sys/vfs/hammer/hammer_ondisk.c

   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 1.71 2008/07/18 00:19:53 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 "hammer.h"
  43 #include <sys/fcntl.h>
  44 #include <sys/nlookup.h>
  45 #include <sys/buf.h>
  46 #include <sys/buf2.h>
  47
  48 static void hammer_free_volume(hammer_volume_t volume);
  49 static int hammer_load_volume(hammer_volume_t volume);
  50 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
  51 static int hammer_load_node(hammer_node_t node, int isnew);
  52
  53 static int
  54 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  55 {
  56         if (vol1->vol_no < vol2->vol_no)
  57                 return(-1);
  58         if (vol1->vol_no > vol2->vol_no)
  59                 return(1);
  60         return(0);
  61 }
  62
  63 static int
  64 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  65 {
  66         if (buf1->zoneX_offset < buf2->zoneX_offset)
  67                 return(-1);
  68         if (buf1->zoneX_offset > buf2->zoneX_offset)
  69                 return(1);
  70         return(0);
  71 }
  72
  73 static int
  74 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  75 {
  76         if (node1->node_offset < node2->node_offset)
  77                 return(-1);
  78         if (node1->node_offset > node2->node_offset)
  79                 return(1);
  80         return(0);
  81 }
  82
  83 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  84              hammer_vol_rb_compare, int32_t, vol_no);
  85 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  86              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  87 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  88              hammer_nod_rb_compare, hammer_off_t, node_offset);
  89
  90 /************************************************************************
  91  *                              VOLUMES                                 *
  92  ************************************************************************
  93  *
  94  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
  95  * code on failure.  Volumes must be loaded at mount time, get_volume() will
  96  * not load a new volume.
  97  *
  98  * Calls made to hammer_load_volume() or single-threaded
  99  */
 100 int
 101 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 102 {
 103         struct mount *mp;
 104         hammer_volume_t volume;
 105         struct hammer_volume_ondisk *ondisk;
 106         struct nlookupdata nd;
 107         struct buf *bp = NULL;
 108         int error;
 109         int ronly;
 110         int setmp = 0;
 111
 112         mp = hmp->mp;
 113         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 114
 115         /*
 116          * Allocate a volume structure
 117          */
 118         ++hammer_count_volumes;
 119         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 120         volume->vol_name = kstrdup(volname, M_HAMMER);
 121         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 122         volume->io.offset = 0LL;
 123         volume->io.bytes = HAMMER_BUFSIZE;
 124
 125         /*
 126          * Get the device vnode
 127          */
 128         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 129         if (error == 0)
 130                 error = nlookup(&nd);
 131         if (error == 0)
 132                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 133         nlookup_done(&nd);
 134         if (error == 0) {
 135                 if (vn_isdisk(volume->devvp, &error)) {
 136                         error = vfs_mountedon(volume->devvp);
 137                 }
 138         }
 139         if (error == 0 &&
 140             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 141                 error = EBUSY;
 142         }
 143         if (error == 0) {
 144                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 145                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 146                 if (error == 0) {
 147                         error = VOP_OPEN(volume->devvp,
 148                                          (ronly ? FREAD : FREAD|FWRITE),
 149                                          FSCRED, NULL);
 150                 }
 151                 vn_unlock(volume->devvp);
 152         }
 153         if (error) {
 154                 hammer_free_volume(volume);
 155                 return(error);
 156         }
 157         volume->devvp->v_rdev->si_mountpoint = mp;
 158         setmp = 1;
 159
 160         /*
 161          * Extract the volume number from the volume header and do various
 162          * sanity checks.
 163          */
 164         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 165         if (error)
 166                 goto late_failure;
 167         ondisk = (void *)bp->b_data;
 168         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 169                 kprintf("hammer_mount: volume %s has an invalid header\n",
 170                         volume->vol_name);
 171                 error = EFTYPE;
 172                 goto late_failure;
 173         }
 174         volume->vol_no = ondisk->vol_no;
 175         volume->buffer_base = ondisk->vol_buf_beg;
 176         volume->vol_flags = ondisk->vol_flags;
 177         volume->nblocks = ondisk->vol_nblocks;
 178         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 179                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 180         volume->maxraw_off = ondisk->vol_buf_end;
 181
 182         if (RB_EMPTY(&hmp->rb_vols_root)) {
 183                 hmp->fsid = ondisk->vol_fsid;
 184         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 185                 kprintf("hammer_mount: volume %s's fsid does not match "
 186                         "other volumes\n", volume->vol_name);
 187                 error = EFTYPE;
 188                 goto late_failure;
 189         }
 190
 191         /*
 192          * Insert the volume structure into the red-black tree.
 193          */
 194         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 195                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 196                         volume->vol_name, volume->vol_no);
 197                 error = EEXIST;
 198         }
 199
 200         /*
 201          * Set the root volume .  HAMMER special cases rootvol the structure.
 202          * We do not hold a ref because this would prevent related I/O
 203          * from being flushed.
 204          */
 205         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 206                 hmp->rootvol = volume;
 207                 hmp->nvolumes = ondisk->vol_count;
 208                 if (bp) {
 209                         brelse(bp);
 210                         bp = NULL;
 211                 }
 212                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 213                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 214                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 215                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 216         }
 217 late_failure:
 218         if (bp)
 219                 brelse(bp);
 220         if (error) {
 221                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 222                 if (setmp)
 223                         volume->devvp->v_rdev->si_mountpoint = NULL;
 224                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 225                 hammer_free_volume(volume);
 226         }
 227         return (error);
 228 }
 229
 230 /*
 231  * This is called for each volume when updating the mount point from
 232  * read-write to read-only or vise-versa.
 233  */
 234 int
 235 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 236 {
 237         if (volume->devvp) {
 238                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 239                 if (volume->io.hmp->ronly) {
 240                         /* do not call vinvalbuf */
 241                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 242                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 243                 } else {
 244                         /* do not call vinvalbuf */
 245                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 246                         VOP_CLOSE(volume->devvp, FREAD);
 247                 }
 248                 vn_unlock(volume->devvp);
 249         }
 250         return(0);
 251 }
 252
 253 /*
 254  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 255  * so returns -1 on failure.
 256  */
 257 int
 258 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 259 {
 260         struct hammer_mount *hmp = volume->io.hmp;
 261         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 262         struct buf *bp;
 263
 264         /*
 265          * Clean up the root volume pointer, which is held unlocked in hmp.
 266          */
 267         if (hmp->rootvol == volume)
 268                 hmp->rootvol = NULL;
 269
 270         /*
 271          * We must not flush a dirty buffer to disk on umount.  It should
 272          * have already been dealt with by the flusher, or we may be in
 273          * catastrophic failure.
 274          */
 275         hammer_io_clear_modify(&volume->io, 1);
 276         volume->io.waitdep = 1;
 277         bp = hammer_io_release(&volume->io, 1);
 278
 279         /*
 280          * Clean up the persistent ref ioerror might have on the volume
 281          */
 282         if (volume->io.ioerror) {
 283                 volume->io.ioerror = 0;
 284                 hammer_unref(&volume->io.lock);
 285         }
 286
 287         /*
 288          * There should be no references on the volume, no clusters, and
 289          * no super-clusters.
 290          */
 291         KKASSERT(volume->io.lock.refs == 0);
 292         if (bp)
 293                 brelse(bp);
 294
 295         volume->ondisk = NULL;
 296         if (volume->devvp) {
 297                 if (volume->devvp->v_rdev &&
 298                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 299                 ) {
 300                         volume->devvp->v_rdev->si_mountpoint = NULL;
 301                 }
 302                 if (ronly) {
 303                         /*
 304                          * Make sure we don't sync anything to disk if we
 305                          * are in read-only mode (1) or critically-errored
 306                          * (2).  Note that there may be dirty buffers in
 307                          * normal read-only mode from crash recovery.
 308                          */
 309                         vinvalbuf(volume->devvp, 0, 0, 0);
 310                         VOP_CLOSE(volume->devvp, FREAD);
 311                 } else {
 312                         /*
 313                          * Normal termination, save any dirty buffers
 314                          * (XXX there really shouldn't be any).
 315                          */
 316                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 317                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 318                 }
 319         }
 320
 321         /*
 322          * Destroy the structure
 323          */
 324         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 325         hammer_free_volume(volume);
 326         return(0);
 327 }
 328
 329 static
 330 void
 331 hammer_free_volume(hammer_volume_t volume)
 332 {
 333         if (volume->vol_name) {
 334                 kfree(volume->vol_name, M_HAMMER);
 335                 volume->vol_name = NULL;
 336         }
 337         if (volume->devvp) {
 338                 vrele(volume->devvp);
 339                 volume->devvp = NULL;
 340         }
 341         --hammer_count_volumes;
 342         kfree(volume, M_HAMMER);
 343 }
 344
 345 /*
 346  * Get a HAMMER volume.  The volume must already exist.
 347  */
 348 hammer_volume_t
 349 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 350 {
 351         struct hammer_volume *volume;
 352
 353         /*
 354          * Locate the volume structure
 355          */
 356         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 357         if (volume == NULL) {
 358                 *errorp = ENOENT;
 359                 return(NULL);
 360         }
 361         hammer_ref(&volume->io.lock);
 362
 363         /*
 364          * Deal with on-disk info
 365          */
 366         if (volume->ondisk == NULL || volume->io.loading) {
 367                 *errorp = hammer_load_volume(volume);
 368                 if (*errorp) {
 369                         hammer_rel_volume(volume, 1);
 370                         volume = NULL;
 371                 }
 372         } else {
 373                 *errorp = 0;
 374         }
 375         return(volume);
 376 }
 377
 378 int
 379 hammer_ref_volume(hammer_volume_t volume)
 380 {
 381         int error;
 382
 383         hammer_ref(&volume->io.lock);
 384
 385         /*
 386          * Deal with on-disk info
 387          */
 388         if (volume->ondisk == NULL || volume->io.loading) {
 389                 error = hammer_load_volume(volume);
 390                 if (error)
 391                         hammer_rel_volume(volume, 1);
 392         } else {
 393                 error = 0;
 394         }
 395         return (error);
 396 }
 397
 398 hammer_volume_t
 399 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 400 {
 401         hammer_volume_t volume;
 402
 403         volume = hmp->rootvol;
 404         KKASSERT(volume != NULL);
 405         hammer_ref(&volume->io.lock);
 406
 407         /*
 408          * Deal with on-disk info
 409          */
 410         if (volume->ondisk == NULL || volume->io.loading) {
 411                 *errorp = hammer_load_volume(volume);
 412                 if (*errorp) {
 413                         hammer_rel_volume(volume, 1);
 414                         volume = NULL;
 415                 }
 416         } else {
 417                 *errorp = 0;
 418         }
 419         return (volume);
 420 }
 421
 422 /*
 423  * Load a volume's on-disk information.  The volume must be referenced and
 424  * not locked.  We temporarily acquire an exclusive lock to interlock
 425  * against releases or multiple get's.
 426  */
 427 static int
 428 hammer_load_volume(hammer_volume_t volume)
 429 {
 430         int error;
 431
 432         ++volume->io.loading;
 433         hammer_lock_ex(&volume->io.lock);
 434
 435         if (volume->ondisk == NULL) {
 436                 error = hammer_io_read(volume->devvp, &volume->io,
 437                                        volume->maxraw_off);
 438                 if (error == 0)
 439                         volume->ondisk = (void *)volume->io.bp->b_data;
 440         } else {
 441                 error = 0;
 442         }
 443         --volume->io.loading;
 444         hammer_unlock(&volume->io.lock);
 445         return(error);
 446 }
 447
 448 /*
 449  * Release a volume.  Call hammer_io_release on the last reference.  We have
 450  * to acquire an exclusive lock to interlock against volume->ondisk tests
 451  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 452  * lock to be held.
 453  *
 454  * Volumes are not unloaded from memory during normal operation.
 455  */
 456 void
 457 hammer_rel_volume(hammer_volume_t volume, int flush)
 458 {
 459         struct buf *bp = NULL;
 460
 461         crit_enter();
 462         if (volume->io.lock.refs == 1) {
 463                 ++volume->io.loading;
 464                 hammer_lock_ex(&volume->io.lock);
 465                 if (volume->io.lock.refs == 1) {
 466                         volume->ondisk = NULL;
 467                         bp = hammer_io_release(&volume->io, flush);
 468                 }
 469                 --volume->io.loading;
 470                 hammer_unlock(&volume->io.lock);
 471         }
 472         hammer_unref(&volume->io.lock);
 473         if (bp)
 474                 brelse(bp);
 475         crit_exit();
 476 }
 477
 478 int
 479 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 480 {
 481         hammer_volume_t vol;
 482         int i;
 483
 484         for (i = 0; i < hmp->nvolumes; ++i) {
 485                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 486                 if (vol == NULL)
 487                         return(EINVAL);
 488         }
 489         return(0);
 490 }
 491
 492 /************************************************************************
 493  *                              BUFFERS                                 *
 494  ************************************************************************
 495  *
 496  * Manage buffers.  Currently all blockmap-backed zones are translated
 497  * to zone-2 buffer offsets.
 498  */
 499 hammer_buffer_t
 500 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 501                   int bytes, int isnew, int *errorp)
 502 {
 503         hammer_buffer_t buffer;
 504         hammer_volume_t volume;
 505         hammer_off_t    zone2_offset;
 506         hammer_io_type_t iotype;
 507         int vol_no;
 508         int zone;
 509
 510         buf_offset &= ~HAMMER_BUFMASK64;
 511 again:
 512         /*
 513          * Shortcut if the buffer is already cached
 514          */
 515         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 516         if (buffer) {
 517                 if (buffer->io.lock.refs == 0)
 518                         ++hammer_count_refedbufs;
 519                 hammer_ref(&buffer->io.lock);
 520
 521                 /*
 522                  * Onced refed the ondisk field will not be cleared by
 523                  * any other action.
 524                  */
 525                 if (buffer->ondisk && buffer->io.loading == 0) {
 526                         *errorp = 0;
 527                         return(buffer);
 528                 }
 529
 530                 /*
 531                  * The buffer is no longer loose if it has a ref, and
 532                  * cannot become loose once it gains a ref.  Loose
 533                  * buffers will never be in a modified state.  This should
 534                  * only occur on the 0->1 transition of refs.
 535                  *
 536                  * lose_list can be modified via a biodone() interrupt.
 537                  */
 538                 if (buffer->io.mod_list == &hmp->lose_list) {
 539                         crit_enter();   /* biodone race against list */
 540                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 541                                      mod_entry);
 542                         crit_exit();
 543                         buffer->io.mod_list = NULL;
 544                         KKASSERT(buffer->io.modified == 0);
 545                 }
 546                 goto found;
 547         }
 548
 549         /*
 550          * What is the buffer class?
 551          */
 552         zone = HAMMER_ZONE_DECODE(buf_offset);
 553
 554         switch(zone) {
 555         case HAMMER_ZONE_LARGE_DATA_INDEX:
 556         case HAMMER_ZONE_SMALL_DATA_INDEX:
 557                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 558                 break;
 559         case HAMMER_ZONE_UNDO_INDEX:
 560                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 561                 break;
 562         case HAMMER_ZONE_META_INDEX:
 563         default:
 564                 /*
 565                  * NOTE: inode data and directory entries are placed in this
 566                  * zone.  inode atime/mtime is updated in-place and thus
 567                  * buffers containing inodes must be synchronized as
 568                  * meta-buffers, same as buffers containing B-Tree info.
 569                  */
 570                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 571                 break;
 572         }
 573
 574         /*
 575          * Handle blockmap offset translations
 576          */
 577         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 578                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 579         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 580                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 581         } else {
 582                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 583                 zone2_offset = buf_offset;
 584                 *errorp = 0;
 585         }
 586         if (*errorp)
 587                 return(NULL);
 588
 589         /*
 590          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 591          * specifications.
 592          */
 593         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 594                  HAMMER_ZONE_RAW_BUFFER);
 595         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 596         volume = hammer_get_volume(hmp, vol_no, errorp);
 597         if (volume == NULL)
 598                 return(NULL);
 599
 600         KKASSERT(zone2_offset < volume->maxbuf_off);
 601
 602         /*
 603          * Allocate a new buffer structure.  We will check for races later.
 604          */
 605         ++hammer_count_buffers;
 606         buffer = kmalloc(sizeof(*buffer), M_HAMMER,
 607                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 608         buffer->zone2_offset = zone2_offset;
 609         buffer->zoneX_offset = buf_offset;
 610         buffer->volume = volume;
 611
 612         hammer_io_init(&buffer->io, hmp, iotype);
 613         buffer->io.offset = volume->ondisk->vol_buf_beg +
 614                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 615         buffer->io.bytes = bytes;
 616         TAILQ_INIT(&buffer->clist);
 617         hammer_ref(&buffer->io.lock);
 618
 619         /*
 620          * Insert the buffer into the RB tree and handle late collisions.
 621          */
 622         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 623                 hammer_unref(&buffer->io.lock);
 624                 --hammer_count_buffers;
 625                 kfree(buffer, M_HAMMER);
 626                 goto again;
 627         }
 628         ++hammer_count_refedbufs;
 629 found:
 630
 631         /*
 632          * Deal with on-disk info and loading races.
 633          */
 634         if (buffer->ondisk == NULL || buffer->io.loading) {
 635                 *errorp = hammer_load_buffer(buffer, isnew);
 636                 if (*errorp) {
 637                         hammer_rel_buffer(buffer, 1);
 638                         buffer = NULL;
 639                 }
 640         } else {
 641                 *errorp = 0;
 642         }
 643         return(buffer);
 644 }
 645
 646 /*
 647  * This is used by the direct-read code to deal with large-data buffers
 648  * created by the reblocker and mirror-write code.  The direct-read code
 649  * bypasses the HAMMER buffer subsystem and so any aliased dirty hammer
 650  * buffers must be fully synced to disk before we can issue the direct-read.
 651  *
 652  * This code path is not considered critical as only the rebocker and
 653  * mirror-write code will create large-data buffers via the HAMMER buffer
 654  * subsystem.  They do that because they operate at the B-Tree level and
 655  * do not access the vnode/inode structures.
 656  */
 657 void
 658 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 659 {
 660         hammer_buffer_t buffer;
 661         int error;
 662
 663         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 664                  HAMMER_ZONE_LARGE_DATA);
 665
 666         while (bytes > 0) {
 667                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 668                                    base_offset);
 669                 if (buffer && buffer->io.modified) {
 670                         error = hammer_ref_buffer(buffer);
 671                         if (error == 0 && buffer->io.modified) {
 672                                 hammer_io_write_interlock(&buffer->io);
 673                                 hammer_io_flush(&buffer->io);
 674                                 hammer_io_done_interlock(&buffer->io);
 675                                 hammer_io_wait(&buffer->io);
 676                                 hammer_rel_buffer(buffer, 0);
 677                         }
 678                 }
 679                 base_offset += HAMMER_BUFSIZE;
 680                 bytes -= HAMMER_BUFSIZE;
 681         }
 682 }
 683
 684 /*
 685  * Destroy all buffers covering the specified zoneX offset range.  This
 686  * is called when the related blockmap layer2 entry is freed or when
 687  * a direct write bypasses our buffer/buffer-cache subsystem.
 688  *
 689  * The buffers may be referenced by the caller itself.  Setting reclaim
 690  * will cause the buffer to be destroyed when it's ref count reaches zero.
 691  */
 692 void
 693 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 694                    hammer_off_t zone2_offset, int bytes)
 695 {
 696         hammer_buffer_t buffer;
 697         hammer_volume_t volume;
 698         int vol_no;
 699         int error;
 700
 701         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 702         volume = hammer_get_volume(hmp, vol_no, &error);
 703         KKASSERT(error == 0);
 704
 705         while (bytes > 0) {
 706                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 707                                    base_offset);
 708                 if (buffer) {
 709                         error = hammer_ref_buffer(buffer);
 710                         if (error == 0) {
 711                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 712                                 hammer_io_clear_modify(&buffer->io, 1);
 713                                 buffer->io.reclaim = 1;
 714                                 KKASSERT(buffer->volume == volume);
 715                                 hammer_rel_buffer(buffer, 0);
 716                         }
 717                 } else {
 718                         hammer_io_inval(volume, zone2_offset);
 719                 }
 720                 base_offset += HAMMER_BUFSIZE;
 721                 zone2_offset += HAMMER_BUFSIZE;
 722                 bytes -= HAMMER_BUFSIZE;
 723         }
 724         hammer_rel_volume(volume, 0);
 725 }
 726
 727 static int
 728 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 729 {
 730         hammer_volume_t volume;
 731         int error;
 732
 733         /*
 734          * Load the buffer's on-disk info
 735          */
 736         volume = buffer->volume;
 737         ++buffer->io.loading;
 738         hammer_lock_ex(&buffer->io.lock);
 739
 740         if (hammer_debug_io & 0x0001) {
 741                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 742                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 743                         buffer->ondisk);
 744         }
 745
 746         if (buffer->ondisk == NULL) {
 747                 if (isnew) {
 748                         error = hammer_io_new(volume->devvp, &buffer->io);
 749                 } else {
 750                         error = hammer_io_read(volume->devvp, &buffer->io,
 751                                                volume->maxraw_off);
 752                 }
 753                 if (error == 0)
 754                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 755         } else if (isnew) {
 756                 error = hammer_io_new(volume->devvp, &buffer->io);
 757         } else {
 758                 error = 0;
 759         }
 760         --buffer->io.loading;
 761         hammer_unlock(&buffer->io.lock);
 762         return (error);
 763 }
 764
 765 /*
 766  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 767  * This routine is only called during unmount.
 768  */
 769 int
 770 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 771 {
 772         /*
 773          * Clean up the persistent ref ioerror might have on the buffer
 774          * and acquire a ref (steal ioerror's if we can).
 775          */
 776         if (buffer->io.ioerror) {
 777                 buffer->io.ioerror = 0;
 778         } else {
 779                 if (buffer->io.lock.refs == 0)
 780                         ++hammer_count_refedbufs;
 781                 hammer_ref(&buffer->io.lock);
 782         }
 783
 784         /*
 785          * We must not flush a dirty buffer to disk on umount.  It should
 786          * have already been dealt with by the flusher, or we may be in
 787          * catastrophic failure.
 788          */
 789         hammer_io_clear_modify(&buffer->io, 1);
 790         hammer_flush_buffer_nodes(buffer);
 791         KKASSERT(buffer->io.lock.refs == 1);
 792         hammer_rel_buffer(buffer, 2);
 793         return(0);
 794 }
 795
 796 /*
 797  * Reference a buffer that is either already referenced or via a specially
 798  * handled pointer (aka cursor->buffer).
 799  */
 800 int
 801 hammer_ref_buffer(hammer_buffer_t buffer)
 802 {
 803         int error;
 804
 805         if (buffer->io.lock.refs == 0)
 806                 ++hammer_count_refedbufs;
 807         hammer_ref(&buffer->io.lock);
 808
 809         /*
 810          * At this point a biodone() will not touch the buffer other then
 811          * incidental bits.  However, lose_list can be modified via
 812          * a biodone() interrupt.
 813          *
 814          * No longer loose
 815          */
 816         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 817                 crit_enter();
 818                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 819                 buffer->io.mod_list = NULL;
 820                 crit_exit();
 821         }
 822
 823         if (buffer->ondisk == NULL || buffer->io.loading) {
 824                 error = hammer_load_buffer(buffer, 0);
 825                 if (error) {
 826                         hammer_rel_buffer(buffer, 1);
 827                         /*
 828                          * NOTE: buffer pointer can become stale after
 829                          * the above release.
 830                          */
 831                 }
 832         } else {
 833                 error = 0;
 834         }
 835         return(error);
 836 }
 837
 838 /*
 839  * Release a buffer.  We have to deal with several places where
 840  * another thread can ref the buffer.
 841  *
 842  * Only destroy the structure itself if the related buffer cache buffer
 843  * was disassociated from it.  This ties the management of the structure
 844  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 845  * embedded io is referenced or not.
 846  */
 847 void
 848 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 849 {
 850         hammer_volume_t volume;
 851         struct buf *bp = NULL;
 852         int freeme = 0;
 853
 854         crit_enter();
 855         if (buffer->io.lock.refs == 1) {
 856                 ++buffer->io.loading;   /* force interlock check */
 857                 hammer_lock_ex(&buffer->io.lock);
 858                 if (buffer->io.lock.refs == 1) {
 859                         bp = hammer_io_release(&buffer->io, flush);
 860
 861                         if (buffer->io.lock.refs == 1)
 862                                 --hammer_count_refedbufs;
 863
 864                         if (buffer->io.bp == NULL &&
 865                             buffer->io.lock.refs == 1) {
 866                                 /*
 867                                  * Final cleanup
 868                                  *
 869                                  * NOTE: It is impossible for any associated
 870                                  * B-Tree nodes to have refs if the buffer
 871                                  * has no additional refs.
 872                                  */
 873                                 RB_REMOVE(hammer_buf_rb_tree,
 874                                           &buffer->io.hmp->rb_bufs_root,
 875                                           buffer);
 876                                 volume = buffer->volume;
 877                                 buffer->volume = NULL; /* sanity */
 878                                 hammer_rel_volume(volume, 0);
 879                                 hammer_io_clear_modlist(&buffer->io);
 880                                 hammer_flush_buffer_nodes(buffer);
 881                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 882                                 freeme = 1;
 883                         }
 884                 }
 885                 --buffer->io.loading;
 886                 hammer_unlock(&buffer->io.lock);
 887         }
 888         hammer_unref(&buffer->io.lock);
 889         crit_exit();
 890         if (bp)
 891                 brelse(bp);
 892         if (freeme) {
 893                 --hammer_count_buffers;
 894                 kfree(buffer, M_HAMMER);
 895         }
 896 }
 897
 898 /*
 899  * Access the filesystem buffer containing the specified hammer offset.
 900  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 901  * relative buffer offset.  It must also have bit 55 set to be valid.
 902  * (see hammer_off_t in hammer_disk.h).
 903  *
 904  * Any prior buffer in *bufferp will be released and replaced by the
 905  * requested buffer.
 906  */
 907 static __inline
 908 void *
 909 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 910              int *errorp, struct hammer_buffer **bufferp)
 911 {
 912         hammer_buffer_t buffer;
 913         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 914
 915         buf_offset &= ~HAMMER_BUFMASK64;
 916         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 917
 918         buffer = *bufferp;
 919         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 920                                buffer->zoneX_offset != buf_offset)) {
 921                 if (buffer)
 922                         hammer_rel_buffer(buffer, 0);
 923                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 924                 *bufferp = buffer;
 925         } else {
 926                 *errorp = 0;
 927         }
 928
 929         /*
 930          * Return a pointer to the buffer data.
 931          */
 932         if (buffer == NULL)
 933                 return(NULL);
 934         else
 935                 return((char *)buffer->ondisk + xoff);
 936 }
 937
 938 void *
 939 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 940              int *errorp, struct hammer_buffer **bufferp)
 941 {
 942         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 943 }
 944
 945 void *
 946 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 947                  int *errorp, struct hammer_buffer **bufferp)
 948 {
 949         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 950         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 951 }
 952
 953 /*
 954  * Access the filesystem buffer containing the specified hammer offset.
 955  * No disk read operation occurs.  The result buffer may contain garbage.
 956  *
 957  * Any prior buffer in *bufferp will be released and replaced by the
 958  * requested buffer.
 959  *
 960  * This function marks the buffer dirty but does not increment its
 961  * modify_refs count.
 962  */
 963 static __inline
 964 void *
 965 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 966              int *errorp, struct hammer_buffer **bufferp)
 967 {
 968         hammer_buffer_t buffer;
 969         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 970
 971         buf_offset &= ~HAMMER_BUFMASK64;
 972
 973         buffer = *bufferp;
 974         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 975                                buffer->zoneX_offset != buf_offset)) {
 976                 if (buffer)
 977                         hammer_rel_buffer(buffer, 0);
 978                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 979                 *bufferp = buffer;
 980         } else {
 981                 *errorp = 0;
 982         }
 983
 984         /*
 985          * Return a pointer to the buffer data.
 986          */
 987         if (buffer == NULL)
 988                 return(NULL);
 989         else
 990                 return((char *)buffer->ondisk + xoff);
 991 }
 992
 993 void *
 994 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 995              int *errorp, struct hammer_buffer **bufferp)
 996 {
 997         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 998 }
 999
1000 void *
1001 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1002                 int *errorp, struct hammer_buffer **bufferp)
1003 {
1004         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1005         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1006 }
1007
1008 /************************************************************************
1009  *                              NODES                                   *
1010  ************************************************************************
1011  *
1012  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1013  * method used by the HAMMER filesystem.
1014  *
1015  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1016  * associated with its buffer, and will only referenced the buffer while
1017  * the node itself is referenced.
1018  *
1019  * A hammer_node can also be passively associated with other HAMMER
1020  * structures, such as inodes, while retaining 0 references.  These
1021  * associations can be cleared backwards using a pointer-to-pointer in
1022  * the hammer_node.
1023  *
1024  * This allows the HAMMER implementation to cache hammer_nodes long-term
1025  * and short-cut a great deal of the infrastructure's complexity.  In
1026  * most cases a cached node can be reacquired without having to dip into
1027  * either the buffer or cluster management code.
1028  *
1029  * The caller must pass a referenced cluster on call and will retain
1030  * ownership of the reference on return.  The node will acquire its own
1031  * additional references, if necessary.
1032  */
1033 hammer_node_t
1034 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
1035                 int isnew, int *errorp)
1036 {
1037         hammer_node_t node;
1038
1039         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1040
1041         /*
1042          * Locate the structure, allocating one if necessary.
1043          */
1044 again:
1045         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1046         if (node == NULL) {
1047                 ++hammer_count_nodes;
1048                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
1049                 node->node_offset = node_offset;
1050                 node->hmp = hmp;
1051                 TAILQ_INIT(&node->cursor_list);
1052                 TAILQ_INIT(&node->cache_list);
1053                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1054                         --hammer_count_nodes;
1055                         kfree(node, M_HAMMER);
1056                         goto again;
1057                 }
1058         }
1059         hammer_ref(&node->lock);
1060         if (node->ondisk)
1061                 *errorp = 0;
1062         else
1063                 *errorp = hammer_load_node(node, isnew);
1064         if (*errorp) {
1065                 hammer_rel_node(node);
1066                 node = NULL;
1067         }
1068         return(node);
1069 }
1070
1071 /*
1072  * Reference an already-referenced node.
1073  */
1074 void
1075 hammer_ref_node(hammer_node_t node)
1076 {
1077         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1078         hammer_ref(&node->lock);
1079 }
1080
1081 /*
1082  * Load a node's on-disk data reference.
1083  */
1084 static int
1085 hammer_load_node(hammer_node_t node, int isnew)
1086 {
1087         hammer_buffer_t buffer;
1088         hammer_off_t buf_offset;
1089         int error;
1090
1091         error = 0;
1092         ++node->loading;
1093         hammer_lock_ex(&node->lock);
1094         if (node->ondisk == NULL) {
1095                 /*
1096                  * This is a little confusing but the jist is that
1097                  * node->buffer determines whether the node is on
1098                  * the buffer's clist and node->ondisk determines
1099                  * whether the buffer is referenced.
1100                  *
1101                  * We could be racing a buffer release, in which case
1102                  * node->buffer may become NULL while we are blocked
1103                  * referencing the buffer.
1104                  */
1105                 if ((buffer = node->buffer) != NULL) {
1106                         error = hammer_ref_buffer(buffer);
1107                         if (error == 0 && node->buffer == NULL) {
1108                                 TAILQ_INSERT_TAIL(&buffer->clist,
1109                                                   node, entry);
1110                                 node->buffer = buffer;
1111                         }
1112                 } else {
1113                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1114                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1115                                                    HAMMER_BUFSIZE, 0, &error);
1116                         if (buffer) {
1117                                 KKASSERT(error == 0);
1118                                 TAILQ_INSERT_TAIL(&buffer->clist,
1119                                                   node, entry);
1120                                 node->buffer = buffer;
1121                         }
1122                 }
1123                 if (error)
1124                         goto failed;
1125                 node->ondisk = (void *)((char *)buffer->ondisk +
1126                                         (node->node_offset & HAMMER_BUFMASK));
1127                 if (isnew == 0 &&
1128                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1129                         if (hammer_crc_test_btree(node->ondisk) == 0)
1130                                 Debugger("CRC FAILED: B-TREE NODE");
1131                         node->flags |= HAMMER_NODE_CRCGOOD;
1132                 }
1133         }
1134 failed:
1135         --node->loading;
1136         hammer_unlock(&node->lock);
1137         return (error);
1138 }
1139
1140 /*
1141  * Safely reference a node, interlock against flushes via the IO subsystem.
1142  */
1143 hammer_node_t
1144 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1145                      int *errorp)
1146 {
1147         hammer_node_t node;
1148
1149         node = cache->node;
1150         if (node != NULL) {
1151                 hammer_ref(&node->lock);
1152                 if (node->ondisk)
1153                         *errorp = 0;
1154                 else
1155                         *errorp = hammer_load_node(node, 0);
1156                 if (*errorp) {
1157                         hammer_rel_node(node);
1158                         node = NULL;
1159                 }
1160         } else {
1161                 *errorp = ENOENT;
1162         }
1163         return(node);
1164 }
1165
1166 /*
1167  * Release a hammer_node.  On the last release the node dereferences
1168  * its underlying buffer and may or may not be destroyed.
1169  */
1170 void
1171 hammer_rel_node(hammer_node_t node)
1172 {
1173         hammer_buffer_t buffer;
1174
1175         /*
1176          * If this isn't the last ref just decrement the ref count and
1177          * return.
1178          */
1179         if (node->lock.refs > 1) {
1180                 hammer_unref(&node->lock);
1181                 return;
1182         }
1183
1184         /*
1185          * If there is no ondisk info or no buffer the node failed to load,
1186          * remove the last reference and destroy the node.
1187          */
1188         if (node->ondisk == NULL) {
1189                 hammer_unref(&node->lock);
1190                 hammer_flush_node(node);
1191                 /* node is stale now */
1192                 return;
1193         }
1194
1195         /*
1196          * Do not disassociate the node from the buffer if it represents
1197          * a modified B-Tree node that still needs its crc to be generated.
1198          */
1199         if (node->flags & HAMMER_NODE_NEEDSCRC)
1200                 return;
1201
1202         /*
1203          * Do final cleanups and then either destroy the node and leave it
1204          * passively cached.  The buffer reference is removed regardless.
1205          */
1206         buffer = node->buffer;
1207         node->ondisk = NULL;
1208
1209         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1210                 hammer_unref(&node->lock);
1211                 hammer_rel_buffer(buffer, 0);
1212                 return;
1213         }
1214
1215         /*
1216          * Destroy the node.
1217          */
1218         hammer_unref(&node->lock);
1219         hammer_flush_node(node);
1220         /* node is stale */
1221         hammer_rel_buffer(buffer, 0);
1222 }
1223
1224 /*
1225  * Free space on-media associated with a B-Tree node.
1226  */
1227 void
1228 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1229 {
1230         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1231         node->flags |= HAMMER_NODE_DELETED;
1232         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1233 }
1234
1235 /*
1236  * Passively cache a referenced hammer_node.  The caller may release
1237  * the node on return.
1238  */
1239 void
1240 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1241 {
1242         /*
1243          * If the node doesn't exist, or is being deleted, don't cache it!
1244          *
1245          * The node can only ever be NULL in the I/O failure path.
1246          */
1247         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1248                 return;
1249         if (cache->node == node)
1250                 return;
1251         while (cache->node)
1252                 hammer_uncache_node(cache);
1253         if (node->flags & HAMMER_NODE_DELETED)
1254                 return;
1255         cache->node = node;
1256         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1257 }
1258
1259 void
1260 hammer_uncache_node(hammer_node_cache_t cache)
1261 {
1262         hammer_node_t node;
1263
1264         if ((node = cache->node) != NULL) {
1265                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1266                 cache->node = NULL;
1267                 if (TAILQ_EMPTY(&node->cache_list))
1268                         hammer_flush_node(node);
1269         }
1270 }
1271
1272 /*
1273  * Remove a node's cache references and destroy the node if it has no
1274  * other references or backing store.
1275  */
1276 void
1277 hammer_flush_node(hammer_node_t node)
1278 {
1279         hammer_node_cache_t cache;
1280         hammer_buffer_t buffer;
1281
1282         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1283                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1284                 cache->node = NULL;
1285         }
1286         if (node->lock.refs == 0 && node->ondisk == NULL) {
1287                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1288                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1289                 if ((buffer = node->buffer) != NULL) {
1290                         node->buffer = NULL;
1291                         TAILQ_REMOVE(&buffer->clist, node, entry);
1292                         /* buffer is unreferenced because ondisk is NULL */
1293                 }
1294                 --hammer_count_nodes;
1295                 kfree(node, M_HAMMER);
1296         }
1297 }
1298
1299 /*
1300  * Flush passively cached B-Tree nodes associated with this buffer.
1301  * This is only called when the buffer is about to be destroyed, so
1302  * none of the nodes should have any references.  The buffer is locked.
1303  *
1304  * We may be interlocked with the buffer.
1305  */
1306 void
1307 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1308 {
1309         hammer_node_t node;
1310
1311         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1312                 KKASSERT(node->ondisk == NULL);
1313                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1314
1315                 if (node->lock.refs == 0) {
1316                         hammer_ref(&node->lock);
1317                         node->flags |= HAMMER_NODE_FLUSH;
1318                         hammer_rel_node(node);
1319                 } else {
1320                         KKASSERT(node->loading != 0);
1321                         KKASSERT(node->buffer != NULL);
1322                         buffer = node->buffer;
1323                         node->buffer = NULL;
1324                         TAILQ_REMOVE(&buffer->clist, node, entry);
1325                         /* buffer is unreferenced because ondisk is NULL */
1326                 }
1327         }
1328 }
1329
1330
1331 /************************************************************************
1332  *                              ALLOCATORS                              *
1333  ************************************************************************/
1334
1335 /*
1336  * Allocate a B-Tree node.
1337  */
1338 hammer_node_t
1339 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1340 {
1341         hammer_buffer_t buffer = NULL;
1342         hammer_node_t node = NULL;
1343         hammer_off_t node_offset;
1344
1345         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1346                                             sizeof(struct hammer_node_ondisk),
1347                                             errorp);
1348         if (*errorp == 0) {
1349                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1350                 hammer_modify_node_noundo(trans, node);
1351                 bzero(node->ondisk, sizeof(*node->ondisk));
1352                 hammer_modify_node_done(node);
1353         }
1354         if (buffer)
1355                 hammer_rel_buffer(buffer, 0);
1356         return(node);
1357 }
1358
1359 /*
1360  * Allocate data.  If the address of a data buffer is supplied then
1361  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1362  * will be set to the related buffer.  The caller must release it when
1363  * finally done.  The initial *data_bufferp should be set to NULL by
1364  * the caller.
1365  *
1366  * The caller is responsible for making hammer_modify*() calls on the
1367  * *data_bufferp.
1368  */
1369 void *
1370 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1371                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1372                   struct hammer_buffer **data_bufferp, int *errorp)
1373 {
1374         void *data;
1375         int zone;
1376
1377         /*
1378          * Allocate data
1379          */
1380         if (data_len) {
1381                 switch(rec_type) {
1382                 case HAMMER_RECTYPE_INODE:
1383                 case HAMMER_RECTYPE_DIRENTRY:
1384                 case HAMMER_RECTYPE_EXT:
1385                 case HAMMER_RECTYPE_FIX:
1386                 case HAMMER_RECTYPE_PFS:
1387                         zone = HAMMER_ZONE_META_INDEX;
1388                         break;
1389                 case HAMMER_RECTYPE_DATA:
1390                 case HAMMER_RECTYPE_DB:
1391                         if (data_len <= HAMMER_BUFSIZE / 2) {
1392                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1393                         } else {
1394                                 data_len = (data_len + HAMMER_BUFMASK) &
1395                                            ~HAMMER_BUFMASK;
1396                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1397                         }
1398                         break;
1399                 default:
1400                         panic("hammer_alloc_data: rec_type %04x unknown",
1401                               rec_type);
1402                         zone = 0;       /* NOT REACHED */
1403                         break;
1404                 }
1405                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1406                                                       data_len, errorp);
1407         } else {
1408                 *data_offsetp = 0;
1409         }
1410         if (*errorp == 0 && data_bufferp) {
1411                 if (data_len) {
1412                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1413                                                 data_len, errorp, data_bufferp);
1414                 } else {
1415                         data = NULL;
1416                 }
1417         } else {
1418                 data = NULL;
1419         }
1420         return(data);
1421 }
1422
1423 /*
1424  * Sync dirty buffers to the media and clean-up any loose ends.
1425  *
1426  * These functions do not start the flusher going, they simply
1427  * queue everything up to the flusher.
1428  */
1429 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1430 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1431
1432 int
1433 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1434 {
1435         struct hammer_sync_info info;
1436
1437         info.error = 0;
1438         info.waitfor = waitfor;
1439         if (waitfor == MNT_WAIT) {
1440                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1441                               hammer_sync_scan1, hammer_sync_scan2, &info);
1442         } else {
1443                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1444                               hammer_sync_scan1, hammer_sync_scan2, &info);
1445         }
1446         return(info.error);
1447 }
1448
1449 /*
1450  * Filesystem sync.  If doing a synchronous sync make a second pass on
1451  * the vnodes in case any were already flushing during the first pass,
1452  * and activate the flusher twice (the second time brings the UNDO FIFO's
1453  * start position up to the end position after the first call).
1454  */
1455 int
1456 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1457 {
1458         struct hammer_sync_info info;
1459
1460         info.error = 0;
1461         info.waitfor = MNT_NOWAIT;
1462         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1463                       hammer_sync_scan1, hammer_sync_scan2, &info);
1464         if (info.error == 0 && waitfor == MNT_WAIT) {
1465                 info.waitfor = waitfor;
1466                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1467                               hammer_sync_scan1, hammer_sync_scan2, &info);
1468         }
1469         if (waitfor == MNT_WAIT) {
1470                 hammer_flusher_sync(hmp);
1471                 hammer_flusher_sync(hmp);
1472         } else {
1473                 hammer_flusher_async(hmp, NULL);
1474         }
1475         return(info.error);
1476 }
1477
1478 static int
1479 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1480 {
1481         struct hammer_inode *ip;
1482
1483         ip = VTOI(vp);
1484         if (vp->v_type == VNON || ip == NULL ||
1485             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1486              RB_EMPTY(&vp->v_rbdirty_tree))) {
1487                 return(-1);
1488         }
1489         return(0);
1490 }
1491
1492 static int
1493 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1494 {
1495         struct hammer_sync_info *info = data;
1496         struct hammer_inode *ip;
1497         int error;
1498
1499         ip = VTOI(vp);
1500         if (vp->v_type == VNON || vp->v_type == VBAD ||
1501             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1502              RB_EMPTY(&vp->v_rbdirty_tree))) {
1503                 return(0);
1504         }
1505         error = VOP_FSYNC(vp, MNT_NOWAIT);
1506         if (error)
1507                 info->error = error;
1508         return(0);
1509 }
1510