1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * $DragonFly: src/sys/vfs/hammer/hammer_ondisk.c,v 1.76 2008/08/29 20:19:08 dillon Exp $ 35 */ 36 /* 37 * Manage HAMMER's on-disk structures. These routines are primarily 38 * responsible for interfacing with the kernel's I/O subsystem and for 39 * managing in-memory structures. 40 */ 41 42 #include <sys/nlookup.h> 43 #include <sys/buf2.h> 44 45 #include "hammer.h" 46 47 static void hammer_free_volume(hammer_volume_t volume); 48 static int hammer_load_volume(hammer_volume_t volume); 49 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew); 50 static int hammer_load_node(hammer_transaction_t trans, 51 hammer_node_t node, int isnew); 52 static void _hammer_rel_node(hammer_node_t node, int locked); 53 54 static int 55 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2) 56 { 57 if (vol1->vol_no < vol2->vol_no) 58 return(-1); 59 if (vol1->vol_no > vol2->vol_no) 60 return(1); 61 return(0); 62 } 63 64 /* 65 * hammer_buffer structures are indexed via their zoneX_offset, not 66 * their zone2_offset. 67 */ 68 static int 69 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2) 70 { 71 if (buf1->zoneX_offset < buf2->zoneX_offset) 72 return(-1); 73 if (buf1->zoneX_offset > buf2->zoneX_offset) 74 return(1); 75 return(0); 76 } 77 78 static int 79 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2) 80 { 81 if (node1->node_offset < node2->node_offset) 82 return(-1); 83 if (node1->node_offset > node2->node_offset) 84 return(1); 85 return(0); 86 } 87 88 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node, 89 hammer_vol_rb_compare, int32_t, vol_no); 90 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node, 91 hammer_buf_rb_compare, hammer_off_t, zoneX_offset); 92 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node, 93 hammer_nod_rb_compare, hammer_off_t, node_offset); 94 95 /************************************************************************ 96 * VOLUMES * 97 ************************************************************************ 98 * 99 * Load a HAMMER volume by name. Returns 0 on success or a positive error 100 * code on failure. Volumes must be loaded at mount time or via hammer 101 * volume-add command, hammer_get_volume() will not load a new volume. 102 * 103 * The passed devvp is vref()'d but not locked. This function consumes the 104 * ref (typically by associating it with the volume structure). 105 * 106 * Calls made to hammer_load_volume() or single-threaded 107 */ 108 int 109 hammer_install_volume(hammer_mount_t hmp, const char *volname, 110 struct vnode *devvp, void *data) 111 { 112 struct mount *mp; 113 hammer_volume_t volume; 114 hammer_volume_ondisk_t ondisk; 115 hammer_volume_ondisk_t img; 116 struct nlookupdata nd; 117 struct buf *bp = NULL; 118 int error; 119 int ronly; 120 int setmp = 0; 121 int i; 122 123 mp = hmp->mp; 124 ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0); 125 126 /* 127 * Allocate a volume structure 128 */ 129 ++hammer_count_volumes; 130 volume = kmalloc(sizeof(*volume), hmp->m_misc, M_WAITOK|M_ZERO); 131 volume->vol_name = kstrdup(volname, hmp->m_misc); 132 volume->io.hmp = hmp; /* bootstrap */ 133 hammer_io_init(&volume->io, volume, HAMMER_IOTYPE_VOLUME); 134 volume->io.offset = 0LL; 135 volume->io.bytes = HAMMER_BUFSIZE; 136 137 /* 138 * Get the device vnode 139 */ 140 if (devvp == NULL) { 141 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW); 142 if (error == 0) 143 error = nlookup(&nd); 144 if (error == 0) 145 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp); 146 nlookup_done(&nd); 147 } else { 148 error = 0; 149 volume->devvp = devvp; 150 } 151 152 if (error == 0) { 153 if (vn_isdisk(volume->devvp, &error)) { 154 error = vfs_mountedon(volume->devvp); 155 } 156 } 157 if (error == 0 && vcount(volume->devvp) > 0) 158 error = EBUSY; 159 if (error == 0) { 160 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY); 161 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0); 162 if (error == 0) { 163 error = VOP_OPEN(volume->devvp, 164 (ronly ? FREAD : FREAD|FWRITE), 165 FSCRED, NULL); 166 } 167 vn_unlock(volume->devvp); 168 } 169 if (error) { 170 hammer_free_volume(volume); 171 return(error); 172 } 173 volume->devvp->v_rdev->si_mountpoint = mp; 174 setmp = 1; 175 176 /* 177 * Extract the volume number from the volume header and do various 178 * sanity checks. 179 */ 180 error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp); 181 if (error) 182 goto late_failure; 183 ondisk = (void *)bp->b_data; 184 185 /* 186 * Initialize the volume header with data if the data is specified. 187 */ 188 if (ronly == 0 && data) { 189 img = (hammer_volume_ondisk_t)data; 190 if (ondisk->vol_signature == HAMMER_FSBUF_VOLUME) { 191 hkprintf("Formatting of valid HAMMER volume %s denied. " 192 "Erase with hammer strip or dd!\n", volname); 193 error = EFTYPE; 194 goto late_failure; 195 } 196 bcopy(img, ondisk, sizeof(*img)); 197 } 198 199 if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) { 200 hkprintf("volume %s has an invalid header\n", volume->vol_name); 201 for (i = 0; i < (int)sizeof(ondisk->vol_signature); i++) { 202 kprintf("%02x", ((char*)&ondisk->vol_signature)[i] & 0xFF); 203 if (i != (int)sizeof(ondisk->vol_signature) - 1) 204 kprintf(" "); 205 } 206 kprintf("\n"); 207 error = EFTYPE; 208 goto late_failure; 209 } 210 volume->vol_no = ondisk->vol_no; 211 volume->vol_flags = ondisk->vol_flags; 212 volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no, 213 HAMMER_VOL_BUF_SIZE(ondisk)); 214 215 if (RB_EMPTY(&hmp->rb_vols_root)) { 216 hmp->fsid = ondisk->vol_fsid; 217 } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) { 218 hkprintf("volume %s's fsid does not match other volumes\n", 219 volume->vol_name); 220 error = EFTYPE; 221 goto late_failure; 222 } 223 224 /* 225 * Insert the volume structure into the red-black tree. 226 */ 227 if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) { 228 hkprintf("volume %s has a duplicate vol_no %d\n", 229 volume->vol_name, volume->vol_no); 230 error = EEXIST; 231 } 232 233 if (error == 0) 234 hammer_volume_number_add(hmp, volume); 235 236 /* 237 * Set the root volume . HAMMER special cases rootvol the structure. 238 * We do not hold a ref because this would prevent related I/O 239 * from being flushed. 240 */ 241 if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) { 242 hmp->rootvol = volume; 243 hmp->nvolumes = ondisk->vol_count; 244 if (bp) { 245 brelse(bp); 246 bp = NULL; 247 } 248 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks * 249 HAMMER_BUFFERS_PER_BIGBLOCK; 250 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks * 251 HAMMER_BUFFERS_PER_BIGBLOCK; 252 } 253 late_failure: 254 if (bp) 255 brelse(bp); 256 if (error) { 257 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/ 258 if (setmp) 259 volume->devvp->v_rdev->si_mountpoint = NULL; 260 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY); 261 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE, NULL); 262 vn_unlock(volume->devvp); 263 hammer_free_volume(volume); 264 } 265 return (error); 266 } 267 268 /* 269 * This is called for each volume when updating the mount point from 270 * read-write to read-only or vise-versa. 271 */ 272 int 273 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused) 274 { 275 if (volume->devvp) { 276 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY); 277 if (volume->io.hmp->ronly) { 278 /* do not call vinvalbuf */ 279 VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL); 280 VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL); 281 } else { 282 /* do not call vinvalbuf */ 283 VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL); 284 VOP_CLOSE(volume->devvp, FREAD, NULL); 285 } 286 vn_unlock(volume->devvp); 287 } 288 return(0); 289 } 290 291 /* 292 * Unload and free a HAMMER volume. Must return >= 0 to continue scan 293 * so returns -1 on failure. 294 */ 295 int 296 hammer_unload_volume(hammer_volume_t volume, void *data) 297 { 298 hammer_mount_t hmp = volume->io.hmp; 299 struct buf *bp = NULL; 300 hammer_volume_ondisk_t img; 301 int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0); 302 int error; 303 304 /* 305 * Clear the volume header with data if the data is specified. 306 */ 307 if (ronly == 0 && data && volume->devvp) { 308 img = (hammer_volume_ondisk_t)data; 309 error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp); 310 if (error || bp->b_bcount < sizeof(*img)) { 311 hmkprintf(hmp, "Failed to read volume header: %d\n", error); 312 brelse(bp); 313 } else { 314 bcopy(img, bp->b_data, sizeof(*img)); 315 error = bwrite(bp); 316 if (error) 317 hmkprintf(hmp, "Failed to clear volume header: %d\n", 318 error); 319 } 320 } 321 322 /* 323 * Clean up the root volume pointer, which is held unlocked in hmp. 324 */ 325 if (hmp->rootvol == volume) 326 hmp->rootvol = NULL; 327 328 /* 329 * We must not flush a dirty buffer to disk on umount. It should 330 * have already been dealt with by the flusher, or we may be in 331 * catastrophic failure. 332 */ 333 hammer_io_clear_modify(&volume->io, 1); 334 volume->io.waitdep = 1; 335 336 /* 337 * Clean up the persistent ref ioerror might have on the volume 338 */ 339 if (volume->io.ioerror) 340 hammer_io_clear_error_noassert(&volume->io); 341 342 /* 343 * This should release the bp. Releasing the volume with flush set 344 * implies the interlock is set. 345 */ 346 hammer_ref_interlock_true(&volume->io.lock); 347 hammer_rel_volume(volume, 1); 348 KKASSERT(volume->io.bp == NULL); 349 350 /* 351 * There should be no references on the volume. 352 */ 353 KKASSERT(hammer_norefs(&volume->io.lock)); 354 355 volume->ondisk = NULL; 356 if (volume->devvp) { 357 if (volume->devvp->v_rdev && 358 volume->devvp->v_rdev->si_mountpoint == hmp->mp) { 359 volume->devvp->v_rdev->si_mountpoint = NULL; 360 } 361 if (ronly) { 362 /* 363 * Make sure we don't sync anything to disk if we 364 * are in read-only mode (1) or critically-errored 365 * (2). Note that there may be dirty buffers in 366 * normal read-only mode from crash recovery. 367 */ 368 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY); 369 vinvalbuf(volume->devvp, 0, 0, 0); 370 VOP_CLOSE(volume->devvp, FREAD, NULL); 371 vn_unlock(volume->devvp); 372 } else { 373 /* 374 * Normal termination, save any dirty buffers 375 * (XXX there really shouldn't be any). 376 */ 377 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY); 378 vinvalbuf(volume->devvp, V_SAVE, 0, 0); 379 VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL); 380 vn_unlock(volume->devvp); 381 } 382 } 383 384 /* 385 * Destroy the structure 386 */ 387 RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume); 388 hammer_volume_number_del(hmp, volume); 389 hammer_free_volume(volume); 390 return(0); 391 } 392 393 static 394 void 395 hammer_free_volume(hammer_volume_t volume) 396 { 397 hammer_mount_t hmp = volume->io.hmp; 398 399 if (volume->vol_name) { 400 kfree(volume->vol_name, hmp->m_misc); 401 volume->vol_name = NULL; 402 } 403 if (volume->devvp) { 404 vrele(volume->devvp); 405 volume->devvp = NULL; 406 } 407 --hammer_count_volumes; 408 kfree(volume, hmp->m_misc); 409 } 410 411 /* 412 * Get a HAMMER volume. The volume must already exist. 413 */ 414 hammer_volume_t 415 hammer_get_volume(hammer_mount_t hmp, int32_t vol_no, int *errorp) 416 { 417 hammer_volume_t volume; 418 419 /* 420 * Locate the volume structure 421 */ 422 volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no); 423 if (volume == NULL) { 424 *errorp = ENOENT; 425 return(NULL); 426 } 427 428 /* 429 * Reference the volume, load/check the data on the 0->1 transition. 430 * hammer_load_volume() will dispose of the interlock on return, 431 * and also clean up the ref count on error. 432 */ 433 if (hammer_ref_interlock(&volume->io.lock)) { 434 *errorp = hammer_load_volume(volume); 435 if (*errorp) 436 volume = NULL; 437 } else { 438 KKASSERT(volume->ondisk); 439 *errorp = 0; 440 } 441 return(volume); 442 } 443 444 int 445 hammer_ref_volume(hammer_volume_t volume) 446 { 447 int error; 448 449 /* 450 * Reference the volume and deal with the check condition used to 451 * load its ondisk info. 452 */ 453 if (hammer_ref_interlock(&volume->io.lock)) { 454 error = hammer_load_volume(volume); 455 } else { 456 KKASSERT(volume->ondisk); 457 error = 0; 458 } 459 return (error); 460 } 461 462 /* 463 * May be called without fs_token 464 */ 465 hammer_volume_t 466 hammer_get_root_volume(hammer_mount_t hmp, int *errorp) 467 { 468 hammer_volume_t volume; 469 470 volume = hmp->rootvol; 471 KKASSERT(volume != NULL); 472 473 /* 474 * Reference the volume and deal with the check condition used to 475 * load its ondisk info. 476 */ 477 if (hammer_ref_interlock(&volume->io.lock)) { 478 lwkt_gettoken(&volume->io.hmp->fs_token); 479 *errorp = hammer_load_volume(volume); 480 lwkt_reltoken(&volume->io.hmp->fs_token); 481 if (*errorp) 482 volume = NULL; 483 } else { 484 KKASSERT(volume->ondisk); 485 *errorp = 0; 486 } 487 return (volume); 488 } 489 490 /* 491 * Load a volume's on-disk information. The volume must be referenced and 492 * the interlock is held on call. The interlock will be released on return. 493 * The reference will also be released on return if an error occurs. 494 */ 495 static int 496 hammer_load_volume(hammer_volume_t volume) 497 { 498 int error; 499 500 if (volume->ondisk == NULL) { 501 error = hammer_io_read(volume->devvp, &volume->io, 502 HAMMER_BUFSIZE); 503 if (error == 0) { 504 volume->ondisk = (void *)volume->io.bp->b_data; 505 hammer_ref_interlock_done(&volume->io.lock); 506 } else { 507 hammer_rel_volume(volume, 1); 508 } 509 } else { 510 error = 0; 511 } 512 return(error); 513 } 514 515 /* 516 * Release a previously acquired reference on the volume. 517 * 518 * Volumes are not unloaded from memory during normal operation. 519 * 520 * May be called without fs_token 521 */ 522 void 523 hammer_rel_volume(hammer_volume_t volume, int locked) 524 { 525 struct buf *bp; 526 527 if (hammer_rel_interlock(&volume->io.lock, locked)) { 528 lwkt_gettoken(&volume->io.hmp->fs_token); 529 volume->ondisk = NULL; 530 bp = hammer_io_release(&volume->io, locked); 531 lwkt_reltoken(&volume->io.hmp->fs_token); 532 hammer_rel_interlock_done(&volume->io.lock, locked); 533 if (bp) 534 brelse(bp); 535 } 536 } 537 538 int 539 hammer_mountcheck_volumes(hammer_mount_t hmp) 540 { 541 hammer_volume_t vol; 542 int i; 543 544 HAMMER_VOLUME_NUMBER_FOREACH(hmp, i) { 545 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i); 546 if (vol == NULL) 547 return(EINVAL); 548 } 549 return(0); 550 } 551 552 int 553 hammer_get_installed_volumes(hammer_mount_t hmp) 554 { 555 int i, ret = 0; 556 557 HAMMER_VOLUME_NUMBER_FOREACH(hmp, i) 558 ret++; 559 return(ret); 560 } 561 562 /************************************************************************ 563 * BUFFERS * 564 ************************************************************************ 565 * 566 * Manage buffers. Currently most blockmap-backed zones are direct-mapped 567 * to zone-2 buffer offsets, without a translation stage. However, the 568 * hammer_buffer structure is indexed by its zoneX_offset, not its 569 * zone2_offset. 570 * 571 * The proper zone must be maintained throughout the code-base all the way 572 * through to the big-block allocator, or routines like hammer_del_buffers() 573 * will not be able to locate all potentially conflicting buffers. 574 */ 575 576 /* 577 * Helper function returns whether a zone offset can be directly translated 578 * to a raw buffer index or not. Really only the volume and undo zones 579 * can't be directly translated. Volumes are special-cased and undo zones 580 * shouldn't be aliased accessed in read-only mode. 581 * 582 * This function is ONLY used to detect aliased zones during a read-only 583 * mount. 584 */ 585 static __inline int 586 hammer_direct_zone(hammer_off_t buf_offset) 587 { 588 return(hammer_is_zone_direct_xlated(buf_offset)); 589 } 590 591 hammer_buffer_t 592 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset, 593 int bytes, int isnew, int *errorp) 594 { 595 hammer_buffer_t buffer; 596 hammer_volume_t volume; 597 hammer_off_t zone2_offset; 598 int vol_no; 599 int zone; 600 601 buf_offset &= ~HAMMER_BUFMASK64; 602 again: 603 /* 604 * Shortcut if the buffer is already cached 605 */ 606 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset); 607 if (buffer) { 608 /* 609 * Once refed the ondisk field will not be cleared by 610 * any other action. Shortcut the operation if the 611 * ondisk structure is valid. 612 */ 613 found_aliased: 614 if (hammer_ref_interlock(&buffer->io.lock) == 0) { 615 hammer_io_advance(&buffer->io); 616 KKASSERT(buffer->ondisk); 617 *errorp = 0; 618 return(buffer); 619 } 620 621 /* 622 * 0->1 transition or defered 0->1 transition (CHECK), 623 * interlock now held. Shortcut if ondisk is already 624 * assigned. 625 */ 626 atomic_add_int(&hammer_count_refedbufs, 1); 627 if (buffer->ondisk) { 628 hammer_io_advance(&buffer->io); 629 hammer_ref_interlock_done(&buffer->io.lock); 630 *errorp = 0; 631 return(buffer); 632 } 633 634 /* 635 * The buffer is no longer loose if it has a ref, and 636 * cannot become loose once it gains a ref. Loose 637 * buffers will never be in a modified state. This should 638 * only occur on the 0->1 transition of refs. 639 * 640 * lose_root can be modified via a biodone() interrupt 641 * so the io_token must be held. 642 */ 643 if (buffer->io.mod_root == &hmp->lose_root) { 644 lwkt_gettoken(&hmp->io_token); 645 if (buffer->io.mod_root == &hmp->lose_root) { 646 RB_REMOVE(hammer_mod_rb_tree, 647 buffer->io.mod_root, &buffer->io); 648 buffer->io.mod_root = NULL; 649 KKASSERT(buffer->io.modified == 0); 650 } 651 lwkt_reltoken(&hmp->io_token); 652 } 653 goto found; 654 } else if (hmp->ronly && hammer_direct_zone(buf_offset)) { 655 /* 656 * If this is a read-only mount there could be an alias 657 * in the raw-zone. If there is we use that buffer instead. 658 * 659 * rw mounts will not have aliases. Also note when going 660 * from ro -> rw the recovered raw buffers are flushed and 661 * reclaimed, so again there will not be any aliases once 662 * the mount is rw. 663 */ 664 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, 665 hammer_xlate_to_zone2(buf_offset)); 666 if (buffer) { 667 if (hammer_debug_general & 0x0001) { 668 hkrateprintf(&hmp->kdiag, 669 "recovered aliased %016jx\n", 670 (intmax_t)buf_offset); 671 } 672 goto found_aliased; 673 } 674 } 675 676 /* 677 * Handle blockmap offset translations 678 */ 679 zone = HAMMER_ZONE_DECODE(buf_offset); 680 if (hammer_is_index_record(zone)) { 681 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp); 682 } else if (zone == HAMMER_ZONE_UNDO_INDEX) { 683 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp); 684 } else { 685 /* Must be zone-2 (not 1 or 4 or 15) */ 686 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX); 687 zone2_offset = buf_offset; 688 *errorp = 0; 689 } 690 if (*errorp) 691 return(NULL); 692 693 /* 694 * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset 695 * specifications. 696 */ 697 KKASSERT(hammer_is_zone_raw_buffer(zone2_offset)); 698 vol_no = HAMMER_VOL_DECODE(zone2_offset); 699 volume = hammer_get_volume(hmp, vol_no, errorp); 700 if (volume == NULL) 701 return(NULL); 702 703 KKASSERT(zone2_offset < volume->maxbuf_off); 704 705 /* 706 * Allocate a new buffer structure. We will check for races later. 707 */ 708 ++hammer_count_buffers; 709 buffer = kmalloc(sizeof(*buffer), hmp->m_misc, 710 M_WAITOK|M_ZERO|M_USE_RESERVE); 711 buffer->zone2_offset = zone2_offset; 712 buffer->zoneX_offset = buf_offset; 713 714 hammer_io_init(&buffer->io, volume, hammer_zone_to_iotype(zone)); 715 buffer->io.offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset); 716 buffer->io.bytes = bytes; 717 TAILQ_INIT(&buffer->node_list); 718 hammer_ref_interlock_true(&buffer->io.lock); 719 720 /* 721 * Insert the buffer into the RB tree and handle late collisions. 722 */ 723 if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) { 724 hammer_rel_volume(volume, 0); 725 buffer->io.volume = NULL; /* safety */ 726 if (hammer_rel_interlock(&buffer->io.lock, 1)) /* safety */ 727 hammer_rel_interlock_done(&buffer->io.lock, 1); 728 --hammer_count_buffers; 729 kfree(buffer, hmp->m_misc); 730 goto again; 731 } 732 atomic_add_int(&hammer_count_refedbufs, 1); 733 found: 734 735 /* 736 * The buffer is referenced and interlocked. Load the buffer 737 * if necessary. hammer_load_buffer() deals with the interlock 738 * and, if an error is returned, also deals with the ref. 739 */ 740 if (buffer->ondisk == NULL) { 741 *errorp = hammer_load_buffer(buffer, isnew); 742 if (*errorp) 743 buffer = NULL; 744 } else { 745 hammer_io_advance(&buffer->io); 746 hammer_ref_interlock_done(&buffer->io.lock); 747 *errorp = 0; 748 } 749 return(buffer); 750 } 751 752 /* 753 * This is used by the direct-read code to deal with large-data buffers 754 * created by the reblocker and mirror-write code. The direct-read code 755 * bypasses the HAMMER buffer subsystem and so any aliased dirty or write- 756 * running hammer buffers must be fully synced to disk before we can issue 757 * the direct-read. 758 * 759 * This code path is not considered critical as only the rebocker and 760 * mirror-write code will create large-data buffers via the HAMMER buffer 761 * subsystem. They do that because they operate at the B-Tree level and 762 * do not access the vnode/inode structures. 763 */ 764 void 765 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes) 766 { 767 hammer_buffer_t buffer; 768 int error; 769 770 KKASSERT(hammer_is_zone_large_data(base_offset)); 771 772 while (bytes > 0) { 773 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, 774 base_offset); 775 if (buffer && (buffer->io.modified || buffer->io.running)) { 776 error = hammer_ref_buffer(buffer); 777 if (error == 0) { 778 hammer_io_wait(&buffer->io); 779 if (buffer->io.modified) { 780 hammer_io_write_interlock(&buffer->io); 781 hammer_io_flush(&buffer->io, 0); 782 hammer_io_done_interlock(&buffer->io); 783 hammer_io_wait(&buffer->io); 784 } 785 hammer_rel_buffer(buffer, 0); 786 } 787 } 788 base_offset += HAMMER_BUFSIZE; 789 bytes -= HAMMER_BUFSIZE; 790 } 791 } 792 793 /* 794 * Destroy all buffers covering the specified zoneX offset range. This 795 * is called when the related blockmap layer2 entry is freed or when 796 * a direct write bypasses our buffer/buffer-cache subsystem. 797 * 798 * The buffers may be referenced by the caller itself. Setting reclaim 799 * will cause the buffer to be destroyed when it's ref count reaches zero. 800 * 801 * Return 0 on success, EAGAIN if some buffers could not be destroyed due 802 * to additional references held by other threads, or some other (typically 803 * fatal) error. 804 */ 805 int 806 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset, 807 hammer_off_t zone2_offset, int bytes, 808 int report_conflicts) 809 { 810 hammer_buffer_t buffer; 811 hammer_volume_t volume; 812 int vol_no; 813 int error; 814 int ret_error; 815 816 vol_no = HAMMER_VOL_DECODE(zone2_offset); 817 volume = hammer_get_volume(hmp, vol_no, &ret_error); 818 KKASSERT(ret_error == 0); 819 820 while (bytes > 0) { 821 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, 822 base_offset); 823 if (buffer) { 824 error = hammer_ref_buffer(buffer); 825 if (hammer_debug_general & 0x20000) { 826 hkprintf("delbufr %016jx rerr=%d 1ref=%d\n", 827 (intmax_t)buffer->zoneX_offset, 828 error, 829 hammer_oneref(&buffer->io.lock)); 830 } 831 if (error == 0 && !hammer_oneref(&buffer->io.lock)) { 832 error = EAGAIN; 833 hammer_rel_buffer(buffer, 0); 834 } 835 if (error == 0) { 836 KKASSERT(buffer->zone2_offset == zone2_offset); 837 hammer_io_clear_modify(&buffer->io, 1); 838 buffer->io.reclaim = 1; 839 buffer->io.waitdep = 1; 840 KKASSERT(buffer->io.volume == volume); 841 hammer_rel_buffer(buffer, 0); 842 } 843 } else { 844 error = hammer_io_inval(volume, zone2_offset); 845 } 846 if (error) { 847 ret_error = error; 848 if (report_conflicts || 849 (hammer_debug_general & 0x8000)) { 850 krateprintf(&hmp->kdiag, 851 "hammer_del_buffers: unable to " 852 "invalidate %016jx buffer=%p " 853 "rep=%d lkrefs=%08x\n", 854 (intmax_t)base_offset, 855 buffer, report_conflicts, 856 (buffer ? buffer->io.lock.refs : -1)); 857 } 858 } 859 base_offset += HAMMER_BUFSIZE; 860 zone2_offset += HAMMER_BUFSIZE; 861 bytes -= HAMMER_BUFSIZE; 862 } 863 hammer_rel_volume(volume, 0); 864 return (ret_error); 865 } 866 867 /* 868 * Given a referenced and interlocked buffer load/validate the data. 869 * 870 * The buffer interlock will be released on return. If an error is 871 * returned the buffer reference will also be released (and the buffer 872 * pointer will thus be stale). 873 */ 874 static int 875 hammer_load_buffer(hammer_buffer_t buffer, int isnew) 876 { 877 hammer_volume_t volume; 878 int error; 879 880 /* 881 * Load the buffer's on-disk info 882 */ 883 volume = buffer->io.volume; 884 885 if (hammer_debug_io & 0x0004) { 886 hdkprintf("load_buffer %016jx %016jx isnew=%d od=%p\n", 887 (intmax_t)buffer->zoneX_offset, 888 (intmax_t)buffer->zone2_offset, 889 isnew, buffer->ondisk); 890 } 891 892 if (buffer->ondisk == NULL) { 893 /* 894 * Issue the read or generate a new buffer. When reading 895 * the limit argument controls any read-ahead clustering 896 * hammer_io_read() is allowed to do. 897 * 898 * We cannot read-ahead in the large-data zone and we cannot 899 * cross a big-block boundary as the next big-block might 900 * use a different buffer size. 901 */ 902 if (isnew) { 903 error = hammer_io_new(volume->devvp, &buffer->io); 904 } else if (hammer_is_zone_large_data(buffer->zoneX_offset)) { 905 error = hammer_io_read(volume->devvp, &buffer->io, 906 buffer->io.bytes); 907 } else { 908 hammer_off_t limit; 909 910 limit = HAMMER_BIGBLOCK_DOALIGN(buffer->zone2_offset); 911 limit -= buffer->zone2_offset; 912 error = hammer_io_read(volume->devvp, &buffer->io, 913 limit); 914 } 915 if (error == 0) 916 buffer->ondisk = (void *)buffer->io.bp->b_data; 917 } else if (isnew) { 918 error = hammer_io_new(volume->devvp, &buffer->io); 919 } else { 920 error = 0; 921 } 922 if (error == 0) { 923 hammer_io_advance(&buffer->io); 924 hammer_ref_interlock_done(&buffer->io.lock); 925 } else { 926 hammer_rel_buffer(buffer, 1); 927 } 928 return (error); 929 } 930 931 /* 932 * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue. 933 * This routine is only called during unmount or when a volume is 934 * removed. 935 * 936 * If data != NULL, it specifies a volume whoose buffers should 937 * be unloaded. 938 */ 939 int 940 hammer_unload_buffer(hammer_buffer_t buffer, void *data) 941 { 942 hammer_volume_t volume = (hammer_volume_t)data; 943 944 /* 945 * If volume != NULL we are only interested in unloading buffers 946 * associated with a particular volume. 947 */ 948 if (volume != NULL && volume != buffer->io.volume) 949 return 0; 950 951 /* 952 * Clean up the persistent ref ioerror might have on the buffer 953 * and acquire a ref. Expect a 0->1 transition. 954 */ 955 if (buffer->io.ioerror) { 956 hammer_io_clear_error_noassert(&buffer->io); 957 atomic_add_int(&hammer_count_refedbufs, -1); 958 } 959 hammer_ref_interlock_true(&buffer->io.lock); 960 atomic_add_int(&hammer_count_refedbufs, 1); 961 962 /* 963 * We must not flush a dirty buffer to disk on umount. It should 964 * have already been dealt with by the flusher, or we may be in 965 * catastrophic failure. 966 * 967 * We must set waitdep to ensure that a running buffer is waited 968 * on and released prior to us trying to unload the volume. 969 */ 970 hammer_io_clear_modify(&buffer->io, 1); 971 hammer_flush_buffer_nodes(buffer); 972 buffer->io.waitdep = 1; 973 hammer_rel_buffer(buffer, 1); 974 return(0); 975 } 976 977 /* 978 * Reference a buffer that is either already referenced or via a specially 979 * handled pointer (aka cursor->buffer). 980 */ 981 int 982 hammer_ref_buffer(hammer_buffer_t buffer) 983 { 984 hammer_mount_t hmp; 985 int error; 986 int locked; 987 988 /* 989 * Acquire a ref, plus the buffer will be interlocked on the 990 * 0->1 transition. 991 */ 992 locked = hammer_ref_interlock(&buffer->io.lock); 993 hmp = buffer->io.hmp; 994 995 /* 996 * At this point a biodone() will not touch the buffer other then 997 * incidental bits. However, lose_root can be modified via 998 * a biodone() interrupt. 999 * 1000 * No longer loose. lose_root requires the io_token. 1001 */ 1002 if (buffer->io.mod_root == &hmp->lose_root) { 1003 lwkt_gettoken(&hmp->io_token); 1004 if (buffer->io.mod_root == &hmp->lose_root) { 1005 RB_REMOVE(hammer_mod_rb_tree, 1006 buffer->io.mod_root, &buffer->io); 1007 buffer->io.mod_root = NULL; 1008 } 1009 lwkt_reltoken(&hmp->io_token); 1010 } 1011 1012 if (locked) { 1013 atomic_add_int(&hammer_count_refedbufs, 1); 1014 error = hammer_load_buffer(buffer, 0); 1015 /* NOTE: on error the buffer pointer is stale */ 1016 } else { 1017 error = 0; 1018 } 1019 return(error); 1020 } 1021 1022 /* 1023 * Release a reference on the buffer. On the 1->0 transition the 1024 * underlying IO will be released but the data reference is left 1025 * cached. 1026 * 1027 * Only destroy the structure itself if the related buffer cache buffer 1028 * was disassociated from it. This ties the management of the structure 1029 * to the buffer cache subsystem. buffer->ondisk determines whether the 1030 * embedded io is referenced or not. 1031 */ 1032 void 1033 hammer_rel_buffer(hammer_buffer_t buffer, int locked) 1034 { 1035 hammer_volume_t volume; 1036 hammer_mount_t hmp; 1037 struct buf *bp = NULL; 1038 int freeme = 0; 1039 1040 hmp = buffer->io.hmp; 1041 1042 if (hammer_rel_interlock(&buffer->io.lock, locked) == 0) 1043 return; 1044 1045 /* 1046 * hammer_count_refedbufs accounting. Decrement if we are in 1047 * the error path or if CHECK is clear. 1048 * 1049 * If we are not in the error path and CHECK is set the caller 1050 * probably just did a hammer_ref() and didn't account for it, 1051 * so we don't account for the loss here. 1052 */ 1053 if (locked || (buffer->io.lock.refs & HAMMER_REFS_CHECK) == 0) 1054 atomic_add_int(&hammer_count_refedbufs, -1); 1055 1056 /* 1057 * If the caller locked us or the normal released transitions 1058 * from 1->0 (and acquired the lock) attempt to release the 1059 * io. If the called locked us we tell hammer_io_release() 1060 * to flush (which would be the unload or failure path). 1061 */ 1062 bp = hammer_io_release(&buffer->io, locked); 1063 1064 /* 1065 * If the buffer has no bp association and no refs we can destroy 1066 * it. 1067 * 1068 * NOTE: It is impossible for any associated B-Tree nodes to have 1069 * refs if the buffer has no additional refs. 1070 */ 1071 if (buffer->io.bp == NULL && hammer_norefs(&buffer->io.lock)) { 1072 RB_REMOVE(hammer_buf_rb_tree, 1073 &buffer->io.hmp->rb_bufs_root, 1074 buffer); 1075 volume = buffer->io.volume; 1076 buffer->io.volume = NULL; /* sanity */ 1077 hammer_rel_volume(volume, 0); 1078 hammer_io_clear_modlist(&buffer->io); 1079 hammer_flush_buffer_nodes(buffer); 1080 KKASSERT(TAILQ_EMPTY(&buffer->node_list)); 1081 freeme = 1; 1082 } 1083 1084 /* 1085 * Cleanup 1086 */ 1087 hammer_rel_interlock_done(&buffer->io.lock, locked); 1088 if (bp) 1089 brelse(bp); 1090 if (freeme) { 1091 --hammer_count_buffers; 1092 kfree(buffer, hmp->m_misc); 1093 } 1094 } 1095 1096 /* 1097 * Access the filesystem buffer containing the specified hammer offset. 1098 * buf_offset is a conglomeration of the volume number and vol_buf_beg 1099 * relative buffer offset. It must also have bit 55 set to be valid. 1100 * (see hammer_off_t in hammer_disk.h). 1101 * 1102 * Any prior buffer in *bufferp will be released and replaced by the 1103 * requested buffer. 1104 * 1105 * NOTE: The buffer is indexed via its zoneX_offset but we allow the 1106 * passed cached *bufferp to match against either zoneX or zone2. 1107 */ 1108 static __inline 1109 void * 1110 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes, 1111 int isnew, int *errorp, hammer_buffer_t *bufferp) 1112 { 1113 hammer_buffer_t buffer; 1114 int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK; 1115 1116 buf_offset &= ~HAMMER_BUFMASK64; 1117 KKASSERT(HAMMER_ZONE(buf_offset) != 0); 1118 1119 buffer = *bufferp; 1120 if (buffer == NULL || (buffer->zone2_offset != buf_offset && 1121 buffer->zoneX_offset != buf_offset)) { 1122 if (buffer) 1123 hammer_rel_buffer(buffer, 0); 1124 buffer = hammer_get_buffer(hmp, buf_offset, bytes, isnew, errorp); 1125 *bufferp = buffer; 1126 } else { 1127 *errorp = 0; 1128 } 1129 1130 /* 1131 * Return a pointer to the buffer data. 1132 */ 1133 if (buffer == NULL) 1134 return(NULL); 1135 else 1136 return((char *)buffer->ondisk + xoff); 1137 } 1138 1139 void * 1140 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, 1141 int *errorp, hammer_buffer_t *bufferp) 1142 { 1143 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, 0, errorp, bufferp)); 1144 } 1145 1146 void * 1147 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes, 1148 int *errorp, hammer_buffer_t *bufferp) 1149 { 1150 bytes = HAMMER_BUFSIZE_DOALIGN(bytes); 1151 return(_hammer_bread(hmp, buf_offset, bytes, 0, errorp, bufferp)); 1152 } 1153 1154 /* 1155 * Access the filesystem buffer containing the specified hammer offset. 1156 * No disk read operation occurs. The result buffer may contain garbage. 1157 * 1158 * Any prior buffer in *bufferp will be released and replaced by the 1159 * requested buffer. 1160 * 1161 * This function marks the buffer dirty but does not increment its 1162 * modify_refs count. 1163 */ 1164 void * 1165 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, 1166 int *errorp, hammer_buffer_t *bufferp) 1167 { 1168 return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, 1, errorp, bufferp)); 1169 } 1170 1171 void * 1172 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes, 1173 int *errorp, hammer_buffer_t *bufferp) 1174 { 1175 bytes = HAMMER_BUFSIZE_DOALIGN(bytes); 1176 return(_hammer_bread(hmp, buf_offset, bytes, 1, errorp, bufferp)); 1177 } 1178 1179 /************************************************************************ 1180 * NODES * 1181 ************************************************************************ 1182 * 1183 * Manage B-Tree nodes. B-Tree nodes represent the primary indexing 1184 * method used by the HAMMER filesystem. 1185 * 1186 * Unlike other HAMMER structures, a hammer_node can be PASSIVELY 1187 * associated with its buffer, and will only referenced the buffer while 1188 * the node itself is referenced. 1189 * 1190 * A hammer_node can also be passively associated with other HAMMER 1191 * structures, such as inodes, while retaining 0 references. These 1192 * associations can be cleared backwards using a pointer-to-pointer in 1193 * the hammer_node. 1194 * 1195 * This allows the HAMMER implementation to cache hammer_nodes long-term 1196 * and short-cut a great deal of the infrastructure's complexity. In 1197 * most cases a cached node can be reacquired without having to dip into 1198 * the B-Tree. 1199 */ 1200 hammer_node_t 1201 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset, 1202 int isnew, int *errorp) 1203 { 1204 hammer_mount_t hmp = trans->hmp; 1205 hammer_node_t node; 1206 int doload; 1207 1208 KKASSERT(hammer_is_zone_btree(node_offset)); 1209 1210 /* 1211 * Locate the structure, allocating one if necessary. 1212 */ 1213 again: 1214 node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset); 1215 if (node == NULL) { 1216 ++hammer_count_nodes; 1217 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE); 1218 node->node_offset = node_offset; 1219 node->hmp = hmp; 1220 TAILQ_INIT(&node->cursor_list); 1221 TAILQ_INIT(&node->cache_list); 1222 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) { 1223 --hammer_count_nodes; 1224 kfree(node, hmp->m_misc); 1225 goto again; 1226 } 1227 doload = hammer_ref_interlock_true(&node->lock); 1228 } else { 1229 doload = hammer_ref_interlock(&node->lock); 1230 } 1231 if (doload) { 1232 *errorp = hammer_load_node(trans, node, isnew); 1233 if (*errorp) 1234 node = NULL; 1235 } else { 1236 KKASSERT(node->ondisk); 1237 *errorp = 0; 1238 hammer_io_advance(&node->buffer->io); 1239 } 1240 return(node); 1241 } 1242 1243 /* 1244 * Reference an already-referenced node. 0->1 transitions should assert 1245 * so we do not have to deal with hammer_ref() setting CHECK. 1246 */ 1247 void 1248 hammer_ref_node(hammer_node_t node) 1249 { 1250 KKASSERT(hammer_isactive(&node->lock) && node->ondisk != NULL); 1251 hammer_ref(&node->lock); 1252 } 1253 1254 /* 1255 * Load a node's on-disk data reference. Called with the node referenced 1256 * and interlocked. 1257 * 1258 * On return the node interlock will be unlocked. If a non-zero error code 1259 * is returned the node will also be dereferenced (and the caller's pointer 1260 * will be stale). 1261 */ 1262 static int 1263 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew) 1264 { 1265 hammer_buffer_t buffer; 1266 hammer_off_t buf_offset; 1267 hammer_mount_t hmp = trans->hmp; 1268 int error; 1269 1270 error = 0; 1271 if (node->ondisk == NULL) { 1272 /* 1273 * This is a little confusing but the jist is that 1274 * node->buffer determines whether the node is on 1275 * the buffer's node_list and node->ondisk determines 1276 * whether the buffer is referenced. 1277 * 1278 * We could be racing a buffer release, in which case 1279 * node->buffer may become NULL while we are blocked 1280 * referencing the buffer. 1281 */ 1282 if ((buffer = node->buffer) != NULL) { 1283 error = hammer_ref_buffer(buffer); 1284 if (error == 0 && node->buffer == NULL) { 1285 TAILQ_INSERT_TAIL(&buffer->node_list, node, entry); 1286 node->buffer = buffer; 1287 } 1288 } else { 1289 buf_offset = node->node_offset & ~HAMMER_BUFMASK64; 1290 buffer = hammer_get_buffer(node->hmp, buf_offset, 1291 HAMMER_BUFSIZE, 0, &error); 1292 if (buffer) { 1293 KKASSERT(error == 0); 1294 TAILQ_INSERT_TAIL(&buffer->node_list, node, entry); 1295 node->buffer = buffer; 1296 } 1297 } 1298 if (error) 1299 goto failed; 1300 node->ondisk = (void *)((char *)buffer->ondisk + 1301 (node->node_offset & HAMMER_BUFMASK)); 1302 1303 /* 1304 * Check CRC. NOTE: Neither flag is set and the CRC is not 1305 * generated on new B-Tree nodes. 1306 */ 1307 if (isnew == 0 && 1308 (node->flags & HAMMER_NODE_CRCANY) == 0) { 1309 if (hammer_crc_test_btree(hmp->version, node->ondisk) == 0) { 1310 hdkprintf("CRC B-TREE NODE @ %016jx/%lu FAILED\n", 1311 (intmax_t)node->node_offset, 1312 sizeof(*node->ondisk)); 1313 if (hammer_debug_critical) 1314 Debugger("CRC FAILED: B-TREE NODE"); 1315 node->flags |= HAMMER_NODE_CRCBAD; 1316 } else { 1317 node->flags |= HAMMER_NODE_CRCGOOD; 1318 } 1319 } 1320 } 1321 if (node->flags & HAMMER_NODE_CRCBAD) { 1322 if (trans->flags & HAMMER_TRANSF_CRCDOM) 1323 error = EDOM; 1324 else 1325 error = EIO; 1326 } 1327 failed: 1328 if (error) { 1329 _hammer_rel_node(node, 1); 1330 } else { 1331 hammer_ref_interlock_done(&node->lock); 1332 } 1333 return (error); 1334 } 1335 1336 /* 1337 * Safely reference a node, interlock against flushes via the IO subsystem. 1338 */ 1339 hammer_node_t 1340 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache, 1341 int *errorp) 1342 { 1343 hammer_node_t node; 1344 int doload; 1345 1346 node = cache->node; 1347 if (node != NULL) { 1348 doload = hammer_ref_interlock(&node->lock); 1349 if (doload) { 1350 *errorp = hammer_load_node(trans, node, 0); 1351 if (*errorp) 1352 node = NULL; 1353 } else { 1354 KKASSERT(node->ondisk); 1355 if (node->flags & HAMMER_NODE_CRCBAD) { 1356 if (trans->flags & HAMMER_TRANSF_CRCDOM) 1357 *errorp = EDOM; 1358 else 1359 *errorp = EIO; 1360 _hammer_rel_node(node, 0); 1361 node = NULL; 1362 } else { 1363 *errorp = 0; 1364 } 1365 } 1366 } else { 1367 *errorp = ENOENT; 1368 } 1369 return(node); 1370 } 1371 1372 /* 1373 * Release a hammer_node. On the last release the node dereferences 1374 * its underlying buffer and may or may not be destroyed. 1375 * 1376 * If locked is non-zero the passed node has been interlocked by the 1377 * caller and we are in the failure/unload path, otherwise it has not and 1378 * we are doing a normal release. 1379 * 1380 * This function will dispose of the interlock and the reference. 1381 * On return the node pointer is stale. 1382 */ 1383 void 1384 _hammer_rel_node(hammer_node_t node, int locked) 1385 { 1386 hammer_buffer_t buffer; 1387 1388 /* 1389 * Deref the node. If this isn't the 1->0 transition we're basically 1390 * done. If locked is non-zero this function will just deref the 1391 * locked node and return 1, otherwise it will deref the locked 1392 * node and either lock and return 1 on the 1->0 transition or 1393 * not lock and return 0. 1394 */ 1395 if (hammer_rel_interlock(&node->lock, locked) == 0) 1396 return; 1397 1398 /* 1399 * Either locked was non-zero and we are interlocked, or the 1400 * hammer_rel_interlock() call returned non-zero and we are 1401 * interlocked. 1402 * 1403 * The ref-count must still be decremented if locked != 0 so 1404 * the cleanup required still varies a bit. 1405 * 1406 * hammer_flush_node() when called with 1 or 2 will dispose of 1407 * the lock and possible ref-count. 1408 */ 1409 if (node->ondisk == NULL) { 1410 hammer_flush_node(node, locked + 1); 1411 /* node is stale now */ 1412 return; 1413 } 1414 1415 /* 1416 * Do not disassociate the node from the buffer if it represents 1417 * a modified B-Tree node that still needs its crc to be generated. 1418 */ 1419 if (node->flags & HAMMER_NODE_NEEDSCRC) { 1420 hammer_rel_interlock_done(&node->lock, locked); 1421 return; 1422 } 1423 1424 /* 1425 * Do final cleanups and then either destroy the node and leave it 1426 * passively cached. The buffer reference is removed regardless. 1427 */ 1428 buffer = node->buffer; 1429 node->ondisk = NULL; 1430 1431 if ((node->flags & HAMMER_NODE_FLUSH) == 0) { 1432 /* 1433 * Normal release. 1434 */ 1435 hammer_rel_interlock_done(&node->lock, locked); 1436 } else { 1437 /* 1438 * Destroy the node. 1439 */ 1440 hammer_flush_node(node, locked + 1); 1441 /* node is stale */ 1442 1443 } 1444 hammer_rel_buffer(buffer, 0); 1445 } 1446 1447 void 1448 hammer_rel_node(hammer_node_t node) 1449 { 1450 _hammer_rel_node(node, 0); 1451 } 1452 1453 /* 1454 * Free space on-media associated with a B-Tree node. 1455 */ 1456 void 1457 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node) 1458 { 1459 KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0); 1460 node->flags |= HAMMER_NODE_DELETED; 1461 hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk)); 1462 } 1463 1464 /* 1465 * Passively cache a referenced hammer_node. The caller may release 1466 * the node on return. 1467 */ 1468 void 1469 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node) 1470 { 1471 /* 1472 * If the node doesn't exist, or is being deleted, don't cache it! 1473 * 1474 * The node can only ever be NULL in the I/O failure path. 1475 */ 1476 if (node == NULL || (node->flags & HAMMER_NODE_DELETED)) 1477 return; 1478 if (cache->node == node) 1479 return; 1480 while (cache->node) 1481 hammer_uncache_node(cache); 1482 if (node->flags & HAMMER_NODE_DELETED) 1483 return; 1484 cache->node = node; 1485 TAILQ_INSERT_TAIL(&node->cache_list, cache, entry); 1486 } 1487 1488 void 1489 hammer_uncache_node(hammer_node_cache_t cache) 1490 { 1491 hammer_node_t node; 1492 1493 if ((node = cache->node) != NULL) { 1494 TAILQ_REMOVE(&node->cache_list, cache, entry); 1495 cache->node = NULL; 1496 if (TAILQ_EMPTY(&node->cache_list)) 1497 hammer_flush_node(node, 0); 1498 } 1499 } 1500 1501 /* 1502 * Remove a node's cache references and destroy the node if it has no 1503 * other references or backing store. 1504 * 1505 * locked == 0 Normal unlocked operation 1506 * locked == 1 Call hammer_rel_interlock_done(..., 0); 1507 * locked == 2 Call hammer_rel_interlock_done(..., 1); 1508 * 1509 * XXX for now this isn't even close to being MPSAFE so the refs check 1510 * is sufficient. 1511 */ 1512 void 1513 hammer_flush_node(hammer_node_t node, int locked) 1514 { 1515 hammer_node_cache_t cache; 1516 hammer_buffer_t buffer; 1517 hammer_mount_t hmp = node->hmp; 1518 int dofree; 1519 1520 while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) { 1521 TAILQ_REMOVE(&node->cache_list, cache, entry); 1522 cache->node = NULL; 1523 } 1524 1525 /* 1526 * NOTE: refs is predisposed if another thread is blocking and 1527 * will be larger than 0 in that case. We aren't MPSAFE 1528 * here. 1529 */ 1530 if (node->ondisk == NULL && hammer_norefs(&node->lock)) { 1531 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0); 1532 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node); 1533 if ((buffer = node->buffer) != NULL) { 1534 node->buffer = NULL; 1535 TAILQ_REMOVE(&buffer->node_list, node, entry); 1536 /* buffer is unreferenced because ondisk is NULL */ 1537 } 1538 dofree = 1; 1539 } else { 1540 dofree = 0; 1541 } 1542 1543 /* 1544 * Deal with the interlock if locked == 1 or locked == 2. 1545 */ 1546 if (locked) 1547 hammer_rel_interlock_done(&node->lock, locked - 1); 1548 1549 /* 1550 * Destroy if requested 1551 */ 1552 if (dofree) { 1553 --hammer_count_nodes; 1554 kfree(node, hmp->m_misc); 1555 } 1556 } 1557 1558 /* 1559 * Flush passively cached B-Tree nodes associated with this buffer. 1560 * This is only called when the buffer is about to be destroyed, so 1561 * none of the nodes should have any references. The buffer is locked. 1562 * 1563 * We may be interlocked with the buffer. 1564 */ 1565 void 1566 hammer_flush_buffer_nodes(hammer_buffer_t buffer) 1567 { 1568 hammer_node_t node; 1569 1570 while ((node = TAILQ_FIRST(&buffer->node_list)) != NULL) { 1571 KKASSERT(node->ondisk == NULL); 1572 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0); 1573 1574 if (hammer_try_interlock_norefs(&node->lock)) { 1575 hammer_ref(&node->lock); 1576 node->flags |= HAMMER_NODE_FLUSH; 1577 _hammer_rel_node(node, 1); 1578 } else { 1579 KKASSERT(node->buffer != NULL); 1580 buffer = node->buffer; 1581 node->buffer = NULL; 1582 TAILQ_REMOVE(&buffer->node_list, node, entry); 1583 /* buffer is unreferenced because ondisk is NULL */ 1584 } 1585 } 1586 } 1587 1588 1589 /************************************************************************ 1590 * ALLOCATORS * 1591 ************************************************************************/ 1592 1593 /* 1594 * Allocate a B-Tree node. 1595 */ 1596 hammer_node_t 1597 hammer_alloc_btree(hammer_transaction_t trans, hammer_off_t hint, int *errorp) 1598 { 1599 hammer_buffer_t buffer = NULL; 1600 hammer_node_t node = NULL; 1601 hammer_off_t node_offset; 1602 1603 node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX, 1604 sizeof(struct hammer_node_ondisk), 1605 hint, errorp); 1606 if (*errorp == 0) { 1607 node = hammer_get_node(trans, node_offset, 1, errorp); 1608 hammer_modify_node_noundo(trans, node); 1609 bzero(node->ondisk, sizeof(*node->ondisk)); 1610 hammer_modify_node_done(node); 1611 } 1612 if (buffer) 1613 hammer_rel_buffer(buffer, 0); 1614 return(node); 1615 } 1616 1617 /* 1618 * Allocate data. If the address of a data buffer is supplied then 1619 * any prior non-NULL *data_bufferp will be released and *data_bufferp 1620 * will be set to the related buffer. The caller must release it when 1621 * finally done. The initial *data_bufferp should be set to NULL by 1622 * the caller. 1623 * 1624 * The caller is responsible for making hammer_modify*() calls on the 1625 * *data_bufferp. 1626 */ 1627 void * 1628 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len, 1629 uint16_t rec_type, hammer_off_t *data_offsetp, 1630 hammer_buffer_t *data_bufferp, 1631 hammer_off_t hint, int *errorp) 1632 { 1633 void *data; 1634 int zone; 1635 1636 /* 1637 * Allocate data directly from blockmap. 1638 */ 1639 if (data_len) { 1640 switch(rec_type) { 1641 case HAMMER_RECTYPE_INODE: 1642 case HAMMER_RECTYPE_DIRENTRY: 1643 case HAMMER_RECTYPE_EXT: 1644 case HAMMER_RECTYPE_FIX: 1645 case HAMMER_RECTYPE_PFS: 1646 case HAMMER_RECTYPE_SNAPSHOT: 1647 case HAMMER_RECTYPE_CONFIG: 1648 zone = HAMMER_ZONE_META_INDEX; 1649 break; 1650 case HAMMER_RECTYPE_DATA: 1651 case HAMMER_RECTYPE_DB: 1652 /* 1653 * Only mirror-write comes here. 1654 * Regular allocation path uses blockmap reservation. 1655 */ 1656 zone = hammer_data_zone_index(data_len); 1657 if (zone == HAMMER_ZONE_LARGE_DATA_INDEX) { 1658 /* round up */ 1659 data_len = HAMMER_BUFSIZE_DOALIGN(data_len); 1660 } 1661 break; 1662 default: 1663 hpanic("rec_type %04x unknown", rec_type); 1664 zone = HAMMER_ZONE_UNAVAIL_INDEX; /* NOT REACHED */ 1665 break; 1666 } 1667 *data_offsetp = hammer_blockmap_alloc(trans, zone, data_len, 1668 hint, errorp); 1669 } else { 1670 *data_offsetp = 0; 1671 } 1672 1673 data = NULL; 1674 if (*errorp == 0 && data_bufferp && data_len) 1675 data = hammer_bread_ext(trans->hmp, *data_offsetp, data_len, 1676 errorp, data_bufferp); 1677 return(data); 1678 } 1679 1680 /* 1681 * Sync dirty buffers to the media and clean-up any loose ends. 1682 * 1683 * These functions do not start the flusher going, they simply 1684 * queue everything up to the flusher. 1685 */ 1686 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data); 1687 1688 struct hammer_sync_info { 1689 int error; 1690 }; 1691 1692 int 1693 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor) 1694 { 1695 struct hammer_sync_info info; 1696 1697 info.error = 0; 1698 if (waitfor == MNT_WAIT) { 1699 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS, 1700 hammer_sync_scan2, &info); 1701 } else { 1702 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS | VMSC_NOWAIT, 1703 hammer_sync_scan2, &info); 1704 } 1705 return(info.error); 1706 } 1707 1708 /* 1709 * Filesystem sync. If doing a synchronous sync make a second pass on 1710 * the vnodes in case any were already flushing during the first pass, 1711 * and activate the flusher twice (the second time brings the UNDO FIFO's 1712 * start position up to the end position after the first call). 1713 * 1714 * If doing a lazy sync make just one pass on the vnode list, ignoring 1715 * any new vnodes added to the list while the sync is in progress. 1716 */ 1717 int 1718 hammer_sync_hmp(hammer_mount_t hmp, int waitfor) 1719 { 1720 struct hammer_sync_info info; 1721 int flags; 1722 1723 flags = VMSC_GETVP; 1724 if (waitfor & MNT_LAZY) 1725 flags |= VMSC_ONEPASS; 1726 1727 info.error = 0; 1728 vsyncscan(hmp->mp, flags | VMSC_NOWAIT, hammer_sync_scan2, &info); 1729 1730 if (info.error == 0 && (waitfor & MNT_WAIT)) { 1731 vsyncscan(hmp->mp, flags, hammer_sync_scan2, &info); 1732 } 1733 if (waitfor == MNT_WAIT) { 1734 hammer_flusher_sync(hmp); 1735 hammer_flusher_sync(hmp); 1736 } else { 1737 hammer_flusher_async(hmp, NULL); 1738 hammer_flusher_async(hmp, NULL); 1739 } 1740 return(info.error); 1741 } 1742 1743 static int 1744 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data) 1745 { 1746 struct hammer_sync_info *info = data; 1747 hammer_inode_t ip; 1748 int error; 1749 1750 ip = VTOI(vp); 1751 if (ip == NULL) 1752 return(0); 1753 if (vp->v_type == VNON || vp->v_type == VBAD) { 1754 vclrisdirty(vp); 1755 return(0); 1756 } 1757 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 && 1758 RB_EMPTY(&vp->v_rbdirty_tree)) { 1759 vclrisdirty(vp); 1760 return(0); 1761 } 1762 error = VOP_FSYNC(vp, MNT_NOWAIT, 0); 1763 if (error) 1764 info->error = error; 1765 return(0); 1766 } 1767