; 3137 { 3138 struct jrefrec *rec; 3139 3140 rec = (struct jrefrec *)data; 3141 rec->jr_op = JOP_REMREF; 3142 inoref_write(&jremref->jr_ref, jseg, rec); 3143 } 3144 3145 static void 3146 jmvref_write(jmvref, jseg, data) 3147 struct jmvref *jmvref; 3148 struct jseg *jseg; 3149 uint8_t *data; 3150 { 3151 struct jmvrec *rec; 3152 3153 rec = (struct jmvrec *)data; 3154 rec->jm_op = JOP_MVREF; 3155 rec->jm_ino = jmvref->jm_ino; 3156 rec->jm_parent = jmvref->jm_parent; 3157 rec->jm_oldoff = jmvref->jm_oldoff; 3158 rec->jm_newoff = jmvref->jm_newoff; 3159 } 3160 3161 static void 3162 jnewblk_write(jnewblk, jseg, data) 3163 struct jnewblk *jnewblk; 3164 struct jseg *jseg; 3165 uint8_t *data; 3166 { 3167 struct jblkrec *rec; 3168 3169 jnewblk->jn_jsegdep->jd_seg = jseg; 3170 rec = (struct jblkrec *)data; 3171 rec->jb_op = JOP_NEWBLK; 3172 rec->jb_ino = jnewblk->jn_ino; 3173 rec->jb_blkno = jnewblk->jn_blkno; 3174 rec->jb_lbn = jnewblk->jn_lbn; 3175 rec->jb_frags = jnewblk->jn_frags; 3176 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3177 } 3178 3179 static void 3180 jfreeblk_write(jfreeblk, jseg, data) 3181 struct jfreeblk *jfreeblk; 3182 struct jseg *jseg; 3183 uint8_t *data; 3184 { 3185 struct jblkrec *rec; 3186 3187 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3188 rec = (struct jblkrec *)data; 3189 rec->jb_op = JOP_FREEBLK; 3190 rec->jb_ino = jfreeblk->jf_ino; 3191 rec->jb_blkno = jfreeblk->jf_blkno; 3192 rec->jb_lbn = jfreeblk->jf_lbn; 3193 rec->jb_frags = jfreeblk->jf_frags; 3194 rec->jb_oldfrags = 0; 3195 } 3196 3197 static void 3198 jfreefrag_write(jfreefrag, jseg, data) 3199 struct jfreefrag *jfreefrag; 3200 struct jseg *jseg; 3201 uint8_t *data; 3202 { 3203 struct jblkrec *rec; 3204 3205 jfreefrag->fr_jsegdep->jd_seg = jseg; 3206 rec = (struct jblkrec *)data; 3207 rec->jb_op = JOP_FREEBLK; 3208 rec->jb_ino = jfreefrag->fr_ino; 3209 rec->jb_blkno = jfreefrag->fr_blkno; 3210 rec->jb_lbn = jfreefrag->fr_lbn; 3211 rec->jb_frags = jfreefrag->fr_frags; 3212 rec->jb_oldfrags = 0; 3213 } 3214 3215 static void 3216 jtrunc_write(jtrunc, jseg, data) 3217 struct jtrunc *jtrunc; 3218 struct jseg *jseg; 3219 uint8_t *data; 3220 { 3221 struct jtrncrec *rec; 3222 3223 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3224 rec = (struct jtrncrec *)data; 3225 rec->jt_op = JOP_TRUNC; 3226 rec->jt_ino = jtrunc->jt_ino; 3227 rec->jt_size = jtrunc->jt_size; 3228 rec->jt_extsize = jtrunc->jt_extsize; 3229 } 3230 3231 static void 3232 jfsync_write(jfsync, jseg, data) 3233 struct jfsync *jfsync; 3234 struct jseg *jseg; 3235 uint8_t *data; 3236 { 3237 struct jtrncrec *rec; 3238 3239 rec = (struct jtrncrec *)data; 3240 rec->jt_op = JOP_SYNC; 3241 rec->jt_ino = jfsync->jfs_ino; 3242 rec->jt_size = jfsync->jfs_size; 3243 rec->jt_extsize = jfsync->jfs_extsize; 3244 } 3245 3246 static void 3247 softdep_flushjournal(mp) 3248 struct mount *mp; 3249 { 3250 struct jblocks *jblocks; 3251 struct ufsmount *ump; 3252 3253 if (MOUNTEDSUJ(mp) == 0) 3254 return; 3255 ump = VFSTOUFS(mp); 3256 jblocks = ump->softdep_jblocks; 3257 ACQUIRE_LOCK(ump); 3258 while (ump->softdep_on_journal) { 3259 jblocks->jb_needseg = 1; 3260 softdep_process_journal(mp, NULL, MNT_WAIT); 3261 } 3262 FREE_LOCK(ump); 3263 } 3264 3265 static void softdep_synchronize_completed(struct bio *); 3266 static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3267 3268 static void 3269 softdep_synchronize_completed(bp) 3270 struct bio *bp; 3271 { 3272 struct jseg *oldest; 3273 struct jseg *jseg; 3274 struct ufsmount *ump; 3275 3276 /* 3277 * caller1 marks the last segment written before we issued the 3278 * synchronize cache. 3279 */ 3280 jseg = bp->bio_caller1; 3281 if (jseg == NULL) { 3282 g_destroy_bio(bp); 3283 return; 3284 } 3285 ump = VFSTOUFS(jseg->js_list.wk_mp); 3286 ACQUIRE_LOCK(ump); 3287 oldest = NULL; 3288 /* 3289 * Mark all the journal entries waiting on the synchronize cache 3290 * as completed so they may continue on. 3291 */ 3292 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3293 jseg->js_state |= COMPLETE; 3294 oldest = jseg; 3295 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3296 } 3297 /* 3298 * Restart deferred journal entry processing from the oldest 3299 * completed jseg. 3300 */ 3301 if (oldest) 3302 complete_jsegs(oldest); 3303 3304 FREE_LOCK(ump); 3305 g_destroy_bio(bp); 3306 } 3307 3308 /* 3309 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3310 * barriers. The journal must be written prior to any blocks that depend 3311 * on it and the journal can not be released until the blocks have be 3312 * written. This code handles both barriers simultaneously. 3313 */ 3314 static void 3315 softdep_synchronize(bp, ump, caller1) 3316 struct bio *bp; 3317 struct ufsmount *ump; 3318 void *caller1; 3319 { 3320 3321 bp->bio_cmd = BIO_FLUSH; 3322 bp->bio_flags |= BIO_ORDERED; 3323 bp->bio_data = NULL; 3324 bp->bio_offset = ump->um_cp->provider->mediasize; 3325 bp->bio_length = 0; 3326 bp->bio_done = softdep_synchronize_completed; 3327 bp->bio_caller1 = caller1; 3328 g_io_request(bp, 3329 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3330 } 3331 3332 /* 3333 * Flush some journal records to disk. 3334 */ 3335 static void 3336 softdep_process_journal(mp, needwk, flags) 3337 struct mount *mp; 3338 struct worklist *needwk; 3339 int flags; 3340 { 3341 struct jblocks *jblocks; 3342 struct ufsmount *ump; 3343 struct worklist *wk; 3344 struct jseg *jseg; 3345 struct buf *bp; 3346 struct bio *bio; 3347 uint8_t *data; 3348 struct fs *fs; 3349 int shouldflush; 3350 int segwritten; 3351 int jrecmin; /* Minimum records per block. */ 3352 int jrecmax; /* Maximum records per block. */ 3353 int size; 3354 int cnt; 3355 int off; 3356 int devbsize; 3357 3358 if (MOUNTEDSUJ(mp) == 0) 3359 return; 3360 shouldflush = softdep_flushcache; 3361 bio = NULL; 3362 jseg = NULL; 3363 ump = VFSTOUFS(mp); 3364 LOCK_OWNED(ump); 3365 fs = ump->um_fs; 3366 jblocks = ump->softdep_jblocks; 3367 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3368 /* 3369 * We write anywhere between a disk block and fs block. The upper 3370 * bound is picked to prevent buffer cache fragmentation and limit 3371 * processing time per I/O. 3372 */ 3373 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3374 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3375 segwritten = 0; 3376 for (;;) { 3377 cnt = ump->softdep_on_journal; 3378 /* 3379 * Criteria for writing a segment: 3380 * 1) We have a full block. 3381 * 2) We're called from jwait() and haven't found the 3382 * journal item yet. 3383 * 3) Always write if needseg is set. 3384 * 4) If we are called from process_worklist and have 3385 * not yet written anything we write a partial block 3386 * to enforce a 1 second maximum latency on journal 3387 * entries. 3388 */ 3389 if (cnt < (jrecmax - 1) && needwk == NULL && 3390 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3391 break; 3392 cnt++; 3393 /* 3394 * Verify some free journal space. softdep_prealloc() should 3395 * guarantee that we don't run out so this is indicative of 3396 * a problem with the flow control. Try to recover 3397 * gracefully in any event. 3398 */ 3399 while (jblocks->jb_free == 0) { 3400 if (flags != MNT_WAIT) 3401 break; 3402 printf("softdep: Out of journal space!\n"); 3403 softdep_speedup(ump); 3404 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3405 } 3406 FREE_LOCK(ump); 3407 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3408 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3409 LIST_INIT(&jseg->js_entries); 3410 LIST_INIT(&jseg->js_indirs); 3411 jseg->js_state = ATTACHED; 3412 if (shouldflush == 0) 3413 jseg->js_state |= COMPLETE; 3414 else if (bio == NULL) 3415 bio = g_alloc_bio(); 3416 jseg->js_jblocks = jblocks; 3417 bp = geteblk(fs->fs_bsize, 0); 3418 ACQUIRE_LOCK(ump); 3419 /* 3420 * If there was a race while we were allocating the block 3421 * and jseg the entry we care about was likely written. 3422 * We bail out in both the WAIT and NOWAIT case and assume 3423 * the caller will loop if the entry it cares about is 3424 * not written. 3425 */ 3426 cnt = ump->softdep_on_journal; 3427 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3428 bp->b_flags |= B_INVAL | B_NOCACHE; 3429 WORKITEM_FREE(jseg, D_JSEG); 3430 FREE_LOCK(ump); 3431 brelse(bp); 3432 ACQUIRE_LOCK(ump); 3433 break; 3434 } 3435 /* 3436 * Calculate the disk block size required for the available 3437 * records rounded to the min size. 3438 */ 3439 if (cnt == 0) 3440 size = devbsize; 3441 else if (cnt < jrecmax) 3442 size = howmany(cnt, jrecmin) * devbsize; 3443 else 3444 size = fs->fs_bsize; 3445 /* 3446 * Allocate a disk block for this journal data and account 3447 * for truncation of the requested size if enough contiguous 3448 * space was not available. 3449 */ 3450 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3451 bp->b_lblkno = bp->b_blkno; 3452 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3453 bp->b_bcount = size; 3454 bp->b_flags &= ~B_INVAL; 3455 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3456 /* 3457 * Initialize our jseg with cnt records. Assign the next 3458 * sequence number to it and link it in-order. 3459 */ 3460 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3461 jseg->js_buf = bp; 3462 jseg->js_cnt = cnt; 3463 jseg->js_refs = cnt + 1; /* Self ref. */ 3464 jseg->js_size = size; 3465 jseg->js_seq = jblocks->jb_nextseq++; 3466 if (jblocks->jb_oldestseg == NULL) 3467 jblocks->jb_oldestseg = jseg; 3468 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3469 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3470 if (jblocks->jb_writeseg == NULL) 3471 jblocks->jb_writeseg = jseg; 3472 /* 3473 * Start filling in records from the pending list. 3474 */ 3475 data = bp->b_data; 3476 off = 0; 3477 3478 /* 3479 * Always put a header on the first block. 3480 * XXX As with below, there might not be a chance to get 3481 * into the loop. Ensure that something valid is written. 3482 */ 3483 jseg_write(ump, jseg, data); 3484 off += JREC_SIZE; 3485 data = bp->b_data + off; 3486 3487 /* 3488 * XXX Something is wrong here. There's no work to do, 3489 * but we need to perform and I/O and allow it to complete 3490 * anyways. 3491 */ 3492 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3493 stat_emptyjblocks++; 3494 3495 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3496 != NULL) { 3497 if (cnt == 0) 3498 break; 3499 /* Place a segment header on every device block. */ 3500 if ((off % devbsize) == 0) { 3501 jseg_write(ump, jseg, data); 3502 off += JREC_SIZE; 3503 data = bp->b_data + off; 3504 } 3505 if (wk == needwk) 3506 needwk = NULL; 3507 remove_from_journal(wk); 3508 wk->wk_state |= INPROGRESS; 3509 WORKLIST_INSERT(&jseg->js_entries, wk); 3510 switch (wk->wk_type) { 3511 case D_JADDREF: 3512 jaddref_write(WK_JADDREF(wk), jseg, data); 3513 break; 3514 case D_JREMREF: 3515 jremref_write(WK_JREMREF(wk), jseg, data); 3516 break; 3517 case D_JMVREF: 3518 jmvref_write(WK_JMVREF(wk), jseg, data); 3519 break; 3520 case D_JNEWBLK: 3521 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3522 break; 3523 case D_JFREEBLK: 3524 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3525 break; 3526 case D_JFREEFRAG: 3527 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3528 break; 3529 case D_JTRUNC: 3530 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3531 break; 3532 case D_JFSYNC: 3533 jfsync_write(WK_JFSYNC(wk), jseg, data); 3534 break; 3535 default: 3536 panic("process_journal: Unknown type %s", 3537 TYPENAME(wk->wk_type)); 3538 /* NOTREACHED */ 3539 } 3540 off += JREC_SIZE; 3541 data = bp->b_data + off; 3542 cnt--; 3543 } 3544 3545 /* Clear any remaining space so we don't leak kernel data */ 3546 if (size > off) 3547 bzero(data, size - off); 3548 3549 /* 3550 * Write this one buffer and continue. 3551 */ 3552 segwritten = 1; 3553 jblocks->jb_needseg = 0; 3554 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3555 FREE_LOCK(ump); 3556 pbgetvp(ump->um_devvp, bp); 3557 /* 3558 * We only do the blocking wait once we find the journal 3559 * entry we're looking for. 3560 */ 3561 if (needwk == NULL && flags == MNT_WAIT) 3562 bwrite(bp); 3563 else 3564 bawrite(bp); 3565 ACQUIRE_LOCK(ump); 3566 } 3567 /* 3568 * If we wrote a segment issue a synchronize cache so the journal 3569 * is reflected on disk before the data is written. Since reclaiming 3570 * journal space also requires writing a journal record this 3571 * process also enforces a barrier before reclamation. 3572 */ 3573 if (segwritten && shouldflush) { 3574 softdep_synchronize(bio, ump, 3575 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3576 } else if (bio) 3577 g_destroy_bio(bio); 3578 /* 3579 * If we've suspended the filesystem because we ran out of journal 3580 * space either try to sync it here to make some progress or 3581 * unsuspend it if we already have. 3582 */ 3583 if (flags == 0 && jblocks->jb_suspended) { 3584 if (journal_unsuspend(ump)) 3585 return; 3586 FREE_LOCK(ump); 3587 VFS_SYNC(mp, MNT_NOWAIT); 3588 ffs_sbupdate(ump, MNT_WAIT, 0); 3589 ACQUIRE_LOCK(ump); 3590 } 3591 } 3592 3593 /* 3594 * Complete a jseg, allowing all dependencies awaiting journal writes 3595 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3596 * structures so that the journal segment can be freed to reclaim space. 3597 */ 3598 static void 3599 complete_jseg(jseg) 3600 struct jseg *jseg; 3601 { 3602 struct worklist *wk; 3603 struct jmvref *jmvref; 3604 #ifdef INVARIANTS 3605 int i = 0; 3606 #endif 3607 3608 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3609 WORKLIST_REMOVE(wk); 3610 wk->wk_state &= ~INPROGRESS; 3611 wk->wk_state |= COMPLETE; 3612 KASSERT(i++ < jseg->js_cnt, 3613 ("handle_written_jseg: overflow %d >= %d", 3614 i - 1, jseg->js_cnt)); 3615 switch (wk->wk_type) { 3616 case D_JADDREF: 3617 handle_written_jaddref(WK_JADDREF(wk)); 3618 break; 3619 case D_JREMREF: 3620 handle_written_jremref(WK_JREMREF(wk)); 3621 break; 3622 case D_JMVREF: 3623 rele_jseg(jseg); /* No jsegdep. */ 3624 jmvref = WK_JMVREF(wk); 3625 LIST_REMOVE(jmvref, jm_deps); 3626 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3627 free_pagedep(jmvref->jm_pagedep); 3628 WORKITEM_FREE(jmvref, D_JMVREF); 3629 break; 3630 case D_JNEWBLK: 3631 handle_written_jnewblk(WK_JNEWBLK(wk)); 3632 break; 3633 case D_JFREEBLK: 3634 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3635 break; 3636 case D_JTRUNC: 3637 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3638 break; 3639 case D_JFSYNC: 3640 rele_jseg(jseg); /* No jsegdep. */ 3641 WORKITEM_FREE(wk, D_JFSYNC); 3642 break; 3643 case D_JFREEFRAG: 3644 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3645 break; 3646 default: 3647 panic("handle_written_jseg: Unknown type %s", 3648 TYPENAME(wk->wk_type)); 3649 /* NOTREACHED */ 3650 } 3651 } 3652 /* Release the self reference so the structure may be freed. */ 3653 rele_jseg(jseg); 3654 } 3655 3656 /* 3657 * Determine which jsegs are ready for completion processing. Waits for 3658 * synchronize cache to complete as well as forcing in-order completion 3659 * of journal entries. 3660 */ 3661 static void 3662 complete_jsegs(jseg) 3663 struct jseg *jseg; 3664 { 3665 struct jblocks *jblocks; 3666 struct jseg *jsegn; 3667 3668 jblocks = jseg->js_jblocks; 3669 /* 3670 * Don't allow out of order completions. If this isn't the first 3671 * block wait for it to write before we're done. 3672 */ 3673 if (jseg != jblocks->jb_writeseg) 3674 return; 3675 /* Iterate through available jsegs processing their entries. */ 3676 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3677 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3678 jsegn = TAILQ_NEXT(jseg, js_next); 3679 complete_jseg(jseg); 3680 jseg = jsegn; 3681 } 3682 jblocks->jb_writeseg = jseg; 3683 /* 3684 * Attempt to free jsegs now that oldestwrseq may have advanced. 3685 */ 3686 free_jsegs(jblocks); 3687 } 3688 3689 /* 3690 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3691 * the final completions. 3692 */ 3693 static void 3694 handle_written_jseg(jseg, bp) 3695 struct jseg *jseg; 3696 struct buf *bp; 3697 { 3698 3699 if (jseg->js_refs == 0) 3700 panic("handle_written_jseg: No self-reference on %p", jseg); 3701 jseg->js_state |= DEPCOMPLETE; 3702 /* 3703 * We'll never need this buffer again, set flags so it will be 3704 * discarded. 3705 */ 3706 bp->b_flags |= B_INVAL | B_NOCACHE; 3707 pbrelvp(bp); 3708 complete_jsegs(jseg); 3709 } 3710 3711 static inline struct jsegdep * 3712 inoref_jseg(inoref) 3713 struct inoref *inoref; 3714 { 3715 struct jsegdep *jsegdep; 3716 3717 jsegdep = inoref->if_jsegdep; 3718 inoref->if_jsegdep = NULL; 3719 3720 return (jsegdep); 3721 } 3722 3723 /* 3724 * Called once a jremref has made it to stable store. The jremref is marked 3725 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3726 * for the jremref to complete will be awoken by free_jremref. 3727 */ 3728 static void 3729 handle_written_jremref(jremref) 3730 struct jremref *jremref; 3731 { 3732 struct inodedep *inodedep; 3733 struct jsegdep *jsegdep; 3734 struct dirrem *dirrem; 3735 3736 /* Grab the jsegdep. */ 3737 jsegdep = inoref_jseg(&jremref->jr_ref); 3738 /* 3739 * Remove us from the inoref list. 3740 */ 3741 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3742 0, &inodedep) == 0) 3743 panic("handle_written_jremref: Lost inodedep"); 3744 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3745 /* 3746 * Complete the dirrem. 3747 */ 3748 dirrem = jremref->jr_dirrem; 3749 jremref->jr_dirrem = NULL; 3750 LIST_REMOVE(jremref, jr_deps); 3751 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3752 jwork_insert(&dirrem->dm_jwork, jsegdep); 3753 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3754 (dirrem->dm_state & COMPLETE) != 0) 3755 add_to_worklist(&dirrem->dm_list, 0); 3756 free_jremref(jremref); 3757 } 3758 3759 /* 3760 * Called once a jaddref has made it to stable store. The dependency is 3761 * marked complete and any dependent structures are added to the inode 3762 * bufwait list to be completed as soon as it is written. If a bitmap write 3763 * depends on this entry we move the inode into the inodedephd of the 3764 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3765 */ 3766 static void 3767 handle_written_jaddref(jaddref) 3768 struct jaddref *jaddref; 3769 { 3770 struct jsegdep *jsegdep; 3771 struct inodedep *inodedep; 3772 struct diradd *diradd; 3773 struct mkdir *mkdir; 3774 3775 /* Grab the jsegdep. */ 3776 jsegdep = inoref_jseg(&jaddref->ja_ref); 3777 mkdir = NULL; 3778 diradd = NULL; 3779 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3780 0, &inodedep) == 0) 3781 panic("handle_written_jaddref: Lost inodedep."); 3782 if (jaddref->ja_diradd == NULL) 3783 panic("handle_written_jaddref: No dependency"); 3784 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3785 diradd = jaddref->ja_diradd; 3786 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3787 } else if (jaddref->ja_state & MKDIR_PARENT) { 3788 mkdir = jaddref->ja_mkdir; 3789 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3790 } else if (jaddref->ja_state & MKDIR_BODY) 3791 mkdir = jaddref->ja_mkdir; 3792 else 3793 panic("handle_written_jaddref: Unknown dependency %p", 3794 jaddref->ja_diradd); 3795 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3796 /* 3797 * Remove us from the inode list. 3798 */ 3799 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3800 /* 3801 * The mkdir may be waiting on the jaddref to clear before freeing. 3802 */ 3803 if (mkdir) { 3804 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3805 ("handle_written_jaddref: Incorrect type for mkdir %s", 3806 TYPENAME(mkdir->md_list.wk_type))); 3807 mkdir->md_jaddref = NULL; 3808 diradd = mkdir->md_diradd; 3809 mkdir->md_state |= DEPCOMPLETE; 3810 complete_mkdir(mkdir); 3811 } 3812 jwork_insert(&diradd->da_jwork, jsegdep); 3813 if (jaddref->ja_state & NEWBLOCK) { 3814 inodedep->id_state |= ONDEPLIST; 3815 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3816 inodedep, id_deps); 3817 } 3818 free_jaddref(jaddref); 3819 } 3820 3821 /* 3822 * Called once a jnewblk journal is written. The allocdirect or allocindir 3823 * is placed in the bmsafemap to await notification of a written bitmap. If 3824 * the operation was canceled we add the segdep to the appropriate 3825 * dependency to free the journal space once the canceling operation 3826 * completes. 3827 */ 3828 static void 3829 handle_written_jnewblk(jnewblk) 3830 struct jnewblk *jnewblk; 3831 { 3832 struct bmsafemap *bmsafemap; 3833 struct freefrag *freefrag; 3834 struct freework *freework; 3835 struct jsegdep *jsegdep; 3836 struct newblk *newblk; 3837 3838 /* Grab the jsegdep. */ 3839 jsegdep = jnewblk->jn_jsegdep; 3840 jnewblk->jn_jsegdep = NULL; 3841 if (jnewblk->jn_dep == NULL) 3842 panic("handle_written_jnewblk: No dependency for the segdep."); 3843 switch (jnewblk->jn_dep->wk_type) { 3844 case D_NEWBLK: 3845 case D_ALLOCDIRECT: 3846 case D_ALLOCINDIR: 3847 /* 3848 * Add the written block to the bmsafemap so it can 3849 * be notified when the bitmap is on disk. 3850 */ 3851 newblk = WK_NEWBLK(jnewblk->jn_dep); 3852 newblk->nb_jnewblk = NULL; 3853 if ((newblk->nb_state & GOINGAWAY) == 0) { 3854 bmsafemap = newblk->nb_bmsafemap; 3855 newblk->nb_state |= ONDEPLIST; 3856 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3857 nb_deps); 3858 } 3859 jwork_insert(&newblk->nb_jwork, jsegdep); 3860 break; 3861 case D_FREEFRAG: 3862 /* 3863 * A newblock being removed by a freefrag when replaced by 3864 * frag extension. 3865 */ 3866 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3867 freefrag->ff_jdep = NULL; 3868 jwork_insert(&freefrag->ff_jwork, jsegdep); 3869 break; 3870 case D_FREEWORK: 3871 /* 3872 * A direct block was removed by truncate. 3873 */ 3874 freework = WK_FREEWORK(jnewblk->jn_dep); 3875 freework->fw_jnewblk = NULL; 3876 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3877 break; 3878 default: 3879 panic("handle_written_jnewblk: Unknown type %d.", 3880 jnewblk->jn_dep->wk_type); 3881 } 3882 jnewblk->jn_dep = NULL; 3883 free_jnewblk(jnewblk); 3884 } 3885 3886 /* 3887 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3888 * an in-flight allocation that has not yet been committed. Divorce us 3889 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3890 * to the worklist. 3891 */ 3892 static void 3893 cancel_jfreefrag(jfreefrag) 3894 struct jfreefrag *jfreefrag; 3895 { 3896 struct freefrag *freefrag; 3897 3898 if (jfreefrag->fr_jsegdep) { 3899 free_jsegdep(jfreefrag->fr_jsegdep); 3900 jfreefrag->fr_jsegdep = NULL; 3901 } 3902 freefrag = jfreefrag->fr_freefrag; 3903 jfreefrag->fr_freefrag = NULL; 3904 free_jfreefrag(jfreefrag); 3905 freefrag->ff_state |= DEPCOMPLETE; 3906 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3907 } 3908 3909 /* 3910 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3911 */ 3912 static void 3913 free_jfreefrag(jfreefrag) 3914 struct jfreefrag *jfreefrag; 3915 { 3916 3917 if (jfreefrag->fr_state & INPROGRESS) 3918 WORKLIST_REMOVE(&jfreefrag->fr_list); 3919 else if (jfreefrag->fr_state & ONWORKLIST) 3920 remove_from_journal(&jfreefrag->fr_list); 3921 if (jfreefrag->fr_freefrag != NULL) 3922 panic("free_jfreefrag: Still attached to a freefrag."); 3923 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3924 } 3925 3926 /* 3927 * Called when the journal write for a jfreefrag completes. The parent 3928 * freefrag is added to the worklist if this completes its dependencies. 3929 */ 3930 static void 3931 handle_written_jfreefrag(jfreefrag) 3932 struct jfreefrag *jfreefrag; 3933 { 3934 struct jsegdep *jsegdep; 3935 struct freefrag *freefrag; 3936 3937 /* Grab the jsegdep. */ 3938 jsegdep = jfreefrag->fr_jsegdep; 3939 jfreefrag->fr_jsegdep = NULL; 3940 freefrag = jfreefrag->fr_freefrag; 3941 if (freefrag == NULL) 3942 panic("handle_written_jfreefrag: No freefrag."); 3943 freefrag->ff_state |= DEPCOMPLETE; 3944 freefrag->ff_jdep = NULL; 3945 jwork_insert(&freefrag->ff_jwork, jsegdep); 3946 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3947 add_to_worklist(&freefrag->ff_list, 0); 3948 jfreefrag->fr_freefrag = NULL; 3949 free_jfreefrag(jfreefrag); 3950 } 3951 3952 /* 3953 * Called when the journal write for a jfreeblk completes. The jfreeblk 3954 * is removed from the freeblks list of pending journal writes and the 3955 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3956 * have been reclaimed. 3957 */ 3958 static void 3959 handle_written_jblkdep(jblkdep) 3960 struct jblkdep *jblkdep; 3961 { 3962 struct freeblks *freeblks; 3963 struct jsegdep *jsegdep; 3964 3965 /* Grab the jsegdep. */ 3966 jsegdep = jblkdep->jb_jsegdep; 3967 jblkdep->jb_jsegdep = NULL; 3968 freeblks = jblkdep->jb_freeblks; 3969 LIST_REMOVE(jblkdep, jb_deps); 3970 jwork_insert(&freeblks->fb_jwork, jsegdep); 3971 /* 3972 * If the freeblks is all journaled, we can add it to the worklist. 3973 */ 3974 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3975 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3976 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3977 3978 free_jblkdep(jblkdep); 3979 } 3980 3981 static struct jsegdep * 3982 newjsegdep(struct worklist *wk) 3983 { 3984 struct jsegdep *jsegdep; 3985 3986 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3987 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3988 jsegdep->jd_seg = NULL; 3989 3990 return (jsegdep); 3991 } 3992 3993 static struct jmvref * 3994 newjmvref(dp, ino, oldoff, newoff) 3995 struct inode *dp; 3996 ino_t ino; 3997 off_t oldoff; 3998 off_t newoff; 3999 { 4000 struct jmvref *jmvref; 4001 4002 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 4003 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp)); 4004 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 4005 jmvref->jm_parent = dp->i_number; 4006 jmvref->jm_ino = ino; 4007 jmvref->jm_oldoff = oldoff; 4008 jmvref->jm_newoff = newoff; 4009 4010 return (jmvref); 4011 } 4012 4013 /* 4014 * Allocate a new jremref that tracks the removal of ip from dp with the 4015 * directory entry offset of diroff. Mark the entry as ATTACHED and 4016 * DEPCOMPLETE as we have all the information required for the journal write 4017 * and the directory has already been removed from the buffer. The caller 4018 * is responsible for linking the jremref into the pagedep and adding it 4019 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4020 * a DOTDOT addition so handle_workitem_remove() can properly assign 4021 * the jsegdep when we're done. 4022 */ 4023 static struct jremref * 4024 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4025 off_t diroff, nlink_t nlink) 4026 { 4027 struct jremref *jremref; 4028 4029 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4030 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp)); 4031 jremref->jr_state = ATTACHED; 4032 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4033 nlink, ip->i_mode); 4034 jremref->jr_dirrem = dirrem; 4035 4036 return (jremref); 4037 } 4038 4039 static inline void 4040 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4041 nlink_t nlink, uint16_t mode) 4042 { 4043 4044 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4045 inoref->if_diroff = diroff; 4046 inoref->if_ino = ino; 4047 inoref->if_parent = parent; 4048 inoref->if_nlink = nlink; 4049 inoref->if_mode = mode; 4050 } 4051 4052 /* 4053 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4054 * directory offset may not be known until later. The caller is responsible 4055 * adding the entry to the journal when this information is available. nlink 4056 * should be the link count prior to the addition and mode is only required 4057 * to have the correct FMT. 4058 */ 4059 static struct jaddref * 4060 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4061 uint16_t mode) 4062 { 4063 struct jaddref *jaddref; 4064 4065 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4066 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp)); 4067 jaddref->ja_state = ATTACHED; 4068 jaddref->ja_mkdir = NULL; 4069 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4070 4071 return (jaddref); 4072 } 4073 4074 /* 4075 * Create a new free dependency for a freework. The caller is responsible 4076 * for adjusting the reference count when it has the lock held. The freedep 4077 * will track an outstanding bitmap write that will ultimately clear the 4078 * freework to continue. 4079 */ 4080 static struct freedep * 4081 newfreedep(struct freework *freework) 4082 { 4083 struct freedep *freedep; 4084 4085 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4086 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4087 freedep->fd_freework = freework; 4088 4089 return (freedep); 4090 } 4091 4092 /* 4093 * Free a freedep structure once the buffer it is linked to is written. If 4094 * this is the last reference to the freework schedule it for completion. 4095 */ 4096 static void 4097 free_freedep(freedep) 4098 struct freedep *freedep; 4099 { 4100 struct freework *freework; 4101 4102 freework = freedep->fd_freework; 4103 freework->fw_freeblks->fb_cgwait--; 4104 if (--freework->fw_ref == 0) 4105 freework_enqueue(freework); 4106 WORKITEM_FREE(freedep, D_FREEDEP); 4107 } 4108 4109 /* 4110 * Allocate a new freework structure that may be a level in an indirect 4111 * when parent is not NULL or a top level block when it is. The top level 4112 * freework structures are allocated without the per-filesystem lock held 4113 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4114 */ 4115 static struct freework * 4116 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4117 struct ufsmount *ump; 4118 struct freeblks *freeblks; 4119 struct freework *parent; 4120 ufs_lbn_t lbn; 4121 ufs2_daddr_t nb; 4122 int frags; 4123 int off; 4124 int journal; 4125 { 4126 struct freework *freework; 4127 4128 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4129 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4130 freework->fw_state = ATTACHED; 4131 freework->fw_jnewblk = NULL; 4132 freework->fw_freeblks = freeblks; 4133 freework->fw_parent = parent; 4134 freework->fw_lbn = lbn; 4135 freework->fw_blkno = nb; 4136 freework->fw_frags = frags; 4137 freework->fw_indir = NULL; 4138 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || 4139 lbn >= -UFS_NXADDR) ? 0 : NINDIR(ump->um_fs) + 1; 4140 freework->fw_start = freework->fw_off = off; 4141 if (journal) 4142 newjfreeblk(freeblks, lbn, nb, frags); 4143 if (parent == NULL) { 4144 ACQUIRE_LOCK(ump); 4145 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4146 freeblks->fb_ref++; 4147 FREE_LOCK(ump); 4148 } 4149 4150 return (freework); 4151 } 4152 4153 /* 4154 * Eliminate a jfreeblk for a block that does not need journaling. 4155 */ 4156 static void 4157 cancel_jfreeblk(freeblks, blkno) 4158 struct freeblks *freeblks; 4159 ufs2_daddr_t blkno; 4160 { 4161 struct jfreeblk *jfreeblk; 4162 struct jblkdep *jblkdep; 4163 4164 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4165 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4166 continue; 4167 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4168 if (jfreeblk->jf_blkno == blkno) 4169 break; 4170 } 4171 if (jblkdep == NULL) 4172 return; 4173 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4174 free_jsegdep(jblkdep->jb_jsegdep); 4175 LIST_REMOVE(jblkdep, jb_deps); 4176 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4177 } 4178 4179 /* 4180 * Allocate a new jfreeblk to journal top level block pointer when truncating 4181 * a file. The caller must add this to the worklist when the per-filesystem 4182 * lock is held. 4183 */ 4184 static struct jfreeblk * 4185 newjfreeblk(freeblks, lbn, blkno, frags) 4186 struct freeblks *freeblks; 4187 ufs_lbn_t lbn; 4188 ufs2_daddr_t blkno; 4189 int frags; 4190 { 4191 struct jfreeblk *jfreeblk; 4192 4193 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4194 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4195 freeblks->fb_list.wk_mp); 4196 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4197 jfreeblk->jf_dep.jb_freeblks = freeblks; 4198 jfreeblk->jf_ino = freeblks->fb_inum; 4199 jfreeblk->jf_lbn = lbn; 4200 jfreeblk->jf_blkno = blkno; 4201 jfreeblk->jf_frags = frags; 4202 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4203 4204 return (jfreeblk); 4205 } 4206 4207 /* 4208 * The journal is only prepared to handle full-size block numbers, so we 4209 * have to adjust the record to reflect the change to a full-size block. 4210 * For example, suppose we have a block made up of fragments 8-15 and 4211 * want to free its last two fragments. We are given a request that says: 4212 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4213 * where frags are the number of fragments to free and oldfrags are the 4214 * number of fragments to keep. To block align it, we have to change it to 4215 * have a valid full-size blkno, so it becomes: 4216 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4217 */ 4218 static void 4219 adjust_newfreework(freeblks, frag_offset) 4220 struct freeblks *freeblks; 4221 int frag_offset; 4222 { 4223 struct jfreeblk *jfreeblk; 4224 4225 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4226 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4227 ("adjust_newfreework: Missing freeblks dependency")); 4228 4229 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4230 jfreeblk->jf_blkno -= frag_offset; 4231 jfreeblk->jf_frags += frag_offset; 4232 } 4233 4234 /* 4235 * Allocate a new jtrunc to track a partial truncation. 4236 */ 4237 static struct jtrunc * 4238 newjtrunc(freeblks, size, extsize) 4239 struct freeblks *freeblks; 4240 off_t size; 4241 int extsize; 4242 { 4243 struct jtrunc *jtrunc; 4244 4245 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4246 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4247 freeblks->fb_list.wk_mp); 4248 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4249 jtrunc->jt_dep.jb_freeblks = freeblks; 4250 jtrunc->jt_ino = freeblks->fb_inum; 4251 jtrunc->jt_size = size; 4252 jtrunc->jt_extsize = extsize; 4253 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4254 4255 return (jtrunc); 4256 } 4257 4258 /* 4259 * If we're canceling a new bitmap we have to search for another ref 4260 * to move into the bmsafemap dep. This might be better expressed 4261 * with another structure. 4262 */ 4263 static void 4264 move_newblock_dep(jaddref, inodedep) 4265 struct jaddref *jaddref; 4266 struct inodedep *inodedep; 4267 { 4268 struct inoref *inoref; 4269 struct jaddref *jaddrefn; 4270 4271 jaddrefn = NULL; 4272 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4273 inoref = TAILQ_NEXT(inoref, if_deps)) { 4274 if ((jaddref->ja_state & NEWBLOCK) && 4275 inoref->if_list.wk_type == D_JADDREF) { 4276 jaddrefn = (struct jaddref *)inoref; 4277 break; 4278 } 4279 } 4280 if (jaddrefn == NULL) 4281 return; 4282 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4283 jaddrefn->ja_state |= jaddref->ja_state & 4284 (ATTACHED | UNDONE | NEWBLOCK); 4285 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4286 jaddref->ja_state |= ATTACHED; 4287 LIST_REMOVE(jaddref, ja_bmdeps); 4288 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4289 ja_bmdeps); 4290 } 4291 4292 /* 4293 * Cancel a jaddref either before it has been written or while it is being 4294 * written. This happens when a link is removed before the add reaches 4295 * the disk. The jaddref dependency is kept linked into the bmsafemap 4296 * and inode to prevent the link count or bitmap from reaching the disk 4297 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4298 * required. 4299 * 4300 * Returns 1 if the canceled addref requires journaling of the remove and 4301 * 0 otherwise. 4302 */ 4303 static int 4304 cancel_jaddref(jaddref, inodedep, wkhd) 4305 struct jaddref *jaddref; 4306 struct inodedep *inodedep; 4307 struct workhead *wkhd; 4308 { 4309 struct inoref *inoref; 4310 struct jsegdep *jsegdep; 4311 int needsj; 4312 4313 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4314 ("cancel_jaddref: Canceling complete jaddref")); 4315 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4316 needsj = 1; 4317 else 4318 needsj = 0; 4319 if (inodedep == NULL) 4320 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4321 0, &inodedep) == 0) 4322 panic("cancel_jaddref: Lost inodedep"); 4323 /* 4324 * We must adjust the nlink of any reference operation that follows 4325 * us so that it is consistent with the in-memory reference. This 4326 * ensures that inode nlink rollbacks always have the correct link. 4327 */ 4328 if (needsj == 0) { 4329 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4330 inoref = TAILQ_NEXT(inoref, if_deps)) { 4331 if (inoref->if_state & GOINGAWAY) 4332 break; 4333 inoref->if_nlink--; 4334 } 4335 } 4336 jsegdep = inoref_jseg(&jaddref->ja_ref); 4337 if (jaddref->ja_state & NEWBLOCK) 4338 move_newblock_dep(jaddref, inodedep); 4339 wake_worklist(&jaddref->ja_list); 4340 jaddref->ja_mkdir = NULL; 4341 if (jaddref->ja_state & INPROGRESS) { 4342 jaddref->ja_state &= ~INPROGRESS; 4343 WORKLIST_REMOVE(&jaddref->ja_list); 4344 jwork_insert(wkhd, jsegdep); 4345 } else { 4346 free_jsegdep(jsegdep); 4347 if (jaddref->ja_state & DEPCOMPLETE) 4348 remove_from_journal(&jaddref->ja_list); 4349 } 4350 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4351 /* 4352 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4353 * can arrange for them to be freed with the bitmap. Otherwise we 4354 * no longer need this addref attached to the inoreflst and it 4355 * will incorrectly adjust nlink if we leave it. 4356 */ 4357 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4358 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4359 if_deps); 4360 jaddref->ja_state |= COMPLETE; 4361 free_jaddref(jaddref); 4362 return (needsj); 4363 } 4364 /* 4365 * Leave the head of the list for jsegdeps for fast merging. 4366 */ 4367 if (LIST_FIRST(wkhd) != NULL) { 4368 jaddref->ja_state |= ONWORKLIST; 4369 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4370 } else 4371 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4372 4373 return (needsj); 4374 } 4375 4376 /* 4377 * Attempt to free a jaddref structure when some work completes. This 4378 * should only succeed once the entry is written and all dependencies have 4379 * been notified. 4380 */ 4381 static void 4382 free_jaddref(jaddref) 4383 struct jaddref *jaddref; 4384 { 4385 4386 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4387 return; 4388 if (jaddref->ja_ref.if_jsegdep) 4389 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4390 jaddref, jaddref->ja_state); 4391 if (jaddref->ja_state & NEWBLOCK) 4392 LIST_REMOVE(jaddref, ja_bmdeps); 4393 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4394 panic("free_jaddref: Bad state %p(0x%X)", 4395 jaddref, jaddref->ja_state); 4396 if (jaddref->ja_mkdir != NULL) 4397 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4398 WORKITEM_FREE(jaddref, D_JADDREF); 4399 } 4400 4401 /* 4402 * Free a jremref structure once it has been written or discarded. 4403 */ 4404 static void 4405 free_jremref(jremref) 4406 struct jremref *jremref; 4407 { 4408 4409 if (jremref->jr_ref.if_jsegdep) 4410 free_jsegdep(jremref->jr_ref.if_jsegdep); 4411 if (jremref->jr_state & INPROGRESS) 4412 panic("free_jremref: IO still pending"); 4413 WORKITEM_FREE(jremref, D_JREMREF); 4414 } 4415 4416 /* 4417 * Free a jnewblk structure. 4418 */ 4419 static void 4420 free_jnewblk(jnewblk) 4421 struct jnewblk *jnewblk; 4422 { 4423 4424 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4425 return; 4426 LIST_REMOVE(jnewblk, jn_deps); 4427 if (jnewblk->jn_dep != NULL) 4428 panic("free_jnewblk: Dependency still attached."); 4429 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4430 } 4431 4432 /* 4433 * Cancel a jnewblk which has been been made redundant by frag extension. 4434 */ 4435 static void 4436 cancel_jnewblk(jnewblk, wkhd) 4437 struct jnewblk *jnewblk; 4438 struct workhead *wkhd; 4439 { 4440 struct jsegdep *jsegdep; 4441 4442 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4443 jsegdep = jnewblk->jn_jsegdep; 4444 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4445 panic("cancel_jnewblk: Invalid state"); 4446 jnewblk->jn_jsegdep = NULL; 4447 jnewblk->jn_dep = NULL; 4448 jnewblk->jn_state |= GOINGAWAY; 4449 if (jnewblk->jn_state & INPROGRESS) { 4450 jnewblk->jn_state &= ~INPROGRESS; 4451 WORKLIST_REMOVE(&jnewblk->jn_list); 4452 jwork_insert(wkhd, jsegdep); 4453 } else { 4454 free_jsegdep(jsegdep); 4455 remove_from_journal(&jnewblk->jn_list); 4456 } 4457 wake_worklist(&jnewblk->jn_list); 4458 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4459 } 4460 4461 static void 4462 free_jblkdep(jblkdep) 4463 struct jblkdep *jblkdep; 4464 { 4465 4466 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4467 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4468 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4469 WORKITEM_FREE(jblkdep, D_JTRUNC); 4470 else 4471 panic("free_jblkdep: Unexpected type %s", 4472 TYPENAME(jblkdep->jb_list.wk_type)); 4473 } 4474 4475 /* 4476 * Free a single jseg once it is no longer referenced in memory or on 4477 * disk. Reclaim journal blocks and dependencies waiting for the segment 4478 * to disappear. 4479 */ 4480 static void 4481 free_jseg(jseg, jblocks) 4482 struct jseg *jseg; 4483 struct jblocks *jblocks; 4484 { 4485 struct freework *freework; 4486 4487 /* 4488 * Free freework structures that were lingering to indicate freed 4489 * indirect blocks that forced journal write ordering on reallocate. 4490 */ 4491 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4492 indirblk_remove(freework); 4493 if (jblocks->jb_oldestseg == jseg) 4494 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4495 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4496 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4497 KASSERT(LIST_EMPTY(&jseg->js_entries), 4498 ("free_jseg: Freed jseg has valid entries.")); 4499 WORKITEM_FREE(jseg, D_JSEG); 4500 } 4501 4502 /* 4503 * Free all jsegs that meet the criteria for being reclaimed and update 4504 * oldestseg. 4505 */ 4506 static void 4507 free_jsegs(jblocks) 4508 struct jblocks *jblocks; 4509 { 4510 struct jseg *jseg; 4511 4512 /* 4513 * Free only those jsegs which have none allocated before them to 4514 * preserve the journal space ordering. 4515 */ 4516 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4517 /* 4518 * Only reclaim space when nothing depends on this journal 4519 * set and another set has written that it is no longer 4520 * valid. 4521 */ 4522 if (jseg->js_refs != 0) { 4523 jblocks->jb_oldestseg = jseg; 4524 return; 4525 } 4526 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4527 break; 4528 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4529 break; 4530 /* 4531 * We can free jsegs that didn't write entries when 4532 * oldestwrseq == js_seq. 4533 */ 4534 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4535 jseg->js_cnt != 0) 4536 break; 4537 free_jseg(jseg, jblocks); 4538 } 4539 /* 4540 * If we exited the loop above we still must discover the 4541 * oldest valid segment. 4542 */ 4543 if (jseg) 4544 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4545 jseg = TAILQ_NEXT(jseg, js_next)) 4546 if (jseg->js_refs != 0) 4547 break; 4548 jblocks->jb_oldestseg = jseg; 4549 /* 4550 * The journal has no valid records but some jsegs may still be 4551 * waiting on oldestwrseq to advance. We force a small record 4552 * out to permit these lingering records to be reclaimed. 4553 */ 4554 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4555 jblocks->jb_needseg = 1; 4556 } 4557 4558 /* 4559 * Release one reference to a jseg and free it if the count reaches 0. This 4560 * should eventually reclaim journal space as well. 4561 */ 4562 static void 4563 rele_jseg(jseg) 4564 struct jseg *jseg; 4565 { 4566 4567 KASSERT(jseg->js_refs > 0, 4568 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4569 if (--jseg->js_refs != 0) 4570 return; 4571 free_jsegs(jseg->js_jblocks); 4572 } 4573 4574 /* 4575 * Release a jsegdep and decrement the jseg count. 4576 */ 4577 static void 4578 free_jsegdep(jsegdep) 4579 struct jsegdep *jsegdep; 4580 { 4581 4582 if (jsegdep->jd_seg) 4583 rele_jseg(jsegdep->jd_seg); 4584 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4585 } 4586 4587 /* 4588 * Wait for a journal item to make it to disk. Initiate journal processing 4589 * if required. 4590 */ 4591 static int 4592 jwait(wk, waitfor) 4593 struct worklist *wk; 4594 int waitfor; 4595 { 4596 4597 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4598 /* 4599 * Blocking journal waits cause slow synchronous behavior. Record 4600 * stats on the frequency of these blocking operations. 4601 */ 4602 if (waitfor == MNT_WAIT) { 4603 stat_journal_wait++; 4604 switch (wk->wk_type) { 4605 case D_JREMREF: 4606 case D_JMVREF: 4607 stat_jwait_filepage++; 4608 break; 4609 case D_JTRUNC: 4610 case D_JFREEBLK: 4611 stat_jwait_freeblks++; 4612 break; 4613 case D_JNEWBLK: 4614 stat_jwait_newblk++; 4615 break; 4616 case D_JADDREF: 4617 stat_jwait_inode++; 4618 break; 4619 default: 4620 break; 4621 } 4622 } 4623 /* 4624 * If IO has not started we process the journal. We can't mark the 4625 * worklist item as IOWAITING because we drop the lock while 4626 * processing the journal and the worklist entry may be freed after 4627 * this point. The caller may call back in and re-issue the request. 4628 */ 4629 if ((wk->wk_state & INPROGRESS) == 0) { 4630 softdep_process_journal(wk->wk_mp, wk, waitfor); 4631 if (waitfor != MNT_WAIT) 4632 return (EBUSY); 4633 return (0); 4634 } 4635 if (waitfor != MNT_WAIT) 4636 return (EBUSY); 4637 wait_worklist(wk, "jwait"); 4638 return (0); 4639 } 4640 4641 /* 4642 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4643 * appropriate. This is a convenience function to reduce duplicate code 4644 * for the setup and revert functions below. 4645 */ 4646 static struct inodedep * 4647 inodedep_lookup_ip(ip) 4648 struct inode *ip; 4649 { 4650 struct inodedep *inodedep; 4651 4652 KASSERT(ip->i_nlink >= ip->i_effnlink, 4653 ("inodedep_lookup_ip: bad delta")); 4654 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC, 4655 &inodedep); 4656 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4657 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4658 4659 return (inodedep); 4660 } 4661 4662 /* 4663 * Called prior to creating a new inode and linking it to a directory. The 4664 * jaddref structure must already be allocated by softdep_setup_inomapdep 4665 * and it is discovered here so we can initialize the mode and update 4666 * nlinkdelta. 4667 */ 4668 void 4669 softdep_setup_create(dp, ip) 4670 struct inode *dp; 4671 struct inode *ip; 4672 { 4673 struct inodedep *inodedep; 4674 struct jaddref *jaddref; 4675 struct vnode *dvp; 4676 4677 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4678 ("softdep_setup_create called on non-softdep filesystem")); 4679 KASSERT(ip->i_nlink == 1, 4680 ("softdep_setup_create: Invalid link count.")); 4681 dvp = ITOV(dp); 4682 ACQUIRE_LOCK(ITOUMP(dp)); 4683 inodedep = inodedep_lookup_ip(ip); 4684 if (DOINGSUJ(dvp)) { 4685 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4686 inoreflst); 4687 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4688 ("softdep_setup_create: No addref structure present.")); 4689 } 4690 softdep_prelink(dvp, NULL); 4691 FREE_LOCK(ITOUMP(dp)); 4692 } 4693 4694 /* 4695 * Create a jaddref structure to track the addition of a DOTDOT link when 4696 * we are reparenting an inode as part of a rename. This jaddref will be 4697 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4698 * non-journaling softdep. 4699 */ 4700 void 4701 softdep_setup_dotdot_link(dp, ip) 4702 struct inode *dp; 4703 struct inode *ip; 4704 { 4705 struct inodedep *inodedep; 4706 struct jaddref *jaddref; 4707 struct vnode *dvp; 4708 4709 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4710 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4711 dvp = ITOV(dp); 4712 jaddref = NULL; 4713 /* 4714 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4715 * is used as a normal link would be. 4716 */ 4717 if (DOINGSUJ(dvp)) 4718 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4719 dp->i_effnlink - 1, dp->i_mode); 4720 ACQUIRE_LOCK(ITOUMP(dp)); 4721 inodedep = inodedep_lookup_ip(dp); 4722 if (jaddref) 4723 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4724 if_deps); 4725 softdep_prelink(dvp, ITOV(ip)); 4726 FREE_LOCK(ITOUMP(dp)); 4727 } 4728 4729 /* 4730 * Create a jaddref structure to track a new link to an inode. The directory 4731 * offset is not known until softdep_setup_directory_add or 4732 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4733 * softdep. 4734 */ 4735 void 4736 softdep_setup_link(dp, ip) 4737 struct inode *dp; 4738 struct inode *ip; 4739 { 4740 struct inodedep *inodedep; 4741 struct jaddref *jaddref; 4742 struct vnode *dvp; 4743 4744 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4745 ("softdep_setup_link called on non-softdep filesystem")); 4746 dvp = ITOV(dp); 4747 jaddref = NULL; 4748 if (DOINGSUJ(dvp)) 4749 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4750 ip->i_mode); 4751 ACQUIRE_LOCK(ITOUMP(dp)); 4752 inodedep = inodedep_lookup_ip(ip); 4753 if (jaddref) 4754 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4755 if_deps); 4756 softdep_prelink(dvp, ITOV(ip)); 4757 FREE_LOCK(ITOUMP(dp)); 4758 } 4759 4760 /* 4761 * Called to create the jaddref structures to track . and .. references as 4762 * well as lookup and further initialize the incomplete jaddref created 4763 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4764 * nlinkdelta for non-journaling softdep. 4765 */ 4766 void 4767 softdep_setup_mkdir(dp, ip) 4768 struct inode *dp; 4769 struct inode *ip; 4770 { 4771 struct inodedep *inodedep; 4772 struct jaddref *dotdotaddref; 4773 struct jaddref *dotaddref; 4774 struct jaddref *jaddref; 4775 struct vnode *dvp; 4776 4777 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4778 ("softdep_setup_mkdir called on non-softdep filesystem")); 4779 dvp = ITOV(dp); 4780 dotaddref = dotdotaddref = NULL; 4781 if (DOINGSUJ(dvp)) { 4782 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4783 ip->i_mode); 4784 dotaddref->ja_state |= MKDIR_BODY; 4785 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4786 dp->i_effnlink - 1, dp->i_mode); 4787 dotdotaddref->ja_state |= MKDIR_PARENT; 4788 } 4789 ACQUIRE_LOCK(ITOUMP(dp)); 4790 inodedep = inodedep_lookup_ip(ip); 4791 if (DOINGSUJ(dvp)) { 4792 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4793 inoreflst); 4794 KASSERT(jaddref != NULL, 4795 ("softdep_setup_mkdir: No addref structure present.")); 4796 KASSERT(jaddref->ja_parent == dp->i_number, 4797 ("softdep_setup_mkdir: bad parent %ju", 4798 (uintmax_t)jaddref->ja_parent)); 4799 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4800 if_deps); 4801 } 4802 inodedep = inodedep_lookup_ip(dp); 4803 if (DOINGSUJ(dvp)) 4804 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4805 &dotdotaddref->ja_ref, if_deps); 4806 softdep_prelink(ITOV(dp), NULL); 4807 FREE_LOCK(ITOUMP(dp)); 4808 } 4809 4810 /* 4811 * Called to track nlinkdelta of the inode and parent directories prior to 4812 * unlinking a directory. 4813 */ 4814 void 4815 softdep_setup_rmdir(dp, ip) 4816 struct inode *dp; 4817 struct inode *ip; 4818 { 4819 struct vnode *dvp; 4820 4821 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4822 ("softdep_setup_rmdir called on non-softdep filesystem")); 4823 dvp = ITOV(dp); 4824 ACQUIRE_LOCK(ITOUMP(dp)); 4825 (void) inodedep_lookup_ip(ip); 4826 (void) inodedep_lookup_ip(dp); 4827 softdep_prelink(dvp, ITOV(ip)); 4828 FREE_LOCK(ITOUMP(dp)); 4829 } 4830 4831 /* 4832 * Called to track nlinkdelta of the inode and parent directories prior to 4833 * unlink. 4834 */ 4835 void 4836 softdep_setup_unlink(dp, ip) 4837 struct inode *dp; 4838 struct inode *ip; 4839 { 4840 struct vnode *dvp; 4841 4842 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4843 ("softdep_setup_unlink called on non-softdep filesystem")); 4844 dvp = ITOV(dp); 4845 ACQUIRE_LOCK(ITOUMP(dp)); 4846 (void) inodedep_lookup_ip(ip); 4847 (void) inodedep_lookup_ip(dp); 4848 softdep_prelink(dvp, ITOV(ip)); 4849 FREE_LOCK(ITOUMP(dp)); 4850 } 4851 4852 /* 4853 * Called to release the journal structures created by a failed non-directory 4854 * creation. Adjusts nlinkdelta for non-journaling softdep. 4855 */ 4856 void 4857 softdep_revert_create(dp, ip) 4858 struct inode *dp; 4859 struct inode *ip; 4860 { 4861 struct inodedep *inodedep; 4862 struct jaddref *jaddref; 4863 struct vnode *dvp; 4864 4865 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0, 4866 ("softdep_revert_create called on non-softdep filesystem")); 4867 dvp = ITOV(dp); 4868 ACQUIRE_LOCK(ITOUMP(dp)); 4869 inodedep = inodedep_lookup_ip(ip); 4870 if (DOINGSUJ(dvp)) { 4871 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4872 inoreflst); 4873 KASSERT(jaddref->ja_parent == dp->i_number, 4874 ("softdep_revert_create: addref parent mismatch")); 4875 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4876 } 4877 FREE_LOCK(ITOUMP(dp)); 4878 } 4879 4880 /* 4881 * Called to release the journal structures created by a failed link 4882 * addition. Adjusts nlinkdelta for non-journaling softdep. 4883 */ 4884 void 4885 softdep_revert_link(dp, ip) 4886 struct inode *dp; 4887 struct inode *ip; 4888 { 4889 struct inodedep *inodedep; 4890 struct jaddref *jaddref; 4891 struct vnode *dvp; 4892 4893 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4894 ("softdep_revert_link called on non-softdep filesystem")); 4895 dvp = ITOV(dp); 4896 ACQUIRE_LOCK(ITOUMP(dp)); 4897 inodedep = inodedep_lookup_ip(ip); 4898 if (DOINGSUJ(dvp)) { 4899 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4900 inoreflst); 4901 KASSERT(jaddref->ja_parent == dp->i_number, 4902 ("softdep_revert_link: addref parent mismatch")); 4903 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4904 } 4905 FREE_LOCK(ITOUMP(dp)); 4906 } 4907 4908 /* 4909 * Called to release the journal structures created by a failed mkdir 4910 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4911 */ 4912 void 4913 softdep_revert_mkdir(dp, ip) 4914 struct inode *dp; 4915 struct inode *ip; 4916 { 4917 struct inodedep *inodedep; 4918 struct jaddref *jaddref; 4919 struct jaddref *dotaddref; 4920 struct vnode *dvp; 4921 4922 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4923 ("softdep_revert_mkdir called on non-softdep filesystem")); 4924 dvp = ITOV(dp); 4925 4926 ACQUIRE_LOCK(ITOUMP(dp)); 4927 inodedep = inodedep_lookup_ip(dp); 4928 if (DOINGSUJ(dvp)) { 4929 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4930 inoreflst); 4931 KASSERT(jaddref->ja_parent == ip->i_number, 4932 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4933 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4934 } 4935 inodedep = inodedep_lookup_ip(ip); 4936 if (DOINGSUJ(dvp)) { 4937 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4938 inoreflst); 4939 KASSERT(jaddref->ja_parent == dp->i_number, 4940 ("softdep_revert_mkdir: addref parent mismatch")); 4941 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4942 inoreflst, if_deps); 4943 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4944 KASSERT(dotaddref->ja_parent == ip->i_number, 4945 ("softdep_revert_mkdir: dot addref parent mismatch")); 4946 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4947 } 4948 FREE_LOCK(ITOUMP(dp)); 4949 } 4950 4951 /* 4952 * Called to correct nlinkdelta after a failed rmdir. 4953 */ 4954 void 4955 softdep_revert_rmdir(dp, ip) 4956 struct inode *dp; 4957 struct inode *ip; 4958 { 4959 4960 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4961 ("softdep_revert_rmdir called on non-softdep filesystem")); 4962 ACQUIRE_LOCK(ITOUMP(dp)); 4963 (void) inodedep_lookup_ip(ip); 4964 (void) inodedep_lookup_ip(dp); 4965 FREE_LOCK(ITOUMP(dp)); 4966 } 4967 4968 /* 4969 * Protecting the freemaps (or bitmaps). 4970 * 4971 * To eliminate the need to execute fsck before mounting a filesystem 4972 * after a power failure, one must (conservatively) guarantee that the 4973 * on-disk copy of the bitmaps never indicate that a live inode or block is 4974 * free. So, when a block or inode is allocated, the bitmap should be 4975 * updated (on disk) before any new pointers. When a block or inode is 4976 * freed, the bitmap should not be updated until all pointers have been 4977 * reset. The latter dependency is handled by the delayed de-allocation 4978 * approach described below for block and inode de-allocation. The former 4979 * dependency is handled by calling the following procedure when a block or 4980 * inode is allocated. When an inode is allocated an "inodedep" is created 4981 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4982 * Each "inodedep" is also inserted into the hash indexing structure so 4983 * that any additional link additions can be made dependent on the inode 4984 * allocation. 4985 * 4986 * The ufs filesystem maintains a number of free block counts (e.g., per 4987 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4988 * in addition to the bitmaps. These counts are used to improve efficiency 4989 * during allocation and therefore must be consistent with the bitmaps. 4990 * There is no convenient way to guarantee post-crash consistency of these 4991 * counts with simple update ordering, for two main reasons: (1) The counts 4992 * and bitmaps for a single cylinder group block are not in the same disk 4993 * sector. If a disk write is interrupted (e.g., by power failure), one may 4994 * be written and the other not. (2) Some of the counts are located in the 4995 * superblock rather than the cylinder group block. So, we focus our soft 4996 * updates implementation on protecting the bitmaps. When mounting a 4997 * filesystem, we recompute the auxiliary counts from the bitmaps. 4998 */ 4999 5000 /* 5001 * Called just after updating the cylinder group block to allocate an inode. 5002 */ 5003 void 5004 softdep_setup_inomapdep(bp, ip, newinum, mode) 5005 struct buf *bp; /* buffer for cylgroup block with inode map */ 5006 struct inode *ip; /* inode related to allocation */ 5007 ino_t newinum; /* new inode number being allocated */ 5008 int mode; 5009 { 5010 struct inodedep *inodedep; 5011 struct bmsafemap *bmsafemap; 5012 struct jaddref *jaddref; 5013 struct mount *mp; 5014 struct fs *fs; 5015 5016 mp = ITOVFS(ip); 5017 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5018 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5019 fs = VFSTOUFS(mp)->um_fs; 5020 jaddref = NULL; 5021 5022 /* 5023 * Allocate the journal reference add structure so that the bitmap 5024 * can be dependent on it. 5025 */ 5026 if (MOUNTEDSUJ(mp)) { 5027 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5028 jaddref->ja_state |= NEWBLOCK; 5029 } 5030 5031 /* 5032 * Create a dependency for the newly allocated inode. 5033 * Panic if it already exists as something is seriously wrong. 5034 * Otherwise add it to the dependency list for the buffer holding 5035 * the cylinder group map from which it was allocated. 5036 * 5037 * We have to preallocate a bmsafemap entry in case it is needed 5038 * in bmsafemap_lookup since once we allocate the inodedep, we 5039 * have to finish initializing it before we can FREE_LOCK(). 5040 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5041 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5042 * creating the inodedep as it can be freed during the time 5043 * that we FREE_LOCK() while allocating the inodedep. We must 5044 * call workitem_alloc() before entering the locked section as 5045 * it also acquires the lock and we must avoid trying doing so 5046 * recursively. 5047 */ 5048 bmsafemap = malloc(sizeof(struct bmsafemap), 5049 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5050 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5051 ACQUIRE_LOCK(ITOUMP(ip)); 5052 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5053 panic("softdep_setup_inomapdep: dependency %p for new" 5054 "inode already exists", inodedep); 5055 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5056 if (jaddref) { 5057 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5058 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5059 if_deps); 5060 } else { 5061 inodedep->id_state |= ONDEPLIST; 5062 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5063 } 5064 inodedep->id_bmsafemap = bmsafemap; 5065 inodedep->id_state &= ~DEPCOMPLETE; 5066 FREE_LOCK(ITOUMP(ip)); 5067 } 5068 5069 /* 5070 * Called just after updating the cylinder group block to 5071 * allocate block or fragment. 5072 */ 5073 void 5074 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5075 struct buf *bp; /* buffer for cylgroup block with block map */ 5076 struct mount *mp; /* filesystem doing allocation */ 5077 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5078 int frags; /* Number of fragments. */ 5079 int oldfrags; /* Previous number of fragments for extend. */ 5080 { 5081 struct newblk *newblk; 5082 struct bmsafemap *bmsafemap; 5083 struct jnewblk *jnewblk; 5084 struct ufsmount *ump; 5085 struct fs *fs; 5086 5087 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5088 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5089 ump = VFSTOUFS(mp); 5090 fs = ump->um_fs; 5091 jnewblk = NULL; 5092 /* 5093 * Create a dependency for the newly allocated block. 5094 * Add it to the dependency list for the buffer holding 5095 * the cylinder group map from which it was allocated. 5096 */ 5097 if (MOUNTEDSUJ(mp)) { 5098 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5099 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5100 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5101 jnewblk->jn_state = ATTACHED; 5102 jnewblk->jn_blkno = newblkno; 5103 jnewblk->jn_frags = frags; 5104 jnewblk->jn_oldfrags = oldfrags; 5105 #ifdef SUJ_DEBUG 5106 { 5107 struct cg *cgp; 5108 uint8_t *blksfree; 5109 long bno; 5110 int i; 5111 5112 cgp = (struct cg *)bp->b_data; 5113 blksfree = cg_blksfree(cgp); 5114 bno = dtogd(fs, jnewblk->jn_blkno); 5115 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5116 i++) { 5117 if (isset(blksfree, bno + i)) 5118 panic("softdep_setup_blkmapdep: " 5119 "free fragment %d from %d-%d " 5120 "state 0x%X dep %p", i, 5121 jnewblk->jn_oldfrags, 5122 jnewblk->jn_frags, 5123 jnewblk->jn_state, 5124 jnewblk->jn_dep); 5125 } 5126 } 5127 #endif 5128 } 5129 5130 CTR3(KTR_SUJ, 5131 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5132 newblkno, frags, oldfrags); 5133 ACQUIRE_LOCK(ump); 5134 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5135 panic("softdep_setup_blkmapdep: found block"); 5136 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5137 dtog(fs, newblkno), NULL); 5138 if (jnewblk) { 5139 jnewblk->jn_dep = (struct worklist *)newblk; 5140 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5141 } else { 5142 newblk->nb_state |= ONDEPLIST; 5143 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5144 } 5145 newblk->nb_bmsafemap = bmsafemap; 5146 newblk->nb_jnewblk = jnewblk; 5147 FREE_LOCK(ump); 5148 } 5149 5150 #define BMSAFEMAP_HASH(ump, cg) \ 5151 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5152 5153 static int 5154 bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5155 struct bmsafemap_hashhead *bmsafemaphd; 5156 int cg; 5157 struct bmsafemap **bmsafemapp; 5158 { 5159 struct bmsafemap *bmsafemap; 5160 5161 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5162 if (bmsafemap->sm_cg == cg) 5163 break; 5164 if (bmsafemap) { 5165 *bmsafemapp = bmsafemap; 5166 return (1); 5167 } 5168 *bmsafemapp = NULL; 5169 5170 return (0); 5171 } 5172 5173 /* 5174 * Find the bmsafemap associated with a cylinder group buffer. 5175 * If none exists, create one. The buffer must be locked when 5176 * this routine is called and this routine must be called with 5177 * the softdep lock held. To avoid giving up the lock while 5178 * allocating a new bmsafemap, a preallocated bmsafemap may be 5179 * provided. If it is provided but not needed, it is freed. 5180 */ 5181 static struct bmsafemap * 5182 bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5183 struct mount *mp; 5184 struct buf *bp; 5185 int cg; 5186 struct bmsafemap *newbmsafemap; 5187 { 5188 struct bmsafemap_hashhead *bmsafemaphd; 5189 struct bmsafemap *bmsafemap, *collision; 5190 struct worklist *wk; 5191 struct ufsmount *ump; 5192 5193 ump = VFSTOUFS(mp); 5194 LOCK_OWNED(ump); 5195 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5196 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5197 if (wk->wk_type == D_BMSAFEMAP) { 5198 if (newbmsafemap) 5199 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5200 return (WK_BMSAFEMAP(wk)); 5201 } 5202 } 5203 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5204 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5205 if (newbmsafemap) 5206 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5207 return (bmsafemap); 5208 } 5209 if (newbmsafemap) { 5210 bmsafemap = newbmsafemap; 5211 } else { 5212 FREE_LOCK(ump); 5213 bmsafemap = malloc(sizeof(struct bmsafemap), 5214 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5215 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5216 ACQUIRE_LOCK(ump); 5217 } 5218 bmsafemap->sm_buf = bp; 5219 LIST_INIT(&bmsafemap->sm_inodedephd); 5220 LIST_INIT(&bmsafemap->sm_inodedepwr); 5221 LIST_INIT(&bmsafemap->sm_newblkhd); 5222 LIST_INIT(&bmsafemap->sm_newblkwr); 5223 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5224 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5225 LIST_INIT(&bmsafemap->sm_freehd); 5226 LIST_INIT(&bmsafemap->sm_freewr); 5227 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5228 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5229 return (collision); 5230 } 5231 bmsafemap->sm_cg = cg; 5232 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5233 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5234 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5235 return (bmsafemap); 5236 } 5237 5238 /* 5239 * Direct block allocation dependencies. 5240 * 5241 * When a new block is allocated, the corresponding disk locations must be 5242 * initialized (with zeros or new data) before the on-disk inode points to 5243 * them. Also, the freemap from which the block was allocated must be 5244 * updated (on disk) before the inode's pointer. These two dependencies are 5245 * independent of each other and are needed for all file blocks and indirect 5246 * blocks that are pointed to directly by the inode. Just before the 5247 * "in-core" version of the inode is updated with a newly allocated block 5248 * number, a procedure (below) is called to setup allocation dependency 5249 * structures. These structures are removed when the corresponding 5250 * dependencies are satisfied or when the block allocation becomes obsolete 5251 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5252 * fragment that gets upgraded). All of these cases are handled in 5253 * procedures described later. 5254 * 5255 * When a file extension causes a fragment to be upgraded, either to a larger 5256 * fragment or to a full block, the on-disk location may change (if the 5257 * previous fragment could not simply be extended). In this case, the old 5258 * fragment must be de-allocated, but not until after the inode's pointer has 5259 * been updated. In most cases, this is handled by later procedures, which 5260 * will construct a "freefrag" structure to be added to the workitem queue 5261 * when the inode update is complete (or obsolete). The main exception to 5262 * this is when an allocation occurs while a pending allocation dependency 5263 * (for the same block pointer) remains. This case is handled in the main 5264 * allocation dependency setup procedure by immediately freeing the 5265 * unreferenced fragments. 5266 */ 5267 void 5268 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5269 struct inode *ip; /* inode to which block is being added */ 5270 ufs_lbn_t off; /* block pointer within inode */ 5271 ufs2_daddr_t newblkno; /* disk block number being added */ 5272 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5273 long newsize; /* size of new block */ 5274 long oldsize; /* size of new block */ 5275 struct buf *bp; /* bp for allocated block */ 5276 { 5277 struct allocdirect *adp, *oldadp; 5278 struct allocdirectlst *adphead; 5279 struct freefrag *freefrag; 5280 struct inodedep *inodedep; 5281 struct pagedep *pagedep; 5282 struct jnewblk *jnewblk; 5283 struct newblk *newblk; 5284 struct mount *mp; 5285 ufs_lbn_t lbn; 5286 5287 lbn = bp->b_lblkno; 5288 mp = ITOVFS(ip); 5289 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5290 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5291 if (oldblkno && oldblkno != newblkno) 5292 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5293 else 5294 freefrag = NULL; 5295 5296 CTR6(KTR_SUJ, 5297 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5298 "off %jd newsize %ld oldsize %d", 5299 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5300 ACQUIRE_LOCK(ITOUMP(ip)); 5301 if (off >= UFS_NDADDR) { 5302 if (lbn > 0) 5303 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5304 lbn, off); 5305 /* allocating an indirect block */ 5306 if (oldblkno != 0) 5307 panic("softdep_setup_allocdirect: non-zero indir"); 5308 } else { 5309 if (off != lbn) 5310 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5311 lbn, off); 5312 /* 5313 * Allocating a direct block. 5314 * 5315 * If we are allocating a directory block, then we must 5316 * allocate an associated pagedep to track additions and 5317 * deletions. 5318 */ 5319 if ((ip->i_mode & IFMT) == IFDIR) 5320 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5321 &pagedep); 5322 } 5323 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5324 panic("softdep_setup_allocdirect: lost block"); 5325 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5326 ("softdep_setup_allocdirect: newblk already initialized")); 5327 /* 5328 * Convert the newblk to an allocdirect. 5329 */ 5330 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5331 adp = (struct allocdirect *)newblk; 5332 newblk->nb_freefrag = freefrag; 5333 adp->ad_offset = off; 5334 adp->ad_oldblkno = oldblkno; 5335 adp->ad_newsize = newsize; 5336 adp->ad_oldsize = oldsize; 5337 5338 /* 5339 * Finish initializing the journal. 5340 */ 5341 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5342 jnewblk->jn_ino = ip->i_number; 5343 jnewblk->jn_lbn = lbn; 5344 add_to_journal(&jnewblk->jn_list); 5345 } 5346 if (freefrag && freefrag->ff_jdep != NULL && 5347 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5348 add_to_journal(freefrag->ff_jdep); 5349 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5350 adp->ad_inodedep = inodedep; 5351 5352 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5353 /* 5354 * The list of allocdirects must be kept in sorted and ascending 5355 * order so that the rollback routines can quickly determine the 5356 * first uncommitted block (the size of the file stored on disk 5357 * ends at the end of the lowest committed fragment, or if there 5358 * are no fragments, at the end of the highest committed block). 5359 * Since files generally grow, the typical case is that the new 5360 * block is to be added at the end of the list. We speed this 5361 * special case by checking against the last allocdirect in the 5362 * list before laboriously traversing the list looking for the 5363 * insertion point. 5364 */ 5365 adphead = &inodedep->id_newinoupdt; 5366 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5367 if (oldadp == NULL || oldadp->ad_offset <= off) { 5368 /* insert at end of list */ 5369 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5370 if (oldadp != NULL && oldadp->ad_offset == off) 5371 allocdirect_merge(adphead, adp, oldadp); 5372 FREE_LOCK(ITOUMP(ip)); 5373 return; 5374 } 5375 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5376 if (oldadp->ad_offset >= off) 5377 break; 5378 } 5379 if (oldadp == NULL) 5380 panic("softdep_setup_allocdirect: lost entry"); 5381 /* insert in middle of list */ 5382 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5383 if (oldadp->ad_offset == off) 5384 allocdirect_merge(adphead, adp, oldadp); 5385 5386 FREE_LOCK(ITOUMP(ip)); 5387 } 5388 5389 /* 5390 * Merge a newer and older journal record to be stored either in a 5391 * newblock or freefrag. This handles aggregating journal records for 5392 * fragment allocation into a second record as well as replacing a 5393 * journal free with an aborted journal allocation. A segment for the 5394 * oldest record will be placed on wkhd if it has been written. If not 5395 * the segment for the newer record will suffice. 5396 */ 5397 static struct worklist * 5398 jnewblk_merge(new, old, wkhd) 5399 struct worklist *new; 5400 struct worklist *old; 5401 struct workhead *wkhd; 5402 { 5403 struct jnewblk *njnewblk; 5404 struct jnewblk *jnewblk; 5405 5406 /* Handle NULLs to simplify callers. */ 5407 if (new == NULL) 5408 return (old); 5409 if (old == NULL) 5410 return (new); 5411 /* Replace a jfreefrag with a jnewblk. */ 5412 if (new->wk_type == D_JFREEFRAG) { 5413 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5414 panic("jnewblk_merge: blkno mismatch: %p, %p", 5415 old, new); 5416 cancel_jfreefrag(WK_JFREEFRAG(new)); 5417 return (old); 5418 } 5419 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5420 panic("jnewblk_merge: Bad type: old %d new %d\n", 5421 old->wk_type, new->wk_type); 5422 /* 5423 * Handle merging of two jnewblk records that describe 5424 * different sets of fragments in the same block. 5425 */ 5426 jnewblk = WK_JNEWBLK(old); 5427 njnewblk = WK_JNEWBLK(new); 5428 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5429 panic("jnewblk_merge: Merging disparate blocks."); 5430 /* 5431 * The record may be rolled back in the cg. 5432 */ 5433 if (jnewblk->jn_state & UNDONE) { 5434 jnewblk->jn_state &= ~UNDONE; 5435 njnewblk->jn_state |= UNDONE; 5436 njnewblk->jn_state &= ~ATTACHED; 5437 } 5438 /* 5439 * We modify the newer addref and free the older so that if neither 5440 * has been written the most up-to-date copy will be on disk. If 5441 * both have been written but rolled back we only temporarily need 5442 * one of them to fix the bits when the cg write completes. 5443 */ 5444 jnewblk->jn_state |= ATTACHED | COMPLETE; 5445 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5446 cancel_jnewblk(jnewblk, wkhd); 5447 WORKLIST_REMOVE(&jnewblk->jn_list); 5448 free_jnewblk(jnewblk); 5449 return (new); 5450 } 5451 5452 /* 5453 * Replace an old allocdirect dependency with a newer one. 5454 * This routine must be called with splbio interrupts blocked. 5455 */ 5456 static void 5457 allocdirect_merge(adphead, newadp, oldadp) 5458 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5459 struct allocdirect *newadp; /* allocdirect being added */ 5460 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5461 { 5462 struct worklist *wk; 5463 struct freefrag *freefrag; 5464 5465 freefrag = NULL; 5466 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5467 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5468 newadp->ad_oldsize != oldadp->ad_newsize || 5469 newadp->ad_offset >= UFS_NDADDR) 5470 panic("%s %jd != new %jd || old size %ld != new %ld", 5471 "allocdirect_merge: old blkno", 5472 (intmax_t)newadp->ad_oldblkno, 5473 (intmax_t)oldadp->ad_newblkno, 5474 newadp->ad_oldsize, oldadp->ad_newsize); 5475 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5476 newadp->ad_oldsize = oldadp->ad_oldsize; 5477 /* 5478 * If the old dependency had a fragment to free or had never 5479 * previously had a block allocated, then the new dependency 5480 * can immediately post its freefrag and adopt the old freefrag. 5481 * This action is done by swapping the freefrag dependencies. 5482 * The new dependency gains the old one's freefrag, and the 5483 * old one gets the new one and then immediately puts it on 5484 * the worklist when it is freed by free_newblk. It is 5485 * not possible to do this swap when the old dependency had a 5486 * non-zero size but no previous fragment to free. This condition 5487 * arises when the new block is an extension of the old block. 5488 * Here, the first part of the fragment allocated to the new 5489 * dependency is part of the block currently claimed on disk by 5490 * the old dependency, so cannot legitimately be freed until the 5491 * conditions for the new dependency are fulfilled. 5492 */ 5493 freefrag = newadp->ad_freefrag; 5494 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5495 newadp->ad_freefrag = oldadp->ad_freefrag; 5496 oldadp->ad_freefrag = freefrag; 5497 } 5498 /* 5499 * If we are tracking a new directory-block allocation, 5500 * move it from the old allocdirect to the new allocdirect. 5501 */ 5502 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5503 WORKLIST_REMOVE(wk); 5504 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5505 panic("allocdirect_merge: extra newdirblk"); 5506 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5507 } 5508 TAILQ_REMOVE(adphead, oldadp, ad_next); 5509 /* 5510 * We need to move any journal dependencies over to the freefrag 5511 * that releases this block if it exists. Otherwise we are 5512 * extending an existing block and we'll wait until that is 5513 * complete to release the journal space and extend the 5514 * new journal to cover this old space as well. 5515 */ 5516 if (freefrag == NULL) { 5517 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5518 panic("allocdirect_merge: %jd != %jd", 5519 oldadp->ad_newblkno, newadp->ad_newblkno); 5520 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5521 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5522 &oldadp->ad_block.nb_jnewblk->jn_list, 5523 &newadp->ad_block.nb_jwork); 5524 oldadp->ad_block.nb_jnewblk = NULL; 5525 cancel_newblk(&oldadp->ad_block, NULL, 5526 &newadp->ad_block.nb_jwork); 5527 } else { 5528 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5529 &freefrag->ff_list, &freefrag->ff_jwork); 5530 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5531 &freefrag->ff_jwork); 5532 } 5533 free_newblk(&oldadp->ad_block); 5534 } 5535 5536 /* 5537 * Allocate a jfreefrag structure to journal a single block free. 5538 */ 5539 static struct jfreefrag * 5540 newjfreefrag(freefrag, ip, blkno, size, lbn) 5541 struct freefrag *freefrag; 5542 struct inode *ip; 5543 ufs2_daddr_t blkno; 5544 long size; 5545 ufs_lbn_t lbn; 5546 { 5547 struct jfreefrag *jfreefrag; 5548 struct fs *fs; 5549 5550 fs = ITOFS(ip); 5551 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5552 M_SOFTDEP_FLAGS); 5553 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip)); 5554 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5555 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5556 jfreefrag->fr_ino = ip->i_number; 5557 jfreefrag->fr_lbn = lbn; 5558 jfreefrag->fr_blkno = blkno; 5559 jfreefrag->fr_frags = numfrags(fs, size); 5560 jfreefrag->fr_freefrag = freefrag; 5561 5562 return (jfreefrag); 5563 } 5564 5565 /* 5566 * Allocate a new freefrag structure. 5567 */ 5568 static struct freefrag * 5569 newfreefrag(ip, blkno, size, lbn) 5570 struct inode *ip; 5571 ufs2_daddr_t blkno; 5572 long size; 5573 ufs_lbn_t lbn; 5574 { 5575 struct freefrag *freefrag; 5576 struct ufsmount *ump; 5577 struct fs *fs; 5578 5579 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5580 ip->i_number, blkno, size, lbn); 5581 ump = ITOUMP(ip); 5582 fs = ump->um_fs; 5583 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5584 panic("newfreefrag: frag size"); 5585 freefrag = malloc(sizeof(struct freefrag), 5586 M_FREEFRAG, M_SOFTDEP_FLAGS); 5587 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump)); 5588 freefrag->ff_state = ATTACHED; 5589 LIST_INIT(&freefrag->ff_jwork); 5590 freefrag->ff_inum = ip->i_number; 5591 freefrag->ff_vtype = ITOV(ip)->v_type; 5592 freefrag->ff_blkno = blkno; 5593 freefrag->ff_fragsize = size; 5594 5595 if (MOUNTEDSUJ(UFSTOVFS(ump))) { 5596 freefrag->ff_jdep = (struct worklist *) 5597 newjfreefrag(freefrag, ip, blkno, size, lbn); 5598 } else { 5599 freefrag->ff_state |= DEPCOMPLETE; 5600 freefrag->ff_jdep = NULL; 5601 } 5602 5603 return (freefrag); 5604 } 5605 5606 /* 5607 * This workitem de-allocates fragments that were replaced during 5608 * file block allocation. 5609 */ 5610 static void 5611 handle_workitem_freefrag(freefrag) 5612 struct freefrag *freefrag; 5613 { 5614 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5615 struct workhead wkhd; 5616 5617 CTR3(KTR_SUJ, 5618 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5619 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5620 /* 5621 * It would be illegal to add new completion items to the 5622 * freefrag after it was schedule to be done so it must be 5623 * safe to modify the list head here. 5624 */ 5625 LIST_INIT(&wkhd); 5626 ACQUIRE_LOCK(ump); 5627 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5628 /* 5629 * If the journal has not been written we must cancel it here. 5630 */ 5631 if (freefrag->ff_jdep) { 5632 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5633 panic("handle_workitem_freefrag: Unexpected type %d\n", 5634 freefrag->ff_jdep->wk_type); 5635 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5636 } 5637 FREE_LOCK(ump); 5638 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5639 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5640 ACQUIRE_LOCK(ump); 5641 WORKITEM_FREE(freefrag, D_FREEFRAG); 5642 FREE_LOCK(ump); 5643 } 5644 5645 /* 5646 * Set up a dependency structure for an external attributes data block. 5647 * This routine follows much of the structure of softdep_setup_allocdirect. 5648 * See the description of softdep_setup_allocdirect above for details. 5649 */ 5650 void 5651 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5652 struct inode *ip; 5653 ufs_lbn_t off; 5654 ufs2_daddr_t newblkno; 5655 ufs2_daddr_t oldblkno; 5656 long newsize; 5657 long oldsize; 5658 struct buf *bp; 5659 { 5660 struct allocdirect *adp, *oldadp; 5661 struct allocdirectlst *adphead; 5662 struct freefrag *freefrag; 5663 struct inodedep *inodedep; 5664 struct jnewblk *jnewblk; 5665 struct newblk *newblk; 5666 struct mount *mp; 5667 struct ufsmount *ump; 5668 ufs_lbn_t lbn; 5669 5670 mp = ITOVFS(ip); 5671 ump = VFSTOUFS(mp); 5672 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5673 ("softdep_setup_allocext called on non-softdep filesystem")); 5674 KASSERT(off < UFS_NXADDR, 5675 ("softdep_setup_allocext: lbn %lld > UFS_NXADDR", (long long)off)); 5676 5677 lbn = bp->b_lblkno; 5678 if (oldblkno && oldblkno != newblkno) 5679 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5680 else 5681 freefrag = NULL; 5682 5683 ACQUIRE_LOCK(ump); 5684 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5685 panic("softdep_setup_allocext: lost block"); 5686 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5687 ("softdep_setup_allocext: newblk already initialized")); 5688 /* 5689 * Convert the newblk to an allocdirect. 5690 */ 5691 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5692 adp = (struct allocdirect *)newblk; 5693 newblk->nb_freefrag = freefrag; 5694 adp->ad_offset = off; 5695 adp->ad_oldblkno = oldblkno; 5696 adp->ad_newsize = newsize; 5697 adp->ad_oldsize = oldsize; 5698 adp->ad_state |= EXTDATA; 5699 5700 /* 5701 * Finish initializing the journal. 5702 */ 5703 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5704 jnewblk->jn_ino = ip->i_number; 5705 jnewblk->jn_lbn = lbn; 5706 add_to_journal(&jnewblk->jn_list); 5707 } 5708 if (freefrag && freefrag->ff_jdep != NULL && 5709 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5710 add_to_journal(freefrag->ff_jdep); 5711 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5712 adp->ad_inodedep = inodedep; 5713 5714 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5715 /* 5716 * The list of allocdirects must be kept in sorted and ascending 5717 * order so that the rollback routines can quickly determine the 5718 * first uncommitted block (the size of the file stored on disk 5719 * ends at the end of the lowest committed fragment, or if there 5720 * are no fragments, at the end of the highest committed block). 5721 * Since files generally grow, the typical case is that the new 5722 * block is to be added at the end of the list. We speed this 5723 * special case by checking against the last allocdirect in the 5724 * list before laboriously traversing the list looking for the 5725 * insertion point. 5726 */ 5727 adphead = &inodedep->id_newextupdt; 5728 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5729 if (oldadp == NULL || oldadp->ad_offset <= off) { 5730 /* insert at end of list */ 5731 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5732 if (oldadp != NULL && oldadp->ad_offset == off) 5733 allocdirect_merge(adphead, adp, oldadp); 5734 FREE_LOCK(ump); 5735 return; 5736 } 5737 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5738 if (oldadp->ad_offset >= off) 5739 break; 5740 } 5741 if (oldadp == NULL) 5742 panic("softdep_setup_allocext: lost entry"); 5743 /* insert in middle of list */ 5744 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5745 if (oldadp->ad_offset == off) 5746 allocdirect_merge(adphead, adp, oldadp); 5747 FREE_LOCK(ump); 5748 } 5749 5750 /* 5751 * Indirect block allocation dependencies. 5752 * 5753 * The same dependencies that exist for a direct block also exist when 5754 * a new block is allocated and pointed to by an entry in a block of 5755 * indirect pointers. The undo/redo states described above are also 5756 * used here. Because an indirect block contains many pointers that 5757 * may have dependencies, a second copy of the entire in-memory indirect 5758 * block is kept. The buffer cache copy is always completely up-to-date. 5759 * The second copy, which is used only as a source for disk writes, 5760 * contains only the safe pointers (i.e., those that have no remaining 5761 * update dependencies). The second copy is freed when all pointers 5762 * are safe. The cache is not allowed to replace indirect blocks with 5763 * pending update dependencies. If a buffer containing an indirect 5764 * block with dependencies is written, these routines will mark it 5765 * dirty again. It can only be successfully written once all the 5766 * dependencies are removed. The ffs_fsync routine in conjunction with 5767 * softdep_sync_metadata work together to get all the dependencies 5768 * removed so that a file can be successfully written to disk. Three 5769 * procedures are used when setting up indirect block pointer 5770 * dependencies. The division is necessary because of the organization 5771 * of the "balloc" routine and because of the distinction between file 5772 * pages and file metadata blocks. 5773 */ 5774 5775 /* 5776 * Allocate a new allocindir structure. 5777 */ 5778 static struct allocindir * 5779 newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5780 struct inode *ip; /* inode for file being extended */ 5781 int ptrno; /* offset of pointer in indirect block */ 5782 ufs2_daddr_t newblkno; /* disk block number being added */ 5783 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5784 ufs_lbn_t lbn; 5785 { 5786 struct newblk *newblk; 5787 struct allocindir *aip; 5788 struct freefrag *freefrag; 5789 struct jnewblk *jnewblk; 5790 5791 if (oldblkno) 5792 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn); 5793 else 5794 freefrag = NULL; 5795 ACQUIRE_LOCK(ITOUMP(ip)); 5796 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0) 5797 panic("new_allocindir: lost block"); 5798 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5799 ("newallocindir: newblk already initialized")); 5800 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5801 newblk->nb_freefrag = freefrag; 5802 aip = (struct allocindir *)newblk; 5803 aip->ai_offset = ptrno; 5804 aip->ai_oldblkno = oldblkno; 5805 aip->ai_lbn = lbn; 5806 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5807 jnewblk->jn_ino = ip->i_number; 5808 jnewblk->jn_lbn = lbn; 5809 add_to_journal(&jnewblk->jn_list); 5810 } 5811 if (freefrag && freefrag->ff_jdep != NULL && 5812 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5813 add_to_journal(freefrag->ff_jdep); 5814 return (aip); 5815 } 5816 5817 /* 5818 * Called just before setting an indirect block pointer 5819 * to a newly allocated file page. 5820 */ 5821 void 5822 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5823 struct inode *ip; /* inode for file being extended */ 5824 ufs_lbn_t lbn; /* allocated block number within file */ 5825 struct buf *bp; /* buffer with indirect blk referencing page */ 5826 int ptrno; /* offset of pointer in indirect block */ 5827 ufs2_daddr_t newblkno; /* disk block number being added */ 5828 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5829 struct buf *nbp; /* buffer holding allocated page */ 5830 { 5831 struct inodedep *inodedep; 5832 struct freefrag *freefrag; 5833 struct allocindir *aip; 5834 struct pagedep *pagedep; 5835 struct mount *mp; 5836 struct ufsmount *ump; 5837 5838 mp = ITOVFS(ip); 5839 ump = VFSTOUFS(mp); 5840 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5841 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5842 KASSERT(lbn == nbp->b_lblkno, 5843 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5844 lbn, bp->b_lblkno)); 5845 CTR4(KTR_SUJ, 5846 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5847 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5848 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5849 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5850 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5851 /* 5852 * If we are allocating a directory page, then we must 5853 * allocate an associated pagedep to track additions and 5854 * deletions. 5855 */ 5856 if ((ip->i_mode & IFMT) == IFDIR) 5857 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5858 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5859 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5860 FREE_LOCK(ump); 5861 if (freefrag) 5862 handle_workitem_freefrag(freefrag); 5863 } 5864 5865 /* 5866 * Called just before setting an indirect block pointer to a 5867 * newly allocated indirect block. 5868 */ 5869 void 5870 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5871 struct buf *nbp; /* newly allocated indirect block */ 5872 struct inode *ip; /* inode for file being extended */ 5873 struct buf *bp; /* indirect block referencing allocated block */ 5874 int ptrno; /* offset of pointer in indirect block */ 5875 ufs2_daddr_t newblkno; /* disk block number being added */ 5876 { 5877 struct inodedep *inodedep; 5878 struct allocindir *aip; 5879 struct ufsmount *ump; 5880 ufs_lbn_t lbn; 5881 5882 ump = ITOUMP(ip); 5883 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 5884 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5885 CTR3(KTR_SUJ, 5886 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5887 ip->i_number, newblkno, ptrno); 5888 lbn = nbp->b_lblkno; 5889 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5890 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5891 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 5892 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5893 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5894 panic("softdep_setup_allocindir_meta: Block already existed"); 5895 FREE_LOCK(ump); 5896 } 5897 5898 static void 5899 indirdep_complete(indirdep) 5900 struct indirdep *indirdep; 5901 { 5902 struct allocindir *aip; 5903 5904 LIST_REMOVE(indirdep, ir_next); 5905 indirdep->ir_state |= DEPCOMPLETE; 5906 5907 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5908 LIST_REMOVE(aip, ai_next); 5909 free_newblk(&aip->ai_block); 5910 } 5911 /* 5912 * If this indirdep is not attached to a buf it was simply waiting 5913 * on completion to clear completehd. free_indirdep() asserts 5914 * that nothing is dangling. 5915 */ 5916 if ((indirdep->ir_state & ONWORKLIST) == 0) 5917 free_indirdep(indirdep); 5918 } 5919 5920 static struct indirdep * 5921 indirdep_lookup(mp, ip, bp) 5922 struct mount *mp; 5923 struct inode *ip; 5924 struct buf *bp; 5925 { 5926 struct indirdep *indirdep, *newindirdep; 5927 struct newblk *newblk; 5928 struct ufsmount *ump; 5929 struct worklist *wk; 5930 struct fs *fs; 5931 ufs2_daddr_t blkno; 5932 5933 ump = VFSTOUFS(mp); 5934 LOCK_OWNED(ump); 5935 indirdep = NULL; 5936 newindirdep = NULL; 5937 fs = ump->um_fs; 5938 for (;;) { 5939 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5940 if (wk->wk_type != D_INDIRDEP) 5941 continue; 5942 indirdep = WK_INDIRDEP(wk); 5943 break; 5944 } 5945 /* Found on the buffer worklist, no new structure to free. */ 5946 if (indirdep != NULL && newindirdep == NULL) 5947 return (indirdep); 5948 if (indirdep != NULL && newindirdep != NULL) 5949 panic("indirdep_lookup: simultaneous create"); 5950 /* None found on the buffer and a new structure is ready. */ 5951 if (indirdep == NULL && newindirdep != NULL) 5952 break; 5953 /* None found and no new structure available. */ 5954 FREE_LOCK(ump); 5955 newindirdep = malloc(sizeof(struct indirdep), 5956 M_INDIRDEP, M_SOFTDEP_FLAGS); 5957 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5958 newindirdep->ir_state = ATTACHED; 5959 if (I_IS_UFS1(ip)) 5960 newindirdep->ir_state |= UFS1FMT; 5961 TAILQ_INIT(&newindirdep->ir_trunc); 5962 newindirdep->ir_saveddata = NULL; 5963 LIST_INIT(&newindirdep->ir_deplisthd); 5964 LIST_INIT(&newindirdep->ir_donehd); 5965 LIST_INIT(&newindirdep->ir_writehd); 5966 LIST_INIT(&newindirdep->ir_completehd); 5967 if (bp->b_blkno == bp->b_lblkno) { 5968 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5969 NULL, NULL); 5970 bp->b_blkno = blkno; 5971 } 5972 newindirdep->ir_freeblks = NULL; 5973 newindirdep->ir_savebp = 5974 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5975 newindirdep->ir_bp = bp; 5976 BUF_KERNPROC(newindirdep->ir_savebp); 5977 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5978 ACQUIRE_LOCK(ump); 5979 } 5980 indirdep = newindirdep; 5981 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5982 /* 5983 * If the block is not yet allocated we don't set DEPCOMPLETE so 5984 * that we don't free dependencies until the pointers are valid. 5985 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5986 * than using the hash. 5987 */ 5988 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5989 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5990 else 5991 indirdep->ir_state |= DEPCOMPLETE; 5992 return (indirdep); 5993 } 5994 5995 /* 5996 * Called to finish the allocation of the "aip" allocated 5997 * by one of the two routines above. 5998 */ 5999 static struct freefrag * 6000 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 6001 struct buf *bp; /* in-memory copy of the indirect block */ 6002 struct inode *ip; /* inode for file being extended */ 6003 struct inodedep *inodedep; /* Inodedep for ip */ 6004 struct allocindir *aip; /* allocindir allocated by the above routines */ 6005 ufs_lbn_t lbn; /* Logical block number for this block. */ 6006 { 6007 struct fs *fs; 6008 struct indirdep *indirdep; 6009 struct allocindir *oldaip; 6010 struct freefrag *freefrag; 6011 struct mount *mp; 6012 struct ufsmount *ump; 6013 6014 mp = ITOVFS(ip); 6015 ump = VFSTOUFS(mp); 6016 LOCK_OWNED(ump); 6017 fs = ump->um_fs; 6018 if (bp->b_lblkno >= 0) 6019 panic("setup_allocindir_phase2: not indir blk"); 6020 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6021 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6022 indirdep = indirdep_lookup(mp, ip, bp); 6023 KASSERT(indirdep->ir_savebp != NULL, 6024 ("setup_allocindir_phase2 NULL ir_savebp")); 6025 aip->ai_indirdep = indirdep; 6026 /* 6027 * Check for an unwritten dependency for this indirect offset. If 6028 * there is, merge the old dependency into the new one. This happens 6029 * as a result of reallocblk only. 6030 */ 6031 freefrag = NULL; 6032 if (aip->ai_oldblkno != 0) { 6033 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6034 if (oldaip->ai_offset == aip->ai_offset) { 6035 freefrag = allocindir_merge(aip, oldaip); 6036 goto done; 6037 } 6038 } 6039 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6040 if (oldaip->ai_offset == aip->ai_offset) { 6041 freefrag = allocindir_merge(aip, oldaip); 6042 goto done; 6043 } 6044 } 6045 } 6046 done: 6047 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6048 return (freefrag); 6049 } 6050 6051 /* 6052 * Merge two allocindirs which refer to the same block. Move newblock 6053 * dependencies and setup the freefrags appropriately. 6054 */ 6055 static struct freefrag * 6056 allocindir_merge(aip, oldaip) 6057 struct allocindir *aip; 6058 struct allocindir *oldaip; 6059 { 6060 struct freefrag *freefrag; 6061 struct worklist *wk; 6062 6063 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6064 panic("allocindir_merge: blkno"); 6065 aip->ai_oldblkno = oldaip->ai_oldblkno; 6066 freefrag = aip->ai_freefrag; 6067 aip->ai_freefrag = oldaip->ai_freefrag; 6068 oldaip->ai_freefrag = NULL; 6069 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6070 /* 6071 * If we are tracking a new directory-block allocation, 6072 * move it from the old allocindir to the new allocindir. 6073 */ 6074 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6075 WORKLIST_REMOVE(wk); 6076 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6077 panic("allocindir_merge: extra newdirblk"); 6078 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6079 } 6080 /* 6081 * We can skip journaling for this freefrag and just complete 6082 * any pending journal work for the allocindir that is being 6083 * removed after the freefrag completes. 6084 */ 6085 if (freefrag->ff_jdep) 6086 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6087 LIST_REMOVE(oldaip, ai_next); 6088 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6089 &freefrag->ff_list, &freefrag->ff_jwork); 6090 free_newblk(&oldaip->ai_block); 6091 6092 return (freefrag); 6093 } 6094 6095 static inline void 6096 setup_freedirect(freeblks, ip, i, needj) 6097 struct freeblks *freeblks; 6098 struct inode *ip; 6099 int i; 6100 int needj; 6101 { 6102 struct ufsmount *ump; 6103 ufs2_daddr_t blkno; 6104 int frags; 6105 6106 blkno = DIP(ip, i_db[i]); 6107 if (blkno == 0) 6108 return; 6109 DIP_SET(ip, i_db[i], 0); 6110 ump = ITOUMP(ip); 6111 frags = sblksize(ump->um_fs, ip->i_size, i); 6112 frags = numfrags(ump->um_fs, frags); 6113 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj); 6114 } 6115 6116 static inline void 6117 setup_freeext(freeblks, ip, i, needj) 6118 struct freeblks *freeblks; 6119 struct inode *ip; 6120 int i; 6121 int needj; 6122 { 6123 struct ufsmount *ump; 6124 ufs2_daddr_t blkno; 6125 int frags; 6126 6127 blkno = ip->i_din2->di_extb[i]; 6128 if (blkno == 0) 6129 return; 6130 ip->i_din2->di_extb[i] = 0; 6131 ump = ITOUMP(ip); 6132 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i); 6133 frags = numfrags(ump->um_fs, frags); 6134 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6135 } 6136 6137 static inline void 6138 setup_freeindir(freeblks, ip, i, lbn, needj) 6139 struct freeblks *freeblks; 6140 struct inode *ip; 6141 int i; 6142 ufs_lbn_t lbn; 6143 int needj; 6144 { 6145 struct ufsmount *ump; 6146 ufs2_daddr_t blkno; 6147 6148 blkno = DIP(ip, i_ib[i]); 6149 if (blkno == 0) 6150 return; 6151 DIP_SET(ip, i_ib[i], 0); 6152 ump = ITOUMP(ip); 6153 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag, 6154 0, needj); 6155 } 6156 6157 static inline struct freeblks * 6158 newfreeblks(mp, ip) 6159 struct mount *mp; 6160 struct inode *ip; 6161 { 6162 struct freeblks *freeblks; 6163 6164 freeblks = malloc(sizeof(struct freeblks), 6165 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6166 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6167 LIST_INIT(&freeblks->fb_jblkdephd); 6168 LIST_INIT(&freeblks->fb_jwork); 6169 freeblks->fb_ref = 0; 6170 freeblks->fb_cgwait = 0; 6171 freeblks->fb_state = ATTACHED; 6172 freeblks->fb_uid = ip->i_uid; 6173 freeblks->fb_inum = ip->i_number; 6174 freeblks->fb_vtype = ITOV(ip)->v_type; 6175 freeblks->fb_modrev = DIP(ip, i_modrev); 6176 freeblks->fb_devvp = ITODEVVP(ip); 6177 freeblks->fb_chkcnt = 0; 6178 freeblks->fb_len = 0; 6179 6180 return (freeblks); 6181 } 6182 6183 static void 6184 trunc_indirdep(indirdep, freeblks, bp, off) 6185 struct indirdep *indirdep; 6186 struct freeblks *freeblks; 6187 struct buf *bp; 6188 int off; 6189 { 6190 struct allocindir *aip, *aipn; 6191 6192 /* 6193 * The first set of allocindirs won't be in savedbp. 6194 */ 6195 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6196 if (aip->ai_offset > off) 6197 cancel_allocindir(aip, bp, freeblks, 1); 6198 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6199 if (aip->ai_offset > off) 6200 cancel_allocindir(aip, bp, freeblks, 1); 6201 /* 6202 * These will exist in savedbp. 6203 */ 6204 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6205 if (aip->ai_offset > off) 6206 cancel_allocindir(aip, NULL, freeblks, 0); 6207 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6208 if (aip->ai_offset > off) 6209 cancel_allocindir(aip, NULL, freeblks, 0); 6210 } 6211 6212 /* 6213 * Follow the chain of indirects down to lastlbn creating a freework 6214 * structure for each. This will be used to start indir_trunc() at 6215 * the right offset and create the journal records for the parrtial 6216 * truncation. A second step will handle the truncated dependencies. 6217 */ 6218 static int 6219 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6220 struct freeblks *freeblks; 6221 struct inode *ip; 6222 ufs_lbn_t lbn; 6223 ufs_lbn_t lastlbn; 6224 ufs2_daddr_t blkno; 6225 { 6226 struct indirdep *indirdep; 6227 struct indirdep *indirn; 6228 struct freework *freework; 6229 struct newblk *newblk; 6230 struct mount *mp; 6231 struct ufsmount *ump; 6232 struct buf *bp; 6233 uint8_t *start; 6234 uint8_t *end; 6235 ufs_lbn_t lbnadd; 6236 int level; 6237 int error; 6238 int off; 6239 6240 6241 freework = NULL; 6242 if (blkno == 0) 6243 return (0); 6244 mp = freeblks->fb_list.wk_mp; 6245 ump = VFSTOUFS(mp); 6246 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6247 if ((bp->b_flags & B_CACHE) == 0) { 6248 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6249 bp->b_iocmd = BIO_READ; 6250 bp->b_flags &= ~B_INVAL; 6251 bp->b_ioflags &= ~BIO_ERROR; 6252 vfs_busy_pages(bp, 0); 6253 bp->b_iooffset = dbtob(bp->b_blkno); 6254 bstrategy(bp); 6255 #ifdef RACCT 6256 if (racct_enable) { 6257 PROC_LOCK(curproc); 6258 racct_add_buf(curproc, bp, 0); 6259 PROC_UNLOCK(curproc); 6260 } 6261 #endif /* RACCT */ 6262 curthread->td_ru.ru_inblock++; 6263 error = bufwait(bp); 6264 if (error) { 6265 brelse(bp); 6266 return (error); 6267 } 6268 } 6269 level = lbn_level(lbn); 6270 lbnadd = lbn_offset(ump->um_fs, level); 6271 /* 6272 * Compute the offset of the last block we want to keep. Store 6273 * in the freework the first block we want to completely free. 6274 */ 6275 off = (lastlbn - -(lbn + level)) / lbnadd; 6276 if (off + 1 == NINDIR(ump->um_fs)) 6277 goto nowork; 6278 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0); 6279 /* 6280 * Link the freework into the indirdep. This will prevent any new 6281 * allocations from proceeding until we are finished with the 6282 * truncate and the block is written. 6283 */ 6284 ACQUIRE_LOCK(ump); 6285 indirdep = indirdep_lookup(mp, ip, bp); 6286 if (indirdep->ir_freeblks) 6287 panic("setup_trunc_indir: indirdep already truncated."); 6288 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6289 freework->fw_indir = indirdep; 6290 /* 6291 * Cancel any allocindirs that will not make it to disk. 6292 * We have to do this for all copies of the indirdep that 6293 * live on this newblk. 6294 */ 6295 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6296 if (newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0, 6297 &newblk) == 0) 6298 panic("setup_trunc_indir: lost block"); 6299 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6300 trunc_indirdep(indirn, freeblks, bp, off); 6301 } else 6302 trunc_indirdep(indirdep, freeblks, bp, off); 6303 FREE_LOCK(ump); 6304 /* 6305 * Creation is protected by the buf lock. The saveddata is only 6306 * needed if a full truncation follows a partial truncation but it 6307 * is difficult to allocate in that case so we fetch it anyway. 6308 */ 6309 if (indirdep->ir_saveddata == NULL) 6310 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6311 M_SOFTDEP_FLAGS); 6312 nowork: 6313 /* Fetch the blkno of the child and the zero start offset. */ 6314 if (I_IS_UFS1(ip)) { 6315 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6316 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6317 } else { 6318 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6319 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6320 } 6321 if (freework) { 6322 /* Zero the truncated pointers. */ 6323 end = bp->b_data + bp->b_bcount; 6324 bzero(start, end - start); 6325 bdwrite(bp); 6326 } else 6327 bqrelse(bp); 6328 if (level == 0) 6329 return (0); 6330 lbn++; /* adjust level */ 6331 lbn -= (off * lbnadd); 6332 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6333 } 6334 6335 /* 6336 * Complete the partial truncation of an indirect block setup by 6337 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6338 * copy and writes them to disk before the freeblks is allowed to complete. 6339 */ 6340 static void 6341 complete_trunc_indir(freework) 6342 struct freework *freework; 6343 { 6344 struct freework *fwn; 6345 struct indirdep *indirdep; 6346 struct ufsmount *ump; 6347 struct buf *bp; 6348 uintptr_t start; 6349 int count; 6350 6351 ump = VFSTOUFS(freework->fw_list.wk_mp); 6352 LOCK_OWNED(ump); 6353 indirdep = freework->fw_indir; 6354 for (;;) { 6355 bp = indirdep->ir_bp; 6356 /* See if the block was discarded. */ 6357 if (bp == NULL) 6358 break; 6359 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6360 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6361 break; 6362 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6363 LOCK_PTR(ump)) == 0) 6364 BUF_UNLOCK(bp); 6365 ACQUIRE_LOCK(ump); 6366 } 6367 freework->fw_state |= DEPCOMPLETE; 6368 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6369 /* 6370 * Zero the pointers in the saved copy. 6371 */ 6372 if (indirdep->ir_state & UFS1FMT) 6373 start = sizeof(ufs1_daddr_t); 6374 else 6375 start = sizeof(ufs2_daddr_t); 6376 start *= freework->fw_start; 6377 count = indirdep->ir_savebp->b_bcount - start; 6378 start += (uintptr_t)indirdep->ir_savebp->b_data; 6379 bzero((char *)start, count); 6380 /* 6381 * We need to start the next truncation in the list if it has not 6382 * been started yet. 6383 */ 6384 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6385 if (fwn != NULL) { 6386 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6387 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6388 if ((fwn->fw_state & ONWORKLIST) == 0) 6389 freework_enqueue(fwn); 6390 } 6391 /* 6392 * If bp is NULL the block was fully truncated, restore 6393 * the saved block list otherwise free it if it is no 6394 * longer needed. 6395 */ 6396 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6397 if (bp == NULL) 6398 bcopy(indirdep->ir_saveddata, 6399 indirdep->ir_savebp->b_data, 6400 indirdep->ir_savebp->b_bcount); 6401 free(indirdep->ir_saveddata, M_INDIRDEP); 6402 indirdep->ir_saveddata = NULL; 6403 } 6404 /* 6405 * When bp is NULL there is a full truncation pending. We 6406 * must wait for this full truncation to be journaled before 6407 * we can release this freework because the disk pointers will 6408 * never be written as zero. 6409 */ 6410 if (bp == NULL) { 6411 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6412 handle_written_freework(freework); 6413 else 6414 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6415 &freework->fw_list); 6416 } else { 6417 /* Complete when the real copy is written. */ 6418 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6419 BUF_UNLOCK(bp); 6420 } 6421 } 6422 6423 /* 6424 * Calculate the number of blocks we are going to release where datablocks 6425 * is the current total and length is the new file size. 6426 */ 6427 static ufs2_daddr_t 6428 blkcount(fs, datablocks, length) 6429 struct fs *fs; 6430 ufs2_daddr_t datablocks; 6431 off_t length; 6432 { 6433 off_t totblks, numblks; 6434 6435 totblks = 0; 6436 numblks = howmany(length, fs->fs_bsize); 6437 if (numblks <= UFS_NDADDR) { 6438 totblks = howmany(length, fs->fs_fsize); 6439 goto out; 6440 } 6441 totblks = blkstofrags(fs, numblks); 6442 numblks -= UFS_NDADDR; 6443 /* 6444 * Count all single, then double, then triple indirects required. 6445 * Subtracting one indirects worth of blocks for each pass 6446 * acknowledges one of each pointed to by the inode. 6447 */ 6448 for (;;) { 6449 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6450 numblks -= NINDIR(fs); 6451 if (numblks <= 0) 6452 break; 6453 numblks = howmany(numblks, NINDIR(fs)); 6454 } 6455 out: 6456 totblks = fsbtodb(fs, totblks); 6457 /* 6458 * Handle sparse files. We can't reclaim more blocks than the inode 6459 * references. We will correct it later in handle_complete_freeblks() 6460 * when we know the real count. 6461 */ 6462 if (totblks > datablocks) 6463 return (0); 6464 return (datablocks - totblks); 6465 } 6466 6467 /* 6468 * Handle freeblocks for journaled softupdate filesystems. 6469 * 6470 * Contrary to normal softupdates, we must preserve the block pointers in 6471 * indirects until their subordinates are free. This is to avoid journaling 6472 * every block that is freed which may consume more space than the journal 6473 * itself. The recovery program will see the free block journals at the 6474 * base of the truncated area and traverse them to reclaim space. The 6475 * pointers in the inode may be cleared immediately after the journal 6476 * records are written because each direct and indirect pointer in the 6477 * inode is recorded in a journal. This permits full truncation to proceed 6478 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6479 * 6480 * The algorithm is as follows: 6481 * 1) Traverse the in-memory state and create journal entries to release 6482 * the relevant blocks and full indirect trees. 6483 * 2) Traverse the indirect block chain adding partial truncation freework 6484 * records to indirects in the path to lastlbn. The freework will 6485 * prevent new allocation dependencies from being satisfied in this 6486 * indirect until the truncation completes. 6487 * 3) Read and lock the inode block, performing an update with the new size 6488 * and pointers. This prevents truncated data from becoming valid on 6489 * disk through step 4. 6490 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6491 * eliminate journal work for those records that do not require it. 6492 * 5) Schedule the journal records to be written followed by the inode block. 6493 * 6) Allocate any necessary frags for the end of file. 6494 * 7) Zero any partially truncated blocks. 6495 * 6496 * From this truncation proceeds asynchronously using the freework and 6497 * indir_trunc machinery. The file will not be extended again into a 6498 * partially truncated indirect block until all work is completed but 6499 * the normal dependency mechanism ensures that it is rolled back/forward 6500 * as appropriate. Further truncation may occur without delay and is 6501 * serialized in indir_trunc(). 6502 */ 6503 void 6504 softdep_journal_freeblocks(ip, cred, length, flags) 6505 struct inode *ip; /* The inode whose length is to be reduced */ 6506 struct ucred *cred; 6507 off_t length; /* The new length for the file */ 6508 int flags; /* IO_EXT and/or IO_NORMAL */ 6509 { 6510 struct freeblks *freeblks, *fbn; 6511 struct worklist *wk, *wkn; 6512 struct inodedep *inodedep; 6513 struct jblkdep *jblkdep; 6514 struct allocdirect *adp, *adpn; 6515 struct ufsmount *ump; 6516 struct fs *fs; 6517 struct buf *bp; 6518 struct vnode *vp; 6519 struct mount *mp; 6520 ufs2_daddr_t extblocks, datablocks; 6521 ufs_lbn_t tmpval, lbn, lastlbn; 6522 int frags, lastoff, iboff, allocblock, needj, error, i; 6523 6524 ump = ITOUMP(ip); 6525 mp = UFSTOVFS(ump); 6526 fs = ump->um_fs; 6527 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6528 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6529 vp = ITOV(ip); 6530 needj = 1; 6531 iboff = -1; 6532 allocblock = 0; 6533 extblocks = 0; 6534 datablocks = 0; 6535 frags = 0; 6536 freeblks = newfreeblks(mp, ip); 6537 ACQUIRE_LOCK(ump); 6538 /* 6539 * If we're truncating a removed file that will never be written 6540 * we don't need to journal the block frees. The canceled journals 6541 * for the allocations will suffice. 6542 */ 6543 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6544 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6545 length == 0) 6546 needj = 0; 6547 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6548 ip->i_number, length, needj); 6549 FREE_LOCK(ump); 6550 /* 6551 * Calculate the lbn that we are truncating to. This results in -1 6552 * if we're truncating the 0 bytes. So it is the last lbn we want 6553 * to keep, not the first lbn we want to truncate. 6554 */ 6555 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6556 lastoff = blkoff(fs, length); 6557 /* 6558 * Compute frags we are keeping in lastlbn. 0 means all. 6559 */ 6560 if (lastlbn >= 0 && lastlbn < UFS_NDADDR) { 6561 frags = fragroundup(fs, lastoff); 6562 /* adp offset of last valid allocdirect. */ 6563 iboff = lastlbn; 6564 } else if (lastlbn > 0) 6565 iboff = UFS_NDADDR; 6566 if (fs->fs_magic == FS_UFS2_MAGIC) 6567 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6568 /* 6569 * Handle normal data blocks and indirects. This section saves 6570 * values used after the inode update to complete frag and indirect 6571 * truncation. 6572 */ 6573 if ((flags & IO_NORMAL) != 0) { 6574 /* 6575 * Handle truncation of whole direct and indirect blocks. 6576 */ 6577 for (i = iboff + 1; i < UFS_NDADDR; i++) 6578 setup_freedirect(freeblks, ip, i, needj); 6579 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6580 i < UFS_NIADDR; 6581 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6582 /* Release a whole indirect tree. */ 6583 if (lbn > lastlbn) { 6584 setup_freeindir(freeblks, ip, i, -lbn -i, 6585 needj); 6586 continue; 6587 } 6588 iboff = i + UFS_NDADDR; 6589 /* 6590 * Traverse partially truncated indirect tree. 6591 */ 6592 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6593 setup_trunc_indir(freeblks, ip, -lbn - i, 6594 lastlbn, DIP(ip, i_ib[i])); 6595 } 6596 /* 6597 * Handle partial truncation to a frag boundary. 6598 */ 6599 if (frags) { 6600 ufs2_daddr_t blkno; 6601 long oldfrags; 6602 6603 oldfrags = blksize(fs, ip, lastlbn); 6604 blkno = DIP(ip, i_db[lastlbn]); 6605 if (blkno && oldfrags != frags) { 6606 oldfrags -= frags; 6607 oldfrags = numfrags(fs, oldfrags); 6608 blkno += numfrags(fs, frags); 6609 newfreework(ump, freeblks, NULL, lastlbn, 6610 blkno, oldfrags, 0, needj); 6611 if (needj) 6612 adjust_newfreework(freeblks, 6613 numfrags(fs, frags)); 6614 } else if (blkno == 0) 6615 allocblock = 1; 6616 } 6617 /* 6618 * Add a journal record for partial truncate if we are 6619 * handling indirect blocks. Non-indirects need no extra 6620 * journaling. 6621 */ 6622 if (length != 0 && lastlbn >= UFS_NDADDR) { 6623 ip->i_flag |= IN_TRUNCATED; 6624 newjtrunc(freeblks, length, 0); 6625 } 6626 ip->i_size = length; 6627 DIP_SET(ip, i_size, ip->i_size); 6628 datablocks = DIP(ip, i_blocks) - extblocks; 6629 if (length != 0) 6630 datablocks = blkcount(fs, datablocks, length); 6631 freeblks->fb_len = length; 6632 } 6633 if ((flags & IO_EXT) != 0) { 6634 for (i = 0; i < UFS_NXADDR; i++) 6635 setup_freeext(freeblks, ip, i, needj); 6636 ip->i_din2->di_extsize = 0; 6637 datablocks += extblocks; 6638 } 6639 #ifdef QUOTA 6640 /* Reference the quotas in case the block count is wrong in the end. */ 6641 quotaref(vp, freeblks->fb_quota); 6642 (void) chkdq(ip, -datablocks, NOCRED, 0); 6643 #endif 6644 freeblks->fb_chkcnt = -datablocks; 6645 UFS_LOCK(ump); 6646 fs->fs_pendingblocks += datablocks; 6647 UFS_UNLOCK(ump); 6648 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6649 /* 6650 * Handle truncation of incomplete alloc direct dependencies. We 6651 * hold the inode block locked to prevent incomplete dependencies 6652 * from reaching the disk while we are eliminating those that 6653 * have been truncated. This is a partially inlined ffs_update(). 6654 */ 6655 ufs_itimes(vp); 6656 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6657 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6658 (int)fs->fs_bsize, cred, &bp); 6659 if (error) { 6660 brelse(bp); 6661 softdep_error("softdep_journal_freeblocks", error); 6662 return; 6663 } 6664 if (bp->b_bufsize == fs->fs_bsize) 6665 bp->b_flags |= B_CLUSTEROK; 6666 softdep_update_inodeblock(ip, bp, 0); 6667 if (ump->um_fstype == UFS1) 6668 *((struct ufs1_dinode *)bp->b_data + 6669 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6670 else 6671 *((struct ufs2_dinode *)bp->b_data + 6672 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6673 ACQUIRE_LOCK(ump); 6674 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6675 if ((inodedep->id_state & IOSTARTED) != 0) 6676 panic("softdep_setup_freeblocks: inode busy"); 6677 /* 6678 * Add the freeblks structure to the list of operations that 6679 * must await the zero'ed inode being written to disk. If we 6680 * still have a bitmap dependency (needj), then the inode 6681 * has never been written to disk, so we can process the 6682 * freeblks below once we have deleted the dependencies. 6683 */ 6684 if (needj) 6685 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6686 else 6687 freeblks->fb_state |= COMPLETE; 6688 if ((flags & IO_NORMAL) != 0) { 6689 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6690 if (adp->ad_offset > iboff) 6691 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6692 freeblks); 6693 /* 6694 * Truncate the allocdirect. We could eliminate 6695 * or modify journal records as well. 6696 */ 6697 else if (adp->ad_offset == iboff && frags) 6698 adp->ad_newsize = frags; 6699 } 6700 } 6701 if ((flags & IO_EXT) != 0) 6702 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6703 cancel_allocdirect(&inodedep->id_extupdt, adp, 6704 freeblks); 6705 /* 6706 * Scan the bufwait list for newblock dependencies that will never 6707 * make it to disk. 6708 */ 6709 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6710 if (wk->wk_type != D_ALLOCDIRECT) 6711 continue; 6712 adp = WK_ALLOCDIRECT(wk); 6713 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6714 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6715 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6716 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6717 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6718 } 6719 } 6720 /* 6721 * Add journal work. 6722 */ 6723 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6724 add_to_journal(&jblkdep->jb_list); 6725 FREE_LOCK(ump); 6726 bdwrite(bp); 6727 /* 6728 * Truncate dependency structures beyond length. 6729 */ 6730 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6731 /* 6732 * This is only set when we need to allocate a fragment because 6733 * none existed at the end of a frag-sized file. It handles only 6734 * allocating a new, zero filled block. 6735 */ 6736 if (allocblock) { 6737 ip->i_size = length - lastoff; 6738 DIP_SET(ip, i_size, ip->i_size); 6739 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6740 if (error != 0) { 6741 softdep_error("softdep_journal_freeblks", error); 6742 return; 6743 } 6744 ip->i_size = length; 6745 DIP_SET(ip, i_size, length); 6746 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6747 allocbuf(bp, frags); 6748 ffs_update(vp, 0); 6749 bawrite(bp); 6750 } else if (lastoff != 0 && vp->v_type != VDIR) { 6751 int size; 6752 6753 /* 6754 * Zero the end of a truncated frag or block. 6755 */ 6756 size = sblksize(fs, length, lastlbn); 6757 error = bread(vp, lastlbn, size, cred, &bp); 6758 if (error) { 6759 softdep_error("softdep_journal_freeblks", error); 6760 return; 6761 } 6762 bzero((char *)bp->b_data + lastoff, size - lastoff); 6763 bawrite(bp); 6764 6765 } 6766 ACQUIRE_LOCK(ump); 6767 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6768 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6769 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6770 /* 6771 * We zero earlier truncations so they don't erroneously 6772 * update i_blocks. 6773 */ 6774 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6775 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6776 fbn->fb_len = 0; 6777 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6778 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6779 freeblks->fb_state |= INPROGRESS; 6780 else 6781 freeblks = NULL; 6782 FREE_LOCK(ump); 6783 if (freeblks) 6784 handle_workitem_freeblocks(freeblks, 0); 6785 trunc_pages(ip, length, extblocks, flags); 6786 6787 } 6788 6789 /* 6790 * Flush a JOP_SYNC to the journal. 6791 */ 6792 void 6793 softdep_journal_fsync(ip) 6794 struct inode *ip; 6795 { 6796 struct jfsync *jfsync; 6797 struct ufsmount *ump; 6798 6799 ump = ITOUMP(ip); 6800 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 6801 ("softdep_journal_fsync called on non-softdep filesystem")); 6802 if ((ip->i_flag & IN_TRUNCATED) == 0) 6803 return; 6804 ip->i_flag &= ~IN_TRUNCATED; 6805 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6806 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump)); 6807 jfsync->jfs_size = ip->i_size; 6808 jfsync->jfs_ino = ip->i_number; 6809 ACQUIRE_LOCK(ump); 6810 add_to_journal(&jfsync->jfs_list); 6811 jwait(&jfsync->jfs_list, MNT_WAIT); 6812 FREE_LOCK(ump); 6813 } 6814 6815 /* 6816 * Block de-allocation dependencies. 6817 * 6818 * When blocks are de-allocated, the on-disk pointers must be nullified before 6819 * the blocks are made available for use by other files. (The true 6820 * requirement is that old pointers must be nullified before new on-disk 6821 * pointers are set. We chose this slightly more stringent requirement to 6822 * reduce complexity.) Our implementation handles this dependency by updating 6823 * the inode (or indirect block) appropriately but delaying the actual block 6824 * de-allocation (i.e., freemap and free space count manipulation) until 6825 * after the updated versions reach stable storage. After the disk is 6826 * updated, the blocks can be safely de-allocated whenever it is convenient. 6827 * This implementation handles only the common case of reducing a file's 6828 * length to zero. Other cases are handled by the conventional synchronous 6829 * write approach. 6830 * 6831 * The ffs implementation with which we worked double-checks 6832 * the state of the block pointers and file size as it reduces 6833 * a file's length. Some of this code is replicated here in our 6834 * soft updates implementation. The freeblks->fb_chkcnt field is 6835 * used to transfer a part of this information to the procedure 6836 * that eventually de-allocates the blocks. 6837 * 6838 * This routine should be called from the routine that shortens 6839 * a file's length, before the inode's size or block pointers 6840 * are modified. It will save the block pointer information for 6841 * later release and zero the inode so that the calling routine 6842 * can release it. 6843 */ 6844 void 6845 softdep_setup_freeblocks(ip, length, flags) 6846 struct inode *ip; /* The inode whose length is to be reduced */ 6847 off_t length; /* The new length for the file */ 6848 int flags; /* IO_EXT and/or IO_NORMAL */ 6849 { 6850 struct ufs1_dinode *dp1; 6851 struct ufs2_dinode *dp2; 6852 struct freeblks *freeblks; 6853 struct inodedep *inodedep; 6854 struct allocdirect *adp; 6855 struct ufsmount *ump; 6856 struct buf *bp; 6857 struct fs *fs; 6858 ufs2_daddr_t extblocks, datablocks; 6859 struct mount *mp; 6860 int i, delay, error; 6861 ufs_lbn_t tmpval; 6862 ufs_lbn_t lbn; 6863 6864 ump = ITOUMP(ip); 6865 mp = UFSTOVFS(ump); 6866 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6867 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6868 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6869 ip->i_number, length); 6870 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6871 fs = ump->um_fs; 6872 if ((error = bread(ump->um_devvp, 6873 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6874 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6875 brelse(bp); 6876 softdep_error("softdep_setup_freeblocks", error); 6877 return; 6878 } 6879 freeblks = newfreeblks(mp, ip); 6880 extblocks = 0; 6881 datablocks = 0; 6882 if (fs->fs_magic == FS_UFS2_MAGIC) 6883 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6884 if ((flags & IO_NORMAL) != 0) { 6885 for (i = 0; i < UFS_NDADDR; i++) 6886 setup_freedirect(freeblks, ip, i, 0); 6887 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6888 i < UFS_NIADDR; 6889 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6890 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6891 ip->i_size = 0; 6892 DIP_SET(ip, i_size, 0); 6893 datablocks = DIP(ip, i_blocks) - extblocks; 6894 } 6895 if ((flags & IO_EXT) != 0) { 6896 for (i = 0; i < UFS_NXADDR; i++) 6897 setup_freeext(freeblks, ip, i, 0); 6898 ip->i_din2->di_extsize = 0; 6899 datablocks += extblocks; 6900 } 6901 #ifdef QUOTA 6902 /* Reference the quotas in case the block count is wrong in the end. */ 6903 quotaref(ITOV(ip), freeblks->fb_quota); 6904 (void) chkdq(ip, -datablocks, NOCRED, 0); 6905 #endif 6906 freeblks->fb_chkcnt = -datablocks; 6907 UFS_LOCK(ump); 6908 fs->fs_pendingblocks += datablocks; 6909 UFS_UNLOCK(ump); 6910 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6911 /* 6912 * Push the zero'ed inode to its disk buffer so that we are free 6913 * to delete its dependencies below. Once the dependencies are gone 6914 * the buffer can be safely released. 6915 */ 6916 if (ump->um_fstype == UFS1) { 6917 dp1 = ((struct ufs1_dinode *)bp->b_data + 6918 ino_to_fsbo(fs, ip->i_number)); 6919 ip->i_din1->di_freelink = dp1->di_freelink; 6920 *dp1 = *ip->i_din1; 6921 } else { 6922 dp2 = ((struct ufs2_dinode *)bp->b_data + 6923 ino_to_fsbo(fs, ip->i_number)); 6924 ip->i_din2->di_freelink = dp2->di_freelink; 6925 *dp2 = *ip->i_din2; 6926 } 6927 /* 6928 * Find and eliminate any inode dependencies. 6929 */ 6930 ACQUIRE_LOCK(ump); 6931 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6932 if ((inodedep->id_state & IOSTARTED) != 0) 6933 panic("softdep_setup_freeblocks: inode busy"); 6934 /* 6935 * Add the freeblks structure to the list of operations that 6936 * must await the zero'ed inode being written to disk. If we 6937 * still have a bitmap dependency (delay == 0), then the inode 6938 * has never been written to disk, so we can process the 6939 * freeblks below once we have deleted the dependencies. 6940 */ 6941 delay = (inodedep->id_state & DEPCOMPLETE); 6942 if (delay) 6943 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6944 else 6945 freeblks->fb_state |= COMPLETE; 6946 /* 6947 * Because the file length has been truncated to zero, any 6948 * pending block allocation dependency structures associated 6949 * with this inode are obsolete and can simply be de-allocated. 6950 * We must first merge the two dependency lists to get rid of 6951 * any duplicate freefrag structures, then purge the merged list. 6952 * If we still have a bitmap dependency, then the inode has never 6953 * been written to disk, so we can free any fragments without delay. 6954 */ 6955 if (flags & IO_NORMAL) { 6956 merge_inode_lists(&inodedep->id_newinoupdt, 6957 &inodedep->id_inoupdt); 6958 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 6959 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6960 freeblks); 6961 } 6962 if (flags & IO_EXT) { 6963 merge_inode_lists(&inodedep->id_newextupdt, 6964 &inodedep->id_extupdt); 6965 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6966 cancel_allocdirect(&inodedep->id_extupdt, adp, 6967 freeblks); 6968 } 6969 FREE_LOCK(ump); 6970 bdwrite(bp); 6971 trunc_dependencies(ip, freeblks, -1, 0, flags); 6972 ACQUIRE_LOCK(ump); 6973 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6974 (void) free_inodedep(inodedep); 6975 freeblks->fb_state |= DEPCOMPLETE; 6976 /* 6977 * If the inode with zeroed block pointers is now on disk 6978 * we can start freeing blocks. 6979 */ 6980 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6981 freeblks->fb_state |= INPROGRESS; 6982 else 6983 freeblks = NULL; 6984 FREE_LOCK(ump); 6985 if (freeblks) 6986 handle_workitem_freeblocks(freeblks, 0); 6987 trunc_pages(ip, length, extblocks, flags); 6988 } 6989 6990 /* 6991 * Eliminate pages from the page cache that back parts of this inode and 6992 * adjust the vnode pager's idea of our size. This prevents stale data 6993 * from hanging around in the page cache. 6994 */ 6995 static void 6996 trunc_pages(ip, length, extblocks, flags) 6997 struct inode *ip; 6998 off_t length; 6999 ufs2_daddr_t extblocks; 7000 int flags; 7001 { 7002 struct vnode *vp; 7003 struct fs *fs; 7004 ufs_lbn_t lbn; 7005 off_t end, extend; 7006 7007 vp = ITOV(ip); 7008 fs = ITOFS(ip); 7009 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 7010 if ((flags & IO_EXT) != 0) 7011 vn_pages_remove(vp, extend, 0); 7012 if ((flags & IO_NORMAL) == 0) 7013 return; 7014 BO_LOCK(&vp->v_bufobj); 7015 drain_output(vp); 7016 BO_UNLOCK(&vp->v_bufobj); 7017 /* 7018 * The vnode pager eliminates file pages we eliminate indirects 7019 * below. 7020 */ 7021 vnode_pager_setsize(vp, length); 7022 /* 7023 * Calculate the end based on the last indirect we want to keep. If 7024 * the block extends into indirects we can just use the negative of 7025 * its lbn. Doubles and triples exist at lower numbers so we must 7026 * be careful not to remove those, if they exist. double and triple 7027 * indirect lbns do not overlap with others so it is not important 7028 * to verify how many levels are required. 7029 */ 7030 lbn = lblkno(fs, length); 7031 if (lbn >= UFS_NDADDR) { 7032 /* Calculate the virtual lbn of the triple indirect. */ 7033 lbn = -lbn - (UFS_NIADDR - 1); 7034 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7035 } else 7036 end = extend; 7037 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7038 } 7039 7040 /* 7041 * See if the buf bp is in the range eliminated by truncation. 7042 */ 7043 static int 7044 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7045 struct buf *bp; 7046 int *blkoffp; 7047 ufs_lbn_t lastlbn; 7048 int lastoff; 7049 int flags; 7050 { 7051 ufs_lbn_t lbn; 7052 7053 *blkoffp = 0; 7054 /* Only match ext/normal blocks as appropriate. */ 7055 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7056 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7057 return (0); 7058 /* ALTDATA is always a full truncation. */ 7059 if ((bp->b_xflags & BX_ALTDATA) != 0) 7060 return (1); 7061 /* -1 is full truncation. */ 7062 if (lastlbn == -1) 7063 return (1); 7064 /* 7065 * If this is a partial truncate we only want those 7066 * blocks and indirect blocks that cover the range 7067 * we're after. 7068 */ 7069 lbn = bp->b_lblkno; 7070 if (lbn < 0) 7071 lbn = -(lbn + lbn_level(lbn)); 7072 if (lbn < lastlbn) 7073 return (0); 7074 /* Here we only truncate lblkno if it's partial. */ 7075 if (lbn == lastlbn) { 7076 if (lastoff == 0) 7077 return (0); 7078 *blkoffp = lastoff; 7079 } 7080 return (1); 7081 } 7082 7083 /* 7084 * Eliminate any dependencies that exist in memory beyond lblkno:off 7085 */ 7086 static void 7087 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7088 struct inode *ip; 7089 struct freeblks *freeblks; 7090 ufs_lbn_t lastlbn; 7091 int lastoff; 7092 int flags; 7093 { 7094 struct bufobj *bo; 7095 struct vnode *vp; 7096 struct buf *bp; 7097 int blkoff; 7098 7099 /* 7100 * We must wait for any I/O in progress to finish so that 7101 * all potential buffers on the dirty list will be visible. 7102 * Once they are all there, walk the list and get rid of 7103 * any dependencies. 7104 */ 7105 vp = ITOV(ip); 7106 bo = &vp->v_bufobj; 7107 BO_LOCK(bo); 7108 drain_output(vp); 7109 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7110 bp->b_vflags &= ~BV_SCANNED; 7111 restart: 7112 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7113 if (bp->b_vflags & BV_SCANNED) 7114 continue; 7115 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7116 bp->b_vflags |= BV_SCANNED; 7117 continue; 7118 } 7119 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7120 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7121 goto restart; 7122 BO_UNLOCK(bo); 7123 if (deallocate_dependencies(bp, freeblks, blkoff)) 7124 bqrelse(bp); 7125 else 7126 brelse(bp); 7127 BO_LOCK(bo); 7128 goto restart; 7129 } 7130 /* 7131 * Now do the work of vtruncbuf while also matching indirect blocks. 7132 */ 7133 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7134 bp->b_vflags &= ~BV_SCANNED; 7135 cleanrestart: 7136 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7137 if (bp->b_vflags & BV_SCANNED) 7138 continue; 7139 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7140 bp->b_vflags |= BV_SCANNED; 7141 continue; 7142 } 7143 if (BUF_LOCK(bp, 7144 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7145 BO_LOCKPTR(bo)) == ENOLCK) { 7146 BO_LOCK(bo); 7147 goto cleanrestart; 7148 } 7149 bp->b_vflags |= BV_SCANNED; 7150 bremfree(bp); 7151 if (blkoff != 0) { 7152 allocbuf(bp, blkoff); 7153 bqrelse(bp); 7154 } else { 7155 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7156 brelse(bp); 7157 } 7158 BO_LOCK(bo); 7159 goto cleanrestart; 7160 } 7161 drain_output(vp); 7162 BO_UNLOCK(bo); 7163 } 7164 7165 static int 7166 cancel_pagedep(pagedep, freeblks, blkoff) 7167 struct pagedep *pagedep; 7168 struct freeblks *freeblks; 7169 int blkoff; 7170 { 7171 struct jremref *jremref; 7172 struct jmvref *jmvref; 7173 struct dirrem *dirrem, *tmp; 7174 int i; 7175 7176 /* 7177 * Copy any directory remove dependencies to the list 7178 * to be processed after the freeblks proceeds. If 7179 * directory entry never made it to disk they 7180 * can be dumped directly onto the work list. 7181 */ 7182 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7183 /* Skip this directory removal if it is intended to remain. */ 7184 if (dirrem->dm_offset < blkoff) 7185 continue; 7186 /* 7187 * If there are any dirrems we wait for the journal write 7188 * to complete and then restart the buf scan as the lock 7189 * has been dropped. 7190 */ 7191 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7192 jwait(&jremref->jr_list, MNT_WAIT); 7193 return (ERESTART); 7194 } 7195 LIST_REMOVE(dirrem, dm_next); 7196 dirrem->dm_dirinum = pagedep->pd_ino; 7197 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7198 } 7199 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7200 jwait(&jmvref->jm_list, MNT_WAIT); 7201 return (ERESTART); 7202 } 7203 /* 7204 * When we're partially truncating a pagedep we just want to flush 7205 * journal entries and return. There can not be any adds in the 7206 * truncated portion of the directory and newblk must remain if 7207 * part of the block remains. 7208 */ 7209 if (blkoff != 0) { 7210 struct diradd *dap; 7211 7212 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7213 if (dap->da_offset > blkoff) 7214 panic("cancel_pagedep: diradd %p off %d > %d", 7215 dap, dap->da_offset, blkoff); 7216 for (i = 0; i < DAHASHSZ; i++) 7217 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7218 if (dap->da_offset > blkoff) 7219 panic("cancel_pagedep: diradd %p off %d > %d", 7220 dap, dap->da_offset, blkoff); 7221 return (0); 7222 } 7223 /* 7224 * There should be no directory add dependencies present 7225 * as the directory could not be truncated until all 7226 * children were removed. 7227 */ 7228 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7229 ("deallocate_dependencies: pendinghd != NULL")); 7230 for (i = 0; i < DAHASHSZ; i++) 7231 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7232 ("deallocate_dependencies: diraddhd != NULL")); 7233 if ((pagedep->pd_state & NEWBLOCK) != 0) 7234 free_newdirblk(pagedep->pd_newdirblk); 7235 if (free_pagedep(pagedep) == 0) 7236 panic("Failed to free pagedep %p", pagedep); 7237 return (0); 7238 } 7239 7240 /* 7241 * Reclaim any dependency structures from a buffer that is about to 7242 * be reallocated to a new vnode. The buffer must be locked, thus, 7243 * no I/O completion operations can occur while we are manipulating 7244 * its associated dependencies. The mutex is held so that other I/O's 7245 * associated with related dependencies do not occur. 7246 */ 7247 static int 7248 deallocate_dependencies(bp, freeblks, off) 7249 struct buf *bp; 7250 struct freeblks *freeblks; 7251 int off; 7252 { 7253 struct indirdep *indirdep; 7254 struct pagedep *pagedep; 7255 struct worklist *wk, *wkn; 7256 struct ufsmount *ump; 7257 7258 ump = softdep_bp_to_mp(bp); 7259 if (ump == NULL) 7260 goto done; 7261 ACQUIRE_LOCK(ump); 7262 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7263 switch (wk->wk_type) { 7264 case D_INDIRDEP: 7265 indirdep = WK_INDIRDEP(wk); 7266 if (bp->b_lblkno >= 0 || 7267 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7268 panic("deallocate_dependencies: not indir"); 7269 cancel_indirdep(indirdep, bp, freeblks); 7270 continue; 7271 7272 case D_PAGEDEP: 7273 pagedep = WK_PAGEDEP(wk); 7274 if (cancel_pagedep(pagedep, freeblks, off)) { 7275 FREE_LOCK(ump); 7276 return (ERESTART); 7277 } 7278 continue; 7279 7280 case D_ALLOCINDIR: 7281 /* 7282 * Simply remove the allocindir, we'll find it via 7283 * the indirdep where we can clear pointers if 7284 * needed. 7285 */ 7286 WORKLIST_REMOVE(wk); 7287 continue; 7288 7289 case D_FREEWORK: 7290 /* 7291 * A truncation is waiting for the zero'd pointers 7292 * to be written. It can be freed when the freeblks 7293 * is journaled. 7294 */ 7295 WORKLIST_REMOVE(wk); 7296 wk->wk_state |= ONDEPLIST; 7297 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7298 break; 7299 7300 case D_ALLOCDIRECT: 7301 if (off != 0) 7302 continue; 7303 /* FALLTHROUGH */ 7304 default: 7305 panic("deallocate_dependencies: Unexpected type %s", 7306 TYPENAME(wk->wk_type)); 7307 /* NOTREACHED */ 7308 } 7309 } 7310 FREE_LOCK(ump); 7311 done: 7312 /* 7313 * Don't throw away this buf, we were partially truncating and 7314 * some deps may always remain. 7315 */ 7316 if (off) { 7317 allocbuf(bp, off); 7318 bp->b_vflags |= BV_SCANNED; 7319 return (EBUSY); 7320 } 7321 bp->b_flags |= B_INVAL | B_NOCACHE; 7322 7323 return (0); 7324 } 7325 7326 /* 7327 * An allocdirect is being canceled due to a truncate. We must make sure 7328 * the journal entry is released in concert with the blkfree that releases 7329 * the storage. Completed journal entries must not be released until the 7330 * space is no longer pointed to by the inode or in the bitmap. 7331 */ 7332 static void 7333 cancel_allocdirect(adphead, adp, freeblks) 7334 struct allocdirectlst *adphead; 7335 struct allocdirect *adp; 7336 struct freeblks *freeblks; 7337 { 7338 struct freework *freework; 7339 struct newblk *newblk; 7340 struct worklist *wk; 7341 7342 TAILQ_REMOVE(adphead, adp, ad_next); 7343 newblk = (struct newblk *)adp; 7344 freework = NULL; 7345 /* 7346 * Find the correct freework structure. 7347 */ 7348 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7349 if (wk->wk_type != D_FREEWORK) 7350 continue; 7351 freework = WK_FREEWORK(wk); 7352 if (freework->fw_blkno == newblk->nb_newblkno) 7353 break; 7354 } 7355 if (freework == NULL) 7356 panic("cancel_allocdirect: Freework not found"); 7357 /* 7358 * If a newblk exists at all we still have the journal entry that 7359 * initiated the allocation so we do not need to journal the free. 7360 */ 7361 cancel_jfreeblk(freeblks, freework->fw_blkno); 7362 /* 7363 * If the journal hasn't been written the jnewblk must be passed 7364 * to the call to ffs_blkfree that reclaims the space. We accomplish 7365 * this by linking the journal dependency into the freework to be 7366 * freed when freework_freeblock() is called. If the journal has 7367 * been written we can simply reclaim the journal space when the 7368 * freeblks work is complete. 7369 */ 7370 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7371 &freeblks->fb_jwork); 7372 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7373 } 7374 7375 7376 /* 7377 * Cancel a new block allocation. May be an indirect or direct block. We 7378 * remove it from various lists and return any journal record that needs to 7379 * be resolved by the caller. 7380 * 7381 * A special consideration is made for indirects which were never pointed 7382 * at on disk and will never be found once this block is released. 7383 */ 7384 static struct jnewblk * 7385 cancel_newblk(newblk, wk, wkhd) 7386 struct newblk *newblk; 7387 struct worklist *wk; 7388 struct workhead *wkhd; 7389 { 7390 struct jnewblk *jnewblk; 7391 7392 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7393 7394 newblk->nb_state |= GOINGAWAY; 7395 /* 7396 * Previously we traversed the completedhd on each indirdep 7397 * attached to this newblk to cancel them and gather journal 7398 * work. Since we need only the oldest journal segment and 7399 * the lowest point on the tree will always have the oldest 7400 * journal segment we are free to release the segments 7401 * of any subordinates and may leave the indirdep list to 7402 * indirdep_complete() when this newblk is freed. 7403 */ 7404 if (newblk->nb_state & ONDEPLIST) { 7405 newblk->nb_state &= ~ONDEPLIST; 7406 LIST_REMOVE(newblk, nb_deps); 7407 } 7408 if (newblk->nb_state & ONWORKLIST) 7409 WORKLIST_REMOVE(&newblk->nb_list); 7410 /* 7411 * If the journal entry hasn't been written we save a pointer to 7412 * the dependency that frees it until it is written or the 7413 * superseding operation completes. 7414 */ 7415 jnewblk = newblk->nb_jnewblk; 7416 if (jnewblk != NULL && wk != NULL) { 7417 newblk->nb_jnewblk = NULL; 7418 jnewblk->jn_dep = wk; 7419 } 7420 if (!LIST_EMPTY(&newblk->nb_jwork)) 7421 jwork_move(wkhd, &newblk->nb_jwork); 7422 /* 7423 * When truncating we must free the newdirblk early to remove 7424 * the pagedep from the hash before returning. 7425 */ 7426 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7427 free_newdirblk(WK_NEWDIRBLK(wk)); 7428 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7429 panic("cancel_newblk: extra newdirblk"); 7430 7431 return (jnewblk); 7432 } 7433 7434 /* 7435 * Schedule the freefrag associated with a newblk to be released once 7436 * the pointers are written and the previous block is no longer needed. 7437 */ 7438 static void 7439 newblk_freefrag(newblk) 7440 struct newblk *newblk; 7441 { 7442 struct freefrag *freefrag; 7443 7444 if (newblk->nb_freefrag == NULL) 7445 return; 7446 freefrag = newblk->nb_freefrag; 7447 newblk->nb_freefrag = NULL; 7448 freefrag->ff_state |= COMPLETE; 7449 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7450 add_to_worklist(&freefrag->ff_list, 0); 7451 } 7452 7453 /* 7454 * Free a newblk. Generate a new freefrag work request if appropriate. 7455 * This must be called after the inode pointer and any direct block pointers 7456 * are valid or fully removed via truncate or frag extension. 7457 */ 7458 static void 7459 free_newblk(newblk) 7460 struct newblk *newblk; 7461 { 7462 struct indirdep *indirdep; 7463 struct worklist *wk; 7464 7465 KASSERT(newblk->nb_jnewblk == NULL, 7466 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7467 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7468 ("free_newblk: unclaimed newblk")); 7469 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7470 newblk_freefrag(newblk); 7471 if (newblk->nb_state & ONDEPLIST) 7472 LIST_REMOVE(newblk, nb_deps); 7473 if (newblk->nb_state & ONWORKLIST) 7474 WORKLIST_REMOVE(&newblk->nb_list); 7475 LIST_REMOVE(newblk, nb_hash); 7476 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7477 free_newdirblk(WK_NEWDIRBLK(wk)); 7478 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7479 panic("free_newblk: extra newdirblk"); 7480 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7481 indirdep_complete(indirdep); 7482 handle_jwork(&newblk->nb_jwork); 7483 WORKITEM_FREE(newblk, D_NEWBLK); 7484 } 7485 7486 /* 7487 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7488 * This routine must be called with splbio interrupts blocked. 7489 */ 7490 static void 7491 free_newdirblk(newdirblk) 7492 struct newdirblk *newdirblk; 7493 { 7494 struct pagedep *pagedep; 7495 struct diradd *dap; 7496 struct worklist *wk; 7497 7498 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7499 WORKLIST_REMOVE(&newdirblk->db_list); 7500 /* 7501 * If the pagedep is still linked onto the directory buffer 7502 * dependency chain, then some of the entries on the 7503 * pd_pendinghd list may not be committed to disk yet. In 7504 * this case, we will simply clear the NEWBLOCK flag and 7505 * let the pd_pendinghd list be processed when the pagedep 7506 * is next written. If the pagedep is no longer on the buffer 7507 * dependency chain, then all the entries on the pd_pending 7508 * list are committed to disk and we can free them here. 7509 */ 7510 pagedep = newdirblk->db_pagedep; 7511 pagedep->pd_state &= ~NEWBLOCK; 7512 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7513 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7514 free_diradd(dap, NULL); 7515 /* 7516 * If no dependencies remain, the pagedep will be freed. 7517 */ 7518 free_pagedep(pagedep); 7519 } 7520 /* Should only ever be one item in the list. */ 7521 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7522 WORKLIST_REMOVE(wk); 7523 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7524 } 7525 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7526 } 7527 7528 /* 7529 * Prepare an inode to be freed. The actual free operation is not 7530 * done until the zero'ed inode has been written to disk. 7531 */ 7532 void 7533 softdep_freefile(pvp, ino, mode) 7534 struct vnode *pvp; 7535 ino_t ino; 7536 int mode; 7537 { 7538 struct inode *ip = VTOI(pvp); 7539 struct inodedep *inodedep; 7540 struct freefile *freefile; 7541 struct freeblks *freeblks; 7542 struct ufsmount *ump; 7543 7544 ump = ITOUMP(ip); 7545 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7546 ("softdep_freefile called on non-softdep filesystem")); 7547 /* 7548 * This sets up the inode de-allocation dependency. 7549 */ 7550 freefile = malloc(sizeof(struct freefile), 7551 M_FREEFILE, M_SOFTDEP_FLAGS); 7552 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7553 freefile->fx_mode = mode; 7554 freefile->fx_oldinum = ino; 7555 freefile->fx_devvp = ump->um_devvp; 7556 LIST_INIT(&freefile->fx_jwork); 7557 UFS_LOCK(ump); 7558 ump->um_fs->fs_pendinginodes += 1; 7559 UFS_UNLOCK(ump); 7560 7561 /* 7562 * If the inodedep does not exist, then the zero'ed inode has 7563 * been written to disk. If the allocated inode has never been 7564 * written to disk, then the on-disk inode is zero'ed. In either 7565 * case we can free the file immediately. If the journal was 7566 * canceled before being written the inode will never make it to 7567 * disk and we must send the canceled journal entrys to 7568 * ffs_freefile() to be cleared in conjunction with the bitmap. 7569 * Any blocks waiting on the inode to write can be safely freed 7570 * here as it will never been written. 7571 */ 7572 ACQUIRE_LOCK(ump); 7573 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7574 if (inodedep) { 7575 /* 7576 * Clear out freeblks that no longer need to reference 7577 * this inode. 7578 */ 7579 while ((freeblks = 7580 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7581 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7582 fb_next); 7583 freeblks->fb_state &= ~ONDEPLIST; 7584 } 7585 /* 7586 * Remove this inode from the unlinked list. 7587 */ 7588 if (inodedep->id_state & UNLINKED) { 7589 /* 7590 * Save the journal work to be freed with the bitmap 7591 * before we clear UNLINKED. Otherwise it can be lost 7592 * if the inode block is written. 7593 */ 7594 handle_bufwait(inodedep, &freefile->fx_jwork); 7595 clear_unlinked_inodedep(inodedep); 7596 /* 7597 * Re-acquire inodedep as we've dropped the 7598 * per-filesystem lock in clear_unlinked_inodedep(). 7599 */ 7600 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7601 } 7602 } 7603 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7604 FREE_LOCK(ump); 7605 handle_workitem_freefile(freefile); 7606 return; 7607 } 7608 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7609 inodedep->id_state |= GOINGAWAY; 7610 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7611 FREE_LOCK(ump); 7612 if (ip->i_number == ino) 7613 ip->i_flag |= IN_MODIFIED; 7614 } 7615 7616 /* 7617 * Check to see if an inode has never been written to disk. If 7618 * so free the inodedep and return success, otherwise return failure. 7619 * This routine must be called with splbio interrupts blocked. 7620 * 7621 * If we still have a bitmap dependency, then the inode has never 7622 * been written to disk. Drop the dependency as it is no longer 7623 * necessary since the inode is being deallocated. We set the 7624 * ALLCOMPLETE flags since the bitmap now properly shows that the 7625 * inode is not allocated. Even if the inode is actively being 7626 * written, it has been rolled back to its zero'ed state, so we 7627 * are ensured that a zero inode is what is on the disk. For short 7628 * lived files, this change will usually result in removing all the 7629 * dependencies from the inode so that it can be freed immediately. 7630 */ 7631 static int 7632 check_inode_unwritten(inodedep) 7633 struct inodedep *inodedep; 7634 { 7635 7636 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7637 7638 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7639 !LIST_EMPTY(&inodedep->id_dirremhd) || 7640 !LIST_EMPTY(&inodedep->id_pendinghd) || 7641 !LIST_EMPTY(&inodedep->id_bufwait) || 7642 !LIST_EMPTY(&inodedep->id_inowait) || 7643 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7644 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7645 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7646 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7647 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7648 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7649 inodedep->id_mkdiradd != NULL || 7650 inodedep->id_nlinkdelta != 0) 7651 return (0); 7652 /* 7653 * Another process might be in initiate_write_inodeblock_ufs[12] 7654 * trying to allocate memory without holding "Softdep Lock". 7655 */ 7656 if ((inodedep->id_state & IOSTARTED) != 0 && 7657 inodedep->id_savedino1 == NULL) 7658 return (0); 7659 7660 if (inodedep->id_state & ONDEPLIST) 7661 LIST_REMOVE(inodedep, id_deps); 7662 inodedep->id_state &= ~ONDEPLIST; 7663 inodedep->id_state |= ALLCOMPLETE; 7664 inodedep->id_bmsafemap = NULL; 7665 if (inodedep->id_state & ONWORKLIST) 7666 WORKLIST_REMOVE(&inodedep->id_list); 7667 if (inodedep->id_savedino1 != NULL) { 7668 free(inodedep->id_savedino1, M_SAVEDINO); 7669 inodedep->id_savedino1 = NULL; 7670 } 7671 if (free_inodedep(inodedep) == 0) 7672 panic("check_inode_unwritten: busy inode"); 7673 return (1); 7674 } 7675 7676 static int 7677 check_inodedep_free(inodedep) 7678 struct inodedep *inodedep; 7679 { 7680 7681 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7682 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7683 !LIST_EMPTY(&inodedep->id_dirremhd) || 7684 !LIST_EMPTY(&inodedep->id_pendinghd) || 7685 !LIST_EMPTY(&inodedep->id_bufwait) || 7686 !LIST_EMPTY(&inodedep->id_inowait) || 7687 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7688 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7689 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7690 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7691 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7692 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7693 inodedep->id_mkdiradd != NULL || 7694 inodedep->id_nlinkdelta != 0 || 7695 inodedep->id_savedino1 != NULL) 7696 return (0); 7697 return (1); 7698 } 7699 7700 /* 7701 * Try to free an inodedep structure. Return 1 if it could be freed. 7702 */ 7703 static int 7704 free_inodedep(inodedep) 7705 struct inodedep *inodedep; 7706 { 7707 7708 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7709 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7710 !check_inodedep_free(inodedep)) 7711 return (0); 7712 if (inodedep->id_state & ONDEPLIST) 7713 LIST_REMOVE(inodedep, id_deps); 7714 LIST_REMOVE(inodedep, id_hash); 7715 WORKITEM_FREE(inodedep, D_INODEDEP); 7716 return (1); 7717 } 7718 7719 /* 7720 * Free the block referenced by a freework structure. The parent freeblks 7721 * structure is released and completed when the final cg bitmap reaches 7722 * the disk. This routine may be freeing a jnewblk which never made it to 7723 * disk in which case we do not have to wait as the operation is undone 7724 * in memory immediately. 7725 */ 7726 static void 7727 freework_freeblock(freework) 7728 struct freework *freework; 7729 { 7730 struct freeblks *freeblks; 7731 struct jnewblk *jnewblk; 7732 struct ufsmount *ump; 7733 struct workhead wkhd; 7734 struct fs *fs; 7735 int bsize; 7736 int needj; 7737 7738 ump = VFSTOUFS(freework->fw_list.wk_mp); 7739 LOCK_OWNED(ump); 7740 /* 7741 * Handle partial truncate separately. 7742 */ 7743 if (freework->fw_indir) { 7744 complete_trunc_indir(freework); 7745 return; 7746 } 7747 freeblks = freework->fw_freeblks; 7748 fs = ump->um_fs; 7749 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7750 bsize = lfragtosize(fs, freework->fw_frags); 7751 LIST_INIT(&wkhd); 7752 /* 7753 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7754 * on the indirblk hashtable and prevents premature freeing. 7755 */ 7756 freework->fw_state |= DEPCOMPLETE; 7757 /* 7758 * SUJ needs to wait for the segment referencing freed indirect 7759 * blocks to expire so that we know the checker will not confuse 7760 * a re-allocated indirect block with its old contents. 7761 */ 7762 if (needj && freework->fw_lbn <= -UFS_NDADDR) 7763 indirblk_insert(freework); 7764 /* 7765 * If we are canceling an existing jnewblk pass it to the free 7766 * routine, otherwise pass the freeblk which will ultimately 7767 * release the freeblks. If we're not journaling, we can just 7768 * free the freeblks immediately. 7769 */ 7770 jnewblk = freework->fw_jnewblk; 7771 if (jnewblk != NULL) { 7772 cancel_jnewblk(jnewblk, &wkhd); 7773 needj = 0; 7774 } else if (needj) { 7775 freework->fw_state |= DELAYEDFREE; 7776 freeblks->fb_cgwait++; 7777 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7778 } 7779 FREE_LOCK(ump); 7780 freeblks_free(ump, freeblks, btodb(bsize)); 7781 CTR4(KTR_SUJ, 7782 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld", 7783 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7784 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7785 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7786 ACQUIRE_LOCK(ump); 7787 /* 7788 * The jnewblk will be discarded and the bits in the map never 7789 * made it to disk. We can immediately free the freeblk. 7790 */ 7791 if (needj == 0) 7792 handle_written_freework(freework); 7793 } 7794 7795 /* 7796 * We enqueue freework items that need processing back on the freeblks and 7797 * add the freeblks to the worklist. This makes it easier to find all work 7798 * required to flush a truncation in process_truncates(). 7799 */ 7800 static void 7801 freework_enqueue(freework) 7802 struct freework *freework; 7803 { 7804 struct freeblks *freeblks; 7805 7806 freeblks = freework->fw_freeblks; 7807 if ((freework->fw_state & INPROGRESS) == 0) 7808 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7809 if ((freeblks->fb_state & 7810 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7811 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7812 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7813 } 7814 7815 /* 7816 * Start, continue, or finish the process of freeing an indirect block tree. 7817 * The free operation may be paused at any point with fw_off containing the 7818 * offset to restart from. This enables us to implement some flow control 7819 * for large truncates which may fan out and generate a huge number of 7820 * dependencies. 7821 */ 7822 static void 7823 handle_workitem_indirblk(freework) 7824 struct freework *freework; 7825 { 7826 struct freeblks *freeblks; 7827 struct ufsmount *ump; 7828 struct fs *fs; 7829 7830 freeblks = freework->fw_freeblks; 7831 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7832 fs = ump->um_fs; 7833 if (freework->fw_state & DEPCOMPLETE) { 7834 handle_written_freework(freework); 7835 return; 7836 } 7837 if (freework->fw_off == NINDIR(fs)) { 7838 freework_freeblock(freework); 7839 return; 7840 } 7841 freework->fw_state |= INPROGRESS; 7842 FREE_LOCK(ump); 7843 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7844 freework->fw_lbn); 7845 ACQUIRE_LOCK(ump); 7846 } 7847 7848 /* 7849 * Called when a freework structure attached to a cg buf is written. The 7850 * ref on either the parent or the freeblks structure is released and 7851 * the freeblks is added back to the worklist if there is more work to do. 7852 */ 7853 static void 7854 handle_written_freework(freework) 7855 struct freework *freework; 7856 { 7857 struct freeblks *freeblks; 7858 struct freework *parent; 7859 7860 freeblks = freework->fw_freeblks; 7861 parent = freework->fw_parent; 7862 if (freework->fw_state & DELAYEDFREE) 7863 freeblks->fb_cgwait--; 7864 freework->fw_state |= COMPLETE; 7865 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7866 WORKITEM_FREE(freework, D_FREEWORK); 7867 if (parent) { 7868 if (--parent->fw_ref == 0) 7869 freework_enqueue(parent); 7870 return; 7871 } 7872 if (--freeblks->fb_ref != 0) 7873 return; 7874 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7875 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7876 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7877 } 7878 7879 /* 7880 * This workitem routine performs the block de-allocation. 7881 * The workitem is added to the pending list after the updated 7882 * inode block has been written to disk. As mentioned above, 7883 * checks regarding the number of blocks de-allocated (compared 7884 * to the number of blocks allocated for the file) are also 7885 * performed in this function. 7886 */ 7887 static int 7888 handle_workitem_freeblocks(freeblks, flags) 7889 struct freeblks *freeblks; 7890 int flags; 7891 { 7892 struct freework *freework; 7893 struct newblk *newblk; 7894 struct allocindir *aip; 7895 struct ufsmount *ump; 7896 struct worklist *wk; 7897 7898 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7899 ("handle_workitem_freeblocks: Journal entries not written.")); 7900 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7901 ACQUIRE_LOCK(ump); 7902 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7903 WORKLIST_REMOVE(wk); 7904 switch (wk->wk_type) { 7905 case D_DIRREM: 7906 wk->wk_state |= COMPLETE; 7907 add_to_worklist(wk, 0); 7908 continue; 7909 7910 case D_ALLOCDIRECT: 7911 free_newblk(WK_NEWBLK(wk)); 7912 continue; 7913 7914 case D_ALLOCINDIR: 7915 aip = WK_ALLOCINDIR(wk); 7916 freework = NULL; 7917 if (aip->ai_state & DELAYEDFREE) { 7918 FREE_LOCK(ump); 7919 freework = newfreework(ump, freeblks, NULL, 7920 aip->ai_lbn, aip->ai_newblkno, 7921 ump->um_fs->fs_frag, 0, 0); 7922 ACQUIRE_LOCK(ump); 7923 } 7924 newblk = WK_NEWBLK(wk); 7925 if (newblk->nb_jnewblk) { 7926 freework->fw_jnewblk = newblk->nb_jnewblk; 7927 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7928 newblk->nb_jnewblk = NULL; 7929 } 7930 free_newblk(newblk); 7931 continue; 7932 7933 case D_FREEWORK: 7934 freework = WK_FREEWORK(wk); 7935 if (freework->fw_lbn <= -UFS_NDADDR) 7936 handle_workitem_indirblk(freework); 7937 else 7938 freework_freeblock(freework); 7939 continue; 7940 default: 7941 panic("handle_workitem_freeblocks: Unknown type %s", 7942 TYPENAME(wk->wk_type)); 7943 } 7944 } 7945 if (freeblks->fb_ref != 0) { 7946 freeblks->fb_state &= ~INPROGRESS; 7947 wake_worklist(&freeblks->fb_list); 7948 freeblks = NULL; 7949 } 7950 FREE_LOCK(ump); 7951 if (freeblks) 7952 return handle_complete_freeblocks(freeblks, flags); 7953 return (0); 7954 } 7955 7956 /* 7957 * Handle completion of block free via truncate. This allows fs_pending 7958 * to track the actual free block count more closely than if we only updated 7959 * it at the end. We must be careful to handle cases where the block count 7960 * on free was incorrect. 7961 */ 7962 static void 7963 freeblks_free(ump, freeblks, blocks) 7964 struct ufsmount *ump; 7965 struct freeblks *freeblks; 7966 int blocks; 7967 { 7968 struct fs *fs; 7969 ufs2_daddr_t remain; 7970 7971 UFS_LOCK(ump); 7972 remain = -freeblks->fb_chkcnt; 7973 freeblks->fb_chkcnt += blocks; 7974 if (remain > 0) { 7975 if (remain < blocks) 7976 blocks = remain; 7977 fs = ump->um_fs; 7978 fs->fs_pendingblocks -= blocks; 7979 } 7980 UFS_UNLOCK(ump); 7981 } 7982 7983 /* 7984 * Once all of the freework workitems are complete we can retire the 7985 * freeblocks dependency and any journal work awaiting completion. This 7986 * can not be called until all other dependencies are stable on disk. 7987 */ 7988 static int 7989 handle_complete_freeblocks(freeblks, flags) 7990 struct freeblks *freeblks; 7991 int flags; 7992 { 7993 struct inodedep *inodedep; 7994 struct inode *ip; 7995 struct vnode *vp; 7996 struct fs *fs; 7997 struct ufsmount *ump; 7998 ufs2_daddr_t spare; 7999 8000 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8001 fs = ump->um_fs; 8002 flags = LK_EXCLUSIVE | flags; 8003 spare = freeblks->fb_chkcnt; 8004 8005 /* 8006 * If we did not release the expected number of blocks we may have 8007 * to adjust the inode block count here. Only do so if it wasn't 8008 * a truncation to zero and the modrev still matches. 8009 */ 8010 if (spare && freeblks->fb_len != 0) { 8011 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8012 flags, &vp, FFSV_FORCEINSMQ) != 0) 8013 return (EBUSY); 8014 ip = VTOI(vp); 8015 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 8016 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 8017 ip->i_flag |= IN_CHANGE; 8018 /* 8019 * We must wait so this happens before the 8020 * journal is reclaimed. 8021 */ 8022 ffs_update(vp, 1); 8023 } 8024 vput(vp); 8025 } 8026 if (spare < 0) { 8027 UFS_LOCK(ump); 8028 fs->fs_pendingblocks += spare; 8029 UFS_UNLOCK(ump); 8030 } 8031 #ifdef QUOTA 8032 /* Handle spare. */ 8033 if (spare) 8034 quotaadj(freeblks->fb_quota, ump, -spare); 8035 quotarele(freeblks->fb_quota); 8036 #endif 8037 ACQUIRE_LOCK(ump); 8038 if (freeblks->fb_state & ONDEPLIST) { 8039 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8040 0, &inodedep); 8041 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8042 freeblks->fb_state &= ~ONDEPLIST; 8043 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8044 free_inodedep(inodedep); 8045 } 8046 /* 8047 * All of the freeblock deps must be complete prior to this call 8048 * so it's now safe to complete earlier outstanding journal entries. 8049 */ 8050 handle_jwork(&freeblks->fb_jwork); 8051 WORKITEM_FREE(freeblks, D_FREEBLKS); 8052 FREE_LOCK(ump); 8053 return (0); 8054 } 8055 8056 /* 8057 * Release blocks associated with the freeblks and stored in the indirect 8058 * block dbn. If level is greater than SINGLE, the block is an indirect block 8059 * and recursive calls to indirtrunc must be used to cleanse other indirect 8060 * blocks. 8061 * 8062 * This handles partial and complete truncation of blocks. Partial is noted 8063 * with goingaway == 0. In this case the freework is completed after the 8064 * zero'd indirects are written to disk. For full truncation the freework 8065 * is completed after the block is freed. 8066 */ 8067 static void 8068 indir_trunc(freework, dbn, lbn) 8069 struct freework *freework; 8070 ufs2_daddr_t dbn; 8071 ufs_lbn_t lbn; 8072 { 8073 struct freework *nfreework; 8074 struct workhead wkhd; 8075 struct freeblks *freeblks; 8076 struct buf *bp; 8077 struct fs *fs; 8078 struct indirdep *indirdep; 8079 struct ufsmount *ump; 8080 ufs1_daddr_t *bap1; 8081 ufs2_daddr_t nb, nnb, *bap2; 8082 ufs_lbn_t lbnadd, nlbn; 8083 int i, nblocks, ufs1fmt; 8084 int freedblocks; 8085 int goingaway; 8086 int freedeps; 8087 int needj; 8088 int level; 8089 int cnt; 8090 8091 freeblks = freework->fw_freeblks; 8092 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8093 fs = ump->um_fs; 8094 /* 8095 * Get buffer of block pointers to be freed. There are three cases: 8096 * 8097 * 1) Partial truncate caches the indirdep pointer in the freework 8098 * which provides us a back copy to the save bp which holds the 8099 * pointers we want to clear. When this completes the zero 8100 * pointers are written to the real copy. 8101 * 2) The indirect is being completely truncated, cancel_indirdep() 8102 * eliminated the real copy and placed the indirdep on the saved 8103 * copy. The indirdep and buf are discarded when this completes. 8104 * 3) The indirect was not in memory, we read a copy off of the disk 8105 * using the devvp and drop and invalidate the buffer when we're 8106 * done. 8107 */ 8108 goingaway = 1; 8109 indirdep = NULL; 8110 if (freework->fw_indir != NULL) { 8111 goingaway = 0; 8112 indirdep = freework->fw_indir; 8113 bp = indirdep->ir_savebp; 8114 if (bp == NULL || bp->b_blkno != dbn) 8115 panic("indir_trunc: Bad saved buf %p blkno %jd", 8116 bp, (intmax_t)dbn); 8117 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8118 /* 8119 * The lock prevents the buf dep list from changing and 8120 * indirects on devvp should only ever have one dependency. 8121 */ 8122 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8123 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8124 panic("indir_trunc: Bad indirdep %p from buf %p", 8125 indirdep, bp); 8126 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8127 NOCRED, &bp) != 0) { 8128 brelse(bp); 8129 return; 8130 } 8131 ACQUIRE_LOCK(ump); 8132 /* Protects against a race with complete_trunc_indir(). */ 8133 freework->fw_state &= ~INPROGRESS; 8134 /* 8135 * If we have an indirdep we need to enforce the truncation order 8136 * and discard it when it is complete. 8137 */ 8138 if (indirdep) { 8139 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8140 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8141 /* 8142 * Add the complete truncate to the list on the 8143 * indirdep to enforce in-order processing. 8144 */ 8145 if (freework->fw_indir == NULL) 8146 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8147 freework, fw_next); 8148 FREE_LOCK(ump); 8149 return; 8150 } 8151 /* 8152 * If we're goingaway, free the indirdep. Otherwise it will 8153 * linger until the write completes. 8154 */ 8155 if (goingaway) 8156 free_indirdep(indirdep); 8157 } 8158 FREE_LOCK(ump); 8159 /* Initialize pointers depending on block size. */ 8160 if (ump->um_fstype == UFS1) { 8161 bap1 = (ufs1_daddr_t *)bp->b_data; 8162 nb = bap1[freework->fw_off]; 8163 ufs1fmt = 1; 8164 bap2 = NULL; 8165 } else { 8166 bap2 = (ufs2_daddr_t *)bp->b_data; 8167 nb = bap2[freework->fw_off]; 8168 ufs1fmt = 0; 8169 bap1 = NULL; 8170 } 8171 level = lbn_level(lbn); 8172 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8173 lbnadd = lbn_offset(fs, level); 8174 nblocks = btodb(fs->fs_bsize); 8175 nfreework = freework; 8176 freedeps = 0; 8177 cnt = 0; 8178 /* 8179 * Reclaim blocks. Traverses into nested indirect levels and 8180 * arranges for the current level to be freed when subordinates 8181 * are free when journaling. 8182 */ 8183 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8184 if (i != NINDIR(fs) - 1) { 8185 if (ufs1fmt) 8186 nnb = bap1[i+1]; 8187 else 8188 nnb = bap2[i+1]; 8189 } else 8190 nnb = 0; 8191 if (nb == 0) 8192 continue; 8193 cnt++; 8194 if (level != 0) { 8195 nlbn = (lbn + 1) - (i * lbnadd); 8196 if (needj != 0) { 8197 nfreework = newfreework(ump, freeblks, freework, 8198 nlbn, nb, fs->fs_frag, 0, 0); 8199 freedeps++; 8200 } 8201 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8202 } else { 8203 struct freedep *freedep; 8204 8205 /* 8206 * Attempt to aggregate freedep dependencies for 8207 * all blocks being released to the same CG. 8208 */ 8209 LIST_INIT(&wkhd); 8210 if (needj != 0 && 8211 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8212 freedep = newfreedep(freework); 8213 WORKLIST_INSERT_UNLOCKED(&wkhd, 8214 &freedep->fd_list); 8215 freedeps++; 8216 } 8217 CTR3(KTR_SUJ, 8218 "indir_trunc: ino %d blkno %jd size %ld", 8219 freeblks->fb_inum, nb, fs->fs_bsize); 8220 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8221 fs->fs_bsize, freeblks->fb_inum, 8222 freeblks->fb_vtype, &wkhd); 8223 } 8224 } 8225 if (goingaway) { 8226 bp->b_flags |= B_INVAL | B_NOCACHE; 8227 brelse(bp); 8228 } 8229 freedblocks = 0; 8230 if (level == 0) 8231 freedblocks = (nblocks * cnt); 8232 if (needj == 0) 8233 freedblocks += nblocks; 8234 freeblks_free(ump, freeblks, freedblocks); 8235 /* 8236 * If we are journaling set up the ref counts and offset so this 8237 * indirect can be completed when its children are free. 8238 */ 8239 if (needj) { 8240 ACQUIRE_LOCK(ump); 8241 freework->fw_off = i; 8242 freework->fw_ref += freedeps; 8243 freework->fw_ref -= NINDIR(fs) + 1; 8244 if (level == 0) 8245 freeblks->fb_cgwait += freedeps; 8246 if (freework->fw_ref == 0) 8247 freework_freeblock(freework); 8248 FREE_LOCK(ump); 8249 return; 8250 } 8251 /* 8252 * If we're not journaling we can free the indirect now. 8253 */ 8254 dbn = dbtofsb(fs, dbn); 8255 CTR3(KTR_SUJ, 8256 "indir_trunc 2: ino %d blkno %jd size %ld", 8257 freeblks->fb_inum, dbn, fs->fs_bsize); 8258 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8259 freeblks->fb_inum, freeblks->fb_vtype, NULL); 8260 /* Non SUJ softdep does single-threaded truncations. */ 8261 if (freework->fw_blkno == dbn) { 8262 freework->fw_state |= ALLCOMPLETE; 8263 ACQUIRE_LOCK(ump); 8264 handle_written_freework(freework); 8265 FREE_LOCK(ump); 8266 } 8267 return; 8268 } 8269 8270 /* 8271 * Cancel an allocindir when it is removed via truncation. When bp is not 8272 * NULL the indirect never appeared on disk and is scheduled to be freed 8273 * independently of the indir so we can more easily track journal work. 8274 */ 8275 static void 8276 cancel_allocindir(aip, bp, freeblks, trunc) 8277 struct allocindir *aip; 8278 struct buf *bp; 8279 struct freeblks *freeblks; 8280 int trunc; 8281 { 8282 struct indirdep *indirdep; 8283 struct freefrag *freefrag; 8284 struct newblk *newblk; 8285 8286 newblk = (struct newblk *)aip; 8287 LIST_REMOVE(aip, ai_next); 8288 /* 8289 * We must eliminate the pointer in bp if it must be freed on its 8290 * own due to partial truncate or pending journal work. 8291 */ 8292 if (bp && (trunc || newblk->nb_jnewblk)) { 8293 /* 8294 * Clear the pointer and mark the aip to be freed 8295 * directly if it never existed on disk. 8296 */ 8297 aip->ai_state |= DELAYEDFREE; 8298 indirdep = aip->ai_indirdep; 8299 if (indirdep->ir_state & UFS1FMT) 8300 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8301 else 8302 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8303 } 8304 /* 8305 * When truncating the previous pointer will be freed via 8306 * savedbp. Eliminate the freefrag which would dup free. 8307 */ 8308 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8309 newblk->nb_freefrag = NULL; 8310 if (freefrag->ff_jdep) 8311 cancel_jfreefrag( 8312 WK_JFREEFRAG(freefrag->ff_jdep)); 8313 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8314 WORKITEM_FREE(freefrag, D_FREEFRAG); 8315 } 8316 /* 8317 * If the journal hasn't been written the jnewblk must be passed 8318 * to the call to ffs_blkfree that reclaims the space. We accomplish 8319 * this by leaving the journal dependency on the newblk to be freed 8320 * when a freework is created in handle_workitem_freeblocks(). 8321 */ 8322 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8323 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8324 } 8325 8326 /* 8327 * Create the mkdir dependencies for . and .. in a new directory. Link them 8328 * in to a newdirblk so any subsequent additions are tracked properly. The 8329 * caller is responsible for adding the mkdir1 dependency to the journal 8330 * and updating id_mkdiradd. This function returns with the per-filesystem 8331 * lock held. 8332 */ 8333 static struct mkdir * 8334 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8335 struct diradd *dap; 8336 ino_t newinum; 8337 ino_t dinum; 8338 struct buf *newdirbp; 8339 struct mkdir **mkdirp; 8340 { 8341 struct newblk *newblk; 8342 struct pagedep *pagedep; 8343 struct inodedep *inodedep; 8344 struct newdirblk *newdirblk; 8345 struct mkdir *mkdir1, *mkdir2; 8346 struct worklist *wk; 8347 struct jaddref *jaddref; 8348 struct ufsmount *ump; 8349 struct mount *mp; 8350 8351 mp = dap->da_list.wk_mp; 8352 ump = VFSTOUFS(mp); 8353 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8354 M_SOFTDEP_FLAGS); 8355 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8356 LIST_INIT(&newdirblk->db_mkdir); 8357 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8358 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8359 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8360 mkdir1->md_diradd = dap; 8361 mkdir1->md_jaddref = NULL; 8362 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8363 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8364 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8365 mkdir2->md_diradd = dap; 8366 mkdir2->md_jaddref = NULL; 8367 if (MOUNTEDSUJ(mp) == 0) { 8368 mkdir1->md_state |= DEPCOMPLETE; 8369 mkdir2->md_state |= DEPCOMPLETE; 8370 } 8371 /* 8372 * Dependency on "." and ".." being written to disk. 8373 */ 8374 mkdir1->md_buf = newdirbp; 8375 ACQUIRE_LOCK(VFSTOUFS(mp)); 8376 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8377 /* 8378 * We must link the pagedep, allocdirect, and newdirblk for 8379 * the initial file page so the pointer to the new directory 8380 * is not written until the directory contents are live and 8381 * any subsequent additions are not marked live until the 8382 * block is reachable via the inode. 8383 */ 8384 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8385 panic("setup_newdir: lost pagedep"); 8386 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8387 if (wk->wk_type == D_ALLOCDIRECT) 8388 break; 8389 if (wk == NULL) 8390 panic("setup_newdir: lost allocdirect"); 8391 if (pagedep->pd_state & NEWBLOCK) 8392 panic("setup_newdir: NEWBLOCK already set"); 8393 newblk = WK_NEWBLK(wk); 8394 pagedep->pd_state |= NEWBLOCK; 8395 pagedep->pd_newdirblk = newdirblk; 8396 newdirblk->db_pagedep = pagedep; 8397 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8398 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8399 /* 8400 * Look up the inodedep for the parent directory so that we 8401 * can link mkdir2 into the pending dotdot jaddref or 8402 * the inode write if there is none. If the inode is 8403 * ALLCOMPLETE and no jaddref is present all dependencies have 8404 * been satisfied and mkdir2 can be freed. 8405 */ 8406 inodedep_lookup(mp, dinum, 0, &inodedep); 8407 if (MOUNTEDSUJ(mp)) { 8408 if (inodedep == NULL) 8409 panic("setup_newdir: Lost parent."); 8410 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8411 inoreflst); 8412 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8413 (jaddref->ja_state & MKDIR_PARENT), 8414 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8415 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8416 mkdir2->md_jaddref = jaddref; 8417 jaddref->ja_mkdir = mkdir2; 8418 } else if (inodedep == NULL || 8419 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8420 dap->da_state &= ~MKDIR_PARENT; 8421 WORKITEM_FREE(mkdir2, D_MKDIR); 8422 mkdir2 = NULL; 8423 } else { 8424 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8425 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8426 } 8427 *mkdirp = mkdir2; 8428 8429 return (mkdir1); 8430 } 8431 8432 /* 8433 * Directory entry addition dependencies. 8434 * 8435 * When adding a new directory entry, the inode (with its incremented link 8436 * count) must be written to disk before the directory entry's pointer to it. 8437 * Also, if the inode is newly allocated, the corresponding freemap must be 8438 * updated (on disk) before the directory entry's pointer. These requirements 8439 * are met via undo/redo on the directory entry's pointer, which consists 8440 * simply of the inode number. 8441 * 8442 * As directory entries are added and deleted, the free space within a 8443 * directory block can become fragmented. The ufs filesystem will compact 8444 * a fragmented directory block to make space for a new entry. When this 8445 * occurs, the offsets of previously added entries change. Any "diradd" 8446 * dependency structures corresponding to these entries must be updated with 8447 * the new offsets. 8448 */ 8449 8450 /* 8451 * This routine is called after the in-memory inode's link 8452 * count has been incremented, but before the directory entry's 8453 * pointer to the inode has been set. 8454 */ 8455 int 8456 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8457 struct buf *bp; /* buffer containing directory block */ 8458 struct inode *dp; /* inode for directory */ 8459 off_t diroffset; /* offset of new entry in directory */ 8460 ino_t newinum; /* inode referenced by new directory entry */ 8461 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8462 int isnewblk; /* entry is in a newly allocated block */ 8463 { 8464 int offset; /* offset of new entry within directory block */ 8465 ufs_lbn_t lbn; /* block in directory containing new entry */ 8466 struct fs *fs; 8467 struct diradd *dap; 8468 struct newblk *newblk; 8469 struct pagedep *pagedep; 8470 struct inodedep *inodedep; 8471 struct newdirblk *newdirblk; 8472 struct mkdir *mkdir1, *mkdir2; 8473 struct jaddref *jaddref; 8474 struct ufsmount *ump; 8475 struct mount *mp; 8476 int isindir; 8477 8478 mp = ITOVFS(dp); 8479 ump = VFSTOUFS(mp); 8480 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8481 ("softdep_setup_directory_add called on non-softdep filesystem")); 8482 /* 8483 * Whiteouts have no dependencies. 8484 */ 8485 if (newinum == UFS_WINO) { 8486 if (newdirbp != NULL) 8487 bdwrite(newdirbp); 8488 return (0); 8489 } 8490 jaddref = NULL; 8491 mkdir1 = mkdir2 = NULL; 8492 fs = ump->um_fs; 8493 lbn = lblkno(fs, diroffset); 8494 offset = blkoff(fs, diroffset); 8495 dap = malloc(sizeof(struct diradd), M_DIRADD, 8496 M_SOFTDEP_FLAGS|M_ZERO); 8497 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8498 dap->da_offset = offset; 8499 dap->da_newinum = newinum; 8500 dap->da_state = ATTACHED; 8501 LIST_INIT(&dap->da_jwork); 8502 isindir = bp->b_lblkno >= UFS_NDADDR; 8503 newdirblk = NULL; 8504 if (isnewblk && 8505 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8506 newdirblk = malloc(sizeof(struct newdirblk), 8507 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8508 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8509 LIST_INIT(&newdirblk->db_mkdir); 8510 } 8511 /* 8512 * If we're creating a new directory setup the dependencies and set 8513 * the dap state to wait for them. Otherwise it's COMPLETE and 8514 * we can move on. 8515 */ 8516 if (newdirbp == NULL) { 8517 dap->da_state |= DEPCOMPLETE; 8518 ACQUIRE_LOCK(ump); 8519 } else { 8520 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8521 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8522 &mkdir2); 8523 } 8524 /* 8525 * Link into parent directory pagedep to await its being written. 8526 */ 8527 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8528 #ifdef DEBUG 8529 if (diradd_lookup(pagedep, offset) != NULL) 8530 panic("softdep_setup_directory_add: %p already at off %d\n", 8531 diradd_lookup(pagedep, offset), offset); 8532 #endif 8533 dap->da_pagedep = pagedep; 8534 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8535 da_pdlist); 8536 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8537 /* 8538 * If we're journaling, link the diradd into the jaddref so it 8539 * may be completed after the journal entry is written. Otherwise, 8540 * link the diradd into its inodedep. If the inode is not yet 8541 * written place it on the bufwait list, otherwise do the post-inode 8542 * write processing to put it on the id_pendinghd list. 8543 */ 8544 if (MOUNTEDSUJ(mp)) { 8545 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8546 inoreflst); 8547 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8548 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8549 jaddref->ja_diroff = diroffset; 8550 jaddref->ja_diradd = dap; 8551 add_to_journal(&jaddref->ja_list); 8552 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8553 diradd_inode_written(dap, inodedep); 8554 else 8555 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8556 /* 8557 * Add the journal entries for . and .. links now that the primary 8558 * link is written. 8559 */ 8560 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8561 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8562 inoreflst, if_deps); 8563 KASSERT(jaddref != NULL && 8564 jaddref->ja_ino == jaddref->ja_parent && 8565 (jaddref->ja_state & MKDIR_BODY), 8566 ("softdep_setup_directory_add: bad dot jaddref %p", 8567 jaddref)); 8568 mkdir1->md_jaddref = jaddref; 8569 jaddref->ja_mkdir = mkdir1; 8570 /* 8571 * It is important that the dotdot journal entry 8572 * is added prior to the dot entry since dot writes 8573 * both the dot and dotdot links. These both must 8574 * be added after the primary link for the journal 8575 * to remain consistent. 8576 */ 8577 add_to_journal(&mkdir2->md_jaddref->ja_list); 8578 add_to_journal(&jaddref->ja_list); 8579 } 8580 /* 8581 * If we are adding a new directory remember this diradd so that if 8582 * we rename it we can keep the dot and dotdot dependencies. If 8583 * we are adding a new name for an inode that has a mkdiradd we 8584 * must be in rename and we have to move the dot and dotdot 8585 * dependencies to this new name. The old name is being orphaned 8586 * soon. 8587 */ 8588 if (mkdir1 != NULL) { 8589 if (inodedep->id_mkdiradd != NULL) 8590 panic("softdep_setup_directory_add: Existing mkdir"); 8591 inodedep->id_mkdiradd = dap; 8592 } else if (inodedep->id_mkdiradd) 8593 merge_diradd(inodedep, dap); 8594 if (newdirblk != NULL) { 8595 /* 8596 * There is nothing to do if we are already tracking 8597 * this block. 8598 */ 8599 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8600 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8601 FREE_LOCK(ump); 8602 return (0); 8603 } 8604 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8605 == 0) 8606 panic("softdep_setup_directory_add: lost entry"); 8607 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8608 pagedep->pd_state |= NEWBLOCK; 8609 pagedep->pd_newdirblk = newdirblk; 8610 newdirblk->db_pagedep = pagedep; 8611 FREE_LOCK(ump); 8612 /* 8613 * If we extended into an indirect signal direnter to sync. 8614 */ 8615 if (isindir) 8616 return (1); 8617 return (0); 8618 } 8619 FREE_LOCK(ump); 8620 return (0); 8621 } 8622 8623 /* 8624 * This procedure is called to change the offset of a directory 8625 * entry when compacting a directory block which must be owned 8626 * exclusively by the caller. Note that the actual entry movement 8627 * must be done in this procedure to ensure that no I/O completions 8628 * occur while the move is in progress. 8629 */ 8630 void 8631 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8632 struct buf *bp; /* Buffer holding directory block. */ 8633 struct inode *dp; /* inode for directory */ 8634 caddr_t base; /* address of dp->i_offset */ 8635 caddr_t oldloc; /* address of old directory location */ 8636 caddr_t newloc; /* address of new directory location */ 8637 int entrysize; /* size of directory entry */ 8638 { 8639 int offset, oldoffset, newoffset; 8640 struct pagedep *pagedep; 8641 struct jmvref *jmvref; 8642 struct diradd *dap; 8643 struct direct *de; 8644 struct mount *mp; 8645 struct ufsmount *ump; 8646 ufs_lbn_t lbn; 8647 int flags; 8648 8649 mp = ITOVFS(dp); 8650 ump = VFSTOUFS(mp); 8651 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8652 ("softdep_change_directoryentry_offset called on " 8653 "non-softdep filesystem")); 8654 de = (struct direct *)oldloc; 8655 jmvref = NULL; 8656 flags = 0; 8657 /* 8658 * Moves are always journaled as it would be too complex to 8659 * determine if any affected adds or removes are present in the 8660 * journal. 8661 */ 8662 if (MOUNTEDSUJ(mp)) { 8663 flags = DEPALLOC; 8664 jmvref = newjmvref(dp, de->d_ino, 8665 dp->i_offset + (oldloc - base), 8666 dp->i_offset + (newloc - base)); 8667 } 8668 lbn = lblkno(ump->um_fs, dp->i_offset); 8669 offset = blkoff(ump->um_fs, dp->i_offset); 8670 oldoffset = offset + (oldloc - base); 8671 newoffset = offset + (newloc - base); 8672 ACQUIRE_LOCK(ump); 8673 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8674 goto done; 8675 dap = diradd_lookup(pagedep, oldoffset); 8676 if (dap) { 8677 dap->da_offset = newoffset; 8678 newoffset = DIRADDHASH(newoffset); 8679 oldoffset = DIRADDHASH(oldoffset); 8680 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8681 newoffset != oldoffset) { 8682 LIST_REMOVE(dap, da_pdlist); 8683 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8684 dap, da_pdlist); 8685 } 8686 } 8687 done: 8688 if (jmvref) { 8689 jmvref->jm_pagedep = pagedep; 8690 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8691 add_to_journal(&jmvref->jm_list); 8692 } 8693 bcopy(oldloc, newloc, entrysize); 8694 FREE_LOCK(ump); 8695 } 8696 8697 /* 8698 * Move the mkdir dependencies and journal work from one diradd to another 8699 * when renaming a directory. The new name must depend on the mkdir deps 8700 * completing as the old name did. Directories can only have one valid link 8701 * at a time so one must be canonical. 8702 */ 8703 static void 8704 merge_diradd(inodedep, newdap) 8705 struct inodedep *inodedep; 8706 struct diradd *newdap; 8707 { 8708 struct diradd *olddap; 8709 struct mkdir *mkdir, *nextmd; 8710 struct ufsmount *ump; 8711 short state; 8712 8713 olddap = inodedep->id_mkdiradd; 8714 inodedep->id_mkdiradd = newdap; 8715 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8716 newdap->da_state &= ~DEPCOMPLETE; 8717 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8718 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8719 mkdir = nextmd) { 8720 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8721 if (mkdir->md_diradd != olddap) 8722 continue; 8723 mkdir->md_diradd = newdap; 8724 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8725 newdap->da_state |= state; 8726 olddap->da_state &= ~state; 8727 if ((olddap->da_state & 8728 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8729 break; 8730 } 8731 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8732 panic("merge_diradd: unfound ref"); 8733 } 8734 /* 8735 * Any mkdir related journal items are not safe to be freed until 8736 * the new name is stable. 8737 */ 8738 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8739 olddap->da_state |= DEPCOMPLETE; 8740 complete_diradd(olddap); 8741 } 8742 8743 /* 8744 * Move the diradd to the pending list when all diradd dependencies are 8745 * complete. 8746 */ 8747 static void 8748 complete_diradd(dap) 8749 struct diradd *dap; 8750 { 8751 struct pagedep *pagedep; 8752 8753 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8754 if (dap->da_state & DIRCHG) 8755 pagedep = dap->da_previous->dm_pagedep; 8756 else 8757 pagedep = dap->da_pagedep; 8758 LIST_REMOVE(dap, da_pdlist); 8759 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8760 } 8761 } 8762 8763 /* 8764 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8765 * add entries and conditonally journal the remove. 8766 */ 8767 static void 8768 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8769 struct diradd *dap; 8770 struct dirrem *dirrem; 8771 struct jremref *jremref; 8772 struct jremref *dotremref; 8773 struct jremref *dotdotremref; 8774 { 8775 struct inodedep *inodedep; 8776 struct jaddref *jaddref; 8777 struct inoref *inoref; 8778 struct ufsmount *ump; 8779 struct mkdir *mkdir; 8780 8781 /* 8782 * If no remove references were allocated we're on a non-journaled 8783 * filesystem and can skip the cancel step. 8784 */ 8785 if (jremref == NULL) { 8786 free_diradd(dap, NULL); 8787 return; 8788 } 8789 /* 8790 * Cancel the primary name an free it if it does not require 8791 * journaling. 8792 */ 8793 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8794 0, &inodedep) != 0) { 8795 /* Abort the addref that reference this diradd. */ 8796 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8797 if (inoref->if_list.wk_type != D_JADDREF) 8798 continue; 8799 jaddref = (struct jaddref *)inoref; 8800 if (jaddref->ja_diradd != dap) 8801 continue; 8802 if (cancel_jaddref(jaddref, inodedep, 8803 &dirrem->dm_jwork) == 0) { 8804 free_jremref(jremref); 8805 jremref = NULL; 8806 } 8807 break; 8808 } 8809 } 8810 /* 8811 * Cancel subordinate names and free them if they do not require 8812 * journaling. 8813 */ 8814 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8815 ump = VFSTOUFS(dap->da_list.wk_mp); 8816 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8817 if (mkdir->md_diradd != dap) 8818 continue; 8819 if ((jaddref = mkdir->md_jaddref) == NULL) 8820 continue; 8821 mkdir->md_jaddref = NULL; 8822 if (mkdir->md_state & MKDIR_PARENT) { 8823 if (cancel_jaddref(jaddref, NULL, 8824 &dirrem->dm_jwork) == 0) { 8825 free_jremref(dotdotremref); 8826 dotdotremref = NULL; 8827 } 8828 } else { 8829 if (cancel_jaddref(jaddref, inodedep, 8830 &dirrem->dm_jwork) == 0) { 8831 free_jremref(dotremref); 8832 dotremref = NULL; 8833 } 8834 } 8835 } 8836 } 8837 8838 if (jremref) 8839 journal_jremref(dirrem, jremref, inodedep); 8840 if (dotremref) 8841 journal_jremref(dirrem, dotremref, inodedep); 8842 if (dotdotremref) 8843 journal_jremref(dirrem, dotdotremref, NULL); 8844 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8845 free_diradd(dap, &dirrem->dm_jwork); 8846 } 8847 8848 /* 8849 * Free a diradd dependency structure. This routine must be called 8850 * with splbio interrupts blocked. 8851 */ 8852 static void 8853 free_diradd(dap, wkhd) 8854 struct diradd *dap; 8855 struct workhead *wkhd; 8856 { 8857 struct dirrem *dirrem; 8858 struct pagedep *pagedep; 8859 struct inodedep *inodedep; 8860 struct mkdir *mkdir, *nextmd; 8861 struct ufsmount *ump; 8862 8863 ump = VFSTOUFS(dap->da_list.wk_mp); 8864 LOCK_OWNED(ump); 8865 LIST_REMOVE(dap, da_pdlist); 8866 if (dap->da_state & ONWORKLIST) 8867 WORKLIST_REMOVE(&dap->da_list); 8868 if ((dap->da_state & DIRCHG) == 0) { 8869 pagedep = dap->da_pagedep; 8870 } else { 8871 dirrem = dap->da_previous; 8872 pagedep = dirrem->dm_pagedep; 8873 dirrem->dm_dirinum = pagedep->pd_ino; 8874 dirrem->dm_state |= COMPLETE; 8875 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8876 add_to_worklist(&dirrem->dm_list, 0); 8877 } 8878 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8879 0, &inodedep) != 0) 8880 if (inodedep->id_mkdiradd == dap) 8881 inodedep->id_mkdiradd = NULL; 8882 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8883 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8884 mkdir = nextmd) { 8885 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8886 if (mkdir->md_diradd != dap) 8887 continue; 8888 dap->da_state &= 8889 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8890 LIST_REMOVE(mkdir, md_mkdirs); 8891 if (mkdir->md_state & ONWORKLIST) 8892 WORKLIST_REMOVE(&mkdir->md_list); 8893 if (mkdir->md_jaddref != NULL) 8894 panic("free_diradd: Unexpected jaddref"); 8895 WORKITEM_FREE(mkdir, D_MKDIR); 8896 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8897 break; 8898 } 8899 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8900 panic("free_diradd: unfound ref"); 8901 } 8902 if (inodedep) 8903 free_inodedep(inodedep); 8904 /* 8905 * Free any journal segments waiting for the directory write. 8906 */ 8907 handle_jwork(&dap->da_jwork); 8908 WORKITEM_FREE(dap, D_DIRADD); 8909 } 8910 8911 /* 8912 * Directory entry removal dependencies. 8913 * 8914 * When removing a directory entry, the entry's inode pointer must be 8915 * zero'ed on disk before the corresponding inode's link count is decremented 8916 * (possibly freeing the inode for re-use). This dependency is handled by 8917 * updating the directory entry but delaying the inode count reduction until 8918 * after the directory block has been written to disk. After this point, the 8919 * inode count can be decremented whenever it is convenient. 8920 */ 8921 8922 /* 8923 * This routine should be called immediately after removing 8924 * a directory entry. The inode's link count should not be 8925 * decremented by the calling procedure -- the soft updates 8926 * code will do this task when it is safe. 8927 */ 8928 void 8929 softdep_setup_remove(bp, dp, ip, isrmdir) 8930 struct buf *bp; /* buffer containing directory block */ 8931 struct inode *dp; /* inode for the directory being modified */ 8932 struct inode *ip; /* inode for directory entry being removed */ 8933 int isrmdir; /* indicates if doing RMDIR */ 8934 { 8935 struct dirrem *dirrem, *prevdirrem; 8936 struct inodedep *inodedep; 8937 struct ufsmount *ump; 8938 int direct; 8939 8940 ump = ITOUMP(ip); 8941 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 8942 ("softdep_setup_remove called on non-softdep filesystem")); 8943 /* 8944 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8945 * newdirrem() to setup the full directory remove which requires 8946 * isrmdir > 1. 8947 */ 8948 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8949 /* 8950 * Add the dirrem to the inodedep's pending remove list for quick 8951 * discovery later. 8952 */ 8953 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) 8954 panic("softdep_setup_remove: Lost inodedep."); 8955 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8956 dirrem->dm_state |= ONDEPLIST; 8957 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8958 8959 /* 8960 * If the COMPLETE flag is clear, then there were no active 8961 * entries and we want to roll back to a zeroed entry until 8962 * the new inode is committed to disk. If the COMPLETE flag is 8963 * set then we have deleted an entry that never made it to 8964 * disk. If the entry we deleted resulted from a name change, 8965 * then the old name still resides on disk. We cannot delete 8966 * its inode (returned to us in prevdirrem) until the zeroed 8967 * directory entry gets to disk. The new inode has never been 8968 * referenced on the disk, so can be deleted immediately. 8969 */ 8970 if ((dirrem->dm_state & COMPLETE) == 0) { 8971 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8972 dm_next); 8973 FREE_LOCK(ump); 8974 } else { 8975 if (prevdirrem != NULL) 8976 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8977 prevdirrem, dm_next); 8978 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8979 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8980 FREE_LOCK(ump); 8981 if (direct) 8982 handle_workitem_remove(dirrem, 0); 8983 } 8984 } 8985 8986 /* 8987 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8988 * pd_pendinghd list of a pagedep. 8989 */ 8990 static struct diradd * 8991 diradd_lookup(pagedep, offset) 8992 struct pagedep *pagedep; 8993 int offset; 8994 { 8995 struct diradd *dap; 8996 8997 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8998 if (dap->da_offset == offset) 8999 return (dap); 9000 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 9001 if (dap->da_offset == offset) 9002 return (dap); 9003 return (NULL); 9004 } 9005 9006 /* 9007 * Search for a .. diradd dependency in a directory that is being removed. 9008 * If the directory was renamed to a new parent we have a diradd rather 9009 * than a mkdir for the .. entry. We need to cancel it now before 9010 * it is found in truncate(). 9011 */ 9012 static struct jremref * 9013 cancel_diradd_dotdot(ip, dirrem, jremref) 9014 struct inode *ip; 9015 struct dirrem *dirrem; 9016 struct jremref *jremref; 9017 { 9018 struct pagedep *pagedep; 9019 struct diradd *dap; 9020 struct worklist *wk; 9021 9022 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0) 9023 return (jremref); 9024 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 9025 if (dap == NULL) 9026 return (jremref); 9027 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 9028 /* 9029 * Mark any journal work as belonging to the parent so it is freed 9030 * with the .. reference. 9031 */ 9032 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9033 wk->wk_state |= MKDIR_PARENT; 9034 return (NULL); 9035 } 9036 9037 /* 9038 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 9039 * replace it with a dirrem/diradd pair as a result of re-parenting a 9040 * directory. This ensures that we don't simultaneously have a mkdir and 9041 * a diradd for the same .. entry. 9042 */ 9043 static struct jremref * 9044 cancel_mkdir_dotdot(ip, dirrem, jremref) 9045 struct inode *ip; 9046 struct dirrem *dirrem; 9047 struct jremref *jremref; 9048 { 9049 struct inodedep *inodedep; 9050 struct jaddref *jaddref; 9051 struct ufsmount *ump; 9052 struct mkdir *mkdir; 9053 struct diradd *dap; 9054 struct mount *mp; 9055 9056 mp = ITOVFS(ip); 9057 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9058 return (jremref); 9059 dap = inodedep->id_mkdiradd; 9060 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9061 return (jremref); 9062 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9063 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9064 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9065 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9066 break; 9067 if (mkdir == NULL) 9068 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9069 if ((jaddref = mkdir->md_jaddref) != NULL) { 9070 mkdir->md_jaddref = NULL; 9071 jaddref->ja_state &= ~MKDIR_PARENT; 9072 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0) 9073 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9074 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9075 journal_jremref(dirrem, jremref, inodedep); 9076 jremref = NULL; 9077 } 9078 } 9079 if (mkdir->md_state & ONWORKLIST) 9080 WORKLIST_REMOVE(&mkdir->md_list); 9081 mkdir->md_state |= ALLCOMPLETE; 9082 complete_mkdir(mkdir); 9083 return (jremref); 9084 } 9085 9086 static void 9087 journal_jremref(dirrem, jremref, inodedep) 9088 struct dirrem *dirrem; 9089 struct jremref *jremref; 9090 struct inodedep *inodedep; 9091 { 9092 9093 if (inodedep == NULL) 9094 if (inodedep_lookup(jremref->jr_list.wk_mp, 9095 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9096 panic("journal_jremref: Lost inodedep"); 9097 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9098 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9099 add_to_journal(&jremref->jr_list); 9100 } 9101 9102 static void 9103 dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9104 struct dirrem *dirrem; 9105 struct jremref *jremref; 9106 struct jremref *dotremref; 9107 struct jremref *dotdotremref; 9108 { 9109 struct inodedep *inodedep; 9110 9111 9112 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9113 &inodedep) == 0) 9114 panic("dirrem_journal: Lost inodedep"); 9115 journal_jremref(dirrem, jremref, inodedep); 9116 if (dotremref) 9117 journal_jremref(dirrem, dotremref, inodedep); 9118 if (dotdotremref) 9119 journal_jremref(dirrem, dotdotremref, NULL); 9120 } 9121 9122 /* 9123 * Allocate a new dirrem if appropriate and return it along with 9124 * its associated pagedep. Called without a lock, returns with lock. 9125 */ 9126 static struct dirrem * 9127 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9128 struct buf *bp; /* buffer containing directory block */ 9129 struct inode *dp; /* inode for the directory being modified */ 9130 struct inode *ip; /* inode for directory entry being removed */ 9131 int isrmdir; /* indicates if doing RMDIR */ 9132 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9133 { 9134 int offset; 9135 ufs_lbn_t lbn; 9136 struct diradd *dap; 9137 struct dirrem *dirrem; 9138 struct pagedep *pagedep; 9139 struct jremref *jremref; 9140 struct jremref *dotremref; 9141 struct jremref *dotdotremref; 9142 struct vnode *dvp; 9143 struct ufsmount *ump; 9144 9145 /* 9146 * Whiteouts have no deletion dependencies. 9147 */ 9148 if (ip == NULL) 9149 panic("newdirrem: whiteout"); 9150 dvp = ITOV(dp); 9151 ump = ITOUMP(dp); 9152 9153 /* 9154 * If the system is over its limit and our filesystem is 9155 * responsible for more than our share of that usage and 9156 * we are not a snapshot, request some inodedep cleanup. 9157 * Limiting the number of dirrem structures will also limit 9158 * the number of freefile and freeblks structures. 9159 */ 9160 ACQUIRE_LOCK(ump); 9161 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM)) 9162 schedule_cleanup(UFSTOVFS(ump)); 9163 else 9164 FREE_LOCK(ump); 9165 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9166 M_ZERO); 9167 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9168 LIST_INIT(&dirrem->dm_jremrefhd); 9169 LIST_INIT(&dirrem->dm_jwork); 9170 dirrem->dm_state = isrmdir ? RMDIR : 0; 9171 dirrem->dm_oldinum = ip->i_number; 9172 *prevdirremp = NULL; 9173 /* 9174 * Allocate remove reference structures to track journal write 9175 * dependencies. We will always have one for the link and 9176 * when doing directories we will always have one more for dot. 9177 * When renaming a directory we skip the dotdot link change so 9178 * this is not needed. 9179 */ 9180 jremref = dotremref = dotdotremref = NULL; 9181 if (DOINGSUJ(dvp)) { 9182 if (isrmdir) { 9183 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9184 ip->i_effnlink + 2); 9185 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9186 ip->i_effnlink + 1); 9187 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9188 dp->i_effnlink + 1); 9189 dotdotremref->jr_state |= MKDIR_PARENT; 9190 } else 9191 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9192 ip->i_effnlink + 1); 9193 } 9194 ACQUIRE_LOCK(ump); 9195 lbn = lblkno(ump->um_fs, dp->i_offset); 9196 offset = blkoff(ump->um_fs, dp->i_offset); 9197 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC, 9198 &pagedep); 9199 dirrem->dm_pagedep = pagedep; 9200 dirrem->dm_offset = offset; 9201 /* 9202 * If we're renaming a .. link to a new directory, cancel any 9203 * existing MKDIR_PARENT mkdir. If it has already been canceled 9204 * the jremref is preserved for any potential diradd in this 9205 * location. This can not coincide with a rmdir. 9206 */ 9207 if (dp->i_offset == DOTDOT_OFFSET) { 9208 if (isrmdir) 9209 panic("newdirrem: .. directory change during remove?"); 9210 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9211 } 9212 /* 9213 * If we're removing a directory search for the .. dependency now and 9214 * cancel it. Any pending journal work will be added to the dirrem 9215 * to be completed when the workitem remove completes. 9216 */ 9217 if (isrmdir) 9218 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9219 /* 9220 * Check for a diradd dependency for the same directory entry. 9221 * If present, then both dependencies become obsolete and can 9222 * be de-allocated. 9223 */ 9224 dap = diradd_lookup(pagedep, offset); 9225 if (dap == NULL) { 9226 /* 9227 * Link the jremref structures into the dirrem so they are 9228 * written prior to the pagedep. 9229 */ 9230 if (jremref) 9231 dirrem_journal(dirrem, jremref, dotremref, 9232 dotdotremref); 9233 return (dirrem); 9234 } 9235 /* 9236 * Must be ATTACHED at this point. 9237 */ 9238 if ((dap->da_state & ATTACHED) == 0) 9239 panic("newdirrem: not ATTACHED"); 9240 if (dap->da_newinum != ip->i_number) 9241 panic("newdirrem: inum %ju should be %ju", 9242 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9243 /* 9244 * If we are deleting a changed name that never made it to disk, 9245 * then return the dirrem describing the previous inode (which 9246 * represents the inode currently referenced from this entry on disk). 9247 */ 9248 if ((dap->da_state & DIRCHG) != 0) { 9249 *prevdirremp = dap->da_previous; 9250 dap->da_state &= ~DIRCHG; 9251 dap->da_pagedep = pagedep; 9252 } 9253 /* 9254 * We are deleting an entry that never made it to disk. 9255 * Mark it COMPLETE so we can delete its inode immediately. 9256 */ 9257 dirrem->dm_state |= COMPLETE; 9258 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9259 #ifdef SUJ_DEBUG 9260 if (isrmdir == 0) { 9261 struct worklist *wk; 9262 9263 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9264 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9265 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9266 } 9267 #endif 9268 9269 return (dirrem); 9270 } 9271 9272 /* 9273 * Directory entry change dependencies. 9274 * 9275 * Changing an existing directory entry requires that an add operation 9276 * be completed first followed by a deletion. The semantics for the addition 9277 * are identical to the description of adding a new entry above except 9278 * that the rollback is to the old inode number rather than zero. Once 9279 * the addition dependency is completed, the removal is done as described 9280 * in the removal routine above. 9281 */ 9282 9283 /* 9284 * This routine should be called immediately after changing 9285 * a directory entry. The inode's link count should not be 9286 * decremented by the calling procedure -- the soft updates 9287 * code will perform this task when it is safe. 9288 */ 9289 void 9290 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9291 struct buf *bp; /* buffer containing directory block */ 9292 struct inode *dp; /* inode for the directory being modified */ 9293 struct inode *ip; /* inode for directory entry being removed */ 9294 ino_t newinum; /* new inode number for changed entry */ 9295 int isrmdir; /* indicates if doing RMDIR */ 9296 { 9297 int offset; 9298 struct diradd *dap = NULL; 9299 struct dirrem *dirrem, *prevdirrem; 9300 struct pagedep *pagedep; 9301 struct inodedep *inodedep; 9302 struct jaddref *jaddref; 9303 struct mount *mp; 9304 struct ufsmount *ump; 9305 9306 mp = ITOVFS(dp); 9307 ump = VFSTOUFS(mp); 9308 offset = blkoff(ump->um_fs, dp->i_offset); 9309 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9310 ("softdep_setup_directory_change called on non-softdep filesystem")); 9311 9312 /* 9313 * Whiteouts do not need diradd dependencies. 9314 */ 9315 if (newinum != UFS_WINO) { 9316 dap = malloc(sizeof(struct diradd), 9317 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9318 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9319 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9320 dap->da_offset = offset; 9321 dap->da_newinum = newinum; 9322 LIST_INIT(&dap->da_jwork); 9323 } 9324 9325 /* 9326 * Allocate a new dirrem and ACQUIRE_LOCK. 9327 */ 9328 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9329 pagedep = dirrem->dm_pagedep; 9330 /* 9331 * The possible values for isrmdir: 9332 * 0 - non-directory file rename 9333 * 1 - directory rename within same directory 9334 * inum - directory rename to new directory of given inode number 9335 * When renaming to a new directory, we are both deleting and 9336 * creating a new directory entry, so the link count on the new 9337 * directory should not change. Thus we do not need the followup 9338 * dirrem which is usually done in handle_workitem_remove. We set 9339 * the DIRCHG flag to tell handle_workitem_remove to skip the 9340 * followup dirrem. 9341 */ 9342 if (isrmdir > 1) 9343 dirrem->dm_state |= DIRCHG; 9344 9345 /* 9346 * Whiteouts have no additional dependencies, 9347 * so just put the dirrem on the correct list. 9348 */ 9349 if (newinum == UFS_WINO) { 9350 if ((dirrem->dm_state & COMPLETE) == 0) { 9351 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9352 dm_next); 9353 } else { 9354 dirrem->dm_dirinum = pagedep->pd_ino; 9355 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9356 add_to_worklist(&dirrem->dm_list, 0); 9357 } 9358 FREE_LOCK(ump); 9359 return; 9360 } 9361 /* 9362 * Add the dirrem to the inodedep's pending remove list for quick 9363 * discovery later. A valid nlinkdelta ensures that this lookup 9364 * will not fail. 9365 */ 9366 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9367 panic("softdep_setup_directory_change: Lost inodedep."); 9368 dirrem->dm_state |= ONDEPLIST; 9369 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9370 9371 /* 9372 * If the COMPLETE flag is clear, then there were no active 9373 * entries and we want to roll back to the previous inode until 9374 * the new inode is committed to disk. If the COMPLETE flag is 9375 * set, then we have deleted an entry that never made it to disk. 9376 * If the entry we deleted resulted from a name change, then the old 9377 * inode reference still resides on disk. Any rollback that we do 9378 * needs to be to that old inode (returned to us in prevdirrem). If 9379 * the entry we deleted resulted from a create, then there is 9380 * no entry on the disk, so we want to roll back to zero rather 9381 * than the uncommitted inode. In either of the COMPLETE cases we 9382 * want to immediately free the unwritten and unreferenced inode. 9383 */ 9384 if ((dirrem->dm_state & COMPLETE) == 0) { 9385 dap->da_previous = dirrem; 9386 } else { 9387 if (prevdirrem != NULL) { 9388 dap->da_previous = prevdirrem; 9389 } else { 9390 dap->da_state &= ~DIRCHG; 9391 dap->da_pagedep = pagedep; 9392 } 9393 dirrem->dm_dirinum = pagedep->pd_ino; 9394 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9395 add_to_worklist(&dirrem->dm_list, 0); 9396 } 9397 /* 9398 * Lookup the jaddref for this journal entry. We must finish 9399 * initializing it and make the diradd write dependent on it. 9400 * If we're not journaling, put it on the id_bufwait list if the 9401 * inode is not yet written. If it is written, do the post-inode 9402 * write processing to put it on the id_pendinghd list. 9403 */ 9404 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9405 if (MOUNTEDSUJ(mp)) { 9406 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9407 inoreflst); 9408 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9409 ("softdep_setup_directory_change: bad jaddref %p", 9410 jaddref)); 9411 jaddref->ja_diroff = dp->i_offset; 9412 jaddref->ja_diradd = dap; 9413 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9414 dap, da_pdlist); 9415 add_to_journal(&jaddref->ja_list); 9416 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9417 dap->da_state |= COMPLETE; 9418 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9419 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9420 } else { 9421 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9422 dap, da_pdlist); 9423 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9424 } 9425 /* 9426 * If we're making a new name for a directory that has not been 9427 * committed when need to move the dot and dotdot references to 9428 * this new name. 9429 */ 9430 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9431 merge_diradd(inodedep, dap); 9432 FREE_LOCK(ump); 9433 } 9434 9435 /* 9436 * Called whenever the link count on an inode is changed. 9437 * It creates an inode dependency so that the new reference(s) 9438 * to the inode cannot be committed to disk until the updated 9439 * inode has been written. 9440 */ 9441 void 9442 softdep_change_linkcnt(ip) 9443 struct inode *ip; /* the inode with the increased link count */ 9444 { 9445 struct inodedep *inodedep; 9446 struct ufsmount *ump; 9447 9448 ump = ITOUMP(ip); 9449 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9450 ("softdep_change_linkcnt called on non-softdep filesystem")); 9451 ACQUIRE_LOCK(ump); 9452 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 9453 if (ip->i_nlink < ip->i_effnlink) 9454 panic("softdep_change_linkcnt: bad delta"); 9455 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9456 FREE_LOCK(ump); 9457 } 9458 9459 /* 9460 * Attach a sbdep dependency to the superblock buf so that we can keep 9461 * track of the head of the linked list of referenced but unlinked inodes. 9462 */ 9463 void 9464 softdep_setup_sbupdate(ump, fs, bp) 9465 struct ufsmount *ump; 9466 struct fs *fs; 9467 struct buf *bp; 9468 { 9469 struct sbdep *sbdep; 9470 struct worklist *wk; 9471 9472 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9473 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9474 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9475 if (wk->wk_type == D_SBDEP) 9476 break; 9477 if (wk != NULL) 9478 return; 9479 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9480 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9481 sbdep->sb_fs = fs; 9482 sbdep->sb_ump = ump; 9483 ACQUIRE_LOCK(ump); 9484 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9485 FREE_LOCK(ump); 9486 } 9487 9488 /* 9489 * Return the first unlinked inodedep which is ready to be the head of the 9490 * list. The inodedep and all those after it must have valid next pointers. 9491 */ 9492 static struct inodedep * 9493 first_unlinked_inodedep(ump) 9494 struct ufsmount *ump; 9495 { 9496 struct inodedep *inodedep; 9497 struct inodedep *idp; 9498 9499 LOCK_OWNED(ump); 9500 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9501 inodedep; inodedep = idp) { 9502 if ((inodedep->id_state & UNLINKNEXT) == 0) 9503 return (NULL); 9504 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9505 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9506 break; 9507 if ((inodedep->id_state & UNLINKPREV) == 0) 9508 break; 9509 } 9510 return (inodedep); 9511 } 9512 9513 /* 9514 * Set the sujfree unlinked head pointer prior to writing a superblock. 9515 */ 9516 static void 9517 initiate_write_sbdep(sbdep) 9518 struct sbdep *sbdep; 9519 { 9520 struct inodedep *inodedep; 9521 struct fs *bpfs; 9522 struct fs *fs; 9523 9524 bpfs = sbdep->sb_fs; 9525 fs = sbdep->sb_ump->um_fs; 9526 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9527 if (inodedep) { 9528 fs->fs_sujfree = inodedep->id_ino; 9529 inodedep->id_state |= UNLINKPREV; 9530 } else 9531 fs->fs_sujfree = 0; 9532 bpfs->fs_sujfree = fs->fs_sujfree; 9533 } 9534 9535 /* 9536 * After a superblock is written determine whether it must be written again 9537 * due to a changing unlinked list head. 9538 */ 9539 static int 9540 handle_written_sbdep(sbdep, bp) 9541 struct sbdep *sbdep; 9542 struct buf *bp; 9543 { 9544 struct inodedep *inodedep; 9545 struct fs *fs; 9546 9547 LOCK_OWNED(sbdep->sb_ump); 9548 fs = sbdep->sb_fs; 9549 /* 9550 * If the superblock doesn't match the in-memory list start over. 9551 */ 9552 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9553 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9554 (inodedep == NULL && fs->fs_sujfree != 0)) { 9555 bdirty(bp); 9556 return (1); 9557 } 9558 WORKITEM_FREE(sbdep, D_SBDEP); 9559 if (fs->fs_sujfree == 0) 9560 return (0); 9561 /* 9562 * Now that we have a record of this inode in stable store allow it 9563 * to be written to free up pending work. Inodes may see a lot of 9564 * write activity after they are unlinked which we must not hold up. 9565 */ 9566 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9567 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9568 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9569 inodedep, inodedep->id_state); 9570 if (inodedep->id_state & UNLINKONLIST) 9571 break; 9572 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9573 } 9574 9575 return (0); 9576 } 9577 9578 /* 9579 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9580 */ 9581 static void 9582 unlinked_inodedep(mp, inodedep) 9583 struct mount *mp; 9584 struct inodedep *inodedep; 9585 { 9586 struct ufsmount *ump; 9587 9588 ump = VFSTOUFS(mp); 9589 LOCK_OWNED(ump); 9590 if (MOUNTEDSUJ(mp) == 0) 9591 return; 9592 ump->um_fs->fs_fmod = 1; 9593 if (inodedep->id_state & UNLINKED) 9594 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9595 inodedep->id_state |= UNLINKED; 9596 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9597 } 9598 9599 /* 9600 * Remove an inodedep from the unlinked inodedep list. This may require 9601 * disk writes if the inode has made it that far. 9602 */ 9603 static void 9604 clear_unlinked_inodedep(inodedep) 9605 struct inodedep *inodedep; 9606 { 9607 struct ufsmount *ump; 9608 struct inodedep *idp; 9609 struct inodedep *idn; 9610 struct fs *fs; 9611 struct buf *bp; 9612 ino_t ino; 9613 ino_t nino; 9614 ino_t pino; 9615 int error; 9616 9617 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9618 fs = ump->um_fs; 9619 ino = inodedep->id_ino; 9620 error = 0; 9621 for (;;) { 9622 LOCK_OWNED(ump); 9623 KASSERT((inodedep->id_state & UNLINKED) != 0, 9624 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9625 inodedep)); 9626 /* 9627 * If nothing has yet been written simply remove us from 9628 * the in memory list and return. This is the most common 9629 * case where handle_workitem_remove() loses the final 9630 * reference. 9631 */ 9632 if ((inodedep->id_state & UNLINKLINKS) == 0) 9633 break; 9634 /* 9635 * If we have a NEXT pointer and no PREV pointer we can simply 9636 * clear NEXT's PREV and remove ourselves from the list. Be 9637 * careful not to clear PREV if the superblock points at 9638 * next as well. 9639 */ 9640 idn = TAILQ_NEXT(inodedep, id_unlinked); 9641 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9642 if (idn && fs->fs_sujfree != idn->id_ino) 9643 idn->id_state &= ~UNLINKPREV; 9644 break; 9645 } 9646 /* 9647 * Here we have an inodedep which is actually linked into 9648 * the list. We must remove it by forcing a write to the 9649 * link before us, whether it be the superblock or an inode. 9650 * Unfortunately the list may change while we're waiting 9651 * on the buf lock for either resource so we must loop until 9652 * we lock the right one. If both the superblock and an 9653 * inode point to this inode we must clear the inode first 9654 * followed by the superblock. 9655 */ 9656 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9657 pino = 0; 9658 if (idp && (idp->id_state & UNLINKNEXT)) 9659 pino = idp->id_ino; 9660 FREE_LOCK(ump); 9661 if (pino == 0) { 9662 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9663 (int)fs->fs_sbsize, 0, 0, 0); 9664 } else { 9665 error = bread(ump->um_devvp, 9666 fsbtodb(fs, ino_to_fsba(fs, pino)), 9667 (int)fs->fs_bsize, NOCRED, &bp); 9668 if (error) 9669 brelse(bp); 9670 } 9671 ACQUIRE_LOCK(ump); 9672 if (error) 9673 break; 9674 /* If the list has changed restart the loop. */ 9675 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9676 nino = 0; 9677 if (idp && (idp->id_state & UNLINKNEXT)) 9678 nino = idp->id_ino; 9679 if (nino != pino || 9680 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9681 FREE_LOCK(ump); 9682 brelse(bp); 9683 ACQUIRE_LOCK(ump); 9684 continue; 9685 } 9686 nino = 0; 9687 idn = TAILQ_NEXT(inodedep, id_unlinked); 9688 if (idn) 9689 nino = idn->id_ino; 9690 /* 9691 * Remove us from the in memory list. After this we cannot 9692 * access the inodedep. 9693 */ 9694 KASSERT((inodedep->id_state & UNLINKED) != 0, 9695 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9696 inodedep)); 9697 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9698 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9699 FREE_LOCK(ump); 9700 /* 9701 * The predecessor's next pointer is manually updated here 9702 * so that the NEXT flag is never cleared for an element 9703 * that is in the list. 9704 */ 9705 if (pino == 0) { 9706 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9707 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9708 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9709 bp); 9710 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9711 ((struct ufs1_dinode *)bp->b_data + 9712 ino_to_fsbo(fs, pino))->di_freelink = nino; 9713 else 9714 ((struct ufs2_dinode *)bp->b_data + 9715 ino_to_fsbo(fs, pino))->di_freelink = nino; 9716 /* 9717 * If the bwrite fails we have no recourse to recover. The 9718 * filesystem is corrupted already. 9719 */ 9720 bwrite(bp); 9721 ACQUIRE_LOCK(ump); 9722 /* 9723 * If the superblock pointer still needs to be cleared force 9724 * a write here. 9725 */ 9726 if (fs->fs_sujfree == ino) { 9727 FREE_LOCK(ump); 9728 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9729 (int)fs->fs_sbsize, 0, 0, 0); 9730 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9731 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9732 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9733 bp); 9734 bwrite(bp); 9735 ACQUIRE_LOCK(ump); 9736 } 9737 9738 if (fs->fs_sujfree != ino) 9739 return; 9740 panic("clear_unlinked_inodedep: Failed to clear free head"); 9741 } 9742 if (inodedep->id_ino == fs->fs_sujfree) 9743 panic("clear_unlinked_inodedep: Freeing head of free list"); 9744 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9745 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9746 return; 9747 } 9748 9749 /* 9750 * This workitem decrements the inode's link count. 9751 * If the link count reaches zero, the file is removed. 9752 */ 9753 static int 9754 handle_workitem_remove(dirrem, flags) 9755 struct dirrem *dirrem; 9756 int flags; 9757 { 9758 struct inodedep *inodedep; 9759 struct workhead dotdotwk; 9760 struct worklist *wk; 9761 struct ufsmount *ump; 9762 struct mount *mp; 9763 struct vnode *vp; 9764 struct inode *ip; 9765 ino_t oldinum; 9766 9767 if (dirrem->dm_state & ONWORKLIST) 9768 panic("handle_workitem_remove: dirrem %p still on worklist", 9769 dirrem); 9770 oldinum = dirrem->dm_oldinum; 9771 mp = dirrem->dm_list.wk_mp; 9772 ump = VFSTOUFS(mp); 9773 flags |= LK_EXCLUSIVE; 9774 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9775 return (EBUSY); 9776 ip = VTOI(vp); 9777 ACQUIRE_LOCK(ump); 9778 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9779 panic("handle_workitem_remove: lost inodedep"); 9780 if (dirrem->dm_state & ONDEPLIST) 9781 LIST_REMOVE(dirrem, dm_inonext); 9782 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9783 ("handle_workitem_remove: Journal entries not written.")); 9784 9785 /* 9786 * Move all dependencies waiting on the remove to complete 9787 * from the dirrem to the inode inowait list to be completed 9788 * after the inode has been updated and written to disk. Any 9789 * marked MKDIR_PARENT are saved to be completed when the .. ref 9790 * is removed. 9791 */ 9792 LIST_INIT(&dotdotwk); 9793 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9794 WORKLIST_REMOVE(wk); 9795 if (wk->wk_state & MKDIR_PARENT) { 9796 wk->wk_state &= ~MKDIR_PARENT; 9797 WORKLIST_INSERT(&dotdotwk, wk); 9798 continue; 9799 } 9800 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9801 } 9802 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9803 /* 9804 * Normal file deletion. 9805 */ 9806 if ((dirrem->dm_state & RMDIR) == 0) { 9807 ip->i_nlink--; 9808 DIP_SET(ip, i_nlink, ip->i_nlink); 9809 ip->i_flag |= IN_CHANGE; 9810 if (ip->i_nlink < ip->i_effnlink) 9811 panic("handle_workitem_remove: bad file delta"); 9812 if (ip->i_nlink == 0) 9813 unlinked_inodedep(mp, inodedep); 9814 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9815 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9816 ("handle_workitem_remove: worklist not empty. %s", 9817 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9818 WORKITEM_FREE(dirrem, D_DIRREM); 9819 FREE_LOCK(ump); 9820 goto out; 9821 } 9822 /* 9823 * Directory deletion. Decrement reference count for both the 9824 * just deleted parent directory entry and the reference for ".". 9825 * Arrange to have the reference count on the parent decremented 9826 * to account for the loss of "..". 9827 */ 9828 ip->i_nlink -= 2; 9829 DIP_SET(ip, i_nlink, ip->i_nlink); 9830 ip->i_flag |= IN_CHANGE; 9831 if (ip->i_nlink < ip->i_effnlink) 9832 panic("handle_workitem_remove: bad dir delta"); 9833 if (ip->i_nlink == 0) 9834 unlinked_inodedep(mp, inodedep); 9835 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9836 /* 9837 * Rename a directory to a new parent. Since, we are both deleting 9838 * and creating a new directory entry, the link count on the new 9839 * directory should not change. Thus we skip the followup dirrem. 9840 */ 9841 if (dirrem->dm_state & DIRCHG) { 9842 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9843 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9844 WORKITEM_FREE(dirrem, D_DIRREM); 9845 FREE_LOCK(ump); 9846 goto out; 9847 } 9848 dirrem->dm_state = ONDEPLIST; 9849 dirrem->dm_oldinum = dirrem->dm_dirinum; 9850 /* 9851 * Place the dirrem on the parent's diremhd list. 9852 */ 9853 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9854 panic("handle_workitem_remove: lost dir inodedep"); 9855 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9856 /* 9857 * If the allocated inode has never been written to disk, then 9858 * the on-disk inode is zero'ed and we can remove the file 9859 * immediately. When journaling if the inode has been marked 9860 * unlinked and not DEPCOMPLETE we know it can never be written. 9861 */ 9862 inodedep_lookup(mp, oldinum, 0, &inodedep); 9863 if (inodedep == NULL || 9864 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9865 check_inode_unwritten(inodedep)) { 9866 FREE_LOCK(ump); 9867 vput(vp); 9868 return handle_workitem_remove(dirrem, flags); 9869 } 9870 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9871 FREE_LOCK(ump); 9872 ip->i_flag |= IN_CHANGE; 9873 out: 9874 ffs_update(vp, 0); 9875 vput(vp); 9876 return (0); 9877 } 9878 9879 /* 9880 * Inode de-allocation dependencies. 9881 * 9882 * When an inode's link count is reduced to zero, it can be de-allocated. We 9883 * found it convenient to postpone de-allocation until after the inode is 9884 * written to disk with its new link count (zero). At this point, all of the 9885 * on-disk inode's block pointers are nullified and, with careful dependency 9886 * list ordering, all dependencies related to the inode will be satisfied and 9887 * the corresponding dependency structures de-allocated. So, if/when the 9888 * inode is reused, there will be no mixing of old dependencies with new 9889 * ones. This artificial dependency is set up by the block de-allocation 9890 * procedure above (softdep_setup_freeblocks) and completed by the 9891 * following procedure. 9892 */ 9893 static void 9894 handle_workitem_freefile(freefile) 9895 struct freefile *freefile; 9896 { 9897 struct workhead wkhd; 9898 struct fs *fs; 9899 struct inodedep *idp; 9900 struct ufsmount *ump; 9901 int error; 9902 9903 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9904 fs = ump->um_fs; 9905 #ifdef DEBUG 9906 ACQUIRE_LOCK(ump); 9907 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9908 FREE_LOCK(ump); 9909 if (error) 9910 panic("handle_workitem_freefile: inodedep %p survived", idp); 9911 #endif 9912 UFS_LOCK(ump); 9913 fs->fs_pendinginodes -= 1; 9914 UFS_UNLOCK(ump); 9915 LIST_INIT(&wkhd); 9916 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9917 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9918 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9919 softdep_error("handle_workitem_freefile", error); 9920 ACQUIRE_LOCK(ump); 9921 WORKITEM_FREE(freefile, D_FREEFILE); 9922 FREE_LOCK(ump); 9923 } 9924 9925 9926 /* 9927 * Helper function which unlinks marker element from work list and returns 9928 * the next element on the list. 9929 */ 9930 static __inline struct worklist * 9931 markernext(struct worklist *marker) 9932 { 9933 struct worklist *next; 9934 9935 next = LIST_NEXT(marker, wk_list); 9936 LIST_REMOVE(marker, wk_list); 9937 return next; 9938 } 9939 9940 /* 9941 * Disk writes. 9942 * 9943 * The dependency structures constructed above are most actively used when file 9944 * system blocks are written to disk. No constraints are placed on when a 9945 * block can be written, but unsatisfied update dependencies are made safe by 9946 * modifying (or replacing) the source memory for the duration of the disk 9947 * write. When the disk write completes, the memory block is again brought 9948 * up-to-date. 9949 * 9950 * In-core inode structure reclamation. 9951 * 9952 * Because there are a finite number of "in-core" inode structures, they are 9953 * reused regularly. By transferring all inode-related dependencies to the 9954 * in-memory inode block and indexing them separately (via "inodedep"s), we 9955 * can allow "in-core" inode structures to be reused at any time and avoid 9956 * any increase in contention. 9957 * 9958 * Called just before entering the device driver to initiate a new disk I/O. 9959 * The buffer must be locked, thus, no I/O completion operations can occur 9960 * while we are manipulating its associated dependencies. 9961 */ 9962 static void 9963 softdep_disk_io_initiation(bp) 9964 struct buf *bp; /* structure describing disk write to occur */ 9965 { 9966 struct worklist *wk; 9967 struct worklist marker; 9968 struct inodedep *inodedep; 9969 struct freeblks *freeblks; 9970 struct jblkdep *jblkdep; 9971 struct newblk *newblk; 9972 struct ufsmount *ump; 9973 9974 /* 9975 * We only care about write operations. There should never 9976 * be dependencies for reads. 9977 */ 9978 if (bp->b_iocmd != BIO_WRITE) 9979 panic("softdep_disk_io_initiation: not write"); 9980 9981 if (bp->b_vflags & BV_BKGRDINPROG) 9982 panic("softdep_disk_io_initiation: Writing buffer with " 9983 "background write in progress: %p", bp); 9984 9985 ump = softdep_bp_to_mp(bp); 9986 if (ump == NULL) 9987 return; 9988 9989 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9990 PHOLD(curproc); /* Don't swap out kernel stack */ 9991 ACQUIRE_LOCK(ump); 9992 /* 9993 * Do any necessary pre-I/O processing. 9994 */ 9995 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9996 wk = markernext(&marker)) { 9997 LIST_INSERT_AFTER(wk, &marker, wk_list); 9998 switch (wk->wk_type) { 9999 10000 case D_PAGEDEP: 10001 initiate_write_filepage(WK_PAGEDEP(wk), bp); 10002 continue; 10003 10004 case D_INODEDEP: 10005 inodedep = WK_INODEDEP(wk); 10006 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 10007 initiate_write_inodeblock_ufs1(inodedep, bp); 10008 else 10009 initiate_write_inodeblock_ufs2(inodedep, bp); 10010 continue; 10011 10012 case D_INDIRDEP: 10013 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 10014 continue; 10015 10016 case D_BMSAFEMAP: 10017 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 10018 continue; 10019 10020 case D_JSEG: 10021 WK_JSEG(wk)->js_buf = NULL; 10022 continue; 10023 10024 case D_FREEBLKS: 10025 freeblks = WK_FREEBLKS(wk); 10026 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 10027 /* 10028 * We have to wait for the freeblks to be journaled 10029 * before we can write an inodeblock with updated 10030 * pointers. Be careful to arrange the marker so 10031 * we revisit the freeblks if it's not removed by 10032 * the first jwait(). 10033 */ 10034 if (jblkdep != NULL) { 10035 LIST_REMOVE(&marker, wk_list); 10036 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10037 jwait(&jblkdep->jb_list, MNT_WAIT); 10038 } 10039 continue; 10040 case D_ALLOCDIRECT: 10041 case D_ALLOCINDIR: 10042 /* 10043 * We have to wait for the jnewblk to be journaled 10044 * before we can write to a block if the contents 10045 * may be confused with an earlier file's indirect 10046 * at recovery time. Handle the marker as described 10047 * above. 10048 */ 10049 newblk = WK_NEWBLK(wk); 10050 if (newblk->nb_jnewblk != NULL && 10051 indirblk_lookup(newblk->nb_list.wk_mp, 10052 newblk->nb_newblkno)) { 10053 LIST_REMOVE(&marker, wk_list); 10054 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10055 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10056 } 10057 continue; 10058 10059 case D_SBDEP: 10060 initiate_write_sbdep(WK_SBDEP(wk)); 10061 continue; 10062 10063 case D_MKDIR: 10064 case D_FREEWORK: 10065 case D_FREEDEP: 10066 case D_JSEGDEP: 10067 continue; 10068 10069 default: 10070 panic("handle_disk_io_initiation: Unexpected type %s", 10071 TYPENAME(wk->wk_type)); 10072 /* NOTREACHED */ 10073 } 10074 } 10075 FREE_LOCK(ump); 10076 PRELE(curproc); /* Allow swapout of kernel stack */ 10077 } 10078 10079 /* 10080 * Called from within the procedure above to deal with unsatisfied 10081 * allocation dependencies in a directory. The buffer must be locked, 10082 * thus, no I/O completion operations can occur while we are 10083 * manipulating its associated dependencies. 10084 */ 10085 static void 10086 initiate_write_filepage(pagedep, bp) 10087 struct pagedep *pagedep; 10088 struct buf *bp; 10089 { 10090 struct jremref *jremref; 10091 struct jmvref *jmvref; 10092 struct dirrem *dirrem; 10093 struct diradd *dap; 10094 struct direct *ep; 10095 int i; 10096 10097 if (pagedep->pd_state & IOSTARTED) { 10098 /* 10099 * This can only happen if there is a driver that does not 10100 * understand chaining. Here biodone will reissue the call 10101 * to strategy for the incomplete buffers. 10102 */ 10103 printf("initiate_write_filepage: already started\n"); 10104 return; 10105 } 10106 pagedep->pd_state |= IOSTARTED; 10107 /* 10108 * Wait for all journal remove dependencies to hit the disk. 10109 * We can not allow any potentially conflicting directory adds 10110 * to be visible before removes and rollback is too difficult. 10111 * The per-filesystem lock may be dropped and re-acquired, however 10112 * we hold the buf locked so the dependency can not go away. 10113 */ 10114 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10115 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10116 jwait(&jremref->jr_list, MNT_WAIT); 10117 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10118 jwait(&jmvref->jm_list, MNT_WAIT); 10119 for (i = 0; i < DAHASHSZ; i++) { 10120 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10121 ep = (struct direct *) 10122 ((char *)bp->b_data + dap->da_offset); 10123 if (ep->d_ino != dap->da_newinum) 10124 panic("%s: dir inum %ju != new %ju", 10125 "initiate_write_filepage", 10126 (uintmax_t)ep->d_ino, 10127 (uintmax_t)dap->da_newinum); 10128 if (dap->da_state & DIRCHG) 10129 ep->d_ino = dap->da_previous->dm_oldinum; 10130 else 10131 ep->d_ino = 0; 10132 dap->da_state &= ~ATTACHED; 10133 dap->da_state |= UNDONE; 10134 } 10135 } 10136 } 10137 10138 /* 10139 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10140 * Note that any bug fixes made to this routine must be done in the 10141 * version found below. 10142 * 10143 * Called from within the procedure above to deal with unsatisfied 10144 * allocation dependencies in an inodeblock. The buffer must be 10145 * locked, thus, no I/O completion operations can occur while we 10146 * are manipulating its associated dependencies. 10147 */ 10148 static void 10149 initiate_write_inodeblock_ufs1(inodedep, bp) 10150 struct inodedep *inodedep; 10151 struct buf *bp; /* The inode block */ 10152 { 10153 struct allocdirect *adp, *lastadp; 10154 struct ufs1_dinode *dp; 10155 struct ufs1_dinode *sip; 10156 struct inoref *inoref; 10157 struct ufsmount *ump; 10158 struct fs *fs; 10159 ufs_lbn_t i; 10160 #ifdef INVARIANTS 10161 ufs_lbn_t prevlbn = 0; 10162 #endif 10163 int deplist; 10164 10165 if (inodedep->id_state & IOSTARTED) 10166 panic("initiate_write_inodeblock_ufs1: already started"); 10167 inodedep->id_state |= IOSTARTED; 10168 fs = inodedep->id_fs; 10169 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10170 LOCK_OWNED(ump); 10171 dp = (struct ufs1_dinode *)bp->b_data + 10172 ino_to_fsbo(fs, inodedep->id_ino); 10173 10174 /* 10175 * If we're on the unlinked list but have not yet written our 10176 * next pointer initialize it here. 10177 */ 10178 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10179 struct inodedep *inon; 10180 10181 inon = TAILQ_NEXT(inodedep, id_unlinked); 10182 dp->di_freelink = inon ? inon->id_ino : 0; 10183 } 10184 /* 10185 * If the bitmap is not yet written, then the allocated 10186 * inode cannot be written to disk. 10187 */ 10188 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10189 if (inodedep->id_savedino1 != NULL) 10190 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10191 FREE_LOCK(ump); 10192 sip = malloc(sizeof(struct ufs1_dinode), 10193 M_SAVEDINO, M_SOFTDEP_FLAGS); 10194 ACQUIRE_LOCK(ump); 10195 inodedep->id_savedino1 = sip; 10196 *inodedep->id_savedino1 = *dp; 10197 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10198 dp->di_gen = inodedep->id_savedino1->di_gen; 10199 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10200 return; 10201 } 10202 /* 10203 * If no dependencies, then there is nothing to roll back. 10204 */ 10205 inodedep->id_savedsize = dp->di_size; 10206 inodedep->id_savedextsize = 0; 10207 inodedep->id_savednlink = dp->di_nlink; 10208 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10209 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10210 return; 10211 /* 10212 * Revert the link count to that of the first unwritten journal entry. 10213 */ 10214 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10215 if (inoref) 10216 dp->di_nlink = inoref->if_nlink; 10217 /* 10218 * Set the dependencies to busy. 10219 */ 10220 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10221 adp = TAILQ_NEXT(adp, ad_next)) { 10222 #ifdef INVARIANTS 10223 if (deplist != 0 && prevlbn >= adp->ad_offset) 10224 panic("softdep_write_inodeblock: lbn order"); 10225 prevlbn = adp->ad_offset; 10226 if (adp->ad_offset < UFS_NDADDR && 10227 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10228 panic("%s: direct pointer #%jd mismatch %d != %jd", 10229 "softdep_write_inodeblock", 10230 (intmax_t)adp->ad_offset, 10231 dp->di_db[adp->ad_offset], 10232 (intmax_t)adp->ad_newblkno); 10233 if (adp->ad_offset >= UFS_NDADDR && 10234 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10235 panic("%s: indirect pointer #%jd mismatch %d != %jd", 10236 "softdep_write_inodeblock", 10237 (intmax_t)adp->ad_offset - UFS_NDADDR, 10238 dp->di_ib[adp->ad_offset - UFS_NDADDR], 10239 (intmax_t)adp->ad_newblkno); 10240 deplist |= 1 << adp->ad_offset; 10241 if ((adp->ad_state & ATTACHED) == 0) 10242 panic("softdep_write_inodeblock: Unknown state 0x%x", 10243 adp->ad_state); 10244 #endif /* INVARIANTS */ 10245 adp->ad_state &= ~ATTACHED; 10246 adp->ad_state |= UNDONE; 10247 } 10248 /* 10249 * The on-disk inode cannot claim to be any larger than the last 10250 * fragment that has been written. Otherwise, the on-disk inode 10251 * might have fragments that were not the last block in the file 10252 * which would corrupt the filesystem. 10253 */ 10254 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10255 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10256 if (adp->ad_offset >= UFS_NDADDR) 10257 break; 10258 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10259 /* keep going until hitting a rollback to a frag */ 10260 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10261 continue; 10262 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10263 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10264 #ifdef INVARIANTS 10265 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10266 panic("softdep_write_inodeblock: lost dep1"); 10267 #endif /* INVARIANTS */ 10268 dp->di_db[i] = 0; 10269 } 10270 for (i = 0; i < UFS_NIADDR; i++) { 10271 #ifdef INVARIANTS 10272 if (dp->di_ib[i] != 0 && 10273 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10274 panic("softdep_write_inodeblock: lost dep2"); 10275 #endif /* INVARIANTS */ 10276 dp->di_ib[i] = 0; 10277 } 10278 return; 10279 } 10280 /* 10281 * If we have zero'ed out the last allocated block of the file, 10282 * roll back the size to the last currently allocated block. 10283 * We know that this last allocated block is a full-sized as 10284 * we already checked for fragments in the loop above. 10285 */ 10286 if (lastadp != NULL && 10287 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10288 for (i = lastadp->ad_offset; i >= 0; i--) 10289 if (dp->di_db[i] != 0) 10290 break; 10291 dp->di_size = (i + 1) * fs->fs_bsize; 10292 } 10293 /* 10294 * The only dependencies are for indirect blocks. 10295 * 10296 * The file size for indirect block additions is not guaranteed. 10297 * Such a guarantee would be non-trivial to achieve. The conventional 10298 * synchronous write implementation also does not make this guarantee. 10299 * Fsck should catch and fix discrepancies. Arguably, the file size 10300 * can be over-estimated without destroying integrity when the file 10301 * moves into the indirect blocks (i.e., is large). If we want to 10302 * postpone fsck, we are stuck with this argument. 10303 */ 10304 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10305 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10306 } 10307 10308 /* 10309 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10310 * Note that any bug fixes made to this routine must be done in the 10311 * version found above. 10312 * 10313 * Called from within the procedure above to deal with unsatisfied 10314 * allocation dependencies in an inodeblock. The buffer must be 10315 * locked, thus, no I/O completion operations can occur while we 10316 * are manipulating its associated dependencies. 10317 */ 10318 static void 10319 initiate_write_inodeblock_ufs2(inodedep, bp) 10320 struct inodedep *inodedep; 10321 struct buf *bp; /* The inode block */ 10322 { 10323 struct allocdirect *adp, *lastadp; 10324 struct ufs2_dinode *dp; 10325 struct ufs2_dinode *sip; 10326 struct inoref *inoref; 10327 struct ufsmount *ump; 10328 struct fs *fs; 10329 ufs_lbn_t i; 10330 #ifdef INVARIANTS 10331 ufs_lbn_t prevlbn = 0; 10332 #endif 10333 int deplist; 10334 10335 if (inodedep->id_state & IOSTARTED) 10336 panic("initiate_write_inodeblock_ufs2: already started"); 10337 inodedep->id_state |= IOSTARTED; 10338 fs = inodedep->id_fs; 10339 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10340 LOCK_OWNED(ump); 10341 dp = (struct ufs2_dinode *)bp->b_data + 10342 ino_to_fsbo(fs, inodedep->id_ino); 10343 10344 /* 10345 * If we're on the unlinked list but have not yet written our 10346 * next pointer initialize it here. 10347 */ 10348 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10349 struct inodedep *inon; 10350 10351 inon = TAILQ_NEXT(inodedep, id_unlinked); 10352 dp->di_freelink = inon ? inon->id_ino : 0; 10353 } 10354 /* 10355 * If the bitmap is not yet written, then the allocated 10356 * inode cannot be written to disk. 10357 */ 10358 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10359 if (inodedep->id_savedino2 != NULL) 10360 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10361 FREE_LOCK(ump); 10362 sip = malloc(sizeof(struct ufs2_dinode), 10363 M_SAVEDINO, M_SOFTDEP_FLAGS); 10364 ACQUIRE_LOCK(ump); 10365 inodedep->id_savedino2 = sip; 10366 *inodedep->id_savedino2 = *dp; 10367 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10368 dp->di_gen = inodedep->id_savedino2->di_gen; 10369 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10370 return; 10371 } 10372 /* 10373 * If no dependencies, then there is nothing to roll back. 10374 */ 10375 inodedep->id_savedsize = dp->di_size; 10376 inodedep->id_savedextsize = dp->di_extsize; 10377 inodedep->id_savednlink = dp->di_nlink; 10378 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10379 TAILQ_EMPTY(&inodedep->id_extupdt) && 10380 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10381 return; 10382 /* 10383 * Revert the link count to that of the first unwritten journal entry. 10384 */ 10385 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10386 if (inoref) 10387 dp->di_nlink = inoref->if_nlink; 10388 10389 /* 10390 * Set the ext data dependencies to busy. 10391 */ 10392 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10393 adp = TAILQ_NEXT(adp, ad_next)) { 10394 #ifdef INVARIANTS 10395 if (deplist != 0 && prevlbn >= adp->ad_offset) 10396 panic("softdep_write_inodeblock: lbn order"); 10397 prevlbn = adp->ad_offset; 10398 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10399 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10400 "softdep_write_inodeblock", 10401 (intmax_t)adp->ad_offset, 10402 (intmax_t)dp->di_extb[adp->ad_offset], 10403 (intmax_t)adp->ad_newblkno); 10404 deplist |= 1 << adp->ad_offset; 10405 if ((adp->ad_state & ATTACHED) == 0) 10406 panic("softdep_write_inodeblock: Unknown state 0x%x", 10407 adp->ad_state); 10408 #endif /* INVARIANTS */ 10409 adp->ad_state &= ~ATTACHED; 10410 adp->ad_state |= UNDONE; 10411 } 10412 /* 10413 * The on-disk inode cannot claim to be any larger than the last 10414 * fragment that has been written. Otherwise, the on-disk inode 10415 * might have fragments that were not the last block in the ext 10416 * data which would corrupt the filesystem. 10417 */ 10418 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10419 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10420 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10421 /* keep going until hitting a rollback to a frag */ 10422 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10423 continue; 10424 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10425 for (i = adp->ad_offset + 1; i < UFS_NXADDR; i++) { 10426 #ifdef INVARIANTS 10427 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10428 panic("softdep_write_inodeblock: lost dep1"); 10429 #endif /* INVARIANTS */ 10430 dp->di_extb[i] = 0; 10431 } 10432 lastadp = NULL; 10433 break; 10434 } 10435 /* 10436 * If we have zero'ed out the last allocated block of the ext 10437 * data, roll back the size to the last currently allocated block. 10438 * We know that this last allocated block is a full-sized as 10439 * we already checked for fragments in the loop above. 10440 */ 10441 if (lastadp != NULL && 10442 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10443 for (i = lastadp->ad_offset; i >= 0; i--) 10444 if (dp->di_extb[i] != 0) 10445 break; 10446 dp->di_extsize = (i + 1) * fs->fs_bsize; 10447 } 10448 /* 10449 * Set the file data dependencies to busy. 10450 */ 10451 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10452 adp = TAILQ_NEXT(adp, ad_next)) { 10453 #ifdef INVARIANTS 10454 if (deplist != 0 && prevlbn >= adp->ad_offset) 10455 panic("softdep_write_inodeblock: lbn order"); 10456 if ((adp->ad_state & ATTACHED) == 0) 10457 panic("inodedep %p and adp %p not attached", inodedep, adp); 10458 prevlbn = adp->ad_offset; 10459 if (adp->ad_offset < UFS_NDADDR && 10460 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10461 panic("%s: direct pointer #%jd mismatch %jd != %jd", 10462 "softdep_write_inodeblock", 10463 (intmax_t)adp->ad_offset, 10464 (intmax_t)dp->di_db[adp->ad_offset], 10465 (intmax_t)adp->ad_newblkno); 10466 if (adp->ad_offset >= UFS_NDADDR && 10467 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10468 panic("%s indirect pointer #%jd mismatch %jd != %jd", 10469 "softdep_write_inodeblock:", 10470 (intmax_t)adp->ad_offset - UFS_NDADDR, 10471 (intmax_t)dp->di_ib[adp->ad_offset - UFS_NDADDR], 10472 (intmax_t)adp->ad_newblkno); 10473 deplist |= 1 << adp->ad_offset; 10474 if ((adp->ad_state & ATTACHED) == 0) 10475 panic("softdep_write_inodeblock: Unknown state 0x%x", 10476 adp->ad_state); 10477 #endif /* INVARIANTS */ 10478 adp->ad_state &= ~ATTACHED; 10479 adp->ad_state |= UNDONE; 10480 } 10481 /* 10482 * The on-disk inode cannot claim to be any larger than the last 10483 * fragment that has been written. Otherwise, the on-disk inode 10484 * might have fragments that were not the last block in the file 10485 * which would corrupt the filesystem. 10486 */ 10487 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10488 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10489 if (adp->ad_offset >= UFS_NDADDR) 10490 break; 10491 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10492 /* keep going until hitting a rollback to a frag */ 10493 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10494 continue; 10495 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10496 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10497 #ifdef INVARIANTS 10498 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10499 panic("softdep_write_inodeblock: lost dep2"); 10500 #endif /* INVARIANTS */ 10501 dp->di_db[i] = 0; 10502 } 10503 for (i = 0; i < UFS_NIADDR; i++) { 10504 #ifdef INVARIANTS 10505 if (dp->di_ib[i] != 0 && 10506 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10507 panic("softdep_write_inodeblock: lost dep3"); 10508 #endif /* INVARIANTS */ 10509 dp->di_ib[i] = 0; 10510 } 10511 return; 10512 } 10513 /* 10514 * If we have zero'ed out the last allocated block of the file, 10515 * roll back the size to the last currently allocated block. 10516 * We know that this last allocated block is a full-sized as 10517 * we already checked for fragments in the loop above. 10518 */ 10519 if (lastadp != NULL && 10520 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10521 for (i = lastadp->ad_offset; i >= 0; i--) 10522 if (dp->di_db[i] != 0) 10523 break; 10524 dp->di_size = (i + 1) * fs->fs_bsize; 10525 } 10526 /* 10527 * The only dependencies are for indirect blocks. 10528 * 10529 * The file size for indirect block additions is not guaranteed. 10530 * Such a guarantee would be non-trivial to achieve. The conventional 10531 * synchronous write implementation also does not make this guarantee. 10532 * Fsck should catch and fix discrepancies. Arguably, the file size 10533 * can be over-estimated without destroying integrity when the file 10534 * moves into the indirect blocks (i.e., is large). If we want to 10535 * postpone fsck, we are stuck with this argument. 10536 */ 10537 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10538 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10539 } 10540 10541 /* 10542 * Cancel an indirdep as a result of truncation. Release all of the 10543 * children allocindirs and place their journal work on the appropriate 10544 * list. 10545 */ 10546 static void 10547 cancel_indirdep(indirdep, bp, freeblks) 10548 struct indirdep *indirdep; 10549 struct buf *bp; 10550 struct freeblks *freeblks; 10551 { 10552 struct allocindir *aip; 10553 10554 /* 10555 * None of the indirect pointers will ever be visible, 10556 * so they can simply be tossed. GOINGAWAY ensures 10557 * that allocated pointers will be saved in the buffer 10558 * cache until they are freed. Note that they will 10559 * only be able to be found by their physical address 10560 * since the inode mapping the logical address will 10561 * be gone. The save buffer used for the safe copy 10562 * was allocated in setup_allocindir_phase2 using 10563 * the physical address so it could be used for this 10564 * purpose. Hence we swap the safe copy with the real 10565 * copy, allowing the safe copy to be freed and holding 10566 * on to the real copy for later use in indir_trunc. 10567 */ 10568 if (indirdep->ir_state & GOINGAWAY) 10569 panic("cancel_indirdep: already gone"); 10570 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10571 indirdep->ir_state |= DEPCOMPLETE; 10572 LIST_REMOVE(indirdep, ir_next); 10573 } 10574 indirdep->ir_state |= GOINGAWAY; 10575 /* 10576 * Pass in bp for blocks still have journal writes 10577 * pending so we can cancel them on their own. 10578 */ 10579 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 10580 cancel_allocindir(aip, bp, freeblks, 0); 10581 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) 10582 cancel_allocindir(aip, NULL, freeblks, 0); 10583 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) 10584 cancel_allocindir(aip, NULL, freeblks, 0); 10585 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) 10586 cancel_allocindir(aip, NULL, freeblks, 0); 10587 /* 10588 * If there are pending partial truncations we need to keep the 10589 * old block copy around until they complete. This is because 10590 * the current b_data is not a perfect superset of the available 10591 * blocks. 10592 */ 10593 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10594 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10595 else 10596 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10597 WORKLIST_REMOVE(&indirdep->ir_list); 10598 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10599 indirdep->ir_bp = NULL; 10600 indirdep->ir_freeblks = freeblks; 10601 } 10602 10603 /* 10604 * Free an indirdep once it no longer has new pointers to track. 10605 */ 10606 static void 10607 free_indirdep(indirdep) 10608 struct indirdep *indirdep; 10609 { 10610 10611 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10612 ("free_indirdep: Indir trunc list not empty.")); 10613 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10614 ("free_indirdep: Complete head not empty.")); 10615 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10616 ("free_indirdep: write head not empty.")); 10617 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10618 ("free_indirdep: done head not empty.")); 10619 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10620 ("free_indirdep: deplist head not empty.")); 10621 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10622 ("free_indirdep: %p still on newblk list.", indirdep)); 10623 KASSERT(indirdep->ir_saveddata == NULL, 10624 ("free_indirdep: %p still has saved data.", indirdep)); 10625 if (indirdep->ir_state & ONWORKLIST) 10626 WORKLIST_REMOVE(&indirdep->ir_list); 10627 WORKITEM_FREE(indirdep, D_INDIRDEP); 10628 } 10629 10630 /* 10631 * Called before a write to an indirdep. This routine is responsible for 10632 * rolling back pointers to a safe state which includes only those 10633 * allocindirs which have been completed. 10634 */ 10635 static void 10636 initiate_write_indirdep(indirdep, bp) 10637 struct indirdep *indirdep; 10638 struct buf *bp; 10639 { 10640 struct ufsmount *ump; 10641 10642 indirdep->ir_state |= IOSTARTED; 10643 if (indirdep->ir_state & GOINGAWAY) 10644 panic("disk_io_initiation: indirdep gone"); 10645 /* 10646 * If there are no remaining dependencies, this will be writing 10647 * the real pointers. 10648 */ 10649 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10650 TAILQ_EMPTY(&indirdep->ir_trunc)) 10651 return; 10652 /* 10653 * Replace up-to-date version with safe version. 10654 */ 10655 if (indirdep->ir_saveddata == NULL) { 10656 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10657 LOCK_OWNED(ump); 10658 FREE_LOCK(ump); 10659 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10660 M_SOFTDEP_FLAGS); 10661 ACQUIRE_LOCK(ump); 10662 } 10663 indirdep->ir_state &= ~ATTACHED; 10664 indirdep->ir_state |= UNDONE; 10665 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10666 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10667 bp->b_bcount); 10668 } 10669 10670 /* 10671 * Called when an inode has been cleared in a cg bitmap. This finally 10672 * eliminates any canceled jaddrefs 10673 */ 10674 void 10675 softdep_setup_inofree(mp, bp, ino, wkhd) 10676 struct mount *mp; 10677 struct buf *bp; 10678 ino_t ino; 10679 struct workhead *wkhd; 10680 { 10681 struct worklist *wk, *wkn; 10682 struct inodedep *inodedep; 10683 struct ufsmount *ump; 10684 uint8_t *inosused; 10685 struct cg *cgp; 10686 struct fs *fs; 10687 10688 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10689 ("softdep_setup_inofree called on non-softdep filesystem")); 10690 ump = VFSTOUFS(mp); 10691 ACQUIRE_LOCK(ump); 10692 fs = ump->um_fs; 10693 cgp = (struct cg *)bp->b_data; 10694 inosused = cg_inosused(cgp); 10695 if (isset(inosused, ino % fs->fs_ipg)) 10696 panic("softdep_setup_inofree: inode %ju not freed.", 10697 (uintmax_t)ino); 10698 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10699 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10700 (uintmax_t)ino, inodedep); 10701 if (wkhd) { 10702 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10703 if (wk->wk_type != D_JADDREF) 10704 continue; 10705 WORKLIST_REMOVE(wk); 10706 /* 10707 * We can free immediately even if the jaddref 10708 * isn't attached in a background write as now 10709 * the bitmaps are reconciled. 10710 */ 10711 wk->wk_state |= COMPLETE | ATTACHED; 10712 free_jaddref(WK_JADDREF(wk)); 10713 } 10714 jwork_move(&bp->b_dep, wkhd); 10715 } 10716 FREE_LOCK(ump); 10717 } 10718 10719 10720 /* 10721 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10722 * map. Any dependencies waiting for the write to clear are added to the 10723 * buf's list and any jnewblks that are being canceled are discarded 10724 * immediately. 10725 */ 10726 void 10727 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10728 struct mount *mp; 10729 struct buf *bp; 10730 ufs2_daddr_t blkno; 10731 int frags; 10732 struct workhead *wkhd; 10733 { 10734 struct bmsafemap *bmsafemap; 10735 struct jnewblk *jnewblk; 10736 struct ufsmount *ump; 10737 struct worklist *wk; 10738 struct fs *fs; 10739 #ifdef SUJ_DEBUG 10740 uint8_t *blksfree; 10741 struct cg *cgp; 10742 ufs2_daddr_t jstart; 10743 ufs2_daddr_t jend; 10744 ufs2_daddr_t end; 10745 long bno; 10746 int i; 10747 #endif 10748 10749 CTR3(KTR_SUJ, 10750 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10751 blkno, frags, wkhd); 10752 10753 ump = VFSTOUFS(mp); 10754 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10755 ("softdep_setup_blkfree called on non-softdep filesystem")); 10756 ACQUIRE_LOCK(ump); 10757 /* Lookup the bmsafemap so we track when it is dirty. */ 10758 fs = ump->um_fs; 10759 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10760 /* 10761 * Detach any jnewblks which have been canceled. They must linger 10762 * until the bitmap is cleared again by ffs_blkfree() to prevent 10763 * an unjournaled allocation from hitting the disk. 10764 */ 10765 if (wkhd) { 10766 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10767 CTR2(KTR_SUJ, 10768 "softdep_setup_blkfree: blkno %jd wk type %d", 10769 blkno, wk->wk_type); 10770 WORKLIST_REMOVE(wk); 10771 if (wk->wk_type != D_JNEWBLK) { 10772 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10773 continue; 10774 } 10775 jnewblk = WK_JNEWBLK(wk); 10776 KASSERT(jnewblk->jn_state & GOINGAWAY, 10777 ("softdep_setup_blkfree: jnewblk not canceled.")); 10778 #ifdef SUJ_DEBUG 10779 /* 10780 * Assert that this block is free in the bitmap 10781 * before we discard the jnewblk. 10782 */ 10783 cgp = (struct cg *)bp->b_data; 10784 blksfree = cg_blksfree(cgp); 10785 bno = dtogd(fs, jnewblk->jn_blkno); 10786 for (i = jnewblk->jn_oldfrags; 10787 i < jnewblk->jn_frags; i++) { 10788 if (isset(blksfree, bno + i)) 10789 continue; 10790 panic("softdep_setup_blkfree: not free"); 10791 } 10792 #endif 10793 /* 10794 * Even if it's not attached we can free immediately 10795 * as the new bitmap is correct. 10796 */ 10797 wk->wk_state |= COMPLETE | ATTACHED; 10798 free_jnewblk(jnewblk); 10799 } 10800 } 10801 10802 #ifdef SUJ_DEBUG 10803 /* 10804 * Assert that we are not freeing a block which has an outstanding 10805 * allocation dependency. 10806 */ 10807 fs = VFSTOUFS(mp)->um_fs; 10808 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10809 end = blkno + frags; 10810 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10811 /* 10812 * Don't match against blocks that will be freed when the 10813 * background write is done. 10814 */ 10815 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10816 (COMPLETE | DEPCOMPLETE)) 10817 continue; 10818 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10819 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10820 if ((blkno >= jstart && blkno < jend) || 10821 (end > jstart && end <= jend)) { 10822 printf("state 0x%X %jd - %d %d dep %p\n", 10823 jnewblk->jn_state, jnewblk->jn_blkno, 10824 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10825 jnewblk->jn_dep); 10826 panic("softdep_setup_blkfree: " 10827 "%jd-%jd(%d) overlaps with %jd-%jd", 10828 blkno, end, frags, jstart, jend); 10829 } 10830 } 10831 #endif 10832 FREE_LOCK(ump); 10833 } 10834 10835 /* 10836 * Revert a block allocation when the journal record that describes it 10837 * is not yet written. 10838 */ 10839 static int 10840 jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10841 struct jnewblk *jnewblk; 10842 struct fs *fs; 10843 struct cg *cgp; 10844 uint8_t *blksfree; 10845 { 10846 ufs1_daddr_t fragno; 10847 long cgbno, bbase; 10848 int frags, blk; 10849 int i; 10850 10851 frags = 0; 10852 cgbno = dtogd(fs, jnewblk->jn_blkno); 10853 /* 10854 * We have to test which frags need to be rolled back. We may 10855 * be operating on a stale copy when doing background writes. 10856 */ 10857 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10858 if (isclr(blksfree, cgbno + i)) 10859 frags++; 10860 if (frags == 0) 10861 return (0); 10862 /* 10863 * This is mostly ffs_blkfree() sans some validation and 10864 * superblock updates. 10865 */ 10866 if (frags == fs->fs_frag) { 10867 fragno = fragstoblks(fs, cgbno); 10868 ffs_setblock(fs, blksfree, fragno); 10869 ffs_clusteracct(fs, cgp, fragno, 1); 10870 cgp->cg_cs.cs_nbfree++; 10871 } else { 10872 cgbno += jnewblk->jn_oldfrags; 10873 bbase = cgbno - fragnum(fs, cgbno); 10874 /* Decrement the old frags. */ 10875 blk = blkmap(fs, blksfree, bbase); 10876 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10877 /* Deallocate the fragment */ 10878 for (i = 0; i < frags; i++) 10879 setbit(blksfree, cgbno + i); 10880 cgp->cg_cs.cs_nffree += frags; 10881 /* Add back in counts associated with the new frags */ 10882 blk = blkmap(fs, blksfree, bbase); 10883 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10884 /* If a complete block has been reassembled, account for it. */ 10885 fragno = fragstoblks(fs, bbase); 10886 if (ffs_isblock(fs, blksfree, fragno)) { 10887 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10888 ffs_clusteracct(fs, cgp, fragno, 1); 10889 cgp->cg_cs.cs_nbfree++; 10890 } 10891 } 10892 stat_jnewblk++; 10893 jnewblk->jn_state &= ~ATTACHED; 10894 jnewblk->jn_state |= UNDONE; 10895 10896 return (frags); 10897 } 10898 10899 static void 10900 initiate_write_bmsafemap(bmsafemap, bp) 10901 struct bmsafemap *bmsafemap; 10902 struct buf *bp; /* The cg block. */ 10903 { 10904 struct jaddref *jaddref; 10905 struct jnewblk *jnewblk; 10906 uint8_t *inosused; 10907 uint8_t *blksfree; 10908 struct cg *cgp; 10909 struct fs *fs; 10910 ino_t ino; 10911 10912 /* 10913 * If this is a background write, we did this at the time that 10914 * the copy was made, so do not need to do it again. 10915 */ 10916 if (bmsafemap->sm_state & IOSTARTED) 10917 return; 10918 bmsafemap->sm_state |= IOSTARTED; 10919 /* 10920 * Clear any inode allocations which are pending journal writes. 10921 */ 10922 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10923 cgp = (struct cg *)bp->b_data; 10924 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10925 inosused = cg_inosused(cgp); 10926 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10927 ino = jaddref->ja_ino % fs->fs_ipg; 10928 if (isset(inosused, ino)) { 10929 if ((jaddref->ja_mode & IFMT) == IFDIR) 10930 cgp->cg_cs.cs_ndir--; 10931 cgp->cg_cs.cs_nifree++; 10932 clrbit(inosused, ino); 10933 jaddref->ja_state &= ~ATTACHED; 10934 jaddref->ja_state |= UNDONE; 10935 stat_jaddref++; 10936 } else 10937 panic("initiate_write_bmsafemap: inode %ju " 10938 "marked free", (uintmax_t)jaddref->ja_ino); 10939 } 10940 } 10941 /* 10942 * Clear any block allocations which are pending journal writes. 10943 */ 10944 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10945 cgp = (struct cg *)bp->b_data; 10946 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10947 blksfree = cg_blksfree(cgp); 10948 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10949 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10950 continue; 10951 panic("initiate_write_bmsafemap: block %jd " 10952 "marked free", jnewblk->jn_blkno); 10953 } 10954 } 10955 /* 10956 * Move allocation lists to the written lists so they can be 10957 * cleared once the block write is complete. 10958 */ 10959 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10960 inodedep, id_deps); 10961 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10962 newblk, nb_deps); 10963 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10964 wk_list); 10965 } 10966 10967 /* 10968 * This routine is called during the completion interrupt 10969 * service routine for a disk write (from the procedure called 10970 * by the device driver to inform the filesystem caches of 10971 * a request completion). It should be called early in this 10972 * procedure, before the block is made available to other 10973 * processes or other routines are called. 10974 * 10975 */ 10976 static void 10977 softdep_disk_write_complete(bp) 10978 struct buf *bp; /* describes the completed disk write */ 10979 { 10980 struct worklist *wk; 10981 struct worklist *owk; 10982 struct ufsmount *ump; 10983 struct workhead reattach; 10984 struct freeblks *freeblks; 10985 struct buf *sbp; 10986 10987 ump = softdep_bp_to_mp(bp); 10988 if (ump == NULL) 10989 return; 10990 10991 sbp = NULL; 10992 10993 /* 10994 * If an error occurred while doing the write, then the data 10995 * has not hit the disk and the dependencies cannot be processed. 10996 * But we do have to go through and roll forward any dependencies 10997 * that were rolled back before the disk write. 10998 */ 10999 ACQUIRE_LOCK(ump); 11000 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) { 11001 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11002 switch (wk->wk_type) { 11003 11004 case D_PAGEDEP: 11005 handle_written_filepage(WK_PAGEDEP(wk), bp, 0); 11006 continue; 11007 11008 case D_INODEDEP: 11009 handle_written_inodeblock(WK_INODEDEP(wk), 11010 bp, 0); 11011 continue; 11012 11013 case D_BMSAFEMAP: 11014 handle_written_bmsafemap(WK_BMSAFEMAP(wk), 11015 bp, 0); 11016 continue; 11017 11018 case D_INDIRDEP: 11019 handle_written_indirdep(WK_INDIRDEP(wk), 11020 bp, &sbp, 0); 11021 continue; 11022 default: 11023 /* nothing to roll forward */ 11024 continue; 11025 } 11026 } 11027 FREE_LOCK(ump); 11028 return; 11029 } 11030 LIST_INIT(&reattach); 11031 11032 /* 11033 * Ump SU lock must not be released anywhere in this code segment. 11034 */ 11035 owk = NULL; 11036 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 11037 WORKLIST_REMOVE(wk); 11038 atomic_add_long(&dep_write[wk->wk_type], 1); 11039 if (wk == owk) 11040 panic("duplicate worklist: %p\n", wk); 11041 owk = wk; 11042 switch (wk->wk_type) { 11043 11044 case D_PAGEDEP: 11045 if (handle_written_filepage(WK_PAGEDEP(wk), bp, 11046 WRITESUCCEEDED)) 11047 WORKLIST_INSERT(&reattach, wk); 11048 continue; 11049 11050 case D_INODEDEP: 11051 if (handle_written_inodeblock(WK_INODEDEP(wk), bp, 11052 WRITESUCCEEDED)) 11053 WORKLIST_INSERT(&reattach, wk); 11054 continue; 11055 11056 case D_BMSAFEMAP: 11057 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp, 11058 WRITESUCCEEDED)) 11059 WORKLIST_INSERT(&reattach, wk); 11060 continue; 11061 11062 case D_MKDIR: 11063 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 11064 continue; 11065 11066 case D_ALLOCDIRECT: 11067 wk->wk_state |= COMPLETE; 11068 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 11069 continue; 11070 11071 case D_ALLOCINDIR: 11072 wk->wk_state |= COMPLETE; 11073 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 11074 continue; 11075 11076 case D_INDIRDEP: 11077 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp, 11078 WRITESUCCEEDED)) 11079 WORKLIST_INSERT(&reattach, wk); 11080 continue; 11081 11082 case D_FREEBLKS: 11083 wk->wk_state |= COMPLETE; 11084 freeblks = WK_FREEBLKS(wk); 11085 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 11086 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11087 add_to_worklist(wk, WK_NODELAY); 11088 continue; 11089 11090 case D_FREEWORK: 11091 handle_written_freework(WK_FREEWORK(wk)); 11092 break; 11093 11094 case D_JSEGDEP: 11095 free_jsegdep(WK_JSEGDEP(wk)); 11096 continue; 11097 11098 case D_JSEG: 11099 handle_written_jseg(WK_JSEG(wk), bp); 11100 continue; 11101 11102 case D_SBDEP: 11103 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11104 WORKLIST_INSERT(&reattach, wk); 11105 continue; 11106 11107 case D_FREEDEP: 11108 free_freedep(WK_FREEDEP(wk)); 11109 continue; 11110 11111 default: 11112 panic("handle_disk_write_complete: Unknown type %s", 11113 TYPENAME(wk->wk_type)); 11114 /* NOTREACHED */ 11115 } 11116 } 11117 /* 11118 * Reattach any requests that must be redone. 11119 */ 11120 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11121 WORKLIST_REMOVE(wk); 11122 WORKLIST_INSERT(&bp->b_dep, wk); 11123 } 11124 FREE_LOCK(ump); 11125 if (sbp) 11126 brelse(sbp); 11127 } 11128 11129 /* 11130 * Called from within softdep_disk_write_complete above. Note that 11131 * this routine is always called from interrupt level with further 11132 * splbio interrupts blocked. 11133 */ 11134 static void 11135 handle_allocdirect_partdone(adp, wkhd) 11136 struct allocdirect *adp; /* the completed allocdirect */ 11137 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11138 { 11139 struct allocdirectlst *listhead; 11140 struct allocdirect *listadp; 11141 struct inodedep *inodedep; 11142 long bsize; 11143 11144 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11145 return; 11146 /* 11147 * The on-disk inode cannot claim to be any larger than the last 11148 * fragment that has been written. Otherwise, the on-disk inode 11149 * might have fragments that were not the last block in the file 11150 * which would corrupt the filesystem. Thus, we cannot free any 11151 * allocdirects after one whose ad_oldblkno claims a fragment as 11152 * these blocks must be rolled back to zero before writing the inode. 11153 * We check the currently active set of allocdirects in id_inoupdt 11154 * or id_extupdt as appropriate. 11155 */ 11156 inodedep = adp->ad_inodedep; 11157 bsize = inodedep->id_fs->fs_bsize; 11158 if (adp->ad_state & EXTDATA) 11159 listhead = &inodedep->id_extupdt; 11160 else 11161 listhead = &inodedep->id_inoupdt; 11162 TAILQ_FOREACH(listadp, listhead, ad_next) { 11163 /* found our block */ 11164 if (listadp == adp) 11165 break; 11166 /* continue if ad_oldlbn is not a fragment */ 11167 if (listadp->ad_oldsize == 0 || 11168 listadp->ad_oldsize == bsize) 11169 continue; 11170 /* hit a fragment */ 11171 return; 11172 } 11173 /* 11174 * If we have reached the end of the current list without 11175 * finding the just finished dependency, then it must be 11176 * on the future dependency list. Future dependencies cannot 11177 * be freed until they are moved to the current list. 11178 */ 11179 if (listadp == NULL) { 11180 #ifdef DEBUG 11181 if (adp->ad_state & EXTDATA) 11182 listhead = &inodedep->id_newextupdt; 11183 else 11184 listhead = &inodedep->id_newinoupdt; 11185 TAILQ_FOREACH(listadp, listhead, ad_next) 11186 /* found our block */ 11187 if (listadp == adp) 11188 break; 11189 if (listadp == NULL) 11190 panic("handle_allocdirect_partdone: lost dep"); 11191 #endif /* DEBUG */ 11192 return; 11193 } 11194 /* 11195 * If we have found the just finished dependency, then queue 11196 * it along with anything that follows it that is complete. 11197 * Since the pointer has not yet been written in the inode 11198 * as the dependency prevents it, place the allocdirect on the 11199 * bufwait list where it will be freed once the pointer is 11200 * valid. 11201 */ 11202 if (wkhd == NULL) 11203 wkhd = &inodedep->id_bufwait; 11204 for (; adp; adp = listadp) { 11205 listadp = TAILQ_NEXT(adp, ad_next); 11206 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11207 return; 11208 TAILQ_REMOVE(listhead, adp, ad_next); 11209 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11210 } 11211 } 11212 11213 /* 11214 * Called from within softdep_disk_write_complete above. This routine 11215 * completes successfully written allocindirs. 11216 */ 11217 static void 11218 handle_allocindir_partdone(aip) 11219 struct allocindir *aip; /* the completed allocindir */ 11220 { 11221 struct indirdep *indirdep; 11222 11223 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11224 return; 11225 indirdep = aip->ai_indirdep; 11226 LIST_REMOVE(aip, ai_next); 11227 /* 11228 * Don't set a pointer while the buffer is undergoing IO or while 11229 * we have active truncations. 11230 */ 11231 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11232 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11233 return; 11234 } 11235 if (indirdep->ir_state & UFS1FMT) 11236 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11237 aip->ai_newblkno; 11238 else 11239 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11240 aip->ai_newblkno; 11241 /* 11242 * Await the pointer write before freeing the allocindir. 11243 */ 11244 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11245 } 11246 11247 /* 11248 * Release segments held on a jwork list. 11249 */ 11250 static void 11251 handle_jwork(wkhd) 11252 struct workhead *wkhd; 11253 { 11254 struct worklist *wk; 11255 11256 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11257 WORKLIST_REMOVE(wk); 11258 switch (wk->wk_type) { 11259 case D_JSEGDEP: 11260 free_jsegdep(WK_JSEGDEP(wk)); 11261 continue; 11262 case D_FREEDEP: 11263 free_freedep(WK_FREEDEP(wk)); 11264 continue; 11265 case D_FREEFRAG: 11266 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11267 WORKITEM_FREE(wk, D_FREEFRAG); 11268 continue; 11269 case D_FREEWORK: 11270 handle_written_freework(WK_FREEWORK(wk)); 11271 continue; 11272 default: 11273 panic("handle_jwork: Unknown type %s\n", 11274 TYPENAME(wk->wk_type)); 11275 } 11276 } 11277 } 11278 11279 /* 11280 * Handle the bufwait list on an inode when it is safe to release items 11281 * held there. This normally happens after an inode block is written but 11282 * may be delayed and handled later if there are pending journal items that 11283 * are not yet safe to be released. 11284 */ 11285 static struct freefile * 11286 handle_bufwait(inodedep, refhd) 11287 struct inodedep *inodedep; 11288 struct workhead *refhd; 11289 { 11290 struct jaddref *jaddref; 11291 struct freefile *freefile; 11292 struct worklist *wk; 11293 11294 freefile = NULL; 11295 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11296 WORKLIST_REMOVE(wk); 11297 switch (wk->wk_type) { 11298 case D_FREEFILE: 11299 /* 11300 * We defer adding freefile to the worklist 11301 * until all other additions have been made to 11302 * ensure that it will be done after all the 11303 * old blocks have been freed. 11304 */ 11305 if (freefile != NULL) 11306 panic("handle_bufwait: freefile"); 11307 freefile = WK_FREEFILE(wk); 11308 continue; 11309 11310 case D_MKDIR: 11311 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11312 continue; 11313 11314 case D_DIRADD: 11315 diradd_inode_written(WK_DIRADD(wk), inodedep); 11316 continue; 11317 11318 case D_FREEFRAG: 11319 wk->wk_state |= COMPLETE; 11320 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11321 add_to_worklist(wk, 0); 11322 continue; 11323 11324 case D_DIRREM: 11325 wk->wk_state |= COMPLETE; 11326 add_to_worklist(wk, 0); 11327 continue; 11328 11329 case D_ALLOCDIRECT: 11330 case D_ALLOCINDIR: 11331 free_newblk(WK_NEWBLK(wk)); 11332 continue; 11333 11334 case D_JNEWBLK: 11335 wk->wk_state |= COMPLETE; 11336 free_jnewblk(WK_JNEWBLK(wk)); 11337 continue; 11338 11339 /* 11340 * Save freed journal segments and add references on 11341 * the supplied list which will delay their release 11342 * until the cg bitmap is cleared on disk. 11343 */ 11344 case D_JSEGDEP: 11345 if (refhd == NULL) 11346 free_jsegdep(WK_JSEGDEP(wk)); 11347 else 11348 WORKLIST_INSERT(refhd, wk); 11349 continue; 11350 11351 case D_JADDREF: 11352 jaddref = WK_JADDREF(wk); 11353 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11354 if_deps); 11355 /* 11356 * Transfer any jaddrefs to the list to be freed with 11357 * the bitmap if we're handling a removed file. 11358 */ 11359 if (refhd == NULL) { 11360 wk->wk_state |= COMPLETE; 11361 free_jaddref(jaddref); 11362 } else 11363 WORKLIST_INSERT(refhd, wk); 11364 continue; 11365 11366 default: 11367 panic("handle_bufwait: Unknown type %p(%s)", 11368 wk, TYPENAME(wk->wk_type)); 11369 /* NOTREACHED */ 11370 } 11371 } 11372 return (freefile); 11373 } 11374 /* 11375 * Called from within softdep_disk_write_complete above to restore 11376 * in-memory inode block contents to their most up-to-date state. Note 11377 * that this routine is always called from interrupt level with further 11378 * interrupts from this device blocked. 11379 * 11380 * If the write did not succeed, we will do all the roll-forward 11381 * operations, but we will not take the actions that will allow its 11382 * dependencies to be processed. 11383 */ 11384 static int 11385 handle_written_inodeblock(inodedep, bp, flags) 11386 struct inodedep *inodedep; 11387 struct buf *bp; /* buffer containing the inode block */ 11388 int flags; 11389 { 11390 struct freefile *freefile; 11391 struct allocdirect *adp, *nextadp; 11392 struct ufs1_dinode *dp1 = NULL; 11393 struct ufs2_dinode *dp2 = NULL; 11394 struct workhead wkhd; 11395 int hadchanges, fstype; 11396 ino_t freelink; 11397 11398 LIST_INIT(&wkhd); 11399 hadchanges = 0; 11400 freefile = NULL; 11401 if ((inodedep->id_state & IOSTARTED) == 0) 11402 panic("handle_written_inodeblock: not started"); 11403 inodedep->id_state &= ~IOSTARTED; 11404 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11405 fstype = UFS1; 11406 dp1 = (struct ufs1_dinode *)bp->b_data + 11407 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11408 freelink = dp1->di_freelink; 11409 } else { 11410 fstype = UFS2; 11411 dp2 = (struct ufs2_dinode *)bp->b_data + 11412 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11413 freelink = dp2->di_freelink; 11414 } 11415 /* 11416 * Leave this inodeblock dirty until it's in the list. 11417 */ 11418 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED && 11419 (flags & WRITESUCCEEDED)) { 11420 struct inodedep *inon; 11421 11422 inon = TAILQ_NEXT(inodedep, id_unlinked); 11423 if ((inon == NULL && freelink == 0) || 11424 (inon && inon->id_ino == freelink)) { 11425 if (inon) 11426 inon->id_state |= UNLINKPREV; 11427 inodedep->id_state |= UNLINKNEXT; 11428 } 11429 hadchanges = 1; 11430 } 11431 /* 11432 * If we had to rollback the inode allocation because of 11433 * bitmaps being incomplete, then simply restore it. 11434 * Keep the block dirty so that it will not be reclaimed until 11435 * all associated dependencies have been cleared and the 11436 * corresponding updates written to disk. 11437 */ 11438 if (inodedep->id_savedino1 != NULL) { 11439 hadchanges = 1; 11440 if (fstype == UFS1) 11441 *dp1 = *inodedep->id_savedino1; 11442 else 11443 *dp2 = *inodedep->id_savedino2; 11444 free(inodedep->id_savedino1, M_SAVEDINO); 11445 inodedep->id_savedino1 = NULL; 11446 if ((bp->b_flags & B_DELWRI) == 0) 11447 stat_inode_bitmap++; 11448 bdirty(bp); 11449 /* 11450 * If the inode is clear here and GOINGAWAY it will never 11451 * be written. Process the bufwait and clear any pending 11452 * work which may include the freefile. 11453 */ 11454 if (inodedep->id_state & GOINGAWAY) 11455 goto bufwait; 11456 return (1); 11457 } 11458 if (flags & WRITESUCCEEDED) 11459 inodedep->id_state |= COMPLETE; 11460 /* 11461 * Roll forward anything that had to be rolled back before 11462 * the inode could be updated. 11463 */ 11464 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11465 nextadp = TAILQ_NEXT(adp, ad_next); 11466 if (adp->ad_state & ATTACHED) 11467 panic("handle_written_inodeblock: new entry"); 11468 if (fstype == UFS1) { 11469 if (adp->ad_offset < UFS_NDADDR) { 11470 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11471 panic("%s %s #%jd mismatch %d != %jd", 11472 "handle_written_inodeblock:", 11473 "direct pointer", 11474 (intmax_t)adp->ad_offset, 11475 dp1->di_db[adp->ad_offset], 11476 (intmax_t)adp->ad_oldblkno); 11477 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11478 } else { 11479 if (dp1->di_ib[adp->ad_offset - UFS_NDADDR] != 11480 0) 11481 panic("%s: %s #%jd allocated as %d", 11482 "handle_written_inodeblock", 11483 "indirect pointer", 11484 (intmax_t)adp->ad_offset - 11485 UFS_NDADDR, 11486 dp1->di_ib[adp->ad_offset - 11487 UFS_NDADDR]); 11488 dp1->di_ib[adp->ad_offset - UFS_NDADDR] = 11489 adp->ad_newblkno; 11490 } 11491 } else { 11492 if (adp->ad_offset < UFS_NDADDR) { 11493 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11494 panic("%s: %s #%jd %s %jd != %jd", 11495 "handle_written_inodeblock", 11496 "direct pointer", 11497 (intmax_t)adp->ad_offset, "mismatch", 11498 (intmax_t)dp2->di_db[adp->ad_offset], 11499 (intmax_t)adp->ad_oldblkno); 11500 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11501 } else { 11502 if (dp2->di_ib[adp->ad_offset - UFS_NDADDR] != 11503 0) 11504 panic("%s: %s #%jd allocated as %jd", 11505 "handle_written_inodeblock", 11506 "indirect pointer", 11507 (intmax_t)adp->ad_offset - 11508 UFS_NDADDR, 11509 (intmax_t) 11510 dp2->di_ib[adp->ad_offset - 11511 UFS_NDADDR]); 11512 dp2->di_ib[adp->ad_offset - UFS_NDADDR] = 11513 adp->ad_newblkno; 11514 } 11515 } 11516 adp->ad_state &= ~UNDONE; 11517 adp->ad_state |= ATTACHED; 11518 hadchanges = 1; 11519 } 11520 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11521 nextadp = TAILQ_NEXT(adp, ad_next); 11522 if (adp->ad_state & ATTACHED) 11523 panic("handle_written_inodeblock: new entry"); 11524 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11525 panic("%s: direct pointers #%jd %s %jd != %jd", 11526 "handle_written_inodeblock", 11527 (intmax_t)adp->ad_offset, "mismatch", 11528 (intmax_t)dp2->di_extb[adp->ad_offset], 11529 (intmax_t)adp->ad_oldblkno); 11530 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11531 adp->ad_state &= ~UNDONE; 11532 adp->ad_state |= ATTACHED; 11533 hadchanges = 1; 11534 } 11535 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11536 stat_direct_blk_ptrs++; 11537 /* 11538 * Reset the file size to its most up-to-date value. 11539 */ 11540 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11541 panic("handle_written_inodeblock: bad size"); 11542 if (inodedep->id_savednlink > UFS_LINK_MAX) 11543 panic("handle_written_inodeblock: Invalid link count " 11544 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink, 11545 inodedep); 11546 if (fstype == UFS1) { 11547 if (dp1->di_nlink != inodedep->id_savednlink) { 11548 dp1->di_nlink = inodedep->id_savednlink; 11549 hadchanges = 1; 11550 } 11551 if (dp1->di_size != inodedep->id_savedsize) { 11552 dp1->di_size = inodedep->id_savedsize; 11553 hadchanges = 1; 11554 } 11555 } else { 11556 if (dp2->di_nlink != inodedep->id_savednlink) { 11557 dp2->di_nlink = inodedep->id_savednlink; 11558 hadchanges = 1; 11559 } 11560 if (dp2->di_size != inodedep->id_savedsize) { 11561 dp2->di_size = inodedep->id_savedsize; 11562 hadchanges = 1; 11563 } 11564 if (dp2->di_extsize != inodedep->id_savedextsize) { 11565 dp2->di_extsize = inodedep->id_savedextsize; 11566 hadchanges = 1; 11567 } 11568 } 11569 inodedep->id_savedsize = -1; 11570 inodedep->id_savedextsize = -1; 11571 inodedep->id_savednlink = -1; 11572 /* 11573 * If there were any rollbacks in the inode block, then it must be 11574 * marked dirty so that its will eventually get written back in 11575 * its correct form. 11576 */ 11577 if (hadchanges) 11578 bdirty(bp); 11579 bufwait: 11580 /* 11581 * If the write did not succeed, we have done all the roll-forward 11582 * operations, but we cannot take the actions that will allow its 11583 * dependencies to be processed. 11584 */ 11585 if ((flags & WRITESUCCEEDED) == 0) 11586 return (hadchanges); 11587 /* 11588 * Process any allocdirects that completed during the update. 11589 */ 11590 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11591 handle_allocdirect_partdone(adp, &wkhd); 11592 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11593 handle_allocdirect_partdone(adp, &wkhd); 11594 /* 11595 * Process deallocations that were held pending until the 11596 * inode had been written to disk. Freeing of the inode 11597 * is delayed until after all blocks have been freed to 11598 * avoid creation of new <vfsid, inum, lbn> triples 11599 * before the old ones have been deleted. Completely 11600 * unlinked inodes are not processed until the unlinked 11601 * inode list is written or the last reference is removed. 11602 */ 11603 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11604 freefile = handle_bufwait(inodedep, NULL); 11605 if (freefile && !LIST_EMPTY(&wkhd)) { 11606 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11607 freefile = NULL; 11608 } 11609 } 11610 /* 11611 * Move rolled forward dependency completions to the bufwait list 11612 * now that those that were already written have been processed. 11613 */ 11614 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11615 panic("handle_written_inodeblock: bufwait but no changes"); 11616 jwork_move(&inodedep->id_bufwait, &wkhd); 11617 11618 if (freefile != NULL) { 11619 /* 11620 * If the inode is goingaway it was never written. Fake up 11621 * the state here so free_inodedep() can succeed. 11622 */ 11623 if (inodedep->id_state & GOINGAWAY) 11624 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11625 if (free_inodedep(inodedep) == 0) 11626 panic("handle_written_inodeblock: live inodedep %p", 11627 inodedep); 11628 add_to_worklist(&freefile->fx_list, 0); 11629 return (0); 11630 } 11631 11632 /* 11633 * If no outstanding dependencies, free it. 11634 */ 11635 if (free_inodedep(inodedep) || 11636 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11637 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11638 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11639 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11640 return (0); 11641 return (hadchanges); 11642 } 11643 11644 /* 11645 * Perform needed roll-forwards and kick off any dependencies that 11646 * can now be processed. 11647 * 11648 * If the write did not succeed, we will do all the roll-forward 11649 * operations, but we will not take the actions that will allow its 11650 * dependencies to be processed. 11651 */ 11652 static int 11653 handle_written_indirdep(indirdep, bp, bpp, flags) 11654 struct indirdep *indirdep; 11655 struct buf *bp; 11656 struct buf **bpp; 11657 int flags; 11658 { 11659 struct allocindir *aip; 11660 struct buf *sbp; 11661 int chgs; 11662 11663 if (indirdep->ir_state & GOINGAWAY) 11664 panic("handle_written_indirdep: indirdep gone"); 11665 if ((indirdep->ir_state & IOSTARTED) == 0) 11666 panic("handle_written_indirdep: IO not started"); 11667 chgs = 0; 11668 /* 11669 * If there were rollbacks revert them here. 11670 */ 11671 if (indirdep->ir_saveddata) { 11672 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11673 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11674 free(indirdep->ir_saveddata, M_INDIRDEP); 11675 indirdep->ir_saveddata = NULL; 11676 } 11677 chgs = 1; 11678 } 11679 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11680 indirdep->ir_state |= ATTACHED; 11681 /* 11682 * If the write did not succeed, we have done all the roll-forward 11683 * operations, but we cannot take the actions that will allow its 11684 * dependencies to be processed. 11685 */ 11686 if ((flags & WRITESUCCEEDED) == 0) { 11687 stat_indir_blk_ptrs++; 11688 bdirty(bp); 11689 return (1); 11690 } 11691 /* 11692 * Move allocindirs with written pointers to the completehd if 11693 * the indirdep's pointer is not yet written. Otherwise 11694 * free them here. 11695 */ 11696 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) { 11697 LIST_REMOVE(aip, ai_next); 11698 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11699 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11700 ai_next); 11701 newblk_freefrag(&aip->ai_block); 11702 continue; 11703 } 11704 free_newblk(&aip->ai_block); 11705 } 11706 /* 11707 * Move allocindirs that have finished dependency processing from 11708 * the done list to the write list after updating the pointers. 11709 */ 11710 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11711 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 11712 handle_allocindir_partdone(aip); 11713 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11714 panic("disk_write_complete: not gone"); 11715 chgs = 1; 11716 } 11717 } 11718 /* 11719 * Preserve the indirdep if there were any changes or if it is not 11720 * yet valid on disk. 11721 */ 11722 if (chgs) { 11723 stat_indir_blk_ptrs++; 11724 bdirty(bp); 11725 return (1); 11726 } 11727 /* 11728 * If there were no changes we can discard the savedbp and detach 11729 * ourselves from the buf. We are only carrying completed pointers 11730 * in this case. 11731 */ 11732 sbp = indirdep->ir_savebp; 11733 sbp->b_flags |= B_INVAL | B_NOCACHE; 11734 indirdep->ir_savebp = NULL; 11735 indirdep->ir_bp = NULL; 11736 if (*bpp != NULL) 11737 panic("handle_written_indirdep: bp already exists."); 11738 *bpp = sbp; 11739 /* 11740 * The indirdep may not be freed until its parent points at it. 11741 */ 11742 if (indirdep->ir_state & DEPCOMPLETE) 11743 free_indirdep(indirdep); 11744 11745 return (0); 11746 } 11747 11748 /* 11749 * Process a diradd entry after its dependent inode has been written. 11750 * This routine must be called with splbio interrupts blocked. 11751 */ 11752 static void 11753 diradd_inode_written(dap, inodedep) 11754 struct diradd *dap; 11755 struct inodedep *inodedep; 11756 { 11757 11758 dap->da_state |= COMPLETE; 11759 complete_diradd(dap); 11760 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11761 } 11762 11763 /* 11764 * Returns true if the bmsafemap will have rollbacks when written. Must only 11765 * be called with the per-filesystem lock and the buf lock on the cg held. 11766 */ 11767 static int 11768 bmsafemap_backgroundwrite(bmsafemap, bp) 11769 struct bmsafemap *bmsafemap; 11770 struct buf *bp; 11771 { 11772 int dirty; 11773 11774 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11775 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11776 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11777 /* 11778 * If we're initiating a background write we need to process the 11779 * rollbacks as they exist now, not as they exist when IO starts. 11780 * No other consumers will look at the contents of the shadowed 11781 * buf so this is safe to do here. 11782 */ 11783 if (bp->b_xflags & BX_BKGRDMARKER) 11784 initiate_write_bmsafemap(bmsafemap, bp); 11785 11786 return (dirty); 11787 } 11788 11789 /* 11790 * Re-apply an allocation when a cg write is complete. 11791 */ 11792 static int 11793 jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11794 struct jnewblk *jnewblk; 11795 struct fs *fs; 11796 struct cg *cgp; 11797 uint8_t *blksfree; 11798 { 11799 ufs1_daddr_t fragno; 11800 ufs2_daddr_t blkno; 11801 long cgbno, bbase; 11802 int frags, blk; 11803 int i; 11804 11805 frags = 0; 11806 cgbno = dtogd(fs, jnewblk->jn_blkno); 11807 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11808 if (isclr(blksfree, cgbno + i)) 11809 panic("jnewblk_rollforward: re-allocated fragment"); 11810 frags++; 11811 } 11812 if (frags == fs->fs_frag) { 11813 blkno = fragstoblks(fs, cgbno); 11814 ffs_clrblock(fs, blksfree, (long)blkno); 11815 ffs_clusteracct(fs, cgp, blkno, -1); 11816 cgp->cg_cs.cs_nbfree--; 11817 } else { 11818 bbase = cgbno - fragnum(fs, cgbno); 11819 cgbno += jnewblk->jn_oldfrags; 11820 /* If a complete block had been reassembled, account for it. */ 11821 fragno = fragstoblks(fs, bbase); 11822 if (ffs_isblock(fs, blksfree, fragno)) { 11823 cgp->cg_cs.cs_nffree += fs->fs_frag; 11824 ffs_clusteracct(fs, cgp, fragno, -1); 11825 cgp->cg_cs.cs_nbfree--; 11826 } 11827 /* Decrement the old frags. */ 11828 blk = blkmap(fs, blksfree, bbase); 11829 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11830 /* Allocate the fragment */ 11831 for (i = 0; i < frags; i++) 11832 clrbit(blksfree, cgbno + i); 11833 cgp->cg_cs.cs_nffree -= frags; 11834 /* Add back in counts associated with the new frags */ 11835 blk = blkmap(fs, blksfree, bbase); 11836 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11837 } 11838 return (frags); 11839 } 11840 11841 /* 11842 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11843 * changes if it's not a background write. Set all written dependencies 11844 * to DEPCOMPLETE and free the structure if possible. 11845 * 11846 * If the write did not succeed, we will do all the roll-forward 11847 * operations, but we will not take the actions that will allow its 11848 * dependencies to be processed. 11849 */ 11850 static int 11851 handle_written_bmsafemap(bmsafemap, bp, flags) 11852 struct bmsafemap *bmsafemap; 11853 struct buf *bp; 11854 int flags; 11855 { 11856 struct newblk *newblk; 11857 struct inodedep *inodedep; 11858 struct jaddref *jaddref, *jatmp; 11859 struct jnewblk *jnewblk, *jntmp; 11860 struct ufsmount *ump; 11861 uint8_t *inosused; 11862 uint8_t *blksfree; 11863 struct cg *cgp; 11864 struct fs *fs; 11865 ino_t ino; 11866 int foreground; 11867 int chgs; 11868 11869 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11870 panic("handle_written_bmsafemap: Not started\n"); 11871 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11872 chgs = 0; 11873 bmsafemap->sm_state &= ~IOSTARTED; 11874 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11875 /* 11876 * If write was successful, release journal work that was waiting 11877 * on the write. Otherwise move the work back. 11878 */ 11879 if (flags & WRITESUCCEEDED) 11880 handle_jwork(&bmsafemap->sm_freewr); 11881 else 11882 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11883 worklist, wk_list); 11884 11885 /* 11886 * Restore unwritten inode allocation pending jaddref writes. 11887 */ 11888 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11889 cgp = (struct cg *)bp->b_data; 11890 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11891 inosused = cg_inosused(cgp); 11892 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11893 ja_bmdeps, jatmp) { 11894 if ((jaddref->ja_state & UNDONE) == 0) 11895 continue; 11896 ino = jaddref->ja_ino % fs->fs_ipg; 11897 if (isset(inosused, ino)) 11898 panic("handle_written_bmsafemap: " 11899 "re-allocated inode"); 11900 /* Do the roll-forward only if it's a real copy. */ 11901 if (foreground) { 11902 if ((jaddref->ja_mode & IFMT) == IFDIR) 11903 cgp->cg_cs.cs_ndir++; 11904 cgp->cg_cs.cs_nifree--; 11905 setbit(inosused, ino); 11906 chgs = 1; 11907 } 11908 jaddref->ja_state &= ~UNDONE; 11909 jaddref->ja_state |= ATTACHED; 11910 free_jaddref(jaddref); 11911 } 11912 } 11913 /* 11914 * Restore any block allocations which are pending journal writes. 11915 */ 11916 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11917 cgp = (struct cg *)bp->b_data; 11918 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11919 blksfree = cg_blksfree(cgp); 11920 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11921 jntmp) { 11922 if ((jnewblk->jn_state & UNDONE) == 0) 11923 continue; 11924 /* Do the roll-forward only if it's a real copy. */ 11925 if (foreground && 11926 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11927 chgs = 1; 11928 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11929 jnewblk->jn_state |= ATTACHED; 11930 free_jnewblk(jnewblk); 11931 } 11932 } 11933 /* 11934 * If the write did not succeed, we have done all the roll-forward 11935 * operations, but we cannot take the actions that will allow its 11936 * dependencies to be processed. 11937 */ 11938 if ((flags & WRITESUCCEEDED) == 0) { 11939 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11940 newblk, nb_deps); 11941 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11942 worklist, wk_list); 11943 if (foreground) 11944 bdirty(bp); 11945 return (1); 11946 } 11947 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11948 newblk->nb_state |= DEPCOMPLETE; 11949 newblk->nb_state &= ~ONDEPLIST; 11950 newblk->nb_bmsafemap = NULL; 11951 LIST_REMOVE(newblk, nb_deps); 11952 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11953 handle_allocdirect_partdone( 11954 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11955 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11956 handle_allocindir_partdone( 11957 WK_ALLOCINDIR(&newblk->nb_list)); 11958 else if (newblk->nb_list.wk_type != D_NEWBLK) 11959 panic("handle_written_bmsafemap: Unexpected type: %s", 11960 TYPENAME(newblk->nb_list.wk_type)); 11961 } 11962 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11963 inodedep->id_state |= DEPCOMPLETE; 11964 inodedep->id_state &= ~ONDEPLIST; 11965 LIST_REMOVE(inodedep, id_deps); 11966 inodedep->id_bmsafemap = NULL; 11967 } 11968 LIST_REMOVE(bmsafemap, sm_next); 11969 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11970 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11971 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11972 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11973 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11974 LIST_REMOVE(bmsafemap, sm_hash); 11975 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11976 return (0); 11977 } 11978 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11979 if (foreground) 11980 bdirty(bp); 11981 return (1); 11982 } 11983 11984 /* 11985 * Try to free a mkdir dependency. 11986 */ 11987 static void 11988 complete_mkdir(mkdir) 11989 struct mkdir *mkdir; 11990 { 11991 struct diradd *dap; 11992 11993 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11994 return; 11995 LIST_REMOVE(mkdir, md_mkdirs); 11996 dap = mkdir->md_diradd; 11997 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11998 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11999 dap->da_state |= DEPCOMPLETE; 12000 complete_diradd(dap); 12001 } 12002 WORKITEM_FREE(mkdir, D_MKDIR); 12003 } 12004 12005 /* 12006 * Handle the completion of a mkdir dependency. 12007 */ 12008 static void 12009 handle_written_mkdir(mkdir, type) 12010 struct mkdir *mkdir; 12011 int type; 12012 { 12013 12014 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 12015 panic("handle_written_mkdir: bad type"); 12016 mkdir->md_state |= COMPLETE; 12017 complete_mkdir(mkdir); 12018 } 12019 12020 static int 12021 free_pagedep(pagedep) 12022 struct pagedep *pagedep; 12023 { 12024 int i; 12025 12026 if (pagedep->pd_state & NEWBLOCK) 12027 return (0); 12028 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 12029 return (0); 12030 for (i = 0; i < DAHASHSZ; i++) 12031 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 12032 return (0); 12033 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 12034 return (0); 12035 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 12036 return (0); 12037 if (pagedep->pd_state & ONWORKLIST) 12038 WORKLIST_REMOVE(&pagedep->pd_list); 12039 LIST_REMOVE(pagedep, pd_hash); 12040 WORKITEM_FREE(pagedep, D_PAGEDEP); 12041 12042 return (1); 12043 } 12044 12045 /* 12046 * Called from within softdep_disk_write_complete above. 12047 * A write operation was just completed. Removed inodes can 12048 * now be freed and associated block pointers may be committed. 12049 * Note that this routine is always called from interrupt level 12050 * with further interrupts from this device blocked. 12051 * 12052 * If the write did not succeed, we will do all the roll-forward 12053 * operations, but we will not take the actions that will allow its 12054 * dependencies to be processed. 12055 */ 12056 static int 12057 handle_written_filepage(pagedep, bp, flags) 12058 struct pagedep *pagedep; 12059 struct buf *bp; /* buffer containing the written page */ 12060 int flags; 12061 { 12062 struct dirrem *dirrem; 12063 struct diradd *dap, *nextdap; 12064 struct direct *ep; 12065 int i, chgs; 12066 12067 if ((pagedep->pd_state & IOSTARTED) == 0) 12068 panic("handle_written_filepage: not started"); 12069 pagedep->pd_state &= ~IOSTARTED; 12070 if ((flags & WRITESUCCEEDED) == 0) 12071 goto rollforward; 12072 /* 12073 * Process any directory removals that have been committed. 12074 */ 12075 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 12076 LIST_REMOVE(dirrem, dm_next); 12077 dirrem->dm_state |= COMPLETE; 12078 dirrem->dm_dirinum = pagedep->pd_ino; 12079 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 12080 ("handle_written_filepage: Journal entries not written.")); 12081 add_to_worklist(&dirrem->dm_list, 0); 12082 } 12083 /* 12084 * Free any directory additions that have been committed. 12085 * If it is a newly allocated block, we have to wait until 12086 * the on-disk directory inode claims the new block. 12087 */ 12088 if ((pagedep->pd_state & NEWBLOCK) == 0) 12089 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 12090 free_diradd(dap, NULL); 12091 rollforward: 12092 /* 12093 * Uncommitted directory entries must be restored. 12094 */ 12095 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 12096 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 12097 dap = nextdap) { 12098 nextdap = LIST_NEXT(dap, da_pdlist); 12099 if (dap->da_state & ATTACHED) 12100 panic("handle_written_filepage: attached"); 12101 ep = (struct direct *) 12102 ((char *)bp->b_data + dap->da_offset); 12103 ep->d_ino = dap->da_newinum; 12104 dap->da_state &= ~UNDONE; 12105 dap->da_state |= ATTACHED; 12106 chgs = 1; 12107 /* 12108 * If the inode referenced by the directory has 12109 * been written out, then the dependency can be 12110 * moved to the pending list. 12111 */ 12112 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 12113 LIST_REMOVE(dap, da_pdlist); 12114 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 12115 da_pdlist); 12116 } 12117 } 12118 } 12119 /* 12120 * If there were any rollbacks in the directory, then it must be 12121 * marked dirty so that its will eventually get written back in 12122 * its correct form. 12123 */ 12124 if (chgs || (flags & WRITESUCCEEDED) == 0) { 12125 if ((bp->b_flags & B_DELWRI) == 0) 12126 stat_dir_entry++; 12127 bdirty(bp); 12128 return (1); 12129 } 12130 /* 12131 * If we are not waiting for a new directory block to be 12132 * claimed by its inode, then the pagedep will be freed. 12133 * Otherwise it will remain to track any new entries on 12134 * the page in case they are fsync'ed. 12135 */ 12136 free_pagedep(pagedep); 12137 return (0); 12138 } 12139 12140 /* 12141 * Writing back in-core inode structures. 12142 * 12143 * The filesystem only accesses an inode's contents when it occupies an 12144 * "in-core" inode structure. These "in-core" structures are separate from 12145 * the page frames used to cache inode blocks. Only the latter are 12146 * transferred to/from the disk. So, when the updated contents of the 12147 * "in-core" inode structure are copied to the corresponding in-memory inode 12148 * block, the dependencies are also transferred. The following procedure is 12149 * called when copying a dirty "in-core" inode to a cached inode block. 12150 */ 12151 12152 /* 12153 * Called when an inode is loaded from disk. If the effective link count 12154 * differed from the actual link count when it was last flushed, then we 12155 * need to ensure that the correct effective link count is put back. 12156 */ 12157 void 12158 softdep_load_inodeblock(ip) 12159 struct inode *ip; /* the "in_core" copy of the inode */ 12160 { 12161 struct inodedep *inodedep; 12162 struct ufsmount *ump; 12163 12164 ump = ITOUMP(ip); 12165 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 12166 ("softdep_load_inodeblock called on non-softdep filesystem")); 12167 /* 12168 * Check for alternate nlink count. 12169 */ 12170 ip->i_effnlink = ip->i_nlink; 12171 ACQUIRE_LOCK(ump); 12172 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) { 12173 FREE_LOCK(ump); 12174 return; 12175 } 12176 ip->i_effnlink -= inodedep->id_nlinkdelta; 12177 FREE_LOCK(ump); 12178 } 12179 12180 /* 12181 * This routine is called just before the "in-core" inode 12182 * information is to be copied to the in-memory inode block. 12183 * Recall that an inode block contains several inodes. If 12184 * the force flag is set, then the dependencies will be 12185 * cleared so that the update can always be made. Note that 12186 * the buffer is locked when this routine is called, so we 12187 * will never be in the middle of writing the inode block 12188 * to disk. 12189 */ 12190 void 12191 softdep_update_inodeblock(ip, bp, waitfor) 12192 struct inode *ip; /* the "in_core" copy of the inode */ 12193 struct buf *bp; /* the buffer containing the inode block */ 12194 int waitfor; /* nonzero => update must be allowed */ 12195 { 12196 struct inodedep *inodedep; 12197 struct inoref *inoref; 12198 struct ufsmount *ump; 12199 struct worklist *wk; 12200 struct mount *mp; 12201 struct buf *ibp; 12202 struct fs *fs; 12203 int error; 12204 12205 ump = ITOUMP(ip); 12206 mp = UFSTOVFS(ump); 12207 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12208 ("softdep_update_inodeblock called on non-softdep filesystem")); 12209 fs = ump->um_fs; 12210 /* 12211 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12212 * does not have access to the in-core ip so must write directly into 12213 * the inode block buffer when setting freelink. 12214 */ 12215 if (fs->fs_magic == FS_UFS1_MAGIC) 12216 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12217 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12218 else 12219 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12220 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12221 /* 12222 * If the effective link count is not equal to the actual link 12223 * count, then we must track the difference in an inodedep while 12224 * the inode is (potentially) tossed out of the cache. Otherwise, 12225 * if there is no existing inodedep, then there are no dependencies 12226 * to track. 12227 */ 12228 ACQUIRE_LOCK(ump); 12229 again: 12230 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12231 FREE_LOCK(ump); 12232 if (ip->i_effnlink != ip->i_nlink) 12233 panic("softdep_update_inodeblock: bad link count"); 12234 return; 12235 } 12236 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12237 panic("softdep_update_inodeblock: bad delta"); 12238 /* 12239 * If we're flushing all dependencies we must also move any waiting 12240 * for journal writes onto the bufwait list prior to I/O. 12241 */ 12242 if (waitfor) { 12243 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12244 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12245 == DEPCOMPLETE) { 12246 jwait(&inoref->if_list, MNT_WAIT); 12247 goto again; 12248 } 12249 } 12250 } 12251 /* 12252 * Changes have been initiated. Anything depending on these 12253 * changes cannot occur until this inode has been written. 12254 */ 12255 inodedep->id_state &= ~COMPLETE; 12256 if ((inodedep->id_state & ONWORKLIST) == 0) 12257 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12258 /* 12259 * Any new dependencies associated with the incore inode must 12260 * now be moved to the list associated with the buffer holding 12261 * the in-memory copy of the inode. Once merged process any 12262 * allocdirects that are completed by the merger. 12263 */ 12264 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12265 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12266 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12267 NULL); 12268 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12269 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12270 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12271 NULL); 12272 /* 12273 * Now that the inode has been pushed into the buffer, the 12274 * operations dependent on the inode being written to disk 12275 * can be moved to the id_bufwait so that they will be 12276 * processed when the buffer I/O completes. 12277 */ 12278 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12279 WORKLIST_REMOVE(wk); 12280 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12281 } 12282 /* 12283 * Newly allocated inodes cannot be written until the bitmap 12284 * that allocates them have been written (indicated by 12285 * DEPCOMPLETE being set in id_state). If we are doing a 12286 * forced sync (e.g., an fsync on a file), we force the bitmap 12287 * to be written so that the update can be done. 12288 */ 12289 if (waitfor == 0) { 12290 FREE_LOCK(ump); 12291 return; 12292 } 12293 retry: 12294 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12295 FREE_LOCK(ump); 12296 return; 12297 } 12298 ibp = inodedep->id_bmsafemap->sm_buf; 12299 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12300 if (ibp == NULL) { 12301 /* 12302 * If ibp came back as NULL, the dependency could have been 12303 * freed while we slept. Look it up again, and check to see 12304 * that it has completed. 12305 */ 12306 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12307 goto retry; 12308 FREE_LOCK(ump); 12309 return; 12310 } 12311 FREE_LOCK(ump); 12312 if ((error = bwrite(ibp)) != 0) 12313 softdep_error("softdep_update_inodeblock: bwrite", error); 12314 } 12315 12316 /* 12317 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12318 * old inode dependency list (such as id_inoupdt). This routine must be 12319 * called with splbio interrupts blocked. 12320 */ 12321 static void 12322 merge_inode_lists(newlisthead, oldlisthead) 12323 struct allocdirectlst *newlisthead; 12324 struct allocdirectlst *oldlisthead; 12325 { 12326 struct allocdirect *listadp, *newadp; 12327 12328 newadp = TAILQ_FIRST(newlisthead); 12329 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12330 if (listadp->ad_offset < newadp->ad_offset) { 12331 listadp = TAILQ_NEXT(listadp, ad_next); 12332 continue; 12333 } 12334 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12335 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12336 if (listadp->ad_offset == newadp->ad_offset) { 12337 allocdirect_merge(oldlisthead, newadp, 12338 listadp); 12339 listadp = newadp; 12340 } 12341 newadp = TAILQ_FIRST(newlisthead); 12342 } 12343 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12344 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12345 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12346 } 12347 } 12348 12349 /* 12350 * If we are doing an fsync, then we must ensure that any directory 12351 * entries for the inode have been written after the inode gets to disk. 12352 */ 12353 int 12354 softdep_fsync(vp) 12355 struct vnode *vp; /* the "in_core" copy of the inode */ 12356 { 12357 struct inodedep *inodedep; 12358 struct pagedep *pagedep; 12359 struct inoref *inoref; 12360 struct ufsmount *ump; 12361 struct worklist *wk; 12362 struct diradd *dap; 12363 struct mount *mp; 12364 struct vnode *pvp; 12365 struct inode *ip; 12366 struct buf *bp; 12367 struct fs *fs; 12368 struct thread *td = curthread; 12369 int error, flushparent, pagedep_new_block; 12370 ino_t parentino; 12371 ufs_lbn_t lbn; 12372 12373 ip = VTOI(vp); 12374 mp = vp->v_mount; 12375 ump = VFSTOUFS(mp); 12376 fs = ump->um_fs; 12377 if (MOUNTEDSOFTDEP(mp) == 0) 12378 return (0); 12379 ACQUIRE_LOCK(ump); 12380 restart: 12381 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12382 FREE_LOCK(ump); 12383 return (0); 12384 } 12385 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12386 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12387 == DEPCOMPLETE) { 12388 jwait(&inoref->if_list, MNT_WAIT); 12389 goto restart; 12390 } 12391 } 12392 if (!LIST_EMPTY(&inodedep->id_inowait) || 12393 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12394 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12395 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12396 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12397 panic("softdep_fsync: pending ops %p", inodedep); 12398 for (error = 0, flushparent = 0; ; ) { 12399 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12400 break; 12401 if (wk->wk_type != D_DIRADD) 12402 panic("softdep_fsync: Unexpected type %s", 12403 TYPENAME(wk->wk_type)); 12404 dap = WK_DIRADD(wk); 12405 /* 12406 * Flush our parent if this directory entry has a MKDIR_PARENT 12407 * dependency or is contained in a newly allocated block. 12408 */ 12409 if (dap->da_state & DIRCHG) 12410 pagedep = dap->da_previous->dm_pagedep; 12411 else 12412 pagedep = dap->da_pagedep; 12413 parentino = pagedep->pd_ino; 12414 lbn = pagedep->pd_lbn; 12415 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12416 panic("softdep_fsync: dirty"); 12417 if ((dap->da_state & MKDIR_PARENT) || 12418 (pagedep->pd_state & NEWBLOCK)) 12419 flushparent = 1; 12420 else 12421 flushparent = 0; 12422 /* 12423 * If we are being fsync'ed as part of vgone'ing this vnode, 12424 * then we will not be able to release and recover the 12425 * vnode below, so we just have to give up on writing its 12426 * directory entry out. It will eventually be written, just 12427 * not now, but then the user was not asking to have it 12428 * written, so we are not breaking any promises. 12429 */ 12430 if (vp->v_iflag & VI_DOOMED) 12431 break; 12432 /* 12433 * We prevent deadlock by always fetching inodes from the 12434 * root, moving down the directory tree. Thus, when fetching 12435 * our parent directory, we first try to get the lock. If 12436 * that fails, we must unlock ourselves before requesting 12437 * the lock on our parent. See the comment in ufs_lookup 12438 * for details on possible races. 12439 */ 12440 FREE_LOCK(ump); 12441 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12442 FFSV_FORCEINSMQ)) { 12443 error = vfs_busy(mp, MBF_NOWAIT); 12444 if (error != 0) { 12445 vfs_ref(mp); 12446 VOP_UNLOCK(vp, 0); 12447 error = vfs_busy(mp, 0); 12448 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12449 vfs_rel(mp); 12450 if (error != 0) 12451 return (ENOENT); 12452 if (vp->v_iflag & VI_DOOMED) { 12453 vfs_unbusy(mp); 12454 return (ENOENT); 12455 } 12456 } 12457 VOP_UNLOCK(vp, 0); 12458 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12459 &pvp, FFSV_FORCEINSMQ); 12460 vfs_unbusy(mp); 12461 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12462 if (vp->v_iflag & VI_DOOMED) { 12463 if (error == 0) 12464 vput(pvp); 12465 error = ENOENT; 12466 } 12467 if (error != 0) 12468 return (error); 12469 } 12470 /* 12471 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12472 * that are contained in direct blocks will be resolved by 12473 * doing a ffs_update. Pagedeps contained in indirect blocks 12474 * may require a complete sync'ing of the directory. So, we 12475 * try the cheap and fast ffs_update first, and if that fails, 12476 * then we do the slower ffs_syncvnode of the directory. 12477 */ 12478 if (flushparent) { 12479 int locked; 12480 12481 if ((error = ffs_update(pvp, 1)) != 0) { 12482 vput(pvp); 12483 return (error); 12484 } 12485 ACQUIRE_LOCK(ump); 12486 locked = 1; 12487 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12488 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12489 if (wk->wk_type != D_DIRADD) 12490 panic("softdep_fsync: Unexpected type %s", 12491 TYPENAME(wk->wk_type)); 12492 dap = WK_DIRADD(wk); 12493 if (dap->da_state & DIRCHG) 12494 pagedep = dap->da_previous->dm_pagedep; 12495 else 12496 pagedep = dap->da_pagedep; 12497 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12498 FREE_LOCK(ump); 12499 locked = 0; 12500 if (pagedep_new_block && (error = 12501 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12502 vput(pvp); 12503 return (error); 12504 } 12505 } 12506 } 12507 if (locked) 12508 FREE_LOCK(ump); 12509 } 12510 /* 12511 * Flush directory page containing the inode's name. 12512 */ 12513 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12514 &bp); 12515 if (error == 0) 12516 error = bwrite(bp); 12517 else 12518 brelse(bp); 12519 vput(pvp); 12520 if (error != 0) 12521 return (error); 12522 ACQUIRE_LOCK(ump); 12523 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12524 break; 12525 } 12526 FREE_LOCK(ump); 12527 return (0); 12528 } 12529 12530 /* 12531 * Flush all the dirty bitmaps associated with the block device 12532 * before flushing the rest of the dirty blocks so as to reduce 12533 * the number of dependencies that will have to be rolled back. 12534 * 12535 * XXX Unused? 12536 */ 12537 void 12538 softdep_fsync_mountdev(vp) 12539 struct vnode *vp; 12540 { 12541 struct buf *bp, *nbp; 12542 struct worklist *wk; 12543 struct bufobj *bo; 12544 12545 if (!vn_isdisk(vp, NULL)) 12546 panic("softdep_fsync_mountdev: vnode not a disk"); 12547 bo = &vp->v_bufobj; 12548 restart: 12549 BO_LOCK(bo); 12550 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12551 /* 12552 * If it is already scheduled, skip to the next buffer. 12553 */ 12554 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12555 continue; 12556 12557 if ((bp->b_flags & B_DELWRI) == 0) 12558 panic("softdep_fsync_mountdev: not dirty"); 12559 /* 12560 * We are only interested in bitmaps with outstanding 12561 * dependencies. 12562 */ 12563 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12564 wk->wk_type != D_BMSAFEMAP || 12565 (bp->b_vflags & BV_BKGRDINPROG)) { 12566 BUF_UNLOCK(bp); 12567 continue; 12568 } 12569 BO_UNLOCK(bo); 12570 bremfree(bp); 12571 (void) bawrite(bp); 12572 goto restart; 12573 } 12574 drain_output(vp); 12575 BO_UNLOCK(bo); 12576 } 12577 12578 /* 12579 * Sync all cylinder groups that were dirty at the time this function is 12580 * called. Newly dirtied cgs will be inserted before the sentinel. This 12581 * is used to flush freedep activity that may be holding up writes to a 12582 * indirect block. 12583 */ 12584 static int 12585 sync_cgs(mp, waitfor) 12586 struct mount *mp; 12587 int waitfor; 12588 { 12589 struct bmsafemap *bmsafemap; 12590 struct bmsafemap *sentinel; 12591 struct ufsmount *ump; 12592 struct buf *bp; 12593 int error; 12594 12595 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12596 sentinel->sm_cg = -1; 12597 ump = VFSTOUFS(mp); 12598 error = 0; 12599 ACQUIRE_LOCK(ump); 12600 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12601 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12602 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12603 /* Skip sentinels and cgs with no work to release. */ 12604 if (bmsafemap->sm_cg == -1 || 12605 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12606 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12607 LIST_REMOVE(sentinel, sm_next); 12608 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12609 continue; 12610 } 12611 /* 12612 * If we don't get the lock and we're waiting try again, if 12613 * not move on to the next buf and try to sync it. 12614 */ 12615 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12616 if (bp == NULL && waitfor == MNT_WAIT) 12617 continue; 12618 LIST_REMOVE(sentinel, sm_next); 12619 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12620 if (bp == NULL) 12621 continue; 12622 FREE_LOCK(ump); 12623 if (waitfor == MNT_NOWAIT) 12624 bawrite(bp); 12625 else 12626 error = bwrite(bp); 12627 ACQUIRE_LOCK(ump); 12628 if (error) 12629 break; 12630 } 12631 LIST_REMOVE(sentinel, sm_next); 12632 FREE_LOCK(ump); 12633 free(sentinel, M_BMSAFEMAP); 12634 return (error); 12635 } 12636 12637 /* 12638 * This routine is called when we are trying to synchronously flush a 12639 * file. This routine must eliminate any filesystem metadata dependencies 12640 * so that the syncing routine can succeed. 12641 */ 12642 int 12643 softdep_sync_metadata(struct vnode *vp) 12644 { 12645 struct inode *ip; 12646 int error; 12647 12648 ip = VTOI(vp); 12649 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12650 ("softdep_sync_metadata called on non-softdep filesystem")); 12651 /* 12652 * Ensure that any direct block dependencies have been cleared, 12653 * truncations are started, and inode references are journaled. 12654 */ 12655 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount)); 12656 /* 12657 * Write all journal records to prevent rollbacks on devvp. 12658 */ 12659 if (vp->v_type == VCHR) 12660 softdep_flushjournal(vp->v_mount); 12661 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12662 /* 12663 * Ensure that all truncates are written so we won't find deps on 12664 * indirect blocks. 12665 */ 12666 process_truncates(vp); 12667 FREE_LOCK(VFSTOUFS(vp->v_mount)); 12668 12669 return (error); 12670 } 12671 12672 /* 12673 * This routine is called when we are attempting to sync a buf with 12674 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12675 * other IO it can but returns EBUSY if the buffer is not yet able to 12676 * be written. Dependencies which will not cause rollbacks will always 12677 * return 0. 12678 */ 12679 int 12680 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12681 { 12682 struct indirdep *indirdep; 12683 struct pagedep *pagedep; 12684 struct allocindir *aip; 12685 struct newblk *newblk; 12686 struct ufsmount *ump; 12687 struct buf *nbp; 12688 struct worklist *wk; 12689 int i, error; 12690 12691 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12692 ("softdep_sync_buf called on non-softdep filesystem")); 12693 /* 12694 * For VCHR we just don't want to force flush any dependencies that 12695 * will cause rollbacks. 12696 */ 12697 if (vp->v_type == VCHR) { 12698 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12699 return (EBUSY); 12700 return (0); 12701 } 12702 ump = VFSTOUFS(vp->v_mount); 12703 ACQUIRE_LOCK(ump); 12704 /* 12705 * As we hold the buffer locked, none of its dependencies 12706 * will disappear. 12707 */ 12708 error = 0; 12709 top: 12710 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12711 switch (wk->wk_type) { 12712 12713 case D_ALLOCDIRECT: 12714 case D_ALLOCINDIR: 12715 newblk = WK_NEWBLK(wk); 12716 if (newblk->nb_jnewblk != NULL) { 12717 if (waitfor == MNT_NOWAIT) { 12718 error = EBUSY; 12719 goto out_unlock; 12720 } 12721 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12722 goto top; 12723 } 12724 if (newblk->nb_state & DEPCOMPLETE || 12725 waitfor == MNT_NOWAIT) 12726 continue; 12727 nbp = newblk->nb_bmsafemap->sm_buf; 12728 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12729 if (nbp == NULL) 12730 goto top; 12731 FREE_LOCK(ump); 12732 if ((error = bwrite(nbp)) != 0) 12733 goto out; 12734 ACQUIRE_LOCK(ump); 12735 continue; 12736 12737 case D_INDIRDEP: 12738 indirdep = WK_INDIRDEP(wk); 12739 if (waitfor == MNT_NOWAIT) { 12740 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12741 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12742 error = EBUSY; 12743 goto out_unlock; 12744 } 12745 } 12746 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12747 panic("softdep_sync_buf: truncation pending."); 12748 restart: 12749 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12750 newblk = (struct newblk *)aip; 12751 if (newblk->nb_jnewblk != NULL) { 12752 jwait(&newblk->nb_jnewblk->jn_list, 12753 waitfor); 12754 goto restart; 12755 } 12756 if (newblk->nb_state & DEPCOMPLETE) 12757 continue; 12758 nbp = newblk->nb_bmsafemap->sm_buf; 12759 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12760 if (nbp == NULL) 12761 goto restart; 12762 FREE_LOCK(ump); 12763 if ((error = bwrite(nbp)) != 0) 12764 goto out; 12765 ACQUIRE_LOCK(ump); 12766 goto restart; 12767 } 12768 continue; 12769 12770 case D_PAGEDEP: 12771 /* 12772 * Only flush directory entries in synchronous passes. 12773 */ 12774 if (waitfor != MNT_WAIT) { 12775 error = EBUSY; 12776 goto out_unlock; 12777 } 12778 /* 12779 * While syncing snapshots, we must allow recursive 12780 * lookups. 12781 */ 12782 BUF_AREC(bp); 12783 /* 12784 * We are trying to sync a directory that may 12785 * have dependencies on both its own metadata 12786 * and/or dependencies on the inodes of any 12787 * recently allocated files. We walk its diradd 12788 * lists pushing out the associated inode. 12789 */ 12790 pagedep = WK_PAGEDEP(wk); 12791 for (i = 0; i < DAHASHSZ; i++) { 12792 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12793 continue; 12794 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12795 &pagedep->pd_diraddhd[i]))) { 12796 BUF_NOREC(bp); 12797 goto out_unlock; 12798 } 12799 } 12800 BUF_NOREC(bp); 12801 continue; 12802 12803 case D_FREEWORK: 12804 case D_FREEDEP: 12805 case D_JSEGDEP: 12806 case D_JNEWBLK: 12807 continue; 12808 12809 default: 12810 panic("softdep_sync_buf: Unknown type %s", 12811 TYPENAME(wk->wk_type)); 12812 /* NOTREACHED */ 12813 } 12814 } 12815 out_unlock: 12816 FREE_LOCK(ump); 12817 out: 12818 return (error); 12819 } 12820 12821 /* 12822 * Flush the dependencies associated with an inodedep. 12823 * Called with splbio blocked. 12824 */ 12825 static int 12826 flush_inodedep_deps(vp, mp, ino) 12827 struct vnode *vp; 12828 struct mount *mp; 12829 ino_t ino; 12830 { 12831 struct inodedep *inodedep; 12832 struct inoref *inoref; 12833 struct ufsmount *ump; 12834 int error, waitfor; 12835 12836 /* 12837 * This work is done in two passes. The first pass grabs most 12838 * of the buffers and begins asynchronously writing them. The 12839 * only way to wait for these asynchronous writes is to sleep 12840 * on the filesystem vnode which may stay busy for a long time 12841 * if the filesystem is active. So, instead, we make a second 12842 * pass over the dependencies blocking on each write. In the 12843 * usual case we will be blocking against a write that we 12844 * initiated, so when it is done the dependency will have been 12845 * resolved. Thus the second pass is expected to end quickly. 12846 * We give a brief window at the top of the loop to allow 12847 * any pending I/O to complete. 12848 */ 12849 ump = VFSTOUFS(mp); 12850 LOCK_OWNED(ump); 12851 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12852 if (error) 12853 return (error); 12854 FREE_LOCK(ump); 12855 ACQUIRE_LOCK(ump); 12856 restart: 12857 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12858 return (0); 12859 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12860 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12861 == DEPCOMPLETE) { 12862 jwait(&inoref->if_list, MNT_WAIT); 12863 goto restart; 12864 } 12865 } 12866 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12867 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12868 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12869 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12870 continue; 12871 /* 12872 * If pass2, we are done, otherwise do pass 2. 12873 */ 12874 if (waitfor == MNT_WAIT) 12875 break; 12876 waitfor = MNT_WAIT; 12877 } 12878 /* 12879 * Try freeing inodedep in case all dependencies have been removed. 12880 */ 12881 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12882 (void) free_inodedep(inodedep); 12883 return (0); 12884 } 12885 12886 /* 12887 * Flush an inode dependency list. 12888 * Called with splbio blocked. 12889 */ 12890 static int 12891 flush_deplist(listhead, waitfor, errorp) 12892 struct allocdirectlst *listhead; 12893 int waitfor; 12894 int *errorp; 12895 { 12896 struct allocdirect *adp; 12897 struct newblk *newblk; 12898 struct ufsmount *ump; 12899 struct buf *bp; 12900 12901 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12902 return (0); 12903 ump = VFSTOUFS(adp->ad_list.wk_mp); 12904 LOCK_OWNED(ump); 12905 TAILQ_FOREACH(adp, listhead, ad_next) { 12906 newblk = (struct newblk *)adp; 12907 if (newblk->nb_jnewblk != NULL) { 12908 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12909 return (1); 12910 } 12911 if (newblk->nb_state & DEPCOMPLETE) 12912 continue; 12913 bp = newblk->nb_bmsafemap->sm_buf; 12914 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12915 if (bp == NULL) { 12916 if (waitfor == MNT_NOWAIT) 12917 continue; 12918 return (1); 12919 } 12920 FREE_LOCK(ump); 12921 if (waitfor == MNT_NOWAIT) 12922 bawrite(bp); 12923 else 12924 *errorp = bwrite(bp); 12925 ACQUIRE_LOCK(ump); 12926 return (1); 12927 } 12928 return (0); 12929 } 12930 12931 /* 12932 * Flush dependencies associated with an allocdirect block. 12933 */ 12934 static int 12935 flush_newblk_dep(vp, mp, lbn) 12936 struct vnode *vp; 12937 struct mount *mp; 12938 ufs_lbn_t lbn; 12939 { 12940 struct newblk *newblk; 12941 struct ufsmount *ump; 12942 struct bufobj *bo; 12943 struct inode *ip; 12944 struct buf *bp; 12945 ufs2_daddr_t blkno; 12946 int error; 12947 12948 error = 0; 12949 bo = &vp->v_bufobj; 12950 ip = VTOI(vp); 12951 blkno = DIP(ip, i_db[lbn]); 12952 if (blkno == 0) 12953 panic("flush_newblk_dep: Missing block"); 12954 ump = VFSTOUFS(mp); 12955 ACQUIRE_LOCK(ump); 12956 /* 12957 * Loop until all dependencies related to this block are satisfied. 12958 * We must be careful to restart after each sleep in case a write 12959 * completes some part of this process for us. 12960 */ 12961 for (;;) { 12962 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12963 FREE_LOCK(ump); 12964 break; 12965 } 12966 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12967 panic("flush_newblk_dep: Bad newblk %p", newblk); 12968 /* 12969 * Flush the journal. 12970 */ 12971 if (newblk->nb_jnewblk != NULL) { 12972 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12973 continue; 12974 } 12975 /* 12976 * Write the bitmap dependency. 12977 */ 12978 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12979 bp = newblk->nb_bmsafemap->sm_buf; 12980 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12981 if (bp == NULL) 12982 continue; 12983 FREE_LOCK(ump); 12984 error = bwrite(bp); 12985 if (error) 12986 break; 12987 ACQUIRE_LOCK(ump); 12988 continue; 12989 } 12990 /* 12991 * Write the buffer. 12992 */ 12993 FREE_LOCK(ump); 12994 BO_LOCK(bo); 12995 bp = gbincore(bo, lbn); 12996 if (bp != NULL) { 12997 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12998 LK_INTERLOCK, BO_LOCKPTR(bo)); 12999 if (error == ENOLCK) { 13000 ACQUIRE_LOCK(ump); 13001 error = 0; 13002 continue; /* Slept, retry */ 13003 } 13004 if (error != 0) 13005 break; /* Failed */ 13006 if (bp->b_flags & B_DELWRI) { 13007 bremfree(bp); 13008 error = bwrite(bp); 13009 if (error) 13010 break; 13011 } else 13012 BUF_UNLOCK(bp); 13013 } else 13014 BO_UNLOCK(bo); 13015 /* 13016 * We have to wait for the direct pointers to 13017 * point at the newdirblk before the dependency 13018 * will go away. 13019 */ 13020 error = ffs_update(vp, 1); 13021 if (error) 13022 break; 13023 ACQUIRE_LOCK(ump); 13024 } 13025 return (error); 13026 } 13027 13028 /* 13029 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 13030 * Called with splbio blocked. 13031 */ 13032 static int 13033 flush_pagedep_deps(pvp, mp, diraddhdp) 13034 struct vnode *pvp; 13035 struct mount *mp; 13036 struct diraddhd *diraddhdp; 13037 { 13038 struct inodedep *inodedep; 13039 struct inoref *inoref; 13040 struct ufsmount *ump; 13041 struct diradd *dap; 13042 struct vnode *vp; 13043 int error = 0; 13044 struct buf *bp; 13045 ino_t inum; 13046 struct diraddhd unfinished; 13047 13048 LIST_INIT(&unfinished); 13049 ump = VFSTOUFS(mp); 13050 LOCK_OWNED(ump); 13051 restart: 13052 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 13053 /* 13054 * Flush ourselves if this directory entry 13055 * has a MKDIR_PARENT dependency. 13056 */ 13057 if (dap->da_state & MKDIR_PARENT) { 13058 FREE_LOCK(ump); 13059 if ((error = ffs_update(pvp, 1)) != 0) 13060 break; 13061 ACQUIRE_LOCK(ump); 13062 /* 13063 * If that cleared dependencies, go on to next. 13064 */ 13065 if (dap != LIST_FIRST(diraddhdp)) 13066 continue; 13067 /* 13068 * All MKDIR_PARENT dependencies and all the 13069 * NEWBLOCK pagedeps that are contained in direct 13070 * blocks were resolved by doing above ffs_update. 13071 * Pagedeps contained in indirect blocks may 13072 * require a complete sync'ing of the directory. 13073 * We are in the midst of doing a complete sync, 13074 * so if they are not resolved in this pass we 13075 * defer them for now as they will be sync'ed by 13076 * our caller shortly. 13077 */ 13078 LIST_REMOVE(dap, da_pdlist); 13079 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 13080 continue; 13081 } 13082 /* 13083 * A newly allocated directory must have its "." and 13084 * ".." entries written out before its name can be 13085 * committed in its parent. 13086 */ 13087 inum = dap->da_newinum; 13088 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13089 panic("flush_pagedep_deps: lost inode1"); 13090 /* 13091 * Wait for any pending journal adds to complete so we don't 13092 * cause rollbacks while syncing. 13093 */ 13094 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13095 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13096 == DEPCOMPLETE) { 13097 jwait(&inoref->if_list, MNT_WAIT); 13098 goto restart; 13099 } 13100 } 13101 if (dap->da_state & MKDIR_BODY) { 13102 FREE_LOCK(ump); 13103 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13104 FFSV_FORCEINSMQ))) 13105 break; 13106 error = flush_newblk_dep(vp, mp, 0); 13107 /* 13108 * If we still have the dependency we might need to 13109 * update the vnode to sync the new link count to 13110 * disk. 13111 */ 13112 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 13113 error = ffs_update(vp, 1); 13114 vput(vp); 13115 if (error != 0) 13116 break; 13117 ACQUIRE_LOCK(ump); 13118 /* 13119 * If that cleared dependencies, go on to next. 13120 */ 13121 if (dap != LIST_FIRST(diraddhdp)) 13122 continue; 13123 if (dap->da_state & MKDIR_BODY) { 13124 inodedep_lookup(UFSTOVFS(ump), inum, 0, 13125 &inodedep); 13126 panic("flush_pagedep_deps: MKDIR_BODY " 13127 "inodedep %p dap %p vp %p", 13128 inodedep, dap, vp); 13129 } 13130 } 13131 /* 13132 * Flush the inode on which the directory entry depends. 13133 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 13134 * the only remaining dependency is that the updated inode 13135 * count must get pushed to disk. The inode has already 13136 * been pushed into its inode buffer (via VOP_UPDATE) at 13137 * the time of the reference count change. So we need only 13138 * locate that buffer, ensure that there will be no rollback 13139 * caused by a bitmap dependency, then write the inode buffer. 13140 */ 13141 retry: 13142 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13143 panic("flush_pagedep_deps: lost inode"); 13144 /* 13145 * If the inode still has bitmap dependencies, 13146 * push them to disk. 13147 */ 13148 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 13149 bp = inodedep->id_bmsafemap->sm_buf; 13150 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13151 if (bp == NULL) 13152 goto retry; 13153 FREE_LOCK(ump); 13154 if ((error = bwrite(bp)) != 0) 13155 break; 13156 ACQUIRE_LOCK(ump); 13157 if (dap != LIST_FIRST(diraddhdp)) 13158 continue; 13159 } 13160 /* 13161 * If the inode is still sitting in a buffer waiting 13162 * to be written or waiting for the link count to be 13163 * adjusted update it here to flush it to disk. 13164 */ 13165 if (dap == LIST_FIRST(diraddhdp)) { 13166 FREE_LOCK(ump); 13167 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13168 FFSV_FORCEINSMQ))) 13169 break; 13170 error = ffs_update(vp, 1); 13171 vput(vp); 13172 if (error) 13173 break; 13174 ACQUIRE_LOCK(ump); 13175 } 13176 /* 13177 * If we have failed to get rid of all the dependencies 13178 * then something is seriously wrong. 13179 */ 13180 if (dap == LIST_FIRST(diraddhdp)) { 13181 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13182 panic("flush_pagedep_deps: failed to flush " 13183 "inodedep %p ino %ju dap %p", 13184 inodedep, (uintmax_t)inum, dap); 13185 } 13186 } 13187 if (error) 13188 ACQUIRE_LOCK(ump); 13189 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13190 LIST_REMOVE(dap, da_pdlist); 13191 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13192 } 13193 return (error); 13194 } 13195 13196 /* 13197 * A large burst of file addition or deletion activity can drive the 13198 * memory load excessively high. First attempt to slow things down 13199 * using the techniques below. If that fails, this routine requests 13200 * the offending operations to fall back to running synchronously 13201 * until the memory load returns to a reasonable level. 13202 */ 13203 int 13204 softdep_slowdown(vp) 13205 struct vnode *vp; 13206 { 13207 struct ufsmount *ump; 13208 int jlow; 13209 int max_softdeps_hard; 13210 13211 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13212 ("softdep_slowdown called on non-softdep filesystem")); 13213 ump = VFSTOUFS(vp->v_mount); 13214 ACQUIRE_LOCK(ump); 13215 jlow = 0; 13216 /* 13217 * Check for journal space if needed. 13218 */ 13219 if (DOINGSUJ(vp)) { 13220 if (journal_space(ump, 0) == 0) 13221 jlow = 1; 13222 } 13223 /* 13224 * If the system is under its limits and our filesystem is 13225 * not responsible for more than our share of the usage and 13226 * we are not low on journal space, then no need to slow down. 13227 */ 13228 max_softdeps_hard = max_softdeps * 11 / 10; 13229 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13230 dep_current[D_INODEDEP] < max_softdeps_hard && 13231 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13232 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13233 ump->softdep_curdeps[D_DIRREM] < 13234 (max_softdeps_hard / 2) / stat_flush_threads && 13235 ump->softdep_curdeps[D_INODEDEP] < 13236 max_softdeps_hard / stat_flush_threads && 13237 ump->softdep_curdeps[D_INDIRDEP] < 13238 (max_softdeps_hard / 1000) / stat_flush_threads && 13239 ump->softdep_curdeps[D_FREEBLKS] < 13240 max_softdeps_hard / stat_flush_threads) { 13241 FREE_LOCK(ump); 13242 return (0); 13243 } 13244 /* 13245 * If the journal is low or our filesystem is over its limit 13246 * then speedup the cleanup. 13247 */ 13248 if (ump->softdep_curdeps[D_INDIRDEP] < 13249 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13250 softdep_speedup(ump); 13251 stat_sync_limit_hit += 1; 13252 FREE_LOCK(ump); 13253 /* 13254 * We only slow down the rate at which new dependencies are 13255 * generated if we are not using journaling. With journaling, 13256 * the cleanup should always be sufficient to keep things 13257 * under control. 13258 */ 13259 if (DOINGSUJ(vp)) 13260 return (0); 13261 return (1); 13262 } 13263 13264 /* 13265 * Called by the allocation routines when they are about to fail 13266 * in the hope that we can free up the requested resource (inodes 13267 * or disk space). 13268 * 13269 * First check to see if the work list has anything on it. If it has, 13270 * clean up entries until we successfully free the requested resource. 13271 * Because this process holds inodes locked, we cannot handle any remove 13272 * requests that might block on a locked inode as that could lead to 13273 * deadlock. If the worklist yields none of the requested resource, 13274 * start syncing out vnodes to free up the needed space. 13275 */ 13276 int 13277 softdep_request_cleanup(fs, vp, cred, resource) 13278 struct fs *fs; 13279 struct vnode *vp; 13280 struct ucred *cred; 13281 int resource; 13282 { 13283 struct ufsmount *ump; 13284 struct mount *mp; 13285 long starttime; 13286 ufs2_daddr_t needed; 13287 int error, failed_vnode; 13288 13289 /* 13290 * If we are being called because of a process doing a 13291 * copy-on-write, then it is not safe to process any 13292 * worklist items as we will recurse into the copyonwrite 13293 * routine. This will result in an incoherent snapshot. 13294 * If the vnode that we hold is a snapshot, we must avoid 13295 * handling other resources that could cause deadlock. 13296 */ 13297 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13298 return (0); 13299 13300 if (resource == FLUSH_BLOCKS_WAIT) 13301 stat_cleanup_blkrequests += 1; 13302 else 13303 stat_cleanup_inorequests += 1; 13304 13305 mp = vp->v_mount; 13306 ump = VFSTOUFS(mp); 13307 mtx_assert(UFS_MTX(ump), MA_OWNED); 13308 UFS_UNLOCK(ump); 13309 error = ffs_update(vp, 1); 13310 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13311 UFS_LOCK(ump); 13312 return (0); 13313 } 13314 /* 13315 * If we are in need of resources, start by cleaning up 13316 * any block removals associated with our inode. 13317 */ 13318 ACQUIRE_LOCK(ump); 13319 process_removes(vp); 13320 process_truncates(vp); 13321 FREE_LOCK(ump); 13322 /* 13323 * Now clean up at least as many resources as we will need. 13324 * 13325 * When requested to clean up inodes, the number that are needed 13326 * is set by the number of simultaneous writers (mnt_writeopcount) 13327 * plus a bit of slop (2) in case some more writers show up while 13328 * we are cleaning. 13329 * 13330 * When requested to free up space, the amount of space that 13331 * we need is enough blocks to allocate a full-sized segment 13332 * (fs_contigsumsize). The number of such segments that will 13333 * be needed is set by the number of simultaneous writers 13334 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13335 * writers show up while we are cleaning. 13336 * 13337 * Additionally, if we are unpriviledged and allocating space, 13338 * we need to ensure that we clean up enough blocks to get the 13339 * needed number of blocks over the threshold of the minimum 13340 * number of blocks required to be kept free by the filesystem 13341 * (fs_minfree). 13342 */ 13343 if (resource == FLUSH_INODES_WAIT) { 13344 needed = vp->v_mount->mnt_writeopcount + 2; 13345 } else if (resource == FLUSH_BLOCKS_WAIT) { 13346 needed = (vp->v_mount->mnt_writeopcount + 2) * 13347 fs->fs_contigsumsize; 13348 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 13349 needed += fragstoblks(fs, 13350 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13351 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13352 } else { 13353 UFS_LOCK(ump); 13354 printf("softdep_request_cleanup: Unknown resource type %d\n", 13355 resource); 13356 return (0); 13357 } 13358 starttime = time_second; 13359 retry: 13360 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13361 fs->fs_cstotal.cs_nbfree <= needed) || 13362 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13363 fs->fs_cstotal.cs_nifree <= needed)) { 13364 ACQUIRE_LOCK(ump); 13365 if (ump->softdep_on_worklist > 0 && 13366 process_worklist_item(UFSTOVFS(ump), 13367 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13368 stat_worklist_push += 1; 13369 FREE_LOCK(ump); 13370 } 13371 /* 13372 * If we still need resources and there are no more worklist 13373 * entries to process to obtain them, we have to start flushing 13374 * the dirty vnodes to force the release of additional requests 13375 * to the worklist that we can then process to reap addition 13376 * resources. We walk the vnodes associated with the mount point 13377 * until we get the needed worklist requests that we can reap. 13378 * 13379 * If there are several threads all needing to clean the same 13380 * mount point, only one is allowed to walk the mount list. 13381 * When several threads all try to walk the same mount list, 13382 * they end up competing with each other and often end up in 13383 * livelock. This approach ensures that forward progress is 13384 * made at the cost of occational ENOSPC errors being returned 13385 * that might otherwise have been avoided. 13386 */ 13387 error = 1; 13388 if ((resource == FLUSH_BLOCKS_WAIT && 13389 fs->fs_cstotal.cs_nbfree <= needed) || 13390 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13391 fs->fs_cstotal.cs_nifree <= needed)) { 13392 ACQUIRE_LOCK(ump); 13393 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) { 13394 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE; 13395 FREE_LOCK(ump); 13396 failed_vnode = softdep_request_cleanup_flush(mp, ump); 13397 ACQUIRE_LOCK(ump); 13398 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE; 13399 FREE_LOCK(ump); 13400 if (ump->softdep_on_worklist > 0) { 13401 stat_cleanup_retries += 1; 13402 if (!failed_vnode) 13403 goto retry; 13404 } 13405 } else { 13406 FREE_LOCK(ump); 13407 error = 0; 13408 } 13409 stat_cleanup_failures += 1; 13410 } 13411 if (time_second - starttime > stat_cleanup_high_delay) 13412 stat_cleanup_high_delay = time_second - starttime; 13413 UFS_LOCK(ump); 13414 return (error); 13415 } 13416 13417 /* 13418 * Scan the vnodes for the specified mount point flushing out any 13419 * vnodes that can be locked without waiting. Finally, try to flush 13420 * the device associated with the mount point if it can be locked 13421 * without waiting. 13422 * 13423 * We return 0 if we were able to lock every vnode in our scan. 13424 * If we had to skip one or more vnodes, we return 1. 13425 */ 13426 static int 13427 softdep_request_cleanup_flush(mp, ump) 13428 struct mount *mp; 13429 struct ufsmount *ump; 13430 { 13431 struct thread *td; 13432 struct vnode *lvp, *mvp; 13433 int failed_vnode; 13434 13435 failed_vnode = 0; 13436 td = curthread; 13437 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13438 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13439 VI_UNLOCK(lvp); 13440 continue; 13441 } 13442 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13443 td) != 0) { 13444 failed_vnode = 1; 13445 continue; 13446 } 13447 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13448 vput(lvp); 13449 continue; 13450 } 13451 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13452 vput(lvp); 13453 } 13454 lvp = ump->um_devvp; 13455 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13456 VOP_FSYNC(lvp, MNT_NOWAIT, td); 13457 VOP_UNLOCK(lvp, 0); 13458 } 13459 return (failed_vnode); 13460 } 13461 13462 static bool 13463 softdep_excess_items(struct ufsmount *ump, int item) 13464 { 13465 13466 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13467 return (dep_current[item] > max_softdeps && 13468 ump->softdep_curdeps[item] > max_softdeps / 13469 stat_flush_threads); 13470 } 13471 13472 static void 13473 schedule_cleanup(struct mount *mp) 13474 { 13475 struct ufsmount *ump; 13476 struct thread *td; 13477 13478 ump = VFSTOUFS(mp); 13479 LOCK_OWNED(ump); 13480 FREE_LOCK(ump); 13481 td = curthread; 13482 if ((td->td_pflags & TDP_KTHREAD) != 0 && 13483 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 13484 /* 13485 * No ast is delivered to kernel threads, so nobody 13486 * would deref the mp. Some kernel threads 13487 * explicitely check for AST, e.g. NFS daemon does 13488 * this in the serving loop. 13489 */ 13490 return; 13491 } 13492 if (td->td_su != NULL) 13493 vfs_rel(td->td_su); 13494 vfs_ref(mp); 13495 td->td_su = mp; 13496 thread_lock(td); 13497 td->td_flags |= TDF_ASTPENDING; 13498 thread_unlock(td); 13499 } 13500 13501 static void 13502 softdep_ast_cleanup_proc(struct thread *td) 13503 { 13504 struct mount *mp; 13505 struct ufsmount *ump; 13506 int error; 13507 bool req; 13508 13509 while ((mp = td->td_su) != NULL) { 13510 td->td_su = NULL; 13511 error = vfs_busy(mp, MBF_NOWAIT); 13512 vfs_rel(mp); 13513 if (error != 0) 13514 return; 13515 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 13516 ump = VFSTOUFS(mp); 13517 for (;;) { 13518 req = false; 13519 ACQUIRE_LOCK(ump); 13520 if (softdep_excess_items(ump, D_INODEDEP)) { 13521 req = true; 13522 request_cleanup(mp, FLUSH_INODES); 13523 } 13524 if (softdep_excess_items(ump, D_DIRREM)) { 13525 req = true; 13526 request_cleanup(mp, FLUSH_BLOCKS); 13527 } 13528 FREE_LOCK(ump); 13529 if (softdep_excess_items(ump, D_NEWBLK) || 13530 softdep_excess_items(ump, D_ALLOCDIRECT) || 13531 softdep_excess_items(ump, D_ALLOCINDIR)) { 13532 error = vn_start_write(NULL, &mp, 13533 V_WAIT); 13534 if (error == 0) { 13535 req = true; 13536 VFS_SYNC(mp, MNT_WAIT); 13537 vn_finished_write(mp); 13538 } 13539 } 13540 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 13541 break; 13542 } 13543 } 13544 vfs_unbusy(mp); 13545 } 13546 if ((mp = td->td_su) != NULL) { 13547 td->td_su = NULL; 13548 vfs_rel(mp); 13549 } 13550 } 13551 13552 /* 13553 * If memory utilization has gotten too high, deliberately slow things 13554 * down and speed up the I/O processing. 13555 */ 13556 static int 13557 request_cleanup(mp, resource) 13558 struct mount *mp; 13559 int resource; 13560 { 13561 struct thread *td = curthread; 13562 struct ufsmount *ump; 13563 13564 ump = VFSTOUFS(mp); 13565 LOCK_OWNED(ump); 13566 /* 13567 * We never hold up the filesystem syncer or buf daemon. 13568 */ 13569 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13570 return (0); 13571 /* 13572 * First check to see if the work list has gotten backlogged. 13573 * If it has, co-opt this process to help clean up two entries. 13574 * Because this process may hold inodes locked, we cannot 13575 * handle any remove requests that might block on a locked 13576 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13577 * to avoid recursively processing the worklist. 13578 */ 13579 if (ump->softdep_on_worklist > max_softdeps / 10) { 13580 td->td_pflags |= TDP_SOFTDEP; 13581 process_worklist_item(mp, 2, LK_NOWAIT); 13582 td->td_pflags &= ~TDP_SOFTDEP; 13583 stat_worklist_push += 2; 13584 return(1); 13585 } 13586 /* 13587 * Next, we attempt to speed up the syncer process. If that 13588 * is successful, then we allow the process to continue. 13589 */ 13590 if (softdep_speedup(ump) && 13591 resource != FLUSH_BLOCKS_WAIT && 13592 resource != FLUSH_INODES_WAIT) 13593 return(0); 13594 /* 13595 * If we are resource constrained on inode dependencies, try 13596 * flushing some dirty inodes. Otherwise, we are constrained 13597 * by file deletions, so try accelerating flushes of directories 13598 * with removal dependencies. We would like to do the cleanup 13599 * here, but we probably hold an inode locked at this point and 13600 * that might deadlock against one that we try to clean. So, 13601 * the best that we can do is request the syncer daemon to do 13602 * the cleanup for us. 13603 */ 13604 switch (resource) { 13605 13606 case FLUSH_INODES: 13607 case FLUSH_INODES_WAIT: 13608 ACQUIRE_GBLLOCK(&lk); 13609 stat_ino_limit_push += 1; 13610 req_clear_inodedeps += 1; 13611 FREE_GBLLOCK(&lk); 13612 stat_countp = &stat_ino_limit_hit; 13613 break; 13614 13615 case FLUSH_BLOCKS: 13616 case FLUSH_BLOCKS_WAIT: 13617 ACQUIRE_GBLLOCK(&lk); 13618 stat_blk_limit_push += 1; 13619 req_clear_remove += 1; 13620 FREE_GBLLOCK(&lk); 13621 stat_countp = &stat_blk_limit_hit; 13622 break; 13623 13624 default: 13625 panic("request_cleanup: unknown type"); 13626 } 13627 /* 13628 * Hopefully the syncer daemon will catch up and awaken us. 13629 * We wait at most tickdelay before proceeding in any case. 13630 */ 13631 ACQUIRE_GBLLOCK(&lk); 13632 FREE_LOCK(ump); 13633 proc_waiting += 1; 13634 if (callout_pending(&softdep_callout) == FALSE) 13635 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13636 pause_timer, 0); 13637 13638 if ((td->td_pflags & TDP_KTHREAD) == 0) 13639 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13640 proc_waiting -= 1; 13641 FREE_GBLLOCK(&lk); 13642 ACQUIRE_LOCK(ump); 13643 return (1); 13644 } 13645 13646 /* 13647 * Awaken processes pausing in request_cleanup and clear proc_waiting 13648 * to indicate that there is no longer a timer running. Pause_timer 13649 * will be called with the global softdep mutex (&lk) locked. 13650 */ 13651 static void 13652 pause_timer(arg) 13653 void *arg; 13654 { 13655 13656 GBLLOCK_OWNED(&lk); 13657 /* 13658 * The callout_ API has acquired mtx and will hold it around this 13659 * function call. 13660 */ 13661 *stat_countp += proc_waiting; 13662 wakeup(&proc_waiting); 13663 } 13664 13665 /* 13666 * If requested, try removing inode or removal dependencies. 13667 */ 13668 static void 13669 check_clear_deps(mp) 13670 struct mount *mp; 13671 { 13672 13673 /* 13674 * If we are suspended, it may be because of our using 13675 * too many inodedeps, so help clear them out. 13676 */ 13677 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13678 clear_inodedeps(mp); 13679 /* 13680 * General requests for cleanup of backed up dependencies 13681 */ 13682 ACQUIRE_GBLLOCK(&lk); 13683 if (req_clear_inodedeps) { 13684 req_clear_inodedeps -= 1; 13685 FREE_GBLLOCK(&lk); 13686 clear_inodedeps(mp); 13687 ACQUIRE_GBLLOCK(&lk); 13688 wakeup(&proc_waiting); 13689 } 13690 if (req_clear_remove) { 13691 req_clear_remove -= 1; 13692 FREE_GBLLOCK(&lk); 13693 clear_remove(mp); 13694 ACQUIRE_GBLLOCK(&lk); 13695 wakeup(&proc_waiting); 13696 } 13697 FREE_GBLLOCK(&lk); 13698 } 13699 13700 /* 13701 * Flush out a directory with at least one removal dependency in an effort to 13702 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13703 */ 13704 static void 13705 clear_remove(mp) 13706 struct mount *mp; 13707 { 13708 struct pagedep_hashhead *pagedephd; 13709 struct pagedep *pagedep; 13710 struct ufsmount *ump; 13711 struct vnode *vp; 13712 struct bufobj *bo; 13713 int error, cnt; 13714 ino_t ino; 13715 13716 ump = VFSTOUFS(mp); 13717 LOCK_OWNED(ump); 13718 13719 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13720 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13721 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13722 ump->pagedep_nextclean = 0; 13723 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13724 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13725 continue; 13726 ino = pagedep->pd_ino; 13727 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13728 continue; 13729 FREE_LOCK(ump); 13730 13731 /* 13732 * Let unmount clear deps 13733 */ 13734 error = vfs_busy(mp, MBF_NOWAIT); 13735 if (error != 0) 13736 goto finish_write; 13737 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13738 FFSV_FORCEINSMQ); 13739 vfs_unbusy(mp); 13740 if (error != 0) { 13741 softdep_error("clear_remove: vget", error); 13742 goto finish_write; 13743 } 13744 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13745 softdep_error("clear_remove: fsync", error); 13746 bo = &vp->v_bufobj; 13747 BO_LOCK(bo); 13748 drain_output(vp); 13749 BO_UNLOCK(bo); 13750 vput(vp); 13751 finish_write: 13752 vn_finished_write(mp); 13753 ACQUIRE_LOCK(ump); 13754 return; 13755 } 13756 } 13757 } 13758 13759 /* 13760 * Clear out a block of dirty inodes in an effort to reduce 13761 * the number of inodedep dependency structures. 13762 */ 13763 static void 13764 clear_inodedeps(mp) 13765 struct mount *mp; 13766 { 13767 struct inodedep_hashhead *inodedephd; 13768 struct inodedep *inodedep; 13769 struct ufsmount *ump; 13770 struct vnode *vp; 13771 struct fs *fs; 13772 int error, cnt; 13773 ino_t firstino, lastino, ino; 13774 13775 ump = VFSTOUFS(mp); 13776 fs = ump->um_fs; 13777 LOCK_OWNED(ump); 13778 /* 13779 * Pick a random inode dependency to be cleared. 13780 * We will then gather up all the inodes in its block 13781 * that have dependencies and flush them out. 13782 */ 13783 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13784 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13785 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13786 ump->inodedep_nextclean = 0; 13787 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13788 break; 13789 } 13790 if (inodedep == NULL) 13791 return; 13792 /* 13793 * Find the last inode in the block with dependencies. 13794 */ 13795 firstino = rounddown2(inodedep->id_ino, INOPB(fs)); 13796 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13797 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13798 break; 13799 /* 13800 * Asynchronously push all but the last inode with dependencies. 13801 * Synchronously push the last inode with dependencies to ensure 13802 * that the inode block gets written to free up the inodedeps. 13803 */ 13804 for (ino = firstino; ino <= lastino; ino++) { 13805 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13806 continue; 13807 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13808 continue; 13809 FREE_LOCK(ump); 13810 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13811 if (error != 0) { 13812 vn_finished_write(mp); 13813 ACQUIRE_LOCK(ump); 13814 return; 13815 } 13816 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13817 FFSV_FORCEINSMQ)) != 0) { 13818 softdep_error("clear_inodedeps: vget", error); 13819 vfs_unbusy(mp); 13820 vn_finished_write(mp); 13821 ACQUIRE_LOCK(ump); 13822 return; 13823 } 13824 vfs_unbusy(mp); 13825 if (ino == lastino) { 13826 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13827 softdep_error("clear_inodedeps: fsync1", error); 13828 } else { 13829 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13830 softdep_error("clear_inodedeps: fsync2", error); 13831 BO_LOCK(&vp->v_bufobj); 13832 drain_output(vp); 13833 BO_UNLOCK(&vp->v_bufobj); 13834 } 13835 vput(vp); 13836 vn_finished_write(mp); 13837 ACQUIRE_LOCK(ump); 13838 } 13839 } 13840 13841 void 13842 softdep_buf_append(bp, wkhd) 13843 struct buf *bp; 13844 struct workhead *wkhd; 13845 { 13846 struct worklist *wk; 13847 struct ufsmount *ump; 13848 13849 if ((wk = LIST_FIRST(wkhd)) == NULL) 13850 return; 13851 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13852 ("softdep_buf_append called on non-softdep filesystem")); 13853 ump = VFSTOUFS(wk->wk_mp); 13854 ACQUIRE_LOCK(ump); 13855 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13856 WORKLIST_REMOVE(wk); 13857 WORKLIST_INSERT(&bp->b_dep, wk); 13858 } 13859 FREE_LOCK(ump); 13860 13861 } 13862 13863 void 13864 softdep_inode_append(ip, cred, wkhd) 13865 struct inode *ip; 13866 struct ucred *cred; 13867 struct workhead *wkhd; 13868 { 13869 struct buf *bp; 13870 struct fs *fs; 13871 struct ufsmount *ump; 13872 int error; 13873 13874 ump = ITOUMP(ip); 13875 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 13876 ("softdep_inode_append called on non-softdep filesystem")); 13877 fs = ump->um_fs; 13878 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13879 (int)fs->fs_bsize, cred, &bp); 13880 if (error) { 13881 bqrelse(bp); 13882 softdep_freework(wkhd); 13883 return; 13884 } 13885 softdep_buf_append(bp, wkhd); 13886 bqrelse(bp); 13887 } 13888 13889 void 13890 softdep_freework(wkhd) 13891 struct workhead *wkhd; 13892 { 13893 struct worklist *wk; 13894 struct ufsmount *ump; 13895 13896 if ((wk = LIST_FIRST(wkhd)) == NULL) 13897 return; 13898 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13899 ("softdep_freework called on non-softdep filesystem")); 13900 ump = VFSTOUFS(wk->wk_mp); 13901 ACQUIRE_LOCK(ump); 13902 handle_jwork(wkhd); 13903 FREE_LOCK(ump); 13904 } 13905 13906 static struct ufsmount * 13907 softdep_bp_to_mp(bp) 13908 struct buf *bp; 13909 { 13910 struct mount *mp; 13911 struct vnode *vp; 13912 13913 if (LIST_EMPTY(&bp->b_dep)) 13914 return (NULL); 13915 vp = bp->b_vp; 13916 13917 /* 13918 * The ump mount point is stable after we get a correct 13919 * pointer, since bp is locked and this prevents unmount from 13920 * proceeding. But to get to it, we cannot dereference bp->b_dep 13921 * head wk_mp, because we do not yet own SU ump lock and 13922 * workitem might be freed while dereferenced. 13923 */ 13924 retry: 13925 if (vp->v_type == VCHR) { 13926 VI_LOCK(vp); 13927 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL; 13928 VI_UNLOCK(vp); 13929 if (mp == NULL) 13930 goto retry; 13931 } else if (vp->v_type == VREG || vp->v_type == VDIR || 13932 vp->v_type == VLNK) { 13933 mp = vp->v_mount; 13934 } else { 13935 return (NULL); 13936 } 13937 return (VFSTOUFS(mp)); 13938 } 13939 13940 /* 13941 * Function to determine if the buffer has outstanding dependencies 13942 * that will cause a roll-back if the buffer is written. If wantcount 13943 * is set, return number of dependencies, otherwise just yes or no. 13944 */ 13945 static int 13946 softdep_count_dependencies(bp, wantcount) 13947 struct buf *bp; 13948 int wantcount; 13949 { 13950 struct worklist *wk; 13951 struct ufsmount *ump; 13952 struct bmsafemap *bmsafemap; 13953 struct freework *freework; 13954 struct inodedep *inodedep; 13955 struct indirdep *indirdep; 13956 struct freeblks *freeblks; 13957 struct allocindir *aip; 13958 struct pagedep *pagedep; 13959 struct dirrem *dirrem; 13960 struct newblk *newblk; 13961 struct mkdir *mkdir; 13962 struct diradd *dap; 13963 int i, retval; 13964 13965 ump = softdep_bp_to_mp(bp); 13966 if (ump == NULL) 13967 return (0); 13968 retval = 0; 13969 ACQUIRE_LOCK(ump); 13970 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13971 switch (wk->wk_type) { 13972 13973 case D_INODEDEP: 13974 inodedep = WK_INODEDEP(wk); 13975 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13976 /* bitmap allocation dependency */ 13977 retval += 1; 13978 if (!wantcount) 13979 goto out; 13980 } 13981 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13982 /* direct block pointer dependency */ 13983 retval += 1; 13984 if (!wantcount) 13985 goto out; 13986 } 13987 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13988 /* direct block pointer dependency */ 13989 retval += 1; 13990 if (!wantcount) 13991 goto out; 13992 } 13993 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 13994 /* Add reference dependency. */ 13995 retval += 1; 13996 if (!wantcount) 13997 goto out; 13998 } 13999 continue; 14000 14001 case D_INDIRDEP: 14002 indirdep = WK_INDIRDEP(wk); 14003 14004 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 14005 /* indirect truncation dependency */ 14006 retval += 1; 14007 if (!wantcount) 14008 goto out; 14009 } 14010 14011 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 14012 /* indirect block pointer dependency */ 14013 retval += 1; 14014 if (!wantcount) 14015 goto out; 14016 } 14017 continue; 14018 14019 case D_PAGEDEP: 14020 pagedep = WK_PAGEDEP(wk); 14021 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 14022 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 14023 /* Journal remove ref dependency. */ 14024 retval += 1; 14025 if (!wantcount) 14026 goto out; 14027 } 14028 } 14029 for (i = 0; i < DAHASHSZ; i++) { 14030 14031 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 14032 /* directory entry dependency */ 14033 retval += 1; 14034 if (!wantcount) 14035 goto out; 14036 } 14037 } 14038 continue; 14039 14040 case D_BMSAFEMAP: 14041 bmsafemap = WK_BMSAFEMAP(wk); 14042 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 14043 /* Add reference dependency. */ 14044 retval += 1; 14045 if (!wantcount) 14046 goto out; 14047 } 14048 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 14049 /* Allocate block dependency. */ 14050 retval += 1; 14051 if (!wantcount) 14052 goto out; 14053 } 14054 continue; 14055 14056 case D_FREEBLKS: 14057 freeblks = WK_FREEBLKS(wk); 14058 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 14059 /* Freeblk journal dependency. */ 14060 retval += 1; 14061 if (!wantcount) 14062 goto out; 14063 } 14064 continue; 14065 14066 case D_ALLOCDIRECT: 14067 case D_ALLOCINDIR: 14068 newblk = WK_NEWBLK(wk); 14069 if (newblk->nb_jnewblk) { 14070 /* Journal allocate dependency. */ 14071 retval += 1; 14072 if (!wantcount) 14073 goto out; 14074 } 14075 continue; 14076 14077 case D_MKDIR: 14078 mkdir = WK_MKDIR(wk); 14079 if (mkdir->md_jaddref) { 14080 /* Journal reference dependency. */ 14081 retval += 1; 14082 if (!wantcount) 14083 goto out; 14084 } 14085 continue; 14086 14087 case D_FREEWORK: 14088 case D_FREEDEP: 14089 case D_JSEGDEP: 14090 case D_JSEG: 14091 case D_SBDEP: 14092 /* never a dependency on these blocks */ 14093 continue; 14094 14095 default: 14096 panic("softdep_count_dependencies: Unexpected type %s", 14097 TYPENAME(wk->wk_type)); 14098 /* NOTREACHED */ 14099 } 14100 } 14101 out: 14102 FREE_LOCK(ump); 14103 return (retval); 14104 } 14105 14106 /* 14107 * Acquire exclusive access to a buffer. 14108 * Must be called with a locked mtx parameter. 14109 * Return acquired buffer or NULL on failure. 14110 */ 14111 static struct buf * 14112 getdirtybuf(bp, lock, waitfor) 14113 struct buf *bp; 14114 struct rwlock *lock; 14115 int waitfor; 14116 { 14117 int error; 14118 14119 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 14120 if (waitfor != MNT_WAIT) 14121 return (NULL); 14122 error = BUF_LOCK(bp, 14123 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 14124 /* 14125 * Even if we successfully acquire bp here, we have dropped 14126 * lock, which may violates our guarantee. 14127 */ 14128 if (error == 0) 14129 BUF_UNLOCK(bp); 14130 else if (error != ENOLCK) 14131 panic("getdirtybuf: inconsistent lock: %d", error); 14132 rw_wlock(lock); 14133 return (NULL); 14134 } 14135 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14136 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 14137 rw_wunlock(lock); 14138 BO_LOCK(bp->b_bufobj); 14139 BUF_UNLOCK(bp); 14140 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14141 bp->b_vflags |= BV_BKGRDWAIT; 14142 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 14143 PRIBIO | PDROP, "getbuf", 0); 14144 } else 14145 BO_UNLOCK(bp->b_bufobj); 14146 rw_wlock(lock); 14147 return (NULL); 14148 } 14149 BUF_UNLOCK(bp); 14150 if (waitfor != MNT_WAIT) 14151 return (NULL); 14152 #ifdef DEBUG_VFS_LOCKS 14153 if (bp->b_vp->v_type != VCHR) 14154 ASSERT_BO_WLOCKED(bp->b_bufobj); 14155 #endif 14156 bp->b_vflags |= BV_BKGRDWAIT; 14157 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 14158 return (NULL); 14159 } 14160 if ((bp->b_flags & B_DELWRI) == 0) { 14161 BUF_UNLOCK(bp); 14162 return (NULL); 14163 } 14164 bremfree(bp); 14165 return (bp); 14166 } 14167 14168 14169 /* 14170 * Check if it is safe to suspend the file system now. On entry, 14171 * the vnode interlock for devvp should be held. Return 0 with 14172 * the mount interlock held if the file system can be suspended now, 14173 * otherwise return EAGAIN with the mount interlock held. 14174 */ 14175 int 14176 softdep_check_suspend(struct mount *mp, 14177 struct vnode *devvp, 14178 int softdep_depcnt, 14179 int softdep_accdepcnt, 14180 int secondary_writes, 14181 int secondary_accwrites) 14182 { 14183 struct bufobj *bo; 14184 struct ufsmount *ump; 14185 struct inodedep *inodedep; 14186 int error, unlinked; 14187 14188 bo = &devvp->v_bufobj; 14189 ASSERT_BO_WLOCKED(bo); 14190 14191 /* 14192 * If we are not running with soft updates, then we need only 14193 * deal with secondary writes as we try to suspend. 14194 */ 14195 if (MOUNTEDSOFTDEP(mp) == 0) { 14196 MNT_ILOCK(mp); 14197 while (mp->mnt_secondary_writes != 0) { 14198 BO_UNLOCK(bo); 14199 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 14200 (PUSER - 1) | PDROP, "secwr", 0); 14201 BO_LOCK(bo); 14202 MNT_ILOCK(mp); 14203 } 14204 14205 /* 14206 * Reasons for needing more work before suspend: 14207 * - Dirty buffers on devvp. 14208 * - Secondary writes occurred after start of vnode sync loop 14209 */ 14210 error = 0; 14211 if (bo->bo_numoutput > 0 || 14212 bo->bo_dirty.bv_cnt > 0 || 14213 secondary_writes != 0 || 14214 mp->mnt_secondary_writes != 0 || 14215 secondary_accwrites != mp->mnt_secondary_accwrites) 14216 error = EAGAIN; 14217 BO_UNLOCK(bo); 14218 return (error); 14219 } 14220 14221 /* 14222 * If we are running with soft updates, then we need to coordinate 14223 * with them as we try to suspend. 14224 */ 14225 ump = VFSTOUFS(mp); 14226 for (;;) { 14227 if (!TRY_ACQUIRE_LOCK(ump)) { 14228 BO_UNLOCK(bo); 14229 ACQUIRE_LOCK(ump); 14230 FREE_LOCK(ump); 14231 BO_LOCK(bo); 14232 continue; 14233 } 14234 MNT_ILOCK(mp); 14235 if (mp->mnt_secondary_writes != 0) { 14236 FREE_LOCK(ump); 14237 BO_UNLOCK(bo); 14238 msleep(&mp->mnt_secondary_writes, 14239 MNT_MTX(mp), 14240 (PUSER - 1) | PDROP, "secwr", 0); 14241 BO_LOCK(bo); 14242 continue; 14243 } 14244 break; 14245 } 14246 14247 unlinked = 0; 14248 if (MOUNTEDSUJ(mp)) { 14249 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14250 inodedep != NULL; 14251 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14252 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14253 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14254 UNLINKONLIST) || 14255 !check_inodedep_free(inodedep)) 14256 continue; 14257 unlinked++; 14258 } 14259 } 14260 14261 /* 14262 * Reasons for needing more work before suspend: 14263 * - Dirty buffers on devvp. 14264 * - Softdep activity occurred after start of vnode sync loop 14265 * - Secondary writes occurred after start of vnode sync loop 14266 */ 14267 error = 0; 14268 if (bo->bo_numoutput > 0 || 14269 bo->bo_dirty.bv_cnt > 0 || 14270 softdep_depcnt != unlinked || 14271 ump->softdep_deps != unlinked || 14272 softdep_accdepcnt != ump->softdep_accdeps || 14273 secondary_writes != 0 || 14274 mp->mnt_secondary_writes != 0 || 14275 secondary_accwrites != mp->mnt_secondary_accwrites) 14276 error = EAGAIN; 14277 FREE_LOCK(ump); 14278 BO_UNLOCK(bo); 14279 return (error); 14280 } 14281 14282 14283 /* 14284 * Get the number of dependency structures for the file system, both 14285 * the current number and the total number allocated. These will 14286 * later be used to detect that softdep processing has occurred. 14287 */ 14288 void 14289 softdep_get_depcounts(struct mount *mp, 14290 int *softdep_depsp, 14291 int *softdep_accdepsp) 14292 { 14293 struct ufsmount *ump; 14294 14295 if (MOUNTEDSOFTDEP(mp) == 0) { 14296 *softdep_depsp = 0; 14297 *softdep_accdepsp = 0; 14298 return; 14299 } 14300 ump = VFSTOUFS(mp); 14301 ACQUIRE_LOCK(ump); 14302 *softdep_depsp = ump->softdep_deps; 14303 *softdep_accdepsp = ump->softdep_accdeps; 14304 FREE_LOCK(ump); 14305 } 14306 14307 /* 14308 * Wait for pending output on a vnode to complete. 14309 */ 14310 static void 14311 drain_output(vp) 14312 struct vnode *vp; 14313 { 14314 14315 ASSERT_VOP_LOCKED(vp, "drain_output"); 14316 (void)bufobj_wwait(&vp->v_bufobj, 0, 0); 14317 } 14318 14319 /* 14320 * Called whenever a buffer that is being invalidated or reallocated 14321 * contains dependencies. This should only happen if an I/O error has 14322 * occurred. The routine is called with the buffer locked. 14323 */ 14324 static void 14325 softdep_deallocate_dependencies(bp) 14326 struct buf *bp; 14327 { 14328 14329 if ((bp->b_ioflags & BIO_ERROR) == 0) 14330 panic("softdep_deallocate_dependencies: dangling deps"); 14331 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14332 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14333 else 14334 printf("softdep_deallocate_dependencies: " 14335 "got error %d while accessing filesystem\n", bp->b_error); 14336 if (bp->b_error != ENXIO) 14337 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14338 } 14339 14340 /* 14341 * Function to handle asynchronous write errors in the filesystem. 14342 */ 14343 static void 14344 softdep_error(func, error) 14345 char *func; 14346 int error; 14347 { 14348 14349 /* XXX should do something better! */ 14350 printf("%s: got error %d while accessing filesystem\n", func, error); 14351 } 14352 14353 #ifdef DDB 14354 14355 static void 14356 inodedep_print(struct inodedep *inodedep, int verbose) 14357 { 14358 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %jd nlink %jd" 14359 " saveino %p\n", 14360 inodedep, inodedep->id_fs, inodedep->id_state, 14361 (intmax_t)inodedep->id_ino, 14362 (intmax_t)fsbtodb(inodedep->id_fs, 14363 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14364 (intmax_t)inodedep->id_nlinkdelta, 14365 (intmax_t)inodedep->id_savednlink, 14366 inodedep->id_savedino1); 14367 14368 if (verbose == 0) 14369 return; 14370 14371 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 14372 "mkdiradd %p\n", 14373 LIST_FIRST(&inodedep->id_pendinghd), 14374 LIST_FIRST(&inodedep->id_bufwait), 14375 LIST_FIRST(&inodedep->id_inowait), 14376 TAILQ_FIRST(&inodedep->id_inoreflst), 14377 inodedep->id_mkdiradd); 14378 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 14379 TAILQ_FIRST(&inodedep->id_inoupdt), 14380 TAILQ_FIRST(&inodedep->id_newinoupdt), 14381 TAILQ_FIRST(&inodedep->id_extupdt), 14382 TAILQ_FIRST(&inodedep->id_newextupdt)); 14383 } 14384 14385 DB_SHOW_COMMAND(inodedep, db_show_inodedep) 14386 { 14387 14388 if (have_addr == 0) { 14389 db_printf("Address required\n"); 14390 return; 14391 } 14392 inodedep_print((struct inodedep*)addr, 1); 14393 } 14394 14395 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 14396 { 14397 struct inodedep_hashhead *inodedephd; 14398 struct inodedep *inodedep; 14399 struct ufsmount *ump; 14400 int cnt; 14401 14402 if (have_addr == 0) { 14403 db_printf("Address required\n"); 14404 return; 14405 } 14406 ump = (struct ufsmount *)addr; 14407 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14408 inodedephd = &ump->inodedep_hashtbl[cnt]; 14409 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14410 inodedep_print(inodedep, 0); 14411 } 14412 } 14413 } 14414 14415 DB_SHOW_COMMAND(worklist, db_show_worklist) 14416 { 14417 struct worklist *wk; 14418 14419 if (have_addr == 0) { 14420 db_printf("Address required\n"); 14421 return; 14422 } 14423 wk = (struct worklist *)addr; 14424 printf("worklist: %p type %s state 0x%X\n", 14425 wk, TYPENAME(wk->wk_type), wk->wk_state); 14426 } 14427 14428 DB_SHOW_COMMAND(workhead, db_show_workhead) 14429 { 14430 struct workhead *wkhd; 14431 struct worklist *wk; 14432 int i; 14433 14434 if (have_addr == 0) { 14435 db_printf("Address required\n"); 14436 return; 14437 } 14438 wkhd = (struct workhead *)addr; 14439 wk = LIST_FIRST(wkhd); 14440 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 14441 db_printf("worklist: %p type %s state 0x%X", 14442 wk, TYPENAME(wk->wk_type), wk->wk_state); 14443 if (i == 100) 14444 db_printf("workhead overflow"); 14445 printf("\n"); 14446 } 14447 14448 14449 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 14450 { 14451 struct mkdirlist *mkdirlisthd; 14452 struct jaddref *jaddref; 14453 struct diradd *diradd; 14454 struct mkdir *mkdir; 14455 14456 if (have_addr == 0) { 14457 db_printf("Address required\n"); 14458 return; 14459 } 14460 mkdirlisthd = (struct mkdirlist *)addr; 14461 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14462 diradd = mkdir->md_diradd; 14463 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 14464 mkdir, mkdir->md_state, diradd, diradd->da_state); 14465 if ((jaddref = mkdir->md_jaddref) != NULL) 14466 db_printf(" jaddref %p jaddref state 0x%X", 14467 jaddref, jaddref->ja_state); 14468 db_printf("\n"); 14469 } 14470 } 14471 14472 /* exported to ffs_vfsops.c */ 14473 extern void db_print_ffs(struct ufsmount *ump); 14474 void 14475 db_print_ffs(struct ufsmount *ump) 14476 { 14477 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n", 14478 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname, 14479 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist, 14480 ump->softdep_deps, ump->softdep_req); 14481 } 14482 14483 #endif /* DDB */ 14484 14485 #endif /* SOFTUPDATES */ 14486