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