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 1.62 2008/06/21 20:21:58 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                 if (bp) {
 208                         brelse(bp);
 209                         bp = NULL;
 210                 }
 211                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 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
 263         /*
 264          * Clean up the root volume pointer, which is held unlocked in hmp.
 265          */
 266         if (hmp->rootvol == volume)
 267                 hmp->rootvol = NULL;
 268
 269         /*
 270          * Release our buffer and flush anything left in the buffer cache.
 271          */
 272         volume->io.waitdep = 1;
 273         hammer_io_release(&volume->io, 1);
 274         hammer_io_clear_modlist(&volume->io);
 275
 276         /*
 277          * There should be no references on the volume, no clusters, and
 278          * no super-clusters.
 279          */
 280         KKASSERT(volume->io.lock.refs == 0);
 281
 282         volume->ondisk = NULL;
 283         if (volume->devvp) {
 284                 if (volume->devvp->v_rdev &&
 285                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 286                 ) {
 287                         volume->devvp->v_rdev->si_mountpoint = NULL;
 288                 }
 289                 if (ronly) {
 290                         vinvalbuf(volume->devvp, 0, 0, 0);
 291                         VOP_CLOSE(volume->devvp, FREAD);
 292                 } else {
 293                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 294                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 295                 }
 296         }
 297
 298         /*
 299          * Destroy the structure
 300          */
 301         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 302         hammer_free_volume(volume);
 303         return(0);
 304 }
 305
 306 static
 307 void
 308 hammer_free_volume(hammer_volume_t volume)
 309 {
 310         if (volume->vol_name) {
 311                 kfree(volume->vol_name, M_HAMMER);
 312                 volume->vol_name = NULL;
 313         }
 314         if (volume->devvp) {
 315                 vrele(volume->devvp);
 316                 volume->devvp = NULL;
 317         }
 318         --hammer_count_volumes;
 319         kfree(volume, M_HAMMER);
 320 }
 321
 322 /*
 323  * Get a HAMMER volume.  The volume must already exist.
 324  */
 325 hammer_volume_t
 326 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 327 {
 328         struct hammer_volume *volume;
 329
 330         /*
 331          * Locate the volume structure
 332          */
 333         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 334         if (volume == NULL) {
 335                 *errorp = ENOENT;
 336                 return(NULL);
 337         }
 338         hammer_ref(&volume->io.lock);
 339
 340         /*
 341          * Deal with on-disk info
 342          */
 343         if (volume->ondisk == NULL || volume->io.loading) {
 344                 *errorp = hammer_load_volume(volume);
 345                 if (*errorp) {
 346                         hammer_rel_volume(volume, 1);
 347                         volume = NULL;
 348                 }
 349         } else {
 350                 *errorp = 0;
 351         }
 352         return(volume);
 353 }
 354
 355 int
 356 hammer_ref_volume(hammer_volume_t volume)
 357 {
 358         int error;
 359
 360         hammer_ref(&volume->io.lock);
 361
 362         /*
 363          * Deal with on-disk info
 364          */
 365         if (volume->ondisk == NULL || volume->io.loading) {
 366                 error = hammer_load_volume(volume);
 367                 if (error)
 368                         hammer_rel_volume(volume, 1);
 369         } else {
 370                 error = 0;
 371         }
 372         return (error);
 373 }
 374
 375 hammer_volume_t
 376 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 377 {
 378         hammer_volume_t volume;
 379
 380         volume = hmp->rootvol;
 381         KKASSERT(volume != NULL);
 382         hammer_ref(&volume->io.lock);
 383
 384         /*
 385          * Deal with on-disk info
 386          */
 387         if (volume->ondisk == NULL || volume->io.loading) {
 388                 *errorp = hammer_load_volume(volume);
 389                 if (*errorp) {
 390                         hammer_rel_volume(volume, 1);
 391                         volume = NULL;
 392                 }
 393         } else {
 394                 *errorp = 0;
 395         }
 396         return (volume);
 397 }
 398
 399 /*
 400  * Load a volume's on-disk information.  The volume must be referenced and
 401  * not locked.  We temporarily acquire an exclusive lock to interlock
 402  * against releases or multiple get's.
 403  */
 404 static int
 405 hammer_load_volume(hammer_volume_t volume)
 406 {
 407         int error;
 408
 409         ++volume->io.loading;
 410         hammer_lock_ex(&volume->io.lock);
 411
 412         if (volume->ondisk == NULL) {
 413                 error = hammer_io_read(volume->devvp, &volume->io,
 414                                        volume->maxraw_off);
 415                 if (error == 0)
 416                         volume->ondisk = (void *)volume->io.bp->b_data;
 417         } else {
 418                 error = 0;
 419         }
 420         --volume->io.loading;
 421         hammer_unlock(&volume->io.lock);
 422         return(error);
 423 }
 424
 425 /*
 426  * Release a volume.  Call hammer_io_release on the last reference.  We have
 427  * to acquire an exclusive lock to interlock against volume->ondisk tests
 428  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 429  * lock to be held.
 430  *
 431  * Volumes are not unloaded from memory during normal operation.
 432  */
 433 void
 434 hammer_rel_volume(hammer_volume_t volume, int flush)
 435 {
 436         crit_enter();
 437         if (volume->io.lock.refs == 1) {
 438                 ++volume->io.loading;
 439                 hammer_lock_ex(&volume->io.lock);
 440                 if (volume->io.lock.refs == 1) {
 441                         volume->ondisk = NULL;
 442                         hammer_io_release(&volume->io, flush);
 443                 }
 444                 --volume->io.loading;
 445                 hammer_unlock(&volume->io.lock);
 446         }
 447         hammer_unref(&volume->io.lock);
 448         crit_exit();
 449 }
 450
 451 /************************************************************************
 452  *                              BUFFERS                                 *
 453  ************************************************************************
 454  *
 455  * Manage buffers.  Currently all blockmap-backed zones are translated
 456  * to zone-2 buffer offsets.
 457  */
 458 hammer_buffer_t
 459 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 460                   int bytes, int isnew, int *errorp)
 461 {
 462         hammer_buffer_t buffer;
 463         hammer_volume_t volume;
 464         hammer_off_t    zone2_offset;
 465         hammer_io_type_t iotype;
 466         int vol_no;
 467         int zone;
 468
 469         buf_offset &= ~HAMMER_BUFMASK64;
 470 again:
 471         /*
 472          * Shortcut if the buffer is already cached
 473          */
 474         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 475         if (buffer) {
 476                 if (buffer->io.lock.refs == 0)
 477                         ++hammer_count_refedbufs;
 478                 hammer_ref(&buffer->io.lock);
 479
 480                 /*
 481                  * Onced refed the ondisk field will not be cleared by
 482                  * any other action.
 483                  */
 484                 if (buffer->ondisk && buffer->io.loading == 0) {
 485                         *errorp = 0;
 486                         return(buffer);
 487                 }
 488
 489                 /*
 490                  * The buffer is no longer loose if it has a ref, and
 491                  * cannot become loose once it gains a ref.  Loose
 492                  * buffers will never be in a modified state.  This should
 493                  * only occur on the 0->1 transition of refs.
 494                  *
 495                  * lose_list can be modified via a biodone() interrupt.
 496                  */
 497                 if (buffer->io.mod_list == &hmp->lose_list) {
 498                         crit_enter();   /* biodone race against list */
 499                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 500                                      mod_entry);
 501                         crit_exit();
 502                         buffer->io.mod_list = NULL;
 503                         KKASSERT(buffer->io.modified == 0);
 504                 }
 505                 goto found;
 506         }
 507
 508         /*
 509          * What is the buffer class?
 510          */
 511         zone = HAMMER_ZONE_DECODE(buf_offset);
 512
 513         switch(zone) {
 514         case HAMMER_ZONE_LARGE_DATA_INDEX:
 515         case HAMMER_ZONE_SMALL_DATA_INDEX:
 516                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 517                 break;
 518         case HAMMER_ZONE_UNDO_INDEX:
 519                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 520                 break;
 521         case HAMMER_ZONE_META_INDEX:
 522         default:
 523                 /*
 524                  * NOTE: inode data and directory entries are placed in this
 525                  * zone.  inode atime/mtime is updated in-place and thus
 526                  * buffers containing inodes must be synchronized as
 527                  * meta-buffers, same as buffers containing B-Tree info.
 528                  */
 529                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 530                 break;
 531         }
 532
 533         /*
 534          * Handle blockmap offset translations
 535          */
 536         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 537                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 538         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 539                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 540         } else {
 541                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 542                 zone2_offset = buf_offset;
 543                 *errorp = 0;
 544         }
 545         if (*errorp)
 546                 return(NULL);
 547
 548         /*
 549          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 550          * specifications.
 551          */
 552         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 553                  HAMMER_ZONE_RAW_BUFFER);
 554         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 555         volume = hammer_get_volume(hmp, vol_no, errorp);
 556         if (volume == NULL)
 557                 return(NULL);
 558
 559         KKASSERT(zone2_offset < volume->maxbuf_off);
 560
 561         /*
 562          * Allocate a new buffer structure.  We will check for races later.
 563          */
 564         ++hammer_count_buffers;
 565         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 566         buffer->zone2_offset = zone2_offset;
 567         buffer->zoneX_offset = buf_offset;
 568         buffer->volume = volume;
 569
 570         hammer_io_init(&buffer->io, hmp, iotype);
 571         buffer->io.offset = volume->ondisk->vol_buf_beg +
 572                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 573         buffer->io.bytes = bytes;
 574         TAILQ_INIT(&buffer->clist);
 575         hammer_ref(&buffer->io.lock);
 576
 577         /*
 578          * Insert the buffer into the RB tree and handle late collisions.
 579          */
 580         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 581                 hammer_unref(&buffer->io.lock);
 582                 --hammer_count_buffers;
 583                 kfree(buffer, M_HAMMER);
 584                 goto again;
 585         }
 586         ++hammer_count_refedbufs;
 587 found:
 588
 589         /*
 590          * Deal with on-disk info and loading races.
 591          */
 592         if (buffer->ondisk == NULL || buffer->io.loading) {
 593                 *errorp = hammer_load_buffer(buffer, isnew);
 594                 if (*errorp) {
 595                         hammer_rel_buffer(buffer, 1);
 596                         buffer = NULL;
 597                 }
 598         } else {
 599                 *errorp = 0;
 600         }
 601         return(buffer);
 602 }
 603
 604 /*
 605  * Destroy all buffers covering the specified zoneX offset range.  This
 606  * is called when the related blockmap layer2 entry is freed or when
 607  * a direct write bypasses our buffer/buffer-cache subsystem.
 608  *
 609  * The buffers may be referenced by the caller itself.  Setting reclaim
 610  * will cause the buffer to be destroyed when it's ref count reaches zero.
 611  */
 612 void
 613 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 614                    hammer_off_t zone2_offset, int bytes)
 615 {
 616         hammer_buffer_t buffer;
 617         hammer_volume_t volume;
 618         int vol_no;
 619         int error;
 620
 621         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 622         volume = hammer_get_volume(hmp, vol_no, &error);
 623         KKASSERT(error == 0);
 624
 625         while (bytes > 0) {
 626                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 627                                    base_offset);
 628                 if (buffer) {
 629                         KKASSERT(buffer->zone2_offset == zone2_offset);
 630                         hammer_io_clear_modify(&buffer->io, 1);
 631                         buffer->io.reclaim = 1;
 632                         KKASSERT(buffer->volume == volume);
 633                         if (buffer->io.lock.refs == 0)
 634                                 hammer_unload_buffer(buffer, NULL);
 635                 } else {
 636                         hammer_io_inval(volume, zone2_offset);
 637                 }
 638                 base_offset += HAMMER_BUFSIZE;
 639                 zone2_offset += HAMMER_BUFSIZE;
 640                 bytes -= HAMMER_BUFSIZE;
 641         }
 642         hammer_rel_volume(volume, 0);
 643 }
 644
 645 static int
 646 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 647 {
 648         hammer_volume_t volume;
 649         int error;
 650
 651         /*
 652          * Load the buffer's on-disk info
 653          */
 654         volume = buffer->volume;
 655         ++buffer->io.loading;
 656         hammer_lock_ex(&buffer->io.lock);
 657
 658         if (hammer_debug_io & 0x0001) {
 659                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 660                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 661                         buffer->ondisk);
 662         }
 663
 664         if (buffer->ondisk == NULL) {
 665                 if (isnew) {
 666                         error = hammer_io_new(volume->devvp, &buffer->io);
 667                 } else {
 668                         error = hammer_io_read(volume->devvp, &buffer->io,
 669                                                volume->maxraw_off);
 670                 }
 671                 if (error == 0)
 672                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 673         } else if (isnew) {
 674                 error = hammer_io_new(volume->devvp, &buffer->io);
 675         } else {
 676                 error = 0;
 677         }
 678         --buffer->io.loading;
 679         hammer_unlock(&buffer->io.lock);
 680         return (error);
 681 }
 682
 683 /*
 684  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 685  */
 686 int
 687 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 688 {
 689         ++hammer_count_refedbufs;
 690         hammer_ref(&buffer->io.lock);
 691         hammer_flush_buffer_nodes(buffer);
 692         KKASSERT(buffer->io.lock.refs == 1);
 693         hammer_rel_buffer(buffer, 2);
 694         return(0);
 695 }
 696
 697 /*
 698  * Reference a buffer that is either already referenced or via a specially
 699  * handled pointer (aka cursor->buffer).
 700  */
 701 int
 702 hammer_ref_buffer(hammer_buffer_t buffer)
 703 {
 704         int error;
 705
 706         if (buffer->io.lock.refs == 0)
 707                 ++hammer_count_refedbufs;
 708         hammer_ref(&buffer->io.lock);
 709
 710         /*
 711          * At this point a biodone() will not touch the buffer other then
 712          * incidental bits.  However, lose_list can be modified via
 713          * a biodone() interrupt.
 714          *
 715          * No longer loose
 716          */
 717         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 718                 crit_enter();
 719                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 720                 buffer->io.mod_list = NULL;
 721                 crit_exit();
 722         }
 723
 724         if (buffer->ondisk == NULL || buffer->io.loading) {
 725                 error = hammer_load_buffer(buffer, 0);
 726                 if (error) {
 727                         hammer_rel_buffer(buffer, 1);
 728                         /*
 729                          * NOTE: buffer pointer can become stale after
 730                          * the above release.
 731                          */
 732                 }
 733         } else {
 734                 error = 0;
 735         }
 736         return(error);
 737 }
 738
 739 /*
 740  * Release a buffer.  We have to deal with several places where
 741  * another thread can ref the buffer.
 742  *
 743  * Only destroy the structure itself if the related buffer cache buffer
 744  * was disassociated from it.  This ties the management of the structure
 745  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 746  * embedded io is referenced or not.
 747  */
 748 void
 749 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 750 {
 751         hammer_volume_t volume;
 752         int freeme = 0;
 753
 754         crit_enter();
 755         if (buffer->io.lock.refs == 1) {
 756                 ++buffer->io.loading;   /* force interlock check */
 757                 hammer_lock_ex(&buffer->io.lock);
 758                 if (buffer->io.lock.refs == 1) {
 759                         hammer_io_release(&buffer->io, flush);
 760
 761                         if (buffer->io.lock.refs == 1)
 762                                 --hammer_count_refedbufs;
 763
 764                         if (buffer->io.bp == NULL &&
 765                             buffer->io.lock.refs == 1) {
 766                                 /*
 767                                  * Final cleanup
 768                                  *
 769                                  * NOTE: It is impossible for any associated
 770                                  * B-Tree nodes to have refs if the buffer
 771                                  * has no additional refs.
 772                                  */
 773                                 RB_REMOVE(hammer_buf_rb_tree,
 774                                           &buffer->io.hmp->rb_bufs_root,
 775                                           buffer);
 776                                 volume = buffer->volume;
 777                                 buffer->volume = NULL; /* sanity */
 778                                 hammer_rel_volume(volume, 0);
 779                                 hammer_io_clear_modlist(&buffer->io);
 780                                 hammer_flush_buffer_nodes(buffer);
 781                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 782                                 freeme = 1;
 783                         }
 784                 }
 785                 --buffer->io.loading;
 786                 hammer_unlock(&buffer->io.lock);
 787         }
 788         hammer_unref(&buffer->io.lock);
 789         crit_exit();
 790         if (freeme) {
 791                 --hammer_count_buffers;
 792                 kfree(buffer, M_HAMMER);
 793         }
 794 }
 795
 796 /*
 797  * Access the filesystem buffer containing the specified hammer offset.
 798  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 799  * relative buffer offset.  It must also have bit 55 set to be valid.
 800  * (see hammer_off_t in hammer_disk.h).
 801  *
 802  * Any prior buffer in *bufferp will be released and replaced by the
 803  * requested buffer.
 804  */
 805 static __inline
 806 void *
 807 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 808              int *errorp, struct hammer_buffer **bufferp)
 809 {
 810         hammer_buffer_t buffer;
 811         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 812
 813         buf_offset &= ~HAMMER_BUFMASK64;
 814         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 815
 816         buffer = *bufferp;
 817         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 818                                buffer->zoneX_offset != buf_offset)) {
 819                 if (buffer)
 820                         hammer_rel_buffer(buffer, 0);
 821                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 822                 *bufferp = buffer;
 823         } else {
 824                 *errorp = 0;
 825         }
 826
 827         /*
 828          * Return a pointer to the buffer data.
 829          */
 830         if (buffer == NULL)
 831                 return(NULL);
 832         else
 833                 return((char *)buffer->ondisk + xoff);
 834 }
 835
 836 void *
 837 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 838              int *errorp, struct hammer_buffer **bufferp)
 839 {
 840         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 841 }
 842
 843 void *
 844 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 845                  int *errorp, struct hammer_buffer **bufferp)
 846 {
 847         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 848         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 849 }
 850
 851 /*
 852  * Access the filesystem buffer containing the specified hammer offset.
 853  * No disk read operation occurs.  The result buffer may contain garbage.
 854  *
 855  * Any prior buffer in *bufferp will be released and replaced by the
 856  * requested buffer.
 857  *
 858  * This function marks the buffer dirty but does not increment its
 859  * modify_refs count.
 860  */
 861 static __inline
 862 void *
 863 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 864              int *errorp, struct hammer_buffer **bufferp)
 865 {
 866         hammer_buffer_t buffer;
 867         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 868
 869         buf_offset &= ~HAMMER_BUFMASK64;
 870
 871         buffer = *bufferp;
 872         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 873                                buffer->zoneX_offset != buf_offset)) {
 874                 if (buffer)
 875                         hammer_rel_buffer(buffer, 0);
 876                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 877                 *bufferp = buffer;
 878         } else {
 879                 *errorp = 0;
 880         }
 881
 882         /*
 883          * Return a pointer to the buffer data.
 884          */
 885         if (buffer == NULL)
 886                 return(NULL);
 887         else
 888                 return((char *)buffer->ondisk + xoff);
 889 }
 890
 891 void *
 892 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 893              int *errorp, struct hammer_buffer **bufferp)
 894 {
 895         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 896 }
 897
 898 void *
 899 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 900                 int *errorp, struct hammer_buffer **bufferp)
 901 {
 902         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 903         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
 904 }
 905
 906 /************************************************************************
 907  *                              NODES                                   *
 908  ************************************************************************
 909  *
 910  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 911  * method used by the HAMMER filesystem.
 912  *
 913  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 914  * associated with its buffer, and will only referenced the buffer while
 915  * the node itself is referenced.
 916  *
 917  * A hammer_node can also be passively associated with other HAMMER
 918  * structures, such as inodes, while retaining 0 references.  These
 919  * associations can be cleared backwards using a pointer-to-pointer in
 920  * the hammer_node.
 921  *
 922  * This allows the HAMMER implementation to cache hammer_nodes long-term
 923  * and short-cut a great deal of the infrastructure's complexity.  In
 924  * most cases a cached node can be reacquired without having to dip into
 925  * either the buffer or cluster management code.
 926  *
 927  * The caller must pass a referenced cluster on call and will retain
 928  * ownership of the reference on return.  The node will acquire its own
 929  * additional references, if necessary.
 930  */
 931 hammer_node_t
 932 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 933                 int isnew, int *errorp)
 934 {
 935         hammer_node_t node;
 936
 937         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 938
 939         /*
 940          * Locate the structure, allocating one if necessary.
 941          */
 942 again:
 943         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 944         if (node == NULL) {
 945                 ++hammer_count_nodes;
 946                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 947                 node->node_offset = node_offset;
 948                 node->hmp = hmp;
 949                 TAILQ_INIT(&node->cache_list);
 950                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 951                         --hammer_count_nodes;
 952                         kfree(node, M_HAMMER);
 953                         goto again;
 954                 }
 955         }
 956         hammer_ref(&node->lock);
 957         if (node->ondisk)
 958                 *errorp = 0;
 959         else
 960                 *errorp = hammer_load_node(node, isnew);
 961         if (*errorp) {
 962                 hammer_rel_node(node);
 963                 node = NULL;
 964         }
 965         return(node);
 966 }
 967
 968 /*
 969  * Reference an already-referenced node.
 970  */
 971 void
 972 hammer_ref_node(hammer_node_t node)
 973 {
 974         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 975         hammer_ref(&node->lock);
 976 }
 977
 978 /*
 979  * Load a node's on-disk data reference.
 980  */
 981 static int
 982 hammer_load_node(hammer_node_t node, int isnew)
 983 {
 984         hammer_buffer_t buffer;
 985         hammer_off_t buf_offset;
 986         int error;
 987
 988         error = 0;
 989         ++node->loading;
 990         hammer_lock_ex(&node->lock);
 991         if (node->ondisk == NULL) {
 992                 /*
 993                  * This is a little confusing but the jist is that
 994                  * node->buffer determines whether the node is on
 995                  * the buffer's clist and node->ondisk determines
 996                  * whether the buffer is referenced.
 997                  *
 998                  * We could be racing a buffer release, in which case
 999                  * node->buffer may become NULL while we are blocked
1000                  * referencing the buffer.
1001                  */
1002                 if ((buffer = node->buffer) != NULL) {
1003                         error = hammer_ref_buffer(buffer);
1004                         if (error == 0 && node->buffer == NULL) {
1005                                 TAILQ_INSERT_TAIL(&buffer->clist,
1006                                                   node, entry);
1007                                 node->buffer = buffer;
1008                         }
1009                 } else {
1010                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1011                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1012                                                    HAMMER_BUFSIZE, 0, &error);
1013                         if (buffer) {
1014                                 KKASSERT(error == 0);
1015                                 TAILQ_INSERT_TAIL(&buffer->clist,
1016                                                   node, entry);
1017                                 node->buffer = buffer;
1018                         }
1019                 }
1020                 if (error)
1021                         goto failed;
1022                 node->ondisk = (void *)((char *)buffer->ondisk +
1023                                         (node->node_offset & HAMMER_BUFMASK));
1024                 if (isnew == 0 &&
1025                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1026                         if (hammer_crc_test_btree(node->ondisk) == 0)
1027                                 Debugger("CRC FAILED: B-TREE NODE");
1028                         node->flags |= HAMMER_NODE_CRCGOOD;
1029                 }
1030         }
1031 failed:
1032         --node->loading;
1033         hammer_unlock(&node->lock);
1034         return (error);
1035 }
1036
1037 /*
1038  * Safely reference a node, interlock against flushes via the IO subsystem.
1039  */
1040 hammer_node_t
1041 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1042                      int *errorp)
1043 {
1044         hammer_node_t node;
1045
1046         node = cache->node;
1047         if (node != NULL) {
1048                 hammer_ref(&node->lock);
1049                 if (node->ondisk)
1050                         *errorp = 0;
1051                 else
1052                         *errorp = hammer_load_node(node, 0);
1053                 if (*errorp) {
1054                         hammer_rel_node(node);
1055                         node = NULL;
1056                 }
1057         } else {
1058                 *errorp = ENOENT;
1059         }
1060         return(node);
1061 }
1062
1063 /*
1064  * Release a hammer_node.  On the last release the node dereferences
1065  * its underlying buffer and may or may not be destroyed.
1066  */
1067 void
1068 hammer_rel_node(hammer_node_t node)
1069 {
1070         hammer_buffer_t buffer;
1071
1072         /*
1073          * If this isn't the last ref just decrement the ref count and
1074          * return.
1075          */
1076         if (node->lock.refs > 1) {
1077                 hammer_unref(&node->lock);
1078                 return;
1079         }
1080
1081         /*
1082          * If there is no ondisk info or no buffer the node failed to load,
1083          * remove the last reference and destroy the node.
1084          */
1085         if (node->ondisk == NULL) {
1086                 hammer_unref(&node->lock);
1087                 hammer_flush_node(node);
1088                 /* node is stale now */
1089                 return;
1090         }
1091
1092         /*
1093          * Do not disassociate the node from the buffer if it represents
1094          * a modified B-Tree node that still needs its crc to be generated.
1095          */
1096         if (node->flags & HAMMER_NODE_NEEDSCRC)
1097                 return;
1098
1099         /*
1100          * Do final cleanups and then either destroy the node and leave it
1101          * passively cached.  The buffer reference is removed regardless.
1102          */
1103         buffer = node->buffer;
1104         node->ondisk = NULL;
1105
1106         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1107                 hammer_unref(&node->lock);
1108                 hammer_rel_buffer(buffer, 0);
1109                 return;
1110         }
1111
1112         /*
1113          * Destroy the node.
1114          */
1115         hammer_unref(&node->lock);
1116         hammer_flush_node(node);
1117         /* node is stale */
1118         hammer_rel_buffer(buffer, 0);
1119 }
1120
1121 /*
1122  * Free space on-media associated with a B-Tree node.
1123  */
1124 void
1125 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1126 {
1127         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1128         node->flags |= HAMMER_NODE_DELETED;
1129         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1130 }
1131
1132 /*
1133  * Passively cache a referenced hammer_node.  The caller may release
1134  * the node on return.
1135  */
1136 void
1137 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1138 {
1139         /*
1140          * If the node is being deleted, don't cache it!
1141          */
1142         if (node->flags & HAMMER_NODE_DELETED)
1143                 return;
1144         if (cache->node == node)
1145                 return;
1146         while (cache->node)
1147                 hammer_uncache_node(cache);
1148         if (node->flags & HAMMER_NODE_DELETED)
1149                 return;
1150         cache->node = node;
1151         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1152 }
1153
1154 void
1155 hammer_uncache_node(hammer_node_cache_t cache)
1156 {
1157         hammer_node_t node;
1158
1159         if ((node = cache->node) != NULL) {
1160                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1161                 cache->node = NULL;
1162                 if (TAILQ_EMPTY(&node->cache_list))
1163                         hammer_flush_node(node);
1164         }
1165 }
1166
1167 /*
1168  * Remove a node's cache references and destroy the node if it has no
1169  * other references or backing store.
1170  */
1171 void
1172 hammer_flush_node(hammer_node_t node)
1173 {
1174         hammer_node_cache_t cache;
1175         hammer_buffer_t buffer;
1176
1177         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1178                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1179                 cache->node = NULL;
1180         }
1181         if (node->lock.refs == 0 && node->ondisk == NULL) {
1182                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1183                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1184                 if ((buffer = node->buffer) != NULL) {
1185                         node->buffer = NULL;
1186                         TAILQ_REMOVE(&buffer->clist, node, entry);
1187                         /* buffer is unreferenced because ondisk is NULL */
1188                 }
1189                 --hammer_count_nodes;
1190                 kfree(node, M_HAMMER);
1191         }
1192 }
1193
1194 /*
1195  * Flush passively cached B-Tree nodes associated with this buffer.
1196  * This is only called when the buffer is about to be destroyed, so
1197  * none of the nodes should have any references.  The buffer is locked.
1198  *
1199  * We may be interlocked with the buffer.
1200  */
1201 void
1202 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1203 {
1204         hammer_node_t node;
1205
1206         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1207                 KKASSERT(node->ondisk == NULL);
1208                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1209
1210                 if (node->lock.refs == 0) {
1211                         hammer_ref(&node->lock);
1212                         node->flags |= HAMMER_NODE_FLUSH;
1213                         hammer_rel_node(node);
1214                 } else {
1215                         KKASSERT(node->loading != 0);
1216                         KKASSERT(node->buffer != NULL);
1217                         buffer = node->buffer;
1218                         node->buffer = NULL;
1219                         TAILQ_REMOVE(&buffer->clist, node, entry);
1220                         /* buffer is unreferenced because ondisk is NULL */
1221                 }
1222         }
1223 }
1224
1225
1226 /************************************************************************
1227  *                              ALLOCATORS                              *
1228  ************************************************************************/
1229
1230 /*
1231  * Allocate a B-Tree node.
1232  */
1233 hammer_node_t
1234 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1235 {
1236         hammer_buffer_t buffer = NULL;
1237         hammer_node_t node = NULL;
1238         hammer_off_t node_offset;
1239
1240         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1241                                             sizeof(struct hammer_node_ondisk),
1242                                             errorp);
1243         if (*errorp == 0) {
1244                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1245                 hammer_modify_node_noundo(trans, node);
1246                 bzero(node->ondisk, sizeof(*node->ondisk));
1247                 hammer_modify_node_done(node);
1248         }
1249         if (buffer)
1250                 hammer_rel_buffer(buffer, 0);
1251         return(node);
1252 }
1253
1254 /*
1255  * Allocate data.  If the address of a data buffer is supplied then
1256  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1257  * will be set to the related buffer.  The caller must release it when
1258  * finally done.  The initial *data_bufferp should be set to NULL by
1259  * the caller.
1260  *
1261  * The caller is responsible for making hammer_modify*() calls on the
1262  * *data_bufferp.
1263  */
1264 void *
1265 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1266                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1267                   struct hammer_buffer **data_bufferp, int *errorp)
1268 {
1269         void *data;
1270         int zone;
1271
1272         /*
1273          * Allocate data
1274          */
1275         if (data_len) {
1276                 switch(rec_type) {
1277                 case HAMMER_RECTYPE_INODE:
1278                 case HAMMER_RECTYPE_DIRENTRY:
1279                 case HAMMER_RECTYPE_EXT:
1280                 case HAMMER_RECTYPE_FIX:
1281                         zone = HAMMER_ZONE_META_INDEX;
1282                         break;
1283                 case HAMMER_RECTYPE_DATA:
1284                 case HAMMER_RECTYPE_DB:
1285                         if (data_len <= HAMMER_BUFSIZE / 2) {
1286                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1287                         } else {
1288                                 data_len = (data_len + HAMMER_BUFMASK) &
1289                                            ~HAMMER_BUFMASK;
1290                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1291                         }
1292                         break;
1293                 default:
1294                         panic("hammer_alloc_data: rec_type %04x unknown",
1295                               rec_type);
1296                         zone = 0;       /* NOT REACHED */
1297                         break;
1298                 }
1299                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1300                                                       data_len, errorp);
1301         } else {
1302                 *data_offsetp = 0;
1303         }
1304         if (*errorp == 0 && data_bufferp) {
1305                 if (data_len) {
1306                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1307                                                 data_len, errorp, data_bufferp);
1308                         KKASSERT(*errorp == 0);
1309                 } else {
1310                         data = NULL;
1311                 }
1312         } else {
1313                 data = NULL;
1314         }
1315         KKASSERT(*errorp == 0);
1316         return(data);
1317 }
1318
1319 /*
1320  * Sync dirty buffers to the media and clean-up any loose ends.
1321  */
1322 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1323 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1324
1325 int
1326 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1327 {
1328         struct hammer_sync_info info;
1329
1330         info.error = 0;
1331         info.waitfor = waitfor;
1332         if (waitfor == MNT_WAIT) {
1333                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1334                               hammer_sync_scan1, hammer_sync_scan2, &info);
1335         } else {
1336                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1337                               hammer_sync_scan1, hammer_sync_scan2, &info);
1338         }
1339         return(info.error);
1340 }
1341
1342 /*
1343  * Filesystem sync.  If doing a synchronous sync make a second pass on
1344  * the vnodes in case any were already flushing during the first pass,
1345  * and activate the flusher twice (the second time brings the UNDO FIFO's
1346  * start position up to the end position after the first call).
1347  */
1348 int
1349 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1350 {
1351         struct hammer_sync_info info;
1352
1353         info.error = 0;
1354         info.waitfor = MNT_NOWAIT;
1355         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1356                       hammer_sync_scan1, hammer_sync_scan2, &info);
1357         if (info.error == 0 && waitfor == MNT_WAIT) {
1358                 info.waitfor = waitfor;
1359                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1360                               hammer_sync_scan1, hammer_sync_scan2, &info);
1361         }
1362         if (waitfor == MNT_WAIT) {
1363                 hammer_flusher_sync(hmp);
1364                 hammer_flusher_sync(hmp);
1365         } else {
1366                 hammer_flusher_async(hmp);
1367         }
1368         return(info.error);
1369 }
1370
1371 static int
1372 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1373 {
1374         struct hammer_inode *ip;
1375
1376         ip = VTOI(vp);
1377         if (vp->v_type == VNON || ip == NULL ||
1378             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1379              RB_EMPTY(&vp->v_rbdirty_tree))) {
1380                 return(-1);
1381         }
1382         return(0);
1383 }
1384
1385 static int
1386 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1387 {
1388         struct hammer_sync_info *info = data;
1389         struct hammer_inode *ip;
1390         int error;
1391
1392         ip = VTOI(vp);
1393         if (vp->v_type == VNON || vp->v_type == VBAD ||
1394             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1395              RB_EMPTY(&vp->v_rbdirty_tree))) {
1396                 return(0);
1397         }
1398         error = VOP_FSYNC(vp, info->waitfor);
1399         if (error)
1400                 info->error = error;
1401         return(0);
1402 }
1403