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 #include <sys/param.h> 52 #include <sys/kernel.h> 53 #include <sys/systm.h> 54 #include <sys/bio.h> 55 #include <sys/buf.h> 56 #include <sys/kdb.h> 57 #include <sys/kthread.h> 58 #include <sys/ktr.h> 59 #include <sys/limits.h> 60 #include <sys/lock.h> 61 #include <sys/malloc.h> 62 #include <sys/mount.h> 63 #include <sys/mutex.h> 64 #include <sys/namei.h> 65 #include <sys/priv.h> 66 #include <sys/proc.h> 67 #include <sys/racct.h> 68 #include <sys/rwlock.h> 69 #include <sys/stat.h> 70 #include <sys/sysctl.h> 71 #include <sys/syslog.h> 72 #include <sys/vnode.h> 73 #include <sys/conf.h> 74 75 #include <ufs/ufs/dir.h> 76 #include <ufs/ufs/extattr.h> 77 #include <ufs/ufs/quota.h> 78 #include <ufs/ufs/inode.h> 79 #include <ufs/ufs/ufsmount.h> 80 #include <ufs/ffs/fs.h> 81 #include <ufs/ffs/softdep.h> 82 #include <ufs/ffs/ffs_extern.h> 83 #include <ufs/ufs/ufs_extern.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_extern.h> 87 #include <vm/vm_object.h> 88 89 #include <geom/geom.h> 90 91 #include <ddb/ddb.h> 92 93 #define KTR_SUJ 0 /* Define to KTR_SPARE. */ 94 95 #ifndef SOFTUPDATES 96 97 int 98 softdep_flushfiles(oldmnt, flags, td) 99 struct mount *oldmnt; 100 int flags; 101 struct thread *td; 102 { 103 104 panic("softdep_flushfiles called"); 105 } 106 107 int 108 softdep_mount(devvp, mp, fs, cred) 109 struct vnode *devvp; 110 struct mount *mp; 111 struct fs *fs; 112 struct ucred *cred; 113 { 114 115 return (0); 116 } 117 118 void 119 softdep_initialize() 120 { 121 122 return; 123 } 124 125 void 126 softdep_uninitialize() 127 { 128 129 return; 130 } 131 132 void 133 softdep_unmount(mp) 134 struct mount *mp; 135 { 136 137 panic("softdep_unmount called"); 138 } 139 140 void 141 softdep_setup_sbupdate(ump, fs, bp) 142 struct ufsmount *ump; 143 struct fs *fs; 144 struct buf *bp; 145 { 146 147 panic("softdep_setup_sbupdate called"); 148 } 149 150 void 151 softdep_setup_inomapdep(bp, ip, newinum, mode) 152 struct buf *bp; 153 struct inode *ip; 154 ino_t newinum; 155 int mode; 156 { 157 158 panic("softdep_setup_inomapdep called"); 159 } 160 161 void 162 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 163 struct buf *bp; 164 struct mount *mp; 165 ufs2_daddr_t newblkno; 166 int frags; 167 int oldfrags; 168 { 169 170 panic("softdep_setup_blkmapdep called"); 171 } 172 173 void 174 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 175 struct inode *ip; 176 ufs_lbn_t lbn; 177 ufs2_daddr_t newblkno; 178 ufs2_daddr_t oldblkno; 179 long newsize; 180 long oldsize; 181 struct buf *bp; 182 { 183 184 panic("softdep_setup_allocdirect called"); 185 } 186 187 void 188 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 189 struct inode *ip; 190 ufs_lbn_t lbn; 191 ufs2_daddr_t newblkno; 192 ufs2_daddr_t oldblkno; 193 long newsize; 194 long oldsize; 195 struct buf *bp; 196 { 197 198 panic("softdep_setup_allocext called"); 199 } 200 201 void 202 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 203 struct inode *ip; 204 ufs_lbn_t lbn; 205 struct buf *bp; 206 int ptrno; 207 ufs2_daddr_t newblkno; 208 ufs2_daddr_t oldblkno; 209 struct buf *nbp; 210 { 211 212 panic("softdep_setup_allocindir_page called"); 213 } 214 215 void 216 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 217 struct buf *nbp; 218 struct inode *ip; 219 struct buf *bp; 220 int ptrno; 221 ufs2_daddr_t newblkno; 222 { 223 224 panic("softdep_setup_allocindir_meta called"); 225 } 226 227 void 228 softdep_journal_freeblocks(ip, cred, length, flags) 229 struct inode *ip; 230 struct ucred *cred; 231 off_t length; 232 int flags; 233 { 234 235 panic("softdep_journal_freeblocks called"); 236 } 237 238 void 239 softdep_journal_fsync(ip) 240 struct inode *ip; 241 { 242 243 panic("softdep_journal_fsync called"); 244 } 245 246 void 247 softdep_setup_freeblocks(ip, length, flags) 248 struct inode *ip; 249 off_t length; 250 int flags; 251 { 252 253 panic("softdep_setup_freeblocks called"); 254 } 255 256 void 257 softdep_freefile(pvp, ino, mode) 258 struct vnode *pvp; 259 ino_t ino; 260 int mode; 261 { 262 263 panic("softdep_freefile called"); 264 } 265 266 int 267 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 268 struct buf *bp; 269 struct inode *dp; 270 off_t diroffset; 271 ino_t newinum; 272 struct buf *newdirbp; 273 int isnewblk; 274 { 275 276 panic("softdep_setup_directory_add called"); 277 } 278 279 void 280 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 281 struct buf *bp; 282 struct inode *dp; 283 caddr_t base; 284 caddr_t oldloc; 285 caddr_t newloc; 286 int entrysize; 287 { 288 289 panic("softdep_change_directoryentry_offset called"); 290 } 291 292 void 293 softdep_setup_remove(bp, dp, ip, isrmdir) 294 struct buf *bp; 295 struct inode *dp; 296 struct inode *ip; 297 int isrmdir; 298 { 299 300 panic("softdep_setup_remove called"); 301 } 302 303 void 304 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 305 struct buf *bp; 306 struct inode *dp; 307 struct inode *ip; 308 ino_t newinum; 309 int isrmdir; 310 { 311 312 panic("softdep_setup_directory_change called"); 313 } 314 315 void 316 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 317 struct mount *mp; 318 struct buf *bp; 319 ufs2_daddr_t blkno; 320 int frags; 321 struct workhead *wkhd; 322 { 323 324 panic("%s called", __FUNCTION__); 325 } 326 327 void 328 softdep_setup_inofree(mp, bp, ino, wkhd) 329 struct mount *mp; 330 struct buf *bp; 331 ino_t ino; 332 struct workhead *wkhd; 333 { 334 335 panic("%s called", __FUNCTION__); 336 } 337 338 void 339 softdep_setup_unlink(dp, ip) 340 struct inode *dp; 341 struct inode *ip; 342 { 343 344 panic("%s called", __FUNCTION__); 345 } 346 347 void 348 softdep_setup_link(dp, ip) 349 struct inode *dp; 350 struct inode *ip; 351 { 352 353 panic("%s called", __FUNCTION__); 354 } 355 356 void 357 softdep_revert_link(dp, ip) 358 struct inode *dp; 359 struct inode *ip; 360 { 361 362 panic("%s called", __FUNCTION__); 363 } 364 365 void 366 softdep_setup_rmdir(dp, ip) 367 struct inode *dp; 368 struct inode *ip; 369 { 370 371 panic("%s called", __FUNCTION__); 372 } 373 374 void 375 softdep_revert_rmdir(dp, ip) 376 struct inode *dp; 377 struct inode *ip; 378 { 379 380 panic("%s called", __FUNCTION__); 381 } 382 383 void 384 softdep_setup_create(dp, ip) 385 struct inode *dp; 386 struct inode *ip; 387 { 388 389 panic("%s called", __FUNCTION__); 390 } 391 392 void 393 softdep_revert_create(dp, ip) 394 struct inode *dp; 395 struct inode *ip; 396 { 397 398 panic("%s called", __FUNCTION__); 399 } 400 401 void 402 softdep_setup_mkdir(dp, ip) 403 struct inode *dp; 404 struct inode *ip; 405 { 406 407 panic("%s called", __FUNCTION__); 408 } 409 410 void 411 softdep_revert_mkdir(dp, ip) 412 struct inode *dp; 413 struct inode *ip; 414 { 415 416 panic("%s called", __FUNCTION__); 417 } 418 419 void 420 softdep_setup_dotdot_link(dp, ip) 421 struct inode *dp; 422 struct inode *ip; 423 { 424 425 panic("%s called", __FUNCTION__); 426 } 427 428 int 429 softdep_prealloc(vp, waitok) 430 struct vnode *vp; 431 int waitok; 432 { 433 434 panic("%s called", __FUNCTION__); 435 } 436 437 int 438 softdep_journal_lookup(mp, vpp) 439 struct mount *mp; 440 struct vnode **vpp; 441 { 442 443 return (ENOENT); 444 } 445 446 void 447 softdep_change_linkcnt(ip) 448 struct inode *ip; 449 { 450 451 panic("softdep_change_linkcnt called"); 452 } 453 454 void 455 softdep_load_inodeblock(ip) 456 struct inode *ip; 457 { 458 459 panic("softdep_load_inodeblock called"); 460 } 461 462 void 463 softdep_update_inodeblock(ip, bp, waitfor) 464 struct inode *ip; 465 struct buf *bp; 466 int waitfor; 467 { 468 469 panic("softdep_update_inodeblock called"); 470 } 471 472 int 473 softdep_fsync(vp) 474 struct vnode *vp; /* the "in_core" copy of the inode */ 475 { 476 477 return (0); 478 } 479 480 void 481 softdep_fsync_mountdev(vp) 482 struct vnode *vp; 483 { 484 485 return; 486 } 487 488 int 489 softdep_flushworklist(oldmnt, countp, td) 490 struct mount *oldmnt; 491 int *countp; 492 struct thread *td; 493 { 494 495 *countp = 0; 496 return (0); 497 } 498 499 int 500 softdep_sync_metadata(struct vnode *vp) 501 { 502 503 panic("softdep_sync_metadata called"); 504 } 505 506 int 507 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 508 { 509 510 panic("softdep_sync_buf called"); 511 } 512 513 int 514 softdep_slowdown(vp) 515 struct vnode *vp; 516 { 517 518 panic("softdep_slowdown called"); 519 } 520 521 int 522 softdep_request_cleanup(fs, vp, cred, resource) 523 struct fs *fs; 524 struct vnode *vp; 525 struct ucred *cred; 526 int resource; 527 { 528 529 return (0); 530 } 531 532 int 533 softdep_check_suspend(struct mount *mp, 534 struct vnode *devvp, 535 int softdep_depcnt, 536 int softdep_accdepcnt, 537 int secondary_writes, 538 int secondary_accwrites) 539 { 540 struct bufobj *bo; 541 int error; 542 543 (void) softdep_depcnt, 544 (void) softdep_accdepcnt; 545 546 bo = &devvp->v_bufobj; 547 ASSERT_BO_WLOCKED(bo); 548 549 MNT_ILOCK(mp); 550 while (mp->mnt_secondary_writes != 0) { 551 BO_UNLOCK(bo); 552 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 553 (PUSER - 1) | PDROP, "secwr", 0); 554 BO_LOCK(bo); 555 MNT_ILOCK(mp); 556 } 557 558 /* 559 * Reasons for needing more work before suspend: 560 * - Dirty buffers on devvp. 561 * - Secondary writes occurred after start of vnode sync loop 562 */ 563 error = 0; 564 if (bo->bo_numoutput > 0 || 565 bo->bo_dirty.bv_cnt > 0 || 566 secondary_writes != 0 || 567 mp->mnt_secondary_writes != 0 || 568 secondary_accwrites != mp->mnt_secondary_accwrites) 569 error = EAGAIN; 570 BO_UNLOCK(bo); 571 return (error); 572 } 573 574 void 575 softdep_get_depcounts(struct mount *mp, 576 int *softdepactivep, 577 int *softdepactiveaccp) 578 { 579 (void) mp; 580 *softdepactivep = 0; 581 *softdepactiveaccp = 0; 582 } 583 584 void 585 softdep_buf_append(bp, wkhd) 586 struct buf *bp; 587 struct workhead *wkhd; 588 { 589 590 panic("softdep_buf_appendwork called"); 591 } 592 593 void 594 softdep_inode_append(ip, cred, wkhd) 595 struct inode *ip; 596 struct ucred *cred; 597 struct workhead *wkhd; 598 { 599 600 panic("softdep_inode_appendwork called"); 601 } 602 603 void 604 softdep_freework(wkhd) 605 struct workhead *wkhd; 606 { 607 608 panic("softdep_freework called"); 609 } 610 611 #else 612 613 FEATURE(softupdates, "FFS soft-updates support"); 614 615 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 616 "soft updates stats"); 617 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 618 "total dependencies allocated"); 619 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 620 "high use dependencies allocated"); 621 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 622 "current dependencies allocated"); 623 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 624 "current dependencies written"); 625 626 unsigned long dep_current[D_LAST + 1]; 627 unsigned long dep_highuse[D_LAST + 1]; 628 unsigned long dep_total[D_LAST + 1]; 629 unsigned long dep_write[D_LAST + 1]; 630 631 #define SOFTDEP_TYPE(type, str, long) \ 632 static MALLOC_DEFINE(M_ ## type, #str, long); \ 633 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 634 &dep_total[D_ ## type], 0, ""); \ 635 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 636 &dep_current[D_ ## type], 0, ""); \ 637 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 638 &dep_highuse[D_ ## type], 0, ""); \ 639 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 640 &dep_write[D_ ## type], 0, ""); 641 642 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 643 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 644 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 645 "Block or frag allocated from cyl group map"); 646 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 647 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 648 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 649 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 650 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 651 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 652 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 653 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 654 SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 655 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 656 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 657 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 658 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 659 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 660 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 661 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 662 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 663 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 664 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 665 SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 666 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 667 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 668 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 669 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 670 671 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 672 673 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 674 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 675 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 676 677 #define M_SOFTDEP_FLAGS (M_WAITOK) 678 679 /* 680 * translate from workitem type to memory type 681 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 682 */ 683 static struct malloc_type *memtype[] = { 684 NULL, 685 M_PAGEDEP, 686 M_INODEDEP, 687 M_BMSAFEMAP, 688 M_NEWBLK, 689 M_ALLOCDIRECT, 690 M_INDIRDEP, 691 M_ALLOCINDIR, 692 M_FREEFRAG, 693 M_FREEBLKS, 694 M_FREEFILE, 695 M_DIRADD, 696 M_MKDIR, 697 M_DIRREM, 698 M_NEWDIRBLK, 699 M_FREEWORK, 700 M_FREEDEP, 701 M_JADDREF, 702 M_JREMREF, 703 M_JMVREF, 704 M_JNEWBLK, 705 M_JFREEBLK, 706 M_JFREEFRAG, 707 M_JSEG, 708 M_JSEGDEP, 709 M_SBDEP, 710 M_JTRUNC, 711 M_JFSYNC, 712 M_SENTINEL 713 }; 714 715 #define DtoM(type) (memtype[type]) 716 717 /* 718 * Names of malloc types. 719 */ 720 #define TYPENAME(type) \ 721 ((unsigned)(type) <= D_LAST && (unsigned)(type) >= D_FIRST ? \ 722 memtype[type]->ks_shortdesc : "???") 723 /* 724 * End system adaptation definitions. 725 */ 726 727 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 728 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 729 730 /* 731 * Internal function prototypes. 732 */ 733 static void check_clear_deps(struct mount *); 734 static void softdep_error(char *, int); 735 static int softdep_process_worklist(struct mount *, int); 736 static int softdep_waitidle(struct mount *, int); 737 static void drain_output(struct vnode *); 738 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 739 static int check_inodedep_free(struct inodedep *); 740 static void clear_remove(struct mount *); 741 static void clear_inodedeps(struct mount *); 742 static void unlinked_inodedep(struct mount *, struct inodedep *); 743 static void clear_unlinked_inodedep(struct inodedep *); 744 static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 745 static int flush_pagedep_deps(struct vnode *, struct mount *, 746 struct diraddhd *); 747 static int free_pagedep(struct pagedep *); 748 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 749 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 750 static int flush_deplist(struct allocdirectlst *, int, int *); 751 static int sync_cgs(struct mount *, int); 752 static int handle_written_filepage(struct pagedep *, struct buf *, int); 753 static int handle_written_sbdep(struct sbdep *, struct buf *); 754 static void initiate_write_sbdep(struct sbdep *); 755 static void diradd_inode_written(struct diradd *, struct inodedep *); 756 static int handle_written_indirdep(struct indirdep *, struct buf *, 757 struct buf**, int); 758 static int handle_written_inodeblock(struct inodedep *, struct buf *, int); 759 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 760 uint8_t *); 761 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int); 762 static void handle_written_jaddref(struct jaddref *); 763 static void handle_written_jremref(struct jremref *); 764 static void handle_written_jseg(struct jseg *, struct buf *); 765 static void handle_written_jnewblk(struct jnewblk *); 766 static void handle_written_jblkdep(struct jblkdep *); 767 static void handle_written_jfreefrag(struct jfreefrag *); 768 static void complete_jseg(struct jseg *); 769 static void complete_jsegs(struct jseg *); 770 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 771 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 772 static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 773 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 774 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 775 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 776 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 777 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 778 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 779 static inline void inoref_write(struct inoref *, struct jseg *, 780 struct jrefrec *); 781 static void handle_allocdirect_partdone(struct allocdirect *, 782 struct workhead *); 783 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 784 struct workhead *); 785 static void indirdep_complete(struct indirdep *); 786 static int indirblk_lookup(struct mount *, ufs2_daddr_t); 787 static void indirblk_insert(struct freework *); 788 static void indirblk_remove(struct freework *); 789 static void handle_allocindir_partdone(struct allocindir *); 790 static void initiate_write_filepage(struct pagedep *, struct buf *); 791 static void initiate_write_indirdep(struct indirdep*, struct buf *); 792 static void handle_written_mkdir(struct mkdir *, int); 793 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 794 uint8_t *); 795 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 796 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 797 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 798 static void handle_workitem_freefile(struct freefile *); 799 static int handle_workitem_remove(struct dirrem *, int); 800 static struct dirrem *newdirrem(struct buf *, struct inode *, 801 struct inode *, int, struct dirrem **); 802 static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 803 struct buf *); 804 static void cancel_indirdep(struct indirdep *, struct buf *, 805 struct freeblks *); 806 static void free_indirdep(struct indirdep *); 807 static void free_diradd(struct diradd *, struct workhead *); 808 static void merge_diradd(struct inodedep *, struct diradd *); 809 static void complete_diradd(struct diradd *); 810 static struct diradd *diradd_lookup(struct pagedep *, int); 811 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 812 struct jremref *); 813 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 814 struct jremref *); 815 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 816 struct jremref *, struct jremref *); 817 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 818 struct jremref *); 819 static void cancel_allocindir(struct allocindir *, struct buf *bp, 820 struct freeblks *, int); 821 static int setup_trunc_indir(struct freeblks *, struct inode *, 822 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 823 static void complete_trunc_indir(struct freework *); 824 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 825 int); 826 static void complete_mkdir(struct mkdir *); 827 static void free_newdirblk(struct newdirblk *); 828 static void free_jremref(struct jremref *); 829 static void free_jaddref(struct jaddref *); 830 static void free_jsegdep(struct jsegdep *); 831 static void free_jsegs(struct jblocks *); 832 static void rele_jseg(struct jseg *); 833 static void free_jseg(struct jseg *, struct jblocks *); 834 static void free_jnewblk(struct jnewblk *); 835 static void free_jblkdep(struct jblkdep *); 836 static void free_jfreefrag(struct jfreefrag *); 837 static void free_freedep(struct freedep *); 838 static void journal_jremref(struct dirrem *, struct jremref *, 839 struct inodedep *); 840 static void cancel_jnewblk(struct jnewblk *, struct workhead *); 841 static int cancel_jaddref(struct jaddref *, struct inodedep *, 842 struct workhead *); 843 static void cancel_jfreefrag(struct jfreefrag *); 844 static inline void setup_freedirect(struct freeblks *, struct inode *, 845 int, int); 846 static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 847 static inline void setup_freeindir(struct freeblks *, struct inode *, int, 848 ufs_lbn_t, int); 849 static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 850 static void freeblks_free(struct ufsmount *, struct freeblks *, int); 851 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 852 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 853 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 854 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 855 int, int); 856 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 857 static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 858 static int deallocate_dependencies(struct buf *, struct freeblks *, int); 859 static void newblk_freefrag(struct newblk*); 860 static void free_newblk(struct newblk *); 861 static void cancel_allocdirect(struct allocdirectlst *, 862 struct allocdirect *, struct freeblks *); 863 static int check_inode_unwritten(struct inodedep *); 864 static int free_inodedep(struct inodedep *); 865 static void freework_freeblock(struct freework *, u_long); 866 static void freework_enqueue(struct freework *); 867 static int handle_workitem_freeblocks(struct freeblks *, int); 868 static int handle_complete_freeblocks(struct freeblks *, int); 869 static void handle_workitem_indirblk(struct freework *); 870 static void handle_written_freework(struct freework *); 871 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 872 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 873 struct workhead *); 874 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 875 struct inodedep *, struct allocindir *, ufs_lbn_t); 876 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 877 ufs2_daddr_t, ufs_lbn_t); 878 static void handle_workitem_freefrag(struct freefrag *); 879 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 880 ufs_lbn_t, u_long); 881 static void allocdirect_merge(struct allocdirectlst *, 882 struct allocdirect *, struct allocdirect *); 883 static struct freefrag *allocindir_merge(struct allocindir *, 884 struct allocindir *); 885 static int bmsafemap_find(struct bmsafemap_hashhead *, int, 886 struct bmsafemap **); 887 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 888 int cg, struct bmsafemap *); 889 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 890 struct newblk **); 891 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 892 static int inodedep_find(struct inodedep_hashhead *, ino_t, 893 struct inodedep **); 894 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 895 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 896 int, struct pagedep **); 897 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 898 struct pagedep **); 899 static void pause_timer(void *); 900 static int request_cleanup(struct mount *, int); 901 static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *); 902 static void schedule_cleanup(struct mount *); 903 static void softdep_ast_cleanup_proc(struct thread *); 904 static struct ufsmount *softdep_bp_to_mp(struct buf *bp); 905 static int process_worklist_item(struct mount *, int, int); 906 static void process_removes(struct vnode *); 907 static void process_truncates(struct vnode *); 908 static void jwork_move(struct workhead *, struct workhead *); 909 static void jwork_insert(struct workhead *, struct jsegdep *); 910 static void add_to_worklist(struct worklist *, int); 911 static void wake_worklist(struct worklist *); 912 static void wait_worklist(struct worklist *, char *); 913 static void remove_from_worklist(struct worklist *); 914 static void softdep_flush(void *); 915 static void softdep_flushjournal(struct mount *); 916 static int softdep_speedup(struct ufsmount *); 917 static void worklist_speedup(struct mount *); 918 static int journal_mount(struct mount *, struct fs *, struct ucred *); 919 static void journal_unmount(struct ufsmount *); 920 static int journal_space(struct ufsmount *, int); 921 static void journal_suspend(struct ufsmount *); 922 static int journal_unsuspend(struct ufsmount *ump); 923 static void softdep_prelink(struct vnode *, struct vnode *); 924 static void add_to_journal(struct worklist *); 925 static void remove_from_journal(struct worklist *); 926 static bool softdep_excess_items(struct ufsmount *, int); 927 static void softdep_process_journal(struct mount *, struct worklist *, int); 928 static struct jremref *newjremref(struct dirrem *, struct inode *, 929 struct inode *ip, off_t, nlink_t); 930 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 931 uint16_t); 932 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 933 uint16_t); 934 static inline struct jsegdep *inoref_jseg(struct inoref *); 935 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 936 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 937 ufs2_daddr_t, int); 938 static void adjust_newfreework(struct freeblks *, int); 939 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 940 static void move_newblock_dep(struct jaddref *, struct inodedep *); 941 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 942 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 943 ufs2_daddr_t, long, ufs_lbn_t); 944 static struct freework *newfreework(struct ufsmount *, struct freeblks *, 945 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 946 static int jwait(struct worklist *, int); 947 static struct inodedep *inodedep_lookup_ip(struct inode *); 948 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 949 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 950 static void handle_jwork(struct workhead *); 951 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 952 struct mkdir **); 953 static struct jblocks *jblocks_create(void); 954 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 955 static void jblocks_free(struct jblocks *, struct mount *, int); 956 static void jblocks_destroy(struct jblocks *); 957 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 958 959 /* 960 * Exported softdep operations. 961 */ 962 static void softdep_disk_io_initiation(struct buf *); 963 static void softdep_disk_write_complete(struct buf *); 964 static void softdep_deallocate_dependencies(struct buf *); 965 static int softdep_count_dependencies(struct buf *bp, int); 966 967 /* 968 * Global lock over all of soft updates. 969 */ 970 static struct mtx lk; 971 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF); 972 973 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 974 #define FREE_GBLLOCK(lk) mtx_unlock(lk) 975 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 976 977 /* 978 * Per-filesystem soft-updates locking. 979 */ 980 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 981 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 982 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 983 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 984 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 985 RA_WLOCKED) 986 987 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 988 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 989 990 /* 991 * Worklist queue management. 992 * These routines require that the lock be held. 993 */ 994 #ifndef /* NOT */ INVARIANTS 995 #define WORKLIST_INSERT(head, item) do { \ 996 (item)->wk_state |= ONWORKLIST; \ 997 LIST_INSERT_HEAD(head, item, wk_list); \ 998 } while (0) 999 #define WORKLIST_REMOVE(item) do { \ 1000 (item)->wk_state &= ~ONWORKLIST; \ 1001 LIST_REMOVE(item, wk_list); \ 1002 } while (0) 1003 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1004 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1005 1006 #else /* INVARIANTS */ 1007 static void worklist_insert(struct workhead *, struct worklist *, int, 1008 const char *, int); 1009 static void worklist_remove(struct worklist *, int, const char *, int); 1010 1011 #define WORKLIST_INSERT(head, item) \ 1012 worklist_insert(head, item, 1, __func__, __LINE__) 1013 #define WORKLIST_INSERT_UNLOCKED(head, item)\ 1014 worklist_insert(head, item, 0, __func__, __LINE__) 1015 #define WORKLIST_REMOVE(item)\ 1016 worklist_remove(item, 1, __func__, __LINE__) 1017 #define WORKLIST_REMOVE_UNLOCKED(item)\ 1018 worklist_remove(item, 0, __func__, __LINE__) 1019 1020 static void 1021 worklist_insert(head, item, locked, func, line) 1022 struct workhead *head; 1023 struct worklist *item; 1024 int locked; 1025 const char *func; 1026 int line; 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 "added in function %s at line %d", 1034 item, TYPENAME(item->wk_type), item->wk_state, 1035 item->wk_func, item->wk_line); 1036 item->wk_state |= ONWORKLIST; 1037 item->wk_func = func; 1038 item->wk_line = line; 1039 LIST_INSERT_HEAD(head, item, wk_list); 1040 } 1041 1042 static void 1043 worklist_remove(item, locked, func, line) 1044 struct worklist *item; 1045 int locked; 1046 const char *func; 1047 int line; 1048 { 1049 1050 if (locked) 1051 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1052 if ((item->wk_state & ONWORKLIST) == 0) 1053 panic("worklist_remove: %p %s(0x%X) not on list, " 1054 "removed in function %s at line %d", 1055 item, TYPENAME(item->wk_type), item->wk_state, 1056 item->wk_func, item->wk_line); 1057 item->wk_state &= ~ONWORKLIST; 1058 item->wk_func = func; 1059 item->wk_line = line; 1060 LIST_REMOVE(item, wk_list); 1061 } 1062 #endif /* INVARIANTS */ 1063 1064 /* 1065 * Merge two jsegdeps keeping only the oldest one as newer references 1066 * can't be discarded until after older references. 1067 */ 1068 static inline struct jsegdep * 1069 jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1070 { 1071 struct jsegdep *swp; 1072 1073 if (two == NULL) 1074 return (one); 1075 1076 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1077 swp = one; 1078 one = two; 1079 two = swp; 1080 } 1081 WORKLIST_REMOVE(&two->jd_list); 1082 free_jsegdep(two); 1083 1084 return (one); 1085 } 1086 1087 /* 1088 * If two freedeps are compatible free one to reduce list size. 1089 */ 1090 static inline struct freedep * 1091 freedep_merge(struct freedep *one, struct freedep *two) 1092 { 1093 if (two == NULL) 1094 return (one); 1095 1096 if (one->fd_freework == two->fd_freework) { 1097 WORKLIST_REMOVE(&two->fd_list); 1098 free_freedep(two); 1099 } 1100 return (one); 1101 } 1102 1103 /* 1104 * Move journal work from one list to another. Duplicate freedeps and 1105 * jsegdeps are coalesced to keep the lists as small as possible. 1106 */ 1107 static void 1108 jwork_move(dst, src) 1109 struct workhead *dst; 1110 struct workhead *src; 1111 { 1112 struct freedep *freedep; 1113 struct jsegdep *jsegdep; 1114 struct worklist *wkn; 1115 struct worklist *wk; 1116 1117 KASSERT(dst != src, 1118 ("jwork_move: dst == src")); 1119 freedep = NULL; 1120 jsegdep = NULL; 1121 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1122 if (wk->wk_type == D_JSEGDEP) 1123 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1124 else if (wk->wk_type == D_FREEDEP) 1125 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1126 } 1127 1128 while ((wk = LIST_FIRST(src)) != NULL) { 1129 WORKLIST_REMOVE(wk); 1130 WORKLIST_INSERT(dst, wk); 1131 if (wk->wk_type == D_JSEGDEP) { 1132 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1133 continue; 1134 } 1135 if (wk->wk_type == D_FREEDEP) 1136 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1137 } 1138 } 1139 1140 static void 1141 jwork_insert(dst, jsegdep) 1142 struct workhead *dst; 1143 struct jsegdep *jsegdep; 1144 { 1145 struct jsegdep *jsegdepn; 1146 struct worklist *wk; 1147 1148 LIST_FOREACH(wk, dst, wk_list) 1149 if (wk->wk_type == D_JSEGDEP) 1150 break; 1151 if (wk == NULL) { 1152 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1153 return; 1154 } 1155 jsegdepn = WK_JSEGDEP(wk); 1156 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1157 WORKLIST_REMOVE(wk); 1158 free_jsegdep(jsegdepn); 1159 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1160 } else 1161 free_jsegdep(jsegdep); 1162 } 1163 1164 /* 1165 * Routines for tracking and managing workitems. 1166 */ 1167 static void workitem_free(struct worklist *, int); 1168 static void workitem_alloc(struct worklist *, int, struct mount *); 1169 static void workitem_reassign(struct worklist *, int); 1170 1171 #define WORKITEM_FREE(item, type) \ 1172 workitem_free((struct worklist *)(item), (type)) 1173 #define WORKITEM_REASSIGN(item, type) \ 1174 workitem_reassign((struct worklist *)(item), (type)) 1175 1176 static void 1177 workitem_free(item, type) 1178 struct worklist *item; 1179 int type; 1180 { 1181 struct ufsmount *ump; 1182 1183 #ifdef INVARIANTS 1184 if (item->wk_state & ONWORKLIST) 1185 panic("workitem_free: %s(0x%X) still on list, " 1186 "added in function %s at line %d", 1187 TYPENAME(item->wk_type), item->wk_state, 1188 item->wk_func, item->wk_line); 1189 if (item->wk_type != type && type != D_NEWBLK) 1190 panic("workitem_free: type mismatch %s != %s", 1191 TYPENAME(item->wk_type), TYPENAME(type)); 1192 #endif 1193 if (item->wk_state & IOWAITING) 1194 wakeup(item); 1195 ump = VFSTOUFS(item->wk_mp); 1196 LOCK_OWNED(ump); 1197 KASSERT(ump->softdep_deps > 0, 1198 ("workitem_free: %s: softdep_deps going negative", 1199 ump->um_fs->fs_fsmnt)); 1200 if (--ump->softdep_deps == 0 && ump->softdep_req) 1201 wakeup(&ump->softdep_deps); 1202 KASSERT(dep_current[item->wk_type] > 0, 1203 ("workitem_free: %s: dep_current[%s] going negative", 1204 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1205 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1206 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1207 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1208 atomic_subtract_long(&dep_current[item->wk_type], 1); 1209 ump->softdep_curdeps[item->wk_type] -= 1; 1210 free(item, DtoM(type)); 1211 } 1212 1213 static void 1214 workitem_alloc(item, type, mp) 1215 struct worklist *item; 1216 int type; 1217 struct mount *mp; 1218 { 1219 struct ufsmount *ump; 1220 1221 item->wk_type = type; 1222 item->wk_mp = mp; 1223 item->wk_state = 0; 1224 1225 ump = VFSTOUFS(mp); 1226 ACQUIRE_GBLLOCK(&lk); 1227 dep_current[type]++; 1228 if (dep_current[type] > dep_highuse[type]) 1229 dep_highuse[type] = dep_current[type]; 1230 dep_total[type]++; 1231 FREE_GBLLOCK(&lk); 1232 ACQUIRE_LOCK(ump); 1233 ump->softdep_curdeps[type] += 1; 1234 ump->softdep_deps++; 1235 ump->softdep_accdeps++; 1236 FREE_LOCK(ump); 1237 } 1238 1239 static void 1240 workitem_reassign(item, newtype) 1241 struct worklist *item; 1242 int newtype; 1243 { 1244 struct ufsmount *ump; 1245 1246 ump = VFSTOUFS(item->wk_mp); 1247 LOCK_OWNED(ump); 1248 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1249 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1250 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1251 ump->softdep_curdeps[item->wk_type] -= 1; 1252 ump->softdep_curdeps[newtype] += 1; 1253 KASSERT(dep_current[item->wk_type] > 0, 1254 ("workitem_reassign: %s: dep_current[%s] going negative", 1255 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1256 ACQUIRE_GBLLOCK(&lk); 1257 dep_current[newtype]++; 1258 dep_current[item->wk_type]--; 1259 if (dep_current[newtype] > dep_highuse[newtype]) 1260 dep_highuse[newtype] = dep_current[newtype]; 1261 dep_total[newtype]++; 1262 FREE_GBLLOCK(&lk); 1263 item->wk_type = newtype; 1264 } 1265 1266 /* 1267 * Workitem queue management 1268 */ 1269 static int max_softdeps; /* maximum number of structs before slowdown */ 1270 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1271 static int proc_waiting; /* tracks whether we have a timeout posted */ 1272 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1273 static struct callout softdep_callout; 1274 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1275 static int req_clear_remove; /* syncer process flush some freeblks */ 1276 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1277 1278 /* 1279 * runtime statistics 1280 */ 1281 static int stat_flush_threads; /* number of softdep flushing threads */ 1282 static int stat_worklist_push; /* number of worklist cleanups */ 1283 static int stat_blk_limit_push; /* number of times block limit neared */ 1284 static int stat_ino_limit_push; /* number of times inode limit neared */ 1285 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1286 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1287 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1288 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1289 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1290 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1291 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1292 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1293 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1294 static int stat_journal_min; /* Times hit journal min threshold */ 1295 static int stat_journal_low; /* Times hit journal low threshold */ 1296 static int stat_journal_wait; /* Times blocked in jwait(). */ 1297 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1298 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1299 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1300 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1301 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1302 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1303 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1304 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1305 static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1306 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1307 1308 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1309 &max_softdeps, 0, ""); 1310 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1311 &tickdelay, 0, ""); 1312 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1313 &stat_flush_threads, 0, ""); 1314 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1315 &stat_worklist_push, 0,""); 1316 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1317 &stat_blk_limit_push, 0,""); 1318 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1319 &stat_ino_limit_push, 0,""); 1320 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1321 &stat_blk_limit_hit, 0, ""); 1322 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1323 &stat_ino_limit_hit, 0, ""); 1324 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1325 &stat_sync_limit_hit, 0, ""); 1326 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1327 &stat_indir_blk_ptrs, 0, ""); 1328 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1329 &stat_inode_bitmap, 0, ""); 1330 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1331 &stat_direct_blk_ptrs, 0, ""); 1332 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1333 &stat_dir_entry, 0, ""); 1334 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1335 &stat_jaddref, 0, ""); 1336 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1337 &stat_jnewblk, 0, ""); 1338 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1339 &stat_journal_low, 0, ""); 1340 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1341 &stat_journal_min, 0, ""); 1342 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1343 &stat_journal_wait, 0, ""); 1344 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1345 &stat_jwait_filepage, 0, ""); 1346 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1347 &stat_jwait_freeblks, 0, ""); 1348 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1349 &stat_jwait_inode, 0, ""); 1350 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1351 &stat_jwait_newblk, 0, ""); 1352 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1353 &stat_cleanup_blkrequests, 0, ""); 1354 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1355 &stat_cleanup_inorequests, 0, ""); 1356 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1357 &stat_cleanup_high_delay, 0, ""); 1358 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1359 &stat_cleanup_retries, 0, ""); 1360 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1361 &stat_cleanup_failures, 0, ""); 1362 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1363 &softdep_flushcache, 0, ""); 1364 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1365 &stat_emptyjblocks, 0, ""); 1366 1367 SYSCTL_DECL(_vfs_ffs); 1368 1369 /* Whether to recompute the summary at mount time */ 1370 static int compute_summary_at_mount = 0; 1371 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1372 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1373 static int print_threads = 0; 1374 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1375 &print_threads, 0, "Notify flusher thread start/stop"); 1376 1377 /* List of all filesystems mounted with soft updates */ 1378 static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1379 1380 /* 1381 * This function cleans the worklist for a filesystem. 1382 * Each filesystem running with soft dependencies gets its own 1383 * thread to run in this function. The thread is started up in 1384 * softdep_mount and shutdown in softdep_unmount. They show up 1385 * as part of the kernel "bufdaemon" process whose process 1386 * entry is available in bufdaemonproc. 1387 */ 1388 static int searchfailed; 1389 extern struct proc *bufdaemonproc; 1390 static void 1391 softdep_flush(addr) 1392 void *addr; 1393 { 1394 struct mount *mp; 1395 struct thread *td; 1396 struct ufsmount *ump; 1397 1398 td = curthread; 1399 td->td_pflags |= TDP_NORUNNINGBUF; 1400 mp = (struct mount *)addr; 1401 ump = VFSTOUFS(mp); 1402 atomic_add_int(&stat_flush_threads, 1); 1403 ACQUIRE_LOCK(ump); 1404 ump->softdep_flags &= ~FLUSH_STARTING; 1405 wakeup(&ump->softdep_flushtd); 1406 FREE_LOCK(ump); 1407 if (print_threads) { 1408 if (stat_flush_threads == 1) 1409 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1410 bufdaemonproc->p_pid); 1411 printf("Start thread %s\n", td->td_name); 1412 } 1413 for (;;) { 1414 while (softdep_process_worklist(mp, 0) > 0 || 1415 (MOUNTEDSUJ(mp) && 1416 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1417 kthread_suspend_check(); 1418 ACQUIRE_LOCK(ump); 1419 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1420 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1421 "sdflush", hz / 2); 1422 ump->softdep_flags &= ~FLUSH_CLEANUP; 1423 /* 1424 * Check to see if we are done and need to exit. 1425 */ 1426 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1427 FREE_LOCK(ump); 1428 continue; 1429 } 1430 ump->softdep_flags &= ~FLUSH_EXIT; 1431 FREE_LOCK(ump); 1432 wakeup(&ump->softdep_flags); 1433 if (print_threads) 1434 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1435 atomic_subtract_int(&stat_flush_threads, 1); 1436 kthread_exit(); 1437 panic("kthread_exit failed\n"); 1438 } 1439 } 1440 1441 static void 1442 worklist_speedup(mp) 1443 struct mount *mp; 1444 { 1445 struct ufsmount *ump; 1446 1447 ump = VFSTOUFS(mp); 1448 LOCK_OWNED(ump); 1449 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1450 ump->softdep_flags |= FLUSH_CLEANUP; 1451 wakeup(&ump->softdep_flushtd); 1452 } 1453 1454 static int 1455 softdep_speedup(ump) 1456 struct ufsmount *ump; 1457 { 1458 struct ufsmount *altump; 1459 struct mount_softdeps *sdp; 1460 1461 LOCK_OWNED(ump); 1462 worklist_speedup(ump->um_mountp); 1463 bd_speedup(); 1464 /* 1465 * If we have global shortages, then we need other 1466 * filesystems to help with the cleanup. Here we wakeup a 1467 * flusher thread for a filesystem that is over its fair 1468 * share of resources. 1469 */ 1470 if (req_clear_inodedeps || req_clear_remove) { 1471 ACQUIRE_GBLLOCK(&lk); 1472 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1473 if ((altump = sdp->sd_ump) == ump) 1474 continue; 1475 if (((req_clear_inodedeps && 1476 altump->softdep_curdeps[D_INODEDEP] > 1477 max_softdeps / stat_flush_threads) || 1478 (req_clear_remove && 1479 altump->softdep_curdeps[D_DIRREM] > 1480 (max_softdeps / 2) / stat_flush_threads)) && 1481 TRY_ACQUIRE_LOCK(altump)) 1482 break; 1483 } 1484 if (sdp == NULL) { 1485 searchfailed++; 1486 FREE_GBLLOCK(&lk); 1487 } else { 1488 /* 1489 * Move to the end of the list so we pick a 1490 * different one on out next try. 1491 */ 1492 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1493 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1494 FREE_GBLLOCK(&lk); 1495 if ((altump->softdep_flags & 1496 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1497 altump->softdep_flags |= FLUSH_CLEANUP; 1498 altump->um_softdep->sd_cleanups++; 1499 wakeup(&altump->softdep_flushtd); 1500 FREE_LOCK(altump); 1501 } 1502 } 1503 return (speedup_syncer()); 1504 } 1505 1506 /* 1507 * Add an item to the end of the work queue. 1508 * This routine requires that the lock be held. 1509 * This is the only routine that adds items to the list. 1510 * The following routine is the only one that removes items 1511 * and does so in order from first to last. 1512 */ 1513 1514 #define WK_HEAD 0x0001 /* Add to HEAD. */ 1515 #define WK_NODELAY 0x0002 /* Process immediately. */ 1516 1517 static void 1518 add_to_worklist(wk, flags) 1519 struct worklist *wk; 1520 int flags; 1521 { 1522 struct ufsmount *ump; 1523 1524 ump = VFSTOUFS(wk->wk_mp); 1525 LOCK_OWNED(ump); 1526 if (wk->wk_state & ONWORKLIST) 1527 panic("add_to_worklist: %s(0x%X) already on list", 1528 TYPENAME(wk->wk_type), wk->wk_state); 1529 wk->wk_state |= ONWORKLIST; 1530 if (ump->softdep_on_worklist == 0) { 1531 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1532 ump->softdep_worklist_tail = wk; 1533 } else if (flags & WK_HEAD) { 1534 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1535 } else { 1536 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1537 ump->softdep_worklist_tail = wk; 1538 } 1539 ump->softdep_on_worklist += 1; 1540 if (flags & WK_NODELAY) 1541 worklist_speedup(wk->wk_mp); 1542 } 1543 1544 /* 1545 * Remove the item to be processed. If we are removing the last 1546 * item on the list, we need to recalculate the tail pointer. 1547 */ 1548 static void 1549 remove_from_worklist(wk) 1550 struct worklist *wk; 1551 { 1552 struct ufsmount *ump; 1553 1554 ump = VFSTOUFS(wk->wk_mp); 1555 if (ump->softdep_worklist_tail == wk) 1556 ump->softdep_worklist_tail = 1557 (struct worklist *)wk->wk_list.le_prev; 1558 WORKLIST_REMOVE(wk); 1559 ump->softdep_on_worklist -= 1; 1560 } 1561 1562 static void 1563 wake_worklist(wk) 1564 struct worklist *wk; 1565 { 1566 if (wk->wk_state & IOWAITING) { 1567 wk->wk_state &= ~IOWAITING; 1568 wakeup(wk); 1569 } 1570 } 1571 1572 static void 1573 wait_worklist(wk, wmesg) 1574 struct worklist *wk; 1575 char *wmesg; 1576 { 1577 struct ufsmount *ump; 1578 1579 ump = VFSTOUFS(wk->wk_mp); 1580 wk->wk_state |= IOWAITING; 1581 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1582 } 1583 1584 /* 1585 * Process that runs once per second to handle items in the background queue. 1586 * 1587 * Note that we ensure that everything is done in the order in which they 1588 * appear in the queue. The code below depends on this property to ensure 1589 * that blocks of a file are freed before the inode itself is freed. This 1590 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1591 * until all the old ones have been purged from the dependency lists. 1592 */ 1593 static int 1594 softdep_process_worklist(mp, full) 1595 struct mount *mp; 1596 int full; 1597 { 1598 int cnt, matchcnt; 1599 struct ufsmount *ump; 1600 long starttime; 1601 1602 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1603 if (MOUNTEDSOFTDEP(mp) == 0) 1604 return (0); 1605 matchcnt = 0; 1606 ump = VFSTOUFS(mp); 1607 ACQUIRE_LOCK(ump); 1608 starttime = time_second; 1609 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1610 check_clear_deps(mp); 1611 while (ump->softdep_on_worklist > 0) { 1612 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1613 break; 1614 else 1615 matchcnt += cnt; 1616 check_clear_deps(mp); 1617 /* 1618 * We do not generally want to stop for buffer space, but if 1619 * we are really being a buffer hog, we will stop and wait. 1620 */ 1621 if (should_yield()) { 1622 FREE_LOCK(ump); 1623 kern_yield(PRI_USER); 1624 bwillwrite(); 1625 ACQUIRE_LOCK(ump); 1626 } 1627 /* 1628 * Never allow processing to run for more than one 1629 * second. This gives the syncer thread the opportunity 1630 * to pause if appropriate. 1631 */ 1632 if (!full && starttime != time_second) 1633 break; 1634 } 1635 if (full == 0) 1636 journal_unsuspend(ump); 1637 FREE_LOCK(ump); 1638 return (matchcnt); 1639 } 1640 1641 /* 1642 * Process all removes associated with a vnode if we are running out of 1643 * journal space. Any other process which attempts to flush these will 1644 * be unable as we have the vnodes locked. 1645 */ 1646 static void 1647 process_removes(vp) 1648 struct vnode *vp; 1649 { 1650 struct inodedep *inodedep; 1651 struct dirrem *dirrem; 1652 struct ufsmount *ump; 1653 struct mount *mp; 1654 ino_t inum; 1655 1656 mp = vp->v_mount; 1657 ump = VFSTOUFS(mp); 1658 LOCK_OWNED(ump); 1659 inum = VTOI(vp)->i_number; 1660 for (;;) { 1661 top: 1662 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1663 return; 1664 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1665 /* 1666 * If another thread is trying to lock this vnode 1667 * it will fail but we must wait for it to do so 1668 * before we can proceed. 1669 */ 1670 if (dirrem->dm_state & INPROGRESS) { 1671 wait_worklist(&dirrem->dm_list, "pwrwait"); 1672 goto top; 1673 } 1674 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1675 (COMPLETE | ONWORKLIST)) 1676 break; 1677 } 1678 if (dirrem == NULL) 1679 return; 1680 remove_from_worklist(&dirrem->dm_list); 1681 FREE_LOCK(ump); 1682 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1683 panic("process_removes: suspended filesystem"); 1684 handle_workitem_remove(dirrem, 0); 1685 vn_finished_secondary_write(mp); 1686 ACQUIRE_LOCK(ump); 1687 } 1688 } 1689 1690 /* 1691 * Process all truncations associated with a vnode if we are running out 1692 * of journal space. This is called when the vnode lock is already held 1693 * and no other process can clear the truncation. This function returns 1694 * a value greater than zero if it did any work. 1695 */ 1696 static void 1697 process_truncates(vp) 1698 struct vnode *vp; 1699 { 1700 struct inodedep *inodedep; 1701 struct freeblks *freeblks; 1702 struct ufsmount *ump; 1703 struct mount *mp; 1704 ino_t inum; 1705 int cgwait; 1706 1707 mp = vp->v_mount; 1708 ump = VFSTOUFS(mp); 1709 LOCK_OWNED(ump); 1710 inum = VTOI(vp)->i_number; 1711 for (;;) { 1712 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1713 return; 1714 cgwait = 0; 1715 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) { 1716 /* Journal entries not yet written. */ 1717 if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) { 1718 jwait(&LIST_FIRST( 1719 &freeblks->fb_jblkdephd)->jb_list, 1720 MNT_WAIT); 1721 break; 1722 } 1723 /* Another thread is executing this item. */ 1724 if (freeblks->fb_state & INPROGRESS) { 1725 wait_worklist(&freeblks->fb_list, "ptrwait"); 1726 break; 1727 } 1728 /* Freeblks is waiting on a inode write. */ 1729 if ((freeblks->fb_state & COMPLETE) == 0) { 1730 FREE_LOCK(ump); 1731 ffs_update(vp, 1); 1732 ACQUIRE_LOCK(ump); 1733 break; 1734 } 1735 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) == 1736 (ALLCOMPLETE | ONWORKLIST)) { 1737 remove_from_worklist(&freeblks->fb_list); 1738 freeblks->fb_state |= INPROGRESS; 1739 FREE_LOCK(ump); 1740 if (vn_start_secondary_write(NULL, &mp, 1741 V_NOWAIT)) 1742 panic("process_truncates: " 1743 "suspended filesystem"); 1744 handle_workitem_freeblocks(freeblks, 0); 1745 vn_finished_secondary_write(mp); 1746 ACQUIRE_LOCK(ump); 1747 break; 1748 } 1749 if (freeblks->fb_cgwait) 1750 cgwait++; 1751 } 1752 if (cgwait) { 1753 FREE_LOCK(ump); 1754 sync_cgs(mp, MNT_WAIT); 1755 ffs_sync_snap(mp, MNT_WAIT); 1756 ACQUIRE_LOCK(ump); 1757 continue; 1758 } 1759 if (freeblks == NULL) 1760 break; 1761 } 1762 return; 1763 } 1764 1765 /* 1766 * Process one item on the worklist. 1767 */ 1768 static int 1769 process_worklist_item(mp, target, flags) 1770 struct mount *mp; 1771 int target; 1772 int flags; 1773 { 1774 struct worklist sentinel; 1775 struct worklist *wk; 1776 struct ufsmount *ump; 1777 int matchcnt; 1778 int error; 1779 1780 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1781 /* 1782 * If we are being called because of a process doing a 1783 * copy-on-write, then it is not safe to write as we may 1784 * recurse into the copy-on-write routine. 1785 */ 1786 if (curthread->td_pflags & TDP_COWINPROGRESS) 1787 return (-1); 1788 PHOLD(curproc); /* Don't let the stack go away. */ 1789 ump = VFSTOUFS(mp); 1790 LOCK_OWNED(ump); 1791 matchcnt = 0; 1792 sentinel.wk_mp = NULL; 1793 sentinel.wk_type = D_SENTINEL; 1794 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list); 1795 for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL; 1796 wk = LIST_NEXT(&sentinel, wk_list)) { 1797 if (wk->wk_type == D_SENTINEL) { 1798 LIST_REMOVE(&sentinel, wk_list); 1799 LIST_INSERT_AFTER(wk, &sentinel, wk_list); 1800 continue; 1801 } 1802 if (wk->wk_state & INPROGRESS) 1803 panic("process_worklist_item: %p already in progress.", 1804 wk); 1805 wk->wk_state |= INPROGRESS; 1806 remove_from_worklist(wk); 1807 FREE_LOCK(ump); 1808 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1809 panic("process_worklist_item: suspended filesystem"); 1810 switch (wk->wk_type) { 1811 case D_DIRREM: 1812 /* removal of a directory entry */ 1813 error = handle_workitem_remove(WK_DIRREM(wk), flags); 1814 break; 1815 1816 case D_FREEBLKS: 1817 /* releasing blocks and/or fragments from a file */ 1818 error = handle_workitem_freeblocks(WK_FREEBLKS(wk), 1819 flags); 1820 break; 1821 1822 case D_FREEFRAG: 1823 /* releasing a fragment when replaced as a file grows */ 1824 handle_workitem_freefrag(WK_FREEFRAG(wk)); 1825 error = 0; 1826 break; 1827 1828 case D_FREEFILE: 1829 /* releasing an inode when its link count drops to 0 */ 1830 handle_workitem_freefile(WK_FREEFILE(wk)); 1831 error = 0; 1832 break; 1833 1834 default: 1835 panic("%s_process_worklist: Unknown type %s", 1836 "softdep", TYPENAME(wk->wk_type)); 1837 /* NOTREACHED */ 1838 } 1839 vn_finished_secondary_write(mp); 1840 ACQUIRE_LOCK(ump); 1841 if (error == 0) { 1842 if (++matchcnt == target) 1843 break; 1844 continue; 1845 } 1846 /* 1847 * We have to retry the worklist item later. Wake up any 1848 * waiters who may be able to complete it immediately and 1849 * add the item back to the head so we don't try to execute 1850 * it again. 1851 */ 1852 wk->wk_state &= ~INPROGRESS; 1853 wake_worklist(wk); 1854 add_to_worklist(wk, WK_HEAD); 1855 } 1856 /* Sentinal could've become the tail from remove_from_worklist. */ 1857 if (ump->softdep_worklist_tail == &sentinel) 1858 ump->softdep_worklist_tail = 1859 (struct worklist *)sentinel.wk_list.le_prev; 1860 LIST_REMOVE(&sentinel, wk_list); 1861 PRELE(curproc); 1862 return (matchcnt); 1863 } 1864 1865 /* 1866 * Move dependencies from one buffer to another. 1867 */ 1868 int 1869 softdep_move_dependencies(oldbp, newbp) 1870 struct buf *oldbp; 1871 struct buf *newbp; 1872 { 1873 struct worklist *wk, *wktail; 1874 struct ufsmount *ump; 1875 int dirty; 1876 1877 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1878 return (0); 1879 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1880 ("softdep_move_dependencies called on non-softdep filesystem")); 1881 dirty = 0; 1882 wktail = NULL; 1883 ump = VFSTOUFS(wk->wk_mp); 1884 ACQUIRE_LOCK(ump); 1885 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1886 LIST_REMOVE(wk, wk_list); 1887 if (wk->wk_type == D_BMSAFEMAP && 1888 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1889 dirty = 1; 1890 if (wktail == NULL) 1891 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1892 else 1893 LIST_INSERT_AFTER(wktail, wk, wk_list); 1894 wktail = wk; 1895 } 1896 FREE_LOCK(ump); 1897 1898 return (dirty); 1899 } 1900 1901 /* 1902 * Purge the work list of all items associated with a particular mount point. 1903 */ 1904 int 1905 softdep_flushworklist(oldmnt, countp, td) 1906 struct mount *oldmnt; 1907 int *countp; 1908 struct thread *td; 1909 { 1910 struct vnode *devvp; 1911 struct ufsmount *ump; 1912 int count, error; 1913 1914 /* 1915 * Alternately flush the block device associated with the mount 1916 * point and process any dependencies that the flushing 1917 * creates. We continue until no more worklist dependencies 1918 * are found. 1919 */ 1920 *countp = 0; 1921 error = 0; 1922 ump = VFSTOUFS(oldmnt); 1923 devvp = ump->um_devvp; 1924 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1925 *countp += count; 1926 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1927 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1928 VOP_UNLOCK(devvp, 0); 1929 if (error != 0) 1930 break; 1931 } 1932 return (error); 1933 } 1934 1935 #define SU_WAITIDLE_RETRIES 20 1936 static int 1937 softdep_waitidle(struct mount *mp, int flags __unused) 1938 { 1939 struct ufsmount *ump; 1940 struct vnode *devvp; 1941 struct thread *td; 1942 int error, i; 1943 1944 ump = VFSTOUFS(mp); 1945 devvp = ump->um_devvp; 1946 td = curthread; 1947 error = 0; 1948 ACQUIRE_LOCK(ump); 1949 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 1950 ump->softdep_req = 1; 1951 KASSERT((flags & FORCECLOSE) == 0 || 1952 ump->softdep_on_worklist == 0, 1953 ("softdep_waitidle: work added after flush")); 1954 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 1955 "softdeps", 10 * hz); 1956 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1957 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1958 VOP_UNLOCK(devvp, 0); 1959 ACQUIRE_LOCK(ump); 1960 if (error != 0) 1961 break; 1962 } 1963 ump->softdep_req = 0; 1964 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 1965 error = EBUSY; 1966 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1967 mp); 1968 } 1969 FREE_LOCK(ump); 1970 return (error); 1971 } 1972 1973 /* 1974 * Flush all vnodes and worklist items associated with a specified mount point. 1975 */ 1976 int 1977 softdep_flushfiles(oldmnt, flags, td) 1978 struct mount *oldmnt; 1979 int flags; 1980 struct thread *td; 1981 { 1982 #ifdef QUOTA 1983 struct ufsmount *ump; 1984 int i; 1985 #endif 1986 int error, early, depcount, loopcnt, retry_flush_count, retry; 1987 int morework; 1988 1989 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1990 ("softdep_flushfiles called on non-softdep filesystem")); 1991 loopcnt = 10; 1992 retry_flush_count = 3; 1993 retry_flush: 1994 error = 0; 1995 1996 /* 1997 * Alternately flush the vnodes associated with the mount 1998 * point and process any dependencies that the flushing 1999 * creates. In theory, this loop can happen at most twice, 2000 * but we give it a few extra just to be sure. 2001 */ 2002 for (; loopcnt > 0; loopcnt--) { 2003 /* 2004 * Do another flush in case any vnodes were brought in 2005 * as part of the cleanup operations. 2006 */ 2007 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 2008 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 2009 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 2010 break; 2011 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 2012 depcount == 0) 2013 break; 2014 } 2015 /* 2016 * If we are unmounting then it is an error to fail. If we 2017 * are simply trying to downgrade to read-only, then filesystem 2018 * activity can keep us busy forever, so we just fail with EBUSY. 2019 */ 2020 if (loopcnt == 0) { 2021 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2022 panic("softdep_flushfiles: looping"); 2023 error = EBUSY; 2024 } 2025 if (!error) 2026 error = softdep_waitidle(oldmnt, flags); 2027 if (!error) { 2028 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2029 retry = 0; 2030 MNT_ILOCK(oldmnt); 2031 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 2032 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 2033 morework = oldmnt->mnt_nvnodelistsize > 0; 2034 #ifdef QUOTA 2035 ump = VFSTOUFS(oldmnt); 2036 UFS_LOCK(ump); 2037 for (i = 0; i < MAXQUOTAS; i++) { 2038 if (ump->um_quotas[i] != NULLVP) 2039 morework = 1; 2040 } 2041 UFS_UNLOCK(ump); 2042 #endif 2043 if (morework) { 2044 if (--retry_flush_count > 0) { 2045 retry = 1; 2046 loopcnt = 3; 2047 } else 2048 error = EBUSY; 2049 } 2050 MNT_IUNLOCK(oldmnt); 2051 if (retry) 2052 goto retry_flush; 2053 } 2054 } 2055 return (error); 2056 } 2057 2058 /* 2059 * Structure hashing. 2060 * 2061 * There are four types of structures that can be looked up: 2062 * 1) pagedep structures identified by mount point, inode number, 2063 * and logical block. 2064 * 2) inodedep structures identified by mount point and inode number. 2065 * 3) newblk structures identified by mount point and 2066 * physical block number. 2067 * 4) bmsafemap structures identified by mount point and 2068 * cylinder group number. 2069 * 2070 * The "pagedep" and "inodedep" dependency structures are hashed 2071 * separately from the file blocks and inodes to which they correspond. 2072 * This separation helps when the in-memory copy of an inode or 2073 * file block must be replaced. It also obviates the need to access 2074 * an inode or file page when simply updating (or de-allocating) 2075 * dependency structures. Lookup of newblk structures is needed to 2076 * find newly allocated blocks when trying to associate them with 2077 * their allocdirect or allocindir structure. 2078 * 2079 * The lookup routines optionally create and hash a new instance when 2080 * an existing entry is not found. The bmsafemap lookup routine always 2081 * allocates a new structure if an existing one is not found. 2082 */ 2083 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2084 2085 /* 2086 * Structures and routines associated with pagedep caching. 2087 */ 2088 #define PAGEDEP_HASH(ump, inum, lbn) \ 2089 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2090 2091 static int 2092 pagedep_find(pagedephd, ino, lbn, pagedeppp) 2093 struct pagedep_hashhead *pagedephd; 2094 ino_t ino; 2095 ufs_lbn_t lbn; 2096 struct pagedep **pagedeppp; 2097 { 2098 struct pagedep *pagedep; 2099 2100 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2101 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2102 *pagedeppp = pagedep; 2103 return (1); 2104 } 2105 } 2106 *pagedeppp = NULL; 2107 return (0); 2108 } 2109 /* 2110 * Look up a pagedep. Return 1 if found, 0 otherwise. 2111 * If not found, allocate if DEPALLOC flag is passed. 2112 * Found or allocated entry is returned in pagedeppp. 2113 */ 2114 static int 2115 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2116 struct mount *mp; 2117 struct buf *bp; 2118 ino_t ino; 2119 ufs_lbn_t lbn; 2120 int flags; 2121 struct pagedep **pagedeppp; 2122 { 2123 struct pagedep *pagedep; 2124 struct pagedep_hashhead *pagedephd; 2125 struct worklist *wk; 2126 struct ufsmount *ump; 2127 int ret; 2128 int i; 2129 2130 ump = VFSTOUFS(mp); 2131 LOCK_OWNED(ump); 2132 if (bp) { 2133 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2134 if (wk->wk_type == D_PAGEDEP) { 2135 *pagedeppp = WK_PAGEDEP(wk); 2136 return (1); 2137 } 2138 } 2139 } 2140 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2141 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2142 if (ret) { 2143 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2144 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2145 return (1); 2146 } 2147 if ((flags & DEPALLOC) == 0) 2148 return (0); 2149 FREE_LOCK(ump); 2150 pagedep = malloc(sizeof(struct pagedep), 2151 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2152 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2153 ACQUIRE_LOCK(ump); 2154 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2155 if (*pagedeppp) { 2156 /* 2157 * This should never happen since we only create pagedeps 2158 * with the vnode lock held. Could be an assert. 2159 */ 2160 WORKITEM_FREE(pagedep, D_PAGEDEP); 2161 return (ret); 2162 } 2163 pagedep->pd_ino = ino; 2164 pagedep->pd_lbn = lbn; 2165 LIST_INIT(&pagedep->pd_dirremhd); 2166 LIST_INIT(&pagedep->pd_pendinghd); 2167 for (i = 0; i < DAHASHSZ; i++) 2168 LIST_INIT(&pagedep->pd_diraddhd[i]); 2169 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2170 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2171 *pagedeppp = pagedep; 2172 return (0); 2173 } 2174 2175 /* 2176 * Structures and routines associated with inodedep caching. 2177 */ 2178 #define INODEDEP_HASH(ump, inum) \ 2179 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2180 2181 static int 2182 inodedep_find(inodedephd, inum, inodedeppp) 2183 struct inodedep_hashhead *inodedephd; 2184 ino_t inum; 2185 struct inodedep **inodedeppp; 2186 { 2187 struct inodedep *inodedep; 2188 2189 LIST_FOREACH(inodedep, inodedephd, id_hash) 2190 if (inum == inodedep->id_ino) 2191 break; 2192 if (inodedep) { 2193 *inodedeppp = inodedep; 2194 return (1); 2195 } 2196 *inodedeppp = NULL; 2197 2198 return (0); 2199 } 2200 /* 2201 * Look up an inodedep. Return 1 if found, 0 if not found. 2202 * If not found, allocate if DEPALLOC flag is passed. 2203 * Found or allocated entry is returned in inodedeppp. 2204 */ 2205 static int 2206 inodedep_lookup(mp, inum, flags, inodedeppp) 2207 struct mount *mp; 2208 ino_t inum; 2209 int flags; 2210 struct inodedep **inodedeppp; 2211 { 2212 struct inodedep *inodedep; 2213 struct inodedep_hashhead *inodedephd; 2214 struct ufsmount *ump; 2215 struct fs *fs; 2216 2217 ump = VFSTOUFS(mp); 2218 LOCK_OWNED(ump); 2219 fs = ump->um_fs; 2220 inodedephd = INODEDEP_HASH(ump, inum); 2221 2222 if (inodedep_find(inodedephd, inum, inodedeppp)) 2223 return (1); 2224 if ((flags & DEPALLOC) == 0) 2225 return (0); 2226 /* 2227 * If the system is over its limit and our filesystem is 2228 * responsible for more than our share of that usage and 2229 * we are not in a rush, request some inodedep cleanup. 2230 */ 2231 if (softdep_excess_items(ump, D_INODEDEP)) 2232 schedule_cleanup(mp); 2233 else 2234 FREE_LOCK(ump); 2235 inodedep = malloc(sizeof(struct inodedep), 2236 M_INODEDEP, M_SOFTDEP_FLAGS); 2237 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2238 ACQUIRE_LOCK(ump); 2239 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2240 WORKITEM_FREE(inodedep, D_INODEDEP); 2241 return (1); 2242 } 2243 inodedep->id_fs = fs; 2244 inodedep->id_ino = inum; 2245 inodedep->id_state = ALLCOMPLETE; 2246 inodedep->id_nlinkdelta = 0; 2247 inodedep->id_savedino1 = NULL; 2248 inodedep->id_savedsize = -1; 2249 inodedep->id_savedextsize = -1; 2250 inodedep->id_savednlink = -1; 2251 inodedep->id_bmsafemap = NULL; 2252 inodedep->id_mkdiradd = NULL; 2253 LIST_INIT(&inodedep->id_dirremhd); 2254 LIST_INIT(&inodedep->id_pendinghd); 2255 LIST_INIT(&inodedep->id_inowait); 2256 LIST_INIT(&inodedep->id_bufwait); 2257 TAILQ_INIT(&inodedep->id_inoreflst); 2258 TAILQ_INIT(&inodedep->id_inoupdt); 2259 TAILQ_INIT(&inodedep->id_newinoupdt); 2260 TAILQ_INIT(&inodedep->id_extupdt); 2261 TAILQ_INIT(&inodedep->id_newextupdt); 2262 TAILQ_INIT(&inodedep->id_freeblklst); 2263 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2264 *inodedeppp = inodedep; 2265 return (0); 2266 } 2267 2268 /* 2269 * Structures and routines associated with newblk caching. 2270 */ 2271 #define NEWBLK_HASH(ump, inum) \ 2272 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2273 2274 static int 2275 newblk_find(newblkhd, newblkno, flags, newblkpp) 2276 struct newblk_hashhead *newblkhd; 2277 ufs2_daddr_t newblkno; 2278 int flags; 2279 struct newblk **newblkpp; 2280 { 2281 struct newblk *newblk; 2282 2283 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2284 if (newblkno != newblk->nb_newblkno) 2285 continue; 2286 /* 2287 * If we're creating a new dependency don't match those that 2288 * have already been converted to allocdirects. This is for 2289 * a frag extend. 2290 */ 2291 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2292 continue; 2293 break; 2294 } 2295 if (newblk) { 2296 *newblkpp = newblk; 2297 return (1); 2298 } 2299 *newblkpp = NULL; 2300 return (0); 2301 } 2302 2303 /* 2304 * Look up a newblk. Return 1 if found, 0 if not found. 2305 * If not found, allocate if DEPALLOC flag is passed. 2306 * Found or allocated entry is returned in newblkpp. 2307 */ 2308 static int 2309 newblk_lookup(mp, newblkno, flags, newblkpp) 2310 struct mount *mp; 2311 ufs2_daddr_t newblkno; 2312 int flags; 2313 struct newblk **newblkpp; 2314 { 2315 struct newblk *newblk; 2316 struct newblk_hashhead *newblkhd; 2317 struct ufsmount *ump; 2318 2319 ump = VFSTOUFS(mp); 2320 LOCK_OWNED(ump); 2321 newblkhd = NEWBLK_HASH(ump, newblkno); 2322 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2323 return (1); 2324 if ((flags & DEPALLOC) == 0) 2325 return (0); 2326 if (softdep_excess_items(ump, D_NEWBLK) || 2327 softdep_excess_items(ump, D_ALLOCDIRECT) || 2328 softdep_excess_items(ump, D_ALLOCINDIR)) 2329 schedule_cleanup(mp); 2330 else 2331 FREE_LOCK(ump); 2332 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2333 M_SOFTDEP_FLAGS | M_ZERO); 2334 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2335 ACQUIRE_LOCK(ump); 2336 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2337 WORKITEM_FREE(newblk, D_NEWBLK); 2338 return (1); 2339 } 2340 newblk->nb_freefrag = NULL; 2341 LIST_INIT(&newblk->nb_indirdeps); 2342 LIST_INIT(&newblk->nb_newdirblk); 2343 LIST_INIT(&newblk->nb_jwork); 2344 newblk->nb_state = ATTACHED; 2345 newblk->nb_newblkno = newblkno; 2346 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2347 *newblkpp = newblk; 2348 return (0); 2349 } 2350 2351 /* 2352 * Structures and routines associated with freed indirect block caching. 2353 */ 2354 #define INDIR_HASH(ump, blkno) \ 2355 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2356 2357 /* 2358 * Lookup an indirect block in the indir hash table. The freework is 2359 * removed and potentially freed. The caller must do a blocking journal 2360 * write before writing to the blkno. 2361 */ 2362 static int 2363 indirblk_lookup(mp, blkno) 2364 struct mount *mp; 2365 ufs2_daddr_t blkno; 2366 { 2367 struct freework *freework; 2368 struct indir_hashhead *wkhd; 2369 struct ufsmount *ump; 2370 2371 ump = VFSTOUFS(mp); 2372 wkhd = INDIR_HASH(ump, blkno); 2373 TAILQ_FOREACH(freework, wkhd, fw_next) { 2374 if (freework->fw_blkno != blkno) 2375 continue; 2376 indirblk_remove(freework); 2377 return (1); 2378 } 2379 return (0); 2380 } 2381 2382 /* 2383 * Insert an indirect block represented by freework into the indirblk 2384 * hash table so that it may prevent the block from being re-used prior 2385 * to the journal being written. 2386 */ 2387 static void 2388 indirblk_insert(freework) 2389 struct freework *freework; 2390 { 2391 struct jblocks *jblocks; 2392 struct jseg *jseg; 2393 struct ufsmount *ump; 2394 2395 ump = VFSTOUFS(freework->fw_list.wk_mp); 2396 jblocks = ump->softdep_jblocks; 2397 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2398 if (jseg == NULL) 2399 return; 2400 2401 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2402 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2403 fw_next); 2404 freework->fw_state &= ~DEPCOMPLETE; 2405 } 2406 2407 static void 2408 indirblk_remove(freework) 2409 struct freework *freework; 2410 { 2411 struct ufsmount *ump; 2412 2413 ump = VFSTOUFS(freework->fw_list.wk_mp); 2414 LIST_REMOVE(freework, fw_segs); 2415 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2416 freework->fw_state |= DEPCOMPLETE; 2417 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2418 WORKITEM_FREE(freework, D_FREEWORK); 2419 } 2420 2421 /* 2422 * Executed during filesystem system initialization before 2423 * mounting any filesystems. 2424 */ 2425 void 2426 softdep_initialize() 2427 { 2428 2429 TAILQ_INIT(&softdepmounts); 2430 #ifdef __LP64__ 2431 max_softdeps = desiredvnodes * 4; 2432 #else 2433 max_softdeps = desiredvnodes * 2; 2434 #endif 2435 2436 /* initialise bioops hack */ 2437 bioops.io_start = softdep_disk_io_initiation; 2438 bioops.io_complete = softdep_disk_write_complete; 2439 bioops.io_deallocate = softdep_deallocate_dependencies; 2440 bioops.io_countdeps = softdep_count_dependencies; 2441 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2442 2443 /* Initialize the callout with an mtx. */ 2444 callout_init_mtx(&softdep_callout, &lk, 0); 2445 } 2446 2447 /* 2448 * Executed after all filesystems have been unmounted during 2449 * filesystem module unload. 2450 */ 2451 void 2452 softdep_uninitialize() 2453 { 2454 2455 /* clear bioops hack */ 2456 bioops.io_start = NULL; 2457 bioops.io_complete = NULL; 2458 bioops.io_deallocate = NULL; 2459 bioops.io_countdeps = NULL; 2460 softdep_ast_cleanup = NULL; 2461 2462 callout_drain(&softdep_callout); 2463 } 2464 2465 /* 2466 * Called at mount time to notify the dependency code that a 2467 * filesystem wishes to use it. 2468 */ 2469 int 2470 softdep_mount(devvp, mp, fs, cred) 2471 struct vnode *devvp; 2472 struct mount *mp; 2473 struct fs *fs; 2474 struct ucred *cred; 2475 { 2476 struct csum_total cstotal; 2477 struct mount_softdeps *sdp; 2478 struct ufsmount *ump; 2479 struct cg *cgp; 2480 struct buf *bp; 2481 u_int cyl, i; 2482 int error; 2483 2484 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2485 M_WAITOK | M_ZERO); 2486 MNT_ILOCK(mp); 2487 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2488 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2489 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2490 MNTK_SOFTDEP | MNTK_NOASYNC; 2491 } 2492 ump = VFSTOUFS(mp); 2493 ump->um_softdep = sdp; 2494 MNT_IUNLOCK(mp); 2495 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock"); 2496 sdp->sd_ump = ump; 2497 LIST_INIT(&ump->softdep_workitem_pending); 2498 LIST_INIT(&ump->softdep_journal_pending); 2499 TAILQ_INIT(&ump->softdep_unlinked); 2500 LIST_INIT(&ump->softdep_dirtycg); 2501 ump->softdep_worklist_tail = NULL; 2502 ump->softdep_on_worklist = 0; 2503 ump->softdep_deps = 0; 2504 LIST_INIT(&ump->softdep_mkdirlisthd); 2505 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2506 &ump->pagedep_hash_size); 2507 ump->pagedep_nextclean = 0; 2508 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2509 &ump->inodedep_hash_size); 2510 ump->inodedep_nextclean = 0; 2511 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2512 &ump->newblk_hash_size); 2513 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2514 &ump->bmsafemap_hash_size); 2515 i = 1 << (ffs(desiredvnodes / 10) - 1); 2516 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2517 M_FREEWORK, M_WAITOK); 2518 ump->indir_hash_size = i - 1; 2519 for (i = 0; i <= ump->indir_hash_size; i++) 2520 TAILQ_INIT(&ump->indir_hashtbl[i]); 2521 ACQUIRE_GBLLOCK(&lk); 2522 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2523 FREE_GBLLOCK(&lk); 2524 if ((fs->fs_flags & FS_SUJ) && 2525 (error = journal_mount(mp, fs, cred)) != 0) { 2526 printf("Failed to start journal: %d\n", error); 2527 softdep_unmount(mp); 2528 return (error); 2529 } 2530 /* 2531 * Start our flushing thread in the bufdaemon process. 2532 */ 2533 ACQUIRE_LOCK(ump); 2534 ump->softdep_flags |= FLUSH_STARTING; 2535 FREE_LOCK(ump); 2536 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2537 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2538 mp->mnt_stat.f_mntonname); 2539 ACQUIRE_LOCK(ump); 2540 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2541 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2542 hz / 2); 2543 } 2544 FREE_LOCK(ump); 2545 /* 2546 * When doing soft updates, the counters in the 2547 * superblock may have gotten out of sync. Recomputation 2548 * can take a long time and can be deferred for background 2549 * fsck. However, the old behavior of scanning the cylinder 2550 * groups and recalculating them at mount time is available 2551 * by setting vfs.ffs.compute_summary_at_mount to one. 2552 */ 2553 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2554 return (0); 2555 bzero(&cstotal, sizeof cstotal); 2556 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2557 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2558 fs->fs_cgsize, cred, &bp)) != 0) { 2559 brelse(bp); 2560 softdep_unmount(mp); 2561 return (error); 2562 } 2563 cgp = (struct cg *)bp->b_data; 2564 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2565 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2566 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2567 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2568 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2569 brelse(bp); 2570 } 2571 #ifdef INVARIANTS 2572 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2573 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2574 #endif 2575 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2576 return (0); 2577 } 2578 2579 void 2580 softdep_unmount(mp) 2581 struct mount *mp; 2582 { 2583 struct ufsmount *ump; 2584 #ifdef INVARIANTS 2585 int i; 2586 #endif 2587 2588 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2589 ("softdep_unmount called on non-softdep filesystem")); 2590 ump = VFSTOUFS(mp); 2591 MNT_ILOCK(mp); 2592 mp->mnt_flag &= ~MNT_SOFTDEP; 2593 if (MOUNTEDSUJ(mp) == 0) { 2594 MNT_IUNLOCK(mp); 2595 } else { 2596 mp->mnt_flag &= ~MNT_SUJ; 2597 MNT_IUNLOCK(mp); 2598 journal_unmount(ump); 2599 } 2600 /* 2601 * Shut down our flushing thread. Check for NULL is if 2602 * softdep_mount errors out before the thread has been created. 2603 */ 2604 if (ump->softdep_flushtd != NULL) { 2605 ACQUIRE_LOCK(ump); 2606 ump->softdep_flags |= FLUSH_EXIT; 2607 wakeup(&ump->softdep_flushtd); 2608 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2609 "sdwait", 0); 2610 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2611 ("Thread shutdown failed")); 2612 } 2613 /* 2614 * Free up our resources. 2615 */ 2616 ACQUIRE_GBLLOCK(&lk); 2617 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2618 FREE_GBLLOCK(&lk); 2619 rw_destroy(LOCK_PTR(ump)); 2620 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2621 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2622 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2623 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2624 ump->bmsafemap_hash_size); 2625 free(ump->indir_hashtbl, M_FREEWORK); 2626 #ifdef INVARIANTS 2627 for (i = 0; i <= D_LAST; i++) 2628 KASSERT(ump->softdep_curdeps[i] == 0, 2629 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2630 TYPENAME(i), ump->softdep_curdeps[i])); 2631 #endif 2632 free(ump->um_softdep, M_MOUNTDATA); 2633 } 2634 2635 static struct jblocks * 2636 jblocks_create(void) 2637 { 2638 struct jblocks *jblocks; 2639 2640 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2641 TAILQ_INIT(&jblocks->jb_segs); 2642 jblocks->jb_avail = 10; 2643 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2644 M_JBLOCKS, M_WAITOK | M_ZERO); 2645 2646 return (jblocks); 2647 } 2648 2649 static ufs2_daddr_t 2650 jblocks_alloc(jblocks, bytes, actual) 2651 struct jblocks *jblocks; 2652 int bytes; 2653 int *actual; 2654 { 2655 ufs2_daddr_t daddr; 2656 struct jextent *jext; 2657 int freecnt; 2658 int blocks; 2659 2660 blocks = bytes / DEV_BSIZE; 2661 jext = &jblocks->jb_extent[jblocks->jb_head]; 2662 freecnt = jext->je_blocks - jblocks->jb_off; 2663 if (freecnt == 0) { 2664 jblocks->jb_off = 0; 2665 if (++jblocks->jb_head > jblocks->jb_used) 2666 jblocks->jb_head = 0; 2667 jext = &jblocks->jb_extent[jblocks->jb_head]; 2668 freecnt = jext->je_blocks; 2669 } 2670 if (freecnt > blocks) 2671 freecnt = blocks; 2672 *actual = freecnt * DEV_BSIZE; 2673 daddr = jext->je_daddr + jblocks->jb_off; 2674 jblocks->jb_off += freecnt; 2675 jblocks->jb_free -= freecnt; 2676 2677 return (daddr); 2678 } 2679 2680 static void 2681 jblocks_free(jblocks, mp, bytes) 2682 struct jblocks *jblocks; 2683 struct mount *mp; 2684 int bytes; 2685 { 2686 2687 LOCK_OWNED(VFSTOUFS(mp)); 2688 jblocks->jb_free += bytes / DEV_BSIZE; 2689 if (jblocks->jb_suspended) 2690 worklist_speedup(mp); 2691 wakeup(jblocks); 2692 } 2693 2694 static void 2695 jblocks_destroy(jblocks) 2696 struct jblocks *jblocks; 2697 { 2698 2699 if (jblocks->jb_extent) 2700 free(jblocks->jb_extent, M_JBLOCKS); 2701 free(jblocks, M_JBLOCKS); 2702 } 2703 2704 static void 2705 jblocks_add(jblocks, daddr, blocks) 2706 struct jblocks *jblocks; 2707 ufs2_daddr_t daddr; 2708 int blocks; 2709 { 2710 struct jextent *jext; 2711 2712 jblocks->jb_blocks += blocks; 2713 jblocks->jb_free += blocks; 2714 jext = &jblocks->jb_extent[jblocks->jb_used]; 2715 /* Adding the first block. */ 2716 if (jext->je_daddr == 0) { 2717 jext->je_daddr = daddr; 2718 jext->je_blocks = blocks; 2719 return; 2720 } 2721 /* Extending the last extent. */ 2722 if (jext->je_daddr + jext->je_blocks == daddr) { 2723 jext->je_blocks += blocks; 2724 return; 2725 } 2726 /* Adding a new extent. */ 2727 if (++jblocks->jb_used == jblocks->jb_avail) { 2728 jblocks->jb_avail *= 2; 2729 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2730 M_JBLOCKS, M_WAITOK | M_ZERO); 2731 memcpy(jext, jblocks->jb_extent, 2732 sizeof(struct jextent) * jblocks->jb_used); 2733 free(jblocks->jb_extent, M_JBLOCKS); 2734 jblocks->jb_extent = jext; 2735 } 2736 jext = &jblocks->jb_extent[jblocks->jb_used]; 2737 jext->je_daddr = daddr; 2738 jext->je_blocks = blocks; 2739 return; 2740 } 2741 2742 int 2743 softdep_journal_lookup(mp, vpp) 2744 struct mount *mp; 2745 struct vnode **vpp; 2746 { 2747 struct componentname cnp; 2748 struct vnode *dvp; 2749 ino_t sujournal; 2750 int error; 2751 2752 error = VFS_VGET(mp, UFS_ROOTINO, LK_EXCLUSIVE, &dvp); 2753 if (error) 2754 return (error); 2755 bzero(&cnp, sizeof(cnp)); 2756 cnp.cn_nameiop = LOOKUP; 2757 cnp.cn_flags = ISLASTCN; 2758 cnp.cn_thread = curthread; 2759 cnp.cn_cred = curthread->td_ucred; 2760 cnp.cn_pnbuf = SUJ_FILE; 2761 cnp.cn_nameptr = SUJ_FILE; 2762 cnp.cn_namelen = strlen(SUJ_FILE); 2763 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2764 vput(dvp); 2765 if (error != 0) 2766 return (error); 2767 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2768 return (error); 2769 } 2770 2771 /* 2772 * Open and verify the journal file. 2773 */ 2774 static int 2775 journal_mount(mp, fs, cred) 2776 struct mount *mp; 2777 struct fs *fs; 2778 struct ucred *cred; 2779 { 2780 struct jblocks *jblocks; 2781 struct ufsmount *ump; 2782 struct vnode *vp; 2783 struct inode *ip; 2784 ufs2_daddr_t blkno; 2785 int bcount; 2786 int error; 2787 int i; 2788 2789 ump = VFSTOUFS(mp); 2790 ump->softdep_journal_tail = NULL; 2791 ump->softdep_on_journal = 0; 2792 ump->softdep_accdeps = 0; 2793 ump->softdep_req = 0; 2794 ump->softdep_jblocks = NULL; 2795 error = softdep_journal_lookup(mp, &vp); 2796 if (error != 0) { 2797 printf("Failed to find journal. Use tunefs to create one\n"); 2798 return (error); 2799 } 2800 ip = VTOI(vp); 2801 if (ip->i_size < SUJ_MIN) { 2802 error = ENOSPC; 2803 goto out; 2804 } 2805 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2806 jblocks = jblocks_create(); 2807 for (i = 0; i < bcount; i++) { 2808 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2809 if (error) 2810 break; 2811 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2812 } 2813 if (error) { 2814 jblocks_destroy(jblocks); 2815 goto out; 2816 } 2817 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2818 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2819 ump->softdep_jblocks = jblocks; 2820 out: 2821 if (error == 0) { 2822 MNT_ILOCK(mp); 2823 mp->mnt_flag |= MNT_SUJ; 2824 mp->mnt_flag &= ~MNT_SOFTDEP; 2825 MNT_IUNLOCK(mp); 2826 /* 2827 * Only validate the journal contents if the 2828 * filesystem is clean, otherwise we write the logs 2829 * but they'll never be used. If the filesystem was 2830 * still dirty when we mounted it the journal is 2831 * invalid and a new journal can only be valid if it 2832 * starts from a clean mount. 2833 */ 2834 if (fs->fs_clean) { 2835 DIP_SET(ip, i_modrev, fs->fs_mtime); 2836 ip->i_flags |= IN_MODIFIED; 2837 ffs_update(vp, 1); 2838 } 2839 } 2840 vput(vp); 2841 return (error); 2842 } 2843 2844 static void 2845 journal_unmount(ump) 2846 struct ufsmount *ump; 2847 { 2848 2849 if (ump->softdep_jblocks) 2850 jblocks_destroy(ump->softdep_jblocks); 2851 ump->softdep_jblocks = NULL; 2852 } 2853 2854 /* 2855 * Called when a journal record is ready to be written. Space is allocated 2856 * and the journal entry is created when the journal is flushed to stable 2857 * store. 2858 */ 2859 static void 2860 add_to_journal(wk) 2861 struct worklist *wk; 2862 { 2863 struct ufsmount *ump; 2864 2865 ump = VFSTOUFS(wk->wk_mp); 2866 LOCK_OWNED(ump); 2867 if (wk->wk_state & ONWORKLIST) 2868 panic("add_to_journal: %s(0x%X) already on list", 2869 TYPENAME(wk->wk_type), wk->wk_state); 2870 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2871 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2872 ump->softdep_jblocks->jb_age = ticks; 2873 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2874 } else 2875 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2876 ump->softdep_journal_tail = wk; 2877 ump->softdep_on_journal += 1; 2878 } 2879 2880 /* 2881 * Remove an arbitrary item for the journal worklist maintain the tail 2882 * pointer. This happens when a new operation obviates the need to 2883 * journal an old operation. 2884 */ 2885 static void 2886 remove_from_journal(wk) 2887 struct worklist *wk; 2888 { 2889 struct ufsmount *ump; 2890 2891 ump = VFSTOUFS(wk->wk_mp); 2892 LOCK_OWNED(ump); 2893 #ifdef INVARIANTS 2894 { 2895 struct worklist *wkn; 2896 2897 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2898 if (wkn == wk) 2899 break; 2900 if (wkn == NULL) 2901 panic("remove_from_journal: %p is not in journal", wk); 2902 } 2903 #endif 2904 /* 2905 * We emulate a TAILQ to save space in most structures which do not 2906 * require TAILQ semantics. Here we must update the tail position 2907 * when removing the tail which is not the final entry. This works 2908 * only if the worklist linkage are at the beginning of the structure. 2909 */ 2910 if (ump->softdep_journal_tail == wk) 2911 ump->softdep_journal_tail = 2912 (struct worklist *)wk->wk_list.le_prev; 2913 WORKLIST_REMOVE(wk); 2914 ump->softdep_on_journal -= 1; 2915 } 2916 2917 /* 2918 * Check for journal space as well as dependency limits so the prelink 2919 * code can throttle both journaled and non-journaled filesystems. 2920 * Threshold is 0 for low and 1 for min. 2921 */ 2922 static int 2923 journal_space(ump, thresh) 2924 struct ufsmount *ump; 2925 int thresh; 2926 { 2927 struct jblocks *jblocks; 2928 int limit, avail; 2929 2930 jblocks = ump->softdep_jblocks; 2931 if (jblocks == NULL) 2932 return (1); 2933 /* 2934 * We use a tighter restriction here to prevent request_cleanup() 2935 * running in threads from running into locks we currently hold. 2936 * We have to be over the limit and our filesystem has to be 2937 * responsible for more than our share of that usage. 2938 */ 2939 limit = (max_softdeps / 10) * 9; 2940 if (dep_current[D_INODEDEP] > limit && 2941 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2942 return (0); 2943 if (thresh) 2944 thresh = jblocks->jb_min; 2945 else 2946 thresh = jblocks->jb_low; 2947 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2948 avail = jblocks->jb_free - avail; 2949 2950 return (avail > thresh); 2951 } 2952 2953 static void 2954 journal_suspend(ump) 2955 struct ufsmount *ump; 2956 { 2957 struct jblocks *jblocks; 2958 struct mount *mp; 2959 2960 mp = UFSTOVFS(ump); 2961 jblocks = ump->softdep_jblocks; 2962 MNT_ILOCK(mp); 2963 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2964 stat_journal_min++; 2965 mp->mnt_kern_flag |= MNTK_SUSPEND; 2966 mp->mnt_susp_owner = ump->softdep_flushtd; 2967 } 2968 jblocks->jb_suspended = 1; 2969 MNT_IUNLOCK(mp); 2970 } 2971 2972 static int 2973 journal_unsuspend(struct ufsmount *ump) 2974 { 2975 struct jblocks *jblocks; 2976 struct mount *mp; 2977 2978 mp = UFSTOVFS(ump); 2979 jblocks = ump->softdep_jblocks; 2980 2981 if (jblocks != NULL && jblocks->jb_suspended && 2982 journal_space(ump, jblocks->jb_min)) { 2983 jblocks->jb_suspended = 0; 2984 FREE_LOCK(ump); 2985 mp->mnt_susp_owner = curthread; 2986 vfs_write_resume(mp, 0); 2987 ACQUIRE_LOCK(ump); 2988 return (1); 2989 } 2990 return (0); 2991 } 2992 2993 /* 2994 * Called before any allocation function to be certain that there is 2995 * sufficient space in the journal prior to creating any new records. 2996 * Since in the case of block allocation we may have multiple locked 2997 * buffers at the time of the actual allocation we can not block 2998 * when the journal records are created. Doing so would create a deadlock 2999 * if any of these buffers needed to be flushed to reclaim space. Instead 3000 * we require a sufficiently large amount of available space such that 3001 * each thread in the system could have passed this allocation check and 3002 * still have sufficient free space. With 20% of a minimum journal size 3003 * of 1MB we have 6553 records available. 3004 */ 3005 int 3006 softdep_prealloc(vp, waitok) 3007 struct vnode *vp; 3008 int waitok; 3009 { 3010 struct ufsmount *ump; 3011 3012 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 3013 ("softdep_prealloc called on non-softdep filesystem")); 3014 /* 3015 * Nothing to do if we are not running journaled soft updates. 3016 * If we currently hold the snapshot lock, we must avoid 3017 * handling other resources that could cause deadlock. Do not 3018 * touch quotas vnode since it is typically recursed with 3019 * other vnode locks held. 3020 */ 3021 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) || 3022 (vp->v_vflag & VV_SYSTEM) != 0) 3023 return (0); 3024 ump = VFSTOUFS(vp->v_mount); 3025 ACQUIRE_LOCK(ump); 3026 if (journal_space(ump, 0)) { 3027 FREE_LOCK(ump); 3028 return (0); 3029 } 3030 stat_journal_low++; 3031 FREE_LOCK(ump); 3032 if (waitok == MNT_NOWAIT) 3033 return (ENOSPC); 3034 /* 3035 * Attempt to sync this vnode once to flush any journal 3036 * work attached to it. 3037 */ 3038 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3039 ffs_syncvnode(vp, waitok, 0); 3040 ACQUIRE_LOCK(ump); 3041 process_removes(vp); 3042 process_truncates(vp); 3043 if (journal_space(ump, 0) == 0) { 3044 softdep_speedup(ump); 3045 if (journal_space(ump, 1) == 0) 3046 journal_suspend(ump); 3047 } 3048 FREE_LOCK(ump); 3049 3050 return (0); 3051 } 3052 3053 /* 3054 * Before adjusting a link count on a vnode verify that we have sufficient 3055 * journal space. If not, process operations that depend on the currently 3056 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3057 * and softdep flush threads can not acquire these locks to reclaim space. 3058 */ 3059 static void 3060 softdep_prelink(dvp, vp) 3061 struct vnode *dvp; 3062 struct vnode *vp; 3063 { 3064 struct ufsmount *ump; 3065 3066 ump = VFSTOUFS(dvp->v_mount); 3067 LOCK_OWNED(ump); 3068 /* 3069 * Nothing to do if we have sufficient journal space. 3070 * If we currently hold the snapshot lock, we must avoid 3071 * handling other resources that could cause deadlock. 3072 */ 3073 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3074 return; 3075 stat_journal_low++; 3076 FREE_LOCK(ump); 3077 if (vp) 3078 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3079 ffs_syncvnode(dvp, MNT_WAIT, 0); 3080 ACQUIRE_LOCK(ump); 3081 /* Process vp before dvp as it may create .. removes. */ 3082 if (vp) { 3083 process_removes(vp); 3084 process_truncates(vp); 3085 } 3086 process_removes(dvp); 3087 process_truncates(dvp); 3088 softdep_speedup(ump); 3089 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3090 if (journal_space(ump, 0) == 0) { 3091 softdep_speedup(ump); 3092 if (journal_space(ump, 1) == 0) 3093 journal_suspend(ump); 3094 } 3095 } 3096 3097 static void 3098 jseg_write(ump, jseg, data) 3099 struct ufsmount *ump; 3100 struct jseg *jseg; 3101 uint8_t *data; 3102 { 3103 struct jsegrec *rec; 3104 3105 rec = (struct jsegrec *)data; 3106 rec->jsr_seq = jseg->js_seq; 3107 rec->jsr_oldest = jseg->js_oldseq; 3108 rec->jsr_cnt = jseg->js_cnt; 3109 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3110 rec->jsr_crc = 0; 3111 rec->jsr_time = ump->um_fs->fs_mtime; 3112 } 3113 3114 static inline void 3115 inoref_write(inoref, jseg, rec) 3116 struct inoref *inoref; 3117 struct jseg *jseg; 3118 struct jrefrec *rec; 3119 { 3120 3121 inoref->if_jsegdep->jd_seg = jseg; 3122 rec->jr_ino = inoref->if_ino; 3123 rec->jr_parent = inoref->if_parent; 3124 rec->jr_nlink = inoref->if_nlink; 3125 rec->jr_mode = inoref->if_mode; 3126 rec->jr_diroff = inoref->if_diroff; 3127 } 3128 3129 static void 3130 jaddref_write(jaddref, jseg, data) 3131 struct jaddref *jaddref; 3132 struct jseg *jseg; 3133 uint8_t *data; 3134 { 3135 struct jrefrec *rec; 3136 3137 rec = (struct jrefrec *)data; 3138 rec->jr_op = JOP_ADDREF; 3139 inoref_write(&jaddref->ja_ref, jseg, rec); 3140 } 3141 3142 static void 3143 jremref_write(jremref, jseg, data) 3144 struct jremref *jremref; 3145 struct jseg *jseg; 3146 uint8_t *data; 3147 { 3148 struct jrefrec *rec; 3149 3150 rec = (struct jrefrec *)data; 3151 rec->jr_op = JOP_REMREF; 3152 inoref_write(&jremref->jr_ref, jseg, rec); 3153 } 3154 3155 static void 3156 jmvref_write(jmvref, jseg, data) 3157 struct jmvref *jmvref; 3158 struct jseg *jseg; 3159 uint8_t *data; 3160 { 3161 struct jmvrec *rec; 3162 3163 rec = (struct jmvrec *)data; 3164 rec->jm_op = JOP_MVREF; 3165 rec->jm_ino = jmvref->jm_ino; 3166 rec->jm_parent = jmvref->jm_parent; 3167 rec->jm_oldoff = jmvref->jm_oldoff; 3168 rec->jm_newoff = jmvref->jm_newoff; 3169 } 3170 3171 static void 3172 jnewblk_write(jnewblk, jseg, data) 3173 struct jnewblk *jnewblk; 3174 struct jseg *jseg; 3175 uint8_t *data; 3176 { 3177 struct jblkrec *rec; 3178 3179 jnewblk->jn_jsegdep->jd_seg = jseg; 3180 rec = (struct jblkrec *)data; 3181 rec->jb_op = JOP_NEWBLK; 3182 rec->jb_ino = jnewblk->jn_ino; 3183 rec->jb_blkno = jnewblk->jn_blkno; 3184 rec->jb_lbn = jnewblk->jn_lbn; 3185 rec->jb_frags = jnewblk->jn_frags; 3186 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3187 } 3188 3189 static void 3190 jfreeblk_write(jfreeblk, jseg, data) 3191 struct jfreeblk *jfreeblk; 3192 struct jseg *jseg; 3193 uint8_t *data; 3194 { 3195 struct jblkrec *rec; 3196 3197 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3198 rec = (struct jblkrec *)data; 3199 rec->jb_op = JOP_FREEBLK; 3200 rec->jb_ino = jfreeblk->jf_ino; 3201 rec->jb_blkno = jfreeblk->jf_blkno; 3202 rec->jb_lbn = jfreeblk->jf_lbn; 3203 rec->jb_frags = jfreeblk->jf_frags; 3204 rec->jb_oldfrags = 0; 3205 } 3206 3207 static void 3208 jfreefrag_write(jfreefrag, jseg, data) 3209 struct jfreefrag *jfreefrag; 3210 struct jseg *jseg; 3211 uint8_t *data; 3212 { 3213 struct jblkrec *rec; 3214 3215 jfreefrag->fr_jsegdep->jd_seg = jseg; 3216 rec = (struct jblkrec *)data; 3217 rec->jb_op = JOP_FREEBLK; 3218 rec->jb_ino = jfreefrag->fr_ino; 3219 rec->jb_blkno = jfreefrag->fr_blkno; 3220 rec->jb_lbn = jfreefrag->fr_lbn; 3221 rec->jb_frags = jfreefrag->fr_frags; 3222 rec->jb_oldfrags = 0; 3223 } 3224 3225 static void 3226 jtrunc_write(jtrunc, jseg, data) 3227 struct jtrunc *jtrunc; 3228 struct jseg *jseg; 3229 uint8_t *data; 3230 { 3231 struct jtrncrec *rec; 3232 3233 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3234 rec = (struct jtrncrec *)data; 3235 rec->jt_op = JOP_TRUNC; 3236 rec->jt_ino = jtrunc->jt_ino; 3237 rec->jt_size = jtrunc->jt_size; 3238 rec->jt_extsize = jtrunc->jt_extsize; 3239 } 3240 3241 static void 3242 jfsync_write(jfsync, jseg, data) 3243 struct jfsync *jfsync; 3244 struct jseg *jseg; 3245 uint8_t *data; 3246 { 3247 struct jtrncrec *rec; 3248 3249 rec = (struct jtrncrec *)data; 3250 rec->jt_op = JOP_SYNC; 3251 rec->jt_ino = jfsync->jfs_ino; 3252 rec->jt_size = jfsync->jfs_size; 3253 rec->jt_extsize = jfsync->jfs_extsize; 3254 } 3255 3256 static void 3257 softdep_flushjournal(mp) 3258 struct mount *mp; 3259 { 3260 struct jblocks *jblocks; 3261 struct ufsmount *ump; 3262 3263 if (MOUNTEDSUJ(mp) == 0) 3264 return; 3265 ump = VFSTOUFS(mp); 3266 jblocks = ump->softdep_jblocks; 3267 ACQUIRE_LOCK(ump); 3268 while (ump->softdep_on_journal) { 3269 jblocks->jb_needseg = 1; 3270 softdep_process_journal(mp, NULL, MNT_WAIT); 3271 } 3272 FREE_LOCK(ump); 3273 } 3274 3275 static void softdep_synchronize_completed(struct bio *); 3276 static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3277 3278 static void 3279 softdep_synchronize_completed(bp) 3280 struct bio *bp; 3281 { 3282 struct jseg *oldest; 3283 struct jseg *jseg; 3284 struct ufsmount *ump; 3285 3286 /* 3287 * caller1 marks the last segment written before we issued the 3288 * synchronize cache. 3289 */ 3290 jseg = bp->bio_caller1; 3291 if (jseg == NULL) { 3292 g_destroy_bio(bp); 3293 return; 3294 } 3295 ump = VFSTOUFS(jseg->js_list.wk_mp); 3296 ACQUIRE_LOCK(ump); 3297 oldest = NULL; 3298 /* 3299 * Mark all the journal entries waiting on the synchronize cache 3300 * as completed so they may continue on. 3301 */ 3302 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3303 jseg->js_state |= COMPLETE; 3304 oldest = jseg; 3305 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3306 } 3307 /* 3308 * Restart deferred journal entry processing from the oldest 3309 * completed jseg. 3310 */ 3311 if (oldest) 3312 complete_jsegs(oldest); 3313 3314 FREE_LOCK(ump); 3315 g_destroy_bio(bp); 3316 } 3317 3318 /* 3319 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3320 * barriers. The journal must be written prior to any blocks that depend 3321 * on it and the journal can not be released until the blocks have be 3322 * written. This code handles both barriers simultaneously. 3323 */ 3324 static void 3325 softdep_synchronize(bp, ump, caller1) 3326 struct bio *bp; 3327 struct ufsmount *ump; 3328 void *caller1; 3329 { 3330 3331 bp->bio_cmd = BIO_FLUSH; 3332 bp->bio_flags |= BIO_ORDERED; 3333 bp->bio_data = NULL; 3334 bp->bio_offset = ump->um_cp->provider->mediasize; 3335 bp->bio_length = 0; 3336 bp->bio_done = softdep_synchronize_completed; 3337 bp->bio_caller1 = caller1; 3338 g_io_request(bp, 3339 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3340 } 3341 3342 /* 3343 * Flush some journal records to disk. 3344 */ 3345 static void 3346 softdep_process_journal(mp, needwk, flags) 3347 struct mount *mp; 3348 struct worklist *needwk; 3349 int flags; 3350 { 3351 struct jblocks *jblocks; 3352 struct ufsmount *ump; 3353 struct worklist *wk; 3354 struct jseg *jseg; 3355 struct buf *bp; 3356 struct bio *bio; 3357 uint8_t *data; 3358 struct fs *fs; 3359 int shouldflush; 3360 int segwritten; 3361 int jrecmin; /* Minimum records per block. */ 3362 int jrecmax; /* Maximum records per block. */ 3363 int size; 3364 int cnt; 3365 int off; 3366 int devbsize; 3367 3368 if (MOUNTEDSUJ(mp) == 0) 3369 return; 3370 shouldflush = softdep_flushcache; 3371 bio = NULL; 3372 jseg = NULL; 3373 ump = VFSTOUFS(mp); 3374 LOCK_OWNED(ump); 3375 fs = ump->um_fs; 3376 jblocks = ump->softdep_jblocks; 3377 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3378 /* 3379 * We write anywhere between a disk block and fs block. The upper 3380 * bound is picked to prevent buffer cache fragmentation and limit 3381 * processing time per I/O. 3382 */ 3383 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3384 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3385 segwritten = 0; 3386 for (;;) { 3387 cnt = ump->softdep_on_journal; 3388 /* 3389 * Criteria for writing a segment: 3390 * 1) We have a full block. 3391 * 2) We're called from jwait() and haven't found the 3392 * journal item yet. 3393 * 3) Always write if needseg is set. 3394 * 4) If we are called from process_worklist and have 3395 * not yet written anything we write a partial block 3396 * to enforce a 1 second maximum latency on journal 3397 * entries. 3398 */ 3399 if (cnt < (jrecmax - 1) && needwk == NULL && 3400 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3401 break; 3402 cnt++; 3403 /* 3404 * Verify some free journal space. softdep_prealloc() should 3405 * guarantee that we don't run out so this is indicative of 3406 * a problem with the flow control. Try to recover 3407 * gracefully in any event. 3408 */ 3409 while (jblocks->jb_free == 0) { 3410 if (flags != MNT_WAIT) 3411 break; 3412 printf("softdep: Out of journal space!\n"); 3413 softdep_speedup(ump); 3414 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3415 } 3416 FREE_LOCK(ump); 3417 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3418 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3419 LIST_INIT(&jseg->js_entries); 3420 LIST_INIT(&jseg->js_indirs); 3421 jseg->js_state = ATTACHED; 3422 if (shouldflush == 0) 3423 jseg->js_state |= COMPLETE; 3424 else if (bio == NULL) 3425 bio = g_alloc_bio(); 3426 jseg->js_jblocks = jblocks; 3427 bp = geteblk(fs->fs_bsize, 0); 3428 ACQUIRE_LOCK(ump); 3429 /* 3430 * If there was a race while we were allocating the block 3431 * and jseg the entry we care about was likely written. 3432 * We bail out in both the WAIT and NOWAIT case and assume 3433 * the caller will loop if the entry it cares about is 3434 * not written. 3435 */ 3436 cnt = ump->softdep_on_journal; 3437 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3438 bp->b_flags |= B_INVAL | B_NOCACHE; 3439 WORKITEM_FREE(jseg, D_JSEG); 3440 FREE_LOCK(ump); 3441 brelse(bp); 3442 ACQUIRE_LOCK(ump); 3443 break; 3444 } 3445 /* 3446 * Calculate the disk block size required for the available 3447 * records rounded to the min size. 3448 */ 3449 if (cnt == 0) 3450 size = devbsize; 3451 else if (cnt < jrecmax) 3452 size = howmany(cnt, jrecmin) * devbsize; 3453 else 3454 size = fs->fs_bsize; 3455 /* 3456 * Allocate a disk block for this journal data and account 3457 * for truncation of the requested size if enough contiguous 3458 * space was not available. 3459 */ 3460 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3461 bp->b_lblkno = bp->b_blkno; 3462 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3463 bp->b_bcount = size; 3464 bp->b_flags &= ~B_INVAL; 3465 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3466 /* 3467 * Initialize our jseg with cnt records. Assign the next 3468 * sequence number to it and link it in-order. 3469 */ 3470 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3471 jseg->js_buf = bp; 3472 jseg->js_cnt = cnt; 3473 jseg->js_refs = cnt + 1; /* Self ref. */ 3474 jseg->js_size = size; 3475 jseg->js_seq = jblocks->jb_nextseq++; 3476 if (jblocks->jb_oldestseg == NULL) 3477 jblocks->jb_oldestseg = jseg; 3478 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3479 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3480 if (jblocks->jb_writeseg == NULL) 3481 jblocks->jb_writeseg = jseg; 3482 /* 3483 * Start filling in records from the pending list. 3484 */ 3485 data = bp->b_data; 3486 off = 0; 3487 3488 /* 3489 * Always put a header on the first block. 3490 * XXX As with below, there might not be a chance to get 3491 * into the loop. Ensure that something valid is written. 3492 */ 3493 jseg_write(ump, jseg, data); 3494 off += JREC_SIZE; 3495 data = bp->b_data + off; 3496 3497 /* 3498 * XXX Something is wrong here. There's no work to do, 3499 * but we need to perform and I/O and allow it to complete 3500 * anyways. 3501 */ 3502 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3503 stat_emptyjblocks++; 3504 3505 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3506 != NULL) { 3507 if (cnt == 0) 3508 break; 3509 /* Place a segment header on every device block. */ 3510 if ((off % devbsize) == 0) { 3511 jseg_write(ump, jseg, data); 3512 off += JREC_SIZE; 3513 data = bp->b_data + off; 3514 } 3515 if (wk == needwk) 3516 needwk = NULL; 3517 remove_from_journal(wk); 3518 wk->wk_state |= INPROGRESS; 3519 WORKLIST_INSERT(&jseg->js_entries, wk); 3520 switch (wk->wk_type) { 3521 case D_JADDREF: 3522 jaddref_write(WK_JADDREF(wk), jseg, data); 3523 break; 3524 case D_JREMREF: 3525 jremref_write(WK_JREMREF(wk), jseg, data); 3526 break; 3527 case D_JMVREF: 3528 jmvref_write(WK_JMVREF(wk), jseg, data); 3529 break; 3530 case D_JNEWBLK: 3531 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3532 break; 3533 case D_JFREEBLK: 3534 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3535 break; 3536 case D_JFREEFRAG: 3537 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3538 break; 3539 case D_JTRUNC: 3540 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3541 break; 3542 case D_JFSYNC: 3543 jfsync_write(WK_JFSYNC(wk), jseg, data); 3544 break; 3545 default: 3546 panic("process_journal: Unknown type %s", 3547 TYPENAME(wk->wk_type)); 3548 /* NOTREACHED */ 3549 } 3550 off += JREC_SIZE; 3551 data = bp->b_data + off; 3552 cnt--; 3553 } 3554 3555 /* Clear any remaining space so we don't leak kernel data */ 3556 if (size > off) 3557 bzero(data, size - off); 3558 3559 /* 3560 * Write this one buffer and continue. 3561 */ 3562 segwritten = 1; 3563 jblocks->jb_needseg = 0; 3564 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3565 FREE_LOCK(ump); 3566 pbgetvp(ump->um_devvp, bp); 3567 /* 3568 * We only do the blocking wait once we find the journal 3569 * entry we're looking for. 3570 */ 3571 if (needwk == NULL && flags == MNT_WAIT) 3572 bwrite(bp); 3573 else 3574 bawrite(bp); 3575 ACQUIRE_LOCK(ump); 3576 } 3577 /* 3578 * If we wrote a segment issue a synchronize cache so the journal 3579 * is reflected on disk before the data is written. Since reclaiming 3580 * journal space also requires writing a journal record this 3581 * process also enforces a barrier before reclamation. 3582 */ 3583 if (segwritten && shouldflush) { 3584 softdep_synchronize(bio, ump, 3585 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3586 } else if (bio) 3587 g_destroy_bio(bio); 3588 /* 3589 * If we've suspended the filesystem because we ran out of journal 3590 * space either try to sync it here to make some progress or 3591 * unsuspend it if we already have. 3592 */ 3593 if (flags == 0 && jblocks->jb_suspended) { 3594 if (journal_unsuspend(ump)) 3595 return; 3596 FREE_LOCK(ump); 3597 VFS_SYNC(mp, MNT_NOWAIT); 3598 ffs_sbupdate(ump, MNT_WAIT, 0); 3599 ACQUIRE_LOCK(ump); 3600 } 3601 } 3602 3603 /* 3604 * Complete a jseg, allowing all dependencies awaiting journal writes 3605 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3606 * structures so that the journal segment can be freed to reclaim space. 3607 */ 3608 static void 3609 complete_jseg(jseg) 3610 struct jseg *jseg; 3611 { 3612 struct worklist *wk; 3613 struct jmvref *jmvref; 3614 #ifdef INVARIANTS 3615 int i = 0; 3616 #endif 3617 3618 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3619 WORKLIST_REMOVE(wk); 3620 wk->wk_state &= ~INPROGRESS; 3621 wk->wk_state |= COMPLETE; 3622 KASSERT(i++ < jseg->js_cnt, 3623 ("handle_written_jseg: overflow %d >= %d", 3624 i - 1, jseg->js_cnt)); 3625 switch (wk->wk_type) { 3626 case D_JADDREF: 3627 handle_written_jaddref(WK_JADDREF(wk)); 3628 break; 3629 case D_JREMREF: 3630 handle_written_jremref(WK_JREMREF(wk)); 3631 break; 3632 case D_JMVREF: 3633 rele_jseg(jseg); /* No jsegdep. */ 3634 jmvref = WK_JMVREF(wk); 3635 LIST_REMOVE(jmvref, jm_deps); 3636 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3637 free_pagedep(jmvref->jm_pagedep); 3638 WORKITEM_FREE(jmvref, D_JMVREF); 3639 break; 3640 case D_JNEWBLK: 3641 handle_written_jnewblk(WK_JNEWBLK(wk)); 3642 break; 3643 case D_JFREEBLK: 3644 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3645 break; 3646 case D_JTRUNC: 3647 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3648 break; 3649 case D_JFSYNC: 3650 rele_jseg(jseg); /* No jsegdep. */ 3651 WORKITEM_FREE(wk, D_JFSYNC); 3652 break; 3653 case D_JFREEFRAG: 3654 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3655 break; 3656 default: 3657 panic("handle_written_jseg: Unknown type %s", 3658 TYPENAME(wk->wk_type)); 3659 /* NOTREACHED */ 3660 } 3661 } 3662 /* Release the self reference so the structure may be freed. */ 3663 rele_jseg(jseg); 3664 } 3665 3666 /* 3667 * Determine which jsegs are ready for completion processing. Waits for 3668 * synchronize cache to complete as well as forcing in-order completion 3669 * of journal entries. 3670 */ 3671 static void 3672 complete_jsegs(jseg) 3673 struct jseg *jseg; 3674 { 3675 struct jblocks *jblocks; 3676 struct jseg *jsegn; 3677 3678 jblocks = jseg->js_jblocks; 3679 /* 3680 * Don't allow out of order completions. If this isn't the first 3681 * block wait for it to write before we're done. 3682 */ 3683 if (jseg != jblocks->jb_writeseg) 3684 return; 3685 /* Iterate through available jsegs processing their entries. */ 3686 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3687 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3688 jsegn = TAILQ_NEXT(jseg, js_next); 3689 complete_jseg(jseg); 3690 jseg = jsegn; 3691 } 3692 jblocks->jb_writeseg = jseg; 3693 /* 3694 * Attempt to free jsegs now that oldestwrseq may have advanced. 3695 */ 3696 free_jsegs(jblocks); 3697 } 3698 3699 /* 3700 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3701 * the final completions. 3702 */ 3703 static void 3704 handle_written_jseg(jseg, bp) 3705 struct jseg *jseg; 3706 struct buf *bp; 3707 { 3708 3709 if (jseg->js_refs == 0) 3710 panic("handle_written_jseg: No self-reference on %p", jseg); 3711 jseg->js_state |= DEPCOMPLETE; 3712 /* 3713 * We'll never need this buffer again, set flags so it will be 3714 * discarded. 3715 */ 3716 bp->b_flags |= B_INVAL | B_NOCACHE; 3717 pbrelvp(bp); 3718 complete_jsegs(jseg); 3719 } 3720 3721 static inline struct jsegdep * 3722 inoref_jseg(inoref) 3723 struct inoref *inoref; 3724 { 3725 struct jsegdep *jsegdep; 3726 3727 jsegdep = inoref->if_jsegdep; 3728 inoref->if_jsegdep = NULL; 3729 3730 return (jsegdep); 3731 } 3732 3733 /* 3734 * Called once a jremref has made it to stable store. The jremref is marked 3735 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3736 * for the jremref to complete will be awoken by free_jremref. 3737 */ 3738 static void 3739 handle_written_jremref(jremref) 3740 struct jremref *jremref; 3741 { 3742 struct inodedep *inodedep; 3743 struct jsegdep *jsegdep; 3744 struct dirrem *dirrem; 3745 3746 /* Grab the jsegdep. */ 3747 jsegdep = inoref_jseg(&jremref->jr_ref); 3748 /* 3749 * Remove us from the inoref list. 3750 */ 3751 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3752 0, &inodedep) == 0) 3753 panic("handle_written_jremref: Lost inodedep"); 3754 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3755 /* 3756 * Complete the dirrem. 3757 */ 3758 dirrem = jremref->jr_dirrem; 3759 jremref->jr_dirrem = NULL; 3760 LIST_REMOVE(jremref, jr_deps); 3761 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3762 jwork_insert(&dirrem->dm_jwork, jsegdep); 3763 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3764 (dirrem->dm_state & COMPLETE) != 0) 3765 add_to_worklist(&dirrem->dm_list, 0); 3766 free_jremref(jremref); 3767 } 3768 3769 /* 3770 * Called once a jaddref has made it to stable store. The dependency is 3771 * marked complete and any dependent structures are added to the inode 3772 * bufwait list to be completed as soon as it is written. If a bitmap write 3773 * depends on this entry we move the inode into the inodedephd of the 3774 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3775 */ 3776 static void 3777 handle_written_jaddref(jaddref) 3778 struct jaddref *jaddref; 3779 { 3780 struct jsegdep *jsegdep; 3781 struct inodedep *inodedep; 3782 struct diradd *diradd; 3783 struct mkdir *mkdir; 3784 3785 /* Grab the jsegdep. */ 3786 jsegdep = inoref_jseg(&jaddref->ja_ref); 3787 mkdir = NULL; 3788 diradd = NULL; 3789 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3790 0, &inodedep) == 0) 3791 panic("handle_written_jaddref: Lost inodedep."); 3792 if (jaddref->ja_diradd == NULL) 3793 panic("handle_written_jaddref: No dependency"); 3794 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3795 diradd = jaddref->ja_diradd; 3796 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3797 } else if (jaddref->ja_state & MKDIR_PARENT) { 3798 mkdir = jaddref->ja_mkdir; 3799 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3800 } else if (jaddref->ja_state & MKDIR_BODY) 3801 mkdir = jaddref->ja_mkdir; 3802 else 3803 panic("handle_written_jaddref: Unknown dependency %p", 3804 jaddref->ja_diradd); 3805 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3806 /* 3807 * Remove us from the inode list. 3808 */ 3809 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3810 /* 3811 * The mkdir may be waiting on the jaddref to clear before freeing. 3812 */ 3813 if (mkdir) { 3814 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3815 ("handle_written_jaddref: Incorrect type for mkdir %s", 3816 TYPENAME(mkdir->md_list.wk_type))); 3817 mkdir->md_jaddref = NULL; 3818 diradd = mkdir->md_diradd; 3819 mkdir->md_state |= DEPCOMPLETE; 3820 complete_mkdir(mkdir); 3821 } 3822 jwork_insert(&diradd->da_jwork, jsegdep); 3823 if (jaddref->ja_state & NEWBLOCK) { 3824 inodedep->id_state |= ONDEPLIST; 3825 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3826 inodedep, id_deps); 3827 } 3828 free_jaddref(jaddref); 3829 } 3830 3831 /* 3832 * Called once a jnewblk journal is written. The allocdirect or allocindir 3833 * is placed in the bmsafemap to await notification of a written bitmap. If 3834 * the operation was canceled we add the segdep to the appropriate 3835 * dependency to free the journal space once the canceling operation 3836 * completes. 3837 */ 3838 static void 3839 handle_written_jnewblk(jnewblk) 3840 struct jnewblk *jnewblk; 3841 { 3842 struct bmsafemap *bmsafemap; 3843 struct freefrag *freefrag; 3844 struct freework *freework; 3845 struct jsegdep *jsegdep; 3846 struct newblk *newblk; 3847 3848 /* Grab the jsegdep. */ 3849 jsegdep = jnewblk->jn_jsegdep; 3850 jnewblk->jn_jsegdep = NULL; 3851 if (jnewblk->jn_dep == NULL) 3852 panic("handle_written_jnewblk: No dependency for the segdep."); 3853 switch (jnewblk->jn_dep->wk_type) { 3854 case D_NEWBLK: 3855 case D_ALLOCDIRECT: 3856 case D_ALLOCINDIR: 3857 /* 3858 * Add the written block to the bmsafemap so it can 3859 * be notified when the bitmap is on disk. 3860 */ 3861 newblk = WK_NEWBLK(jnewblk->jn_dep); 3862 newblk->nb_jnewblk = NULL; 3863 if ((newblk->nb_state & GOINGAWAY) == 0) { 3864 bmsafemap = newblk->nb_bmsafemap; 3865 newblk->nb_state |= ONDEPLIST; 3866 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3867 nb_deps); 3868 } 3869 jwork_insert(&newblk->nb_jwork, jsegdep); 3870 break; 3871 case D_FREEFRAG: 3872 /* 3873 * A newblock being removed by a freefrag when replaced by 3874 * frag extension. 3875 */ 3876 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3877 freefrag->ff_jdep = NULL; 3878 jwork_insert(&freefrag->ff_jwork, jsegdep); 3879 break; 3880 case D_FREEWORK: 3881 /* 3882 * A direct block was removed by truncate. 3883 */ 3884 freework = WK_FREEWORK(jnewblk->jn_dep); 3885 freework->fw_jnewblk = NULL; 3886 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3887 break; 3888 default: 3889 panic("handle_written_jnewblk: Unknown type %d.", 3890 jnewblk->jn_dep->wk_type); 3891 } 3892 jnewblk->jn_dep = NULL; 3893 free_jnewblk(jnewblk); 3894 } 3895 3896 /* 3897 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3898 * an in-flight allocation that has not yet been committed. Divorce us 3899 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3900 * to the worklist. 3901 */ 3902 static void 3903 cancel_jfreefrag(jfreefrag) 3904 struct jfreefrag *jfreefrag; 3905 { 3906 struct freefrag *freefrag; 3907 3908 if (jfreefrag->fr_jsegdep) { 3909 free_jsegdep(jfreefrag->fr_jsegdep); 3910 jfreefrag->fr_jsegdep = NULL; 3911 } 3912 freefrag = jfreefrag->fr_freefrag; 3913 jfreefrag->fr_freefrag = NULL; 3914 free_jfreefrag(jfreefrag); 3915 freefrag->ff_state |= DEPCOMPLETE; 3916 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3917 } 3918 3919 /* 3920 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3921 */ 3922 static void 3923 free_jfreefrag(jfreefrag) 3924 struct jfreefrag *jfreefrag; 3925 { 3926 3927 if (jfreefrag->fr_state & INPROGRESS) 3928 WORKLIST_REMOVE(&jfreefrag->fr_list); 3929 else if (jfreefrag->fr_state & ONWORKLIST) 3930 remove_from_journal(&jfreefrag->fr_list); 3931 if (jfreefrag->fr_freefrag != NULL) 3932 panic("free_jfreefrag: Still attached to a freefrag."); 3933 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3934 } 3935 3936 /* 3937 * Called when the journal write for a jfreefrag completes. The parent 3938 * freefrag is added to the worklist if this completes its dependencies. 3939 */ 3940 static void 3941 handle_written_jfreefrag(jfreefrag) 3942 struct jfreefrag *jfreefrag; 3943 { 3944 struct jsegdep *jsegdep; 3945 struct freefrag *freefrag; 3946 3947 /* Grab the jsegdep. */ 3948 jsegdep = jfreefrag->fr_jsegdep; 3949 jfreefrag->fr_jsegdep = NULL; 3950 freefrag = jfreefrag->fr_freefrag; 3951 if (freefrag == NULL) 3952 panic("handle_written_jfreefrag: No freefrag."); 3953 freefrag->ff_state |= DEPCOMPLETE; 3954 freefrag->ff_jdep = NULL; 3955 jwork_insert(&freefrag->ff_jwork, jsegdep); 3956 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3957 add_to_worklist(&freefrag->ff_list, 0); 3958 jfreefrag->fr_freefrag = NULL; 3959 free_jfreefrag(jfreefrag); 3960 } 3961 3962 /* 3963 * Called when the journal write for a jfreeblk completes. The jfreeblk 3964 * is removed from the freeblks list of pending journal writes and the 3965 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3966 * have been reclaimed. 3967 */ 3968 static void 3969 handle_written_jblkdep(jblkdep) 3970 struct jblkdep *jblkdep; 3971 { 3972 struct freeblks *freeblks; 3973 struct jsegdep *jsegdep; 3974 3975 /* Grab the jsegdep. */ 3976 jsegdep = jblkdep->jb_jsegdep; 3977 jblkdep->jb_jsegdep = NULL; 3978 freeblks = jblkdep->jb_freeblks; 3979 LIST_REMOVE(jblkdep, jb_deps); 3980 jwork_insert(&freeblks->fb_jwork, jsegdep); 3981 /* 3982 * If the freeblks is all journaled, we can add it to the worklist. 3983 */ 3984 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3985 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3986 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3987 3988 free_jblkdep(jblkdep); 3989 } 3990 3991 static struct jsegdep * 3992 newjsegdep(struct worklist *wk) 3993 { 3994 struct jsegdep *jsegdep; 3995 3996 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3997 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3998 jsegdep->jd_seg = NULL; 3999 4000 return (jsegdep); 4001 } 4002 4003 static struct jmvref * 4004 newjmvref(dp, ino, oldoff, newoff) 4005 struct inode *dp; 4006 ino_t ino; 4007 off_t oldoff; 4008 off_t newoff; 4009 { 4010 struct jmvref *jmvref; 4011 4012 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 4013 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp)); 4014 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 4015 jmvref->jm_parent = dp->i_number; 4016 jmvref->jm_ino = ino; 4017 jmvref->jm_oldoff = oldoff; 4018 jmvref->jm_newoff = newoff; 4019 4020 return (jmvref); 4021 } 4022 4023 /* 4024 * Allocate a new jremref that tracks the removal of ip from dp with the 4025 * directory entry offset of diroff. Mark the entry as ATTACHED and 4026 * DEPCOMPLETE as we have all the information required for the journal write 4027 * and the directory has already been removed from the buffer. The caller 4028 * is responsible for linking the jremref into the pagedep and adding it 4029 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4030 * a DOTDOT addition so handle_workitem_remove() can properly assign 4031 * the jsegdep when we're done. 4032 */ 4033 static struct jremref * 4034 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4035 off_t diroff, nlink_t nlink) 4036 { 4037 struct jremref *jremref; 4038 4039 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4040 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp)); 4041 jremref->jr_state = ATTACHED; 4042 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4043 nlink, ip->i_mode); 4044 jremref->jr_dirrem = dirrem; 4045 4046 return (jremref); 4047 } 4048 4049 static inline void 4050 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4051 nlink_t nlink, uint16_t mode) 4052 { 4053 4054 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4055 inoref->if_diroff = diroff; 4056 inoref->if_ino = ino; 4057 inoref->if_parent = parent; 4058 inoref->if_nlink = nlink; 4059 inoref->if_mode = mode; 4060 } 4061 4062 /* 4063 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4064 * directory offset may not be known until later. The caller is responsible 4065 * adding the entry to the journal when this information is available. nlink 4066 * should be the link count prior to the addition and mode is only required 4067 * to have the correct FMT. 4068 */ 4069 static struct jaddref * 4070 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4071 uint16_t mode) 4072 { 4073 struct jaddref *jaddref; 4074 4075 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4076 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp)); 4077 jaddref->ja_state = ATTACHED; 4078 jaddref->ja_mkdir = NULL; 4079 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4080 4081 return (jaddref); 4082 } 4083 4084 /* 4085 * Create a new free dependency for a freework. The caller is responsible 4086 * for adjusting the reference count when it has the lock held. The freedep 4087 * will track an outstanding bitmap write that will ultimately clear the 4088 * freework to continue. 4089 */ 4090 static struct freedep * 4091 newfreedep(struct freework *freework) 4092 { 4093 struct freedep *freedep; 4094 4095 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4096 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4097 freedep->fd_freework = freework; 4098 4099 return (freedep); 4100 } 4101 4102 /* 4103 * Free a freedep structure once the buffer it is linked to is written. If 4104 * this is the last reference to the freework schedule it for completion. 4105 */ 4106 static void 4107 free_freedep(freedep) 4108 struct freedep *freedep; 4109 { 4110 struct freework *freework; 4111 4112 freework = freedep->fd_freework; 4113 freework->fw_freeblks->fb_cgwait--; 4114 if (--freework->fw_ref == 0) 4115 freework_enqueue(freework); 4116 WORKITEM_FREE(freedep, D_FREEDEP); 4117 } 4118 4119 /* 4120 * Allocate a new freework structure that may be a level in an indirect 4121 * when parent is not NULL or a top level block when it is. The top level 4122 * freework structures are allocated without the per-filesystem lock held 4123 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4124 */ 4125 static struct freework * 4126 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4127 struct ufsmount *ump; 4128 struct freeblks *freeblks; 4129 struct freework *parent; 4130 ufs_lbn_t lbn; 4131 ufs2_daddr_t nb; 4132 int frags; 4133 int off; 4134 int journal; 4135 { 4136 struct freework *freework; 4137 4138 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4139 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4140 freework->fw_state = ATTACHED; 4141 freework->fw_jnewblk = NULL; 4142 freework->fw_freeblks = freeblks; 4143 freework->fw_parent = parent; 4144 freework->fw_lbn = lbn; 4145 freework->fw_blkno = nb; 4146 freework->fw_frags = frags; 4147 freework->fw_indir = NULL; 4148 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || 4149 lbn >= -UFS_NXADDR) ? 0 : NINDIR(ump->um_fs) + 1; 4150 freework->fw_start = freework->fw_off = off; 4151 if (journal) 4152 newjfreeblk(freeblks, lbn, nb, frags); 4153 if (parent == NULL) { 4154 ACQUIRE_LOCK(ump); 4155 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4156 freeblks->fb_ref++; 4157 FREE_LOCK(ump); 4158 } 4159 4160 return (freework); 4161 } 4162 4163 /* 4164 * Eliminate a jfreeblk for a block that does not need journaling. 4165 */ 4166 static void 4167 cancel_jfreeblk(freeblks, blkno) 4168 struct freeblks *freeblks; 4169 ufs2_daddr_t blkno; 4170 { 4171 struct jfreeblk *jfreeblk; 4172 struct jblkdep *jblkdep; 4173 4174 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4175 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4176 continue; 4177 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4178 if (jfreeblk->jf_blkno == blkno) 4179 break; 4180 } 4181 if (jblkdep == NULL) 4182 return; 4183 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4184 free_jsegdep(jblkdep->jb_jsegdep); 4185 LIST_REMOVE(jblkdep, jb_deps); 4186 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4187 } 4188 4189 /* 4190 * Allocate a new jfreeblk to journal top level block pointer when truncating 4191 * a file. The caller must add this to the worklist when the per-filesystem 4192 * lock is held. 4193 */ 4194 static struct jfreeblk * 4195 newjfreeblk(freeblks, lbn, blkno, frags) 4196 struct freeblks *freeblks; 4197 ufs_lbn_t lbn; 4198 ufs2_daddr_t blkno; 4199 int frags; 4200 { 4201 struct jfreeblk *jfreeblk; 4202 4203 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4204 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4205 freeblks->fb_list.wk_mp); 4206 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4207 jfreeblk->jf_dep.jb_freeblks = freeblks; 4208 jfreeblk->jf_ino = freeblks->fb_inum; 4209 jfreeblk->jf_lbn = lbn; 4210 jfreeblk->jf_blkno = blkno; 4211 jfreeblk->jf_frags = frags; 4212 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4213 4214 return (jfreeblk); 4215 } 4216 4217 /* 4218 * The journal is only prepared to handle full-size block numbers, so we 4219 * have to adjust the record to reflect the change to a full-size block. 4220 * For example, suppose we have a block made up of fragments 8-15 and 4221 * want to free its last two fragments. We are given a request that says: 4222 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4223 * where frags are the number of fragments to free and oldfrags are the 4224 * number of fragments to keep. To block align it, we have to change it to 4225 * have a valid full-size blkno, so it becomes: 4226 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4227 */ 4228 static void 4229 adjust_newfreework(freeblks, frag_offset) 4230 struct freeblks *freeblks; 4231 int frag_offset; 4232 { 4233 struct jfreeblk *jfreeblk; 4234 4235 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4236 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4237 ("adjust_newfreework: Missing freeblks dependency")); 4238 4239 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4240 jfreeblk->jf_blkno -= frag_offset; 4241 jfreeblk->jf_frags += frag_offset; 4242 } 4243 4244 /* 4245 * Allocate a new jtrunc to track a partial truncation. 4246 */ 4247 static struct jtrunc * 4248 newjtrunc(freeblks, size, extsize) 4249 struct freeblks *freeblks; 4250 off_t size; 4251 int extsize; 4252 { 4253 struct jtrunc *jtrunc; 4254 4255 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4256 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4257 freeblks->fb_list.wk_mp); 4258 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4259 jtrunc->jt_dep.jb_freeblks = freeblks; 4260 jtrunc->jt_ino = freeblks->fb_inum; 4261 jtrunc->jt_size = size; 4262 jtrunc->jt_extsize = extsize; 4263 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4264 4265 return (jtrunc); 4266 } 4267 4268 /* 4269 * If we're canceling a new bitmap we have to search for another ref 4270 * to move into the bmsafemap dep. This might be better expressed 4271 * with another structure. 4272 */ 4273 static void 4274 move_newblock_dep(jaddref, inodedep) 4275 struct jaddref *jaddref; 4276 struct inodedep *inodedep; 4277 { 4278 struct inoref *inoref; 4279 struct jaddref *jaddrefn; 4280 4281 jaddrefn = NULL; 4282 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4283 inoref = TAILQ_NEXT(inoref, if_deps)) { 4284 if ((jaddref->ja_state & NEWBLOCK) && 4285 inoref->if_list.wk_type == D_JADDREF) { 4286 jaddrefn = (struct jaddref *)inoref; 4287 break; 4288 } 4289 } 4290 if (jaddrefn == NULL) 4291 return; 4292 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4293 jaddrefn->ja_state |= jaddref->ja_state & 4294 (ATTACHED | UNDONE | NEWBLOCK); 4295 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4296 jaddref->ja_state |= ATTACHED; 4297 LIST_REMOVE(jaddref, ja_bmdeps); 4298 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4299 ja_bmdeps); 4300 } 4301 4302 /* 4303 * Cancel a jaddref either before it has been written or while it is being 4304 * written. This happens when a link is removed before the add reaches 4305 * the disk. The jaddref dependency is kept linked into the bmsafemap 4306 * and inode to prevent the link count or bitmap from reaching the disk 4307 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4308 * required. 4309 * 4310 * Returns 1 if the canceled addref requires journaling of the remove and 4311 * 0 otherwise. 4312 */ 4313 static int 4314 cancel_jaddref(jaddref, inodedep, wkhd) 4315 struct jaddref *jaddref; 4316 struct inodedep *inodedep; 4317 struct workhead *wkhd; 4318 { 4319 struct inoref *inoref; 4320 struct jsegdep *jsegdep; 4321 int needsj; 4322 4323 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4324 ("cancel_jaddref: Canceling complete jaddref")); 4325 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4326 needsj = 1; 4327 else 4328 needsj = 0; 4329 if (inodedep == NULL) 4330 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4331 0, &inodedep) == 0) 4332 panic("cancel_jaddref: Lost inodedep"); 4333 /* 4334 * We must adjust the nlink of any reference operation that follows 4335 * us so that it is consistent with the in-memory reference. This 4336 * ensures that inode nlink rollbacks always have the correct link. 4337 */ 4338 if (needsj == 0) { 4339 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4340 inoref = TAILQ_NEXT(inoref, if_deps)) { 4341 if (inoref->if_state & GOINGAWAY) 4342 break; 4343 inoref->if_nlink--; 4344 } 4345 } 4346 jsegdep = inoref_jseg(&jaddref->ja_ref); 4347 if (jaddref->ja_state & NEWBLOCK) 4348 move_newblock_dep(jaddref, inodedep); 4349 wake_worklist(&jaddref->ja_list); 4350 jaddref->ja_mkdir = NULL; 4351 if (jaddref->ja_state & INPROGRESS) { 4352 jaddref->ja_state &= ~INPROGRESS; 4353 WORKLIST_REMOVE(&jaddref->ja_list); 4354 jwork_insert(wkhd, jsegdep); 4355 } else { 4356 free_jsegdep(jsegdep); 4357 if (jaddref->ja_state & DEPCOMPLETE) 4358 remove_from_journal(&jaddref->ja_list); 4359 } 4360 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4361 /* 4362 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4363 * can arrange for them to be freed with the bitmap. Otherwise we 4364 * no longer need this addref attached to the inoreflst and it 4365 * will incorrectly adjust nlink if we leave it. 4366 */ 4367 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4368 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4369 if_deps); 4370 jaddref->ja_state |= COMPLETE; 4371 free_jaddref(jaddref); 4372 return (needsj); 4373 } 4374 /* 4375 * Leave the head of the list for jsegdeps for fast merging. 4376 */ 4377 if (LIST_FIRST(wkhd) != NULL) { 4378 jaddref->ja_state |= ONWORKLIST; 4379 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4380 } else 4381 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4382 4383 return (needsj); 4384 } 4385 4386 /* 4387 * Attempt to free a jaddref structure when some work completes. This 4388 * should only succeed once the entry is written and all dependencies have 4389 * been notified. 4390 */ 4391 static void 4392 free_jaddref(jaddref) 4393 struct jaddref *jaddref; 4394 { 4395 4396 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4397 return; 4398 if (jaddref->ja_ref.if_jsegdep) 4399 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4400 jaddref, jaddref->ja_state); 4401 if (jaddref->ja_state & NEWBLOCK) 4402 LIST_REMOVE(jaddref, ja_bmdeps); 4403 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4404 panic("free_jaddref: Bad state %p(0x%X)", 4405 jaddref, jaddref->ja_state); 4406 if (jaddref->ja_mkdir != NULL) 4407 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4408 WORKITEM_FREE(jaddref, D_JADDREF); 4409 } 4410 4411 /* 4412 * Free a jremref structure once it has been written or discarded. 4413 */ 4414 static void 4415 free_jremref(jremref) 4416 struct jremref *jremref; 4417 { 4418 4419 if (jremref->jr_ref.if_jsegdep) 4420 free_jsegdep(jremref->jr_ref.if_jsegdep); 4421 if (jremref->jr_state & INPROGRESS) 4422 panic("free_jremref: IO still pending"); 4423 WORKITEM_FREE(jremref, D_JREMREF); 4424 } 4425 4426 /* 4427 * Free a jnewblk structure. 4428 */ 4429 static void 4430 free_jnewblk(jnewblk) 4431 struct jnewblk *jnewblk; 4432 { 4433 4434 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4435 return; 4436 LIST_REMOVE(jnewblk, jn_deps); 4437 if (jnewblk->jn_dep != NULL) 4438 panic("free_jnewblk: Dependency still attached."); 4439 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4440 } 4441 4442 /* 4443 * Cancel a jnewblk which has been been made redundant by frag extension. 4444 */ 4445 static void 4446 cancel_jnewblk(jnewblk, wkhd) 4447 struct jnewblk *jnewblk; 4448 struct workhead *wkhd; 4449 { 4450 struct jsegdep *jsegdep; 4451 4452 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4453 jsegdep = jnewblk->jn_jsegdep; 4454 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4455 panic("cancel_jnewblk: Invalid state"); 4456 jnewblk->jn_jsegdep = NULL; 4457 jnewblk->jn_dep = NULL; 4458 jnewblk->jn_state |= GOINGAWAY; 4459 if (jnewblk->jn_state & INPROGRESS) { 4460 jnewblk->jn_state &= ~INPROGRESS; 4461 WORKLIST_REMOVE(&jnewblk->jn_list); 4462 jwork_insert(wkhd, jsegdep); 4463 } else { 4464 free_jsegdep(jsegdep); 4465 remove_from_journal(&jnewblk->jn_list); 4466 } 4467 wake_worklist(&jnewblk->jn_list); 4468 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4469 } 4470 4471 static void 4472 free_jblkdep(jblkdep) 4473 struct jblkdep *jblkdep; 4474 { 4475 4476 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4477 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4478 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4479 WORKITEM_FREE(jblkdep, D_JTRUNC); 4480 else 4481 panic("free_jblkdep: Unexpected type %s", 4482 TYPENAME(jblkdep->jb_list.wk_type)); 4483 } 4484 4485 /* 4486 * Free a single jseg once it is no longer referenced in memory or on 4487 * disk. Reclaim journal blocks and dependencies waiting for the segment 4488 * to disappear. 4489 */ 4490 static void 4491 free_jseg(jseg, jblocks) 4492 struct jseg *jseg; 4493 struct jblocks *jblocks; 4494 { 4495 struct freework *freework; 4496 4497 /* 4498 * Free freework structures that were lingering to indicate freed 4499 * indirect blocks that forced journal write ordering on reallocate. 4500 */ 4501 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4502 indirblk_remove(freework); 4503 if (jblocks->jb_oldestseg == jseg) 4504 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4505 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4506 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4507 KASSERT(LIST_EMPTY(&jseg->js_entries), 4508 ("free_jseg: Freed jseg has valid entries.")); 4509 WORKITEM_FREE(jseg, D_JSEG); 4510 } 4511 4512 /* 4513 * Free all jsegs that meet the criteria for being reclaimed and update 4514 * oldestseg. 4515 */ 4516 static void 4517 free_jsegs(jblocks) 4518 struct jblocks *jblocks; 4519 { 4520 struct jseg *jseg; 4521 4522 /* 4523 * Free only those jsegs which have none allocated before them to 4524 * preserve the journal space ordering. 4525 */ 4526 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4527 /* 4528 * Only reclaim space when nothing depends on this journal 4529 * set and another set has written that it is no longer 4530 * valid. 4531 */ 4532 if (jseg->js_refs != 0) { 4533 jblocks->jb_oldestseg = jseg; 4534 return; 4535 } 4536 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4537 break; 4538 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4539 break; 4540 /* 4541 * We can free jsegs that didn't write entries when 4542 * oldestwrseq == js_seq. 4543 */ 4544 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4545 jseg->js_cnt != 0) 4546 break; 4547 free_jseg(jseg, jblocks); 4548 } 4549 /* 4550 * If we exited the loop above we still must discover the 4551 * oldest valid segment. 4552 */ 4553 if (jseg) 4554 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4555 jseg = TAILQ_NEXT(jseg, js_next)) 4556 if (jseg->js_refs != 0) 4557 break; 4558 jblocks->jb_oldestseg = jseg; 4559 /* 4560 * The journal has no valid records but some jsegs may still be 4561 * waiting on oldestwrseq to advance. We force a small record 4562 * out to permit these lingering records to be reclaimed. 4563 */ 4564 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4565 jblocks->jb_needseg = 1; 4566 } 4567 4568 /* 4569 * Release one reference to a jseg and free it if the count reaches 0. This 4570 * should eventually reclaim journal space as well. 4571 */ 4572 static void 4573 rele_jseg(jseg) 4574 struct jseg *jseg; 4575 { 4576 4577 KASSERT(jseg->js_refs > 0, 4578 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4579 if (--jseg->js_refs != 0) 4580 return; 4581 free_jsegs(jseg->js_jblocks); 4582 } 4583 4584 /* 4585 * Release a jsegdep and decrement the jseg count. 4586 */ 4587 static void 4588 free_jsegdep(jsegdep) 4589 struct jsegdep *jsegdep; 4590 { 4591 4592 if (jsegdep->jd_seg) 4593 rele_jseg(jsegdep->jd_seg); 4594 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4595 } 4596 4597 /* 4598 * Wait for a journal item to make it to disk. Initiate journal processing 4599 * if required. 4600 */ 4601 static int 4602 jwait(wk, waitfor) 4603 struct worklist *wk; 4604 int waitfor; 4605 { 4606 4607 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4608 /* 4609 * Blocking journal waits cause slow synchronous behavior. Record 4610 * stats on the frequency of these blocking operations. 4611 */ 4612 if (waitfor == MNT_WAIT) { 4613 stat_journal_wait++; 4614 switch (wk->wk_type) { 4615 case D_JREMREF: 4616 case D_JMVREF: 4617 stat_jwait_filepage++; 4618 break; 4619 case D_JTRUNC: 4620 case D_JFREEBLK: 4621 stat_jwait_freeblks++; 4622 break; 4623 case D_JNEWBLK: 4624 stat_jwait_newblk++; 4625 break; 4626 case D_JADDREF: 4627 stat_jwait_inode++; 4628 break; 4629 default: 4630 break; 4631 } 4632 } 4633 /* 4634 * If IO has not started we process the journal. We can't mark the 4635 * worklist item as IOWAITING because we drop the lock while 4636 * processing the journal and the worklist entry may be freed after 4637 * this point. The caller may call back in and re-issue the request. 4638 */ 4639 if ((wk->wk_state & INPROGRESS) == 0) { 4640 softdep_process_journal(wk->wk_mp, wk, waitfor); 4641 if (waitfor != MNT_WAIT) 4642 return (EBUSY); 4643 return (0); 4644 } 4645 if (waitfor != MNT_WAIT) 4646 return (EBUSY); 4647 wait_worklist(wk, "jwait"); 4648 return (0); 4649 } 4650 4651 /* 4652 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4653 * appropriate. This is a convenience function to reduce duplicate code 4654 * for the setup and revert functions below. 4655 */ 4656 static struct inodedep * 4657 inodedep_lookup_ip(ip) 4658 struct inode *ip; 4659 { 4660 struct inodedep *inodedep; 4661 4662 KASSERT(ip->i_nlink >= ip->i_effnlink, 4663 ("inodedep_lookup_ip: bad delta")); 4664 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC, 4665 &inodedep); 4666 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4667 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4668 4669 return (inodedep); 4670 } 4671 4672 /* 4673 * Called prior to creating a new inode and linking it to a directory. The 4674 * jaddref structure must already be allocated by softdep_setup_inomapdep 4675 * and it is discovered here so we can initialize the mode and update 4676 * nlinkdelta. 4677 */ 4678 void 4679 softdep_setup_create(dp, ip) 4680 struct inode *dp; 4681 struct inode *ip; 4682 { 4683 struct inodedep *inodedep; 4684 struct jaddref *jaddref; 4685 struct vnode *dvp; 4686 4687 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4688 ("softdep_setup_create called on non-softdep filesystem")); 4689 KASSERT(ip->i_nlink == 1, 4690 ("softdep_setup_create: Invalid link count.")); 4691 dvp = ITOV(dp); 4692 ACQUIRE_LOCK(ITOUMP(dp)); 4693 inodedep = inodedep_lookup_ip(ip); 4694 if (DOINGSUJ(dvp)) { 4695 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4696 inoreflst); 4697 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4698 ("softdep_setup_create: No addref structure present.")); 4699 } 4700 softdep_prelink(dvp, NULL); 4701 FREE_LOCK(ITOUMP(dp)); 4702 } 4703 4704 /* 4705 * Create a jaddref structure to track the addition of a DOTDOT link when 4706 * we are reparenting an inode as part of a rename. This jaddref will be 4707 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4708 * non-journaling softdep. 4709 */ 4710 void 4711 softdep_setup_dotdot_link(dp, ip) 4712 struct inode *dp; 4713 struct inode *ip; 4714 { 4715 struct inodedep *inodedep; 4716 struct jaddref *jaddref; 4717 struct vnode *dvp; 4718 4719 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4720 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4721 dvp = ITOV(dp); 4722 jaddref = NULL; 4723 /* 4724 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4725 * is used as a normal link would be. 4726 */ 4727 if (DOINGSUJ(dvp)) 4728 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4729 dp->i_effnlink - 1, dp->i_mode); 4730 ACQUIRE_LOCK(ITOUMP(dp)); 4731 inodedep = inodedep_lookup_ip(dp); 4732 if (jaddref) 4733 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4734 if_deps); 4735 softdep_prelink(dvp, ITOV(ip)); 4736 FREE_LOCK(ITOUMP(dp)); 4737 } 4738 4739 /* 4740 * Create a jaddref structure to track a new link to an inode. The directory 4741 * offset is not known until softdep_setup_directory_add or 4742 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4743 * softdep. 4744 */ 4745 void 4746 softdep_setup_link(dp, ip) 4747 struct inode *dp; 4748 struct inode *ip; 4749 { 4750 struct inodedep *inodedep; 4751 struct jaddref *jaddref; 4752 struct vnode *dvp; 4753 4754 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4755 ("softdep_setup_link called on non-softdep filesystem")); 4756 dvp = ITOV(dp); 4757 jaddref = NULL; 4758 if (DOINGSUJ(dvp)) 4759 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4760 ip->i_mode); 4761 ACQUIRE_LOCK(ITOUMP(dp)); 4762 inodedep = inodedep_lookup_ip(ip); 4763 if (jaddref) 4764 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4765 if_deps); 4766 softdep_prelink(dvp, ITOV(ip)); 4767 FREE_LOCK(ITOUMP(dp)); 4768 } 4769 4770 /* 4771 * Called to create the jaddref structures to track . and .. references as 4772 * well as lookup and further initialize the incomplete jaddref created 4773 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4774 * nlinkdelta for non-journaling softdep. 4775 */ 4776 void 4777 softdep_setup_mkdir(dp, ip) 4778 struct inode *dp; 4779 struct inode *ip; 4780 { 4781 struct inodedep *inodedep; 4782 struct jaddref *dotdotaddref; 4783 struct jaddref *dotaddref; 4784 struct jaddref *jaddref; 4785 struct vnode *dvp; 4786 4787 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4788 ("softdep_setup_mkdir called on non-softdep filesystem")); 4789 dvp = ITOV(dp); 4790 dotaddref = dotdotaddref = NULL; 4791 if (DOINGSUJ(dvp)) { 4792 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4793 ip->i_mode); 4794 dotaddref->ja_state |= MKDIR_BODY; 4795 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4796 dp->i_effnlink - 1, dp->i_mode); 4797 dotdotaddref->ja_state |= MKDIR_PARENT; 4798 } 4799 ACQUIRE_LOCK(ITOUMP(dp)); 4800 inodedep = inodedep_lookup_ip(ip); 4801 if (DOINGSUJ(dvp)) { 4802 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4803 inoreflst); 4804 KASSERT(jaddref != NULL, 4805 ("softdep_setup_mkdir: No addref structure present.")); 4806 KASSERT(jaddref->ja_parent == dp->i_number, 4807 ("softdep_setup_mkdir: bad parent %ju", 4808 (uintmax_t)jaddref->ja_parent)); 4809 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4810 if_deps); 4811 } 4812 inodedep = inodedep_lookup_ip(dp); 4813 if (DOINGSUJ(dvp)) 4814 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4815 &dotdotaddref->ja_ref, if_deps); 4816 softdep_prelink(ITOV(dp), NULL); 4817 FREE_LOCK(ITOUMP(dp)); 4818 } 4819 4820 /* 4821 * Called to track nlinkdelta of the inode and parent directories prior to 4822 * unlinking a directory. 4823 */ 4824 void 4825 softdep_setup_rmdir(dp, ip) 4826 struct inode *dp; 4827 struct inode *ip; 4828 { 4829 struct vnode *dvp; 4830 4831 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4832 ("softdep_setup_rmdir called on non-softdep filesystem")); 4833 dvp = ITOV(dp); 4834 ACQUIRE_LOCK(ITOUMP(dp)); 4835 (void) inodedep_lookup_ip(ip); 4836 (void) inodedep_lookup_ip(dp); 4837 softdep_prelink(dvp, ITOV(ip)); 4838 FREE_LOCK(ITOUMP(dp)); 4839 } 4840 4841 /* 4842 * Called to track nlinkdelta of the inode and parent directories prior to 4843 * unlink. 4844 */ 4845 void 4846 softdep_setup_unlink(dp, ip) 4847 struct inode *dp; 4848 struct inode *ip; 4849 { 4850 struct vnode *dvp; 4851 4852 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4853 ("softdep_setup_unlink called on non-softdep filesystem")); 4854 dvp = ITOV(dp); 4855 ACQUIRE_LOCK(ITOUMP(dp)); 4856 (void) inodedep_lookup_ip(ip); 4857 (void) inodedep_lookup_ip(dp); 4858 softdep_prelink(dvp, ITOV(ip)); 4859 FREE_LOCK(ITOUMP(dp)); 4860 } 4861 4862 /* 4863 * Called to release the journal structures created by a failed non-directory 4864 * creation. Adjusts nlinkdelta for non-journaling softdep. 4865 */ 4866 void 4867 softdep_revert_create(dp, ip) 4868 struct inode *dp; 4869 struct inode *ip; 4870 { 4871 struct inodedep *inodedep; 4872 struct jaddref *jaddref; 4873 struct vnode *dvp; 4874 4875 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0, 4876 ("softdep_revert_create called on non-softdep filesystem")); 4877 dvp = ITOV(dp); 4878 ACQUIRE_LOCK(ITOUMP(dp)); 4879 inodedep = inodedep_lookup_ip(ip); 4880 if (DOINGSUJ(dvp)) { 4881 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4882 inoreflst); 4883 KASSERT(jaddref->ja_parent == dp->i_number, 4884 ("softdep_revert_create: addref parent mismatch")); 4885 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4886 } 4887 FREE_LOCK(ITOUMP(dp)); 4888 } 4889 4890 /* 4891 * Called to release the journal structures created by a failed link 4892 * addition. Adjusts nlinkdelta for non-journaling softdep. 4893 */ 4894 void 4895 softdep_revert_link(dp, ip) 4896 struct inode *dp; 4897 struct inode *ip; 4898 { 4899 struct inodedep *inodedep; 4900 struct jaddref *jaddref; 4901 struct vnode *dvp; 4902 4903 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4904 ("softdep_revert_link called on non-softdep filesystem")); 4905 dvp = ITOV(dp); 4906 ACQUIRE_LOCK(ITOUMP(dp)); 4907 inodedep = inodedep_lookup_ip(ip); 4908 if (DOINGSUJ(dvp)) { 4909 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4910 inoreflst); 4911 KASSERT(jaddref->ja_parent == dp->i_number, 4912 ("softdep_revert_link: addref parent mismatch")); 4913 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4914 } 4915 FREE_LOCK(ITOUMP(dp)); 4916 } 4917 4918 /* 4919 * Called to release the journal structures created by a failed mkdir 4920 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4921 */ 4922 void 4923 softdep_revert_mkdir(dp, ip) 4924 struct inode *dp; 4925 struct inode *ip; 4926 { 4927 struct inodedep *inodedep; 4928 struct jaddref *jaddref; 4929 struct jaddref *dotaddref; 4930 struct vnode *dvp; 4931 4932 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4933 ("softdep_revert_mkdir called on non-softdep filesystem")); 4934 dvp = ITOV(dp); 4935 4936 ACQUIRE_LOCK(ITOUMP(dp)); 4937 inodedep = inodedep_lookup_ip(dp); 4938 if (DOINGSUJ(dvp)) { 4939 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4940 inoreflst); 4941 KASSERT(jaddref->ja_parent == ip->i_number, 4942 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4943 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4944 } 4945 inodedep = inodedep_lookup_ip(ip); 4946 if (DOINGSUJ(dvp)) { 4947 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4948 inoreflst); 4949 KASSERT(jaddref->ja_parent == dp->i_number, 4950 ("softdep_revert_mkdir: addref parent mismatch")); 4951 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4952 inoreflst, if_deps); 4953 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4954 KASSERT(dotaddref->ja_parent == ip->i_number, 4955 ("softdep_revert_mkdir: dot addref parent mismatch")); 4956 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4957 } 4958 FREE_LOCK(ITOUMP(dp)); 4959 } 4960 4961 /* 4962 * Called to correct nlinkdelta after a failed rmdir. 4963 */ 4964 void 4965 softdep_revert_rmdir(dp, ip) 4966 struct inode *dp; 4967 struct inode *ip; 4968 { 4969 4970 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4971 ("softdep_revert_rmdir called on non-softdep filesystem")); 4972 ACQUIRE_LOCK(ITOUMP(dp)); 4973 (void) inodedep_lookup_ip(ip); 4974 (void) inodedep_lookup_ip(dp); 4975 FREE_LOCK(ITOUMP(dp)); 4976 } 4977 4978 /* 4979 * Protecting the freemaps (or bitmaps). 4980 * 4981 * To eliminate the need to execute fsck before mounting a filesystem 4982 * after a power failure, one must (conservatively) guarantee that the 4983 * on-disk copy of the bitmaps never indicate that a live inode or block is 4984 * free. So, when a block or inode is allocated, the bitmap should be 4985 * updated (on disk) before any new pointers. When a block or inode is 4986 * freed, the bitmap should not be updated until all pointers have been 4987 * reset. The latter dependency is handled by the delayed de-allocation 4988 * approach described below for block and inode de-allocation. The former 4989 * dependency is handled by calling the following procedure when a block or 4990 * inode is allocated. When an inode is allocated an "inodedep" is created 4991 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4992 * Each "inodedep" is also inserted into the hash indexing structure so 4993 * that any additional link additions can be made dependent on the inode 4994 * allocation. 4995 * 4996 * The ufs filesystem maintains a number of free block counts (e.g., per 4997 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4998 * in addition to the bitmaps. These counts are used to improve efficiency 4999 * during allocation and therefore must be consistent with the bitmaps. 5000 * There is no convenient way to guarantee post-crash consistency of these 5001 * counts with simple update ordering, for two main reasons: (1) The counts 5002 * and bitmaps for a single cylinder group block are not in the same disk 5003 * sector. If a disk write is interrupted (e.g., by power failure), one may 5004 * be written and the other not. (2) Some of the counts are located in the 5005 * superblock rather than the cylinder group block. So, we focus our soft 5006 * updates implementation on protecting the bitmaps. When mounting a 5007 * filesystem, we recompute the auxiliary counts from the bitmaps. 5008 */ 5009 5010 /* 5011 * Called just after updating the cylinder group block to allocate an inode. 5012 */ 5013 void 5014 softdep_setup_inomapdep(bp, ip, newinum, mode) 5015 struct buf *bp; /* buffer for cylgroup block with inode map */ 5016 struct inode *ip; /* inode related to allocation */ 5017 ino_t newinum; /* new inode number being allocated */ 5018 int mode; 5019 { 5020 struct inodedep *inodedep; 5021 struct bmsafemap *bmsafemap; 5022 struct jaddref *jaddref; 5023 struct mount *mp; 5024 struct fs *fs; 5025 5026 mp = ITOVFS(ip); 5027 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5028 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5029 fs = VFSTOUFS(mp)->um_fs; 5030 jaddref = NULL; 5031 5032 /* 5033 * Allocate the journal reference add structure so that the bitmap 5034 * can be dependent on it. 5035 */ 5036 if (MOUNTEDSUJ(mp)) { 5037 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5038 jaddref->ja_state |= NEWBLOCK; 5039 } 5040 5041 /* 5042 * Create a dependency for the newly allocated inode. 5043 * Panic if it already exists as something is seriously wrong. 5044 * Otherwise add it to the dependency list for the buffer holding 5045 * the cylinder group map from which it was allocated. 5046 * 5047 * We have to preallocate a bmsafemap entry in case it is needed 5048 * in bmsafemap_lookup since once we allocate the inodedep, we 5049 * have to finish initializing it before we can FREE_LOCK(). 5050 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5051 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5052 * creating the inodedep as it can be freed during the time 5053 * that we FREE_LOCK() while allocating the inodedep. We must 5054 * call workitem_alloc() before entering the locked section as 5055 * it also acquires the lock and we must avoid trying doing so 5056 * recursively. 5057 */ 5058 bmsafemap = malloc(sizeof(struct bmsafemap), 5059 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5060 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5061 ACQUIRE_LOCK(ITOUMP(ip)); 5062 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5063 panic("softdep_setup_inomapdep: dependency %p for new" 5064 "inode already exists", inodedep); 5065 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5066 if (jaddref) { 5067 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5068 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5069 if_deps); 5070 } else { 5071 inodedep->id_state |= ONDEPLIST; 5072 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5073 } 5074 inodedep->id_bmsafemap = bmsafemap; 5075 inodedep->id_state &= ~DEPCOMPLETE; 5076 FREE_LOCK(ITOUMP(ip)); 5077 } 5078 5079 /* 5080 * Called just after updating the cylinder group block to 5081 * allocate block or fragment. 5082 */ 5083 void 5084 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5085 struct buf *bp; /* buffer for cylgroup block with block map */ 5086 struct mount *mp; /* filesystem doing allocation */ 5087 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5088 int frags; /* Number of fragments. */ 5089 int oldfrags; /* Previous number of fragments for extend. */ 5090 { 5091 struct newblk *newblk; 5092 struct bmsafemap *bmsafemap; 5093 struct jnewblk *jnewblk; 5094 struct ufsmount *ump; 5095 struct fs *fs; 5096 5097 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5098 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5099 ump = VFSTOUFS(mp); 5100 fs = ump->um_fs; 5101 jnewblk = NULL; 5102 /* 5103 * Create a dependency for the newly allocated block. 5104 * Add it to the dependency list for the buffer holding 5105 * the cylinder group map from which it was allocated. 5106 */ 5107 if (MOUNTEDSUJ(mp)) { 5108 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5109 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5110 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5111 jnewblk->jn_state = ATTACHED; 5112 jnewblk->jn_blkno = newblkno; 5113 jnewblk->jn_frags = frags; 5114 jnewblk->jn_oldfrags = oldfrags; 5115 #ifdef INVARIANTS 5116 { 5117 struct cg *cgp; 5118 uint8_t *blksfree; 5119 long bno; 5120 int i; 5121 5122 cgp = (struct cg *)bp->b_data; 5123 blksfree = cg_blksfree(cgp); 5124 bno = dtogd(fs, jnewblk->jn_blkno); 5125 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5126 i++) { 5127 if (isset(blksfree, bno + i)) 5128 panic("softdep_setup_blkmapdep: " 5129 "free fragment %d from %d-%d " 5130 "state 0x%X dep %p", i, 5131 jnewblk->jn_oldfrags, 5132 jnewblk->jn_frags, 5133 jnewblk->jn_state, 5134 jnewblk->jn_dep); 5135 } 5136 } 5137 #endif 5138 } 5139 5140 CTR3(KTR_SUJ, 5141 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5142 newblkno, frags, oldfrags); 5143 ACQUIRE_LOCK(ump); 5144 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5145 panic("softdep_setup_blkmapdep: found block"); 5146 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5147 dtog(fs, newblkno), NULL); 5148 if (jnewblk) { 5149 jnewblk->jn_dep = (struct worklist *)newblk; 5150 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5151 } else { 5152 newblk->nb_state |= ONDEPLIST; 5153 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5154 } 5155 newblk->nb_bmsafemap = bmsafemap; 5156 newblk->nb_jnewblk = jnewblk; 5157 FREE_LOCK(ump); 5158 } 5159 5160 #define BMSAFEMAP_HASH(ump, cg) \ 5161 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5162 5163 static int 5164 bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5165 struct bmsafemap_hashhead *bmsafemaphd; 5166 int cg; 5167 struct bmsafemap **bmsafemapp; 5168 { 5169 struct bmsafemap *bmsafemap; 5170 5171 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5172 if (bmsafemap->sm_cg == cg) 5173 break; 5174 if (bmsafemap) { 5175 *bmsafemapp = bmsafemap; 5176 return (1); 5177 } 5178 *bmsafemapp = NULL; 5179 5180 return (0); 5181 } 5182 5183 /* 5184 * Find the bmsafemap associated with a cylinder group buffer. 5185 * If none exists, create one. The buffer must be locked when 5186 * this routine is called and this routine must be called with 5187 * the softdep lock held. To avoid giving up the lock while 5188 * allocating a new bmsafemap, a preallocated bmsafemap may be 5189 * provided. If it is provided but not needed, it is freed. 5190 */ 5191 static struct bmsafemap * 5192 bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5193 struct mount *mp; 5194 struct buf *bp; 5195 int cg; 5196 struct bmsafemap *newbmsafemap; 5197 { 5198 struct bmsafemap_hashhead *bmsafemaphd; 5199 struct bmsafemap *bmsafemap, *collision; 5200 struct worklist *wk; 5201 struct ufsmount *ump; 5202 5203 ump = VFSTOUFS(mp); 5204 LOCK_OWNED(ump); 5205 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5206 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5207 if (wk->wk_type == D_BMSAFEMAP) { 5208 if (newbmsafemap) 5209 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5210 return (WK_BMSAFEMAP(wk)); 5211 } 5212 } 5213 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5214 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5215 if (newbmsafemap) 5216 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5217 return (bmsafemap); 5218 } 5219 if (newbmsafemap) { 5220 bmsafemap = newbmsafemap; 5221 } else { 5222 FREE_LOCK(ump); 5223 bmsafemap = malloc(sizeof(struct bmsafemap), 5224 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5225 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5226 ACQUIRE_LOCK(ump); 5227 } 5228 bmsafemap->sm_buf = bp; 5229 LIST_INIT(&bmsafemap->sm_inodedephd); 5230 LIST_INIT(&bmsafemap->sm_inodedepwr); 5231 LIST_INIT(&bmsafemap->sm_newblkhd); 5232 LIST_INIT(&bmsafemap->sm_newblkwr); 5233 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5234 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5235 LIST_INIT(&bmsafemap->sm_freehd); 5236 LIST_INIT(&bmsafemap->sm_freewr); 5237 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5238 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5239 return (collision); 5240 } 5241 bmsafemap->sm_cg = cg; 5242 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5243 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5244 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5245 return (bmsafemap); 5246 } 5247 5248 /* 5249 * Direct block allocation dependencies. 5250 * 5251 * When a new block is allocated, the corresponding disk locations must be 5252 * initialized (with zeros or new data) before the on-disk inode points to 5253 * them. Also, the freemap from which the block was allocated must be 5254 * updated (on disk) before the inode's pointer. These two dependencies are 5255 * independent of each other and are needed for all file blocks and indirect 5256 * blocks that are pointed to directly by the inode. Just before the 5257 * "in-core" version of the inode is updated with a newly allocated block 5258 * number, a procedure (below) is called to setup allocation dependency 5259 * structures. These structures are removed when the corresponding 5260 * dependencies are satisfied or when the block allocation becomes obsolete 5261 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5262 * fragment that gets upgraded). All of these cases are handled in 5263 * procedures described later. 5264 * 5265 * When a file extension causes a fragment to be upgraded, either to a larger 5266 * fragment or to a full block, the on-disk location may change (if the 5267 * previous fragment could not simply be extended). In this case, the old 5268 * fragment must be de-allocated, but not until after the inode's pointer has 5269 * been updated. In most cases, this is handled by later procedures, which 5270 * will construct a "freefrag" structure to be added to the workitem queue 5271 * when the inode update is complete (or obsolete). The main exception to 5272 * this is when an allocation occurs while a pending allocation dependency 5273 * (for the same block pointer) remains. This case is handled in the main 5274 * allocation dependency setup procedure by immediately freeing the 5275 * unreferenced fragments. 5276 */ 5277 void 5278 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5279 struct inode *ip; /* inode to which block is being added */ 5280 ufs_lbn_t off; /* block pointer within inode */ 5281 ufs2_daddr_t newblkno; /* disk block number being added */ 5282 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5283 long newsize; /* size of new block */ 5284 long oldsize; /* size of new block */ 5285 struct buf *bp; /* bp for allocated block */ 5286 { 5287 struct allocdirect *adp, *oldadp; 5288 struct allocdirectlst *adphead; 5289 struct freefrag *freefrag; 5290 struct inodedep *inodedep; 5291 struct pagedep *pagedep; 5292 struct jnewblk *jnewblk; 5293 struct newblk *newblk; 5294 struct mount *mp; 5295 ufs_lbn_t lbn; 5296 5297 lbn = bp->b_lblkno; 5298 mp = ITOVFS(ip); 5299 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5300 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5301 if (oldblkno && oldblkno != newblkno) 5302 /* 5303 * The usual case is that a smaller fragment that 5304 * was just allocated has been replaced with a bigger 5305 * fragment or a full-size block. If it is marked as 5306 * B_DELWRI, the current contents have not been written 5307 * to disk. It is possible that the block was written 5308 * earlier, but very uncommon. If the block has never 5309 * been written, there is no need to send a BIO_DELETE 5310 * for it when it is freed. The gain from avoiding the 5311 * TRIMs for the common case of unwritten blocks far 5312 * exceeds the cost of the write amplification for the 5313 * uncommon case of failing to send a TRIM for a block 5314 * that had been written. 5315 */ 5316 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 5317 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 5318 else 5319 freefrag = NULL; 5320 5321 CTR6(KTR_SUJ, 5322 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5323 "off %jd newsize %ld oldsize %d", 5324 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5325 ACQUIRE_LOCK(ITOUMP(ip)); 5326 if (off >= UFS_NDADDR) { 5327 if (lbn > 0) 5328 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5329 lbn, off); 5330 /* allocating an indirect block */ 5331 if (oldblkno != 0) 5332 panic("softdep_setup_allocdirect: non-zero indir"); 5333 } else { 5334 if (off != lbn) 5335 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5336 lbn, off); 5337 /* 5338 * Allocating a direct block. 5339 * 5340 * If we are allocating a directory block, then we must 5341 * allocate an associated pagedep to track additions and 5342 * deletions. 5343 */ 5344 if ((ip->i_mode & IFMT) == IFDIR) 5345 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5346 &pagedep); 5347 } 5348 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5349 panic("softdep_setup_allocdirect: lost block"); 5350 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5351 ("softdep_setup_allocdirect: newblk already initialized")); 5352 /* 5353 * Convert the newblk to an allocdirect. 5354 */ 5355 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5356 adp = (struct allocdirect *)newblk; 5357 newblk->nb_freefrag = freefrag; 5358 adp->ad_offset = off; 5359 adp->ad_oldblkno = oldblkno; 5360 adp->ad_newsize = newsize; 5361 adp->ad_oldsize = oldsize; 5362 5363 /* 5364 * Finish initializing the journal. 5365 */ 5366 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5367 jnewblk->jn_ino = ip->i_number; 5368 jnewblk->jn_lbn = lbn; 5369 add_to_journal(&jnewblk->jn_list); 5370 } 5371 if (freefrag && freefrag->ff_jdep != NULL && 5372 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5373 add_to_journal(freefrag->ff_jdep); 5374 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5375 adp->ad_inodedep = inodedep; 5376 5377 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5378 /* 5379 * The list of allocdirects must be kept in sorted and ascending 5380 * order so that the rollback routines can quickly determine the 5381 * first uncommitted block (the size of the file stored on disk 5382 * ends at the end of the lowest committed fragment, or if there 5383 * are no fragments, at the end of the highest committed block). 5384 * Since files generally grow, the typical case is that the new 5385 * block is to be added at the end of the list. We speed this 5386 * special case by checking against the last allocdirect in the 5387 * list before laboriously traversing the list looking for the 5388 * insertion point. 5389 */ 5390 adphead = &inodedep->id_newinoupdt; 5391 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5392 if (oldadp == NULL || oldadp->ad_offset <= off) { 5393 /* insert at end of list */ 5394 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5395 if (oldadp != NULL && oldadp->ad_offset == off) 5396 allocdirect_merge(adphead, adp, oldadp); 5397 FREE_LOCK(ITOUMP(ip)); 5398 return; 5399 } 5400 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5401 if (oldadp->ad_offset >= off) 5402 break; 5403 } 5404 if (oldadp == NULL) 5405 panic("softdep_setup_allocdirect: lost entry"); 5406 /* insert in middle of list */ 5407 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5408 if (oldadp->ad_offset == off) 5409 allocdirect_merge(adphead, adp, oldadp); 5410 5411 FREE_LOCK(ITOUMP(ip)); 5412 } 5413 5414 /* 5415 * Merge a newer and older journal record to be stored either in a 5416 * newblock or freefrag. This handles aggregating journal records for 5417 * fragment allocation into a second record as well as replacing a 5418 * journal free with an aborted journal allocation. A segment for the 5419 * oldest record will be placed on wkhd if it has been written. If not 5420 * the segment for the newer record will suffice. 5421 */ 5422 static struct worklist * 5423 jnewblk_merge(new, old, wkhd) 5424 struct worklist *new; 5425 struct worklist *old; 5426 struct workhead *wkhd; 5427 { 5428 struct jnewblk *njnewblk; 5429 struct jnewblk *jnewblk; 5430 5431 /* Handle NULLs to simplify callers. */ 5432 if (new == NULL) 5433 return (old); 5434 if (old == NULL) 5435 return (new); 5436 /* Replace a jfreefrag with a jnewblk. */ 5437 if (new->wk_type == D_JFREEFRAG) { 5438 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5439 panic("jnewblk_merge: blkno mismatch: %p, %p", 5440 old, new); 5441 cancel_jfreefrag(WK_JFREEFRAG(new)); 5442 return (old); 5443 } 5444 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5445 panic("jnewblk_merge: Bad type: old %d new %d\n", 5446 old->wk_type, new->wk_type); 5447 /* 5448 * Handle merging of two jnewblk records that describe 5449 * different sets of fragments in the same block. 5450 */ 5451 jnewblk = WK_JNEWBLK(old); 5452 njnewblk = WK_JNEWBLK(new); 5453 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5454 panic("jnewblk_merge: Merging disparate blocks."); 5455 /* 5456 * The record may be rolled back in the cg. 5457 */ 5458 if (jnewblk->jn_state & UNDONE) { 5459 jnewblk->jn_state &= ~UNDONE; 5460 njnewblk->jn_state |= UNDONE; 5461 njnewblk->jn_state &= ~ATTACHED; 5462 } 5463 /* 5464 * We modify the newer addref and free the older so that if neither 5465 * has been written the most up-to-date copy will be on disk. If 5466 * both have been written but rolled back we only temporarily need 5467 * one of them to fix the bits when the cg write completes. 5468 */ 5469 jnewblk->jn_state |= ATTACHED | COMPLETE; 5470 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5471 cancel_jnewblk(jnewblk, wkhd); 5472 WORKLIST_REMOVE(&jnewblk->jn_list); 5473 free_jnewblk(jnewblk); 5474 return (new); 5475 } 5476 5477 /* 5478 * Replace an old allocdirect dependency with a newer one. 5479 */ 5480 static void 5481 allocdirect_merge(adphead, newadp, oldadp) 5482 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5483 struct allocdirect *newadp; /* allocdirect being added */ 5484 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5485 { 5486 struct worklist *wk; 5487 struct freefrag *freefrag; 5488 5489 freefrag = NULL; 5490 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5491 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5492 newadp->ad_oldsize != oldadp->ad_newsize || 5493 newadp->ad_offset >= UFS_NDADDR) 5494 panic("%s %jd != new %jd || old size %ld != new %ld", 5495 "allocdirect_merge: old blkno", 5496 (intmax_t)newadp->ad_oldblkno, 5497 (intmax_t)oldadp->ad_newblkno, 5498 newadp->ad_oldsize, oldadp->ad_newsize); 5499 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5500 newadp->ad_oldsize = oldadp->ad_oldsize; 5501 /* 5502 * If the old dependency had a fragment to free or had never 5503 * previously had a block allocated, then the new dependency 5504 * can immediately post its freefrag and adopt the old freefrag. 5505 * This action is done by swapping the freefrag dependencies. 5506 * The new dependency gains the old one's freefrag, and the 5507 * old one gets the new one and then immediately puts it on 5508 * the worklist when it is freed by free_newblk. It is 5509 * not possible to do this swap when the old dependency had a 5510 * non-zero size but no previous fragment to free. This condition 5511 * arises when the new block is an extension of the old block. 5512 * Here, the first part of the fragment allocated to the new 5513 * dependency is part of the block currently claimed on disk by 5514 * the old dependency, so cannot legitimately be freed until the 5515 * conditions for the new dependency are fulfilled. 5516 */ 5517 freefrag = newadp->ad_freefrag; 5518 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5519 newadp->ad_freefrag = oldadp->ad_freefrag; 5520 oldadp->ad_freefrag = freefrag; 5521 } 5522 /* 5523 * If we are tracking a new directory-block allocation, 5524 * move it from the old allocdirect to the new allocdirect. 5525 */ 5526 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5527 WORKLIST_REMOVE(wk); 5528 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5529 panic("allocdirect_merge: extra newdirblk"); 5530 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5531 } 5532 TAILQ_REMOVE(adphead, oldadp, ad_next); 5533 /* 5534 * We need to move any journal dependencies over to the freefrag 5535 * that releases this block if it exists. Otherwise we are 5536 * extending an existing block and we'll wait until that is 5537 * complete to release the journal space and extend the 5538 * new journal to cover this old space as well. 5539 */ 5540 if (freefrag == NULL) { 5541 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5542 panic("allocdirect_merge: %jd != %jd", 5543 oldadp->ad_newblkno, newadp->ad_newblkno); 5544 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5545 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5546 &oldadp->ad_block.nb_jnewblk->jn_list, 5547 &newadp->ad_block.nb_jwork); 5548 oldadp->ad_block.nb_jnewblk = NULL; 5549 cancel_newblk(&oldadp->ad_block, NULL, 5550 &newadp->ad_block.nb_jwork); 5551 } else { 5552 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5553 &freefrag->ff_list, &freefrag->ff_jwork); 5554 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5555 &freefrag->ff_jwork); 5556 } 5557 free_newblk(&oldadp->ad_block); 5558 } 5559 5560 /* 5561 * Allocate a jfreefrag structure to journal a single block free. 5562 */ 5563 static struct jfreefrag * 5564 newjfreefrag(freefrag, ip, blkno, size, lbn) 5565 struct freefrag *freefrag; 5566 struct inode *ip; 5567 ufs2_daddr_t blkno; 5568 long size; 5569 ufs_lbn_t lbn; 5570 { 5571 struct jfreefrag *jfreefrag; 5572 struct fs *fs; 5573 5574 fs = ITOFS(ip); 5575 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5576 M_SOFTDEP_FLAGS); 5577 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip)); 5578 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5579 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5580 jfreefrag->fr_ino = ip->i_number; 5581 jfreefrag->fr_lbn = lbn; 5582 jfreefrag->fr_blkno = blkno; 5583 jfreefrag->fr_frags = numfrags(fs, size); 5584 jfreefrag->fr_freefrag = freefrag; 5585 5586 return (jfreefrag); 5587 } 5588 5589 /* 5590 * Allocate a new freefrag structure. 5591 */ 5592 static struct freefrag * 5593 newfreefrag(ip, blkno, size, lbn, key) 5594 struct inode *ip; 5595 ufs2_daddr_t blkno; 5596 long size; 5597 ufs_lbn_t lbn; 5598 u_long key; 5599 { 5600 struct freefrag *freefrag; 5601 struct ufsmount *ump; 5602 struct fs *fs; 5603 5604 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5605 ip->i_number, blkno, size, lbn); 5606 ump = ITOUMP(ip); 5607 fs = ump->um_fs; 5608 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5609 panic("newfreefrag: frag size"); 5610 freefrag = malloc(sizeof(struct freefrag), 5611 M_FREEFRAG, M_SOFTDEP_FLAGS); 5612 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump)); 5613 freefrag->ff_state = ATTACHED; 5614 LIST_INIT(&freefrag->ff_jwork); 5615 freefrag->ff_inum = ip->i_number; 5616 freefrag->ff_vtype = ITOV(ip)->v_type; 5617 freefrag->ff_blkno = blkno; 5618 freefrag->ff_fragsize = size; 5619 freefrag->ff_key = key; 5620 5621 if (MOUNTEDSUJ(UFSTOVFS(ump))) { 5622 freefrag->ff_jdep = (struct worklist *) 5623 newjfreefrag(freefrag, ip, blkno, size, lbn); 5624 } else { 5625 freefrag->ff_state |= DEPCOMPLETE; 5626 freefrag->ff_jdep = NULL; 5627 } 5628 5629 return (freefrag); 5630 } 5631 5632 /* 5633 * This workitem de-allocates fragments that were replaced during 5634 * file block allocation. 5635 */ 5636 static void 5637 handle_workitem_freefrag(freefrag) 5638 struct freefrag *freefrag; 5639 { 5640 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5641 struct workhead wkhd; 5642 5643 CTR3(KTR_SUJ, 5644 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5645 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5646 /* 5647 * It would be illegal to add new completion items to the 5648 * freefrag after it was schedule to be done so it must be 5649 * safe to modify the list head here. 5650 */ 5651 LIST_INIT(&wkhd); 5652 ACQUIRE_LOCK(ump); 5653 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5654 /* 5655 * If the journal has not been written we must cancel it here. 5656 */ 5657 if (freefrag->ff_jdep) { 5658 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5659 panic("handle_workitem_freefrag: Unexpected type %d\n", 5660 freefrag->ff_jdep->wk_type); 5661 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5662 } 5663 FREE_LOCK(ump); 5664 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5665 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, 5666 &wkhd, freefrag->ff_key); 5667 ACQUIRE_LOCK(ump); 5668 WORKITEM_FREE(freefrag, D_FREEFRAG); 5669 FREE_LOCK(ump); 5670 } 5671 5672 /* 5673 * Set up a dependency structure for an external attributes data block. 5674 * This routine follows much of the structure of softdep_setup_allocdirect. 5675 * See the description of softdep_setup_allocdirect above for details. 5676 */ 5677 void 5678 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5679 struct inode *ip; 5680 ufs_lbn_t off; 5681 ufs2_daddr_t newblkno; 5682 ufs2_daddr_t oldblkno; 5683 long newsize; 5684 long oldsize; 5685 struct buf *bp; 5686 { 5687 struct allocdirect *adp, *oldadp; 5688 struct allocdirectlst *adphead; 5689 struct freefrag *freefrag; 5690 struct inodedep *inodedep; 5691 struct jnewblk *jnewblk; 5692 struct newblk *newblk; 5693 struct mount *mp; 5694 struct ufsmount *ump; 5695 ufs_lbn_t lbn; 5696 5697 mp = ITOVFS(ip); 5698 ump = VFSTOUFS(mp); 5699 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5700 ("softdep_setup_allocext called on non-softdep filesystem")); 5701 KASSERT(off < UFS_NXADDR, 5702 ("softdep_setup_allocext: lbn %lld > UFS_NXADDR", (long long)off)); 5703 5704 lbn = bp->b_lblkno; 5705 if (oldblkno && oldblkno != newblkno) 5706 /* 5707 * The usual case is that a smaller fragment that 5708 * was just allocated has been replaced with a bigger 5709 * fragment or a full-size block. If it is marked as 5710 * B_DELWRI, the current contents have not been written 5711 * to disk. It is possible that the block was written 5712 * earlier, but very uncommon. If the block has never 5713 * been written, there is no need to send a BIO_DELETE 5714 * for it when it is freed. The gain from avoiding the 5715 * TRIMs for the common case of unwritten blocks far 5716 * exceeds the cost of the write amplification for the 5717 * uncommon case of failing to send a TRIM for a block 5718 * that had been written. 5719 */ 5720 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn, 5721 (bp->b_flags & B_DELWRI) != 0 ? NOTRIM_KEY : SINGLETON_KEY); 5722 else 5723 freefrag = NULL; 5724 5725 ACQUIRE_LOCK(ump); 5726 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5727 panic("softdep_setup_allocext: lost block"); 5728 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5729 ("softdep_setup_allocext: newblk already initialized")); 5730 /* 5731 * Convert the newblk to an allocdirect. 5732 */ 5733 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5734 adp = (struct allocdirect *)newblk; 5735 newblk->nb_freefrag = freefrag; 5736 adp->ad_offset = off; 5737 adp->ad_oldblkno = oldblkno; 5738 adp->ad_newsize = newsize; 5739 adp->ad_oldsize = oldsize; 5740 adp->ad_state |= EXTDATA; 5741 5742 /* 5743 * Finish initializing the journal. 5744 */ 5745 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5746 jnewblk->jn_ino = ip->i_number; 5747 jnewblk->jn_lbn = lbn; 5748 add_to_journal(&jnewblk->jn_list); 5749 } 5750 if (freefrag && freefrag->ff_jdep != NULL && 5751 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5752 add_to_journal(freefrag->ff_jdep); 5753 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5754 adp->ad_inodedep = inodedep; 5755 5756 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5757 /* 5758 * The list of allocdirects must be kept in sorted and ascending 5759 * order so that the rollback routines can quickly determine the 5760 * first uncommitted block (the size of the file stored on disk 5761 * ends at the end of the lowest committed fragment, or if there 5762 * are no fragments, at the end of the highest committed block). 5763 * Since files generally grow, the typical case is that the new 5764 * block is to be added at the end of the list. We speed this 5765 * special case by checking against the last allocdirect in the 5766 * list before laboriously traversing the list looking for the 5767 * insertion point. 5768 */ 5769 adphead = &inodedep->id_newextupdt; 5770 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5771 if (oldadp == NULL || oldadp->ad_offset <= off) { 5772 /* insert at end of list */ 5773 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5774 if (oldadp != NULL && oldadp->ad_offset == off) 5775 allocdirect_merge(adphead, adp, oldadp); 5776 FREE_LOCK(ump); 5777 return; 5778 } 5779 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5780 if (oldadp->ad_offset >= off) 5781 break; 5782 } 5783 if (oldadp == NULL) 5784 panic("softdep_setup_allocext: lost entry"); 5785 /* insert in middle of list */ 5786 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5787 if (oldadp->ad_offset == off) 5788 allocdirect_merge(adphead, adp, oldadp); 5789 FREE_LOCK(ump); 5790 } 5791 5792 /* 5793 * Indirect block allocation dependencies. 5794 * 5795 * The same dependencies that exist for a direct block also exist when 5796 * a new block is allocated and pointed to by an entry in a block of 5797 * indirect pointers. The undo/redo states described above are also 5798 * used here. Because an indirect block contains many pointers that 5799 * may have dependencies, a second copy of the entire in-memory indirect 5800 * block is kept. The buffer cache copy is always completely up-to-date. 5801 * The second copy, which is used only as a source for disk writes, 5802 * contains only the safe pointers (i.e., those that have no remaining 5803 * update dependencies). The second copy is freed when all pointers 5804 * are safe. The cache is not allowed to replace indirect blocks with 5805 * pending update dependencies. If a buffer containing an indirect 5806 * block with dependencies is written, these routines will mark it 5807 * dirty again. It can only be successfully written once all the 5808 * dependencies are removed. The ffs_fsync routine in conjunction with 5809 * softdep_sync_metadata work together to get all the dependencies 5810 * removed so that a file can be successfully written to disk. Three 5811 * procedures are used when setting up indirect block pointer 5812 * dependencies. The division is necessary because of the organization 5813 * of the "balloc" routine and because of the distinction between file 5814 * pages and file metadata blocks. 5815 */ 5816 5817 /* 5818 * Allocate a new allocindir structure. 5819 */ 5820 static struct allocindir * 5821 newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5822 struct inode *ip; /* inode for file being extended */ 5823 int ptrno; /* offset of pointer in indirect block */ 5824 ufs2_daddr_t newblkno; /* disk block number being added */ 5825 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5826 ufs_lbn_t lbn; 5827 { 5828 struct newblk *newblk; 5829 struct allocindir *aip; 5830 struct freefrag *freefrag; 5831 struct jnewblk *jnewblk; 5832 5833 if (oldblkno) 5834 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn, 5835 SINGLETON_KEY); 5836 else 5837 freefrag = NULL; 5838 ACQUIRE_LOCK(ITOUMP(ip)); 5839 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0) 5840 panic("new_allocindir: lost block"); 5841 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5842 ("newallocindir: newblk already initialized")); 5843 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5844 newblk->nb_freefrag = freefrag; 5845 aip = (struct allocindir *)newblk; 5846 aip->ai_offset = ptrno; 5847 aip->ai_oldblkno = oldblkno; 5848 aip->ai_lbn = lbn; 5849 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5850 jnewblk->jn_ino = ip->i_number; 5851 jnewblk->jn_lbn = lbn; 5852 add_to_journal(&jnewblk->jn_list); 5853 } 5854 if (freefrag && freefrag->ff_jdep != NULL && 5855 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5856 add_to_journal(freefrag->ff_jdep); 5857 return (aip); 5858 } 5859 5860 /* 5861 * Called just before setting an indirect block pointer 5862 * to a newly allocated file page. 5863 */ 5864 void 5865 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5866 struct inode *ip; /* inode for file being extended */ 5867 ufs_lbn_t lbn; /* allocated block number within file */ 5868 struct buf *bp; /* buffer with indirect blk referencing page */ 5869 int ptrno; /* offset of pointer in indirect block */ 5870 ufs2_daddr_t newblkno; /* disk block number being added */ 5871 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5872 struct buf *nbp; /* buffer holding allocated page */ 5873 { 5874 struct inodedep *inodedep; 5875 struct freefrag *freefrag; 5876 struct allocindir *aip; 5877 struct pagedep *pagedep; 5878 struct mount *mp; 5879 struct ufsmount *ump; 5880 5881 mp = ITOVFS(ip); 5882 ump = VFSTOUFS(mp); 5883 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5884 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5885 KASSERT(lbn == nbp->b_lblkno, 5886 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5887 lbn, bp->b_lblkno)); 5888 CTR4(KTR_SUJ, 5889 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5890 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5891 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5892 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5893 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5894 /* 5895 * If we are allocating a directory page, then we must 5896 * allocate an associated pagedep to track additions and 5897 * deletions. 5898 */ 5899 if ((ip->i_mode & IFMT) == IFDIR) 5900 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5901 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5902 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5903 FREE_LOCK(ump); 5904 if (freefrag) 5905 handle_workitem_freefrag(freefrag); 5906 } 5907 5908 /* 5909 * Called just before setting an indirect block pointer to a 5910 * newly allocated indirect block. 5911 */ 5912 void 5913 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5914 struct buf *nbp; /* newly allocated indirect block */ 5915 struct inode *ip; /* inode for file being extended */ 5916 struct buf *bp; /* indirect block referencing allocated block */ 5917 int ptrno; /* offset of pointer in indirect block */ 5918 ufs2_daddr_t newblkno; /* disk block number being added */ 5919 { 5920 struct inodedep *inodedep; 5921 struct allocindir *aip; 5922 struct ufsmount *ump; 5923 ufs_lbn_t lbn; 5924 5925 ump = ITOUMP(ip); 5926 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 5927 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5928 CTR3(KTR_SUJ, 5929 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5930 ip->i_number, newblkno, ptrno); 5931 lbn = nbp->b_lblkno; 5932 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5933 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5934 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 5935 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5936 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5937 panic("softdep_setup_allocindir_meta: Block already existed"); 5938 FREE_LOCK(ump); 5939 } 5940 5941 static void 5942 indirdep_complete(indirdep) 5943 struct indirdep *indirdep; 5944 { 5945 struct allocindir *aip; 5946 5947 LIST_REMOVE(indirdep, ir_next); 5948 indirdep->ir_state |= DEPCOMPLETE; 5949 5950 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5951 LIST_REMOVE(aip, ai_next); 5952 free_newblk(&aip->ai_block); 5953 } 5954 /* 5955 * If this indirdep is not attached to a buf it was simply waiting 5956 * on completion to clear completehd. free_indirdep() asserts 5957 * that nothing is dangling. 5958 */ 5959 if ((indirdep->ir_state & ONWORKLIST) == 0) 5960 free_indirdep(indirdep); 5961 } 5962 5963 static struct indirdep * 5964 indirdep_lookup(mp, ip, bp) 5965 struct mount *mp; 5966 struct inode *ip; 5967 struct buf *bp; 5968 { 5969 struct indirdep *indirdep, *newindirdep; 5970 struct newblk *newblk; 5971 struct ufsmount *ump; 5972 struct worklist *wk; 5973 struct fs *fs; 5974 ufs2_daddr_t blkno; 5975 5976 ump = VFSTOUFS(mp); 5977 LOCK_OWNED(ump); 5978 indirdep = NULL; 5979 newindirdep = NULL; 5980 fs = ump->um_fs; 5981 for (;;) { 5982 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5983 if (wk->wk_type != D_INDIRDEP) 5984 continue; 5985 indirdep = WK_INDIRDEP(wk); 5986 break; 5987 } 5988 /* Found on the buffer worklist, no new structure to free. */ 5989 if (indirdep != NULL && newindirdep == NULL) 5990 return (indirdep); 5991 if (indirdep != NULL && newindirdep != NULL) 5992 panic("indirdep_lookup: simultaneous create"); 5993 /* None found on the buffer and a new structure is ready. */ 5994 if (indirdep == NULL && newindirdep != NULL) 5995 break; 5996 /* None found and no new structure available. */ 5997 FREE_LOCK(ump); 5998 newindirdep = malloc(sizeof(struct indirdep), 5999 M_INDIRDEP, M_SOFTDEP_FLAGS); 6000 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 6001 newindirdep->ir_state = ATTACHED; 6002 if (I_IS_UFS1(ip)) 6003 newindirdep->ir_state |= UFS1FMT; 6004 TAILQ_INIT(&newindirdep->ir_trunc); 6005 newindirdep->ir_saveddata = NULL; 6006 LIST_INIT(&newindirdep->ir_deplisthd); 6007 LIST_INIT(&newindirdep->ir_donehd); 6008 LIST_INIT(&newindirdep->ir_writehd); 6009 LIST_INIT(&newindirdep->ir_completehd); 6010 if (bp->b_blkno == bp->b_lblkno) { 6011 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 6012 NULL, NULL); 6013 bp->b_blkno = blkno; 6014 } 6015 newindirdep->ir_freeblks = NULL; 6016 newindirdep->ir_savebp = 6017 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 6018 newindirdep->ir_bp = bp; 6019 BUF_KERNPROC(newindirdep->ir_savebp); 6020 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 6021 ACQUIRE_LOCK(ump); 6022 } 6023 indirdep = newindirdep; 6024 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 6025 /* 6026 * If the block is not yet allocated we don't set DEPCOMPLETE so 6027 * that we don't free dependencies until the pointers are valid. 6028 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 6029 * than using the hash. 6030 */ 6031 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 6032 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 6033 else 6034 indirdep->ir_state |= DEPCOMPLETE; 6035 return (indirdep); 6036 } 6037 6038 /* 6039 * Called to finish the allocation of the "aip" allocated 6040 * by one of the two routines above. 6041 */ 6042 static struct freefrag * 6043 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 6044 struct buf *bp; /* in-memory copy of the indirect block */ 6045 struct inode *ip; /* inode for file being extended */ 6046 struct inodedep *inodedep; /* Inodedep for ip */ 6047 struct allocindir *aip; /* allocindir allocated by the above routines */ 6048 ufs_lbn_t lbn; /* Logical block number for this block. */ 6049 { 6050 struct fs *fs; 6051 struct indirdep *indirdep; 6052 struct allocindir *oldaip; 6053 struct freefrag *freefrag; 6054 struct mount *mp; 6055 struct ufsmount *ump; 6056 6057 mp = ITOVFS(ip); 6058 ump = VFSTOUFS(mp); 6059 LOCK_OWNED(ump); 6060 fs = ump->um_fs; 6061 if (bp->b_lblkno >= 0) 6062 panic("setup_allocindir_phase2: not indir blk"); 6063 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6064 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6065 indirdep = indirdep_lookup(mp, ip, bp); 6066 KASSERT(indirdep->ir_savebp != NULL, 6067 ("setup_allocindir_phase2 NULL ir_savebp")); 6068 aip->ai_indirdep = indirdep; 6069 /* 6070 * Check for an unwritten dependency for this indirect offset. If 6071 * there is, merge the old dependency into the new one. This happens 6072 * as a result of reallocblk only. 6073 */ 6074 freefrag = NULL; 6075 if (aip->ai_oldblkno != 0) { 6076 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6077 if (oldaip->ai_offset == aip->ai_offset) { 6078 freefrag = allocindir_merge(aip, oldaip); 6079 goto done; 6080 } 6081 } 6082 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6083 if (oldaip->ai_offset == aip->ai_offset) { 6084 freefrag = allocindir_merge(aip, oldaip); 6085 goto done; 6086 } 6087 } 6088 } 6089 done: 6090 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6091 return (freefrag); 6092 } 6093 6094 /* 6095 * Merge two allocindirs which refer to the same block. Move newblock 6096 * dependencies and setup the freefrags appropriately. 6097 */ 6098 static struct freefrag * 6099 allocindir_merge(aip, oldaip) 6100 struct allocindir *aip; 6101 struct allocindir *oldaip; 6102 { 6103 struct freefrag *freefrag; 6104 struct worklist *wk; 6105 6106 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6107 panic("allocindir_merge: blkno"); 6108 aip->ai_oldblkno = oldaip->ai_oldblkno; 6109 freefrag = aip->ai_freefrag; 6110 aip->ai_freefrag = oldaip->ai_freefrag; 6111 oldaip->ai_freefrag = NULL; 6112 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6113 /* 6114 * If we are tracking a new directory-block allocation, 6115 * move it from the old allocindir to the new allocindir. 6116 */ 6117 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6118 WORKLIST_REMOVE(wk); 6119 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6120 panic("allocindir_merge: extra newdirblk"); 6121 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6122 } 6123 /* 6124 * We can skip journaling for this freefrag and just complete 6125 * any pending journal work for the allocindir that is being 6126 * removed after the freefrag completes. 6127 */ 6128 if (freefrag->ff_jdep) 6129 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6130 LIST_REMOVE(oldaip, ai_next); 6131 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6132 &freefrag->ff_list, &freefrag->ff_jwork); 6133 free_newblk(&oldaip->ai_block); 6134 6135 return (freefrag); 6136 } 6137 6138 static inline void 6139 setup_freedirect(freeblks, ip, i, needj) 6140 struct freeblks *freeblks; 6141 struct inode *ip; 6142 int i; 6143 int needj; 6144 { 6145 struct ufsmount *ump; 6146 ufs2_daddr_t blkno; 6147 int frags; 6148 6149 blkno = DIP(ip, i_db[i]); 6150 if (blkno == 0) 6151 return; 6152 DIP_SET(ip, i_db[i], 0); 6153 ump = ITOUMP(ip); 6154 frags = sblksize(ump->um_fs, ip->i_size, i); 6155 frags = numfrags(ump->um_fs, frags); 6156 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj); 6157 } 6158 6159 static inline void 6160 setup_freeext(freeblks, ip, i, needj) 6161 struct freeblks *freeblks; 6162 struct inode *ip; 6163 int i; 6164 int needj; 6165 { 6166 struct ufsmount *ump; 6167 ufs2_daddr_t blkno; 6168 int frags; 6169 6170 blkno = ip->i_din2->di_extb[i]; 6171 if (blkno == 0) 6172 return; 6173 ip->i_din2->di_extb[i] = 0; 6174 ump = ITOUMP(ip); 6175 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i); 6176 frags = numfrags(ump->um_fs, frags); 6177 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6178 } 6179 6180 static inline void 6181 setup_freeindir(freeblks, ip, i, lbn, needj) 6182 struct freeblks *freeblks; 6183 struct inode *ip; 6184 int i; 6185 ufs_lbn_t lbn; 6186 int needj; 6187 { 6188 struct ufsmount *ump; 6189 ufs2_daddr_t blkno; 6190 6191 blkno = DIP(ip, i_ib[i]); 6192 if (blkno == 0) 6193 return; 6194 DIP_SET(ip, i_ib[i], 0); 6195 ump = ITOUMP(ip); 6196 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag, 6197 0, needj); 6198 } 6199 6200 static inline struct freeblks * 6201 newfreeblks(mp, ip) 6202 struct mount *mp; 6203 struct inode *ip; 6204 { 6205 struct freeblks *freeblks; 6206 6207 freeblks = malloc(sizeof(struct freeblks), 6208 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6209 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6210 LIST_INIT(&freeblks->fb_jblkdephd); 6211 LIST_INIT(&freeblks->fb_jwork); 6212 freeblks->fb_ref = 0; 6213 freeblks->fb_cgwait = 0; 6214 freeblks->fb_state = ATTACHED; 6215 freeblks->fb_uid = ip->i_uid; 6216 freeblks->fb_inum = ip->i_number; 6217 freeblks->fb_vtype = ITOV(ip)->v_type; 6218 freeblks->fb_modrev = DIP(ip, i_modrev); 6219 freeblks->fb_devvp = ITODEVVP(ip); 6220 freeblks->fb_chkcnt = 0; 6221 freeblks->fb_len = 0; 6222 6223 return (freeblks); 6224 } 6225 6226 static void 6227 trunc_indirdep(indirdep, freeblks, bp, off) 6228 struct indirdep *indirdep; 6229 struct freeblks *freeblks; 6230 struct buf *bp; 6231 int off; 6232 { 6233 struct allocindir *aip, *aipn; 6234 6235 /* 6236 * The first set of allocindirs won't be in savedbp. 6237 */ 6238 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6239 if (aip->ai_offset > off) 6240 cancel_allocindir(aip, bp, freeblks, 1); 6241 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6242 if (aip->ai_offset > off) 6243 cancel_allocindir(aip, bp, freeblks, 1); 6244 /* 6245 * These will exist in savedbp. 6246 */ 6247 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6248 if (aip->ai_offset > off) 6249 cancel_allocindir(aip, NULL, freeblks, 0); 6250 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6251 if (aip->ai_offset > off) 6252 cancel_allocindir(aip, NULL, freeblks, 0); 6253 } 6254 6255 /* 6256 * Follow the chain of indirects down to lastlbn creating a freework 6257 * structure for each. This will be used to start indir_trunc() at 6258 * the right offset and create the journal records for the parrtial 6259 * truncation. A second step will handle the truncated dependencies. 6260 */ 6261 static int 6262 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6263 struct freeblks *freeblks; 6264 struct inode *ip; 6265 ufs_lbn_t lbn; 6266 ufs_lbn_t lastlbn; 6267 ufs2_daddr_t blkno; 6268 { 6269 struct indirdep *indirdep; 6270 struct indirdep *indirn; 6271 struct freework *freework; 6272 struct newblk *newblk; 6273 struct mount *mp; 6274 struct ufsmount *ump; 6275 struct buf *bp; 6276 uint8_t *start; 6277 uint8_t *end; 6278 ufs_lbn_t lbnadd; 6279 int level; 6280 int error; 6281 int off; 6282 6283 6284 freework = NULL; 6285 if (blkno == 0) 6286 return (0); 6287 mp = freeblks->fb_list.wk_mp; 6288 ump = VFSTOUFS(mp); 6289 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6290 if ((bp->b_flags & B_CACHE) == 0) { 6291 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6292 bp->b_iocmd = BIO_READ; 6293 bp->b_flags &= ~B_INVAL; 6294 bp->b_ioflags &= ~BIO_ERROR; 6295 vfs_busy_pages(bp, 0); 6296 bp->b_iooffset = dbtob(bp->b_blkno); 6297 bstrategy(bp); 6298 #ifdef RACCT 6299 if (racct_enable) { 6300 PROC_LOCK(curproc); 6301 racct_add_buf(curproc, bp, 0); 6302 PROC_UNLOCK(curproc); 6303 } 6304 #endif /* RACCT */ 6305 curthread->td_ru.ru_inblock++; 6306 error = bufwait(bp); 6307 if (error) { 6308 brelse(bp); 6309 return (error); 6310 } 6311 } 6312 level = lbn_level(lbn); 6313 lbnadd = lbn_offset(ump->um_fs, level); 6314 /* 6315 * Compute the offset of the last block we want to keep. Store 6316 * in the freework the first block we want to completely free. 6317 */ 6318 off = (lastlbn - -(lbn + level)) / lbnadd; 6319 if (off + 1 == NINDIR(ump->um_fs)) 6320 goto nowork; 6321 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0); 6322 /* 6323 * Link the freework into the indirdep. This will prevent any new 6324 * allocations from proceeding until we are finished with the 6325 * truncate and the block is written. 6326 */ 6327 ACQUIRE_LOCK(ump); 6328 indirdep = indirdep_lookup(mp, ip, bp); 6329 if (indirdep->ir_freeblks) 6330 panic("setup_trunc_indir: indirdep already truncated."); 6331 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6332 freework->fw_indir = indirdep; 6333 /* 6334 * Cancel any allocindirs that will not make it to disk. 6335 * We have to do this for all copies of the indirdep that 6336 * live on this newblk. 6337 */ 6338 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6339 if (newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0, 6340 &newblk) == 0) 6341 panic("setup_trunc_indir: lost block"); 6342 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6343 trunc_indirdep(indirn, freeblks, bp, off); 6344 } else 6345 trunc_indirdep(indirdep, freeblks, bp, off); 6346 FREE_LOCK(ump); 6347 /* 6348 * Creation is protected by the buf lock. The saveddata is only 6349 * needed if a full truncation follows a partial truncation but it 6350 * is difficult to allocate in that case so we fetch it anyway. 6351 */ 6352 if (indirdep->ir_saveddata == NULL) 6353 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6354 M_SOFTDEP_FLAGS); 6355 nowork: 6356 /* Fetch the blkno of the child and the zero start offset. */ 6357 if (I_IS_UFS1(ip)) { 6358 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6359 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6360 } else { 6361 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6362 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6363 } 6364 if (freework) { 6365 /* Zero the truncated pointers. */ 6366 end = bp->b_data + bp->b_bcount; 6367 bzero(start, end - start); 6368 bdwrite(bp); 6369 } else 6370 bqrelse(bp); 6371 if (level == 0) 6372 return (0); 6373 lbn++; /* adjust level */ 6374 lbn -= (off * lbnadd); 6375 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6376 } 6377 6378 /* 6379 * Complete the partial truncation of an indirect block setup by 6380 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6381 * copy and writes them to disk before the freeblks is allowed to complete. 6382 */ 6383 static void 6384 complete_trunc_indir(freework) 6385 struct freework *freework; 6386 { 6387 struct freework *fwn; 6388 struct indirdep *indirdep; 6389 struct ufsmount *ump; 6390 struct buf *bp; 6391 uintptr_t start; 6392 int count; 6393 6394 ump = VFSTOUFS(freework->fw_list.wk_mp); 6395 LOCK_OWNED(ump); 6396 indirdep = freework->fw_indir; 6397 for (;;) { 6398 bp = indirdep->ir_bp; 6399 /* See if the block was discarded. */ 6400 if (bp == NULL) 6401 break; 6402 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6403 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6404 break; 6405 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6406 LOCK_PTR(ump)) == 0) 6407 BUF_UNLOCK(bp); 6408 ACQUIRE_LOCK(ump); 6409 } 6410 freework->fw_state |= DEPCOMPLETE; 6411 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6412 /* 6413 * Zero the pointers in the saved copy. 6414 */ 6415 if (indirdep->ir_state & UFS1FMT) 6416 start = sizeof(ufs1_daddr_t); 6417 else 6418 start = sizeof(ufs2_daddr_t); 6419 start *= freework->fw_start; 6420 count = indirdep->ir_savebp->b_bcount - start; 6421 start += (uintptr_t)indirdep->ir_savebp->b_data; 6422 bzero((char *)start, count); 6423 /* 6424 * We need to start the next truncation in the list if it has not 6425 * been started yet. 6426 */ 6427 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6428 if (fwn != NULL) { 6429 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6430 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6431 if ((fwn->fw_state & ONWORKLIST) == 0) 6432 freework_enqueue(fwn); 6433 } 6434 /* 6435 * If bp is NULL the block was fully truncated, restore 6436 * the saved block list otherwise free it if it is no 6437 * longer needed. 6438 */ 6439 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6440 if (bp == NULL) 6441 bcopy(indirdep->ir_saveddata, 6442 indirdep->ir_savebp->b_data, 6443 indirdep->ir_savebp->b_bcount); 6444 free(indirdep->ir_saveddata, M_INDIRDEP); 6445 indirdep->ir_saveddata = NULL; 6446 } 6447 /* 6448 * When bp is NULL there is a full truncation pending. We 6449 * must wait for this full truncation to be journaled before 6450 * we can release this freework because the disk pointers will 6451 * never be written as zero. 6452 */ 6453 if (bp == NULL) { 6454 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6455 handle_written_freework(freework); 6456 else 6457 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6458 &freework->fw_list); 6459 } else { 6460 /* Complete when the real copy is written. */ 6461 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6462 BUF_UNLOCK(bp); 6463 } 6464 } 6465 6466 /* 6467 * Calculate the number of blocks we are going to release where datablocks 6468 * is the current total and length is the new file size. 6469 */ 6470 static ufs2_daddr_t 6471 blkcount(fs, datablocks, length) 6472 struct fs *fs; 6473 ufs2_daddr_t datablocks; 6474 off_t length; 6475 { 6476 off_t totblks, numblks; 6477 6478 totblks = 0; 6479 numblks = howmany(length, fs->fs_bsize); 6480 if (numblks <= UFS_NDADDR) { 6481 totblks = howmany(length, fs->fs_fsize); 6482 goto out; 6483 } 6484 totblks = blkstofrags(fs, numblks); 6485 numblks -= UFS_NDADDR; 6486 /* 6487 * Count all single, then double, then triple indirects required. 6488 * Subtracting one indirects worth of blocks for each pass 6489 * acknowledges one of each pointed to by the inode. 6490 */ 6491 for (;;) { 6492 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6493 numblks -= NINDIR(fs); 6494 if (numblks <= 0) 6495 break; 6496 numblks = howmany(numblks, NINDIR(fs)); 6497 } 6498 out: 6499 totblks = fsbtodb(fs, totblks); 6500 /* 6501 * Handle sparse files. We can't reclaim more blocks than the inode 6502 * references. We will correct it later in handle_complete_freeblks() 6503 * when we know the real count. 6504 */ 6505 if (totblks > datablocks) 6506 return (0); 6507 return (datablocks - totblks); 6508 } 6509 6510 /* 6511 * Handle freeblocks for journaled softupdate filesystems. 6512 * 6513 * Contrary to normal softupdates, we must preserve the block pointers in 6514 * indirects until their subordinates are free. This is to avoid journaling 6515 * every block that is freed which may consume more space than the journal 6516 * itself. The recovery program will see the free block journals at the 6517 * base of the truncated area and traverse them to reclaim space. The 6518 * pointers in the inode may be cleared immediately after the journal 6519 * records are written because each direct and indirect pointer in the 6520 * inode is recorded in a journal. This permits full truncation to proceed 6521 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6522 * 6523 * The algorithm is as follows: 6524 * 1) Traverse the in-memory state and create journal entries to release 6525 * the relevant blocks and full indirect trees. 6526 * 2) Traverse the indirect block chain adding partial truncation freework 6527 * records to indirects in the path to lastlbn. The freework will 6528 * prevent new allocation dependencies from being satisfied in this 6529 * indirect until the truncation completes. 6530 * 3) Read and lock the inode block, performing an update with the new size 6531 * and pointers. This prevents truncated data from becoming valid on 6532 * disk through step 4. 6533 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6534 * eliminate journal work for those records that do not require it. 6535 * 5) Schedule the journal records to be written followed by the inode block. 6536 * 6) Allocate any necessary frags for the end of file. 6537 * 7) Zero any partially truncated blocks. 6538 * 6539 * From this truncation proceeds asynchronously using the freework and 6540 * indir_trunc machinery. The file will not be extended again into a 6541 * partially truncated indirect block until all work is completed but 6542 * the normal dependency mechanism ensures that it is rolled back/forward 6543 * as appropriate. Further truncation may occur without delay and is 6544 * serialized in indir_trunc(). 6545 */ 6546 void 6547 softdep_journal_freeblocks(ip, cred, length, flags) 6548 struct inode *ip; /* The inode whose length is to be reduced */ 6549 struct ucred *cred; 6550 off_t length; /* The new length for the file */ 6551 int flags; /* IO_EXT and/or IO_NORMAL */ 6552 { 6553 struct freeblks *freeblks, *fbn; 6554 struct worklist *wk, *wkn; 6555 struct inodedep *inodedep; 6556 struct jblkdep *jblkdep; 6557 struct allocdirect *adp, *adpn; 6558 struct ufsmount *ump; 6559 struct fs *fs; 6560 struct buf *bp; 6561 struct vnode *vp; 6562 struct mount *mp; 6563 ufs2_daddr_t extblocks, datablocks; 6564 ufs_lbn_t tmpval, lbn, lastlbn; 6565 int frags, lastoff, iboff, allocblock, needj, error, i; 6566 6567 ump = ITOUMP(ip); 6568 mp = UFSTOVFS(ump); 6569 fs = ump->um_fs; 6570 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6571 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6572 vp = ITOV(ip); 6573 needj = 1; 6574 iboff = -1; 6575 allocblock = 0; 6576 extblocks = 0; 6577 datablocks = 0; 6578 frags = 0; 6579 freeblks = newfreeblks(mp, ip); 6580 ACQUIRE_LOCK(ump); 6581 /* 6582 * If we're truncating a removed file that will never be written 6583 * we don't need to journal the block frees. The canceled journals 6584 * for the allocations will suffice. 6585 */ 6586 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6587 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6588 length == 0) 6589 needj = 0; 6590 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6591 ip->i_number, length, needj); 6592 FREE_LOCK(ump); 6593 /* 6594 * Calculate the lbn that we are truncating to. This results in -1 6595 * if we're truncating the 0 bytes. So it is the last lbn we want 6596 * to keep, not the first lbn we want to truncate. 6597 */ 6598 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6599 lastoff = blkoff(fs, length); 6600 /* 6601 * Compute frags we are keeping in lastlbn. 0 means all. 6602 */ 6603 if (lastlbn >= 0 && lastlbn < UFS_NDADDR) { 6604 frags = fragroundup(fs, lastoff); 6605 /* adp offset of last valid allocdirect. */ 6606 iboff = lastlbn; 6607 } else if (lastlbn > 0) 6608 iboff = UFS_NDADDR; 6609 if (fs->fs_magic == FS_UFS2_MAGIC) 6610 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6611 /* 6612 * Handle normal data blocks and indirects. This section saves 6613 * values used after the inode update to complete frag and indirect 6614 * truncation. 6615 */ 6616 if ((flags & IO_NORMAL) != 0) { 6617 /* 6618 * Handle truncation of whole direct and indirect blocks. 6619 */ 6620 for (i = iboff + 1; i < UFS_NDADDR; i++) 6621 setup_freedirect(freeblks, ip, i, needj); 6622 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6623 i < UFS_NIADDR; 6624 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6625 /* Release a whole indirect tree. */ 6626 if (lbn > lastlbn) { 6627 setup_freeindir(freeblks, ip, i, -lbn -i, 6628 needj); 6629 continue; 6630 } 6631 iboff = i + UFS_NDADDR; 6632 /* 6633 * Traverse partially truncated indirect tree. 6634 */ 6635 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6636 setup_trunc_indir(freeblks, ip, -lbn - i, 6637 lastlbn, DIP(ip, i_ib[i])); 6638 } 6639 /* 6640 * Handle partial truncation to a frag boundary. 6641 */ 6642 if (frags) { 6643 ufs2_daddr_t blkno; 6644 long oldfrags; 6645 6646 oldfrags = blksize(fs, ip, lastlbn); 6647 blkno = DIP(ip, i_db[lastlbn]); 6648 if (blkno && oldfrags != frags) { 6649 oldfrags -= frags; 6650 oldfrags = numfrags(fs, oldfrags); 6651 blkno += numfrags(fs, frags); 6652 newfreework(ump, freeblks, NULL, lastlbn, 6653 blkno, oldfrags, 0, needj); 6654 if (needj) 6655 adjust_newfreework(freeblks, 6656 numfrags(fs, frags)); 6657 } else if (blkno == 0) 6658 allocblock = 1; 6659 } 6660 /* 6661 * Add a journal record for partial truncate if we are 6662 * handling indirect blocks. Non-indirects need no extra 6663 * journaling. 6664 */ 6665 if (length != 0 && lastlbn >= UFS_NDADDR) { 6666 ip->i_flag |= IN_TRUNCATED; 6667 newjtrunc(freeblks, length, 0); 6668 } 6669 ip->i_size = length; 6670 DIP_SET(ip, i_size, ip->i_size); 6671 datablocks = DIP(ip, i_blocks) - extblocks; 6672 if (length != 0) 6673 datablocks = blkcount(fs, datablocks, length); 6674 freeblks->fb_len = length; 6675 } 6676 if ((flags & IO_EXT) != 0) { 6677 for (i = 0; i < UFS_NXADDR; i++) 6678 setup_freeext(freeblks, ip, i, needj); 6679 ip->i_din2->di_extsize = 0; 6680 datablocks += extblocks; 6681 } 6682 #ifdef QUOTA 6683 /* Reference the quotas in case the block count is wrong in the end. */ 6684 quotaref(vp, freeblks->fb_quota); 6685 (void) chkdq(ip, -datablocks, NOCRED, 0); 6686 #endif 6687 freeblks->fb_chkcnt = -datablocks; 6688 UFS_LOCK(ump); 6689 fs->fs_pendingblocks += datablocks; 6690 UFS_UNLOCK(ump); 6691 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6692 /* 6693 * Handle truncation of incomplete alloc direct dependencies. We 6694 * hold the inode block locked to prevent incomplete dependencies 6695 * from reaching the disk while we are eliminating those that 6696 * have been truncated. This is a partially inlined ffs_update(). 6697 */ 6698 ufs_itimes(vp); 6699 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6700 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6701 (int)fs->fs_bsize, cred, &bp); 6702 if (error) { 6703 brelse(bp); 6704 softdep_error("softdep_journal_freeblocks", error); 6705 return; 6706 } 6707 if (bp->b_bufsize == fs->fs_bsize) 6708 bp->b_flags |= B_CLUSTEROK; 6709 softdep_update_inodeblock(ip, bp, 0); 6710 if (ump->um_fstype == UFS1) { 6711 *((struct ufs1_dinode *)bp->b_data + 6712 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6713 } else { 6714 ffs_update_dinode_ckhash(fs, ip->i_din2); 6715 *((struct ufs2_dinode *)bp->b_data + 6716 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6717 } 6718 ACQUIRE_LOCK(ump); 6719 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6720 if ((inodedep->id_state & IOSTARTED) != 0) 6721 panic("softdep_setup_freeblocks: inode busy"); 6722 /* 6723 * Add the freeblks structure to the list of operations that 6724 * must await the zero'ed inode being written to disk. If we 6725 * still have a bitmap dependency (needj), then the inode 6726 * has never been written to disk, so we can process the 6727 * freeblks below once we have deleted the dependencies. 6728 */ 6729 if (needj) 6730 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6731 else 6732 freeblks->fb_state |= COMPLETE; 6733 if ((flags & IO_NORMAL) != 0) { 6734 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6735 if (adp->ad_offset > iboff) 6736 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6737 freeblks); 6738 /* 6739 * Truncate the allocdirect. We could eliminate 6740 * or modify journal records as well. 6741 */ 6742 else if (adp->ad_offset == iboff && frags) 6743 adp->ad_newsize = frags; 6744 } 6745 } 6746 if ((flags & IO_EXT) != 0) 6747 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6748 cancel_allocdirect(&inodedep->id_extupdt, adp, 6749 freeblks); 6750 /* 6751 * Scan the bufwait list for newblock dependencies that will never 6752 * make it to disk. 6753 */ 6754 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6755 if (wk->wk_type != D_ALLOCDIRECT) 6756 continue; 6757 adp = WK_ALLOCDIRECT(wk); 6758 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6759 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6760 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6761 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6762 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6763 } 6764 } 6765 /* 6766 * Add journal work. 6767 */ 6768 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6769 add_to_journal(&jblkdep->jb_list); 6770 FREE_LOCK(ump); 6771 bdwrite(bp); 6772 /* 6773 * Truncate dependency structures beyond length. 6774 */ 6775 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6776 /* 6777 * This is only set when we need to allocate a fragment because 6778 * none existed at the end of a frag-sized file. It handles only 6779 * allocating a new, zero filled block. 6780 */ 6781 if (allocblock) { 6782 ip->i_size = length - lastoff; 6783 DIP_SET(ip, i_size, ip->i_size); 6784 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6785 if (error != 0) { 6786 softdep_error("softdep_journal_freeblks", error); 6787 return; 6788 } 6789 ip->i_size = length; 6790 DIP_SET(ip, i_size, length); 6791 ip->i_flag |= IN_CHANGE | IN_UPDATE; 6792 allocbuf(bp, frags); 6793 ffs_update(vp, 0); 6794 bawrite(bp); 6795 } else if (lastoff != 0 && vp->v_type != VDIR) { 6796 int size; 6797 6798 /* 6799 * Zero the end of a truncated frag or block. 6800 */ 6801 size = sblksize(fs, length, lastlbn); 6802 error = bread(vp, lastlbn, size, cred, &bp); 6803 if (error) { 6804 softdep_error("softdep_journal_freeblks", error); 6805 return; 6806 } 6807 bzero((char *)bp->b_data + lastoff, size - lastoff); 6808 bawrite(bp); 6809 6810 } 6811 ACQUIRE_LOCK(ump); 6812 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6813 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6814 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6815 /* 6816 * We zero earlier truncations so they don't erroneously 6817 * update i_blocks. 6818 */ 6819 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6820 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6821 fbn->fb_len = 0; 6822 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6823 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6824 freeblks->fb_state |= INPROGRESS; 6825 else 6826 freeblks = NULL; 6827 FREE_LOCK(ump); 6828 if (freeblks) 6829 handle_workitem_freeblocks(freeblks, 0); 6830 trunc_pages(ip, length, extblocks, flags); 6831 6832 } 6833 6834 /* 6835 * Flush a JOP_SYNC to the journal. 6836 */ 6837 void 6838 softdep_journal_fsync(ip) 6839 struct inode *ip; 6840 { 6841 struct jfsync *jfsync; 6842 struct ufsmount *ump; 6843 6844 ump = ITOUMP(ip); 6845 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 6846 ("softdep_journal_fsync called on non-softdep filesystem")); 6847 if ((ip->i_flag & IN_TRUNCATED) == 0) 6848 return; 6849 ip->i_flag &= ~IN_TRUNCATED; 6850 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6851 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump)); 6852 jfsync->jfs_size = ip->i_size; 6853 jfsync->jfs_ino = ip->i_number; 6854 ACQUIRE_LOCK(ump); 6855 add_to_journal(&jfsync->jfs_list); 6856 jwait(&jfsync->jfs_list, MNT_WAIT); 6857 FREE_LOCK(ump); 6858 } 6859 6860 /* 6861 * Block de-allocation dependencies. 6862 * 6863 * When blocks are de-allocated, the on-disk pointers must be nullified before 6864 * the blocks are made available for use by other files. (The true 6865 * requirement is that old pointers must be nullified before new on-disk 6866 * pointers are set. We chose this slightly more stringent requirement to 6867 * reduce complexity.) Our implementation handles this dependency by updating 6868 * the inode (or indirect block) appropriately but delaying the actual block 6869 * de-allocation (i.e., freemap and free space count manipulation) until 6870 * after the updated versions reach stable storage. After the disk is 6871 * updated, the blocks can be safely de-allocated whenever it is convenient. 6872 * This implementation handles only the common case of reducing a file's 6873 * length to zero. Other cases are handled by the conventional synchronous 6874 * write approach. 6875 * 6876 * The ffs implementation with which we worked double-checks 6877 * the state of the block pointers and file size as it reduces 6878 * a file's length. Some of this code is replicated here in our 6879 * soft updates implementation. The freeblks->fb_chkcnt field is 6880 * used to transfer a part of this information to the procedure 6881 * that eventually de-allocates the blocks. 6882 * 6883 * This routine should be called from the routine that shortens 6884 * a file's length, before the inode's size or block pointers 6885 * are modified. It will save the block pointer information for 6886 * later release and zero the inode so that the calling routine 6887 * can release it. 6888 */ 6889 void 6890 softdep_setup_freeblocks(ip, length, flags) 6891 struct inode *ip; /* The inode whose length is to be reduced */ 6892 off_t length; /* The new length for the file */ 6893 int flags; /* IO_EXT and/or IO_NORMAL */ 6894 { 6895 struct ufs1_dinode *dp1; 6896 struct ufs2_dinode *dp2; 6897 struct freeblks *freeblks; 6898 struct inodedep *inodedep; 6899 struct allocdirect *adp; 6900 struct ufsmount *ump; 6901 struct buf *bp; 6902 struct fs *fs; 6903 ufs2_daddr_t extblocks, datablocks; 6904 struct mount *mp; 6905 int i, delay, error; 6906 ufs_lbn_t tmpval; 6907 ufs_lbn_t lbn; 6908 6909 ump = ITOUMP(ip); 6910 mp = UFSTOVFS(ump); 6911 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6912 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6913 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6914 ip->i_number, length); 6915 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6916 fs = ump->um_fs; 6917 if ((error = bread(ump->um_devvp, 6918 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6919 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6920 brelse(bp); 6921 softdep_error("softdep_setup_freeblocks", error); 6922 return; 6923 } 6924 freeblks = newfreeblks(mp, ip); 6925 extblocks = 0; 6926 datablocks = 0; 6927 if (fs->fs_magic == FS_UFS2_MAGIC) 6928 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6929 if ((flags & IO_NORMAL) != 0) { 6930 for (i = 0; i < UFS_NDADDR; i++) 6931 setup_freedirect(freeblks, ip, i, 0); 6932 for (i = 0, tmpval = NINDIR(fs), lbn = UFS_NDADDR; 6933 i < UFS_NIADDR; 6934 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6935 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6936 ip->i_size = 0; 6937 DIP_SET(ip, i_size, 0); 6938 datablocks = DIP(ip, i_blocks) - extblocks; 6939 } 6940 if ((flags & IO_EXT) != 0) { 6941 for (i = 0; i < UFS_NXADDR; i++) 6942 setup_freeext(freeblks, ip, i, 0); 6943 ip->i_din2->di_extsize = 0; 6944 datablocks += extblocks; 6945 } 6946 #ifdef QUOTA 6947 /* Reference the quotas in case the block count is wrong in the end. */ 6948 quotaref(ITOV(ip), freeblks->fb_quota); 6949 (void) chkdq(ip, -datablocks, NOCRED, 0); 6950 #endif 6951 freeblks->fb_chkcnt = -datablocks; 6952 UFS_LOCK(ump); 6953 fs->fs_pendingblocks += datablocks; 6954 UFS_UNLOCK(ump); 6955 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6956 /* 6957 * Push the zero'ed inode to its disk buffer so that we are free 6958 * to delete its dependencies below. Once the dependencies are gone 6959 * the buffer can be safely released. 6960 */ 6961 if (ump->um_fstype == UFS1) { 6962 dp1 = ((struct ufs1_dinode *)bp->b_data + 6963 ino_to_fsbo(fs, ip->i_number)); 6964 ip->i_din1->di_freelink = dp1->di_freelink; 6965 *dp1 = *ip->i_din1; 6966 } else { 6967 dp2 = ((struct ufs2_dinode *)bp->b_data + 6968 ino_to_fsbo(fs, ip->i_number)); 6969 ip->i_din2->di_freelink = dp2->di_freelink; 6970 ffs_update_dinode_ckhash(fs, ip->i_din2); 6971 *dp2 = *ip->i_din2; 6972 } 6973 /* 6974 * Find and eliminate any inode dependencies. 6975 */ 6976 ACQUIRE_LOCK(ump); 6977 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6978 if ((inodedep->id_state & IOSTARTED) != 0) 6979 panic("softdep_setup_freeblocks: inode busy"); 6980 /* 6981 * Add the freeblks structure to the list of operations that 6982 * must await the zero'ed inode being written to disk. If we 6983 * still have a bitmap dependency (delay == 0), then the inode 6984 * has never been written to disk, so we can process the 6985 * freeblks below once we have deleted the dependencies. 6986 */ 6987 delay = (inodedep->id_state & DEPCOMPLETE); 6988 if (delay) 6989 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6990 else 6991 freeblks->fb_state |= COMPLETE; 6992 /* 6993 * Because the file length has been truncated to zero, any 6994 * pending block allocation dependency structures associated 6995 * with this inode are obsolete and can simply be de-allocated. 6996 * We must first merge the two dependency lists to get rid of 6997 * any duplicate freefrag structures, then purge the merged list. 6998 * If we still have a bitmap dependency, then the inode has never 6999 * been written to disk, so we can free any fragments without delay. 7000 */ 7001 if (flags & IO_NORMAL) { 7002 merge_inode_lists(&inodedep->id_newinoupdt, 7003 &inodedep->id_inoupdt); 7004 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 7005 cancel_allocdirect(&inodedep->id_inoupdt, adp, 7006 freeblks); 7007 } 7008 if (flags & IO_EXT) { 7009 merge_inode_lists(&inodedep->id_newextupdt, 7010 &inodedep->id_extupdt); 7011 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 7012 cancel_allocdirect(&inodedep->id_extupdt, adp, 7013 freeblks); 7014 } 7015 FREE_LOCK(ump); 7016 bdwrite(bp); 7017 trunc_dependencies(ip, freeblks, -1, 0, flags); 7018 ACQUIRE_LOCK(ump); 7019 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 7020 (void) free_inodedep(inodedep); 7021 freeblks->fb_state |= DEPCOMPLETE; 7022 /* 7023 * If the inode with zeroed block pointers is now on disk 7024 * we can start freeing blocks. 7025 */ 7026 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 7027 freeblks->fb_state |= INPROGRESS; 7028 else 7029 freeblks = NULL; 7030 FREE_LOCK(ump); 7031 if (freeblks) 7032 handle_workitem_freeblocks(freeblks, 0); 7033 trunc_pages(ip, length, extblocks, flags); 7034 } 7035 7036 /* 7037 * Eliminate pages from the page cache that back parts of this inode and 7038 * adjust the vnode pager's idea of our size. This prevents stale data 7039 * from hanging around in the page cache. 7040 */ 7041 static void 7042 trunc_pages(ip, length, extblocks, flags) 7043 struct inode *ip; 7044 off_t length; 7045 ufs2_daddr_t extblocks; 7046 int flags; 7047 { 7048 struct vnode *vp; 7049 struct fs *fs; 7050 ufs_lbn_t lbn; 7051 off_t end, extend; 7052 7053 vp = ITOV(ip); 7054 fs = ITOFS(ip); 7055 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 7056 if ((flags & IO_EXT) != 0) 7057 vn_pages_remove(vp, extend, 0); 7058 if ((flags & IO_NORMAL) == 0) 7059 return; 7060 BO_LOCK(&vp->v_bufobj); 7061 drain_output(vp); 7062 BO_UNLOCK(&vp->v_bufobj); 7063 /* 7064 * The vnode pager eliminates file pages we eliminate indirects 7065 * below. 7066 */ 7067 vnode_pager_setsize(vp, length); 7068 /* 7069 * Calculate the end based on the last indirect we want to keep. If 7070 * the block extends into indirects we can just use the negative of 7071 * its lbn. Doubles and triples exist at lower numbers so we must 7072 * be careful not to remove those, if they exist. double and triple 7073 * indirect lbns do not overlap with others so it is not important 7074 * to verify how many levels are required. 7075 */ 7076 lbn = lblkno(fs, length); 7077 if (lbn >= UFS_NDADDR) { 7078 /* Calculate the virtual lbn of the triple indirect. */ 7079 lbn = -lbn - (UFS_NIADDR - 1); 7080 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7081 } else 7082 end = extend; 7083 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7084 } 7085 7086 /* 7087 * See if the buf bp is in the range eliminated by truncation. 7088 */ 7089 static int 7090 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7091 struct buf *bp; 7092 int *blkoffp; 7093 ufs_lbn_t lastlbn; 7094 int lastoff; 7095 int flags; 7096 { 7097 ufs_lbn_t lbn; 7098 7099 *blkoffp = 0; 7100 /* Only match ext/normal blocks as appropriate. */ 7101 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7102 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7103 return (0); 7104 /* ALTDATA is always a full truncation. */ 7105 if ((bp->b_xflags & BX_ALTDATA) != 0) 7106 return (1); 7107 /* -1 is full truncation. */ 7108 if (lastlbn == -1) 7109 return (1); 7110 /* 7111 * If this is a partial truncate we only want those 7112 * blocks and indirect blocks that cover the range 7113 * we're after. 7114 */ 7115 lbn = bp->b_lblkno; 7116 if (lbn < 0) 7117 lbn = -(lbn + lbn_level(lbn)); 7118 if (lbn < lastlbn) 7119 return (0); 7120 /* Here we only truncate lblkno if it's partial. */ 7121 if (lbn == lastlbn) { 7122 if (lastoff == 0) 7123 return (0); 7124 *blkoffp = lastoff; 7125 } 7126 return (1); 7127 } 7128 7129 /* 7130 * Eliminate any dependencies that exist in memory beyond lblkno:off 7131 */ 7132 static void 7133 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7134 struct inode *ip; 7135 struct freeblks *freeblks; 7136 ufs_lbn_t lastlbn; 7137 int lastoff; 7138 int flags; 7139 { 7140 struct bufobj *bo; 7141 struct vnode *vp; 7142 struct buf *bp; 7143 int blkoff; 7144 7145 /* 7146 * We must wait for any I/O in progress to finish so that 7147 * all potential buffers on the dirty list will be visible. 7148 * Once they are all there, walk the list and get rid of 7149 * any dependencies. 7150 */ 7151 vp = ITOV(ip); 7152 bo = &vp->v_bufobj; 7153 BO_LOCK(bo); 7154 drain_output(vp); 7155 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7156 bp->b_vflags &= ~BV_SCANNED; 7157 restart: 7158 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7159 if (bp->b_vflags & BV_SCANNED) 7160 continue; 7161 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7162 bp->b_vflags |= BV_SCANNED; 7163 continue; 7164 } 7165 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7166 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7167 goto restart; 7168 BO_UNLOCK(bo); 7169 if (deallocate_dependencies(bp, freeblks, blkoff)) 7170 bqrelse(bp); 7171 else 7172 brelse(bp); 7173 BO_LOCK(bo); 7174 goto restart; 7175 } 7176 /* 7177 * Now do the work of vtruncbuf while also matching indirect blocks. 7178 */ 7179 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7180 bp->b_vflags &= ~BV_SCANNED; 7181 cleanrestart: 7182 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7183 if (bp->b_vflags & BV_SCANNED) 7184 continue; 7185 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7186 bp->b_vflags |= BV_SCANNED; 7187 continue; 7188 } 7189 if (BUF_LOCK(bp, 7190 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7191 BO_LOCKPTR(bo)) == ENOLCK) { 7192 BO_LOCK(bo); 7193 goto cleanrestart; 7194 } 7195 bp->b_vflags |= BV_SCANNED; 7196 bremfree(bp); 7197 if (blkoff != 0) { 7198 allocbuf(bp, blkoff); 7199 bqrelse(bp); 7200 } else { 7201 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7202 brelse(bp); 7203 } 7204 BO_LOCK(bo); 7205 goto cleanrestart; 7206 } 7207 drain_output(vp); 7208 BO_UNLOCK(bo); 7209 } 7210 7211 static int 7212 cancel_pagedep(pagedep, freeblks, blkoff) 7213 struct pagedep *pagedep; 7214 struct freeblks *freeblks; 7215 int blkoff; 7216 { 7217 struct jremref *jremref; 7218 struct jmvref *jmvref; 7219 struct dirrem *dirrem, *tmp; 7220 int i; 7221 7222 /* 7223 * Copy any directory remove dependencies to the list 7224 * to be processed after the freeblks proceeds. If 7225 * directory entry never made it to disk they 7226 * can be dumped directly onto the work list. 7227 */ 7228 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7229 /* Skip this directory removal if it is intended to remain. */ 7230 if (dirrem->dm_offset < blkoff) 7231 continue; 7232 /* 7233 * If there are any dirrems we wait for the journal write 7234 * to complete and then restart the buf scan as the lock 7235 * has been dropped. 7236 */ 7237 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7238 jwait(&jremref->jr_list, MNT_WAIT); 7239 return (ERESTART); 7240 } 7241 LIST_REMOVE(dirrem, dm_next); 7242 dirrem->dm_dirinum = pagedep->pd_ino; 7243 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7244 } 7245 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7246 jwait(&jmvref->jm_list, MNT_WAIT); 7247 return (ERESTART); 7248 } 7249 /* 7250 * When we're partially truncating a pagedep we just want to flush 7251 * journal entries and return. There can not be any adds in the 7252 * truncated portion of the directory and newblk must remain if 7253 * part of the block remains. 7254 */ 7255 if (blkoff != 0) { 7256 struct diradd *dap; 7257 7258 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7259 if (dap->da_offset > blkoff) 7260 panic("cancel_pagedep: diradd %p off %d > %d", 7261 dap, dap->da_offset, blkoff); 7262 for (i = 0; i < DAHASHSZ; i++) 7263 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7264 if (dap->da_offset > blkoff) 7265 panic("cancel_pagedep: diradd %p off %d > %d", 7266 dap, dap->da_offset, blkoff); 7267 return (0); 7268 } 7269 /* 7270 * There should be no directory add dependencies present 7271 * as the directory could not be truncated until all 7272 * children were removed. 7273 */ 7274 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7275 ("deallocate_dependencies: pendinghd != NULL")); 7276 for (i = 0; i < DAHASHSZ; i++) 7277 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7278 ("deallocate_dependencies: diraddhd != NULL")); 7279 if ((pagedep->pd_state & NEWBLOCK) != 0) 7280 free_newdirblk(pagedep->pd_newdirblk); 7281 if (free_pagedep(pagedep) == 0) 7282 panic("Failed to free pagedep %p", pagedep); 7283 return (0); 7284 } 7285 7286 /* 7287 * Reclaim any dependency structures from a buffer that is about to 7288 * be reallocated to a new vnode. The buffer must be locked, thus, 7289 * no I/O completion operations can occur while we are manipulating 7290 * its associated dependencies. The mutex is held so that other I/O's 7291 * associated with related dependencies do not occur. 7292 */ 7293 static int 7294 deallocate_dependencies(bp, freeblks, off) 7295 struct buf *bp; 7296 struct freeblks *freeblks; 7297 int off; 7298 { 7299 struct indirdep *indirdep; 7300 struct pagedep *pagedep; 7301 struct worklist *wk, *wkn; 7302 struct ufsmount *ump; 7303 7304 ump = softdep_bp_to_mp(bp); 7305 if (ump == NULL) 7306 goto done; 7307 ACQUIRE_LOCK(ump); 7308 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7309 switch (wk->wk_type) { 7310 case D_INDIRDEP: 7311 indirdep = WK_INDIRDEP(wk); 7312 if (bp->b_lblkno >= 0 || 7313 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7314 panic("deallocate_dependencies: not indir"); 7315 cancel_indirdep(indirdep, bp, freeblks); 7316 continue; 7317 7318 case D_PAGEDEP: 7319 pagedep = WK_PAGEDEP(wk); 7320 if (cancel_pagedep(pagedep, freeblks, off)) { 7321 FREE_LOCK(ump); 7322 return (ERESTART); 7323 } 7324 continue; 7325 7326 case D_ALLOCINDIR: 7327 /* 7328 * Simply remove the allocindir, we'll find it via 7329 * the indirdep where we can clear pointers if 7330 * needed. 7331 */ 7332 WORKLIST_REMOVE(wk); 7333 continue; 7334 7335 case D_FREEWORK: 7336 /* 7337 * A truncation is waiting for the zero'd pointers 7338 * to be written. It can be freed when the freeblks 7339 * is journaled. 7340 */ 7341 WORKLIST_REMOVE(wk); 7342 wk->wk_state |= ONDEPLIST; 7343 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7344 break; 7345 7346 case D_ALLOCDIRECT: 7347 if (off != 0) 7348 continue; 7349 /* FALLTHROUGH */ 7350 default: 7351 panic("deallocate_dependencies: Unexpected type %s", 7352 TYPENAME(wk->wk_type)); 7353 /* NOTREACHED */ 7354 } 7355 } 7356 FREE_LOCK(ump); 7357 done: 7358 /* 7359 * Don't throw away this buf, we were partially truncating and 7360 * some deps may always remain. 7361 */ 7362 if (off) { 7363 allocbuf(bp, off); 7364 bp->b_vflags |= BV_SCANNED; 7365 return (EBUSY); 7366 } 7367 bp->b_flags |= B_INVAL | B_NOCACHE; 7368 7369 return (0); 7370 } 7371 7372 /* 7373 * An allocdirect is being canceled due to a truncate. We must make sure 7374 * the journal entry is released in concert with the blkfree that releases 7375 * the storage. Completed journal entries must not be released until the 7376 * space is no longer pointed to by the inode or in the bitmap. 7377 */ 7378 static void 7379 cancel_allocdirect(adphead, adp, freeblks) 7380 struct allocdirectlst *adphead; 7381 struct allocdirect *adp; 7382 struct freeblks *freeblks; 7383 { 7384 struct freework *freework; 7385 struct newblk *newblk; 7386 struct worklist *wk; 7387 7388 TAILQ_REMOVE(adphead, adp, ad_next); 7389 newblk = (struct newblk *)adp; 7390 freework = NULL; 7391 /* 7392 * Find the correct freework structure. 7393 */ 7394 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7395 if (wk->wk_type != D_FREEWORK) 7396 continue; 7397 freework = WK_FREEWORK(wk); 7398 if (freework->fw_blkno == newblk->nb_newblkno) 7399 break; 7400 } 7401 if (freework == NULL) 7402 panic("cancel_allocdirect: Freework not found"); 7403 /* 7404 * If a newblk exists at all we still have the journal entry that 7405 * initiated the allocation so we do not need to journal the free. 7406 */ 7407 cancel_jfreeblk(freeblks, freework->fw_blkno); 7408 /* 7409 * If the journal hasn't been written the jnewblk must be passed 7410 * to the call to ffs_blkfree that reclaims the space. We accomplish 7411 * this by linking the journal dependency into the freework to be 7412 * freed when freework_freeblock() is called. If the journal has 7413 * been written we can simply reclaim the journal space when the 7414 * freeblks work is complete. 7415 */ 7416 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7417 &freeblks->fb_jwork); 7418 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7419 } 7420 7421 7422 /* 7423 * Cancel a new block allocation. May be an indirect or direct block. We 7424 * remove it from various lists and return any journal record that needs to 7425 * be resolved by the caller. 7426 * 7427 * A special consideration is made for indirects which were never pointed 7428 * at on disk and will never be found once this block is released. 7429 */ 7430 static struct jnewblk * 7431 cancel_newblk(newblk, wk, wkhd) 7432 struct newblk *newblk; 7433 struct worklist *wk; 7434 struct workhead *wkhd; 7435 { 7436 struct jnewblk *jnewblk; 7437 7438 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7439 7440 newblk->nb_state |= GOINGAWAY; 7441 /* 7442 * Previously we traversed the completedhd on each indirdep 7443 * attached to this newblk to cancel them and gather journal 7444 * work. Since we need only the oldest journal segment and 7445 * the lowest point on the tree will always have the oldest 7446 * journal segment we are free to release the segments 7447 * of any subordinates and may leave the indirdep list to 7448 * indirdep_complete() when this newblk is freed. 7449 */ 7450 if (newblk->nb_state & ONDEPLIST) { 7451 newblk->nb_state &= ~ONDEPLIST; 7452 LIST_REMOVE(newblk, nb_deps); 7453 } 7454 if (newblk->nb_state & ONWORKLIST) 7455 WORKLIST_REMOVE(&newblk->nb_list); 7456 /* 7457 * If the journal entry hasn't been written we save a pointer to 7458 * the dependency that frees it until it is written or the 7459 * superseding operation completes. 7460 */ 7461 jnewblk = newblk->nb_jnewblk; 7462 if (jnewblk != NULL && wk != NULL) { 7463 newblk->nb_jnewblk = NULL; 7464 jnewblk->jn_dep = wk; 7465 } 7466 if (!LIST_EMPTY(&newblk->nb_jwork)) 7467 jwork_move(wkhd, &newblk->nb_jwork); 7468 /* 7469 * When truncating we must free the newdirblk early to remove 7470 * the pagedep from the hash before returning. 7471 */ 7472 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7473 free_newdirblk(WK_NEWDIRBLK(wk)); 7474 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7475 panic("cancel_newblk: extra newdirblk"); 7476 7477 return (jnewblk); 7478 } 7479 7480 /* 7481 * Schedule the freefrag associated with a newblk to be released once 7482 * the pointers are written and the previous block is no longer needed. 7483 */ 7484 static void 7485 newblk_freefrag(newblk) 7486 struct newblk *newblk; 7487 { 7488 struct freefrag *freefrag; 7489 7490 if (newblk->nb_freefrag == NULL) 7491 return; 7492 freefrag = newblk->nb_freefrag; 7493 newblk->nb_freefrag = NULL; 7494 freefrag->ff_state |= COMPLETE; 7495 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7496 add_to_worklist(&freefrag->ff_list, 0); 7497 } 7498 7499 /* 7500 * Free a newblk. Generate a new freefrag work request if appropriate. 7501 * This must be called after the inode pointer and any direct block pointers 7502 * are valid or fully removed via truncate or frag extension. 7503 */ 7504 static void 7505 free_newblk(newblk) 7506 struct newblk *newblk; 7507 { 7508 struct indirdep *indirdep; 7509 struct worklist *wk; 7510 7511 KASSERT(newblk->nb_jnewblk == NULL, 7512 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7513 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7514 ("free_newblk: unclaimed newblk")); 7515 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7516 newblk_freefrag(newblk); 7517 if (newblk->nb_state & ONDEPLIST) 7518 LIST_REMOVE(newblk, nb_deps); 7519 if (newblk->nb_state & ONWORKLIST) 7520 WORKLIST_REMOVE(&newblk->nb_list); 7521 LIST_REMOVE(newblk, nb_hash); 7522 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7523 free_newdirblk(WK_NEWDIRBLK(wk)); 7524 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7525 panic("free_newblk: extra newdirblk"); 7526 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7527 indirdep_complete(indirdep); 7528 handle_jwork(&newblk->nb_jwork); 7529 WORKITEM_FREE(newblk, D_NEWBLK); 7530 } 7531 7532 /* 7533 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7534 */ 7535 static void 7536 free_newdirblk(newdirblk) 7537 struct newdirblk *newdirblk; 7538 { 7539 struct pagedep *pagedep; 7540 struct diradd *dap; 7541 struct worklist *wk; 7542 7543 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7544 WORKLIST_REMOVE(&newdirblk->db_list); 7545 /* 7546 * If the pagedep is still linked onto the directory buffer 7547 * dependency chain, then some of the entries on the 7548 * pd_pendinghd list may not be committed to disk yet. In 7549 * this case, we will simply clear the NEWBLOCK flag and 7550 * let the pd_pendinghd list be processed when the pagedep 7551 * is next written. If the pagedep is no longer on the buffer 7552 * dependency chain, then all the entries on the pd_pending 7553 * list are committed to disk and we can free them here. 7554 */ 7555 pagedep = newdirblk->db_pagedep; 7556 pagedep->pd_state &= ~NEWBLOCK; 7557 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7558 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7559 free_diradd(dap, NULL); 7560 /* 7561 * If no dependencies remain, the pagedep will be freed. 7562 */ 7563 free_pagedep(pagedep); 7564 } 7565 /* Should only ever be one item in the list. */ 7566 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7567 WORKLIST_REMOVE(wk); 7568 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7569 } 7570 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7571 } 7572 7573 /* 7574 * Prepare an inode to be freed. The actual free operation is not 7575 * done until the zero'ed inode has been written to disk. 7576 */ 7577 void 7578 softdep_freefile(pvp, ino, mode) 7579 struct vnode *pvp; 7580 ino_t ino; 7581 int mode; 7582 { 7583 struct inode *ip = VTOI(pvp); 7584 struct inodedep *inodedep; 7585 struct freefile *freefile; 7586 struct freeblks *freeblks; 7587 struct ufsmount *ump; 7588 7589 ump = ITOUMP(ip); 7590 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7591 ("softdep_freefile called on non-softdep filesystem")); 7592 /* 7593 * This sets up the inode de-allocation dependency. 7594 */ 7595 freefile = malloc(sizeof(struct freefile), 7596 M_FREEFILE, M_SOFTDEP_FLAGS); 7597 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7598 freefile->fx_mode = mode; 7599 freefile->fx_oldinum = ino; 7600 freefile->fx_devvp = ump->um_devvp; 7601 LIST_INIT(&freefile->fx_jwork); 7602 UFS_LOCK(ump); 7603 ump->um_fs->fs_pendinginodes += 1; 7604 UFS_UNLOCK(ump); 7605 7606 /* 7607 * If the inodedep does not exist, then the zero'ed inode has 7608 * been written to disk. If the allocated inode has never been 7609 * written to disk, then the on-disk inode is zero'ed. In either 7610 * case we can free the file immediately. If the journal was 7611 * canceled before being written the inode will never make it to 7612 * disk and we must send the canceled journal entrys to 7613 * ffs_freefile() to be cleared in conjunction with the bitmap. 7614 * Any blocks waiting on the inode to write can be safely freed 7615 * here as it will never been written. 7616 */ 7617 ACQUIRE_LOCK(ump); 7618 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7619 if (inodedep) { 7620 /* 7621 * Clear out freeblks that no longer need to reference 7622 * this inode. 7623 */ 7624 while ((freeblks = 7625 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7626 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7627 fb_next); 7628 freeblks->fb_state &= ~ONDEPLIST; 7629 } 7630 /* 7631 * Remove this inode from the unlinked list. 7632 */ 7633 if (inodedep->id_state & UNLINKED) { 7634 /* 7635 * Save the journal work to be freed with the bitmap 7636 * before we clear UNLINKED. Otherwise it can be lost 7637 * if the inode block is written. 7638 */ 7639 handle_bufwait(inodedep, &freefile->fx_jwork); 7640 clear_unlinked_inodedep(inodedep); 7641 /* 7642 * Re-acquire inodedep as we've dropped the 7643 * per-filesystem lock in clear_unlinked_inodedep(). 7644 */ 7645 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7646 } 7647 } 7648 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7649 FREE_LOCK(ump); 7650 handle_workitem_freefile(freefile); 7651 return; 7652 } 7653 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7654 inodedep->id_state |= GOINGAWAY; 7655 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7656 FREE_LOCK(ump); 7657 if (ip->i_number == ino) 7658 ip->i_flag |= IN_MODIFIED; 7659 } 7660 7661 /* 7662 * Check to see if an inode has never been written to disk. If 7663 * so free the inodedep and return success, otherwise return failure. 7664 * 7665 * If we still have a bitmap dependency, then the inode has never 7666 * been written to disk. Drop the dependency as it is no longer 7667 * necessary since the inode is being deallocated. We set the 7668 * ALLCOMPLETE flags since the bitmap now properly shows that the 7669 * inode is not allocated. Even if the inode is actively being 7670 * written, it has been rolled back to its zero'ed state, so we 7671 * are ensured that a zero inode is what is on the disk. For short 7672 * lived files, this change will usually result in removing all the 7673 * dependencies from the inode so that it can be freed immediately. 7674 */ 7675 static int 7676 check_inode_unwritten(inodedep) 7677 struct inodedep *inodedep; 7678 { 7679 7680 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7681 7682 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 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 return (0); 7696 /* 7697 * Another process might be in initiate_write_inodeblock_ufs[12] 7698 * trying to allocate memory without holding "Softdep Lock". 7699 */ 7700 if ((inodedep->id_state & IOSTARTED) != 0 && 7701 inodedep->id_savedino1 == NULL) 7702 return (0); 7703 7704 if (inodedep->id_state & ONDEPLIST) 7705 LIST_REMOVE(inodedep, id_deps); 7706 inodedep->id_state &= ~ONDEPLIST; 7707 inodedep->id_state |= ALLCOMPLETE; 7708 inodedep->id_bmsafemap = NULL; 7709 if (inodedep->id_state & ONWORKLIST) 7710 WORKLIST_REMOVE(&inodedep->id_list); 7711 if (inodedep->id_savedino1 != NULL) { 7712 free(inodedep->id_savedino1, M_SAVEDINO); 7713 inodedep->id_savedino1 = NULL; 7714 } 7715 if (free_inodedep(inodedep) == 0) 7716 panic("check_inode_unwritten: busy inode"); 7717 return (1); 7718 } 7719 7720 static int 7721 check_inodedep_free(inodedep) 7722 struct inodedep *inodedep; 7723 { 7724 7725 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7726 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7727 !LIST_EMPTY(&inodedep->id_dirremhd) || 7728 !LIST_EMPTY(&inodedep->id_pendinghd) || 7729 !LIST_EMPTY(&inodedep->id_bufwait) || 7730 !LIST_EMPTY(&inodedep->id_inowait) || 7731 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7732 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7733 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7734 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7735 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7736 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7737 inodedep->id_mkdiradd != NULL || 7738 inodedep->id_nlinkdelta != 0 || 7739 inodedep->id_savedino1 != NULL) 7740 return (0); 7741 return (1); 7742 } 7743 7744 /* 7745 * Try to free an inodedep structure. Return 1 if it could be freed. 7746 */ 7747 static int 7748 free_inodedep(inodedep) 7749 struct inodedep *inodedep; 7750 { 7751 7752 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7753 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7754 !check_inodedep_free(inodedep)) 7755 return (0); 7756 if (inodedep->id_state & ONDEPLIST) 7757 LIST_REMOVE(inodedep, id_deps); 7758 LIST_REMOVE(inodedep, id_hash); 7759 WORKITEM_FREE(inodedep, D_INODEDEP); 7760 return (1); 7761 } 7762 7763 /* 7764 * Free the block referenced by a freework structure. The parent freeblks 7765 * structure is released and completed when the final cg bitmap reaches 7766 * the disk. This routine may be freeing a jnewblk which never made it to 7767 * disk in which case we do not have to wait as the operation is undone 7768 * in memory immediately. 7769 */ 7770 static void 7771 freework_freeblock(freework, key) 7772 struct freework *freework; 7773 u_long key; 7774 { 7775 struct freeblks *freeblks; 7776 struct jnewblk *jnewblk; 7777 struct ufsmount *ump; 7778 struct workhead wkhd; 7779 struct fs *fs; 7780 int bsize; 7781 int needj; 7782 7783 ump = VFSTOUFS(freework->fw_list.wk_mp); 7784 LOCK_OWNED(ump); 7785 /* 7786 * Handle partial truncate separately. 7787 */ 7788 if (freework->fw_indir) { 7789 complete_trunc_indir(freework); 7790 return; 7791 } 7792 freeblks = freework->fw_freeblks; 7793 fs = ump->um_fs; 7794 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7795 bsize = lfragtosize(fs, freework->fw_frags); 7796 LIST_INIT(&wkhd); 7797 /* 7798 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7799 * on the indirblk hashtable and prevents premature freeing. 7800 */ 7801 freework->fw_state |= DEPCOMPLETE; 7802 /* 7803 * SUJ needs to wait for the segment referencing freed indirect 7804 * blocks to expire so that we know the checker will not confuse 7805 * a re-allocated indirect block with its old contents. 7806 */ 7807 if (needj && freework->fw_lbn <= -UFS_NDADDR) 7808 indirblk_insert(freework); 7809 /* 7810 * If we are canceling an existing jnewblk pass it to the free 7811 * routine, otherwise pass the freeblk which will ultimately 7812 * release the freeblks. If we're not journaling, we can just 7813 * free the freeblks immediately. 7814 */ 7815 jnewblk = freework->fw_jnewblk; 7816 if (jnewblk != NULL) { 7817 cancel_jnewblk(jnewblk, &wkhd); 7818 needj = 0; 7819 } else if (needj) { 7820 freework->fw_state |= DELAYEDFREE; 7821 freeblks->fb_cgwait++; 7822 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7823 } 7824 FREE_LOCK(ump); 7825 freeblks_free(ump, freeblks, btodb(bsize)); 7826 CTR4(KTR_SUJ, 7827 "freework_freeblock: ino %jd blkno %jd lbn %jd size %d", 7828 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7829 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7830 freeblks->fb_inum, freeblks->fb_vtype, &wkhd, key); 7831 ACQUIRE_LOCK(ump); 7832 /* 7833 * The jnewblk will be discarded and the bits in the map never 7834 * made it to disk. We can immediately free the freeblk. 7835 */ 7836 if (needj == 0) 7837 handle_written_freework(freework); 7838 } 7839 7840 /* 7841 * We enqueue freework items that need processing back on the freeblks and 7842 * add the freeblks to the worklist. This makes it easier to find all work 7843 * required to flush a truncation in process_truncates(). 7844 */ 7845 static void 7846 freework_enqueue(freework) 7847 struct freework *freework; 7848 { 7849 struct freeblks *freeblks; 7850 7851 freeblks = freework->fw_freeblks; 7852 if ((freework->fw_state & INPROGRESS) == 0) 7853 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7854 if ((freeblks->fb_state & 7855 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7856 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7857 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7858 } 7859 7860 /* 7861 * Start, continue, or finish the process of freeing an indirect block tree. 7862 * The free operation may be paused at any point with fw_off containing the 7863 * offset to restart from. This enables us to implement some flow control 7864 * for large truncates which may fan out and generate a huge number of 7865 * dependencies. 7866 */ 7867 static void 7868 handle_workitem_indirblk(freework) 7869 struct freework *freework; 7870 { 7871 struct freeblks *freeblks; 7872 struct ufsmount *ump; 7873 struct fs *fs; 7874 7875 freeblks = freework->fw_freeblks; 7876 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7877 fs = ump->um_fs; 7878 if (freework->fw_state & DEPCOMPLETE) { 7879 handle_written_freework(freework); 7880 return; 7881 } 7882 if (freework->fw_off == NINDIR(fs)) { 7883 freework_freeblock(freework, SINGLETON_KEY); 7884 return; 7885 } 7886 freework->fw_state |= INPROGRESS; 7887 FREE_LOCK(ump); 7888 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7889 freework->fw_lbn); 7890 ACQUIRE_LOCK(ump); 7891 } 7892 7893 /* 7894 * Called when a freework structure attached to a cg buf is written. The 7895 * ref on either the parent or the freeblks structure is released and 7896 * the freeblks is added back to the worklist if there is more work to do. 7897 */ 7898 static void 7899 handle_written_freework(freework) 7900 struct freework *freework; 7901 { 7902 struct freeblks *freeblks; 7903 struct freework *parent; 7904 7905 freeblks = freework->fw_freeblks; 7906 parent = freework->fw_parent; 7907 if (freework->fw_state & DELAYEDFREE) 7908 freeblks->fb_cgwait--; 7909 freework->fw_state |= COMPLETE; 7910 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7911 WORKITEM_FREE(freework, D_FREEWORK); 7912 if (parent) { 7913 if (--parent->fw_ref == 0) 7914 freework_enqueue(parent); 7915 return; 7916 } 7917 if (--freeblks->fb_ref != 0) 7918 return; 7919 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7920 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7921 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7922 } 7923 7924 /* 7925 * This workitem routine performs the block de-allocation. 7926 * The workitem is added to the pending list after the updated 7927 * inode block has been written to disk. As mentioned above, 7928 * checks regarding the number of blocks de-allocated (compared 7929 * to the number of blocks allocated for the file) are also 7930 * performed in this function. 7931 */ 7932 static int 7933 handle_workitem_freeblocks(freeblks, flags) 7934 struct freeblks *freeblks; 7935 int flags; 7936 { 7937 struct freework *freework; 7938 struct newblk *newblk; 7939 struct allocindir *aip; 7940 struct ufsmount *ump; 7941 struct worklist *wk; 7942 u_long key; 7943 7944 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7945 ("handle_workitem_freeblocks: Journal entries not written.")); 7946 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7947 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 7948 ACQUIRE_LOCK(ump); 7949 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7950 WORKLIST_REMOVE(wk); 7951 switch (wk->wk_type) { 7952 case D_DIRREM: 7953 wk->wk_state |= COMPLETE; 7954 add_to_worklist(wk, 0); 7955 continue; 7956 7957 case D_ALLOCDIRECT: 7958 free_newblk(WK_NEWBLK(wk)); 7959 continue; 7960 7961 case D_ALLOCINDIR: 7962 aip = WK_ALLOCINDIR(wk); 7963 freework = NULL; 7964 if (aip->ai_state & DELAYEDFREE) { 7965 FREE_LOCK(ump); 7966 freework = newfreework(ump, freeblks, NULL, 7967 aip->ai_lbn, aip->ai_newblkno, 7968 ump->um_fs->fs_frag, 0, 0); 7969 ACQUIRE_LOCK(ump); 7970 } 7971 newblk = WK_NEWBLK(wk); 7972 if (newblk->nb_jnewblk) { 7973 freework->fw_jnewblk = newblk->nb_jnewblk; 7974 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7975 newblk->nb_jnewblk = NULL; 7976 } 7977 free_newblk(newblk); 7978 continue; 7979 7980 case D_FREEWORK: 7981 freework = WK_FREEWORK(wk); 7982 if (freework->fw_lbn <= -UFS_NDADDR) 7983 handle_workitem_indirblk(freework); 7984 else 7985 freework_freeblock(freework, key); 7986 continue; 7987 default: 7988 panic("handle_workitem_freeblocks: Unknown type %s", 7989 TYPENAME(wk->wk_type)); 7990 } 7991 } 7992 if (freeblks->fb_ref != 0) { 7993 freeblks->fb_state &= ~INPROGRESS; 7994 wake_worklist(&freeblks->fb_list); 7995 freeblks = NULL; 7996 } 7997 FREE_LOCK(ump); 7998 ffs_blkrelease_finish(ump, key); 7999 if (freeblks) 8000 return handle_complete_freeblocks(freeblks, flags); 8001 return (0); 8002 } 8003 8004 /* 8005 * Handle completion of block free via truncate. This allows fs_pending 8006 * to track the actual free block count more closely than if we only updated 8007 * it at the end. We must be careful to handle cases where the block count 8008 * on free was incorrect. 8009 */ 8010 static void 8011 freeblks_free(ump, freeblks, blocks) 8012 struct ufsmount *ump; 8013 struct freeblks *freeblks; 8014 int blocks; 8015 { 8016 struct fs *fs; 8017 ufs2_daddr_t remain; 8018 8019 UFS_LOCK(ump); 8020 remain = -freeblks->fb_chkcnt; 8021 freeblks->fb_chkcnt += blocks; 8022 if (remain > 0) { 8023 if (remain < blocks) 8024 blocks = remain; 8025 fs = ump->um_fs; 8026 fs->fs_pendingblocks -= blocks; 8027 } 8028 UFS_UNLOCK(ump); 8029 } 8030 8031 /* 8032 * Once all of the freework workitems are complete we can retire the 8033 * freeblocks dependency and any journal work awaiting completion. This 8034 * can not be called until all other dependencies are stable on disk. 8035 */ 8036 static int 8037 handle_complete_freeblocks(freeblks, flags) 8038 struct freeblks *freeblks; 8039 int flags; 8040 { 8041 struct inodedep *inodedep; 8042 struct inode *ip; 8043 struct vnode *vp; 8044 struct fs *fs; 8045 struct ufsmount *ump; 8046 ufs2_daddr_t spare; 8047 8048 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8049 fs = ump->um_fs; 8050 flags = LK_EXCLUSIVE | flags; 8051 spare = freeblks->fb_chkcnt; 8052 8053 /* 8054 * If we did not release the expected number of blocks we may have 8055 * to adjust the inode block count here. Only do so if it wasn't 8056 * a truncation to zero and the modrev still matches. 8057 */ 8058 if (spare && freeblks->fb_len != 0) { 8059 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8060 flags, &vp, FFSV_FORCEINSMQ) != 0) 8061 return (EBUSY); 8062 ip = VTOI(vp); 8063 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 8064 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 8065 ip->i_flag |= IN_CHANGE; 8066 /* 8067 * We must wait so this happens before the 8068 * journal is reclaimed. 8069 */ 8070 ffs_update(vp, 1); 8071 } 8072 vput(vp); 8073 } 8074 if (spare < 0) { 8075 UFS_LOCK(ump); 8076 fs->fs_pendingblocks += spare; 8077 UFS_UNLOCK(ump); 8078 } 8079 #ifdef QUOTA 8080 /* Handle spare. */ 8081 if (spare) 8082 quotaadj(freeblks->fb_quota, ump, -spare); 8083 quotarele(freeblks->fb_quota); 8084 #endif 8085 ACQUIRE_LOCK(ump); 8086 if (freeblks->fb_state & ONDEPLIST) { 8087 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8088 0, &inodedep); 8089 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8090 freeblks->fb_state &= ~ONDEPLIST; 8091 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8092 free_inodedep(inodedep); 8093 } 8094 /* 8095 * All of the freeblock deps must be complete prior to this call 8096 * so it's now safe to complete earlier outstanding journal entries. 8097 */ 8098 handle_jwork(&freeblks->fb_jwork); 8099 WORKITEM_FREE(freeblks, D_FREEBLKS); 8100 FREE_LOCK(ump); 8101 return (0); 8102 } 8103 8104 /* 8105 * Release blocks associated with the freeblks and stored in the indirect 8106 * block dbn. If level is greater than SINGLE, the block is an indirect block 8107 * and recursive calls to indirtrunc must be used to cleanse other indirect 8108 * blocks. 8109 * 8110 * This handles partial and complete truncation of blocks. Partial is noted 8111 * with goingaway == 0. In this case the freework is completed after the 8112 * zero'd indirects are written to disk. For full truncation the freework 8113 * is completed after the block is freed. 8114 */ 8115 static void 8116 indir_trunc(freework, dbn, lbn) 8117 struct freework *freework; 8118 ufs2_daddr_t dbn; 8119 ufs_lbn_t lbn; 8120 { 8121 struct freework *nfreework; 8122 struct workhead wkhd; 8123 struct freeblks *freeblks; 8124 struct buf *bp; 8125 struct fs *fs; 8126 struct indirdep *indirdep; 8127 struct mount *mp; 8128 struct ufsmount *ump; 8129 ufs1_daddr_t *bap1; 8130 ufs2_daddr_t nb, nnb, *bap2; 8131 ufs_lbn_t lbnadd, nlbn; 8132 u_long key; 8133 int nblocks, ufs1fmt, freedblocks; 8134 int goingaway, freedeps, needj, level, cnt, i; 8135 8136 freeblks = freework->fw_freeblks; 8137 mp = freeblks->fb_list.wk_mp; 8138 ump = VFSTOUFS(mp); 8139 fs = ump->um_fs; 8140 /* 8141 * Get buffer of block pointers to be freed. There are three cases: 8142 * 8143 * 1) Partial truncate caches the indirdep pointer in the freework 8144 * which provides us a back copy to the save bp which holds the 8145 * pointers we want to clear. When this completes the zero 8146 * pointers are written to the real copy. 8147 * 2) The indirect is being completely truncated, cancel_indirdep() 8148 * eliminated the real copy and placed the indirdep on the saved 8149 * copy. The indirdep and buf are discarded when this completes. 8150 * 3) The indirect was not in memory, we read a copy off of the disk 8151 * using the devvp and drop and invalidate the buffer when we're 8152 * done. 8153 */ 8154 goingaway = 1; 8155 indirdep = NULL; 8156 if (freework->fw_indir != NULL) { 8157 goingaway = 0; 8158 indirdep = freework->fw_indir; 8159 bp = indirdep->ir_savebp; 8160 if (bp == NULL || bp->b_blkno != dbn) 8161 panic("indir_trunc: Bad saved buf %p blkno %jd", 8162 bp, (intmax_t)dbn); 8163 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8164 /* 8165 * The lock prevents the buf dep list from changing and 8166 * indirects on devvp should only ever have one dependency. 8167 */ 8168 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8169 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8170 panic("indir_trunc: Bad indirdep %p from buf %p", 8171 indirdep, bp); 8172 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8173 NOCRED, &bp) != 0) { 8174 brelse(bp); 8175 return; 8176 } 8177 ACQUIRE_LOCK(ump); 8178 /* Protects against a race with complete_trunc_indir(). */ 8179 freework->fw_state &= ~INPROGRESS; 8180 /* 8181 * If we have an indirdep we need to enforce the truncation order 8182 * and discard it when it is complete. 8183 */ 8184 if (indirdep) { 8185 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8186 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8187 /* 8188 * Add the complete truncate to the list on the 8189 * indirdep to enforce in-order processing. 8190 */ 8191 if (freework->fw_indir == NULL) 8192 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8193 freework, fw_next); 8194 FREE_LOCK(ump); 8195 return; 8196 } 8197 /* 8198 * If we're goingaway, free the indirdep. Otherwise it will 8199 * linger until the write completes. 8200 */ 8201 if (goingaway) 8202 free_indirdep(indirdep); 8203 } 8204 FREE_LOCK(ump); 8205 /* Initialize pointers depending on block size. */ 8206 if (ump->um_fstype == UFS1) { 8207 bap1 = (ufs1_daddr_t *)bp->b_data; 8208 nb = bap1[freework->fw_off]; 8209 ufs1fmt = 1; 8210 bap2 = NULL; 8211 } else { 8212 bap2 = (ufs2_daddr_t *)bp->b_data; 8213 nb = bap2[freework->fw_off]; 8214 ufs1fmt = 0; 8215 bap1 = NULL; 8216 } 8217 level = lbn_level(lbn); 8218 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8219 lbnadd = lbn_offset(fs, level); 8220 nblocks = btodb(fs->fs_bsize); 8221 nfreework = freework; 8222 freedeps = 0; 8223 cnt = 0; 8224 /* 8225 * Reclaim blocks. Traverses into nested indirect levels and 8226 * arranges for the current level to be freed when subordinates 8227 * are free when journaling. 8228 */ 8229 key = ffs_blkrelease_start(ump, freeblks->fb_devvp, freeblks->fb_inum); 8230 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8231 if (UFS_CHECK_BLKNO(mp, freeblks->fb_inum, nb, 8232 fs->fs_bsize) != 0) 8233 nb = 0; 8234 if (i != NINDIR(fs) - 1) { 8235 if (ufs1fmt) 8236 nnb = bap1[i+1]; 8237 else 8238 nnb = bap2[i+1]; 8239 } else 8240 nnb = 0; 8241 if (nb == 0) 8242 continue; 8243 cnt++; 8244 if (level != 0) { 8245 nlbn = (lbn + 1) - (i * lbnadd); 8246 if (needj != 0) { 8247 nfreework = newfreework(ump, freeblks, freework, 8248 nlbn, nb, fs->fs_frag, 0, 0); 8249 freedeps++; 8250 } 8251 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8252 } else { 8253 struct freedep *freedep; 8254 8255 /* 8256 * Attempt to aggregate freedep dependencies for 8257 * all blocks being released to the same CG. 8258 */ 8259 LIST_INIT(&wkhd); 8260 if (needj != 0 && 8261 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8262 freedep = newfreedep(freework); 8263 WORKLIST_INSERT_UNLOCKED(&wkhd, 8264 &freedep->fd_list); 8265 freedeps++; 8266 } 8267 CTR3(KTR_SUJ, 8268 "indir_trunc: ino %jd blkno %jd size %d", 8269 freeblks->fb_inum, nb, fs->fs_bsize); 8270 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8271 fs->fs_bsize, freeblks->fb_inum, 8272 freeblks->fb_vtype, &wkhd, key); 8273 } 8274 } 8275 ffs_blkrelease_finish(ump, key); 8276 if (goingaway) { 8277 bp->b_flags |= B_INVAL | B_NOCACHE; 8278 brelse(bp); 8279 } 8280 freedblocks = 0; 8281 if (level == 0) 8282 freedblocks = (nblocks * cnt); 8283 if (needj == 0) 8284 freedblocks += nblocks; 8285 freeblks_free(ump, freeblks, freedblocks); 8286 /* 8287 * If we are journaling set up the ref counts and offset so this 8288 * indirect can be completed when its children are free. 8289 */ 8290 if (needj) { 8291 ACQUIRE_LOCK(ump); 8292 freework->fw_off = i; 8293 freework->fw_ref += freedeps; 8294 freework->fw_ref -= NINDIR(fs) + 1; 8295 if (level == 0) 8296 freeblks->fb_cgwait += freedeps; 8297 if (freework->fw_ref == 0) 8298 freework_freeblock(freework, SINGLETON_KEY); 8299 FREE_LOCK(ump); 8300 return; 8301 } 8302 /* 8303 * If we're not journaling we can free the indirect now. 8304 */ 8305 dbn = dbtofsb(fs, dbn); 8306 CTR3(KTR_SUJ, 8307 "indir_trunc 2: ino %jd blkno %jd size %d", 8308 freeblks->fb_inum, dbn, fs->fs_bsize); 8309 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8310 freeblks->fb_inum, freeblks->fb_vtype, NULL, SINGLETON_KEY); 8311 /* Non SUJ softdep does single-threaded truncations. */ 8312 if (freework->fw_blkno == dbn) { 8313 freework->fw_state |= ALLCOMPLETE; 8314 ACQUIRE_LOCK(ump); 8315 handle_written_freework(freework); 8316 FREE_LOCK(ump); 8317 } 8318 return; 8319 } 8320 8321 /* 8322 * Cancel an allocindir when it is removed via truncation. When bp is not 8323 * NULL the indirect never appeared on disk and is scheduled to be freed 8324 * independently of the indir so we can more easily track journal work. 8325 */ 8326 static void 8327 cancel_allocindir(aip, bp, freeblks, trunc) 8328 struct allocindir *aip; 8329 struct buf *bp; 8330 struct freeblks *freeblks; 8331 int trunc; 8332 { 8333 struct indirdep *indirdep; 8334 struct freefrag *freefrag; 8335 struct newblk *newblk; 8336 8337 newblk = (struct newblk *)aip; 8338 LIST_REMOVE(aip, ai_next); 8339 /* 8340 * We must eliminate the pointer in bp if it must be freed on its 8341 * own due to partial truncate or pending journal work. 8342 */ 8343 if (bp && (trunc || newblk->nb_jnewblk)) { 8344 /* 8345 * Clear the pointer and mark the aip to be freed 8346 * directly if it never existed on disk. 8347 */ 8348 aip->ai_state |= DELAYEDFREE; 8349 indirdep = aip->ai_indirdep; 8350 if (indirdep->ir_state & UFS1FMT) 8351 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8352 else 8353 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8354 } 8355 /* 8356 * When truncating the previous pointer will be freed via 8357 * savedbp. Eliminate the freefrag which would dup free. 8358 */ 8359 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8360 newblk->nb_freefrag = NULL; 8361 if (freefrag->ff_jdep) 8362 cancel_jfreefrag( 8363 WK_JFREEFRAG(freefrag->ff_jdep)); 8364 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8365 WORKITEM_FREE(freefrag, D_FREEFRAG); 8366 } 8367 /* 8368 * If the journal hasn't been written the jnewblk must be passed 8369 * to the call to ffs_blkfree that reclaims the space. We accomplish 8370 * this by leaving the journal dependency on the newblk to be freed 8371 * when a freework is created in handle_workitem_freeblocks(). 8372 */ 8373 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8374 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8375 } 8376 8377 /* 8378 * Create the mkdir dependencies for . and .. in a new directory. Link them 8379 * in to a newdirblk so any subsequent additions are tracked properly. The 8380 * caller is responsible for adding the mkdir1 dependency to the journal 8381 * and updating id_mkdiradd. This function returns with the per-filesystem 8382 * lock held. 8383 */ 8384 static struct mkdir * 8385 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8386 struct diradd *dap; 8387 ino_t newinum; 8388 ino_t dinum; 8389 struct buf *newdirbp; 8390 struct mkdir **mkdirp; 8391 { 8392 struct newblk *newblk; 8393 struct pagedep *pagedep; 8394 struct inodedep *inodedep; 8395 struct newdirblk *newdirblk; 8396 struct mkdir *mkdir1, *mkdir2; 8397 struct worklist *wk; 8398 struct jaddref *jaddref; 8399 struct ufsmount *ump; 8400 struct mount *mp; 8401 8402 mp = dap->da_list.wk_mp; 8403 ump = VFSTOUFS(mp); 8404 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8405 M_SOFTDEP_FLAGS); 8406 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8407 LIST_INIT(&newdirblk->db_mkdir); 8408 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8409 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8410 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8411 mkdir1->md_diradd = dap; 8412 mkdir1->md_jaddref = NULL; 8413 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8414 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8415 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8416 mkdir2->md_diradd = dap; 8417 mkdir2->md_jaddref = NULL; 8418 if (MOUNTEDSUJ(mp) == 0) { 8419 mkdir1->md_state |= DEPCOMPLETE; 8420 mkdir2->md_state |= DEPCOMPLETE; 8421 } 8422 /* 8423 * Dependency on "." and ".." being written to disk. 8424 */ 8425 mkdir1->md_buf = newdirbp; 8426 ACQUIRE_LOCK(VFSTOUFS(mp)); 8427 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8428 /* 8429 * We must link the pagedep, allocdirect, and newdirblk for 8430 * the initial file page so the pointer to the new directory 8431 * is not written until the directory contents are live and 8432 * any subsequent additions are not marked live until the 8433 * block is reachable via the inode. 8434 */ 8435 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8436 panic("setup_newdir: lost pagedep"); 8437 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8438 if (wk->wk_type == D_ALLOCDIRECT) 8439 break; 8440 if (wk == NULL) 8441 panic("setup_newdir: lost allocdirect"); 8442 if (pagedep->pd_state & NEWBLOCK) 8443 panic("setup_newdir: NEWBLOCK already set"); 8444 newblk = WK_NEWBLK(wk); 8445 pagedep->pd_state |= NEWBLOCK; 8446 pagedep->pd_newdirblk = newdirblk; 8447 newdirblk->db_pagedep = pagedep; 8448 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8449 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8450 /* 8451 * Look up the inodedep for the parent directory so that we 8452 * can link mkdir2 into the pending dotdot jaddref or 8453 * the inode write if there is none. If the inode is 8454 * ALLCOMPLETE and no jaddref is present all dependencies have 8455 * been satisfied and mkdir2 can be freed. 8456 */ 8457 inodedep_lookup(mp, dinum, 0, &inodedep); 8458 if (MOUNTEDSUJ(mp)) { 8459 if (inodedep == NULL) 8460 panic("setup_newdir: Lost parent."); 8461 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8462 inoreflst); 8463 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8464 (jaddref->ja_state & MKDIR_PARENT), 8465 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8466 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8467 mkdir2->md_jaddref = jaddref; 8468 jaddref->ja_mkdir = mkdir2; 8469 } else if (inodedep == NULL || 8470 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8471 dap->da_state &= ~MKDIR_PARENT; 8472 WORKITEM_FREE(mkdir2, D_MKDIR); 8473 mkdir2 = NULL; 8474 } else { 8475 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8476 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8477 } 8478 *mkdirp = mkdir2; 8479 8480 return (mkdir1); 8481 } 8482 8483 /* 8484 * Directory entry addition dependencies. 8485 * 8486 * When adding a new directory entry, the inode (with its incremented link 8487 * count) must be written to disk before the directory entry's pointer to it. 8488 * Also, if the inode is newly allocated, the corresponding freemap must be 8489 * updated (on disk) before the directory entry's pointer. These requirements 8490 * are met via undo/redo on the directory entry's pointer, which consists 8491 * simply of the inode number. 8492 * 8493 * As directory entries are added and deleted, the free space within a 8494 * directory block can become fragmented. The ufs filesystem will compact 8495 * a fragmented directory block to make space for a new entry. When this 8496 * occurs, the offsets of previously added entries change. Any "diradd" 8497 * dependency structures corresponding to these entries must be updated with 8498 * the new offsets. 8499 */ 8500 8501 /* 8502 * This routine is called after the in-memory inode's link 8503 * count has been incremented, but before the directory entry's 8504 * pointer to the inode has been set. 8505 */ 8506 int 8507 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8508 struct buf *bp; /* buffer containing directory block */ 8509 struct inode *dp; /* inode for directory */ 8510 off_t diroffset; /* offset of new entry in directory */ 8511 ino_t newinum; /* inode referenced by new directory entry */ 8512 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8513 int isnewblk; /* entry is in a newly allocated block */ 8514 { 8515 int offset; /* offset of new entry within directory block */ 8516 ufs_lbn_t lbn; /* block in directory containing new entry */ 8517 struct fs *fs; 8518 struct diradd *dap; 8519 struct newblk *newblk; 8520 struct pagedep *pagedep; 8521 struct inodedep *inodedep; 8522 struct newdirblk *newdirblk; 8523 struct mkdir *mkdir1, *mkdir2; 8524 struct jaddref *jaddref; 8525 struct ufsmount *ump; 8526 struct mount *mp; 8527 int isindir; 8528 8529 mp = ITOVFS(dp); 8530 ump = VFSTOUFS(mp); 8531 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8532 ("softdep_setup_directory_add called on non-softdep filesystem")); 8533 /* 8534 * Whiteouts have no dependencies. 8535 */ 8536 if (newinum == UFS_WINO) { 8537 if (newdirbp != NULL) 8538 bdwrite(newdirbp); 8539 return (0); 8540 } 8541 jaddref = NULL; 8542 mkdir1 = mkdir2 = NULL; 8543 fs = ump->um_fs; 8544 lbn = lblkno(fs, diroffset); 8545 offset = blkoff(fs, diroffset); 8546 dap = malloc(sizeof(struct diradd), M_DIRADD, 8547 M_SOFTDEP_FLAGS|M_ZERO); 8548 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8549 dap->da_offset = offset; 8550 dap->da_newinum = newinum; 8551 dap->da_state = ATTACHED; 8552 LIST_INIT(&dap->da_jwork); 8553 isindir = bp->b_lblkno >= UFS_NDADDR; 8554 newdirblk = NULL; 8555 if (isnewblk && 8556 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8557 newdirblk = malloc(sizeof(struct newdirblk), 8558 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8559 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8560 LIST_INIT(&newdirblk->db_mkdir); 8561 } 8562 /* 8563 * If we're creating a new directory setup the dependencies and set 8564 * the dap state to wait for them. Otherwise it's COMPLETE and 8565 * we can move on. 8566 */ 8567 if (newdirbp == NULL) { 8568 dap->da_state |= DEPCOMPLETE; 8569 ACQUIRE_LOCK(ump); 8570 } else { 8571 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8572 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8573 &mkdir2); 8574 } 8575 /* 8576 * Link into parent directory pagedep to await its being written. 8577 */ 8578 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8579 #ifdef INVARIANTS 8580 if (diradd_lookup(pagedep, offset) != NULL) 8581 panic("softdep_setup_directory_add: %p already at off %d\n", 8582 diradd_lookup(pagedep, offset), offset); 8583 #endif 8584 dap->da_pagedep = pagedep; 8585 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8586 da_pdlist); 8587 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8588 /* 8589 * If we're journaling, link the diradd into the jaddref so it 8590 * may be completed after the journal entry is written. Otherwise, 8591 * link the diradd into its inodedep. If the inode is not yet 8592 * written place it on the bufwait list, otherwise do the post-inode 8593 * write processing to put it on the id_pendinghd list. 8594 */ 8595 if (MOUNTEDSUJ(mp)) { 8596 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8597 inoreflst); 8598 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8599 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8600 jaddref->ja_diroff = diroffset; 8601 jaddref->ja_diradd = dap; 8602 add_to_journal(&jaddref->ja_list); 8603 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8604 diradd_inode_written(dap, inodedep); 8605 else 8606 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8607 /* 8608 * Add the journal entries for . and .. links now that the primary 8609 * link is written. 8610 */ 8611 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8612 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8613 inoreflst, if_deps); 8614 KASSERT(jaddref != NULL && 8615 jaddref->ja_ino == jaddref->ja_parent && 8616 (jaddref->ja_state & MKDIR_BODY), 8617 ("softdep_setup_directory_add: bad dot jaddref %p", 8618 jaddref)); 8619 mkdir1->md_jaddref = jaddref; 8620 jaddref->ja_mkdir = mkdir1; 8621 /* 8622 * It is important that the dotdot journal entry 8623 * is added prior to the dot entry since dot writes 8624 * both the dot and dotdot links. These both must 8625 * be added after the primary link for the journal 8626 * to remain consistent. 8627 */ 8628 add_to_journal(&mkdir2->md_jaddref->ja_list); 8629 add_to_journal(&jaddref->ja_list); 8630 } 8631 /* 8632 * If we are adding a new directory remember this diradd so that if 8633 * we rename it we can keep the dot and dotdot dependencies. If 8634 * we are adding a new name for an inode that has a mkdiradd we 8635 * must be in rename and we have to move the dot and dotdot 8636 * dependencies to this new name. The old name is being orphaned 8637 * soon. 8638 */ 8639 if (mkdir1 != NULL) { 8640 if (inodedep->id_mkdiradd != NULL) 8641 panic("softdep_setup_directory_add: Existing mkdir"); 8642 inodedep->id_mkdiradd = dap; 8643 } else if (inodedep->id_mkdiradd) 8644 merge_diradd(inodedep, dap); 8645 if (newdirblk != NULL) { 8646 /* 8647 * There is nothing to do if we are already tracking 8648 * this block. 8649 */ 8650 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8651 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8652 FREE_LOCK(ump); 8653 return (0); 8654 } 8655 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8656 == 0) 8657 panic("softdep_setup_directory_add: lost entry"); 8658 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8659 pagedep->pd_state |= NEWBLOCK; 8660 pagedep->pd_newdirblk = newdirblk; 8661 newdirblk->db_pagedep = pagedep; 8662 FREE_LOCK(ump); 8663 /* 8664 * If we extended into an indirect signal direnter to sync. 8665 */ 8666 if (isindir) 8667 return (1); 8668 return (0); 8669 } 8670 FREE_LOCK(ump); 8671 return (0); 8672 } 8673 8674 /* 8675 * This procedure is called to change the offset of a directory 8676 * entry when compacting a directory block which must be owned 8677 * exclusively by the caller. Note that the actual entry movement 8678 * must be done in this procedure to ensure that no I/O completions 8679 * occur while the move is in progress. 8680 */ 8681 void 8682 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8683 struct buf *bp; /* Buffer holding directory block. */ 8684 struct inode *dp; /* inode for directory */ 8685 caddr_t base; /* address of dp->i_offset */ 8686 caddr_t oldloc; /* address of old directory location */ 8687 caddr_t newloc; /* address of new directory location */ 8688 int entrysize; /* size of directory entry */ 8689 { 8690 int offset, oldoffset, newoffset; 8691 struct pagedep *pagedep; 8692 struct jmvref *jmvref; 8693 struct diradd *dap; 8694 struct direct *de; 8695 struct mount *mp; 8696 struct ufsmount *ump; 8697 ufs_lbn_t lbn; 8698 int flags; 8699 8700 mp = ITOVFS(dp); 8701 ump = VFSTOUFS(mp); 8702 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8703 ("softdep_change_directoryentry_offset called on " 8704 "non-softdep filesystem")); 8705 de = (struct direct *)oldloc; 8706 jmvref = NULL; 8707 flags = 0; 8708 /* 8709 * Moves are always journaled as it would be too complex to 8710 * determine if any affected adds or removes are present in the 8711 * journal. 8712 */ 8713 if (MOUNTEDSUJ(mp)) { 8714 flags = DEPALLOC; 8715 jmvref = newjmvref(dp, de->d_ino, 8716 dp->i_offset + (oldloc - base), 8717 dp->i_offset + (newloc - base)); 8718 } 8719 lbn = lblkno(ump->um_fs, dp->i_offset); 8720 offset = blkoff(ump->um_fs, dp->i_offset); 8721 oldoffset = offset + (oldloc - base); 8722 newoffset = offset + (newloc - base); 8723 ACQUIRE_LOCK(ump); 8724 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8725 goto done; 8726 dap = diradd_lookup(pagedep, oldoffset); 8727 if (dap) { 8728 dap->da_offset = newoffset; 8729 newoffset = DIRADDHASH(newoffset); 8730 oldoffset = DIRADDHASH(oldoffset); 8731 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8732 newoffset != oldoffset) { 8733 LIST_REMOVE(dap, da_pdlist); 8734 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8735 dap, da_pdlist); 8736 } 8737 } 8738 done: 8739 if (jmvref) { 8740 jmvref->jm_pagedep = pagedep; 8741 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8742 add_to_journal(&jmvref->jm_list); 8743 } 8744 bcopy(oldloc, newloc, entrysize); 8745 FREE_LOCK(ump); 8746 } 8747 8748 /* 8749 * Move the mkdir dependencies and journal work from one diradd to another 8750 * when renaming a directory. The new name must depend on the mkdir deps 8751 * completing as the old name did. Directories can only have one valid link 8752 * at a time so one must be canonical. 8753 */ 8754 static void 8755 merge_diradd(inodedep, newdap) 8756 struct inodedep *inodedep; 8757 struct diradd *newdap; 8758 { 8759 struct diradd *olddap; 8760 struct mkdir *mkdir, *nextmd; 8761 struct ufsmount *ump; 8762 short state; 8763 8764 olddap = inodedep->id_mkdiradd; 8765 inodedep->id_mkdiradd = newdap; 8766 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8767 newdap->da_state &= ~DEPCOMPLETE; 8768 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8769 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8770 mkdir = nextmd) { 8771 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8772 if (mkdir->md_diradd != olddap) 8773 continue; 8774 mkdir->md_diradd = newdap; 8775 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8776 newdap->da_state |= state; 8777 olddap->da_state &= ~state; 8778 if ((olddap->da_state & 8779 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8780 break; 8781 } 8782 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8783 panic("merge_diradd: unfound ref"); 8784 } 8785 /* 8786 * Any mkdir related journal items are not safe to be freed until 8787 * the new name is stable. 8788 */ 8789 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8790 olddap->da_state |= DEPCOMPLETE; 8791 complete_diradd(olddap); 8792 } 8793 8794 /* 8795 * Move the diradd to the pending list when all diradd dependencies are 8796 * complete. 8797 */ 8798 static void 8799 complete_diradd(dap) 8800 struct diradd *dap; 8801 { 8802 struct pagedep *pagedep; 8803 8804 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8805 if (dap->da_state & DIRCHG) 8806 pagedep = dap->da_previous->dm_pagedep; 8807 else 8808 pagedep = dap->da_pagedep; 8809 LIST_REMOVE(dap, da_pdlist); 8810 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8811 } 8812 } 8813 8814 /* 8815 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8816 * add entries and conditonally journal the remove. 8817 */ 8818 static void 8819 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8820 struct diradd *dap; 8821 struct dirrem *dirrem; 8822 struct jremref *jremref; 8823 struct jremref *dotremref; 8824 struct jremref *dotdotremref; 8825 { 8826 struct inodedep *inodedep; 8827 struct jaddref *jaddref; 8828 struct inoref *inoref; 8829 struct ufsmount *ump; 8830 struct mkdir *mkdir; 8831 8832 /* 8833 * If no remove references were allocated we're on a non-journaled 8834 * filesystem and can skip the cancel step. 8835 */ 8836 if (jremref == NULL) { 8837 free_diradd(dap, NULL); 8838 return; 8839 } 8840 /* 8841 * Cancel the primary name an free it if it does not require 8842 * journaling. 8843 */ 8844 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8845 0, &inodedep) != 0) { 8846 /* Abort the addref that reference this diradd. */ 8847 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8848 if (inoref->if_list.wk_type != D_JADDREF) 8849 continue; 8850 jaddref = (struct jaddref *)inoref; 8851 if (jaddref->ja_diradd != dap) 8852 continue; 8853 if (cancel_jaddref(jaddref, inodedep, 8854 &dirrem->dm_jwork) == 0) { 8855 free_jremref(jremref); 8856 jremref = NULL; 8857 } 8858 break; 8859 } 8860 } 8861 /* 8862 * Cancel subordinate names and free them if they do not require 8863 * journaling. 8864 */ 8865 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8866 ump = VFSTOUFS(dap->da_list.wk_mp); 8867 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8868 if (mkdir->md_diradd != dap) 8869 continue; 8870 if ((jaddref = mkdir->md_jaddref) == NULL) 8871 continue; 8872 mkdir->md_jaddref = NULL; 8873 if (mkdir->md_state & MKDIR_PARENT) { 8874 if (cancel_jaddref(jaddref, NULL, 8875 &dirrem->dm_jwork) == 0) { 8876 free_jremref(dotdotremref); 8877 dotdotremref = NULL; 8878 } 8879 } else { 8880 if (cancel_jaddref(jaddref, inodedep, 8881 &dirrem->dm_jwork) == 0) { 8882 free_jremref(dotremref); 8883 dotremref = NULL; 8884 } 8885 } 8886 } 8887 } 8888 8889 if (jremref) 8890 journal_jremref(dirrem, jremref, inodedep); 8891 if (dotremref) 8892 journal_jremref(dirrem, dotremref, inodedep); 8893 if (dotdotremref) 8894 journal_jremref(dirrem, dotdotremref, NULL); 8895 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8896 free_diradd(dap, &dirrem->dm_jwork); 8897 } 8898 8899 /* 8900 * Free a diradd dependency structure. 8901 */ 8902 static void 8903 free_diradd(dap, wkhd) 8904 struct diradd *dap; 8905 struct workhead *wkhd; 8906 { 8907 struct dirrem *dirrem; 8908 struct pagedep *pagedep; 8909 struct inodedep *inodedep; 8910 struct mkdir *mkdir, *nextmd; 8911 struct ufsmount *ump; 8912 8913 ump = VFSTOUFS(dap->da_list.wk_mp); 8914 LOCK_OWNED(ump); 8915 LIST_REMOVE(dap, da_pdlist); 8916 if (dap->da_state & ONWORKLIST) 8917 WORKLIST_REMOVE(&dap->da_list); 8918 if ((dap->da_state & DIRCHG) == 0) { 8919 pagedep = dap->da_pagedep; 8920 } else { 8921 dirrem = dap->da_previous; 8922 pagedep = dirrem->dm_pagedep; 8923 dirrem->dm_dirinum = pagedep->pd_ino; 8924 dirrem->dm_state |= COMPLETE; 8925 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8926 add_to_worklist(&dirrem->dm_list, 0); 8927 } 8928 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8929 0, &inodedep) != 0) 8930 if (inodedep->id_mkdiradd == dap) 8931 inodedep->id_mkdiradd = NULL; 8932 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8933 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8934 mkdir = nextmd) { 8935 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8936 if (mkdir->md_diradd != dap) 8937 continue; 8938 dap->da_state &= 8939 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8940 LIST_REMOVE(mkdir, md_mkdirs); 8941 if (mkdir->md_state & ONWORKLIST) 8942 WORKLIST_REMOVE(&mkdir->md_list); 8943 if (mkdir->md_jaddref != NULL) 8944 panic("free_diradd: Unexpected jaddref"); 8945 WORKITEM_FREE(mkdir, D_MKDIR); 8946 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8947 break; 8948 } 8949 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8950 panic("free_diradd: unfound ref"); 8951 } 8952 if (inodedep) 8953 free_inodedep(inodedep); 8954 /* 8955 * Free any journal segments waiting for the directory write. 8956 */ 8957 handle_jwork(&dap->da_jwork); 8958 WORKITEM_FREE(dap, D_DIRADD); 8959 } 8960 8961 /* 8962 * Directory entry removal dependencies. 8963 * 8964 * When removing a directory entry, the entry's inode pointer must be 8965 * zero'ed on disk before the corresponding inode's link count is decremented 8966 * (possibly freeing the inode for re-use). This dependency is handled by 8967 * updating the directory entry but delaying the inode count reduction until 8968 * after the directory block has been written to disk. After this point, the 8969 * inode count can be decremented whenever it is convenient. 8970 */ 8971 8972 /* 8973 * This routine should be called immediately after removing 8974 * a directory entry. The inode's link count should not be 8975 * decremented by the calling procedure -- the soft updates 8976 * code will do this task when it is safe. 8977 */ 8978 void 8979 softdep_setup_remove(bp, dp, ip, isrmdir) 8980 struct buf *bp; /* buffer containing directory block */ 8981 struct inode *dp; /* inode for the directory being modified */ 8982 struct inode *ip; /* inode for directory entry being removed */ 8983 int isrmdir; /* indicates if doing RMDIR */ 8984 { 8985 struct dirrem *dirrem, *prevdirrem; 8986 struct inodedep *inodedep; 8987 struct ufsmount *ump; 8988 int direct; 8989 8990 ump = ITOUMP(ip); 8991 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 8992 ("softdep_setup_remove called on non-softdep filesystem")); 8993 /* 8994 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8995 * newdirrem() to setup the full directory remove which requires 8996 * isrmdir > 1. 8997 */ 8998 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8999 /* 9000 * Add the dirrem to the inodedep's pending remove list for quick 9001 * discovery later. 9002 */ 9003 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) 9004 panic("softdep_setup_remove: Lost inodedep."); 9005 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 9006 dirrem->dm_state |= ONDEPLIST; 9007 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9008 9009 /* 9010 * If the COMPLETE flag is clear, then there were no active 9011 * entries and we want to roll back to a zeroed entry until 9012 * the new inode is committed to disk. If the COMPLETE flag is 9013 * set then we have deleted an entry that never made it to 9014 * disk. If the entry we deleted resulted from a name change, 9015 * then the old name still resides on disk. We cannot delete 9016 * its inode (returned to us in prevdirrem) until the zeroed 9017 * directory entry gets to disk. The new inode has never been 9018 * referenced on the disk, so can be deleted immediately. 9019 */ 9020 if ((dirrem->dm_state & COMPLETE) == 0) { 9021 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 9022 dm_next); 9023 FREE_LOCK(ump); 9024 } else { 9025 if (prevdirrem != NULL) 9026 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 9027 prevdirrem, dm_next); 9028 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 9029 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 9030 FREE_LOCK(ump); 9031 if (direct) 9032 handle_workitem_remove(dirrem, 0); 9033 } 9034 } 9035 9036 /* 9037 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 9038 * pd_pendinghd list of a pagedep. 9039 */ 9040 static struct diradd * 9041 diradd_lookup(pagedep, offset) 9042 struct pagedep *pagedep; 9043 int offset; 9044 { 9045 struct diradd *dap; 9046 9047 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 9048 if (dap->da_offset == offset) 9049 return (dap); 9050 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 9051 if (dap->da_offset == offset) 9052 return (dap); 9053 return (NULL); 9054 } 9055 9056 /* 9057 * Search for a .. diradd dependency in a directory that is being removed. 9058 * If the directory was renamed to a new parent we have a diradd rather 9059 * than a mkdir for the .. entry. We need to cancel it now before 9060 * it is found in truncate(). 9061 */ 9062 static struct jremref * 9063 cancel_diradd_dotdot(ip, dirrem, jremref) 9064 struct inode *ip; 9065 struct dirrem *dirrem; 9066 struct jremref *jremref; 9067 { 9068 struct pagedep *pagedep; 9069 struct diradd *dap; 9070 struct worklist *wk; 9071 9072 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0) 9073 return (jremref); 9074 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 9075 if (dap == NULL) 9076 return (jremref); 9077 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 9078 /* 9079 * Mark any journal work as belonging to the parent so it is freed 9080 * with the .. reference. 9081 */ 9082 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9083 wk->wk_state |= MKDIR_PARENT; 9084 return (NULL); 9085 } 9086 9087 /* 9088 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 9089 * replace it with a dirrem/diradd pair as a result of re-parenting a 9090 * directory. This ensures that we don't simultaneously have a mkdir and 9091 * a diradd for the same .. entry. 9092 */ 9093 static struct jremref * 9094 cancel_mkdir_dotdot(ip, dirrem, jremref) 9095 struct inode *ip; 9096 struct dirrem *dirrem; 9097 struct jremref *jremref; 9098 { 9099 struct inodedep *inodedep; 9100 struct jaddref *jaddref; 9101 struct ufsmount *ump; 9102 struct mkdir *mkdir; 9103 struct diradd *dap; 9104 struct mount *mp; 9105 9106 mp = ITOVFS(ip); 9107 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9108 return (jremref); 9109 dap = inodedep->id_mkdiradd; 9110 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9111 return (jremref); 9112 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9113 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9114 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9115 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9116 break; 9117 if (mkdir == NULL) 9118 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9119 if ((jaddref = mkdir->md_jaddref) != NULL) { 9120 mkdir->md_jaddref = NULL; 9121 jaddref->ja_state &= ~MKDIR_PARENT; 9122 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0) 9123 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9124 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9125 journal_jremref(dirrem, jremref, inodedep); 9126 jremref = NULL; 9127 } 9128 } 9129 if (mkdir->md_state & ONWORKLIST) 9130 WORKLIST_REMOVE(&mkdir->md_list); 9131 mkdir->md_state |= ALLCOMPLETE; 9132 complete_mkdir(mkdir); 9133 return (jremref); 9134 } 9135 9136 static void 9137 journal_jremref(dirrem, jremref, inodedep) 9138 struct dirrem *dirrem; 9139 struct jremref *jremref; 9140 struct inodedep *inodedep; 9141 { 9142 9143 if (inodedep == NULL) 9144 if (inodedep_lookup(jremref->jr_list.wk_mp, 9145 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9146 panic("journal_jremref: Lost inodedep"); 9147 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9148 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9149 add_to_journal(&jremref->jr_list); 9150 } 9151 9152 static void 9153 dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9154 struct dirrem *dirrem; 9155 struct jremref *jremref; 9156 struct jremref *dotremref; 9157 struct jremref *dotdotremref; 9158 { 9159 struct inodedep *inodedep; 9160 9161 9162 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9163 &inodedep) == 0) 9164 panic("dirrem_journal: Lost inodedep"); 9165 journal_jremref(dirrem, jremref, inodedep); 9166 if (dotremref) 9167 journal_jremref(dirrem, dotremref, inodedep); 9168 if (dotdotremref) 9169 journal_jremref(dirrem, dotdotremref, NULL); 9170 } 9171 9172 /* 9173 * Allocate a new dirrem if appropriate and return it along with 9174 * its associated pagedep. Called without a lock, returns with lock. 9175 */ 9176 static struct dirrem * 9177 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9178 struct buf *bp; /* buffer containing directory block */ 9179 struct inode *dp; /* inode for the directory being modified */ 9180 struct inode *ip; /* inode for directory entry being removed */ 9181 int isrmdir; /* indicates if doing RMDIR */ 9182 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9183 { 9184 int offset; 9185 ufs_lbn_t lbn; 9186 struct diradd *dap; 9187 struct dirrem *dirrem; 9188 struct pagedep *pagedep; 9189 struct jremref *jremref; 9190 struct jremref *dotremref; 9191 struct jremref *dotdotremref; 9192 struct vnode *dvp; 9193 struct ufsmount *ump; 9194 9195 /* 9196 * Whiteouts have no deletion dependencies. 9197 */ 9198 if (ip == NULL) 9199 panic("newdirrem: whiteout"); 9200 dvp = ITOV(dp); 9201 ump = ITOUMP(dp); 9202 9203 /* 9204 * If the system is over its limit and our filesystem is 9205 * responsible for more than our share of that usage and 9206 * we are not a snapshot, request some inodedep cleanup. 9207 * Limiting the number of dirrem structures will also limit 9208 * the number of freefile and freeblks structures. 9209 */ 9210 ACQUIRE_LOCK(ump); 9211 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM)) 9212 schedule_cleanup(UFSTOVFS(ump)); 9213 else 9214 FREE_LOCK(ump); 9215 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9216 M_ZERO); 9217 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9218 LIST_INIT(&dirrem->dm_jremrefhd); 9219 LIST_INIT(&dirrem->dm_jwork); 9220 dirrem->dm_state = isrmdir ? RMDIR : 0; 9221 dirrem->dm_oldinum = ip->i_number; 9222 *prevdirremp = NULL; 9223 /* 9224 * Allocate remove reference structures to track journal write 9225 * dependencies. We will always have one for the link and 9226 * when doing directories we will always have one more for dot. 9227 * When renaming a directory we skip the dotdot link change so 9228 * this is not needed. 9229 */ 9230 jremref = dotremref = dotdotremref = NULL; 9231 if (DOINGSUJ(dvp)) { 9232 if (isrmdir) { 9233 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9234 ip->i_effnlink + 2); 9235 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9236 ip->i_effnlink + 1); 9237 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9238 dp->i_effnlink + 1); 9239 dotdotremref->jr_state |= MKDIR_PARENT; 9240 } else 9241 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9242 ip->i_effnlink + 1); 9243 } 9244 ACQUIRE_LOCK(ump); 9245 lbn = lblkno(ump->um_fs, dp->i_offset); 9246 offset = blkoff(ump->um_fs, dp->i_offset); 9247 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC, 9248 &pagedep); 9249 dirrem->dm_pagedep = pagedep; 9250 dirrem->dm_offset = offset; 9251 /* 9252 * If we're renaming a .. link to a new directory, cancel any 9253 * existing MKDIR_PARENT mkdir. If it has already been canceled 9254 * the jremref is preserved for any potential diradd in this 9255 * location. This can not coincide with a rmdir. 9256 */ 9257 if (dp->i_offset == DOTDOT_OFFSET) { 9258 if (isrmdir) 9259 panic("newdirrem: .. directory change during remove?"); 9260 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9261 } 9262 /* 9263 * If we're removing a directory search for the .. dependency now and 9264 * cancel it. Any pending journal work will be added to the dirrem 9265 * to be completed when the workitem remove completes. 9266 */ 9267 if (isrmdir) 9268 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9269 /* 9270 * Check for a diradd dependency for the same directory entry. 9271 * If present, then both dependencies become obsolete and can 9272 * be de-allocated. 9273 */ 9274 dap = diradd_lookup(pagedep, offset); 9275 if (dap == NULL) { 9276 /* 9277 * Link the jremref structures into the dirrem so they are 9278 * written prior to the pagedep. 9279 */ 9280 if (jremref) 9281 dirrem_journal(dirrem, jremref, dotremref, 9282 dotdotremref); 9283 return (dirrem); 9284 } 9285 /* 9286 * Must be ATTACHED at this point. 9287 */ 9288 if ((dap->da_state & ATTACHED) == 0) 9289 panic("newdirrem: not ATTACHED"); 9290 if (dap->da_newinum != ip->i_number) 9291 panic("newdirrem: inum %ju should be %ju", 9292 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9293 /* 9294 * If we are deleting a changed name that never made it to disk, 9295 * then return the dirrem describing the previous inode (which 9296 * represents the inode currently referenced from this entry on disk). 9297 */ 9298 if ((dap->da_state & DIRCHG) != 0) { 9299 *prevdirremp = dap->da_previous; 9300 dap->da_state &= ~DIRCHG; 9301 dap->da_pagedep = pagedep; 9302 } 9303 /* 9304 * We are deleting an entry that never made it to disk. 9305 * Mark it COMPLETE so we can delete its inode immediately. 9306 */ 9307 dirrem->dm_state |= COMPLETE; 9308 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9309 #ifdef INVARIANTS 9310 if (isrmdir == 0) { 9311 struct worklist *wk; 9312 9313 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9314 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9315 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9316 } 9317 #endif 9318 9319 return (dirrem); 9320 } 9321 9322 /* 9323 * Directory entry change dependencies. 9324 * 9325 * Changing an existing directory entry requires that an add operation 9326 * be completed first followed by a deletion. The semantics for the addition 9327 * are identical to the description of adding a new entry above except 9328 * that the rollback is to the old inode number rather than zero. Once 9329 * the addition dependency is completed, the removal is done as described 9330 * in the removal routine above. 9331 */ 9332 9333 /* 9334 * This routine should be called immediately after changing 9335 * a directory entry. The inode's link count should not be 9336 * decremented by the calling procedure -- the soft updates 9337 * code will perform this task when it is safe. 9338 */ 9339 void 9340 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9341 struct buf *bp; /* buffer containing directory block */ 9342 struct inode *dp; /* inode for the directory being modified */ 9343 struct inode *ip; /* inode for directory entry being removed */ 9344 ino_t newinum; /* new inode number for changed entry */ 9345 int isrmdir; /* indicates if doing RMDIR */ 9346 { 9347 int offset; 9348 struct diradd *dap = NULL; 9349 struct dirrem *dirrem, *prevdirrem; 9350 struct pagedep *pagedep; 9351 struct inodedep *inodedep; 9352 struct jaddref *jaddref; 9353 struct mount *mp; 9354 struct ufsmount *ump; 9355 9356 mp = ITOVFS(dp); 9357 ump = VFSTOUFS(mp); 9358 offset = blkoff(ump->um_fs, dp->i_offset); 9359 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9360 ("softdep_setup_directory_change called on non-softdep filesystem")); 9361 9362 /* 9363 * Whiteouts do not need diradd dependencies. 9364 */ 9365 if (newinum != UFS_WINO) { 9366 dap = malloc(sizeof(struct diradd), 9367 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9368 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9369 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9370 dap->da_offset = offset; 9371 dap->da_newinum = newinum; 9372 LIST_INIT(&dap->da_jwork); 9373 } 9374 9375 /* 9376 * Allocate a new dirrem and ACQUIRE_LOCK. 9377 */ 9378 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9379 pagedep = dirrem->dm_pagedep; 9380 /* 9381 * The possible values for isrmdir: 9382 * 0 - non-directory file rename 9383 * 1 - directory rename within same directory 9384 * inum - directory rename to new directory of given inode number 9385 * When renaming to a new directory, we are both deleting and 9386 * creating a new directory entry, so the link count on the new 9387 * directory should not change. Thus we do not need the followup 9388 * dirrem which is usually done in handle_workitem_remove. We set 9389 * the DIRCHG flag to tell handle_workitem_remove to skip the 9390 * followup dirrem. 9391 */ 9392 if (isrmdir > 1) 9393 dirrem->dm_state |= DIRCHG; 9394 9395 /* 9396 * Whiteouts have no additional dependencies, 9397 * so just put the dirrem on the correct list. 9398 */ 9399 if (newinum == UFS_WINO) { 9400 if ((dirrem->dm_state & COMPLETE) == 0) { 9401 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9402 dm_next); 9403 } else { 9404 dirrem->dm_dirinum = pagedep->pd_ino; 9405 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9406 add_to_worklist(&dirrem->dm_list, 0); 9407 } 9408 FREE_LOCK(ump); 9409 return; 9410 } 9411 /* 9412 * Add the dirrem to the inodedep's pending remove list for quick 9413 * discovery later. A valid nlinkdelta ensures that this lookup 9414 * will not fail. 9415 */ 9416 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9417 panic("softdep_setup_directory_change: Lost inodedep."); 9418 dirrem->dm_state |= ONDEPLIST; 9419 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9420 9421 /* 9422 * If the COMPLETE flag is clear, then there were no active 9423 * entries and we want to roll back to the previous inode until 9424 * the new inode is committed to disk. If the COMPLETE flag is 9425 * set, then we have deleted an entry that never made it to disk. 9426 * If the entry we deleted resulted from a name change, then the old 9427 * inode reference still resides on disk. Any rollback that we do 9428 * needs to be to that old inode (returned to us in prevdirrem). If 9429 * the entry we deleted resulted from a create, then there is 9430 * no entry on the disk, so we want to roll back to zero rather 9431 * than the uncommitted inode. In either of the COMPLETE cases we 9432 * want to immediately free the unwritten and unreferenced inode. 9433 */ 9434 if ((dirrem->dm_state & COMPLETE) == 0) { 9435 dap->da_previous = dirrem; 9436 } else { 9437 if (prevdirrem != NULL) { 9438 dap->da_previous = prevdirrem; 9439 } else { 9440 dap->da_state &= ~DIRCHG; 9441 dap->da_pagedep = pagedep; 9442 } 9443 dirrem->dm_dirinum = pagedep->pd_ino; 9444 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9445 add_to_worklist(&dirrem->dm_list, 0); 9446 } 9447 /* 9448 * Lookup the jaddref for this journal entry. We must finish 9449 * initializing it and make the diradd write dependent on it. 9450 * If we're not journaling, put it on the id_bufwait list if the 9451 * inode is not yet written. If it is written, do the post-inode 9452 * write processing to put it on the id_pendinghd list. 9453 */ 9454 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9455 if (MOUNTEDSUJ(mp)) { 9456 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9457 inoreflst); 9458 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9459 ("softdep_setup_directory_change: bad jaddref %p", 9460 jaddref)); 9461 jaddref->ja_diroff = dp->i_offset; 9462 jaddref->ja_diradd = dap; 9463 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9464 dap, da_pdlist); 9465 add_to_journal(&jaddref->ja_list); 9466 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9467 dap->da_state |= COMPLETE; 9468 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9469 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9470 } else { 9471 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9472 dap, da_pdlist); 9473 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9474 } 9475 /* 9476 * If we're making a new name for a directory that has not been 9477 * committed when need to move the dot and dotdot references to 9478 * this new name. 9479 */ 9480 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9481 merge_diradd(inodedep, dap); 9482 FREE_LOCK(ump); 9483 } 9484 9485 /* 9486 * Called whenever the link count on an inode is changed. 9487 * It creates an inode dependency so that the new reference(s) 9488 * to the inode cannot be committed to disk until the updated 9489 * inode has been written. 9490 */ 9491 void 9492 softdep_change_linkcnt(ip) 9493 struct inode *ip; /* the inode with the increased link count */ 9494 { 9495 struct inodedep *inodedep; 9496 struct ufsmount *ump; 9497 9498 ump = ITOUMP(ip); 9499 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9500 ("softdep_change_linkcnt called on non-softdep filesystem")); 9501 ACQUIRE_LOCK(ump); 9502 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 9503 if (ip->i_nlink < ip->i_effnlink) 9504 panic("softdep_change_linkcnt: bad delta"); 9505 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9506 FREE_LOCK(ump); 9507 } 9508 9509 /* 9510 * Attach a sbdep dependency to the superblock buf so that we can keep 9511 * track of the head of the linked list of referenced but unlinked inodes. 9512 */ 9513 void 9514 softdep_setup_sbupdate(ump, fs, bp) 9515 struct ufsmount *ump; 9516 struct fs *fs; 9517 struct buf *bp; 9518 { 9519 struct sbdep *sbdep; 9520 struct worklist *wk; 9521 9522 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9523 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9524 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9525 if (wk->wk_type == D_SBDEP) 9526 break; 9527 if (wk != NULL) 9528 return; 9529 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9530 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9531 sbdep->sb_fs = fs; 9532 sbdep->sb_ump = ump; 9533 ACQUIRE_LOCK(ump); 9534 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9535 FREE_LOCK(ump); 9536 } 9537 9538 /* 9539 * Return the first unlinked inodedep which is ready to be the head of the 9540 * list. The inodedep and all those after it must have valid next pointers. 9541 */ 9542 static struct inodedep * 9543 first_unlinked_inodedep(ump) 9544 struct ufsmount *ump; 9545 { 9546 struct inodedep *inodedep; 9547 struct inodedep *idp; 9548 9549 LOCK_OWNED(ump); 9550 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9551 inodedep; inodedep = idp) { 9552 if ((inodedep->id_state & UNLINKNEXT) == 0) 9553 return (NULL); 9554 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9555 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9556 break; 9557 if ((inodedep->id_state & UNLINKPREV) == 0) 9558 break; 9559 } 9560 return (inodedep); 9561 } 9562 9563 /* 9564 * Set the sujfree unlinked head pointer prior to writing a superblock. 9565 */ 9566 static void 9567 initiate_write_sbdep(sbdep) 9568 struct sbdep *sbdep; 9569 { 9570 struct inodedep *inodedep; 9571 struct fs *bpfs; 9572 struct fs *fs; 9573 9574 bpfs = sbdep->sb_fs; 9575 fs = sbdep->sb_ump->um_fs; 9576 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9577 if (inodedep) { 9578 fs->fs_sujfree = inodedep->id_ino; 9579 inodedep->id_state |= UNLINKPREV; 9580 } else 9581 fs->fs_sujfree = 0; 9582 bpfs->fs_sujfree = fs->fs_sujfree; 9583 } 9584 9585 /* 9586 * After a superblock is written determine whether it must be written again 9587 * due to a changing unlinked list head. 9588 */ 9589 static int 9590 handle_written_sbdep(sbdep, bp) 9591 struct sbdep *sbdep; 9592 struct buf *bp; 9593 { 9594 struct inodedep *inodedep; 9595 struct fs *fs; 9596 9597 LOCK_OWNED(sbdep->sb_ump); 9598 fs = sbdep->sb_fs; 9599 /* 9600 * If the superblock doesn't match the in-memory list start over. 9601 */ 9602 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9603 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9604 (inodedep == NULL && fs->fs_sujfree != 0)) { 9605 bdirty(bp); 9606 return (1); 9607 } 9608 WORKITEM_FREE(sbdep, D_SBDEP); 9609 if (fs->fs_sujfree == 0) 9610 return (0); 9611 /* 9612 * Now that we have a record of this inode in stable store allow it 9613 * to be written to free up pending work. Inodes may see a lot of 9614 * write activity after they are unlinked which we must not hold up. 9615 */ 9616 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9617 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9618 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9619 inodedep, inodedep->id_state); 9620 if (inodedep->id_state & UNLINKONLIST) 9621 break; 9622 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9623 } 9624 9625 return (0); 9626 } 9627 9628 /* 9629 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9630 */ 9631 static void 9632 unlinked_inodedep(mp, inodedep) 9633 struct mount *mp; 9634 struct inodedep *inodedep; 9635 { 9636 struct ufsmount *ump; 9637 9638 ump = VFSTOUFS(mp); 9639 LOCK_OWNED(ump); 9640 if (MOUNTEDSUJ(mp) == 0) 9641 return; 9642 ump->um_fs->fs_fmod = 1; 9643 if (inodedep->id_state & UNLINKED) 9644 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9645 inodedep->id_state |= UNLINKED; 9646 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9647 } 9648 9649 /* 9650 * Remove an inodedep from the unlinked inodedep list. This may require 9651 * disk writes if the inode has made it that far. 9652 */ 9653 static void 9654 clear_unlinked_inodedep(inodedep) 9655 struct inodedep *inodedep; 9656 { 9657 struct ufs2_dinode *dip; 9658 struct ufsmount *ump; 9659 struct inodedep *idp; 9660 struct inodedep *idn; 9661 struct fs *fs; 9662 struct buf *bp; 9663 ino_t ino; 9664 ino_t nino; 9665 ino_t pino; 9666 int error; 9667 9668 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9669 fs = ump->um_fs; 9670 ino = inodedep->id_ino; 9671 error = 0; 9672 for (;;) { 9673 LOCK_OWNED(ump); 9674 KASSERT((inodedep->id_state & UNLINKED) != 0, 9675 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9676 inodedep)); 9677 /* 9678 * If nothing has yet been written simply remove us from 9679 * the in memory list and return. This is the most common 9680 * case where handle_workitem_remove() loses the final 9681 * reference. 9682 */ 9683 if ((inodedep->id_state & UNLINKLINKS) == 0) 9684 break; 9685 /* 9686 * If we have a NEXT pointer and no PREV pointer we can simply 9687 * clear NEXT's PREV and remove ourselves from the list. Be 9688 * careful not to clear PREV if the superblock points at 9689 * next as well. 9690 */ 9691 idn = TAILQ_NEXT(inodedep, id_unlinked); 9692 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9693 if (idn && fs->fs_sujfree != idn->id_ino) 9694 idn->id_state &= ~UNLINKPREV; 9695 break; 9696 } 9697 /* 9698 * Here we have an inodedep which is actually linked into 9699 * the list. We must remove it by forcing a write to the 9700 * link before us, whether it be the superblock or an inode. 9701 * Unfortunately the list may change while we're waiting 9702 * on the buf lock for either resource so we must loop until 9703 * we lock the right one. If both the superblock and an 9704 * inode point to this inode we must clear the inode first 9705 * followed by the superblock. 9706 */ 9707 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9708 pino = 0; 9709 if (idp && (idp->id_state & UNLINKNEXT)) 9710 pino = idp->id_ino; 9711 FREE_LOCK(ump); 9712 if (pino == 0) { 9713 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9714 (int)fs->fs_sbsize, 0, 0, 0); 9715 } else { 9716 error = bread(ump->um_devvp, 9717 fsbtodb(fs, ino_to_fsba(fs, pino)), 9718 (int)fs->fs_bsize, NOCRED, &bp); 9719 if (error) 9720 brelse(bp); 9721 } 9722 ACQUIRE_LOCK(ump); 9723 if (error) 9724 break; 9725 /* If the list has changed restart the loop. */ 9726 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9727 nino = 0; 9728 if (idp && (idp->id_state & UNLINKNEXT)) 9729 nino = idp->id_ino; 9730 if (nino != pino || 9731 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9732 FREE_LOCK(ump); 9733 brelse(bp); 9734 ACQUIRE_LOCK(ump); 9735 continue; 9736 } 9737 nino = 0; 9738 idn = TAILQ_NEXT(inodedep, id_unlinked); 9739 if (idn) 9740 nino = idn->id_ino; 9741 /* 9742 * Remove us from the in memory list. After this we cannot 9743 * access the inodedep. 9744 */ 9745 KASSERT((inodedep->id_state & UNLINKED) != 0, 9746 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9747 inodedep)); 9748 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9749 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9750 FREE_LOCK(ump); 9751 /* 9752 * The predecessor's next pointer is manually updated here 9753 * so that the NEXT flag is never cleared for an element 9754 * that is in the list. 9755 */ 9756 if (pino == 0) { 9757 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9758 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9759 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9760 bp); 9761 } else if (fs->fs_magic == FS_UFS1_MAGIC) { 9762 ((struct ufs1_dinode *)bp->b_data + 9763 ino_to_fsbo(fs, pino))->di_freelink = nino; 9764 } else { 9765 dip = (struct ufs2_dinode *)bp->b_data + 9766 ino_to_fsbo(fs, pino); 9767 dip->di_freelink = nino; 9768 ffs_update_dinode_ckhash(fs, dip); 9769 } 9770 /* 9771 * If the bwrite fails we have no recourse to recover. The 9772 * filesystem is corrupted already. 9773 */ 9774 bwrite(bp); 9775 ACQUIRE_LOCK(ump); 9776 /* 9777 * If the superblock pointer still needs to be cleared force 9778 * a write here. 9779 */ 9780 if (fs->fs_sujfree == ino) { 9781 FREE_LOCK(ump); 9782 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9783 (int)fs->fs_sbsize, 0, 0, 0); 9784 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9785 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9786 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9787 bp); 9788 bwrite(bp); 9789 ACQUIRE_LOCK(ump); 9790 } 9791 9792 if (fs->fs_sujfree != ino) 9793 return; 9794 panic("clear_unlinked_inodedep: Failed to clear free head"); 9795 } 9796 if (inodedep->id_ino == fs->fs_sujfree) 9797 panic("clear_unlinked_inodedep: Freeing head of free list"); 9798 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9799 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9800 return; 9801 } 9802 9803 /* 9804 * This workitem decrements the inode's link count. 9805 * If the link count reaches zero, the file is removed. 9806 */ 9807 static int 9808 handle_workitem_remove(dirrem, flags) 9809 struct dirrem *dirrem; 9810 int flags; 9811 { 9812 struct inodedep *inodedep; 9813 struct workhead dotdotwk; 9814 struct worklist *wk; 9815 struct ufsmount *ump; 9816 struct mount *mp; 9817 struct vnode *vp; 9818 struct inode *ip; 9819 ino_t oldinum; 9820 9821 if (dirrem->dm_state & ONWORKLIST) 9822 panic("handle_workitem_remove: dirrem %p still on worklist", 9823 dirrem); 9824 oldinum = dirrem->dm_oldinum; 9825 mp = dirrem->dm_list.wk_mp; 9826 ump = VFSTOUFS(mp); 9827 flags |= LK_EXCLUSIVE; 9828 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9829 return (EBUSY); 9830 ip = VTOI(vp); 9831 ACQUIRE_LOCK(ump); 9832 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9833 panic("handle_workitem_remove: lost inodedep"); 9834 if (dirrem->dm_state & ONDEPLIST) 9835 LIST_REMOVE(dirrem, dm_inonext); 9836 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9837 ("handle_workitem_remove: Journal entries not written.")); 9838 9839 /* 9840 * Move all dependencies waiting on the remove to complete 9841 * from the dirrem to the inode inowait list to be completed 9842 * after the inode has been updated and written to disk. Any 9843 * marked MKDIR_PARENT are saved to be completed when the .. ref 9844 * is removed. 9845 */ 9846 LIST_INIT(&dotdotwk); 9847 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9848 WORKLIST_REMOVE(wk); 9849 if (wk->wk_state & MKDIR_PARENT) { 9850 wk->wk_state &= ~MKDIR_PARENT; 9851 WORKLIST_INSERT(&dotdotwk, wk); 9852 continue; 9853 } 9854 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9855 } 9856 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9857 /* 9858 * Normal file deletion. 9859 */ 9860 if ((dirrem->dm_state & RMDIR) == 0) { 9861 ip->i_nlink--; 9862 DIP_SET(ip, i_nlink, ip->i_nlink); 9863 ip->i_flag |= IN_CHANGE; 9864 if (ip->i_nlink < ip->i_effnlink) 9865 panic("handle_workitem_remove: bad file delta"); 9866 if (ip->i_nlink == 0) 9867 unlinked_inodedep(mp, inodedep); 9868 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9869 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9870 ("handle_workitem_remove: worklist not empty. %s", 9871 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9872 WORKITEM_FREE(dirrem, D_DIRREM); 9873 FREE_LOCK(ump); 9874 goto out; 9875 } 9876 /* 9877 * Directory deletion. Decrement reference count for both the 9878 * just deleted parent directory entry and the reference for ".". 9879 * Arrange to have the reference count on the parent decremented 9880 * to account for the loss of "..". 9881 */ 9882 ip->i_nlink -= 2; 9883 DIP_SET(ip, i_nlink, ip->i_nlink); 9884 ip->i_flag |= IN_CHANGE; 9885 if (ip->i_nlink < ip->i_effnlink) 9886 panic("handle_workitem_remove: bad dir delta"); 9887 if (ip->i_nlink == 0) 9888 unlinked_inodedep(mp, inodedep); 9889 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9890 /* 9891 * Rename a directory to a new parent. Since, we are both deleting 9892 * and creating a new directory entry, the link count on the new 9893 * directory should not change. Thus we skip the followup dirrem. 9894 */ 9895 if (dirrem->dm_state & DIRCHG) { 9896 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9897 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9898 WORKITEM_FREE(dirrem, D_DIRREM); 9899 FREE_LOCK(ump); 9900 goto out; 9901 } 9902 dirrem->dm_state = ONDEPLIST; 9903 dirrem->dm_oldinum = dirrem->dm_dirinum; 9904 /* 9905 * Place the dirrem on the parent's diremhd list. 9906 */ 9907 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9908 panic("handle_workitem_remove: lost dir inodedep"); 9909 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9910 /* 9911 * If the allocated inode has never been written to disk, then 9912 * the on-disk inode is zero'ed and we can remove the file 9913 * immediately. When journaling if the inode has been marked 9914 * unlinked and not DEPCOMPLETE we know it can never be written. 9915 */ 9916 inodedep_lookup(mp, oldinum, 0, &inodedep); 9917 if (inodedep == NULL || 9918 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9919 check_inode_unwritten(inodedep)) { 9920 FREE_LOCK(ump); 9921 vput(vp); 9922 return handle_workitem_remove(dirrem, flags); 9923 } 9924 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9925 FREE_LOCK(ump); 9926 ip->i_flag |= IN_CHANGE; 9927 out: 9928 ffs_update(vp, 0); 9929 vput(vp); 9930 return (0); 9931 } 9932 9933 /* 9934 * Inode de-allocation dependencies. 9935 * 9936 * When an inode's link count is reduced to zero, it can be de-allocated. We 9937 * found it convenient to postpone de-allocation until after the inode is 9938 * written to disk with its new link count (zero). At this point, all of the 9939 * on-disk inode's block pointers are nullified and, with careful dependency 9940 * list ordering, all dependencies related to the inode will be satisfied and 9941 * the corresponding dependency structures de-allocated. So, if/when the 9942 * inode is reused, there will be no mixing of old dependencies with new 9943 * ones. This artificial dependency is set up by the block de-allocation 9944 * procedure above (softdep_setup_freeblocks) and completed by the 9945 * following procedure. 9946 */ 9947 static void 9948 handle_workitem_freefile(freefile) 9949 struct freefile *freefile; 9950 { 9951 struct workhead wkhd; 9952 struct fs *fs; 9953 struct ufsmount *ump; 9954 int error; 9955 #ifdef INVARIANTS 9956 struct inodedep *idp; 9957 #endif 9958 9959 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9960 fs = ump->um_fs; 9961 #ifdef INVARIANTS 9962 ACQUIRE_LOCK(ump); 9963 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9964 FREE_LOCK(ump); 9965 if (error) 9966 panic("handle_workitem_freefile: inodedep %p survived", idp); 9967 #endif 9968 UFS_LOCK(ump); 9969 fs->fs_pendinginodes -= 1; 9970 UFS_UNLOCK(ump); 9971 LIST_INIT(&wkhd); 9972 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9973 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9974 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9975 softdep_error("handle_workitem_freefile", error); 9976 ACQUIRE_LOCK(ump); 9977 WORKITEM_FREE(freefile, D_FREEFILE); 9978 FREE_LOCK(ump); 9979 } 9980 9981 9982 /* 9983 * Helper function which unlinks marker element from work list and returns 9984 * the next element on the list. 9985 */ 9986 static __inline struct worklist * 9987 markernext(struct worklist *marker) 9988 { 9989 struct worklist *next; 9990 9991 next = LIST_NEXT(marker, wk_list); 9992 LIST_REMOVE(marker, wk_list); 9993 return next; 9994 } 9995 9996 /* 9997 * Disk writes. 9998 * 9999 * The dependency structures constructed above are most actively used when file 10000 * system blocks are written to disk. No constraints are placed on when a 10001 * block can be written, but unsatisfied update dependencies are made safe by 10002 * modifying (or replacing) the source memory for the duration of the disk 10003 * write. When the disk write completes, the memory block is again brought 10004 * up-to-date. 10005 * 10006 * In-core inode structure reclamation. 10007 * 10008 * Because there are a finite number of "in-core" inode structures, they are 10009 * reused regularly. By transferring all inode-related dependencies to the 10010 * in-memory inode block and indexing them separately (via "inodedep"s), we 10011 * can allow "in-core" inode structures to be reused at any time and avoid 10012 * any increase in contention. 10013 * 10014 * Called just before entering the device driver to initiate a new disk I/O. 10015 * The buffer must be locked, thus, no I/O completion operations can occur 10016 * while we are manipulating its associated dependencies. 10017 */ 10018 static void 10019 softdep_disk_io_initiation(bp) 10020 struct buf *bp; /* structure describing disk write to occur */ 10021 { 10022 struct worklist *wk; 10023 struct worklist marker; 10024 struct inodedep *inodedep; 10025 struct freeblks *freeblks; 10026 struct jblkdep *jblkdep; 10027 struct newblk *newblk; 10028 struct ufsmount *ump; 10029 10030 /* 10031 * We only care about write operations. There should never 10032 * be dependencies for reads. 10033 */ 10034 if (bp->b_iocmd != BIO_WRITE) 10035 panic("softdep_disk_io_initiation: not write"); 10036 10037 if (bp->b_vflags & BV_BKGRDINPROG) 10038 panic("softdep_disk_io_initiation: Writing buffer with " 10039 "background write in progress: %p", bp); 10040 10041 ump = softdep_bp_to_mp(bp); 10042 if (ump == NULL) 10043 return; 10044 10045 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 10046 PHOLD(curproc); /* Don't swap out kernel stack */ 10047 ACQUIRE_LOCK(ump); 10048 /* 10049 * Do any necessary pre-I/O processing. 10050 */ 10051 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 10052 wk = markernext(&marker)) { 10053 LIST_INSERT_AFTER(wk, &marker, wk_list); 10054 switch (wk->wk_type) { 10055 10056 case D_PAGEDEP: 10057 initiate_write_filepage(WK_PAGEDEP(wk), bp); 10058 continue; 10059 10060 case D_INODEDEP: 10061 inodedep = WK_INODEDEP(wk); 10062 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 10063 initiate_write_inodeblock_ufs1(inodedep, bp); 10064 else 10065 initiate_write_inodeblock_ufs2(inodedep, bp); 10066 continue; 10067 10068 case D_INDIRDEP: 10069 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 10070 continue; 10071 10072 case D_BMSAFEMAP: 10073 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 10074 continue; 10075 10076 case D_JSEG: 10077 WK_JSEG(wk)->js_buf = NULL; 10078 continue; 10079 10080 case D_FREEBLKS: 10081 freeblks = WK_FREEBLKS(wk); 10082 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 10083 /* 10084 * We have to wait for the freeblks to be journaled 10085 * before we can write an inodeblock with updated 10086 * pointers. Be careful to arrange the marker so 10087 * we revisit the freeblks if it's not removed by 10088 * the first jwait(). 10089 */ 10090 if (jblkdep != NULL) { 10091 LIST_REMOVE(&marker, wk_list); 10092 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10093 jwait(&jblkdep->jb_list, MNT_WAIT); 10094 } 10095 continue; 10096 case D_ALLOCDIRECT: 10097 case D_ALLOCINDIR: 10098 /* 10099 * We have to wait for the jnewblk to be journaled 10100 * before we can write to a block if the contents 10101 * may be confused with an earlier file's indirect 10102 * at recovery time. Handle the marker as described 10103 * above. 10104 */ 10105 newblk = WK_NEWBLK(wk); 10106 if (newblk->nb_jnewblk != NULL && 10107 indirblk_lookup(newblk->nb_list.wk_mp, 10108 newblk->nb_newblkno)) { 10109 LIST_REMOVE(&marker, wk_list); 10110 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10111 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10112 } 10113 continue; 10114 10115 case D_SBDEP: 10116 initiate_write_sbdep(WK_SBDEP(wk)); 10117 continue; 10118 10119 case D_MKDIR: 10120 case D_FREEWORK: 10121 case D_FREEDEP: 10122 case D_JSEGDEP: 10123 continue; 10124 10125 default: 10126 panic("handle_disk_io_initiation: Unexpected type %s", 10127 TYPENAME(wk->wk_type)); 10128 /* NOTREACHED */ 10129 } 10130 } 10131 FREE_LOCK(ump); 10132 PRELE(curproc); /* Allow swapout of kernel stack */ 10133 } 10134 10135 /* 10136 * Called from within the procedure above to deal with unsatisfied 10137 * allocation dependencies in a directory. The buffer must be locked, 10138 * thus, no I/O completion operations can occur while we are 10139 * manipulating its associated dependencies. 10140 */ 10141 static void 10142 initiate_write_filepage(pagedep, bp) 10143 struct pagedep *pagedep; 10144 struct buf *bp; 10145 { 10146 struct jremref *jremref; 10147 struct jmvref *jmvref; 10148 struct dirrem *dirrem; 10149 struct diradd *dap; 10150 struct direct *ep; 10151 int i; 10152 10153 if (pagedep->pd_state & IOSTARTED) { 10154 /* 10155 * This can only happen if there is a driver that does not 10156 * understand chaining. Here biodone will reissue the call 10157 * to strategy for the incomplete buffers. 10158 */ 10159 printf("initiate_write_filepage: already started\n"); 10160 return; 10161 } 10162 pagedep->pd_state |= IOSTARTED; 10163 /* 10164 * Wait for all journal remove dependencies to hit the disk. 10165 * We can not allow any potentially conflicting directory adds 10166 * to be visible before removes and rollback is too difficult. 10167 * The per-filesystem lock may be dropped and re-acquired, however 10168 * we hold the buf locked so the dependency can not go away. 10169 */ 10170 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10171 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10172 jwait(&jremref->jr_list, MNT_WAIT); 10173 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10174 jwait(&jmvref->jm_list, MNT_WAIT); 10175 for (i = 0; i < DAHASHSZ; i++) { 10176 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10177 ep = (struct direct *) 10178 ((char *)bp->b_data + dap->da_offset); 10179 if (ep->d_ino != dap->da_newinum) 10180 panic("%s: dir inum %ju != new %ju", 10181 "initiate_write_filepage", 10182 (uintmax_t)ep->d_ino, 10183 (uintmax_t)dap->da_newinum); 10184 if (dap->da_state & DIRCHG) 10185 ep->d_ino = dap->da_previous->dm_oldinum; 10186 else 10187 ep->d_ino = 0; 10188 dap->da_state &= ~ATTACHED; 10189 dap->da_state |= UNDONE; 10190 } 10191 } 10192 } 10193 10194 /* 10195 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10196 * Note that any bug fixes made to this routine must be done in the 10197 * version found below. 10198 * 10199 * Called from within the procedure above to deal with unsatisfied 10200 * allocation dependencies in an inodeblock. The buffer must be 10201 * locked, thus, no I/O completion operations can occur while we 10202 * are manipulating its associated dependencies. 10203 */ 10204 static void 10205 initiate_write_inodeblock_ufs1(inodedep, bp) 10206 struct inodedep *inodedep; 10207 struct buf *bp; /* The inode block */ 10208 { 10209 struct allocdirect *adp, *lastadp; 10210 struct ufs1_dinode *dp; 10211 struct ufs1_dinode *sip; 10212 struct inoref *inoref; 10213 struct ufsmount *ump; 10214 struct fs *fs; 10215 ufs_lbn_t i; 10216 #ifdef INVARIANTS 10217 ufs_lbn_t prevlbn = 0; 10218 #endif 10219 int deplist; 10220 10221 if (inodedep->id_state & IOSTARTED) 10222 panic("initiate_write_inodeblock_ufs1: already started"); 10223 inodedep->id_state |= IOSTARTED; 10224 fs = inodedep->id_fs; 10225 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10226 LOCK_OWNED(ump); 10227 dp = (struct ufs1_dinode *)bp->b_data + 10228 ino_to_fsbo(fs, inodedep->id_ino); 10229 10230 /* 10231 * If we're on the unlinked list but have not yet written our 10232 * next pointer initialize it here. 10233 */ 10234 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10235 struct inodedep *inon; 10236 10237 inon = TAILQ_NEXT(inodedep, id_unlinked); 10238 dp->di_freelink = inon ? inon->id_ino : 0; 10239 } 10240 /* 10241 * If the bitmap is not yet written, then the allocated 10242 * inode cannot be written to disk. 10243 */ 10244 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10245 if (inodedep->id_savedino1 != NULL) 10246 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10247 FREE_LOCK(ump); 10248 sip = malloc(sizeof(struct ufs1_dinode), 10249 M_SAVEDINO, M_SOFTDEP_FLAGS); 10250 ACQUIRE_LOCK(ump); 10251 inodedep->id_savedino1 = sip; 10252 *inodedep->id_savedino1 = *dp; 10253 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10254 dp->di_gen = inodedep->id_savedino1->di_gen; 10255 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10256 return; 10257 } 10258 /* 10259 * If no dependencies, then there is nothing to roll back. 10260 */ 10261 inodedep->id_savedsize = dp->di_size; 10262 inodedep->id_savedextsize = 0; 10263 inodedep->id_savednlink = dp->di_nlink; 10264 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10265 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10266 return; 10267 /* 10268 * Revert the link count to that of the first unwritten journal entry. 10269 */ 10270 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10271 if (inoref) 10272 dp->di_nlink = inoref->if_nlink; 10273 /* 10274 * Set the dependencies to busy. 10275 */ 10276 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10277 adp = TAILQ_NEXT(adp, ad_next)) { 10278 #ifdef INVARIANTS 10279 if (deplist != 0 && prevlbn >= adp->ad_offset) 10280 panic("softdep_write_inodeblock: lbn order"); 10281 prevlbn = adp->ad_offset; 10282 if (adp->ad_offset < UFS_NDADDR && 10283 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10284 panic("initiate_write_inodeblock_ufs1: " 10285 "direct pointer #%jd mismatch %d != %jd", 10286 (intmax_t)adp->ad_offset, 10287 dp->di_db[adp->ad_offset], 10288 (intmax_t)adp->ad_newblkno); 10289 if (adp->ad_offset >= UFS_NDADDR && 10290 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10291 panic("initiate_write_inodeblock_ufs1: " 10292 "indirect pointer #%jd mismatch %d != %jd", 10293 (intmax_t)adp->ad_offset - UFS_NDADDR, 10294 dp->di_ib[adp->ad_offset - UFS_NDADDR], 10295 (intmax_t)adp->ad_newblkno); 10296 deplist |= 1 << adp->ad_offset; 10297 if ((adp->ad_state & ATTACHED) == 0) 10298 panic("initiate_write_inodeblock_ufs1: " 10299 "Unknown state 0x%x", adp->ad_state); 10300 #endif /* INVARIANTS */ 10301 adp->ad_state &= ~ATTACHED; 10302 adp->ad_state |= UNDONE; 10303 } 10304 /* 10305 * The on-disk inode cannot claim to be any larger than the last 10306 * fragment that has been written. Otherwise, the on-disk inode 10307 * might have fragments that were not the last block in the file 10308 * which would corrupt the filesystem. 10309 */ 10310 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10311 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10312 if (adp->ad_offset >= UFS_NDADDR) 10313 break; 10314 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10315 /* keep going until hitting a rollback to a frag */ 10316 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10317 continue; 10318 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10319 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10320 #ifdef INVARIANTS 10321 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10322 panic("initiate_write_inodeblock_ufs1: " 10323 "lost dep1"); 10324 #endif /* INVARIANTS */ 10325 dp->di_db[i] = 0; 10326 } 10327 for (i = 0; i < UFS_NIADDR; i++) { 10328 #ifdef INVARIANTS 10329 if (dp->di_ib[i] != 0 && 10330 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10331 panic("initiate_write_inodeblock_ufs1: " 10332 "lost dep2"); 10333 #endif /* INVARIANTS */ 10334 dp->di_ib[i] = 0; 10335 } 10336 return; 10337 } 10338 /* 10339 * If we have zero'ed out the last allocated block of the file, 10340 * roll back the size to the last currently allocated block. 10341 * We know that this last allocated block is a full-sized as 10342 * we already checked for fragments in the loop above. 10343 */ 10344 if (lastadp != NULL && 10345 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10346 for (i = lastadp->ad_offset; i >= 0; i--) 10347 if (dp->di_db[i] != 0) 10348 break; 10349 dp->di_size = (i + 1) * fs->fs_bsize; 10350 } 10351 /* 10352 * The only dependencies are for indirect blocks. 10353 * 10354 * The file size for indirect block additions is not guaranteed. 10355 * Such a guarantee would be non-trivial to achieve. The conventional 10356 * synchronous write implementation also does not make this guarantee. 10357 * Fsck should catch and fix discrepancies. Arguably, the file size 10358 * can be over-estimated without destroying integrity when the file 10359 * moves into the indirect blocks (i.e., is large). If we want to 10360 * postpone fsck, we are stuck with this argument. 10361 */ 10362 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10363 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10364 } 10365 10366 /* 10367 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10368 * Note that any bug fixes made to this routine must be done in the 10369 * version found above. 10370 * 10371 * Called from within the procedure above to deal with unsatisfied 10372 * allocation dependencies in an inodeblock. The buffer must be 10373 * locked, thus, no I/O completion operations can occur while we 10374 * are manipulating its associated dependencies. 10375 */ 10376 static void 10377 initiate_write_inodeblock_ufs2(inodedep, bp) 10378 struct inodedep *inodedep; 10379 struct buf *bp; /* The inode block */ 10380 { 10381 struct allocdirect *adp, *lastadp; 10382 struct ufs2_dinode *dp; 10383 struct ufs2_dinode *sip; 10384 struct inoref *inoref; 10385 struct ufsmount *ump; 10386 struct fs *fs; 10387 ufs_lbn_t i; 10388 #ifdef INVARIANTS 10389 ufs_lbn_t prevlbn = 0; 10390 #endif 10391 int deplist; 10392 10393 if (inodedep->id_state & IOSTARTED) 10394 panic("initiate_write_inodeblock_ufs2: already started"); 10395 inodedep->id_state |= IOSTARTED; 10396 fs = inodedep->id_fs; 10397 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10398 LOCK_OWNED(ump); 10399 dp = (struct ufs2_dinode *)bp->b_data + 10400 ino_to_fsbo(fs, inodedep->id_ino); 10401 10402 /* 10403 * If we're on the unlinked list but have not yet written our 10404 * next pointer initialize it here. 10405 */ 10406 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10407 struct inodedep *inon; 10408 10409 inon = TAILQ_NEXT(inodedep, id_unlinked); 10410 dp->di_freelink = inon ? inon->id_ino : 0; 10411 ffs_update_dinode_ckhash(fs, dp); 10412 } 10413 /* 10414 * If the bitmap is not yet written, then the allocated 10415 * inode cannot be written to disk. 10416 */ 10417 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10418 if (inodedep->id_savedino2 != NULL) 10419 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10420 FREE_LOCK(ump); 10421 sip = malloc(sizeof(struct ufs2_dinode), 10422 M_SAVEDINO, M_SOFTDEP_FLAGS); 10423 ACQUIRE_LOCK(ump); 10424 inodedep->id_savedino2 = sip; 10425 *inodedep->id_savedino2 = *dp; 10426 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10427 dp->di_gen = inodedep->id_savedino2->di_gen; 10428 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10429 return; 10430 } 10431 /* 10432 * If no dependencies, then there is nothing to roll back. 10433 */ 10434 inodedep->id_savedsize = dp->di_size; 10435 inodedep->id_savedextsize = dp->di_extsize; 10436 inodedep->id_savednlink = dp->di_nlink; 10437 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10438 TAILQ_EMPTY(&inodedep->id_extupdt) && 10439 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10440 return; 10441 /* 10442 * Revert the link count to that of the first unwritten journal entry. 10443 */ 10444 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10445 if (inoref) 10446 dp->di_nlink = inoref->if_nlink; 10447 10448 /* 10449 * Set the ext data dependencies to busy. 10450 */ 10451 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10452 adp = TAILQ_NEXT(adp, ad_next)) { 10453 #ifdef INVARIANTS 10454 if (deplist != 0 && prevlbn >= adp->ad_offset) 10455 panic("initiate_write_inodeblock_ufs2: lbn order"); 10456 prevlbn = adp->ad_offset; 10457 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10458 panic("initiate_write_inodeblock_ufs2: " 10459 "ext pointer #%jd mismatch %jd != %jd", 10460 (intmax_t)adp->ad_offset, 10461 (intmax_t)dp->di_extb[adp->ad_offset], 10462 (intmax_t)adp->ad_newblkno); 10463 deplist |= 1 << adp->ad_offset; 10464 if ((adp->ad_state & ATTACHED) == 0) 10465 panic("initiate_write_inodeblock_ufs2: Unknown " 10466 "state 0x%x", adp->ad_state); 10467 #endif /* INVARIANTS */ 10468 adp->ad_state &= ~ATTACHED; 10469 adp->ad_state |= UNDONE; 10470 } 10471 /* 10472 * The on-disk inode cannot claim to be any larger than the last 10473 * fragment that has been written. Otherwise, the on-disk inode 10474 * might have fragments that were not the last block in the ext 10475 * data which would corrupt the filesystem. 10476 */ 10477 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10478 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10479 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10480 /* keep going until hitting a rollback to a frag */ 10481 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10482 continue; 10483 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10484 for (i = adp->ad_offset + 1; i < UFS_NXADDR; i++) { 10485 #ifdef INVARIANTS 10486 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10487 panic("initiate_write_inodeblock_ufs2: " 10488 "lost dep1"); 10489 #endif /* INVARIANTS */ 10490 dp->di_extb[i] = 0; 10491 } 10492 lastadp = NULL; 10493 break; 10494 } 10495 /* 10496 * If we have zero'ed out the last allocated block of the ext 10497 * data, roll back the size to the last currently allocated block. 10498 * We know that this last allocated block is a full-sized as 10499 * we already checked for fragments in the loop above. 10500 */ 10501 if (lastadp != NULL && 10502 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10503 for (i = lastadp->ad_offset; i >= 0; i--) 10504 if (dp->di_extb[i] != 0) 10505 break; 10506 dp->di_extsize = (i + 1) * fs->fs_bsize; 10507 } 10508 /* 10509 * Set the file data dependencies to busy. 10510 */ 10511 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10512 adp = TAILQ_NEXT(adp, ad_next)) { 10513 #ifdef INVARIANTS 10514 if (deplist != 0 && prevlbn >= adp->ad_offset) 10515 panic("softdep_write_inodeblock: lbn order"); 10516 if ((adp->ad_state & ATTACHED) == 0) 10517 panic("inodedep %p and adp %p not attached", inodedep, adp); 10518 prevlbn = adp->ad_offset; 10519 if (adp->ad_offset < UFS_NDADDR && 10520 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10521 panic("initiate_write_inodeblock_ufs2: " 10522 "direct pointer #%jd mismatch %jd != %jd", 10523 (intmax_t)adp->ad_offset, 10524 (intmax_t)dp->di_db[adp->ad_offset], 10525 (intmax_t)adp->ad_newblkno); 10526 if (adp->ad_offset >= UFS_NDADDR && 10527 dp->di_ib[adp->ad_offset - UFS_NDADDR] != adp->ad_newblkno) 10528 panic("initiate_write_inodeblock_ufs2: " 10529 "indirect pointer #%jd mismatch %jd != %jd", 10530 (intmax_t)adp->ad_offset - UFS_NDADDR, 10531 (intmax_t)dp->di_ib[adp->ad_offset - UFS_NDADDR], 10532 (intmax_t)adp->ad_newblkno); 10533 deplist |= 1 << adp->ad_offset; 10534 if ((adp->ad_state & ATTACHED) == 0) 10535 panic("initiate_write_inodeblock_ufs2: Unknown " 10536 "state 0x%x", adp->ad_state); 10537 #endif /* INVARIANTS */ 10538 adp->ad_state &= ~ATTACHED; 10539 adp->ad_state |= UNDONE; 10540 } 10541 /* 10542 * The on-disk inode cannot claim to be any larger than the last 10543 * fragment that has been written. Otherwise, the on-disk inode 10544 * might have fragments that were not the last block in the file 10545 * which would corrupt the filesystem. 10546 */ 10547 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10548 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10549 if (adp->ad_offset >= UFS_NDADDR) 10550 break; 10551 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10552 /* keep going until hitting a rollback to a frag */ 10553 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10554 continue; 10555 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10556 for (i = adp->ad_offset + 1; i < UFS_NDADDR; i++) { 10557 #ifdef INVARIANTS 10558 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10559 panic("initiate_write_inodeblock_ufs2: " 10560 "lost dep2"); 10561 #endif /* INVARIANTS */ 10562 dp->di_db[i] = 0; 10563 } 10564 for (i = 0; i < UFS_NIADDR; i++) { 10565 #ifdef INVARIANTS 10566 if (dp->di_ib[i] != 0 && 10567 (deplist & ((1 << UFS_NDADDR) << i)) == 0) 10568 panic("initiate_write_inodeblock_ufs2: " 10569 "lost dep3"); 10570 #endif /* INVARIANTS */ 10571 dp->di_ib[i] = 0; 10572 } 10573 ffs_update_dinode_ckhash(fs, dp); 10574 return; 10575 } 10576 /* 10577 * If we have zero'ed out the last allocated block of the file, 10578 * roll back the size to the last currently allocated block. 10579 * We know that this last allocated block is a full-sized as 10580 * we already checked for fragments in the loop above. 10581 */ 10582 if (lastadp != NULL && 10583 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10584 for (i = lastadp->ad_offset; i >= 0; i--) 10585 if (dp->di_db[i] != 0) 10586 break; 10587 dp->di_size = (i + 1) * fs->fs_bsize; 10588 } 10589 /* 10590 * The only dependencies are for indirect blocks. 10591 * 10592 * The file size for indirect block additions is not guaranteed. 10593 * Such a guarantee would be non-trivial to achieve. The conventional 10594 * synchronous write implementation also does not make this guarantee. 10595 * Fsck should catch and fix discrepancies. Arguably, the file size 10596 * can be over-estimated without destroying integrity when the file 10597 * moves into the indirect blocks (i.e., is large). If we want to 10598 * postpone fsck, we are stuck with this argument. 10599 */ 10600 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10601 dp->di_ib[adp->ad_offset - UFS_NDADDR] = 0; 10602 ffs_update_dinode_ckhash(fs, dp); 10603 } 10604 10605 /* 10606 * Cancel an indirdep as a result of truncation. Release all of the 10607 * children allocindirs and place their journal work on the appropriate 10608 * list. 10609 */ 10610 static void 10611 cancel_indirdep(indirdep, bp, freeblks) 10612 struct indirdep *indirdep; 10613 struct buf *bp; 10614 struct freeblks *freeblks; 10615 { 10616 struct allocindir *aip; 10617 10618 /* 10619 * None of the indirect pointers will ever be visible, 10620 * so they can simply be tossed. GOINGAWAY ensures 10621 * that allocated pointers will be saved in the buffer 10622 * cache until they are freed. Note that they will 10623 * only be able to be found by their physical address 10624 * since the inode mapping the logical address will 10625 * be gone. The save buffer used for the safe copy 10626 * was allocated in setup_allocindir_phase2 using 10627 * the physical address so it could be used for this 10628 * purpose. Hence we swap the safe copy with the real 10629 * copy, allowing the safe copy to be freed and holding 10630 * on to the real copy for later use in indir_trunc. 10631 */ 10632 if (indirdep->ir_state & GOINGAWAY) 10633 panic("cancel_indirdep: already gone"); 10634 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10635 indirdep->ir_state |= DEPCOMPLETE; 10636 LIST_REMOVE(indirdep, ir_next); 10637 } 10638 indirdep->ir_state |= GOINGAWAY; 10639 /* 10640 * Pass in bp for blocks still have journal writes 10641 * pending so we can cancel them on their own. 10642 */ 10643 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 10644 cancel_allocindir(aip, bp, freeblks, 0); 10645 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) 10646 cancel_allocindir(aip, NULL, freeblks, 0); 10647 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) 10648 cancel_allocindir(aip, NULL, freeblks, 0); 10649 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) 10650 cancel_allocindir(aip, NULL, freeblks, 0); 10651 /* 10652 * If there are pending partial truncations we need to keep the 10653 * old block copy around until they complete. This is because 10654 * the current b_data is not a perfect superset of the available 10655 * blocks. 10656 */ 10657 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10658 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10659 else 10660 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10661 WORKLIST_REMOVE(&indirdep->ir_list); 10662 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10663 indirdep->ir_bp = NULL; 10664 indirdep->ir_freeblks = freeblks; 10665 } 10666 10667 /* 10668 * Free an indirdep once it no longer has new pointers to track. 10669 */ 10670 static void 10671 free_indirdep(indirdep) 10672 struct indirdep *indirdep; 10673 { 10674 10675 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10676 ("free_indirdep: Indir trunc list not empty.")); 10677 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10678 ("free_indirdep: Complete head not empty.")); 10679 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10680 ("free_indirdep: write head not empty.")); 10681 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10682 ("free_indirdep: done head not empty.")); 10683 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10684 ("free_indirdep: deplist head not empty.")); 10685 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10686 ("free_indirdep: %p still on newblk list.", indirdep)); 10687 KASSERT(indirdep->ir_saveddata == NULL, 10688 ("free_indirdep: %p still has saved data.", indirdep)); 10689 if (indirdep->ir_state & ONWORKLIST) 10690 WORKLIST_REMOVE(&indirdep->ir_list); 10691 WORKITEM_FREE(indirdep, D_INDIRDEP); 10692 } 10693 10694 /* 10695 * Called before a write to an indirdep. This routine is responsible for 10696 * rolling back pointers to a safe state which includes only those 10697 * allocindirs which have been completed. 10698 */ 10699 static void 10700 initiate_write_indirdep(indirdep, bp) 10701 struct indirdep *indirdep; 10702 struct buf *bp; 10703 { 10704 struct ufsmount *ump; 10705 10706 indirdep->ir_state |= IOSTARTED; 10707 if (indirdep->ir_state & GOINGAWAY) 10708 panic("disk_io_initiation: indirdep gone"); 10709 /* 10710 * If there are no remaining dependencies, this will be writing 10711 * the real pointers. 10712 */ 10713 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10714 TAILQ_EMPTY(&indirdep->ir_trunc)) 10715 return; 10716 /* 10717 * Replace up-to-date version with safe version. 10718 */ 10719 if (indirdep->ir_saveddata == NULL) { 10720 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10721 LOCK_OWNED(ump); 10722 FREE_LOCK(ump); 10723 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10724 M_SOFTDEP_FLAGS); 10725 ACQUIRE_LOCK(ump); 10726 } 10727 indirdep->ir_state &= ~ATTACHED; 10728 indirdep->ir_state |= UNDONE; 10729 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10730 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10731 bp->b_bcount); 10732 } 10733 10734 /* 10735 * Called when an inode has been cleared in a cg bitmap. This finally 10736 * eliminates any canceled jaddrefs 10737 */ 10738 void 10739 softdep_setup_inofree(mp, bp, ino, wkhd) 10740 struct mount *mp; 10741 struct buf *bp; 10742 ino_t ino; 10743 struct workhead *wkhd; 10744 { 10745 struct worklist *wk, *wkn; 10746 struct inodedep *inodedep; 10747 struct ufsmount *ump; 10748 uint8_t *inosused; 10749 struct cg *cgp; 10750 struct fs *fs; 10751 10752 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10753 ("softdep_setup_inofree called on non-softdep filesystem")); 10754 ump = VFSTOUFS(mp); 10755 ACQUIRE_LOCK(ump); 10756 fs = ump->um_fs; 10757 cgp = (struct cg *)bp->b_data; 10758 inosused = cg_inosused(cgp); 10759 if (isset(inosused, ino % fs->fs_ipg)) 10760 panic("softdep_setup_inofree: inode %ju not freed.", 10761 (uintmax_t)ino); 10762 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10763 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10764 (uintmax_t)ino, inodedep); 10765 if (wkhd) { 10766 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10767 if (wk->wk_type != D_JADDREF) 10768 continue; 10769 WORKLIST_REMOVE(wk); 10770 /* 10771 * We can free immediately even if the jaddref 10772 * isn't attached in a background write as now 10773 * the bitmaps are reconciled. 10774 */ 10775 wk->wk_state |= COMPLETE | ATTACHED; 10776 free_jaddref(WK_JADDREF(wk)); 10777 } 10778 jwork_move(&bp->b_dep, wkhd); 10779 } 10780 FREE_LOCK(ump); 10781 } 10782 10783 10784 /* 10785 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10786 * map. Any dependencies waiting for the write to clear are added to the 10787 * buf's list and any jnewblks that are being canceled are discarded 10788 * immediately. 10789 */ 10790 void 10791 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10792 struct mount *mp; 10793 struct buf *bp; 10794 ufs2_daddr_t blkno; 10795 int frags; 10796 struct workhead *wkhd; 10797 { 10798 struct bmsafemap *bmsafemap; 10799 struct jnewblk *jnewblk; 10800 struct ufsmount *ump; 10801 struct worklist *wk; 10802 struct fs *fs; 10803 #ifdef INVARIANTS 10804 uint8_t *blksfree; 10805 struct cg *cgp; 10806 ufs2_daddr_t jstart; 10807 ufs2_daddr_t jend; 10808 ufs2_daddr_t end; 10809 long bno; 10810 int i; 10811 #endif 10812 10813 CTR3(KTR_SUJ, 10814 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10815 blkno, frags, wkhd); 10816 10817 ump = VFSTOUFS(mp); 10818 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10819 ("softdep_setup_blkfree called on non-softdep filesystem")); 10820 ACQUIRE_LOCK(ump); 10821 /* Lookup the bmsafemap so we track when it is dirty. */ 10822 fs = ump->um_fs; 10823 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10824 /* 10825 * Detach any jnewblks which have been canceled. They must linger 10826 * until the bitmap is cleared again by ffs_blkfree() to prevent 10827 * an unjournaled allocation from hitting the disk. 10828 */ 10829 if (wkhd) { 10830 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10831 CTR2(KTR_SUJ, 10832 "softdep_setup_blkfree: blkno %jd wk type %d", 10833 blkno, wk->wk_type); 10834 WORKLIST_REMOVE(wk); 10835 if (wk->wk_type != D_JNEWBLK) { 10836 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10837 continue; 10838 } 10839 jnewblk = WK_JNEWBLK(wk); 10840 KASSERT(jnewblk->jn_state & GOINGAWAY, 10841 ("softdep_setup_blkfree: jnewblk not canceled.")); 10842 #ifdef INVARIANTS 10843 /* 10844 * Assert that this block is free in the bitmap 10845 * before we discard the jnewblk. 10846 */ 10847 cgp = (struct cg *)bp->b_data; 10848 blksfree = cg_blksfree(cgp); 10849 bno = dtogd(fs, jnewblk->jn_blkno); 10850 for (i = jnewblk->jn_oldfrags; 10851 i < jnewblk->jn_frags; i++) { 10852 if (isset(blksfree, bno + i)) 10853 continue; 10854 panic("softdep_setup_blkfree: not free"); 10855 } 10856 #endif 10857 /* 10858 * Even if it's not attached we can free immediately 10859 * as the new bitmap is correct. 10860 */ 10861 wk->wk_state |= COMPLETE | ATTACHED; 10862 free_jnewblk(jnewblk); 10863 } 10864 } 10865 10866 #ifdef INVARIANTS 10867 /* 10868 * Assert that we are not freeing a block which has an outstanding 10869 * allocation dependency. 10870 */ 10871 fs = VFSTOUFS(mp)->um_fs; 10872 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10873 end = blkno + frags; 10874 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10875 /* 10876 * Don't match against blocks that will be freed when the 10877 * background write is done. 10878 */ 10879 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10880 (COMPLETE | DEPCOMPLETE)) 10881 continue; 10882 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10883 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10884 if ((blkno >= jstart && blkno < jend) || 10885 (end > jstart && end <= jend)) { 10886 printf("state 0x%X %jd - %d %d dep %p\n", 10887 jnewblk->jn_state, jnewblk->jn_blkno, 10888 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10889 jnewblk->jn_dep); 10890 panic("softdep_setup_blkfree: " 10891 "%jd-%jd(%d) overlaps with %jd-%jd", 10892 blkno, end, frags, jstart, jend); 10893 } 10894 } 10895 #endif 10896 FREE_LOCK(ump); 10897 } 10898 10899 /* 10900 * Revert a block allocation when the journal record that describes it 10901 * is not yet written. 10902 */ 10903 static int 10904 jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10905 struct jnewblk *jnewblk; 10906 struct fs *fs; 10907 struct cg *cgp; 10908 uint8_t *blksfree; 10909 { 10910 ufs1_daddr_t fragno; 10911 long cgbno, bbase; 10912 int frags, blk; 10913 int i; 10914 10915 frags = 0; 10916 cgbno = dtogd(fs, jnewblk->jn_blkno); 10917 /* 10918 * We have to test which frags need to be rolled back. We may 10919 * be operating on a stale copy when doing background writes. 10920 */ 10921 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10922 if (isclr(blksfree, cgbno + i)) 10923 frags++; 10924 if (frags == 0) 10925 return (0); 10926 /* 10927 * This is mostly ffs_blkfree() sans some validation and 10928 * superblock updates. 10929 */ 10930 if (frags == fs->fs_frag) { 10931 fragno = fragstoblks(fs, cgbno); 10932 ffs_setblock(fs, blksfree, fragno); 10933 ffs_clusteracct(fs, cgp, fragno, 1); 10934 cgp->cg_cs.cs_nbfree++; 10935 } else { 10936 cgbno += jnewblk->jn_oldfrags; 10937 bbase = cgbno - fragnum(fs, cgbno); 10938 /* Decrement the old frags. */ 10939 blk = blkmap(fs, blksfree, bbase); 10940 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10941 /* Deallocate the fragment */ 10942 for (i = 0; i < frags; i++) 10943 setbit(blksfree, cgbno + i); 10944 cgp->cg_cs.cs_nffree += frags; 10945 /* Add back in counts associated with the new frags */ 10946 blk = blkmap(fs, blksfree, bbase); 10947 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10948 /* If a complete block has been reassembled, account for it. */ 10949 fragno = fragstoblks(fs, bbase); 10950 if (ffs_isblock(fs, blksfree, fragno)) { 10951 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10952 ffs_clusteracct(fs, cgp, fragno, 1); 10953 cgp->cg_cs.cs_nbfree++; 10954 } 10955 } 10956 stat_jnewblk++; 10957 jnewblk->jn_state &= ~ATTACHED; 10958 jnewblk->jn_state |= UNDONE; 10959 10960 return (frags); 10961 } 10962 10963 static void 10964 initiate_write_bmsafemap(bmsafemap, bp) 10965 struct bmsafemap *bmsafemap; 10966 struct buf *bp; /* The cg block. */ 10967 { 10968 struct jaddref *jaddref; 10969 struct jnewblk *jnewblk; 10970 uint8_t *inosused; 10971 uint8_t *blksfree; 10972 struct cg *cgp; 10973 struct fs *fs; 10974 ino_t ino; 10975 10976 /* 10977 * If this is a background write, we did this at the time that 10978 * the copy was made, so do not need to do it again. 10979 */ 10980 if (bmsafemap->sm_state & IOSTARTED) 10981 return; 10982 bmsafemap->sm_state |= IOSTARTED; 10983 /* 10984 * Clear any inode allocations which are pending journal writes. 10985 */ 10986 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10987 cgp = (struct cg *)bp->b_data; 10988 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10989 inosused = cg_inosused(cgp); 10990 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10991 ino = jaddref->ja_ino % fs->fs_ipg; 10992 if (isset(inosused, ino)) { 10993 if ((jaddref->ja_mode & IFMT) == IFDIR) 10994 cgp->cg_cs.cs_ndir--; 10995 cgp->cg_cs.cs_nifree++; 10996 clrbit(inosused, ino); 10997 jaddref->ja_state &= ~ATTACHED; 10998 jaddref->ja_state |= UNDONE; 10999 stat_jaddref++; 11000 } else 11001 panic("initiate_write_bmsafemap: inode %ju " 11002 "marked free", (uintmax_t)jaddref->ja_ino); 11003 } 11004 } 11005 /* 11006 * Clear any block allocations which are pending journal writes. 11007 */ 11008 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11009 cgp = (struct cg *)bp->b_data; 11010 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11011 blksfree = cg_blksfree(cgp); 11012 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 11013 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 11014 continue; 11015 panic("initiate_write_bmsafemap: block %jd " 11016 "marked free", jnewblk->jn_blkno); 11017 } 11018 } 11019 /* 11020 * Move allocation lists to the written lists so they can be 11021 * cleared once the block write is complete. 11022 */ 11023 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 11024 inodedep, id_deps); 11025 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11026 newblk, nb_deps); 11027 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 11028 wk_list); 11029 } 11030 11031 /* 11032 * This routine is called during the completion interrupt 11033 * service routine for a disk write (from the procedure called 11034 * by the device driver to inform the filesystem caches of 11035 * a request completion). It should be called early in this 11036 * procedure, before the block is made available to other 11037 * processes or other routines are called. 11038 * 11039 */ 11040 static void 11041 softdep_disk_write_complete(bp) 11042 struct buf *bp; /* describes the completed disk write */ 11043 { 11044 struct worklist *wk; 11045 struct worklist *owk; 11046 struct ufsmount *ump; 11047 struct workhead reattach; 11048 struct freeblks *freeblks; 11049 struct buf *sbp; 11050 11051 ump = softdep_bp_to_mp(bp); 11052 KASSERT(LIST_EMPTY(&bp->b_dep) || ump != NULL, 11053 ("softdep_disk_write_complete: softdep_bp_to_mp returned NULL " 11054 "with outstanding dependencies for buffer %p", bp)); 11055 if (ump == NULL) 11056 return; 11057 /* 11058 * If an error occurred while doing the write, then the data 11059 * has not hit the disk and the dependencies cannot be processed. 11060 * But we do have to go through and roll forward any dependencies 11061 * that were rolled back before the disk write. 11062 */ 11063 sbp = NULL; 11064 ACQUIRE_LOCK(ump); 11065 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) { 11066 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11067 switch (wk->wk_type) { 11068 11069 case D_PAGEDEP: 11070 handle_written_filepage(WK_PAGEDEP(wk), bp, 0); 11071 continue; 11072 11073 case D_INODEDEP: 11074 handle_written_inodeblock(WK_INODEDEP(wk), 11075 bp, 0); 11076 continue; 11077 11078 case D_BMSAFEMAP: 11079 handle_written_bmsafemap(WK_BMSAFEMAP(wk), 11080 bp, 0); 11081 continue; 11082 11083 case D_INDIRDEP: 11084 handle_written_indirdep(WK_INDIRDEP(wk), 11085 bp, &sbp, 0); 11086 continue; 11087 default: 11088 /* nothing to roll forward */ 11089 continue; 11090 } 11091 } 11092 FREE_LOCK(ump); 11093 if (sbp) 11094 brelse(sbp); 11095 return; 11096 } 11097 LIST_INIT(&reattach); 11098 11099 /* 11100 * Ump SU lock must not be released anywhere in this code segment. 11101 */ 11102 owk = NULL; 11103 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 11104 WORKLIST_REMOVE(wk); 11105 atomic_add_long(&dep_write[wk->wk_type], 1); 11106 if (wk == owk) 11107 panic("duplicate worklist: %p\n", wk); 11108 owk = wk; 11109 switch (wk->wk_type) { 11110 11111 case D_PAGEDEP: 11112 if (handle_written_filepage(WK_PAGEDEP(wk), bp, 11113 WRITESUCCEEDED)) 11114 WORKLIST_INSERT(&reattach, wk); 11115 continue; 11116 11117 case D_INODEDEP: 11118 if (handle_written_inodeblock(WK_INODEDEP(wk), bp, 11119 WRITESUCCEEDED)) 11120 WORKLIST_INSERT(&reattach, wk); 11121 continue; 11122 11123 case D_BMSAFEMAP: 11124 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp, 11125 WRITESUCCEEDED)) 11126 WORKLIST_INSERT(&reattach, wk); 11127 continue; 11128 11129 case D_MKDIR: 11130 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 11131 continue; 11132 11133 case D_ALLOCDIRECT: 11134 wk->wk_state |= COMPLETE; 11135 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 11136 continue; 11137 11138 case D_ALLOCINDIR: 11139 wk->wk_state |= COMPLETE; 11140 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 11141 continue; 11142 11143 case D_INDIRDEP: 11144 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp, 11145 WRITESUCCEEDED)) 11146 WORKLIST_INSERT(&reattach, wk); 11147 continue; 11148 11149 case D_FREEBLKS: 11150 wk->wk_state |= COMPLETE; 11151 freeblks = WK_FREEBLKS(wk); 11152 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 11153 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11154 add_to_worklist(wk, WK_NODELAY); 11155 continue; 11156 11157 case D_FREEWORK: 11158 handle_written_freework(WK_FREEWORK(wk)); 11159 break; 11160 11161 case D_JSEGDEP: 11162 free_jsegdep(WK_JSEGDEP(wk)); 11163 continue; 11164 11165 case D_JSEG: 11166 handle_written_jseg(WK_JSEG(wk), bp); 11167 continue; 11168 11169 case D_SBDEP: 11170 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11171 WORKLIST_INSERT(&reattach, wk); 11172 continue; 11173 11174 case D_FREEDEP: 11175 free_freedep(WK_FREEDEP(wk)); 11176 continue; 11177 11178 default: 11179 panic("handle_disk_write_complete: Unknown type %s", 11180 TYPENAME(wk->wk_type)); 11181 /* NOTREACHED */ 11182 } 11183 } 11184 /* 11185 * Reattach any requests that must be redone. 11186 */ 11187 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11188 WORKLIST_REMOVE(wk); 11189 WORKLIST_INSERT(&bp->b_dep, wk); 11190 } 11191 FREE_LOCK(ump); 11192 if (sbp) 11193 brelse(sbp); 11194 } 11195 11196 /* 11197 * Called from within softdep_disk_write_complete above. 11198 */ 11199 static void 11200 handle_allocdirect_partdone(adp, wkhd) 11201 struct allocdirect *adp; /* the completed allocdirect */ 11202 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11203 { 11204 struct allocdirectlst *listhead; 11205 struct allocdirect *listadp; 11206 struct inodedep *inodedep; 11207 long bsize; 11208 11209 LOCK_OWNED(VFSTOUFS(adp->ad_block.nb_list.wk_mp)); 11210 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11211 return; 11212 /* 11213 * The on-disk inode cannot claim to be any larger than the last 11214 * fragment that has been written. Otherwise, the on-disk inode 11215 * might have fragments that were not the last block in the file 11216 * which would corrupt the filesystem. Thus, we cannot free any 11217 * allocdirects after one whose ad_oldblkno claims a fragment as 11218 * these blocks must be rolled back to zero before writing the inode. 11219 * We check the currently active set of allocdirects in id_inoupdt 11220 * or id_extupdt as appropriate. 11221 */ 11222 inodedep = adp->ad_inodedep; 11223 bsize = inodedep->id_fs->fs_bsize; 11224 if (adp->ad_state & EXTDATA) 11225 listhead = &inodedep->id_extupdt; 11226 else 11227 listhead = &inodedep->id_inoupdt; 11228 TAILQ_FOREACH(listadp, listhead, ad_next) { 11229 /* found our block */ 11230 if (listadp == adp) 11231 break; 11232 /* continue if ad_oldlbn is not a fragment */ 11233 if (listadp->ad_oldsize == 0 || 11234 listadp->ad_oldsize == bsize) 11235 continue; 11236 /* hit a fragment */ 11237 return; 11238 } 11239 /* 11240 * If we have reached the end of the current list without 11241 * finding the just finished dependency, then it must be 11242 * on the future dependency list. Future dependencies cannot 11243 * be freed until they are moved to the current list. 11244 */ 11245 if (listadp == NULL) { 11246 #ifdef INVARIANTS 11247 if (adp->ad_state & EXTDATA) 11248 listhead = &inodedep->id_newextupdt; 11249 else 11250 listhead = &inodedep->id_newinoupdt; 11251 TAILQ_FOREACH(listadp, listhead, ad_next) 11252 /* found our block */ 11253 if (listadp == adp) 11254 break; 11255 if (listadp == NULL) 11256 panic("handle_allocdirect_partdone: lost dep"); 11257 #endif /* INVARIANTS */ 11258 return; 11259 } 11260 /* 11261 * If we have found the just finished dependency, then queue 11262 * it along with anything that follows it that is complete. 11263 * Since the pointer has not yet been written in the inode 11264 * as the dependency prevents it, place the allocdirect on the 11265 * bufwait list where it will be freed once the pointer is 11266 * valid. 11267 */ 11268 if (wkhd == NULL) 11269 wkhd = &inodedep->id_bufwait; 11270 for (; adp; adp = listadp) { 11271 listadp = TAILQ_NEXT(adp, ad_next); 11272 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11273 return; 11274 TAILQ_REMOVE(listhead, adp, ad_next); 11275 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11276 } 11277 } 11278 11279 /* 11280 * Called from within softdep_disk_write_complete above. This routine 11281 * completes successfully written allocindirs. 11282 */ 11283 static void 11284 handle_allocindir_partdone(aip) 11285 struct allocindir *aip; /* the completed allocindir */ 11286 { 11287 struct indirdep *indirdep; 11288 11289 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11290 return; 11291 indirdep = aip->ai_indirdep; 11292 LIST_REMOVE(aip, ai_next); 11293 /* 11294 * Don't set a pointer while the buffer is undergoing IO or while 11295 * we have active truncations. 11296 */ 11297 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11298 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11299 return; 11300 } 11301 if (indirdep->ir_state & UFS1FMT) 11302 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11303 aip->ai_newblkno; 11304 else 11305 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11306 aip->ai_newblkno; 11307 /* 11308 * Await the pointer write before freeing the allocindir. 11309 */ 11310 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11311 } 11312 11313 /* 11314 * Release segments held on a jwork list. 11315 */ 11316 static void 11317 handle_jwork(wkhd) 11318 struct workhead *wkhd; 11319 { 11320 struct worklist *wk; 11321 11322 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11323 WORKLIST_REMOVE(wk); 11324 switch (wk->wk_type) { 11325 case D_JSEGDEP: 11326 free_jsegdep(WK_JSEGDEP(wk)); 11327 continue; 11328 case D_FREEDEP: 11329 free_freedep(WK_FREEDEP(wk)); 11330 continue; 11331 case D_FREEFRAG: 11332 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11333 WORKITEM_FREE(wk, D_FREEFRAG); 11334 continue; 11335 case D_FREEWORK: 11336 handle_written_freework(WK_FREEWORK(wk)); 11337 continue; 11338 default: 11339 panic("handle_jwork: Unknown type %s\n", 11340 TYPENAME(wk->wk_type)); 11341 } 11342 } 11343 } 11344 11345 /* 11346 * Handle the bufwait list on an inode when it is safe to release items 11347 * held there. This normally happens after an inode block is written but 11348 * may be delayed and handled later if there are pending journal items that 11349 * are not yet safe to be released. 11350 */ 11351 static struct freefile * 11352 handle_bufwait(inodedep, refhd) 11353 struct inodedep *inodedep; 11354 struct workhead *refhd; 11355 { 11356 struct jaddref *jaddref; 11357 struct freefile *freefile; 11358 struct worklist *wk; 11359 11360 freefile = NULL; 11361 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11362 WORKLIST_REMOVE(wk); 11363 switch (wk->wk_type) { 11364 case D_FREEFILE: 11365 /* 11366 * We defer adding freefile to the worklist 11367 * until all other additions have been made to 11368 * ensure that it will be done after all the 11369 * old blocks have been freed. 11370 */ 11371 if (freefile != NULL) 11372 panic("handle_bufwait: freefile"); 11373 freefile = WK_FREEFILE(wk); 11374 continue; 11375 11376 case D_MKDIR: 11377 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11378 continue; 11379 11380 case D_DIRADD: 11381 diradd_inode_written(WK_DIRADD(wk), inodedep); 11382 continue; 11383 11384 case D_FREEFRAG: 11385 wk->wk_state |= COMPLETE; 11386 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11387 add_to_worklist(wk, 0); 11388 continue; 11389 11390 case D_DIRREM: 11391 wk->wk_state |= COMPLETE; 11392 add_to_worklist(wk, 0); 11393 continue; 11394 11395 case D_ALLOCDIRECT: 11396 case D_ALLOCINDIR: 11397 free_newblk(WK_NEWBLK(wk)); 11398 continue; 11399 11400 case D_JNEWBLK: 11401 wk->wk_state |= COMPLETE; 11402 free_jnewblk(WK_JNEWBLK(wk)); 11403 continue; 11404 11405 /* 11406 * Save freed journal segments and add references on 11407 * the supplied list which will delay their release 11408 * until the cg bitmap is cleared on disk. 11409 */ 11410 case D_JSEGDEP: 11411 if (refhd == NULL) 11412 free_jsegdep(WK_JSEGDEP(wk)); 11413 else 11414 WORKLIST_INSERT(refhd, wk); 11415 continue; 11416 11417 case D_JADDREF: 11418 jaddref = WK_JADDREF(wk); 11419 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11420 if_deps); 11421 /* 11422 * Transfer any jaddrefs to the list to be freed with 11423 * the bitmap if we're handling a removed file. 11424 */ 11425 if (refhd == NULL) { 11426 wk->wk_state |= COMPLETE; 11427 free_jaddref(jaddref); 11428 } else 11429 WORKLIST_INSERT(refhd, wk); 11430 continue; 11431 11432 default: 11433 panic("handle_bufwait: Unknown type %p(%s)", 11434 wk, TYPENAME(wk->wk_type)); 11435 /* NOTREACHED */ 11436 } 11437 } 11438 return (freefile); 11439 } 11440 /* 11441 * Called from within softdep_disk_write_complete above to restore 11442 * in-memory inode block contents to their most up-to-date state. Note 11443 * that this routine is always called from interrupt level with further 11444 * interrupts from this device blocked. 11445 * 11446 * If the write did not succeed, we will do all the roll-forward 11447 * operations, but we will not take the actions that will allow its 11448 * dependencies to be processed. 11449 */ 11450 static int 11451 handle_written_inodeblock(inodedep, bp, flags) 11452 struct inodedep *inodedep; 11453 struct buf *bp; /* buffer containing the inode block */ 11454 int flags; 11455 { 11456 struct freefile *freefile; 11457 struct allocdirect *adp, *nextadp; 11458 struct ufs1_dinode *dp1 = NULL; 11459 struct ufs2_dinode *dp2 = NULL; 11460 struct workhead wkhd; 11461 int hadchanges, fstype; 11462 ino_t freelink; 11463 11464 LIST_INIT(&wkhd); 11465 hadchanges = 0; 11466 freefile = NULL; 11467 if ((inodedep->id_state & IOSTARTED) == 0) 11468 panic("handle_written_inodeblock: not started"); 11469 inodedep->id_state &= ~IOSTARTED; 11470 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11471 fstype = UFS1; 11472 dp1 = (struct ufs1_dinode *)bp->b_data + 11473 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11474 freelink = dp1->di_freelink; 11475 } else { 11476 fstype = UFS2; 11477 dp2 = (struct ufs2_dinode *)bp->b_data + 11478 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11479 freelink = dp2->di_freelink; 11480 } 11481 /* 11482 * Leave this inodeblock dirty until it's in the list. 11483 */ 11484 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED && 11485 (flags & WRITESUCCEEDED)) { 11486 struct inodedep *inon; 11487 11488 inon = TAILQ_NEXT(inodedep, id_unlinked); 11489 if ((inon == NULL && freelink == 0) || 11490 (inon && inon->id_ino == freelink)) { 11491 if (inon) 11492 inon->id_state |= UNLINKPREV; 11493 inodedep->id_state |= UNLINKNEXT; 11494 } 11495 hadchanges = 1; 11496 } 11497 /* 11498 * If we had to rollback the inode allocation because of 11499 * bitmaps being incomplete, then simply restore it. 11500 * Keep the block dirty so that it will not be reclaimed until 11501 * all associated dependencies have been cleared and the 11502 * corresponding updates written to disk. 11503 */ 11504 if (inodedep->id_savedino1 != NULL) { 11505 hadchanges = 1; 11506 if (fstype == UFS1) 11507 *dp1 = *inodedep->id_savedino1; 11508 else 11509 *dp2 = *inodedep->id_savedino2; 11510 free(inodedep->id_savedino1, M_SAVEDINO); 11511 inodedep->id_savedino1 = NULL; 11512 if ((bp->b_flags & B_DELWRI) == 0) 11513 stat_inode_bitmap++; 11514 bdirty(bp); 11515 /* 11516 * If the inode is clear here and GOINGAWAY it will never 11517 * be written. Process the bufwait and clear any pending 11518 * work which may include the freefile. 11519 */ 11520 if (inodedep->id_state & GOINGAWAY) 11521 goto bufwait; 11522 return (1); 11523 } 11524 if (flags & WRITESUCCEEDED) 11525 inodedep->id_state |= COMPLETE; 11526 /* 11527 * Roll forward anything that had to be rolled back before 11528 * the inode could be updated. 11529 */ 11530 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11531 nextadp = TAILQ_NEXT(adp, ad_next); 11532 if (adp->ad_state & ATTACHED) 11533 panic("handle_written_inodeblock: new entry"); 11534 if (fstype == UFS1) { 11535 if (adp->ad_offset < UFS_NDADDR) { 11536 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11537 panic("%s %s #%jd mismatch %d != %jd", 11538 "handle_written_inodeblock:", 11539 "direct pointer", 11540 (intmax_t)adp->ad_offset, 11541 dp1->di_db[adp->ad_offset], 11542 (intmax_t)adp->ad_oldblkno); 11543 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11544 } else { 11545 if (dp1->di_ib[adp->ad_offset - UFS_NDADDR] != 11546 0) 11547 panic("%s: %s #%jd allocated as %d", 11548 "handle_written_inodeblock", 11549 "indirect pointer", 11550 (intmax_t)adp->ad_offset - 11551 UFS_NDADDR, 11552 dp1->di_ib[adp->ad_offset - 11553 UFS_NDADDR]); 11554 dp1->di_ib[adp->ad_offset - UFS_NDADDR] = 11555 adp->ad_newblkno; 11556 } 11557 } else { 11558 if (adp->ad_offset < UFS_NDADDR) { 11559 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11560 panic("%s: %s #%jd %s %jd != %jd", 11561 "handle_written_inodeblock", 11562 "direct pointer", 11563 (intmax_t)adp->ad_offset, "mismatch", 11564 (intmax_t)dp2->di_db[adp->ad_offset], 11565 (intmax_t)adp->ad_oldblkno); 11566 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11567 } else { 11568 if (dp2->di_ib[adp->ad_offset - UFS_NDADDR] != 11569 0) 11570 panic("%s: %s #%jd allocated as %jd", 11571 "handle_written_inodeblock", 11572 "indirect pointer", 11573 (intmax_t)adp->ad_offset - 11574 UFS_NDADDR, 11575 (intmax_t) 11576 dp2->di_ib[adp->ad_offset - 11577 UFS_NDADDR]); 11578 dp2->di_ib[adp->ad_offset - UFS_NDADDR] = 11579 adp->ad_newblkno; 11580 } 11581 } 11582 adp->ad_state &= ~UNDONE; 11583 adp->ad_state |= ATTACHED; 11584 hadchanges = 1; 11585 } 11586 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11587 nextadp = TAILQ_NEXT(adp, ad_next); 11588 if (adp->ad_state & ATTACHED) 11589 panic("handle_written_inodeblock: new entry"); 11590 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11591 panic("%s: direct pointers #%jd %s %jd != %jd", 11592 "handle_written_inodeblock", 11593 (intmax_t)adp->ad_offset, "mismatch", 11594 (intmax_t)dp2->di_extb[adp->ad_offset], 11595 (intmax_t)adp->ad_oldblkno); 11596 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11597 adp->ad_state &= ~UNDONE; 11598 adp->ad_state |= ATTACHED; 11599 hadchanges = 1; 11600 } 11601 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11602 stat_direct_blk_ptrs++; 11603 /* 11604 * Reset the file size to its most up-to-date value. 11605 */ 11606 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11607 panic("handle_written_inodeblock: bad size"); 11608 if (inodedep->id_savednlink > UFS_LINK_MAX) 11609 panic("handle_written_inodeblock: Invalid link count " 11610 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink, 11611 inodedep); 11612 if (fstype == UFS1) { 11613 if (dp1->di_nlink != inodedep->id_savednlink) { 11614 dp1->di_nlink = inodedep->id_savednlink; 11615 hadchanges = 1; 11616 } 11617 if (dp1->di_size != inodedep->id_savedsize) { 11618 dp1->di_size = inodedep->id_savedsize; 11619 hadchanges = 1; 11620 } 11621 } else { 11622 if (dp2->di_nlink != inodedep->id_savednlink) { 11623 dp2->di_nlink = inodedep->id_savednlink; 11624 hadchanges = 1; 11625 } 11626 if (dp2->di_size != inodedep->id_savedsize) { 11627 dp2->di_size = inodedep->id_savedsize; 11628 hadchanges = 1; 11629 } 11630 if (dp2->di_extsize != inodedep->id_savedextsize) { 11631 dp2->di_extsize = inodedep->id_savedextsize; 11632 hadchanges = 1; 11633 } 11634 } 11635 inodedep->id_savedsize = -1; 11636 inodedep->id_savedextsize = -1; 11637 inodedep->id_savednlink = -1; 11638 /* 11639 * If there were any rollbacks in the inode block, then it must be 11640 * marked dirty so that its will eventually get written back in 11641 * its correct form. 11642 */ 11643 if (hadchanges) { 11644 if (fstype == UFS2) 11645 ffs_update_dinode_ckhash(inodedep->id_fs, dp2); 11646 bdirty(bp); 11647 } 11648 bufwait: 11649 /* 11650 * If the write did not succeed, we have done all the roll-forward 11651 * operations, but we cannot take the actions that will allow its 11652 * dependencies to be processed. 11653 */ 11654 if ((flags & WRITESUCCEEDED) == 0) 11655 return (hadchanges); 11656 /* 11657 * Process any allocdirects that completed during the update. 11658 */ 11659 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11660 handle_allocdirect_partdone(adp, &wkhd); 11661 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11662 handle_allocdirect_partdone(adp, &wkhd); 11663 /* 11664 * Process deallocations that were held pending until the 11665 * inode had been written to disk. Freeing of the inode 11666 * is delayed until after all blocks have been freed to 11667 * avoid creation of new <vfsid, inum, lbn> triples 11668 * before the old ones have been deleted. Completely 11669 * unlinked inodes are not processed until the unlinked 11670 * inode list is written or the last reference is removed. 11671 */ 11672 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11673 freefile = handle_bufwait(inodedep, NULL); 11674 if (freefile && !LIST_EMPTY(&wkhd)) { 11675 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11676 freefile = NULL; 11677 } 11678 } 11679 /* 11680 * Move rolled forward dependency completions to the bufwait list 11681 * now that those that were already written have been processed. 11682 */ 11683 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11684 panic("handle_written_inodeblock: bufwait but no changes"); 11685 jwork_move(&inodedep->id_bufwait, &wkhd); 11686 11687 if (freefile != NULL) { 11688 /* 11689 * If the inode is goingaway it was never written. Fake up 11690 * the state here so free_inodedep() can succeed. 11691 */ 11692 if (inodedep->id_state & GOINGAWAY) 11693 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11694 if (free_inodedep(inodedep) == 0) 11695 panic("handle_written_inodeblock: live inodedep %p", 11696 inodedep); 11697 add_to_worklist(&freefile->fx_list, 0); 11698 return (0); 11699 } 11700 11701 /* 11702 * If no outstanding dependencies, free it. 11703 */ 11704 if (free_inodedep(inodedep) || 11705 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11706 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11707 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11708 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11709 return (0); 11710 return (hadchanges); 11711 } 11712 11713 /* 11714 * Perform needed roll-forwards and kick off any dependencies that 11715 * can now be processed. 11716 * 11717 * If the write did not succeed, we will do all the roll-forward 11718 * operations, but we will not take the actions that will allow its 11719 * dependencies to be processed. 11720 */ 11721 static int 11722 handle_written_indirdep(indirdep, bp, bpp, flags) 11723 struct indirdep *indirdep; 11724 struct buf *bp; 11725 struct buf **bpp; 11726 int flags; 11727 { 11728 struct allocindir *aip; 11729 struct buf *sbp; 11730 int chgs; 11731 11732 if (indirdep->ir_state & GOINGAWAY) 11733 panic("handle_written_indirdep: indirdep gone"); 11734 if ((indirdep->ir_state & IOSTARTED) == 0) 11735 panic("handle_written_indirdep: IO not started"); 11736 chgs = 0; 11737 /* 11738 * If there were rollbacks revert them here. 11739 */ 11740 if (indirdep->ir_saveddata) { 11741 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11742 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11743 free(indirdep->ir_saveddata, M_INDIRDEP); 11744 indirdep->ir_saveddata = NULL; 11745 } 11746 chgs = 1; 11747 } 11748 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11749 indirdep->ir_state |= ATTACHED; 11750 /* 11751 * If the write did not succeed, we have done all the roll-forward 11752 * operations, but we cannot take the actions that will allow its 11753 * dependencies to be processed. 11754 */ 11755 if ((flags & WRITESUCCEEDED) == 0) { 11756 stat_indir_blk_ptrs++; 11757 bdirty(bp); 11758 return (1); 11759 } 11760 /* 11761 * Move allocindirs with written pointers to the completehd if 11762 * the indirdep's pointer is not yet written. Otherwise 11763 * free them here. 11764 */ 11765 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) { 11766 LIST_REMOVE(aip, ai_next); 11767 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11768 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11769 ai_next); 11770 newblk_freefrag(&aip->ai_block); 11771 continue; 11772 } 11773 free_newblk(&aip->ai_block); 11774 } 11775 /* 11776 * Move allocindirs that have finished dependency processing from 11777 * the done list to the write list after updating the pointers. 11778 */ 11779 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11780 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 11781 handle_allocindir_partdone(aip); 11782 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11783 panic("disk_write_complete: not gone"); 11784 chgs = 1; 11785 } 11786 } 11787 /* 11788 * Preserve the indirdep if there were any changes or if it is not 11789 * yet valid on disk. 11790 */ 11791 if (chgs) { 11792 stat_indir_blk_ptrs++; 11793 bdirty(bp); 11794 return (1); 11795 } 11796 /* 11797 * If there were no changes we can discard the savedbp and detach 11798 * ourselves from the buf. We are only carrying completed pointers 11799 * in this case. 11800 */ 11801 sbp = indirdep->ir_savebp; 11802 sbp->b_flags |= B_INVAL | B_NOCACHE; 11803 indirdep->ir_savebp = NULL; 11804 indirdep->ir_bp = NULL; 11805 if (*bpp != NULL) 11806 panic("handle_written_indirdep: bp already exists."); 11807 *bpp = sbp; 11808 /* 11809 * The indirdep may not be freed until its parent points at it. 11810 */ 11811 if (indirdep->ir_state & DEPCOMPLETE) 11812 free_indirdep(indirdep); 11813 11814 return (0); 11815 } 11816 11817 /* 11818 * Process a diradd entry after its dependent inode has been written. 11819 */ 11820 static void 11821 diradd_inode_written(dap, inodedep) 11822 struct diradd *dap; 11823 struct inodedep *inodedep; 11824 { 11825 11826 LOCK_OWNED(VFSTOUFS(dap->da_list.wk_mp)); 11827 dap->da_state |= COMPLETE; 11828 complete_diradd(dap); 11829 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11830 } 11831 11832 /* 11833 * Returns true if the bmsafemap will have rollbacks when written. Must only 11834 * be called with the per-filesystem lock and the buf lock on the cg held. 11835 */ 11836 static int 11837 bmsafemap_backgroundwrite(bmsafemap, bp) 11838 struct bmsafemap *bmsafemap; 11839 struct buf *bp; 11840 { 11841 int dirty; 11842 11843 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11844 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11845 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11846 /* 11847 * If we're initiating a background write we need to process the 11848 * rollbacks as they exist now, not as they exist when IO starts. 11849 * No other consumers will look at the contents of the shadowed 11850 * buf so this is safe to do here. 11851 */ 11852 if (bp->b_xflags & BX_BKGRDMARKER) 11853 initiate_write_bmsafemap(bmsafemap, bp); 11854 11855 return (dirty); 11856 } 11857 11858 /* 11859 * Re-apply an allocation when a cg write is complete. 11860 */ 11861 static int 11862 jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11863 struct jnewblk *jnewblk; 11864 struct fs *fs; 11865 struct cg *cgp; 11866 uint8_t *blksfree; 11867 { 11868 ufs1_daddr_t fragno; 11869 ufs2_daddr_t blkno; 11870 long cgbno, bbase; 11871 int frags, blk; 11872 int i; 11873 11874 frags = 0; 11875 cgbno = dtogd(fs, jnewblk->jn_blkno); 11876 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11877 if (isclr(blksfree, cgbno + i)) 11878 panic("jnewblk_rollforward: re-allocated fragment"); 11879 frags++; 11880 } 11881 if (frags == fs->fs_frag) { 11882 blkno = fragstoblks(fs, cgbno); 11883 ffs_clrblock(fs, blksfree, (long)blkno); 11884 ffs_clusteracct(fs, cgp, blkno, -1); 11885 cgp->cg_cs.cs_nbfree--; 11886 } else { 11887 bbase = cgbno - fragnum(fs, cgbno); 11888 cgbno += jnewblk->jn_oldfrags; 11889 /* If a complete block had been reassembled, account for it. */ 11890 fragno = fragstoblks(fs, bbase); 11891 if (ffs_isblock(fs, blksfree, fragno)) { 11892 cgp->cg_cs.cs_nffree += fs->fs_frag; 11893 ffs_clusteracct(fs, cgp, fragno, -1); 11894 cgp->cg_cs.cs_nbfree--; 11895 } 11896 /* Decrement the old frags. */ 11897 blk = blkmap(fs, blksfree, bbase); 11898 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11899 /* Allocate the fragment */ 11900 for (i = 0; i < frags; i++) 11901 clrbit(blksfree, cgbno + i); 11902 cgp->cg_cs.cs_nffree -= frags; 11903 /* Add back in counts associated with the new frags */ 11904 blk = blkmap(fs, blksfree, bbase); 11905 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11906 } 11907 return (frags); 11908 } 11909 11910 /* 11911 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11912 * changes if it's not a background write. Set all written dependencies 11913 * to DEPCOMPLETE and free the structure if possible. 11914 * 11915 * If the write did not succeed, we will do all the roll-forward 11916 * operations, but we will not take the actions that will allow its 11917 * dependencies to be processed. 11918 */ 11919 static int 11920 handle_written_bmsafemap(bmsafemap, bp, flags) 11921 struct bmsafemap *bmsafemap; 11922 struct buf *bp; 11923 int flags; 11924 { 11925 struct newblk *newblk; 11926 struct inodedep *inodedep; 11927 struct jaddref *jaddref, *jatmp; 11928 struct jnewblk *jnewblk, *jntmp; 11929 struct ufsmount *ump; 11930 uint8_t *inosused; 11931 uint8_t *blksfree; 11932 struct cg *cgp; 11933 struct fs *fs; 11934 ino_t ino; 11935 int foreground; 11936 int chgs; 11937 11938 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11939 panic("handle_written_bmsafemap: Not started\n"); 11940 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11941 chgs = 0; 11942 bmsafemap->sm_state &= ~IOSTARTED; 11943 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11944 /* 11945 * If write was successful, release journal work that was waiting 11946 * on the write. Otherwise move the work back. 11947 */ 11948 if (flags & WRITESUCCEEDED) 11949 handle_jwork(&bmsafemap->sm_freewr); 11950 else 11951 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11952 worklist, wk_list); 11953 11954 /* 11955 * Restore unwritten inode allocation pending jaddref writes. 11956 */ 11957 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11958 cgp = (struct cg *)bp->b_data; 11959 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11960 inosused = cg_inosused(cgp); 11961 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11962 ja_bmdeps, jatmp) { 11963 if ((jaddref->ja_state & UNDONE) == 0) 11964 continue; 11965 ino = jaddref->ja_ino % fs->fs_ipg; 11966 if (isset(inosused, ino)) 11967 panic("handle_written_bmsafemap: " 11968 "re-allocated inode"); 11969 /* Do the roll-forward only if it's a real copy. */ 11970 if (foreground) { 11971 if ((jaddref->ja_mode & IFMT) == IFDIR) 11972 cgp->cg_cs.cs_ndir++; 11973 cgp->cg_cs.cs_nifree--; 11974 setbit(inosused, ino); 11975 chgs = 1; 11976 } 11977 jaddref->ja_state &= ~UNDONE; 11978 jaddref->ja_state |= ATTACHED; 11979 free_jaddref(jaddref); 11980 } 11981 } 11982 /* 11983 * Restore any block allocations which are pending journal writes. 11984 */ 11985 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11986 cgp = (struct cg *)bp->b_data; 11987 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11988 blksfree = cg_blksfree(cgp); 11989 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11990 jntmp) { 11991 if ((jnewblk->jn_state & UNDONE) == 0) 11992 continue; 11993 /* Do the roll-forward only if it's a real copy. */ 11994 if (foreground && 11995 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11996 chgs = 1; 11997 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11998 jnewblk->jn_state |= ATTACHED; 11999 free_jnewblk(jnewblk); 12000 } 12001 } 12002 /* 12003 * If the write did not succeed, we have done all the roll-forward 12004 * operations, but we cannot take the actions that will allow its 12005 * dependencies to be processed. 12006 */ 12007 if ((flags & WRITESUCCEEDED) == 0) { 12008 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 12009 newblk, nb_deps); 12010 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 12011 worklist, wk_list); 12012 if (foreground) 12013 bdirty(bp); 12014 return (1); 12015 } 12016 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 12017 newblk->nb_state |= DEPCOMPLETE; 12018 newblk->nb_state &= ~ONDEPLIST; 12019 newblk->nb_bmsafemap = NULL; 12020 LIST_REMOVE(newblk, nb_deps); 12021 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 12022 handle_allocdirect_partdone( 12023 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 12024 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 12025 handle_allocindir_partdone( 12026 WK_ALLOCINDIR(&newblk->nb_list)); 12027 else if (newblk->nb_list.wk_type != D_NEWBLK) 12028 panic("handle_written_bmsafemap: Unexpected type: %s", 12029 TYPENAME(newblk->nb_list.wk_type)); 12030 } 12031 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 12032 inodedep->id_state |= DEPCOMPLETE; 12033 inodedep->id_state &= ~ONDEPLIST; 12034 LIST_REMOVE(inodedep, id_deps); 12035 inodedep->id_bmsafemap = NULL; 12036 } 12037 LIST_REMOVE(bmsafemap, sm_next); 12038 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 12039 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 12040 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 12041 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 12042 LIST_EMPTY(&bmsafemap->sm_freehd)) { 12043 LIST_REMOVE(bmsafemap, sm_hash); 12044 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 12045 return (0); 12046 } 12047 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 12048 if (foreground) 12049 bdirty(bp); 12050 return (1); 12051 } 12052 12053 /* 12054 * Try to free a mkdir dependency. 12055 */ 12056 static void 12057 complete_mkdir(mkdir) 12058 struct mkdir *mkdir; 12059 { 12060 struct diradd *dap; 12061 12062 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 12063 return; 12064 LIST_REMOVE(mkdir, md_mkdirs); 12065 dap = mkdir->md_diradd; 12066 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 12067 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 12068 dap->da_state |= DEPCOMPLETE; 12069 complete_diradd(dap); 12070 } 12071 WORKITEM_FREE(mkdir, D_MKDIR); 12072 } 12073 12074 /* 12075 * Handle the completion of a mkdir dependency. 12076 */ 12077 static void 12078 handle_written_mkdir(mkdir, type) 12079 struct mkdir *mkdir; 12080 int type; 12081 { 12082 12083 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 12084 panic("handle_written_mkdir: bad type"); 12085 mkdir->md_state |= COMPLETE; 12086 complete_mkdir(mkdir); 12087 } 12088 12089 static int 12090 free_pagedep(pagedep) 12091 struct pagedep *pagedep; 12092 { 12093 int i; 12094 12095 if (pagedep->pd_state & NEWBLOCK) 12096 return (0); 12097 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 12098 return (0); 12099 for (i = 0; i < DAHASHSZ; i++) 12100 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 12101 return (0); 12102 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 12103 return (0); 12104 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 12105 return (0); 12106 if (pagedep->pd_state & ONWORKLIST) 12107 WORKLIST_REMOVE(&pagedep->pd_list); 12108 LIST_REMOVE(pagedep, pd_hash); 12109 WORKITEM_FREE(pagedep, D_PAGEDEP); 12110 12111 return (1); 12112 } 12113 12114 /* 12115 * Called from within softdep_disk_write_complete above. 12116 * A write operation was just completed. Removed inodes can 12117 * now be freed and associated block pointers may be committed. 12118 * Note that this routine is always called from interrupt level 12119 * with further interrupts from this device blocked. 12120 * 12121 * If the write did not succeed, we will do all the roll-forward 12122 * operations, but we will not take the actions that will allow its 12123 * dependencies to be processed. 12124 */ 12125 static int 12126 handle_written_filepage(pagedep, bp, flags) 12127 struct pagedep *pagedep; 12128 struct buf *bp; /* buffer containing the written page */ 12129 int flags; 12130 { 12131 struct dirrem *dirrem; 12132 struct diradd *dap, *nextdap; 12133 struct direct *ep; 12134 int i, chgs; 12135 12136 if ((pagedep->pd_state & IOSTARTED) == 0) 12137 panic("handle_written_filepage: not started"); 12138 pagedep->pd_state &= ~IOSTARTED; 12139 if ((flags & WRITESUCCEEDED) == 0) 12140 goto rollforward; 12141 /* 12142 * Process any directory removals that have been committed. 12143 */ 12144 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 12145 LIST_REMOVE(dirrem, dm_next); 12146 dirrem->dm_state |= COMPLETE; 12147 dirrem->dm_dirinum = pagedep->pd_ino; 12148 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 12149 ("handle_written_filepage: Journal entries not written.")); 12150 add_to_worklist(&dirrem->dm_list, 0); 12151 } 12152 /* 12153 * Free any directory additions that have been committed. 12154 * If it is a newly allocated block, we have to wait until 12155 * the on-disk directory inode claims the new block. 12156 */ 12157 if ((pagedep->pd_state & NEWBLOCK) == 0) 12158 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 12159 free_diradd(dap, NULL); 12160 rollforward: 12161 /* 12162 * Uncommitted directory entries must be restored. 12163 */ 12164 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 12165 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 12166 dap = nextdap) { 12167 nextdap = LIST_NEXT(dap, da_pdlist); 12168 if (dap->da_state & ATTACHED) 12169 panic("handle_written_filepage: attached"); 12170 ep = (struct direct *) 12171 ((char *)bp->b_data + dap->da_offset); 12172 ep->d_ino = dap->da_newinum; 12173 dap->da_state &= ~UNDONE; 12174 dap->da_state |= ATTACHED; 12175 chgs = 1; 12176 /* 12177 * If the inode referenced by the directory has 12178 * been written out, then the dependency can be 12179 * moved to the pending list. 12180 */ 12181 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 12182 LIST_REMOVE(dap, da_pdlist); 12183 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 12184 da_pdlist); 12185 } 12186 } 12187 } 12188 /* 12189 * If there were any rollbacks in the directory, then it must be 12190 * marked dirty so that its will eventually get written back in 12191 * its correct form. 12192 */ 12193 if (chgs || (flags & WRITESUCCEEDED) == 0) { 12194 if ((bp->b_flags & B_DELWRI) == 0) 12195 stat_dir_entry++; 12196 bdirty(bp); 12197 return (1); 12198 } 12199 /* 12200 * If we are not waiting for a new directory block to be 12201 * claimed by its inode, then the pagedep will be freed. 12202 * Otherwise it will remain to track any new entries on 12203 * the page in case they are fsync'ed. 12204 */ 12205 free_pagedep(pagedep); 12206 return (0); 12207 } 12208 12209 /* 12210 * Writing back in-core inode structures. 12211 * 12212 * The filesystem only accesses an inode's contents when it occupies an 12213 * "in-core" inode structure. These "in-core" structures are separate from 12214 * the page frames used to cache inode blocks. Only the latter are 12215 * transferred to/from the disk. So, when the updated contents of the 12216 * "in-core" inode structure are copied to the corresponding in-memory inode 12217 * block, the dependencies are also transferred. The following procedure is 12218 * called when copying a dirty "in-core" inode to a cached inode block. 12219 */ 12220 12221 /* 12222 * Called when an inode is loaded from disk. If the effective link count 12223 * differed from the actual link count when it was last flushed, then we 12224 * need to ensure that the correct effective link count is put back. 12225 */ 12226 void 12227 softdep_load_inodeblock(ip) 12228 struct inode *ip; /* the "in_core" copy of the inode */ 12229 { 12230 struct inodedep *inodedep; 12231 struct ufsmount *ump; 12232 12233 ump = ITOUMP(ip); 12234 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 12235 ("softdep_load_inodeblock called on non-softdep filesystem")); 12236 /* 12237 * Check for alternate nlink count. 12238 */ 12239 ip->i_effnlink = ip->i_nlink; 12240 ACQUIRE_LOCK(ump); 12241 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) { 12242 FREE_LOCK(ump); 12243 return; 12244 } 12245 ip->i_effnlink -= inodedep->id_nlinkdelta; 12246 FREE_LOCK(ump); 12247 } 12248 12249 /* 12250 * This routine is called just before the "in-core" inode 12251 * information is to be copied to the in-memory inode block. 12252 * Recall that an inode block contains several inodes. If 12253 * the force flag is set, then the dependencies will be 12254 * cleared so that the update can always be made. Note that 12255 * the buffer is locked when this routine is called, so we 12256 * will never be in the middle of writing the inode block 12257 * to disk. 12258 */ 12259 void 12260 softdep_update_inodeblock(ip, bp, waitfor) 12261 struct inode *ip; /* the "in_core" copy of the inode */ 12262 struct buf *bp; /* the buffer containing the inode block */ 12263 int waitfor; /* nonzero => update must be allowed */ 12264 { 12265 struct inodedep *inodedep; 12266 struct inoref *inoref; 12267 struct ufsmount *ump; 12268 struct worklist *wk; 12269 struct mount *mp; 12270 struct buf *ibp; 12271 struct fs *fs; 12272 int error; 12273 12274 ump = ITOUMP(ip); 12275 mp = UFSTOVFS(ump); 12276 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12277 ("softdep_update_inodeblock called on non-softdep filesystem")); 12278 fs = ump->um_fs; 12279 /* 12280 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12281 * does not have access to the in-core ip so must write directly into 12282 * the inode block buffer when setting freelink. 12283 */ 12284 if (fs->fs_magic == FS_UFS1_MAGIC) 12285 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12286 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12287 else 12288 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12289 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12290 /* 12291 * If the effective link count is not equal to the actual link 12292 * count, then we must track the difference in an inodedep while 12293 * the inode is (potentially) tossed out of the cache. Otherwise, 12294 * if there is no existing inodedep, then there are no dependencies 12295 * to track. 12296 */ 12297 ACQUIRE_LOCK(ump); 12298 again: 12299 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12300 FREE_LOCK(ump); 12301 if (ip->i_effnlink != ip->i_nlink) 12302 panic("softdep_update_inodeblock: bad link count"); 12303 return; 12304 } 12305 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12306 panic("softdep_update_inodeblock: bad delta"); 12307 /* 12308 * If we're flushing all dependencies we must also move any waiting 12309 * for journal writes onto the bufwait list prior to I/O. 12310 */ 12311 if (waitfor) { 12312 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12313 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12314 == DEPCOMPLETE) { 12315 jwait(&inoref->if_list, MNT_WAIT); 12316 goto again; 12317 } 12318 } 12319 } 12320 /* 12321 * Changes have been initiated. Anything depending on these 12322 * changes cannot occur until this inode has been written. 12323 */ 12324 inodedep->id_state &= ~COMPLETE; 12325 if ((inodedep->id_state & ONWORKLIST) == 0) 12326 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12327 /* 12328 * Any new dependencies associated with the incore inode must 12329 * now be moved to the list associated with the buffer holding 12330 * the in-memory copy of the inode. Once merged process any 12331 * allocdirects that are completed by the merger. 12332 */ 12333 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12334 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12335 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12336 NULL); 12337 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12338 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12339 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12340 NULL); 12341 /* 12342 * Now that the inode has been pushed into the buffer, the 12343 * operations dependent on the inode being written to disk 12344 * can be moved to the id_bufwait so that they will be 12345 * processed when the buffer I/O completes. 12346 */ 12347 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12348 WORKLIST_REMOVE(wk); 12349 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12350 } 12351 /* 12352 * Newly allocated inodes cannot be written until the bitmap 12353 * that allocates them have been written (indicated by 12354 * DEPCOMPLETE being set in id_state). If we are doing a 12355 * forced sync (e.g., an fsync on a file), we force the bitmap 12356 * to be written so that the update can be done. 12357 */ 12358 if (waitfor == 0) { 12359 FREE_LOCK(ump); 12360 return; 12361 } 12362 retry: 12363 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12364 FREE_LOCK(ump); 12365 return; 12366 } 12367 ibp = inodedep->id_bmsafemap->sm_buf; 12368 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12369 if (ibp == NULL) { 12370 /* 12371 * If ibp came back as NULL, the dependency could have been 12372 * freed while we slept. Look it up again, and check to see 12373 * that it has completed. 12374 */ 12375 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12376 goto retry; 12377 FREE_LOCK(ump); 12378 return; 12379 } 12380 FREE_LOCK(ump); 12381 if ((error = bwrite(ibp)) != 0) 12382 softdep_error("softdep_update_inodeblock: bwrite", error); 12383 } 12384 12385 /* 12386 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12387 * old inode dependency list (such as id_inoupdt). 12388 */ 12389 static void 12390 merge_inode_lists(newlisthead, oldlisthead) 12391 struct allocdirectlst *newlisthead; 12392 struct allocdirectlst *oldlisthead; 12393 { 12394 struct allocdirect *listadp, *newadp; 12395 12396 newadp = TAILQ_FIRST(newlisthead); 12397 if (newadp != NULL) 12398 LOCK_OWNED(VFSTOUFS(newadp->ad_block.nb_list.wk_mp)); 12399 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12400 if (listadp->ad_offset < newadp->ad_offset) { 12401 listadp = TAILQ_NEXT(listadp, ad_next); 12402 continue; 12403 } 12404 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12405 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12406 if (listadp->ad_offset == newadp->ad_offset) { 12407 allocdirect_merge(oldlisthead, newadp, 12408 listadp); 12409 listadp = newadp; 12410 } 12411 newadp = TAILQ_FIRST(newlisthead); 12412 } 12413 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12414 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12415 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12416 } 12417 } 12418 12419 /* 12420 * If we are doing an fsync, then we must ensure that any directory 12421 * entries for the inode have been written after the inode gets to disk. 12422 */ 12423 int 12424 softdep_fsync(vp) 12425 struct vnode *vp; /* the "in_core" copy of the inode */ 12426 { 12427 struct inodedep *inodedep; 12428 struct pagedep *pagedep; 12429 struct inoref *inoref; 12430 struct ufsmount *ump; 12431 struct worklist *wk; 12432 struct diradd *dap; 12433 struct mount *mp; 12434 struct vnode *pvp; 12435 struct inode *ip; 12436 struct buf *bp; 12437 struct fs *fs; 12438 struct thread *td = curthread; 12439 int error, flushparent, pagedep_new_block; 12440 ino_t parentino; 12441 ufs_lbn_t lbn; 12442 12443 ip = VTOI(vp); 12444 mp = vp->v_mount; 12445 ump = VFSTOUFS(mp); 12446 fs = ump->um_fs; 12447 if (MOUNTEDSOFTDEP(mp) == 0) 12448 return (0); 12449 ACQUIRE_LOCK(ump); 12450 restart: 12451 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12452 FREE_LOCK(ump); 12453 return (0); 12454 } 12455 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12456 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12457 == DEPCOMPLETE) { 12458 jwait(&inoref->if_list, MNT_WAIT); 12459 goto restart; 12460 } 12461 } 12462 if (!LIST_EMPTY(&inodedep->id_inowait) || 12463 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12464 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12465 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12466 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12467 panic("softdep_fsync: pending ops %p", inodedep); 12468 for (error = 0, flushparent = 0; ; ) { 12469 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12470 break; 12471 if (wk->wk_type != D_DIRADD) 12472 panic("softdep_fsync: Unexpected type %s", 12473 TYPENAME(wk->wk_type)); 12474 dap = WK_DIRADD(wk); 12475 /* 12476 * Flush our parent if this directory entry has a MKDIR_PARENT 12477 * dependency or is contained in a newly allocated block. 12478 */ 12479 if (dap->da_state & DIRCHG) 12480 pagedep = dap->da_previous->dm_pagedep; 12481 else 12482 pagedep = dap->da_pagedep; 12483 parentino = pagedep->pd_ino; 12484 lbn = pagedep->pd_lbn; 12485 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12486 panic("softdep_fsync: dirty"); 12487 if ((dap->da_state & MKDIR_PARENT) || 12488 (pagedep->pd_state & NEWBLOCK)) 12489 flushparent = 1; 12490 else 12491 flushparent = 0; 12492 /* 12493 * If we are being fsync'ed as part of vgone'ing this vnode, 12494 * then we will not be able to release and recover the 12495 * vnode below, so we just have to give up on writing its 12496 * directory entry out. It will eventually be written, just 12497 * not now, but then the user was not asking to have it 12498 * written, so we are not breaking any promises. 12499 */ 12500 if (vp->v_iflag & VI_DOOMED) 12501 break; 12502 /* 12503 * We prevent deadlock by always fetching inodes from the 12504 * root, moving down the directory tree. Thus, when fetching 12505 * our parent directory, we first try to get the lock. If 12506 * that fails, we must unlock ourselves before requesting 12507 * the lock on our parent. See the comment in ufs_lookup 12508 * for details on possible races. 12509 */ 12510 FREE_LOCK(ump); 12511 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12512 FFSV_FORCEINSMQ)) { 12513 error = vfs_busy(mp, MBF_NOWAIT); 12514 if (error != 0) { 12515 vfs_ref(mp); 12516 VOP_UNLOCK(vp, 0); 12517 error = vfs_busy(mp, 0); 12518 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12519 vfs_rel(mp); 12520 if (error != 0) 12521 return (ENOENT); 12522 if (vp->v_iflag & VI_DOOMED) { 12523 vfs_unbusy(mp); 12524 return (ENOENT); 12525 } 12526 } 12527 VOP_UNLOCK(vp, 0); 12528 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12529 &pvp, FFSV_FORCEINSMQ); 12530 vfs_unbusy(mp); 12531 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12532 if (vp->v_iflag & VI_DOOMED) { 12533 if (error == 0) 12534 vput(pvp); 12535 error = ENOENT; 12536 } 12537 if (error != 0) 12538 return (error); 12539 } 12540 /* 12541 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12542 * that are contained in direct blocks will be resolved by 12543 * doing a ffs_update. Pagedeps contained in indirect blocks 12544 * may require a complete sync'ing of the directory. So, we 12545 * try the cheap and fast ffs_update first, and if that fails, 12546 * then we do the slower ffs_syncvnode of the directory. 12547 */ 12548 if (flushparent) { 12549 int locked; 12550 12551 if ((error = ffs_update(pvp, 1)) != 0) { 12552 vput(pvp); 12553 return (error); 12554 } 12555 ACQUIRE_LOCK(ump); 12556 locked = 1; 12557 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12558 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12559 if (wk->wk_type != D_DIRADD) 12560 panic("softdep_fsync: Unexpected type %s", 12561 TYPENAME(wk->wk_type)); 12562 dap = WK_DIRADD(wk); 12563 if (dap->da_state & DIRCHG) 12564 pagedep = dap->da_previous->dm_pagedep; 12565 else 12566 pagedep = dap->da_pagedep; 12567 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12568 FREE_LOCK(ump); 12569 locked = 0; 12570 if (pagedep_new_block && (error = 12571 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12572 vput(pvp); 12573 return (error); 12574 } 12575 } 12576 } 12577 if (locked) 12578 FREE_LOCK(ump); 12579 } 12580 /* 12581 * Flush directory page containing the inode's name. 12582 */ 12583 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12584 &bp); 12585 if (error == 0) 12586 error = bwrite(bp); 12587 else 12588 brelse(bp); 12589 vput(pvp); 12590 if (error != 0) 12591 return (error); 12592 ACQUIRE_LOCK(ump); 12593 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12594 break; 12595 } 12596 FREE_LOCK(ump); 12597 return (0); 12598 } 12599 12600 /* 12601 * Flush all the dirty bitmaps associated with the block device 12602 * before flushing the rest of the dirty blocks so as to reduce 12603 * the number of dependencies that will have to be rolled back. 12604 * 12605 * XXX Unused? 12606 */ 12607 void 12608 softdep_fsync_mountdev(vp) 12609 struct vnode *vp; 12610 { 12611 struct buf *bp, *nbp; 12612 struct worklist *wk; 12613 struct bufobj *bo; 12614 12615 if (!vn_isdisk(vp, NULL)) 12616 panic("softdep_fsync_mountdev: vnode not a disk"); 12617 bo = &vp->v_bufobj; 12618 restart: 12619 BO_LOCK(bo); 12620 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12621 /* 12622 * If it is already scheduled, skip to the next buffer. 12623 */ 12624 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12625 continue; 12626 12627 if ((bp->b_flags & B_DELWRI) == 0) 12628 panic("softdep_fsync_mountdev: not dirty"); 12629 /* 12630 * We are only interested in bitmaps with outstanding 12631 * dependencies. 12632 */ 12633 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12634 wk->wk_type != D_BMSAFEMAP || 12635 (bp->b_vflags & BV_BKGRDINPROG)) { 12636 BUF_UNLOCK(bp); 12637 continue; 12638 } 12639 BO_UNLOCK(bo); 12640 bremfree(bp); 12641 (void) bawrite(bp); 12642 goto restart; 12643 } 12644 drain_output(vp); 12645 BO_UNLOCK(bo); 12646 } 12647 12648 /* 12649 * Sync all cylinder groups that were dirty at the time this function is 12650 * called. Newly dirtied cgs will be inserted before the sentinel. This 12651 * is used to flush freedep activity that may be holding up writes to a 12652 * indirect block. 12653 */ 12654 static int 12655 sync_cgs(mp, waitfor) 12656 struct mount *mp; 12657 int waitfor; 12658 { 12659 struct bmsafemap *bmsafemap; 12660 struct bmsafemap *sentinel; 12661 struct ufsmount *ump; 12662 struct buf *bp; 12663 int error; 12664 12665 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12666 sentinel->sm_cg = -1; 12667 ump = VFSTOUFS(mp); 12668 error = 0; 12669 ACQUIRE_LOCK(ump); 12670 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12671 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12672 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12673 /* Skip sentinels and cgs with no work to release. */ 12674 if (bmsafemap->sm_cg == -1 || 12675 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12676 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12677 LIST_REMOVE(sentinel, sm_next); 12678 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12679 continue; 12680 } 12681 /* 12682 * If we don't get the lock and we're waiting try again, if 12683 * not move on to the next buf and try to sync it. 12684 */ 12685 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12686 if (bp == NULL && waitfor == MNT_WAIT) 12687 continue; 12688 LIST_REMOVE(sentinel, sm_next); 12689 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12690 if (bp == NULL) 12691 continue; 12692 FREE_LOCK(ump); 12693 if (waitfor == MNT_NOWAIT) 12694 bawrite(bp); 12695 else 12696 error = bwrite(bp); 12697 ACQUIRE_LOCK(ump); 12698 if (error) 12699 break; 12700 } 12701 LIST_REMOVE(sentinel, sm_next); 12702 FREE_LOCK(ump); 12703 free(sentinel, M_BMSAFEMAP); 12704 return (error); 12705 } 12706 12707 /* 12708 * This routine is called when we are trying to synchronously flush a 12709 * file. This routine must eliminate any filesystem metadata dependencies 12710 * so that the syncing routine can succeed. 12711 */ 12712 int 12713 softdep_sync_metadata(struct vnode *vp) 12714 { 12715 struct inode *ip; 12716 int error; 12717 12718 ip = VTOI(vp); 12719 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12720 ("softdep_sync_metadata called on non-softdep filesystem")); 12721 /* 12722 * Ensure that any direct block dependencies have been cleared, 12723 * truncations are started, and inode references are journaled. 12724 */ 12725 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount)); 12726 /* 12727 * Write all journal records to prevent rollbacks on devvp. 12728 */ 12729 if (vp->v_type == VCHR) 12730 softdep_flushjournal(vp->v_mount); 12731 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12732 /* 12733 * Ensure that all truncates are written so we won't find deps on 12734 * indirect blocks. 12735 */ 12736 process_truncates(vp); 12737 FREE_LOCK(VFSTOUFS(vp->v_mount)); 12738 12739 return (error); 12740 } 12741 12742 /* 12743 * This routine is called when we are attempting to sync a buf with 12744 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12745 * other IO it can but returns EBUSY if the buffer is not yet able to 12746 * be written. Dependencies which will not cause rollbacks will always 12747 * return 0. 12748 */ 12749 int 12750 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12751 { 12752 struct indirdep *indirdep; 12753 struct pagedep *pagedep; 12754 struct allocindir *aip; 12755 struct newblk *newblk; 12756 struct ufsmount *ump; 12757 struct buf *nbp; 12758 struct worklist *wk; 12759 int i, error; 12760 12761 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12762 ("softdep_sync_buf called on non-softdep filesystem")); 12763 /* 12764 * For VCHR we just don't want to force flush any dependencies that 12765 * will cause rollbacks. 12766 */ 12767 if (vp->v_type == VCHR) { 12768 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12769 return (EBUSY); 12770 return (0); 12771 } 12772 ump = VFSTOUFS(vp->v_mount); 12773 ACQUIRE_LOCK(ump); 12774 /* 12775 * As we hold the buffer locked, none of its dependencies 12776 * will disappear. 12777 */ 12778 error = 0; 12779 top: 12780 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12781 switch (wk->wk_type) { 12782 12783 case D_ALLOCDIRECT: 12784 case D_ALLOCINDIR: 12785 newblk = WK_NEWBLK(wk); 12786 if (newblk->nb_jnewblk != NULL) { 12787 if (waitfor == MNT_NOWAIT) { 12788 error = EBUSY; 12789 goto out_unlock; 12790 } 12791 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12792 goto top; 12793 } 12794 if (newblk->nb_state & DEPCOMPLETE || 12795 waitfor == MNT_NOWAIT) 12796 continue; 12797 nbp = newblk->nb_bmsafemap->sm_buf; 12798 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12799 if (nbp == NULL) 12800 goto top; 12801 FREE_LOCK(ump); 12802 if ((error = bwrite(nbp)) != 0) 12803 goto out; 12804 ACQUIRE_LOCK(ump); 12805 continue; 12806 12807 case D_INDIRDEP: 12808 indirdep = WK_INDIRDEP(wk); 12809 if (waitfor == MNT_NOWAIT) { 12810 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12811 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12812 error = EBUSY; 12813 goto out_unlock; 12814 } 12815 } 12816 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12817 panic("softdep_sync_buf: truncation pending."); 12818 restart: 12819 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12820 newblk = (struct newblk *)aip; 12821 if (newblk->nb_jnewblk != NULL) { 12822 jwait(&newblk->nb_jnewblk->jn_list, 12823 waitfor); 12824 goto restart; 12825 } 12826 if (newblk->nb_state & DEPCOMPLETE) 12827 continue; 12828 nbp = newblk->nb_bmsafemap->sm_buf; 12829 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12830 if (nbp == NULL) 12831 goto restart; 12832 FREE_LOCK(ump); 12833 if ((error = bwrite(nbp)) != 0) 12834 goto out; 12835 ACQUIRE_LOCK(ump); 12836 goto restart; 12837 } 12838 continue; 12839 12840 case D_PAGEDEP: 12841 /* 12842 * Only flush directory entries in synchronous passes. 12843 */ 12844 if (waitfor != MNT_WAIT) { 12845 error = EBUSY; 12846 goto out_unlock; 12847 } 12848 /* 12849 * While syncing snapshots, we must allow recursive 12850 * lookups. 12851 */ 12852 BUF_AREC(bp); 12853 /* 12854 * We are trying to sync a directory that may 12855 * have dependencies on both its own metadata 12856 * and/or dependencies on the inodes of any 12857 * recently allocated files. We walk its diradd 12858 * lists pushing out the associated inode. 12859 */ 12860 pagedep = WK_PAGEDEP(wk); 12861 for (i = 0; i < DAHASHSZ; i++) { 12862 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12863 continue; 12864 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12865 &pagedep->pd_diraddhd[i]))) { 12866 BUF_NOREC(bp); 12867 goto out_unlock; 12868 } 12869 } 12870 BUF_NOREC(bp); 12871 continue; 12872 12873 case D_FREEWORK: 12874 case D_FREEDEP: 12875 case D_JSEGDEP: 12876 case D_JNEWBLK: 12877 continue; 12878 12879 default: 12880 panic("softdep_sync_buf: Unknown type %s", 12881 TYPENAME(wk->wk_type)); 12882 /* NOTREACHED */ 12883 } 12884 } 12885 out_unlock: 12886 FREE_LOCK(ump); 12887 out: 12888 return (error); 12889 } 12890 12891 /* 12892 * Flush the dependencies associated with an inodedep. 12893 */ 12894 static int 12895 flush_inodedep_deps(vp, mp, ino) 12896 struct vnode *vp; 12897 struct mount *mp; 12898 ino_t ino; 12899 { 12900 struct inodedep *inodedep; 12901 struct inoref *inoref; 12902 struct ufsmount *ump; 12903 int error, waitfor; 12904 12905 /* 12906 * This work is done in two passes. The first pass grabs most 12907 * of the buffers and begins asynchronously writing them. The 12908 * only way to wait for these asynchronous writes is to sleep 12909 * on the filesystem vnode which may stay busy for a long time 12910 * if the filesystem is active. So, instead, we make a second 12911 * pass over the dependencies blocking on each write. In the 12912 * usual case we will be blocking against a write that we 12913 * initiated, so when it is done the dependency will have been 12914 * resolved. Thus the second pass is expected to end quickly. 12915 * We give a brief window at the top of the loop to allow 12916 * any pending I/O to complete. 12917 */ 12918 ump = VFSTOUFS(mp); 12919 LOCK_OWNED(ump); 12920 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12921 if (error) 12922 return (error); 12923 FREE_LOCK(ump); 12924 ACQUIRE_LOCK(ump); 12925 restart: 12926 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12927 return (0); 12928 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12929 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12930 == DEPCOMPLETE) { 12931 jwait(&inoref->if_list, MNT_WAIT); 12932 goto restart; 12933 } 12934 } 12935 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12936 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12937 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12938 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12939 continue; 12940 /* 12941 * If pass2, we are done, otherwise do pass 2. 12942 */ 12943 if (waitfor == MNT_WAIT) 12944 break; 12945 waitfor = MNT_WAIT; 12946 } 12947 /* 12948 * Try freeing inodedep in case all dependencies have been removed. 12949 */ 12950 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12951 (void) free_inodedep(inodedep); 12952 return (0); 12953 } 12954 12955 /* 12956 * Flush an inode dependency list. 12957 */ 12958 static int 12959 flush_deplist(listhead, waitfor, errorp) 12960 struct allocdirectlst *listhead; 12961 int waitfor; 12962 int *errorp; 12963 { 12964 struct allocdirect *adp; 12965 struct newblk *newblk; 12966 struct ufsmount *ump; 12967 struct buf *bp; 12968 12969 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12970 return (0); 12971 ump = VFSTOUFS(adp->ad_list.wk_mp); 12972 LOCK_OWNED(ump); 12973 TAILQ_FOREACH(adp, listhead, ad_next) { 12974 newblk = (struct newblk *)adp; 12975 if (newblk->nb_jnewblk != NULL) { 12976 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12977 return (1); 12978 } 12979 if (newblk->nb_state & DEPCOMPLETE) 12980 continue; 12981 bp = newblk->nb_bmsafemap->sm_buf; 12982 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12983 if (bp == NULL) { 12984 if (waitfor == MNT_NOWAIT) 12985 continue; 12986 return (1); 12987 } 12988 FREE_LOCK(ump); 12989 if (waitfor == MNT_NOWAIT) 12990 bawrite(bp); 12991 else 12992 *errorp = bwrite(bp); 12993 ACQUIRE_LOCK(ump); 12994 return (1); 12995 } 12996 return (0); 12997 } 12998 12999 /* 13000 * Flush dependencies associated with an allocdirect block. 13001 */ 13002 static int 13003 flush_newblk_dep(vp, mp, lbn) 13004 struct vnode *vp; 13005 struct mount *mp; 13006 ufs_lbn_t lbn; 13007 { 13008 struct newblk *newblk; 13009 struct ufsmount *ump; 13010 struct bufobj *bo; 13011 struct inode *ip; 13012 struct buf *bp; 13013 ufs2_daddr_t blkno; 13014 int error; 13015 13016 error = 0; 13017 bo = &vp->v_bufobj; 13018 ip = VTOI(vp); 13019 blkno = DIP(ip, i_db[lbn]); 13020 if (blkno == 0) 13021 panic("flush_newblk_dep: Missing block"); 13022 ump = VFSTOUFS(mp); 13023 ACQUIRE_LOCK(ump); 13024 /* 13025 * Loop until all dependencies related to this block are satisfied. 13026 * We must be careful to restart after each sleep in case a write 13027 * completes some part of this process for us. 13028 */ 13029 for (;;) { 13030 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 13031 FREE_LOCK(ump); 13032 break; 13033 } 13034 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 13035 panic("flush_newblk_dep: Bad newblk %p", newblk); 13036 /* 13037 * Flush the journal. 13038 */ 13039 if (newblk->nb_jnewblk != NULL) { 13040 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 13041 continue; 13042 } 13043 /* 13044 * Write the bitmap dependency. 13045 */ 13046 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 13047 bp = newblk->nb_bmsafemap->sm_buf; 13048 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13049 if (bp == NULL) 13050 continue; 13051 FREE_LOCK(ump); 13052 error = bwrite(bp); 13053 if (error) 13054 break; 13055 ACQUIRE_LOCK(ump); 13056 continue; 13057 } 13058 /* 13059 * Write the buffer. 13060 */ 13061 FREE_LOCK(ump); 13062 BO_LOCK(bo); 13063 bp = gbincore(bo, lbn); 13064 if (bp != NULL) { 13065 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 13066 LK_INTERLOCK, BO_LOCKPTR(bo)); 13067 if (error == ENOLCK) { 13068 ACQUIRE_LOCK(ump); 13069 error = 0; 13070 continue; /* Slept, retry */ 13071 } 13072 if (error != 0) 13073 break; /* Failed */ 13074 if (bp->b_flags & B_DELWRI) { 13075 bremfree(bp); 13076 error = bwrite(bp); 13077 if (error) 13078 break; 13079 } else 13080 BUF_UNLOCK(bp); 13081 } else 13082 BO_UNLOCK(bo); 13083 /* 13084 * We have to wait for the direct pointers to 13085 * point at the newdirblk before the dependency 13086 * will go away. 13087 */ 13088 error = ffs_update(vp, 1); 13089 if (error) 13090 break; 13091 ACQUIRE_LOCK(ump); 13092 } 13093 return (error); 13094 } 13095 13096 /* 13097 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 13098 */ 13099 static int 13100 flush_pagedep_deps(pvp, mp, diraddhdp) 13101 struct vnode *pvp; 13102 struct mount *mp; 13103 struct diraddhd *diraddhdp; 13104 { 13105 struct inodedep *inodedep; 13106 struct inoref *inoref; 13107 struct ufsmount *ump; 13108 struct diradd *dap; 13109 struct vnode *vp; 13110 int error = 0; 13111 struct buf *bp; 13112 ino_t inum; 13113 struct diraddhd unfinished; 13114 13115 LIST_INIT(&unfinished); 13116 ump = VFSTOUFS(mp); 13117 LOCK_OWNED(ump); 13118 restart: 13119 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 13120 /* 13121 * Flush ourselves if this directory entry 13122 * has a MKDIR_PARENT dependency. 13123 */ 13124 if (dap->da_state & MKDIR_PARENT) { 13125 FREE_LOCK(ump); 13126 if ((error = ffs_update(pvp, 1)) != 0) 13127 break; 13128 ACQUIRE_LOCK(ump); 13129 /* 13130 * If that cleared dependencies, go on to next. 13131 */ 13132 if (dap != LIST_FIRST(diraddhdp)) 13133 continue; 13134 /* 13135 * All MKDIR_PARENT dependencies and all the 13136 * NEWBLOCK pagedeps that are contained in direct 13137 * blocks were resolved by doing above ffs_update. 13138 * Pagedeps contained in indirect blocks may 13139 * require a complete sync'ing of the directory. 13140 * We are in the midst of doing a complete sync, 13141 * so if they are not resolved in this pass we 13142 * defer them for now as they will be sync'ed by 13143 * our caller shortly. 13144 */ 13145 LIST_REMOVE(dap, da_pdlist); 13146 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 13147 continue; 13148 } 13149 /* 13150 * A newly allocated directory must have its "." and 13151 * ".." entries written out before its name can be 13152 * committed in its parent. 13153 */ 13154 inum = dap->da_newinum; 13155 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13156 panic("flush_pagedep_deps: lost inode1"); 13157 /* 13158 * Wait for any pending journal adds to complete so we don't 13159 * cause rollbacks while syncing. 13160 */ 13161 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13162 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13163 == DEPCOMPLETE) { 13164 jwait(&inoref->if_list, MNT_WAIT); 13165 goto restart; 13166 } 13167 } 13168 if (dap->da_state & MKDIR_BODY) { 13169 FREE_LOCK(ump); 13170 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13171 FFSV_FORCEINSMQ))) 13172 break; 13173 error = flush_newblk_dep(vp, mp, 0); 13174 /* 13175 * If we still have the dependency we might need to 13176 * update the vnode to sync the new link count to 13177 * disk. 13178 */ 13179 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 13180 error = ffs_update(vp, 1); 13181 vput(vp); 13182 if (error != 0) 13183 break; 13184 ACQUIRE_LOCK(ump); 13185 /* 13186 * If that cleared dependencies, go on to next. 13187 */ 13188 if (dap != LIST_FIRST(diraddhdp)) 13189 continue; 13190 if (dap->da_state & MKDIR_BODY) { 13191 inodedep_lookup(UFSTOVFS(ump), inum, 0, 13192 &inodedep); 13193 panic("flush_pagedep_deps: MKDIR_BODY " 13194 "inodedep %p dap %p vp %p", 13195 inodedep, dap, vp); 13196 } 13197 } 13198 /* 13199 * Flush the inode on which the directory entry depends. 13200 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 13201 * the only remaining dependency is that the updated inode 13202 * count must get pushed to disk. The inode has already 13203 * been pushed into its inode buffer (via VOP_UPDATE) at 13204 * the time of the reference count change. So we need only 13205 * locate that buffer, ensure that there will be no rollback 13206 * caused by a bitmap dependency, then write the inode buffer. 13207 */ 13208 retry: 13209 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13210 panic("flush_pagedep_deps: lost inode"); 13211 /* 13212 * If the inode still has bitmap dependencies, 13213 * push them to disk. 13214 */ 13215 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 13216 bp = inodedep->id_bmsafemap->sm_buf; 13217 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13218 if (bp == NULL) 13219 goto retry; 13220 FREE_LOCK(ump); 13221 if ((error = bwrite(bp)) != 0) 13222 break; 13223 ACQUIRE_LOCK(ump); 13224 if (dap != LIST_FIRST(diraddhdp)) 13225 continue; 13226 } 13227 /* 13228 * If the inode is still sitting in a buffer waiting 13229 * to be written or waiting for the link count to be 13230 * adjusted update it here to flush it to disk. 13231 */ 13232 if (dap == LIST_FIRST(diraddhdp)) { 13233 FREE_LOCK(ump); 13234 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13235 FFSV_FORCEINSMQ))) 13236 break; 13237 error = ffs_update(vp, 1); 13238 vput(vp); 13239 if (error) 13240 break; 13241 ACQUIRE_LOCK(ump); 13242 } 13243 /* 13244 * If we have failed to get rid of all the dependencies 13245 * then something is seriously wrong. 13246 */ 13247 if (dap == LIST_FIRST(diraddhdp)) { 13248 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13249 panic("flush_pagedep_deps: failed to flush " 13250 "inodedep %p ino %ju dap %p", 13251 inodedep, (uintmax_t)inum, dap); 13252 } 13253 } 13254 if (error) 13255 ACQUIRE_LOCK(ump); 13256 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13257 LIST_REMOVE(dap, da_pdlist); 13258 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13259 } 13260 return (error); 13261 } 13262 13263 /* 13264 * A large burst of file addition or deletion activity can drive the 13265 * memory load excessively high. First attempt to slow things down 13266 * using the techniques below. If that fails, this routine requests 13267 * the offending operations to fall back to running synchronously 13268 * until the memory load returns to a reasonable level. 13269 */ 13270 int 13271 softdep_slowdown(vp) 13272 struct vnode *vp; 13273 { 13274 struct ufsmount *ump; 13275 int jlow; 13276 int max_softdeps_hard; 13277 13278 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13279 ("softdep_slowdown called on non-softdep filesystem")); 13280 ump = VFSTOUFS(vp->v_mount); 13281 ACQUIRE_LOCK(ump); 13282 jlow = 0; 13283 /* 13284 * Check for journal space if needed. 13285 */ 13286 if (DOINGSUJ(vp)) { 13287 if (journal_space(ump, 0) == 0) 13288 jlow = 1; 13289 } 13290 /* 13291 * If the system is under its limits and our filesystem is 13292 * not responsible for more than our share of the usage and 13293 * we are not low on journal space, then no need to slow down. 13294 */ 13295 max_softdeps_hard = max_softdeps * 11 / 10; 13296 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13297 dep_current[D_INODEDEP] < max_softdeps_hard && 13298 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13299 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13300 ump->softdep_curdeps[D_DIRREM] < 13301 (max_softdeps_hard / 2) / stat_flush_threads && 13302 ump->softdep_curdeps[D_INODEDEP] < 13303 max_softdeps_hard / stat_flush_threads && 13304 ump->softdep_curdeps[D_INDIRDEP] < 13305 (max_softdeps_hard / 1000) / stat_flush_threads && 13306 ump->softdep_curdeps[D_FREEBLKS] < 13307 max_softdeps_hard / stat_flush_threads) { 13308 FREE_LOCK(ump); 13309 return (0); 13310 } 13311 /* 13312 * If the journal is low or our filesystem is over its limit 13313 * then speedup the cleanup. 13314 */ 13315 if (ump->softdep_curdeps[D_INDIRDEP] < 13316 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13317 softdep_speedup(ump); 13318 stat_sync_limit_hit += 1; 13319 FREE_LOCK(ump); 13320 /* 13321 * We only slow down the rate at which new dependencies are 13322 * generated if we are not using journaling. With journaling, 13323 * the cleanup should always be sufficient to keep things 13324 * under control. 13325 */ 13326 if (DOINGSUJ(vp)) 13327 return (0); 13328 return (1); 13329 } 13330 13331 /* 13332 * Called by the allocation routines when they are about to fail 13333 * in the hope that we can free up the requested resource (inodes 13334 * or disk space). 13335 * 13336 * First check to see if the work list has anything on it. If it has, 13337 * clean up entries until we successfully free the requested resource. 13338 * Because this process holds inodes locked, we cannot handle any remove 13339 * requests that might block on a locked inode as that could lead to 13340 * deadlock. If the worklist yields none of the requested resource, 13341 * start syncing out vnodes to free up the needed space. 13342 */ 13343 int 13344 softdep_request_cleanup(fs, vp, cred, resource) 13345 struct fs *fs; 13346 struct vnode *vp; 13347 struct ucred *cred; 13348 int resource; 13349 { 13350 struct ufsmount *ump; 13351 struct mount *mp; 13352 long starttime; 13353 ufs2_daddr_t needed; 13354 int error, failed_vnode; 13355 13356 /* 13357 * If we are being called because of a process doing a 13358 * copy-on-write, then it is not safe to process any 13359 * worklist items as we will recurse into the copyonwrite 13360 * routine. This will result in an incoherent snapshot. 13361 * If the vnode that we hold is a snapshot, we must avoid 13362 * handling other resources that could cause deadlock. 13363 */ 13364 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13365 return (0); 13366 13367 if (resource == FLUSH_BLOCKS_WAIT) 13368 stat_cleanup_blkrequests += 1; 13369 else 13370 stat_cleanup_inorequests += 1; 13371 13372 mp = vp->v_mount; 13373 ump = VFSTOUFS(mp); 13374 mtx_assert(UFS_MTX(ump), MA_OWNED); 13375 UFS_UNLOCK(ump); 13376 error = ffs_update(vp, 1); 13377 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13378 UFS_LOCK(ump); 13379 return (0); 13380 } 13381 /* 13382 * If we are in need of resources, start by cleaning up 13383 * any block removals associated with our inode. 13384 */ 13385 ACQUIRE_LOCK(ump); 13386 process_removes(vp); 13387 process_truncates(vp); 13388 FREE_LOCK(ump); 13389 /* 13390 * Now clean up at least as many resources as we will need. 13391 * 13392 * When requested to clean up inodes, the number that are needed 13393 * is set by the number of simultaneous writers (mnt_writeopcount) 13394 * plus a bit of slop (2) in case some more writers show up while 13395 * we are cleaning. 13396 * 13397 * When requested to free up space, the amount of space that 13398 * we need is enough blocks to allocate a full-sized segment 13399 * (fs_contigsumsize). The number of such segments that will 13400 * be needed is set by the number of simultaneous writers 13401 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13402 * writers show up while we are cleaning. 13403 * 13404 * Additionally, if we are unpriviledged and allocating space, 13405 * we need to ensure that we clean up enough blocks to get the 13406 * needed number of blocks over the threshold of the minimum 13407 * number of blocks required to be kept free by the filesystem 13408 * (fs_minfree). 13409 */ 13410 if (resource == FLUSH_INODES_WAIT) { 13411 needed = vp->v_mount->mnt_writeopcount + 2; 13412 } else if (resource == FLUSH_BLOCKS_WAIT) { 13413 needed = (vp->v_mount->mnt_writeopcount + 2) * 13414 fs->fs_contigsumsize; 13415 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE)) 13416 needed += fragstoblks(fs, 13417 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13418 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13419 } else { 13420 UFS_LOCK(ump); 13421 printf("softdep_request_cleanup: Unknown resource type %d\n", 13422 resource); 13423 return (0); 13424 } 13425 starttime = time_second; 13426 retry: 13427 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13428 fs->fs_cstotal.cs_nbfree <= needed) || 13429 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13430 fs->fs_cstotal.cs_nifree <= needed)) { 13431 ACQUIRE_LOCK(ump); 13432 if (ump->softdep_on_worklist > 0 && 13433 process_worklist_item(UFSTOVFS(ump), 13434 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13435 stat_worklist_push += 1; 13436 FREE_LOCK(ump); 13437 } 13438 /* 13439 * If we still need resources and there are no more worklist 13440 * entries to process to obtain them, we have to start flushing 13441 * the dirty vnodes to force the release of additional requests 13442 * to the worklist that we can then process to reap addition 13443 * resources. We walk the vnodes associated with the mount point 13444 * until we get the needed worklist requests that we can reap. 13445 * 13446 * If there are several threads all needing to clean the same 13447 * mount point, only one is allowed to walk the mount list. 13448 * When several threads all try to walk the same mount list, 13449 * they end up competing with each other and often end up in 13450 * livelock. This approach ensures that forward progress is 13451 * made at the cost of occational ENOSPC errors being returned 13452 * that might otherwise have been avoided. 13453 */ 13454 error = 1; 13455 if ((resource == FLUSH_BLOCKS_WAIT && 13456 fs->fs_cstotal.cs_nbfree <= needed) || 13457 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13458 fs->fs_cstotal.cs_nifree <= needed)) { 13459 ACQUIRE_LOCK(ump); 13460 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) { 13461 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE; 13462 FREE_LOCK(ump); 13463 failed_vnode = softdep_request_cleanup_flush(mp, ump); 13464 ACQUIRE_LOCK(ump); 13465 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE; 13466 FREE_LOCK(ump); 13467 if (ump->softdep_on_worklist > 0) { 13468 stat_cleanup_retries += 1; 13469 if (!failed_vnode) 13470 goto retry; 13471 } 13472 } else { 13473 FREE_LOCK(ump); 13474 error = 0; 13475 } 13476 stat_cleanup_failures += 1; 13477 } 13478 if (time_second - starttime > stat_cleanup_high_delay) 13479 stat_cleanup_high_delay = time_second - starttime; 13480 UFS_LOCK(ump); 13481 return (error); 13482 } 13483 13484 /* 13485 * Scan the vnodes for the specified mount point flushing out any 13486 * vnodes that can be locked without waiting. Finally, try to flush 13487 * the device associated with the mount point if it can be locked 13488 * without waiting. 13489 * 13490 * We return 0 if we were able to lock every vnode in our scan. 13491 * If we had to skip one or more vnodes, we return 1. 13492 */ 13493 static int 13494 softdep_request_cleanup_flush(mp, ump) 13495 struct mount *mp; 13496 struct ufsmount *ump; 13497 { 13498 struct thread *td; 13499 struct vnode *lvp, *mvp; 13500 int failed_vnode; 13501 13502 failed_vnode = 0; 13503 td = curthread; 13504 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13505 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13506 VI_UNLOCK(lvp); 13507 continue; 13508 } 13509 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13510 td) != 0) { 13511 failed_vnode = 1; 13512 continue; 13513 } 13514 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13515 vput(lvp); 13516 continue; 13517 } 13518 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13519 vput(lvp); 13520 } 13521 lvp = ump->um_devvp; 13522 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13523 VOP_FSYNC(lvp, MNT_NOWAIT, td); 13524 VOP_UNLOCK(lvp, 0); 13525 } 13526 return (failed_vnode); 13527 } 13528 13529 static bool 13530 softdep_excess_items(struct ufsmount *ump, int item) 13531 { 13532 13533 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13534 return (dep_current[item] > max_softdeps && 13535 ump->softdep_curdeps[item] > max_softdeps / 13536 stat_flush_threads); 13537 } 13538 13539 static void 13540 schedule_cleanup(struct mount *mp) 13541 { 13542 struct ufsmount *ump; 13543 struct thread *td; 13544 13545 ump = VFSTOUFS(mp); 13546 LOCK_OWNED(ump); 13547 FREE_LOCK(ump); 13548 td = curthread; 13549 if ((td->td_pflags & TDP_KTHREAD) != 0 && 13550 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 13551 /* 13552 * No ast is delivered to kernel threads, so nobody 13553 * would deref the mp. Some kernel threads 13554 * explicitely check for AST, e.g. NFS daemon does 13555 * this in the serving loop. 13556 */ 13557 return; 13558 } 13559 if (td->td_su != NULL) 13560 vfs_rel(td->td_su); 13561 vfs_ref(mp); 13562 td->td_su = mp; 13563 thread_lock(td); 13564 td->td_flags |= TDF_ASTPENDING; 13565 thread_unlock(td); 13566 } 13567 13568 static void 13569 softdep_ast_cleanup_proc(struct thread *td) 13570 { 13571 struct mount *mp; 13572 struct ufsmount *ump; 13573 int error; 13574 bool req; 13575 13576 while ((mp = td->td_su) != NULL) { 13577 td->td_su = NULL; 13578 error = vfs_busy(mp, MBF_NOWAIT); 13579 vfs_rel(mp); 13580 if (error != 0) 13581 return; 13582 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 13583 ump = VFSTOUFS(mp); 13584 for (;;) { 13585 req = false; 13586 ACQUIRE_LOCK(ump); 13587 if (softdep_excess_items(ump, D_INODEDEP)) { 13588 req = true; 13589 request_cleanup(mp, FLUSH_INODES); 13590 } 13591 if (softdep_excess_items(ump, D_DIRREM)) { 13592 req = true; 13593 request_cleanup(mp, FLUSH_BLOCKS); 13594 } 13595 FREE_LOCK(ump); 13596 if (softdep_excess_items(ump, D_NEWBLK) || 13597 softdep_excess_items(ump, D_ALLOCDIRECT) || 13598 softdep_excess_items(ump, D_ALLOCINDIR)) { 13599 error = vn_start_write(NULL, &mp, 13600 V_WAIT); 13601 if (error == 0) { 13602 req = true; 13603 VFS_SYNC(mp, MNT_WAIT); 13604 vn_finished_write(mp); 13605 } 13606 } 13607 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 13608 break; 13609 } 13610 } 13611 vfs_unbusy(mp); 13612 } 13613 if ((mp = td->td_su) != NULL) { 13614 td->td_su = NULL; 13615 vfs_rel(mp); 13616 } 13617 } 13618 13619 /* 13620 * If memory utilization has gotten too high, deliberately slow things 13621 * down and speed up the I/O processing. 13622 */ 13623 static int 13624 request_cleanup(mp, resource) 13625 struct mount *mp; 13626 int resource; 13627 { 13628 struct thread *td = curthread; 13629 struct ufsmount *ump; 13630 13631 ump = VFSTOUFS(mp); 13632 LOCK_OWNED(ump); 13633 /* 13634 * We never hold up the filesystem syncer or buf daemon. 13635 */ 13636 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13637 return (0); 13638 /* 13639 * First check to see if the work list has gotten backlogged. 13640 * If it has, co-opt this process to help clean up two entries. 13641 * Because this process may hold inodes locked, we cannot 13642 * handle any remove requests that might block on a locked 13643 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13644 * to avoid recursively processing the worklist. 13645 */ 13646 if (ump->softdep_on_worklist > max_softdeps / 10) { 13647 td->td_pflags |= TDP_SOFTDEP; 13648 process_worklist_item(mp, 2, LK_NOWAIT); 13649 td->td_pflags &= ~TDP_SOFTDEP; 13650 stat_worklist_push += 2; 13651 return(1); 13652 } 13653 /* 13654 * Next, we attempt to speed up the syncer process. If that 13655 * is successful, then we allow the process to continue. 13656 */ 13657 if (softdep_speedup(ump) && 13658 resource != FLUSH_BLOCKS_WAIT && 13659 resource != FLUSH_INODES_WAIT) 13660 return(0); 13661 /* 13662 * If we are resource constrained on inode dependencies, try 13663 * flushing some dirty inodes. Otherwise, we are constrained 13664 * by file deletions, so try accelerating flushes of directories 13665 * with removal dependencies. We would like to do the cleanup 13666 * here, but we probably hold an inode locked at this point and 13667 * that might deadlock against one that we try to clean. So, 13668 * the best that we can do is request the syncer daemon to do 13669 * the cleanup for us. 13670 */ 13671 switch (resource) { 13672 13673 case FLUSH_INODES: 13674 case FLUSH_INODES_WAIT: 13675 ACQUIRE_GBLLOCK(&lk); 13676 stat_ino_limit_push += 1; 13677 req_clear_inodedeps += 1; 13678 FREE_GBLLOCK(&lk); 13679 stat_countp = &stat_ino_limit_hit; 13680 break; 13681 13682 case FLUSH_BLOCKS: 13683 case FLUSH_BLOCKS_WAIT: 13684 ACQUIRE_GBLLOCK(&lk); 13685 stat_blk_limit_push += 1; 13686 req_clear_remove += 1; 13687 FREE_GBLLOCK(&lk); 13688 stat_countp = &stat_blk_limit_hit; 13689 break; 13690 13691 default: 13692 panic("request_cleanup: unknown type"); 13693 } 13694 /* 13695 * Hopefully the syncer daemon will catch up and awaken us. 13696 * We wait at most tickdelay before proceeding in any case. 13697 */ 13698 ACQUIRE_GBLLOCK(&lk); 13699 FREE_LOCK(ump); 13700 proc_waiting += 1; 13701 if (callout_pending(&softdep_callout) == FALSE) 13702 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13703 pause_timer, 0); 13704 13705 if ((td->td_pflags & TDP_KTHREAD) == 0) 13706 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13707 proc_waiting -= 1; 13708 FREE_GBLLOCK(&lk); 13709 ACQUIRE_LOCK(ump); 13710 return (1); 13711 } 13712 13713 /* 13714 * Awaken processes pausing in request_cleanup and clear proc_waiting 13715 * to indicate that there is no longer a timer running. Pause_timer 13716 * will be called with the global softdep mutex (&lk) locked. 13717 */ 13718 static void 13719 pause_timer(arg) 13720 void *arg; 13721 { 13722 13723 GBLLOCK_OWNED(&lk); 13724 /* 13725 * The callout_ API has acquired mtx and will hold it around this 13726 * function call. 13727 */ 13728 *stat_countp += proc_waiting; 13729 wakeup(&proc_waiting); 13730 } 13731 13732 /* 13733 * If requested, try removing inode or removal dependencies. 13734 */ 13735 static void 13736 check_clear_deps(mp) 13737 struct mount *mp; 13738 { 13739 13740 /* 13741 * If we are suspended, it may be because of our using 13742 * too many inodedeps, so help clear them out. 13743 */ 13744 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13745 clear_inodedeps(mp); 13746 /* 13747 * General requests for cleanup of backed up dependencies 13748 */ 13749 ACQUIRE_GBLLOCK(&lk); 13750 if (req_clear_inodedeps) { 13751 req_clear_inodedeps -= 1; 13752 FREE_GBLLOCK(&lk); 13753 clear_inodedeps(mp); 13754 ACQUIRE_GBLLOCK(&lk); 13755 wakeup(&proc_waiting); 13756 } 13757 if (req_clear_remove) { 13758 req_clear_remove -= 1; 13759 FREE_GBLLOCK(&lk); 13760 clear_remove(mp); 13761 ACQUIRE_GBLLOCK(&lk); 13762 wakeup(&proc_waiting); 13763 } 13764 FREE_GBLLOCK(&lk); 13765 } 13766 13767 /* 13768 * Flush out a directory with at least one removal dependency in an effort to 13769 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13770 */ 13771 static void 13772 clear_remove(mp) 13773 struct mount *mp; 13774 { 13775 struct pagedep_hashhead *pagedephd; 13776 struct pagedep *pagedep; 13777 struct ufsmount *ump; 13778 struct vnode *vp; 13779 struct bufobj *bo; 13780 int error, cnt; 13781 ino_t ino; 13782 13783 ump = VFSTOUFS(mp); 13784 LOCK_OWNED(ump); 13785 13786 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13787 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13788 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13789 ump->pagedep_nextclean = 0; 13790 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13791 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13792 continue; 13793 ino = pagedep->pd_ino; 13794 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13795 continue; 13796 FREE_LOCK(ump); 13797 13798 /* 13799 * Let unmount clear deps 13800 */ 13801 error = vfs_busy(mp, MBF_NOWAIT); 13802 if (error != 0) 13803 goto finish_write; 13804 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13805 FFSV_FORCEINSMQ); 13806 vfs_unbusy(mp); 13807 if (error != 0) { 13808 softdep_error("clear_remove: vget", error); 13809 goto finish_write; 13810 } 13811 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13812 softdep_error("clear_remove: fsync", error); 13813 bo = &vp->v_bufobj; 13814 BO_LOCK(bo); 13815 drain_output(vp); 13816 BO_UNLOCK(bo); 13817 vput(vp); 13818 finish_write: 13819 vn_finished_write(mp); 13820 ACQUIRE_LOCK(ump); 13821 return; 13822 } 13823 } 13824 } 13825 13826 /* 13827 * Clear out a block of dirty inodes in an effort to reduce 13828 * the number of inodedep dependency structures. 13829 */ 13830 static void 13831 clear_inodedeps(mp) 13832 struct mount *mp; 13833 { 13834 struct inodedep_hashhead *inodedephd; 13835 struct inodedep *inodedep; 13836 struct ufsmount *ump; 13837 struct vnode *vp; 13838 struct fs *fs; 13839 int error, cnt; 13840 ino_t firstino, lastino, ino; 13841 13842 ump = VFSTOUFS(mp); 13843 fs = ump->um_fs; 13844 LOCK_OWNED(ump); 13845 /* 13846 * Pick a random inode dependency to be cleared. 13847 * We will then gather up all the inodes in its block 13848 * that have dependencies and flush them out. 13849 */ 13850 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13851 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13852 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13853 ump->inodedep_nextclean = 0; 13854 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13855 break; 13856 } 13857 if (inodedep == NULL) 13858 return; 13859 /* 13860 * Find the last inode in the block with dependencies. 13861 */ 13862 firstino = rounddown2(inodedep->id_ino, INOPB(fs)); 13863 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13864 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13865 break; 13866 /* 13867 * Asynchronously push all but the last inode with dependencies. 13868 * Synchronously push the last inode with dependencies to ensure 13869 * that the inode block gets written to free up the inodedeps. 13870 */ 13871 for (ino = firstino; ino <= lastino; ino++) { 13872 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13873 continue; 13874 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13875 continue; 13876 FREE_LOCK(ump); 13877 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13878 if (error != 0) { 13879 vn_finished_write(mp); 13880 ACQUIRE_LOCK(ump); 13881 return; 13882 } 13883 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13884 FFSV_FORCEINSMQ)) != 0) { 13885 softdep_error("clear_inodedeps: vget", error); 13886 vfs_unbusy(mp); 13887 vn_finished_write(mp); 13888 ACQUIRE_LOCK(ump); 13889 return; 13890 } 13891 vfs_unbusy(mp); 13892 if (ino == lastino) { 13893 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13894 softdep_error("clear_inodedeps: fsync1", error); 13895 } else { 13896 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13897 softdep_error("clear_inodedeps: fsync2", error); 13898 BO_LOCK(&vp->v_bufobj); 13899 drain_output(vp); 13900 BO_UNLOCK(&vp->v_bufobj); 13901 } 13902 vput(vp); 13903 vn_finished_write(mp); 13904 ACQUIRE_LOCK(ump); 13905 } 13906 } 13907 13908 void 13909 softdep_buf_append(bp, wkhd) 13910 struct buf *bp; 13911 struct workhead *wkhd; 13912 { 13913 struct worklist *wk; 13914 struct ufsmount *ump; 13915 13916 if ((wk = LIST_FIRST(wkhd)) == NULL) 13917 return; 13918 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13919 ("softdep_buf_append called on non-softdep filesystem")); 13920 ump = VFSTOUFS(wk->wk_mp); 13921 ACQUIRE_LOCK(ump); 13922 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13923 WORKLIST_REMOVE(wk); 13924 WORKLIST_INSERT(&bp->b_dep, wk); 13925 } 13926 FREE_LOCK(ump); 13927 13928 } 13929 13930 void 13931 softdep_inode_append(ip, cred, wkhd) 13932 struct inode *ip; 13933 struct ucred *cred; 13934 struct workhead *wkhd; 13935 { 13936 struct buf *bp; 13937 struct fs *fs; 13938 struct ufsmount *ump; 13939 int error; 13940 13941 ump = ITOUMP(ip); 13942 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 13943 ("softdep_inode_append called on non-softdep filesystem")); 13944 fs = ump->um_fs; 13945 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13946 (int)fs->fs_bsize, cred, &bp); 13947 if (error) { 13948 bqrelse(bp); 13949 softdep_freework(wkhd); 13950 return; 13951 } 13952 softdep_buf_append(bp, wkhd); 13953 bqrelse(bp); 13954 } 13955 13956 void 13957 softdep_freework(wkhd) 13958 struct workhead *wkhd; 13959 { 13960 struct worklist *wk; 13961 struct ufsmount *ump; 13962 13963 if ((wk = LIST_FIRST(wkhd)) == NULL) 13964 return; 13965 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13966 ("softdep_freework called on non-softdep filesystem")); 13967 ump = VFSTOUFS(wk->wk_mp); 13968 ACQUIRE_LOCK(ump); 13969 handle_jwork(wkhd); 13970 FREE_LOCK(ump); 13971 } 13972 13973 static struct ufsmount * 13974 softdep_bp_to_mp(bp) 13975 struct buf *bp; 13976 { 13977 struct mount *mp; 13978 struct vnode *vp; 13979 13980 if (LIST_EMPTY(&bp->b_dep)) 13981 return (NULL); 13982 vp = bp->b_vp; 13983 KASSERT(vp != NULL, 13984 ("%s, buffer with dependencies lacks vnode", __func__)); 13985 13986 /* 13987 * The ump mount point is stable after we get a correct 13988 * pointer, since bp is locked and this prevents unmount from 13989 * proceeding. But to get to it, we cannot dereference bp->b_dep 13990 * head wk_mp, because we do not yet own SU ump lock and 13991 * workitem might be freed while dereferenced. 13992 */ 13993 retry: 13994 switch (vp->v_type) { 13995 case VCHR: 13996 VI_LOCK(vp); 13997 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL; 13998 VI_UNLOCK(vp); 13999 if (mp == NULL) 14000 goto retry; 14001 break; 14002 case VREG: 14003 case VDIR: 14004 case VLNK: 14005 case VFIFO: 14006 case VSOCK: 14007 mp = vp->v_mount; 14008 break; 14009 case VBLK: 14010 vn_printf(vp, "softdep_bp_to_mp: unexpected block device\n"); 14011 /* FALLTHROUGH */ 14012 case VNON: 14013 case VBAD: 14014 case VMARKER: 14015 mp = NULL; 14016 break; 14017 default: 14018 vn_printf(vp, "unknown vnode type"); 14019 mp = NULL; 14020 break; 14021 } 14022 return (VFSTOUFS(mp)); 14023 } 14024 14025 /* 14026 * Function to determine if the buffer has outstanding dependencies 14027 * that will cause a roll-back if the buffer is written. If wantcount 14028 * is set, return number of dependencies, otherwise just yes or no. 14029 */ 14030 static int 14031 softdep_count_dependencies(bp, wantcount) 14032 struct buf *bp; 14033 int wantcount; 14034 { 14035 struct worklist *wk; 14036 struct ufsmount *ump; 14037 struct bmsafemap *bmsafemap; 14038 struct freework *freework; 14039 struct inodedep *inodedep; 14040 struct indirdep *indirdep; 14041 struct freeblks *freeblks; 14042 struct allocindir *aip; 14043 struct pagedep *pagedep; 14044 struct dirrem *dirrem; 14045 struct newblk *newblk; 14046 struct mkdir *mkdir; 14047 struct diradd *dap; 14048 int i, retval; 14049 14050 ump = softdep_bp_to_mp(bp); 14051 if (ump == NULL) 14052 return (0); 14053 retval = 0; 14054 ACQUIRE_LOCK(ump); 14055 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 14056 switch (wk->wk_type) { 14057 14058 case D_INODEDEP: 14059 inodedep = WK_INODEDEP(wk); 14060 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 14061 /* bitmap allocation dependency */ 14062 retval += 1; 14063 if (!wantcount) 14064 goto out; 14065 } 14066 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 14067 /* direct block pointer dependency */ 14068 retval += 1; 14069 if (!wantcount) 14070 goto out; 14071 } 14072 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 14073 /* direct block pointer dependency */ 14074 retval += 1; 14075 if (!wantcount) 14076 goto out; 14077 } 14078 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 14079 /* Add reference dependency. */ 14080 retval += 1; 14081 if (!wantcount) 14082 goto out; 14083 } 14084 continue; 14085 14086 case D_INDIRDEP: 14087 indirdep = WK_INDIRDEP(wk); 14088 14089 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 14090 /* indirect truncation dependency */ 14091 retval += 1; 14092 if (!wantcount) 14093 goto out; 14094 } 14095 14096 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 14097 /* indirect block pointer dependency */ 14098 retval += 1; 14099 if (!wantcount) 14100 goto out; 14101 } 14102 continue; 14103 14104 case D_PAGEDEP: 14105 pagedep = WK_PAGEDEP(wk); 14106 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 14107 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 14108 /* Journal remove ref dependency. */ 14109 retval += 1; 14110 if (!wantcount) 14111 goto out; 14112 } 14113 } 14114 for (i = 0; i < DAHASHSZ; i++) { 14115 14116 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 14117 /* directory entry dependency */ 14118 retval += 1; 14119 if (!wantcount) 14120 goto out; 14121 } 14122 } 14123 continue; 14124 14125 case D_BMSAFEMAP: 14126 bmsafemap = WK_BMSAFEMAP(wk); 14127 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 14128 /* Add reference dependency. */ 14129 retval += 1; 14130 if (!wantcount) 14131 goto out; 14132 } 14133 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 14134 /* Allocate block dependency. */ 14135 retval += 1; 14136 if (!wantcount) 14137 goto out; 14138 } 14139 continue; 14140 14141 case D_FREEBLKS: 14142 freeblks = WK_FREEBLKS(wk); 14143 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 14144 /* Freeblk journal dependency. */ 14145 retval += 1; 14146 if (!wantcount) 14147 goto out; 14148 } 14149 continue; 14150 14151 case D_ALLOCDIRECT: 14152 case D_ALLOCINDIR: 14153 newblk = WK_NEWBLK(wk); 14154 if (newblk->nb_jnewblk) { 14155 /* Journal allocate dependency. */ 14156 retval += 1; 14157 if (!wantcount) 14158 goto out; 14159 } 14160 continue; 14161 14162 case D_MKDIR: 14163 mkdir = WK_MKDIR(wk); 14164 if (mkdir->md_jaddref) { 14165 /* Journal reference dependency. */ 14166 retval += 1; 14167 if (!wantcount) 14168 goto out; 14169 } 14170 continue; 14171 14172 case D_FREEWORK: 14173 case D_FREEDEP: 14174 case D_JSEGDEP: 14175 case D_JSEG: 14176 case D_SBDEP: 14177 /* never a dependency on these blocks */ 14178 continue; 14179 14180 default: 14181 panic("softdep_count_dependencies: Unexpected type %s", 14182 TYPENAME(wk->wk_type)); 14183 /* NOTREACHED */ 14184 } 14185 } 14186 out: 14187 FREE_LOCK(ump); 14188 return (retval); 14189 } 14190 14191 /* 14192 * Acquire exclusive access to a buffer. 14193 * Must be called with a locked mtx parameter. 14194 * Return acquired buffer or NULL on failure. 14195 */ 14196 static struct buf * 14197 getdirtybuf(bp, lock, waitfor) 14198 struct buf *bp; 14199 struct rwlock *lock; 14200 int waitfor; 14201 { 14202 int error; 14203 14204 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 14205 if (waitfor != MNT_WAIT) 14206 return (NULL); 14207 error = BUF_LOCK(bp, 14208 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 14209 /* 14210 * Even if we successfully acquire bp here, we have dropped 14211 * lock, which may violates our guarantee. 14212 */ 14213 if (error == 0) 14214 BUF_UNLOCK(bp); 14215 else if (error != ENOLCK) 14216 panic("getdirtybuf: inconsistent lock: %d", error); 14217 rw_wlock(lock); 14218 return (NULL); 14219 } 14220 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14221 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 14222 rw_wunlock(lock); 14223 BO_LOCK(bp->b_bufobj); 14224 BUF_UNLOCK(bp); 14225 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14226 bp->b_vflags |= BV_BKGRDWAIT; 14227 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 14228 PRIBIO | PDROP, "getbuf", 0); 14229 } else 14230 BO_UNLOCK(bp->b_bufobj); 14231 rw_wlock(lock); 14232 return (NULL); 14233 } 14234 BUF_UNLOCK(bp); 14235 if (waitfor != MNT_WAIT) 14236 return (NULL); 14237 #ifdef DEBUG_VFS_LOCKS 14238 if (bp->b_vp->v_type != VCHR) 14239 ASSERT_BO_WLOCKED(bp->b_bufobj); 14240 #endif 14241 bp->b_vflags |= BV_BKGRDWAIT; 14242 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 14243 return (NULL); 14244 } 14245 if ((bp->b_flags & B_DELWRI) == 0) { 14246 BUF_UNLOCK(bp); 14247 return (NULL); 14248 } 14249 bremfree(bp); 14250 return (bp); 14251 } 14252 14253 14254 /* 14255 * Check if it is safe to suspend the file system now. On entry, 14256 * the vnode interlock for devvp should be held. Return 0 with 14257 * the mount interlock held if the file system can be suspended now, 14258 * otherwise return EAGAIN with the mount interlock held. 14259 */ 14260 int 14261 softdep_check_suspend(struct mount *mp, 14262 struct vnode *devvp, 14263 int softdep_depcnt, 14264 int softdep_accdepcnt, 14265 int secondary_writes, 14266 int secondary_accwrites) 14267 { 14268 struct bufobj *bo; 14269 struct ufsmount *ump; 14270 struct inodedep *inodedep; 14271 int error, unlinked; 14272 14273 bo = &devvp->v_bufobj; 14274 ASSERT_BO_WLOCKED(bo); 14275 14276 /* 14277 * If we are not running with soft updates, then we need only 14278 * deal with secondary writes as we try to suspend. 14279 */ 14280 if (MOUNTEDSOFTDEP(mp) == 0) { 14281 MNT_ILOCK(mp); 14282 while (mp->mnt_secondary_writes != 0) { 14283 BO_UNLOCK(bo); 14284 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 14285 (PUSER - 1) | PDROP, "secwr", 0); 14286 BO_LOCK(bo); 14287 MNT_ILOCK(mp); 14288 } 14289 14290 /* 14291 * Reasons for needing more work before suspend: 14292 * - Dirty buffers on devvp. 14293 * - Secondary writes occurred after start of vnode sync loop 14294 */ 14295 error = 0; 14296 if (bo->bo_numoutput > 0 || 14297 bo->bo_dirty.bv_cnt > 0 || 14298 secondary_writes != 0 || 14299 mp->mnt_secondary_writes != 0 || 14300 secondary_accwrites != mp->mnt_secondary_accwrites) 14301 error = EAGAIN; 14302 BO_UNLOCK(bo); 14303 return (error); 14304 } 14305 14306 /* 14307 * If we are running with soft updates, then we need to coordinate 14308 * with them as we try to suspend. 14309 */ 14310 ump = VFSTOUFS(mp); 14311 for (;;) { 14312 if (!TRY_ACQUIRE_LOCK(ump)) { 14313 BO_UNLOCK(bo); 14314 ACQUIRE_LOCK(ump); 14315 FREE_LOCK(ump); 14316 BO_LOCK(bo); 14317 continue; 14318 } 14319 MNT_ILOCK(mp); 14320 if (mp->mnt_secondary_writes != 0) { 14321 FREE_LOCK(ump); 14322 BO_UNLOCK(bo); 14323 msleep(&mp->mnt_secondary_writes, 14324 MNT_MTX(mp), 14325 (PUSER - 1) | PDROP, "secwr", 0); 14326 BO_LOCK(bo); 14327 continue; 14328 } 14329 break; 14330 } 14331 14332 unlinked = 0; 14333 if (MOUNTEDSUJ(mp)) { 14334 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14335 inodedep != NULL; 14336 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14337 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14338 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14339 UNLINKONLIST) || 14340 !check_inodedep_free(inodedep)) 14341 continue; 14342 unlinked++; 14343 } 14344 } 14345 14346 /* 14347 * Reasons for needing more work before suspend: 14348 * - Dirty buffers on devvp. 14349 * - Softdep activity occurred after start of vnode sync loop 14350 * - Secondary writes occurred after start of vnode sync loop 14351 */ 14352 error = 0; 14353 if (bo->bo_numoutput > 0 || 14354 bo->bo_dirty.bv_cnt > 0 || 14355 softdep_depcnt != unlinked || 14356 ump->softdep_deps != unlinked || 14357 softdep_accdepcnt != ump->softdep_accdeps || 14358 secondary_writes != 0 || 14359 mp->mnt_secondary_writes != 0 || 14360 secondary_accwrites != mp->mnt_secondary_accwrites) 14361 error = EAGAIN; 14362 FREE_LOCK(ump); 14363 BO_UNLOCK(bo); 14364 return (error); 14365 } 14366 14367 14368 /* 14369 * Get the number of dependency structures for the file system, both 14370 * the current number and the total number allocated. These will 14371 * later be used to detect that softdep processing has occurred. 14372 */ 14373 void 14374 softdep_get_depcounts(struct mount *mp, 14375 int *softdep_depsp, 14376 int *softdep_accdepsp) 14377 { 14378 struct ufsmount *ump; 14379 14380 if (MOUNTEDSOFTDEP(mp) == 0) { 14381 *softdep_depsp = 0; 14382 *softdep_accdepsp = 0; 14383 return; 14384 } 14385 ump = VFSTOUFS(mp); 14386 ACQUIRE_LOCK(ump); 14387 *softdep_depsp = ump->softdep_deps; 14388 *softdep_accdepsp = ump->softdep_accdeps; 14389 FREE_LOCK(ump); 14390 } 14391 14392 /* 14393 * Wait for pending output on a vnode to complete. 14394 */ 14395 static void 14396 drain_output(vp) 14397 struct vnode *vp; 14398 { 14399 14400 ASSERT_VOP_LOCKED(vp, "drain_output"); 14401 (void)bufobj_wwait(&vp->v_bufobj, 0, 0); 14402 } 14403 14404 /* 14405 * Called whenever a buffer that is being invalidated or reallocated 14406 * contains dependencies. This should only happen if an I/O error has 14407 * occurred. The routine is called with the buffer locked. 14408 */ 14409 static void 14410 softdep_deallocate_dependencies(bp) 14411 struct buf *bp; 14412 { 14413 14414 if ((bp->b_ioflags & BIO_ERROR) == 0) 14415 panic("softdep_deallocate_dependencies: dangling deps"); 14416 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14417 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14418 else 14419 printf("softdep_deallocate_dependencies: " 14420 "got error %d while accessing filesystem\n", bp->b_error); 14421 if (bp->b_error != ENXIO) 14422 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14423 } 14424 14425 /* 14426 * Function to handle asynchronous write errors in the filesystem. 14427 */ 14428 static void 14429 softdep_error(func, error) 14430 char *func; 14431 int error; 14432 { 14433 14434 /* XXX should do something better! */ 14435 printf("%s: got error %d while accessing filesystem\n", func, error); 14436 } 14437 14438 #ifdef DDB 14439 14440 /* exported to ffs_vfsops.c */ 14441 extern void db_print_ffs(struct ufsmount *ump); 14442 void 14443 db_print_ffs(struct ufsmount *ump) 14444 { 14445 db_printf("mp %p (%s) devvp %p\n", ump->um_mountp, 14446 ump->um_mountp->mnt_stat.f_mntonname, ump->um_devvp); 14447 db_printf(" fs %p su_wl %d su_deps %d su_req %d\n", 14448 ump->um_fs, ump->softdep_on_worklist, 14449 ump->softdep_deps, ump->softdep_req); 14450 } 14451 14452 static void 14453 worklist_print(struct worklist *wk, int verbose) 14454 { 14455 14456 if (!verbose) { 14457 db_printf("%s: %p state 0x%b\n", TYPENAME(wk->wk_type), wk, 14458 (u_int)wk->wk_state, PRINT_SOFTDEP_FLAGS); 14459 return; 14460 } 14461 db_printf("worklist: %p type %s state 0x%b next %p\n ", wk, 14462 TYPENAME(wk->wk_type), (u_int)wk->wk_state, PRINT_SOFTDEP_FLAGS, 14463 LIST_NEXT(wk, wk_list)); 14464 db_print_ffs(VFSTOUFS(wk->wk_mp)); 14465 } 14466 14467 static void 14468 inodedep_print(struct inodedep *inodedep, int verbose) 14469 { 14470 14471 worklist_print(&inodedep->id_list, 0); 14472 db_printf(" fs %p ino %jd inoblk %jd delta %jd nlink %jd\n", 14473 inodedep->id_fs, 14474 (intmax_t)inodedep->id_ino, 14475 (intmax_t)fsbtodb(inodedep->id_fs, 14476 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14477 (intmax_t)inodedep->id_nlinkdelta, 14478 (intmax_t)inodedep->id_savednlink); 14479 14480 if (verbose == 0) 14481 return; 14482 14483 db_printf(" bmsafemap %p, mkdiradd %p, inoreflst %p\n", 14484 inodedep->id_bmsafemap, 14485 inodedep->id_mkdiradd, 14486 TAILQ_FIRST(&inodedep->id_inoreflst)); 14487 db_printf(" dirremhd %p, pendinghd %p, bufwait %p\n", 14488 LIST_FIRST(&inodedep->id_dirremhd), 14489 LIST_FIRST(&inodedep->id_pendinghd), 14490 LIST_FIRST(&inodedep->id_bufwait)); 14491 db_printf(" inowait %p, inoupdt %p, newinoupdt %p\n", 14492 LIST_FIRST(&inodedep->id_inowait), 14493 TAILQ_FIRST(&inodedep->id_inoupdt), 14494 TAILQ_FIRST(&inodedep->id_newinoupdt)); 14495 db_printf(" extupdt %p, newextupdt %p, freeblklst %p\n", 14496 TAILQ_FIRST(&inodedep->id_extupdt), 14497 TAILQ_FIRST(&inodedep->id_newextupdt), 14498 TAILQ_FIRST(&inodedep->id_freeblklst)); 14499 db_printf(" saveino %p, savedsize %jd, savedextsize %jd\n", 14500 inodedep->id_savedino1, 14501 (intmax_t)inodedep->id_savedsize, 14502 (intmax_t)inodedep->id_savedextsize); 14503 } 14504 14505 static void 14506 newblk_print(struct newblk *nbp) 14507 { 14508 14509 worklist_print(&nbp->nb_list, 0); 14510 db_printf(" newblkno %jd\n", (intmax_t)nbp->nb_newblkno); 14511 db_printf(" jnewblk %p, bmsafemap %p, freefrag %p\n", 14512 &nbp->nb_jnewblk, 14513 &nbp->nb_bmsafemap, 14514 &nbp->nb_freefrag); 14515 db_printf(" indirdeps %p, newdirblk %p, jwork %p\n", 14516 LIST_FIRST(&nbp->nb_indirdeps), 14517 LIST_FIRST(&nbp->nb_newdirblk), 14518 LIST_FIRST(&nbp->nb_jwork)); 14519 } 14520 14521 static void 14522 allocdirect_print(struct allocdirect *adp) 14523 { 14524 14525 newblk_print(&adp->ad_block); 14526 db_printf(" oldblkno %jd, oldsize %ld, newsize %ld\n", 14527 adp->ad_oldblkno, adp->ad_oldsize, adp->ad_newsize); 14528 db_printf(" offset %d, inodedep %p\n", 14529 adp->ad_offset, adp->ad_inodedep); 14530 } 14531 14532 static void 14533 allocindir_print(struct allocindir *aip) 14534 { 14535 14536 newblk_print(&aip->ai_block); 14537 db_printf(" oldblkno %jd, lbn %jd\n", 14538 (intmax_t)aip->ai_oldblkno, (intmax_t)aip->ai_lbn); 14539 db_printf(" offset %d, indirdep %p\n", 14540 aip->ai_offset, aip->ai_indirdep); 14541 } 14542 14543 static void 14544 mkdir_print(struct mkdir *mkdir) 14545 { 14546 14547 worklist_print(&mkdir->md_list, 0); 14548 db_printf(" diradd %p, jaddref %p, buf %p\n", 14549 mkdir->md_diradd, mkdir->md_jaddref, mkdir->md_buf); 14550 } 14551 14552 DB_SHOW_COMMAND(sd_inodedep, db_show_sd_inodedep) 14553 { 14554 14555 if (have_addr == 0) { 14556 db_printf("inodedep address required\n"); 14557 return; 14558 } 14559 inodedep_print((struct inodedep*)addr, 1); 14560 } 14561 14562 DB_SHOW_COMMAND(sd_allinodedeps, db_show_sd_allinodedeps) 14563 { 14564 struct inodedep_hashhead *inodedephd; 14565 struct inodedep *inodedep; 14566 struct ufsmount *ump; 14567 int cnt; 14568 14569 if (have_addr == 0) { 14570 db_printf("ufsmount address required\n"); 14571 return; 14572 } 14573 ump = (struct ufsmount *)addr; 14574 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14575 inodedephd = &ump->inodedep_hashtbl[cnt]; 14576 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14577 inodedep_print(inodedep, 0); 14578 } 14579 } 14580 } 14581 14582 DB_SHOW_COMMAND(sd_worklist, db_show_sd_worklist) 14583 { 14584 14585 if (have_addr == 0) { 14586 db_printf("worklist address required\n"); 14587 return; 14588 } 14589 worklist_print((struct worklist *)addr, 1); 14590 } 14591 14592 DB_SHOW_COMMAND(sd_workhead, db_show_sd_workhead) 14593 { 14594 struct worklist *wk; 14595 struct workhead *wkhd; 14596 14597 if (have_addr == 0) { 14598 db_printf("worklist address required " 14599 "(for example value in bp->b_dep)\n"); 14600 return; 14601 } 14602 /* 14603 * We often do not have the address of the worklist head but 14604 * instead a pointer to its first entry (e.g., we have the 14605 * contents of bp->b_dep rather than &bp->b_dep). But the back 14606 * pointer of bp->b_dep will point at the head of the list, so 14607 * we cheat and use that instead. If we are in the middle of 14608 * a list we will still get the same result, so nothing 14609 * unexpected will result. 14610 */ 14611 wk = (struct worklist *)addr; 14612 if (wk == NULL) 14613 return; 14614 wkhd = (struct workhead *)wk->wk_list.le_prev; 14615 LIST_FOREACH(wk, wkhd, wk_list) { 14616 switch(wk->wk_type) { 14617 case D_INODEDEP: 14618 inodedep_print(WK_INODEDEP(wk), 0); 14619 continue; 14620 case D_ALLOCDIRECT: 14621 allocdirect_print(WK_ALLOCDIRECT(wk)); 14622 continue; 14623 case D_ALLOCINDIR: 14624 allocindir_print(WK_ALLOCINDIR(wk)); 14625 continue; 14626 case D_MKDIR: 14627 mkdir_print(WK_MKDIR(wk)); 14628 continue; 14629 default: 14630 worklist_print(wk, 0); 14631 continue; 14632 } 14633 } 14634 } 14635 14636 DB_SHOW_COMMAND(sd_mkdir, db_show_sd_mkdir) 14637 { 14638 if (have_addr == 0) { 14639 db_printf("mkdir address required\n"); 14640 return; 14641 } 14642 mkdir_print((struct mkdir *)addr); 14643 } 14644 14645 DB_SHOW_COMMAND(sd_mkdir_list, db_show_sd_mkdir_list) 14646 { 14647 struct mkdirlist *mkdirlisthd; 14648 struct mkdir *mkdir; 14649 14650 if (have_addr == 0) { 14651 db_printf("mkdir listhead address required\n"); 14652 return; 14653 } 14654 mkdirlisthd = (struct mkdirlist *)addr; 14655 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14656 mkdir_print(mkdir); 14657 if (mkdir->md_diradd != NULL) { 14658 db_printf(" "); 14659 worklist_print(&mkdir->md_diradd->da_list, 0); 14660 } 14661 if (mkdir->md_jaddref != NULL) { 14662 db_printf(" "); 14663 worklist_print(&mkdir->md_jaddref->ja_list, 0); 14664 } 14665 } 14666 } 14667 14668 DB_SHOW_COMMAND(sd_allocdirect, db_show_sd_allocdirect) 14669 { 14670 if (have_addr == 0) { 14671 db_printf("allocdirect address required\n"); 14672 return; 14673 } 14674 allocdirect_print((struct allocdirect *)addr); 14675 } 14676 14677 DB_SHOW_COMMAND(sd_allocindir, db_show_sd_allocindir) 14678 { 14679 if (have_addr == 0) { 14680 db_printf("allocindir address required\n"); 14681 return; 14682 } 14683 allocindir_print((struct allocindir *)addr); 14684 } 14685 14686 #endif /* DDB */ 14687 14688 #endif /* SOFTUPDATES */ 14689