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 1.64 2008/07/01 02:08:58 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         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          * Release our buffer and flush anything left in the buffer cache.
 272          */
 273         volume->io.waitdep = 1;
 274         bp = hammer_io_release(&volume->io, 1);
 275         hammer_io_clear_modlist(&volume->io);
 276
 277         /*
 278          * There should be no references on the volume, no clusters, and
 279          * no super-clusters.
 280          */
 281         KKASSERT(volume->io.lock.refs == 0);
 282         if (bp)
 283                 brelse(bp);
 284
 285         volume->ondisk = NULL;
 286         if (volume->devvp) {
 287                 if (volume->devvp->v_rdev &&
 288                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 289                 ) {
 290                         volume->devvp->v_rdev->si_mountpoint = NULL;
 291                 }
 292                 if (ronly) {
 293                         vinvalbuf(volume->devvp, 0, 0, 0);
 294                         VOP_CLOSE(volume->devvp, FREAD);
 295                 } else {
 296                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 297                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 298                 }
 299         }
 300
 301         /*
 302          * Destroy the structure
 303          */
 304         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 305         hammer_free_volume(volume);
 306         return(0);
 307 }
 308
 309 static
 310 void
 311 hammer_free_volume(hammer_volume_t volume)
 312 {
 313         if (volume->vol_name) {
 314                 kfree(volume->vol_name, M_HAMMER);
 315                 volume->vol_name = NULL;
 316         }
 317         if (volume->devvp) {
 318                 vrele(volume->devvp);
 319                 volume->devvp = NULL;
 320         }
 321         --hammer_count_volumes;
 322         kfree(volume, M_HAMMER);
 323 }
 324
 325 /*
 326  * Get a HAMMER volume.  The volume must already exist.
 327  */
 328 hammer_volume_t
 329 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 330 {
 331         struct hammer_volume *volume;
 332
 333         /*
 334          * Locate the volume structure
 335          */
 336         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 337         if (volume == NULL) {
 338                 *errorp = ENOENT;
 339                 return(NULL);
 340         }
 341         hammer_ref(&volume->io.lock);
 342
 343         /*
 344          * Deal with on-disk info
 345          */
 346         if (volume->ondisk == NULL || volume->io.loading) {
 347                 *errorp = hammer_load_volume(volume);
 348                 if (*errorp) {
 349                         hammer_rel_volume(volume, 1);
 350                         volume = NULL;
 351                 }
 352         } else {
 353                 *errorp = 0;
 354         }
 355         return(volume);
 356 }
 357
 358 int
 359 hammer_ref_volume(hammer_volume_t volume)
 360 {
 361         int error;
 362
 363         hammer_ref(&volume->io.lock);
 364
 365         /*
 366          * Deal with on-disk info
 367          */
 368         if (volume->ondisk == NULL || volume->io.loading) {
 369                 error = hammer_load_volume(volume);
 370                 if (error)
 371                         hammer_rel_volume(volume, 1);
 372         } else {
 373                 error = 0;
 374         }
 375         return (error);
 376 }
 377
 378 hammer_volume_t
 379 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 380 {
 381         hammer_volume_t volume;
 382
 383         volume = hmp->rootvol;
 384         KKASSERT(volume != NULL);
 385         hammer_ref(&volume->io.lock);
 386
 387         /*
 388          * Deal with on-disk info
 389          */
 390         if (volume->ondisk == NULL || volume->io.loading) {
 391                 *errorp = hammer_load_volume(volume);
 392                 if (*errorp) {
 393                         hammer_rel_volume(volume, 1);
 394                         volume = NULL;
 395                 }
 396         } else {
 397                 *errorp = 0;
 398         }
 399         return (volume);
 400 }
 401
 402 /*
 403  * Load a volume's on-disk information.  The volume must be referenced and
 404  * not locked.  We temporarily acquire an exclusive lock to interlock
 405  * against releases or multiple get's.
 406  */
 407 static int
 408 hammer_load_volume(hammer_volume_t volume)
 409 {
 410         int error;
 411
 412         ++volume->io.loading;
 413         hammer_lock_ex(&volume->io.lock);
 414
 415         if (volume->ondisk == NULL) {
 416                 error = hammer_io_read(volume->devvp, &volume->io,
 417                                        volume->maxraw_off);
 418                 if (error == 0)
 419                         volume->ondisk = (void *)volume->io.bp->b_data;
 420         } else {
 421                 error = 0;
 422         }
 423         --volume->io.loading;
 424         hammer_unlock(&volume->io.lock);
 425         return(error);
 426 }
 427
 428 /*
 429  * Release a volume.  Call hammer_io_release on the last reference.  We have
 430  * to acquire an exclusive lock to interlock against volume->ondisk tests
 431  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 432  * lock to be held.
 433  *
 434  * Volumes are not unloaded from memory during normal operation.
 435  */
 436 void
 437 hammer_rel_volume(hammer_volume_t volume, int flush)
 438 {
 439         struct buf *bp = NULL;
 440
 441         crit_enter();
 442         if (volume->io.lock.refs == 1) {
 443                 ++volume->io.loading;
 444                 hammer_lock_ex(&volume->io.lock);
 445                 if (volume->io.lock.refs == 1) {
 446                         volume->ondisk = NULL;
 447                         bp = hammer_io_release(&volume->io, flush);
 448                 }
 449                 --volume->io.loading;
 450                 hammer_unlock(&volume->io.lock);
 451         }
 452         hammer_unref(&volume->io.lock);
 453         if (bp)
 454                 brelse(bp);
 455         crit_exit();
 456 }
 457
 458 /************************************************************************
 459  *                              BUFFERS                                 *
 460  ************************************************************************
 461  *
 462  * Manage buffers.  Currently all blockmap-backed zones are translated
 463  * to zone-2 buffer offsets.
 464  */
 465 hammer_buffer_t
 466 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 467                   int bytes, int isnew, int *errorp)
 468 {
 469         hammer_buffer_t buffer;
 470         hammer_volume_t volume;
 471         hammer_off_t    zone2_offset;
 472         hammer_io_type_t iotype;
 473         int vol_no;
 474         int zone;
 475
 476         buf_offset &= ~HAMMER_BUFMASK64;
 477 again:
 478         /*
 479          * Shortcut if the buffer is already cached
 480          */
 481         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 482         if (buffer) {
 483                 if (buffer->io.lock.refs == 0)
 484                         ++hammer_count_refedbufs;
 485                 hammer_ref(&buffer->io.lock);
 486
 487                 /*
 488                  * Onced refed the ondisk field will not be cleared by
 489                  * any other action.
 490                  */
 491                 if (buffer->ondisk && buffer->io.loading == 0) {
 492                         *errorp = 0;
 493                         return(buffer);
 494                 }
 495
 496                 /*
 497                  * The buffer is no longer loose if it has a ref, and
 498                  * cannot become loose once it gains a ref.  Loose
 499                  * buffers will never be in a modified state.  This should
 500                  * only occur on the 0->1 transition of refs.
 501                  *
 502                  * lose_list can be modified via a biodone() interrupt.
 503                  */
 504                 if (buffer->io.mod_list == &hmp->lose_list) {
 505                         crit_enter();   /* biodone race against list */
 506                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 507                                      mod_entry);
 508                         crit_exit();
 509                         buffer->io.mod_list = NULL;
 510                         KKASSERT(buffer->io.modified == 0);
 511                 }
 512                 goto found;
 513         }
 514
 515         /*
 516          * What is the buffer class?
 517          */
 518         zone = HAMMER_ZONE_DECODE(buf_offset);
 519
 520         switch(zone) {
 521         case HAMMER_ZONE_LARGE_DATA_INDEX:
 522         case HAMMER_ZONE_SMALL_DATA_INDEX:
 523                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 524                 break;
 525         case HAMMER_ZONE_UNDO_INDEX:
 526                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 527                 break;
 528         case HAMMER_ZONE_META_INDEX:
 529         default:
 530                 /*
 531                  * NOTE: inode data and directory entries are placed in this
 532                  * zone.  inode atime/mtime is updated in-place and thus
 533                  * buffers containing inodes must be synchronized as
 534                  * meta-buffers, same as buffers containing B-Tree info.
 535                  */
 536                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 537                 break;
 538         }
 539
 540         /*
 541          * Handle blockmap offset translations
 542          */
 543         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 544                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 545         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 546                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 547         } else {
 548                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 549                 zone2_offset = buf_offset;
 550                 *errorp = 0;
 551         }
 552         if (*errorp)
 553                 return(NULL);
 554
 555         /*
 556          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 557          * specifications.
 558          */
 559         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 560                  HAMMER_ZONE_RAW_BUFFER);
 561         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 562         volume = hammer_get_volume(hmp, vol_no, errorp);
 563         if (volume == NULL)
 564                 return(NULL);
 565
 566         KKASSERT(zone2_offset < volume->maxbuf_off);
 567
 568         /*
 569          * Allocate a new buffer structure.  We will check for races later.
 570          */
 571         ++hammer_count_buffers;
 572         buffer = kmalloc(sizeof(*buffer), M_HAMMER,
 573                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 574         buffer->zone2_offset = zone2_offset;
 575         buffer->zoneX_offset = buf_offset;
 576         buffer->volume = volume;
 577
 578         hammer_io_init(&buffer->io, hmp, iotype);
 579         buffer->io.offset = volume->ondisk->vol_buf_beg +
 580                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 581         buffer->io.bytes = bytes;
 582         TAILQ_INIT(&buffer->clist);
 583         hammer_ref(&buffer->io.lock);
 584
 585         /*
 586          * Insert the buffer into the RB tree and handle late collisions.
 587          */
 588         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 589                 hammer_unref(&buffer->io.lock);
 590                 --hammer_count_buffers;
 591                 kfree(buffer, M_HAMMER);
 592                 goto again;
 593         }
 594         ++hammer_count_refedbufs;
 595 found:
 596
 597         /*
 598          * Deal with on-disk info and loading races.
 599          */
 600         if (buffer->ondisk == NULL || buffer->io.loading) {
 601                 *errorp = hammer_load_buffer(buffer, isnew);
 602                 if (*errorp) {
 603                         hammer_rel_buffer(buffer, 1);
 604                         buffer = NULL;
 605                 }
 606         } else {
 607                 *errorp = 0;
 608         }
 609         return(buffer);
 610 }
 611
 612 /*
 613  * Destroy all buffers covering the specified zoneX offset range.  This
 614  * is called when the related blockmap layer2 entry is freed or when
 615  * a direct write bypasses our buffer/buffer-cache subsystem.
 616  *
 617  * The buffers may be referenced by the caller itself.  Setting reclaim
 618  * will cause the buffer to be destroyed when it's ref count reaches zero.
 619  */
 620 void
 621 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 622                    hammer_off_t zone2_offset, int bytes)
 623 {
 624         hammer_buffer_t buffer;
 625         hammer_volume_t volume;
 626         int vol_no;
 627         int error;
 628
 629         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 630         volume = hammer_get_volume(hmp, vol_no, &error);
 631         KKASSERT(error == 0);
 632
 633         while (bytes > 0) {
 634                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 635                                    base_offset);
 636                 if (buffer) {
 637                         KKASSERT(buffer->zone2_offset == zone2_offset);
 638                         hammer_io_clear_modify(&buffer->io, 1);
 639                         buffer->io.reclaim = 1;
 640                         KKASSERT(buffer->volume == volume);
 641                         if (buffer->io.lock.refs == 0)
 642                                 hammer_unload_buffer(buffer, NULL);
 643                 } else {
 644                         hammer_io_inval(volume, zone2_offset);
 645                 }
 646                 base_offset += HAMMER_BUFSIZE;
 647                 zone2_offset += HAMMER_BUFSIZE;
 648                 bytes -= HAMMER_BUFSIZE;
 649         }
 650         hammer_rel_volume(volume, 0);
 651 }
 652
 653 static int
 654 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 655 {
 656         hammer_volume_t volume;
 657         int error;
 658
 659         /*
 660          * Load the buffer's on-disk info
 661          */
 662         volume = buffer->volume;
 663         ++buffer->io.loading;
 664         hammer_lock_ex(&buffer->io.lock);
 665
 666         if (hammer_debug_io & 0x0001) {
 667                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 668                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 669                         buffer->ondisk);
 670         }
 671
 672         if (buffer->ondisk == NULL) {
 673                 if (isnew) {
 674                         error = hammer_io_new(volume->devvp, &buffer->io);
 675                 } else {
 676                         error = hammer_io_read(volume->devvp, &buffer->io,
 677                                                volume->maxraw_off);
 678                 }
 679                 if (error == 0)
 680                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 681         } else if (isnew) {
 682                 error = hammer_io_new(volume->devvp, &buffer->io);
 683         } else {
 684                 error = 0;
 685         }
 686         --buffer->io.loading;
 687         hammer_unlock(&buffer->io.lock);
 688         return (error);
 689 }
 690
 691 /*
 692  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 693  */
 694 int
 695 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 696 {
 697         ++hammer_count_refedbufs;
 698         hammer_ref(&buffer->io.lock);
 699         hammer_flush_buffer_nodes(buffer);
 700         KKASSERT(buffer->io.lock.refs == 1);
 701         hammer_rel_buffer(buffer, 2);
 702         return(0);
 703 }
 704
 705 /*
 706  * Reference a buffer that is either already referenced or via a specially
 707  * handled pointer (aka cursor->buffer).
 708  */
 709 int
 710 hammer_ref_buffer(hammer_buffer_t buffer)
 711 {
 712         int error;
 713
 714         if (buffer->io.lock.refs == 0)
 715                 ++hammer_count_refedbufs;
 716         hammer_ref(&buffer->io.lock);
 717
 718         /*
 719          * At this point a biodone() will not touch the buffer other then
 720          * incidental bits.  However, lose_list can be modified via
 721          * a biodone() interrupt.
 722          *
 723          * No longer loose
 724          */
 725         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 726                 crit_enter();
 727                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 728                 buffer->io.mod_list = NULL;
 729                 crit_exit();
 730         }
 731
 732         if (buffer->ondisk == NULL || buffer->io.loading) {
 733                 error = hammer_load_buffer(buffer, 0);
 734                 if (error) {
 735                         hammer_rel_buffer(buffer, 1);
 736                         /*
 737                          * NOTE: buffer pointer can become stale after
 738                          * the above release.
 739                          */
 740                 }
 741         } else {
 742                 error = 0;
 743         }
 744         return(error);
 745 }
 746
 747 /*
 748  * Release a buffer.  We have to deal with several places where
 749  * another thread can ref the buffer.
 750  *
 751  * Only destroy the structure itself if the related buffer cache buffer
 752  * was disassociated from it.  This ties the management of the structure
 753  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 754  * embedded io is referenced or not.
 755  */
 756 void
 757 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 758 {
 759         hammer_volume_t volume;
 760         struct buf *bp = NULL;
 761         int freeme = 0;
 762
 763         crit_enter();
 764         if (buffer->io.lock.refs == 1) {
 765                 ++buffer->io.loading;   /* force interlock check */
 766                 hammer_lock_ex(&buffer->io.lock);
 767                 if (buffer->io.lock.refs == 1) {
 768                         bp = hammer_io_release(&buffer->io, flush);
 769
 770                         if (buffer->io.lock.refs == 1)
 771                                 --hammer_count_refedbufs;
 772
 773                         if (buffer->io.bp == NULL &&
 774                             buffer->io.lock.refs == 1) {
 775                                 /*
 776                                  * Final cleanup
 777                                  *
 778                                  * NOTE: It is impossible for any associated
 779                                  * B-Tree nodes to have refs if the buffer
 780                                  * has no additional refs.
 781                                  */
 782                                 RB_REMOVE(hammer_buf_rb_tree,
 783                                           &buffer->io.hmp->rb_bufs_root,
 784                                           buffer);
 785                                 volume = buffer->volume;
 786                                 buffer->volume = NULL; /* sanity */
 787                                 hammer_rel_volume(volume, 0);
 788                                 hammer_io_clear_modlist(&buffer->io);
 789                                 hammer_flush_buffer_nodes(buffer);
 790                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 791                                 freeme = 1;
 792                         }
 793                 }
 794                 --buffer->io.loading;
 795                 hammer_unlock(&buffer->io.lock);
 796         }
 797         hammer_unref(&buffer->io.lock);
 798         crit_exit();
 799         if (bp)
 800                 brelse(bp);
 801         if (freeme) {
 802                 --hammer_count_buffers;
 803                 kfree(buffer, M_HAMMER);
 804         }
 805 }
 806
 807 /*
 808  * Access the filesystem buffer containing the specified hammer offset.
 809  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 810  * relative buffer offset.  It must also have bit 55 set to be valid.
 811  * (see hammer_off_t in hammer_disk.h).
 812  *
 813  * Any prior buffer in *bufferp will be released and replaced by the
 814  * requested buffer.
 815  */
 816 static __inline
 817 void *
 818 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 819              int *errorp, struct hammer_buffer **bufferp)
 820 {
 821         hammer_buffer_t buffer;
 822         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 823
 824         buf_offset &= ~HAMMER_BUFMASK64;
 825         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 826
 827         buffer = *bufferp;
 828         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 829                                buffer->zoneX_offset != buf_offset)) {
 830                 if (buffer)
 831                         hammer_rel_buffer(buffer, 0);
 832                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 833                 *bufferp = buffer;
 834         } else {
 835                 *errorp = 0;
 836         }
 837
 838         /*
 839          * Return a pointer to the buffer data.
 840          */
 841         if (buffer == NULL)
 842                 return(NULL);
 843         else
 844                 return((char *)buffer->ondisk + xoff);
 845 }
 846
 847 void *
 848 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 849              int *errorp, struct hammer_buffer **bufferp)
 850 {
 851         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 852 }
 853
 854 void *
 855 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 856                  int *errorp, struct hammer_buffer **bufferp)
 857 {
 858         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 859         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 860 }
 861
 862 /*
 863  * Access the filesystem buffer containing the specified hammer offset.
 864  * No disk read operation occurs.  The result buffer may contain garbage.
 865  *
 866  * Any prior buffer in *bufferp will be released and replaced by the
 867  * requested buffer.
 868  *
 869  * This function marks the buffer dirty but does not increment its
 870  * modify_refs count.
 871  */
 872 static __inline
 873 void *
 874 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 875              int *errorp, struct hammer_buffer **bufferp)
 876 {
 877         hammer_buffer_t buffer;
 878         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 879
 880         buf_offset &= ~HAMMER_BUFMASK64;
 881
 882         buffer = *bufferp;
 883         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 884                                buffer->zoneX_offset != buf_offset)) {
 885                 if (buffer)
 886                         hammer_rel_buffer(buffer, 0);
 887                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 888                 *bufferp = buffer;
 889         } else {
 890                 *errorp = 0;
 891         }
 892
 893         /*
 894          * Return a pointer to the buffer data.
 895          */
 896         if (buffer == NULL)
 897                 return(NULL);
 898         else
 899                 return((char *)buffer->ondisk + xoff);
 900 }
 901
 902 void *
 903 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 904              int *errorp, struct hammer_buffer **bufferp)
 905 {
 906         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 907 }
 908
 909 void *
 910 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 911                 int *errorp, struct hammer_buffer **bufferp)
 912 {
 913         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 914         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
 915 }
 916
 917 /************************************************************************
 918  *                              NODES                                   *
 919  ************************************************************************
 920  *
 921  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 922  * method used by the HAMMER filesystem.
 923  *
 924  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 925  * associated with its buffer, and will only referenced the buffer while
 926  * the node itself is referenced.
 927  *
 928  * A hammer_node can also be passively associated with other HAMMER
 929  * structures, such as inodes, while retaining 0 references.  These
 930  * associations can be cleared backwards using a pointer-to-pointer in
 931  * the hammer_node.
 932  *
 933  * This allows the HAMMER implementation to cache hammer_nodes long-term
 934  * and short-cut a great deal of the infrastructure's complexity.  In
 935  * most cases a cached node can be reacquired without having to dip into
 936  * either the buffer or cluster management code.
 937  *
 938  * The caller must pass a referenced cluster on call and will retain
 939  * ownership of the reference on return.  The node will acquire its own
 940  * additional references, if necessary.
 941  */
 942 hammer_node_t
 943 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 944                 int isnew, int *errorp)
 945 {
 946         hammer_node_t node;
 947
 948         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 949
 950         /*
 951          * Locate the structure, allocating one if necessary.
 952          */
 953 again:
 954         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 955         if (node == NULL) {
 956                 ++hammer_count_nodes;
 957                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
 958                 node->node_offset = node_offset;
 959                 node->hmp = hmp;
 960                 TAILQ_INIT(&node->cache_list);
 961                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 962                         --hammer_count_nodes;
 963                         kfree(node, M_HAMMER);
 964                         goto again;
 965                 }
 966         }
 967         hammer_ref(&node->lock);
 968         if (node->ondisk)
 969                 *errorp = 0;
 970         else
 971                 *errorp = hammer_load_node(node, isnew);
 972         if (*errorp) {
 973                 hammer_rel_node(node);
 974                 node = NULL;
 975         }
 976         return(node);
 977 }
 978
 979 /*
 980  * Reference an already-referenced node.
 981  */
 982 void
 983 hammer_ref_node(hammer_node_t node)
 984 {
 985         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 986         hammer_ref(&node->lock);
 987 }
 988
 989 /*
 990  * Load a node's on-disk data reference.
 991  */
 992 static int
 993 hammer_load_node(hammer_node_t node, int isnew)
 994 {
 995         hammer_buffer_t buffer;
 996         hammer_off_t buf_offset;
 997         int error;
 998
 999         error = 0;
1000         ++node->loading;
1001         hammer_lock_ex(&node->lock);
1002         if (node->ondisk == NULL) {
1003                 /*
1004                  * This is a little confusing but the jist is that
1005                  * node->buffer determines whether the node is on
1006                  * the buffer's clist and node->ondisk determines
1007                  * whether the buffer is referenced.
1008                  *
1009                  * We could be racing a buffer release, in which case
1010                  * node->buffer may become NULL while we are blocked
1011                  * referencing the buffer.
1012                  */
1013                 if ((buffer = node->buffer) != NULL) {
1014                         error = hammer_ref_buffer(buffer);
1015                         if (error == 0 && node->buffer == NULL) {
1016                                 TAILQ_INSERT_TAIL(&buffer->clist,
1017                                                   node, entry);
1018                                 node->buffer = buffer;
1019                         }
1020                 } else {
1021                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1022                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1023                                                    HAMMER_BUFSIZE, 0, &error);
1024                         if (buffer) {
1025                                 KKASSERT(error == 0);
1026                                 TAILQ_INSERT_TAIL(&buffer->clist,
1027                                                   node, entry);
1028                                 node->buffer = buffer;
1029                         }
1030                 }
1031                 if (error)
1032                         goto failed;
1033                 node->ondisk = (void *)((char *)buffer->ondisk +
1034                                         (node->node_offset & HAMMER_BUFMASK));
1035                 if (isnew == 0 &&
1036                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1037                         if (hammer_crc_test_btree(node->ondisk) == 0)
1038                                 Debugger("CRC FAILED: B-TREE NODE");
1039                         node->flags |= HAMMER_NODE_CRCGOOD;
1040                 }
1041         }
1042 failed:
1043         --node->loading;
1044         hammer_unlock(&node->lock);
1045         return (error);
1046 }
1047
1048 /*
1049  * Safely reference a node, interlock against flushes via the IO subsystem.
1050  */
1051 hammer_node_t
1052 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1053                      int *errorp)
1054 {
1055         hammer_node_t node;
1056
1057         node = cache->node;
1058         if (node != NULL) {
1059                 hammer_ref(&node->lock);
1060                 if (node->ondisk)
1061                         *errorp = 0;
1062                 else
1063                         *errorp = hammer_load_node(node, 0);
1064                 if (*errorp) {
1065                         hammer_rel_node(node);
1066                         node = NULL;
1067                 }
1068         } else {
1069                 *errorp = ENOENT;
1070         }
1071         return(node);
1072 }
1073
1074 /*
1075  * Release a hammer_node.  On the last release the node dereferences
1076  * its underlying buffer and may or may not be destroyed.
1077  */
1078 void
1079 hammer_rel_node(hammer_node_t node)
1080 {
1081         hammer_buffer_t buffer;
1082
1083         /*
1084          * If this isn't the last ref just decrement the ref count and
1085          * return.
1086          */
1087         if (node->lock.refs > 1) {
1088                 hammer_unref(&node->lock);
1089                 return;
1090         }
1091
1092         /*
1093          * If there is no ondisk info or no buffer the node failed to load,
1094          * remove the last reference and destroy the node.
1095          */
1096         if (node->ondisk == NULL) {
1097                 hammer_unref(&node->lock);
1098                 hammer_flush_node(node);
1099                 /* node is stale now */
1100                 return;
1101         }
1102
1103         /*
1104          * Do not disassociate the node from the buffer if it represents
1105          * a modified B-Tree node that still needs its crc to be generated.
1106          */
1107         if (node->flags & HAMMER_NODE_NEEDSCRC)
1108                 return;
1109
1110         /*
1111          * Do final cleanups and then either destroy the node and leave it
1112          * passively cached.  The buffer reference is removed regardless.
1113          */
1114         buffer = node->buffer;
1115         node->ondisk = NULL;
1116
1117         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1118                 hammer_unref(&node->lock);
1119                 hammer_rel_buffer(buffer, 0);
1120                 return;
1121         }
1122
1123         /*
1124          * Destroy the node.
1125          */
1126         hammer_unref(&node->lock);
1127         hammer_flush_node(node);
1128         /* node is stale */
1129         hammer_rel_buffer(buffer, 0);
1130 }
1131
1132 /*
1133  * Free space on-media associated with a B-Tree node.
1134  */
1135 void
1136 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1137 {
1138         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1139         node->flags |= HAMMER_NODE_DELETED;
1140         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1141 }
1142
1143 /*
1144  * Passively cache a referenced hammer_node.  The caller may release
1145  * the node on return.
1146  */
1147 void
1148 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1149 {
1150         /*
1151          * If the node is being deleted, don't cache it!
1152          */
1153         if (node->flags & HAMMER_NODE_DELETED)
1154                 return;
1155         if (cache->node == node)
1156                 return;
1157         while (cache->node)
1158                 hammer_uncache_node(cache);
1159         if (node->flags & HAMMER_NODE_DELETED)
1160                 return;
1161         cache->node = node;
1162         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1163 }
1164
1165 void
1166 hammer_uncache_node(hammer_node_cache_t cache)
1167 {
1168         hammer_node_t node;
1169
1170         if ((node = cache->node) != NULL) {
1171                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1172                 cache->node = NULL;
1173                 if (TAILQ_EMPTY(&node->cache_list))
1174                         hammer_flush_node(node);
1175         }
1176 }
1177
1178 /*
1179  * Remove a node's cache references and destroy the node if it has no
1180  * other references or backing store.
1181  */
1182 void
1183 hammer_flush_node(hammer_node_t node)
1184 {
1185         hammer_node_cache_t cache;
1186         hammer_buffer_t buffer;
1187
1188         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1189                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1190                 cache->node = NULL;
1191         }
1192         if (node->lock.refs == 0 && node->ondisk == NULL) {
1193                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1194                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1195                 if ((buffer = node->buffer) != NULL) {
1196                         node->buffer = NULL;
1197                         TAILQ_REMOVE(&buffer->clist, node, entry);
1198                         /* buffer is unreferenced because ondisk is NULL */
1199                 }
1200                 --hammer_count_nodes;
1201                 kfree(node, M_HAMMER);
1202         }
1203 }
1204
1205 /*
1206  * Flush passively cached B-Tree nodes associated with this buffer.
1207  * This is only called when the buffer is about to be destroyed, so
1208  * none of the nodes should have any references.  The buffer is locked.
1209  *
1210  * We may be interlocked with the buffer.
1211  */
1212 void
1213 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1214 {
1215         hammer_node_t node;
1216
1217         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1218                 KKASSERT(node->ondisk == NULL);
1219                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1220
1221                 if (node->lock.refs == 0) {
1222                         hammer_ref(&node->lock);
1223                         node->flags |= HAMMER_NODE_FLUSH;
1224                         hammer_rel_node(node);
1225                 } else {
1226                         KKASSERT(node->loading != 0);
1227                         KKASSERT(node->buffer != NULL);
1228                         buffer = node->buffer;
1229                         node->buffer = NULL;
1230                         TAILQ_REMOVE(&buffer->clist, node, entry);
1231                         /* buffer is unreferenced because ondisk is NULL */
1232                 }
1233         }
1234 }
1235
1236
1237 /************************************************************************
1238  *                              ALLOCATORS                              *
1239  ************************************************************************/
1240
1241 /*
1242  * Allocate a B-Tree node.
1243  */
1244 hammer_node_t
1245 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1246 {
1247         hammer_buffer_t buffer = NULL;
1248         hammer_node_t node = NULL;
1249         hammer_off_t node_offset;
1250
1251         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1252                                             sizeof(struct hammer_node_ondisk),
1253                                             errorp);
1254         if (*errorp == 0) {
1255                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1256                 hammer_modify_node_noundo(trans, node);
1257                 bzero(node->ondisk, sizeof(*node->ondisk));
1258                 hammer_modify_node_done(node);
1259         }
1260         if (buffer)
1261                 hammer_rel_buffer(buffer, 0);
1262         return(node);
1263 }
1264
1265 /*
1266  * Allocate data.  If the address of a data buffer is supplied then
1267  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1268  * will be set to the related buffer.  The caller must release it when
1269  * finally done.  The initial *data_bufferp should be set to NULL by
1270  * the caller.
1271  *
1272  * The caller is responsible for making hammer_modify*() calls on the
1273  * *data_bufferp.
1274  */
1275 void *
1276 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1277                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1278                   struct hammer_buffer **data_bufferp, int *errorp)
1279 {
1280         void *data;
1281         int zone;
1282
1283         /*
1284          * Allocate data
1285          */
1286         if (data_len) {
1287                 switch(rec_type) {
1288                 case HAMMER_RECTYPE_INODE:
1289                 case HAMMER_RECTYPE_DIRENTRY:
1290                 case HAMMER_RECTYPE_EXT:
1291                 case HAMMER_RECTYPE_FIX:
1292                         zone = HAMMER_ZONE_META_INDEX;
1293                         break;
1294                 case HAMMER_RECTYPE_DATA:
1295                 case HAMMER_RECTYPE_DB:
1296                         if (data_len <= HAMMER_BUFSIZE / 2) {
1297                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1298                         } else {
1299                                 data_len = (data_len + HAMMER_BUFMASK) &
1300                                            ~HAMMER_BUFMASK;
1301                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1302                         }
1303                         break;
1304                 default:
1305                         panic("hammer_alloc_data: rec_type %04x unknown",
1306                               rec_type);
1307                         zone = 0;       /* NOT REACHED */
1308                         break;
1309                 }
1310                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1311                                                       data_len, errorp);
1312         } else {
1313                 *data_offsetp = 0;
1314         }
1315         if (*errorp == 0 && data_bufferp) {
1316                 if (data_len) {
1317                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1318                                                 data_len, errorp, data_bufferp);
1319                         KKASSERT(*errorp == 0);
1320                 } else {
1321                         data = NULL;
1322                 }
1323         } else {
1324                 data = NULL;
1325         }
1326         KKASSERT(*errorp == 0);
1327         return(data);
1328 }
1329
1330 /*
1331  * Sync dirty buffers to the media and clean-up any loose ends.
1332  */
1333 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1334 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1335
1336 int
1337 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1338 {
1339         struct hammer_sync_info info;
1340
1341         info.error = 0;
1342         info.waitfor = waitfor;
1343         if (waitfor == MNT_WAIT) {
1344                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1345                               hammer_sync_scan1, hammer_sync_scan2, &info);
1346         } else {
1347                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1348                               hammer_sync_scan1, hammer_sync_scan2, &info);
1349         }
1350         return(info.error);
1351 }
1352
1353 /*
1354  * Filesystem sync.  If doing a synchronous sync make a second pass on
1355  * the vnodes in case any were already flushing during the first pass,
1356  * and activate the flusher twice (the second time brings the UNDO FIFO's
1357  * start position up to the end position after the first call).
1358  */
1359 int
1360 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1361 {
1362         struct hammer_sync_info info;
1363
1364         info.error = 0;
1365         info.waitfor = MNT_NOWAIT;
1366         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1367                       hammer_sync_scan1, hammer_sync_scan2, &info);
1368         if (info.error == 0 && waitfor == MNT_WAIT) {
1369                 info.waitfor = waitfor;
1370                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1371                               hammer_sync_scan1, hammer_sync_scan2, &info);
1372         }
1373         if (waitfor == MNT_WAIT) {
1374                 hammer_flusher_sync(hmp);
1375                 hammer_flusher_sync(hmp);
1376         } else {
1377                 hammer_flusher_async(hmp);
1378         }
1379         return(info.error);
1380 }
1381
1382 static int
1383 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1384 {
1385         struct hammer_inode *ip;
1386
1387         ip = VTOI(vp);
1388         if (vp->v_type == VNON || ip == NULL ||
1389             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1390              RB_EMPTY(&vp->v_rbdirty_tree))) {
1391                 return(-1);
1392         }
1393         return(0);
1394 }
1395
1396 static int
1397 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1398 {
1399         struct hammer_sync_info *info = data;
1400         struct hammer_inode *ip;
1401         int error;
1402
1403         ip = VTOI(vp);
1404         if (vp->v_type == VNON || vp->v_type == VBAD ||
1405             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1406              RB_EMPTY(&vp->v_rbdirty_tree))) {
1407                 return(0);
1408         }
1409         error = VOP_FSYNC(vp, info->waitfor);
1410         if (error)
1411                 info->error = error;
1412         return(0);
1413 }
1414