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 1.53 2008/06/10 22:30:21 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 /*
  54  * Red-Black tree support for various structures
  55  */
  56 int
  57 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  58 {
  59         if (ip1->obj_id < ip2->obj_id)
  60                 return(-1);
  61         if (ip1->obj_id > ip2->obj_id)
  62                 return(1);
  63         if (ip1->obj_asof < ip2->obj_asof)
  64                 return(-1);
  65         if (ip1->obj_asof > ip2->obj_asof)
  66                 return(1);
  67         return(0);
  68 }
  69
  70 static int
  71 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  72 {
  73         if (info->obj_id < ip->obj_id)
  74                 return(-1);
  75         if (info->obj_id > ip->obj_id)
  76                 return(1);
  77         if (info->obj_asof < ip->obj_asof)
  78                 return(-1);
  79         if (info->obj_asof > ip->obj_asof)
  80                 return(1);
  81         return(0);
  82 }
  83
  84 static int
  85 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  86 {
  87         if (vol1->vol_no < vol2->vol_no)
  88                 return(-1);
  89         if (vol1->vol_no > vol2->vol_no)
  90                 return(1);
  91         return(0);
  92 }
  93
  94 static int
  95 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  96 {
  97         if (buf1->zoneX_offset < buf2->zoneX_offset)
  98                 return(-1);
  99         if (buf1->zoneX_offset > buf2->zoneX_offset)
 100                 return(1);
 101         return(0);
 102 }
 103
 104 static int
 105 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 106 {
 107         if (node1->node_offset < node2->node_offset)
 108                 return(-1);
 109         if (node1->node_offset > node2->node_offset)
 110                 return(1);
 111         return(0);
 112 }
 113
 114 /*
 115  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 116  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 117  * functions are normal, e.g. hammer_buf_rb_tree_RB_LOOKUP(root, zone2_offset).
 118  */
 119 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 120 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 121                 hammer_inode_info_cmp, hammer_inode_info_t);
 122 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 123              hammer_vol_rb_compare, int32_t, vol_no);
 124 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 125              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
 126 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 127              hammer_nod_rb_compare, hammer_off_t, node_offset);
 128
 129 /************************************************************************
 130  *                              VOLUMES                                 *
 131  ************************************************************************
 132  *
 133  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 134  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 135  * not load a new volume.
 136  *
 137  * Calls made to hammer_load_volume() or single-threaded
 138  */
 139 int
 140 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 141 {
 142         struct mount *mp;
 143         hammer_volume_t volume;
 144         struct hammer_volume_ondisk *ondisk;
 145         struct nlookupdata nd;
 146         struct buf *bp = NULL;
 147         int error;
 148         int ronly;
 149         int setmp = 0;
 150
 151         mp = hmp->mp;
 152         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 153
 154         /*
 155          * Allocate a volume structure
 156          */
 157         ++hammer_count_volumes;
 158         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 159         volume->vol_name = kstrdup(volname, M_HAMMER);
 160         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 161         volume->io.offset = 0LL;
 162
 163         /*
 164          * Get the device vnode
 165          */
 166         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 167         if (error == 0)
 168                 error = nlookup(&nd);
 169         if (error == 0)
 170                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 171         nlookup_done(&nd);
 172         if (error == 0) {
 173                 if (vn_isdisk(volume->devvp, &error)) {
 174                         error = vfs_mountedon(volume->devvp);
 175                 }
 176         }
 177         if (error == 0 &&
 178             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 179                 error = EBUSY;
 180         }
 181         if (error == 0) {
 182                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 183                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 184                 if (error == 0) {
 185                         error = VOP_OPEN(volume->devvp,
 186                                          (ronly ? FREAD : FREAD|FWRITE),
 187                                          FSCRED, NULL);
 188                 }
 189                 vn_unlock(volume->devvp);
 190         }
 191         if (error) {
 192                 hammer_free_volume(volume);
 193                 return(error);
 194         }
 195         volume->devvp->v_rdev->si_mountpoint = mp;
 196         setmp = 1;
 197
 198         /*
 199          * Extract the volume number from the volume header and do various
 200          * sanity checks.
 201          */
 202         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 203         if (error)
 204                 goto late_failure;
 205         ondisk = (void *)bp->b_data;
 206         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 207                 kprintf("hammer_mount: volume %s has an invalid header\n",
 208                         volume->vol_name);
 209                 error = EFTYPE;
 210                 goto late_failure;
 211         }
 212         volume->vol_no = ondisk->vol_no;
 213         volume->buffer_base = ondisk->vol_buf_beg;
 214         volume->vol_flags = ondisk->vol_flags;
 215         volume->nblocks = ondisk->vol_nblocks;
 216         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 217                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 218         volume->maxraw_off = ondisk->vol_buf_end;
 219
 220         if (RB_EMPTY(&hmp->rb_vols_root)) {
 221                 hmp->fsid = ondisk->vol_fsid;
 222         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 223                 kprintf("hammer_mount: volume %s's fsid does not match "
 224                         "other volumes\n", volume->vol_name);
 225                 error = EFTYPE;
 226                 goto late_failure;
 227         }
 228
 229         /*
 230          * Insert the volume structure into the red-black tree.
 231          */
 232         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 233                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 234                         volume->vol_name, volume->vol_no);
 235                 error = EEXIST;
 236         }
 237
 238         /*
 239          * Set the root volume .  HAMMER special cases rootvol the structure.
 240          * We do not hold a ref because this would prevent related I/O
 241          * from being flushed.
 242          */
 243         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 244                 hmp->rootvol = volume;
 245                 if (bp) {
 246                         brelse(bp);
 247                         bp = NULL;
 248                 }
 249                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 250                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 251                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 252                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 253                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 254         }
 255 late_failure:
 256         if (bp)
 257                 brelse(bp);
 258         if (error) {
 259                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 260                 if (setmp)
 261                         volume->devvp->v_rdev->si_mountpoint = NULL;
 262                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 263                 hammer_free_volume(volume);
 264         }
 265         return (error);
 266 }
 267
 268 /*
 269  * This is called for each volume when updating the mount point from
 270  * read-write to read-only or vise-versa.
 271  */
 272 int
 273 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 274 {
 275         if (volume->devvp) {
 276                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 277                 if (volume->io.hmp->ronly) {
 278                         /* do not call vinvalbuf */
 279                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 280                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 281                 } else {
 282                         /* do not call vinvalbuf */
 283                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 284                         VOP_CLOSE(volume->devvp, FREAD);
 285                 }
 286                 vn_unlock(volume->devvp);
 287         }
 288         return(0);
 289 }
 290
 291 /*
 292  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 293  * so returns -1 on failure.
 294  */
 295 int
 296 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 297 {
 298         struct hammer_mount *hmp = volume->io.hmp;
 299         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 300
 301         /*
 302          * Clean up the root volume pointer, which is held unlocked in hmp.
 303          */
 304         if (hmp->rootvol == volume)
 305                 hmp->rootvol = NULL;
 306
 307         /*
 308          * Release our buffer and flush anything left in the buffer cache.
 309          */
 310         volume->io.waitdep = 1;
 311         hammer_io_release(&volume->io, 1);
 312         hammer_io_clear_modlist(&volume->io);
 313
 314         /*
 315          * There should be no references on the volume, no clusters, and
 316          * no super-clusters.
 317          */
 318         KKASSERT(volume->io.lock.refs == 0);
 319
 320         volume->ondisk = NULL;
 321         if (volume->devvp) {
 322                 if (volume->devvp->v_rdev &&
 323                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 324                 ) {
 325                         volume->devvp->v_rdev->si_mountpoint = NULL;
 326                 }
 327                 if (ronly) {
 328                         vinvalbuf(volume->devvp, 0, 0, 0);
 329                         VOP_CLOSE(volume->devvp, FREAD);
 330                 } else {
 331                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 332                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 333                 }
 334         }
 335
 336         /*
 337          * Destroy the structure
 338          */
 339         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 340         hammer_free_volume(volume);
 341         return(0);
 342 }
 343
 344 static
 345 void
 346 hammer_free_volume(hammer_volume_t volume)
 347 {
 348         if (volume->vol_name) {
 349                 kfree(volume->vol_name, M_HAMMER);
 350                 volume->vol_name = NULL;
 351         }
 352         if (volume->devvp) {
 353                 vrele(volume->devvp);
 354                 volume->devvp = NULL;
 355         }
 356         --hammer_count_volumes;
 357         kfree(volume, M_HAMMER);
 358 }
 359
 360 /*
 361  * Get a HAMMER volume.  The volume must already exist.
 362  */
 363 hammer_volume_t
 364 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 365 {
 366         struct hammer_volume *volume;
 367
 368         /*
 369          * Locate the volume structure
 370          */
 371         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 372         if (volume == NULL) {
 373                 *errorp = ENOENT;
 374                 return(NULL);
 375         }
 376         hammer_ref(&volume->io.lock);
 377
 378         /*
 379          * Deal with on-disk info
 380          */
 381         if (volume->ondisk == NULL || volume->io.loading) {
 382                 *errorp = hammer_load_volume(volume);
 383                 if (*errorp) {
 384                         hammer_rel_volume(volume, 1);
 385                         volume = NULL;
 386                 }
 387         } else {
 388                 *errorp = 0;
 389         }
 390         return(volume);
 391 }
 392
 393 int
 394 hammer_ref_volume(hammer_volume_t volume)
 395 {
 396         int error;
 397
 398         hammer_ref(&volume->io.lock);
 399
 400         /*
 401          * Deal with on-disk info
 402          */
 403         if (volume->ondisk == NULL || volume->io.loading) {
 404                 error = hammer_load_volume(volume);
 405                 if (error)
 406                         hammer_rel_volume(volume, 1);
 407         } else {
 408                 error = 0;
 409         }
 410         return (error);
 411 }
 412
 413 hammer_volume_t
 414 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 415 {
 416         hammer_volume_t volume;
 417
 418         volume = hmp->rootvol;
 419         KKASSERT(volume != NULL);
 420         hammer_ref(&volume->io.lock);
 421
 422         /*
 423          * Deal with on-disk info
 424          */
 425         if (volume->ondisk == NULL || volume->io.loading) {
 426                 *errorp = hammer_load_volume(volume);
 427                 if (*errorp) {
 428                         hammer_rel_volume(volume, 1);
 429                         volume = NULL;
 430                 }
 431         } else {
 432                 *errorp = 0;
 433         }
 434         return (volume);
 435 }
 436
 437 /*
 438  * Load a volume's on-disk information.  The volume must be referenced and
 439  * not locked.  We temporarily acquire an exclusive lock to interlock
 440  * against releases or multiple get's.
 441  */
 442 static int
 443 hammer_load_volume(hammer_volume_t volume)
 444 {
 445         int error;
 446
 447         ++volume->io.loading;
 448         hammer_lock_ex(&volume->io.lock);
 449
 450         if (volume->ondisk == NULL) {
 451                 error = hammer_io_read(volume->devvp, &volume->io,
 452                                        volume->maxraw_off);
 453                 if (error == 0)
 454                         volume->ondisk = (void *)volume->io.bp->b_data;
 455         } else {
 456                 error = 0;
 457         }
 458         --volume->io.loading;
 459         hammer_unlock(&volume->io.lock);
 460         return(error);
 461 }
 462
 463 /*
 464  * Release a volume.  Call hammer_io_release on the last reference.  We have
 465  * to acquire an exclusive lock to interlock against volume->ondisk tests
 466  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 467  * lock to be held.
 468  *
 469  * Volumes are not unloaded from memory during normal operation.
 470  */
 471 void
 472 hammer_rel_volume(hammer_volume_t volume, int flush)
 473 {
 474         crit_enter();
 475         if (volume->io.lock.refs == 1) {
 476                 ++volume->io.loading;
 477                 hammer_lock_ex(&volume->io.lock);
 478                 if (volume->io.lock.refs == 1) {
 479                         volume->ondisk = NULL;
 480                         hammer_io_release(&volume->io, flush);
 481                 }
 482                 --volume->io.loading;
 483                 hammer_unlock(&volume->io.lock);
 484         }
 485         hammer_unref(&volume->io.lock);
 486         crit_exit();
 487 }
 488
 489 /************************************************************************
 490  *                              BUFFERS                                 *
 491  ************************************************************************
 492  *
 493  * Manage buffers.  Currently all blockmap-backed zones are translated
 494  * to zone-2 buffer offsets.
 495  */
 496 hammer_buffer_t
 497 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 498                   int isnew, int *errorp)
 499 {
 500         hammer_buffer_t buffer;
 501         hammer_volume_t volume;
 502         hammer_off_t    zone2_offset;
 503         hammer_io_type_t iotype;
 504         int vol_no;
 505         int zone;
 506
 507 again:
 508         /*
 509          * Shortcut if the buffer is already cached
 510          */
 511         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 512                            buf_offset & ~HAMMER_BUFMASK64);
 513         if (buffer) {
 514                 hammer_ref(&buffer->io.lock);
 515                 if (buffer->ondisk && buffer->io.loading == 0) {
 516                         *errorp = 0;
 517                         return(buffer);
 518                 }
 519
 520                 /*
 521                  * The buffer is no longer loose if it has a ref.  Loose
 522                  * buffers will never be in a modified state.  This should
 523                  * only occur on the 0->1 transition of refs.
 524                  */
 525                 if (buffer->io.mod_list == &hmp->lose_list) {
 526                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 527                                      mod_entry);
 528                         buffer->io.mod_list = NULL;
 529                         KKASSERT(buffer->io.modified == 0);
 530                 }
 531                 goto found;
 532         }
 533
 534         /*
 535          * What is the buffer class?
 536          */
 537         zone = HAMMER_ZONE_DECODE(buf_offset);
 538
 539         switch(zone) {
 540         case HAMMER_ZONE_LARGE_DATA_INDEX:
 541         case HAMMER_ZONE_SMALL_DATA_INDEX:
 542                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 543                 break;
 544         case HAMMER_ZONE_UNDO_INDEX:
 545                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 546                 break;
 547         default:
 548                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 549                 break;
 550         }
 551
 552         /*
 553          * Handle blockmap offset translations
 554          */
 555         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 556                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 557         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 558                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 559         } else {
 560                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 561                 zone2_offset = buf_offset;
 562                 *errorp = 0;
 563         }
 564         if (*errorp)
 565                 return(NULL);
 566
 567         /*
 568          * Calculate the base zone2-offset and acquire the volume
 569          *
 570          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 571          * specifications.
 572          */
 573         zone2_offset &= ~HAMMER_BUFMASK64;
 574         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 575                  HAMMER_ZONE_RAW_BUFFER);
 576         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 577         volume = hammer_get_volume(hmp, vol_no, errorp);
 578         if (volume == NULL)
 579                 return(NULL);
 580
 581         KKASSERT(zone2_offset < volume->maxbuf_off);
 582
 583         /*
 584          * Allocate a new buffer structure.  We will check for races later.
 585          */
 586         ++hammer_count_buffers;
 587         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 588         buffer->zone2_offset = zone2_offset;
 589         buffer->zoneX_offset = buf_offset;
 590         buffer->volume = volume;
 591
 592         hammer_io_init(&buffer->io, hmp, iotype);
 593         buffer->io.offset = volume->ondisk->vol_buf_beg +
 594                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 595         TAILQ_INIT(&buffer->clist);
 596         hammer_ref(&buffer->io.lock);
 597
 598         /*
 599          * Insert the buffer into the RB tree and handle late collisions.
 600          */
 601         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 602                 hammer_unref(&buffer->io.lock);
 603                 --hammer_count_buffers;
 604                 kfree(buffer, M_HAMMER);
 605                 goto again;
 606         }
 607 found:
 608
 609         /*
 610          * Deal with on-disk info and loading races.
 611          */
 612         if (buffer->ondisk == NULL || buffer->io.loading) {
 613                 *errorp = hammer_load_buffer(buffer, isnew);
 614                 if (*errorp) {
 615                         hammer_rel_buffer(buffer, 1);
 616                         buffer = NULL;
 617                 }
 618         } else {
 619                 *errorp = 0;
 620         }
 621         return(buffer);
 622 }
 623
 624 /*
 625  * Destroy all buffers covering the specified zoneX offset range.  This
 626  * is called when the related blockmap layer2 entry is freed or when
 627  * a direct write bypasses our buffer/buffer-cache subsystem.
 628  *
 629  * The buffers may be referenced by the caller itself.  Setting reclaim
 630  * will cause the buffer to be destroyed when it's ref count reaches zero.
 631  */
 632 void
 633 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 634                    hammer_off_t zone2_offset, int bytes)
 635 {
 636         hammer_buffer_t buffer;
 637         hammer_volume_t volume;
 638         int vol_no;
 639         int error;
 640
 641         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 642         volume = hammer_get_volume(hmp, vol_no, &error);
 643         KKASSERT(error == 0);
 644
 645         while (bytes > 0) {
 646                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 647                                    base_offset);
 648                 if (buffer) {
 649                         KKASSERT(buffer->zone2_offset == zone2_offset);
 650                         hammer_io_clear_modify(&buffer->io);
 651                         buffer->io.reclaim = 1;
 652                         KKASSERT(buffer->volume == volume);
 653                         if (buffer->io.lock.refs == 0)
 654                                 hammer_unload_buffer(buffer, NULL);
 655                 } else {
 656                         hammer_io_inval(volume, zone2_offset);
 657                 }
 658                 base_offset += HAMMER_BUFSIZE;
 659                 zone2_offset += HAMMER_BUFSIZE;
 660                 bytes -= HAMMER_BUFSIZE;
 661         }
 662         hammer_rel_volume(volume, 0);
 663 }
 664
 665 static int
 666 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 667 {
 668         hammer_volume_t volume;
 669         int error;
 670
 671         /*
 672          * Load the buffer's on-disk info
 673          */
 674         volume = buffer->volume;
 675         ++buffer->io.loading;
 676         hammer_lock_ex(&buffer->io.lock);
 677
 678         if (hammer_debug_io & 0x0001) {
 679                 kprintf("load_buffer %016llx %016llx\n",
 680                         buffer->zoneX_offset, buffer->zone2_offset);
 681         }
 682
 683         if (buffer->ondisk == NULL) {
 684                 if (isnew) {
 685                         error = hammer_io_new(volume->devvp, &buffer->io);
 686                 } else {
 687                         error = hammer_io_read(volume->devvp, &buffer->io,
 688                                                volume->maxraw_off);
 689                 }
 690                 if (error == 0)
 691                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 692         } else if (isnew) {
 693                 error = hammer_io_new(volume->devvp, &buffer->io);
 694         } else {
 695                 error = 0;
 696         }
 697         --buffer->io.loading;
 698         hammer_unlock(&buffer->io.lock);
 699         return (error);
 700 }
 701
 702 /*
 703  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 704  */
 705 int
 706 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 707 {
 708         hammer_ref(&buffer->io.lock);
 709         hammer_flush_buffer_nodes(buffer);
 710         KKASSERT(buffer->io.lock.refs == 1);
 711         hammer_rel_buffer(buffer, 2);
 712         return(0);
 713 }
 714
 715 /*
 716  * Reference a buffer that is either already referenced or via a specially
 717  * handled pointer (aka cursor->buffer).
 718  */
 719 int
 720 hammer_ref_buffer(hammer_buffer_t buffer)
 721 {
 722         int error;
 723
 724         hammer_ref(&buffer->io.lock);
 725
 726         /*
 727          * No longer loose
 728          */
 729         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 730                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 731                 buffer->io.mod_list = NULL;
 732         }
 733
 734         if (buffer->ondisk == NULL || buffer->io.loading) {
 735                 error = hammer_load_buffer(buffer, 0);
 736                 if (error) {
 737                         hammer_rel_buffer(buffer, 1);
 738                         /*
 739                          * NOTE: buffer pointer can become stale after
 740                          * the above release.
 741                          */
 742                 }
 743         } else {
 744                 error = 0;
 745         }
 746         return(error);
 747 }
 748
 749 /*
 750  * Release a buffer.  We have to deal with several places where
 751  * another thread can ref the buffer.
 752  *
 753  * Only destroy the structure itself if the related buffer cache buffer
 754  * was disassociated from it.  This ties the management of the structure
 755  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 756  * embedded io is referenced or not.
 757  */
 758 void
 759 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 760 {
 761         hammer_volume_t volume;
 762         int freeme = 0;
 763
 764         crit_enter();
 765         if (buffer->io.lock.refs == 1) {
 766                 ++buffer->io.loading;   /* force interlock check */
 767                 hammer_lock_ex(&buffer->io.lock);
 768                 if (buffer->io.lock.refs == 1) {
 769                         hammer_io_release(&buffer->io, flush);
 770                         hammer_flush_buffer_nodes(buffer);
 771                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 772
 773                         if (buffer->io.bp == NULL &&
 774                             buffer->io.lock.refs == 1) {
 775                                 /*
 776                                  * Final cleanup
 777                                  */
 778                                 RB_REMOVE(hammer_buf_rb_tree,
 779                                           &buffer->io.hmp->rb_bufs_root,
 780                                           buffer);
 781                                 volume = buffer->volume;
 782                                 buffer->volume = NULL; /* sanity */
 783                                 hammer_rel_volume(volume, 0);
 784                                 hammer_io_clear_modlist(&buffer->io);
 785                                 freeme = 1;
 786                         }
 787                 }
 788                 --buffer->io.loading;
 789                 hammer_unlock(&buffer->io.lock);
 790         }
 791         hammer_unref(&buffer->io.lock);
 792         crit_exit();
 793         if (freeme) {
 794                 --hammer_count_buffers;
 795                 kfree(buffer, M_HAMMER);
 796         }
 797 }
 798
 799 /*
 800  * Access the filesystem buffer containing the specified hammer offset.
 801  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 802  * relative buffer offset.  It must also have bit 55 set to be valid.
 803  * (see hammer_off_t in hammer_disk.h).
 804  *
 805  * Any prior buffer in *bufferp will be released and replaced by the
 806  * requested buffer.
 807  */
 808 void *
 809 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 810              struct hammer_buffer **bufferp)
 811 {
 812         hammer_buffer_t buffer;
 813         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 814
 815         buf_offset &= ~HAMMER_BUFMASK64;
 816         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 817
 818         buffer = *bufferp;
 819         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 820                                buffer->zoneX_offset != buf_offset)) {
 821                 if (buffer)
 822                         hammer_rel_buffer(buffer, 0);
 823                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 824                 *bufferp = buffer;
 825         } else {
 826                 *errorp = 0;
 827         }
 828
 829         /*
 830          * Return a pointer to the buffer data.
 831          */
 832         if (buffer == NULL)
 833                 return(NULL);
 834         else
 835                 return((char *)buffer->ondisk + xoff);
 836 }
 837
 838 /*
 839  * Access the filesystem buffer containing the specified hammer offset.
 840  * No disk read operation occurs.  The result buffer may contain garbage.
 841  *
 842  * Any prior buffer in *bufferp will be released and replaced by the
 843  * requested buffer.
 844  *
 845  * This function marks the buffer dirty but does not increment its
 846  * modify_refs count.
 847  */
 848 void *
 849 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 850              struct hammer_buffer **bufferp)
 851 {
 852         hammer_buffer_t buffer;
 853         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 854
 855         buf_offset &= ~HAMMER_BUFMASK64;
 856
 857         buffer = *bufferp;
 858         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 859                                buffer->zoneX_offset != buf_offset)) {
 860                 if (buffer)
 861                         hammer_rel_buffer(buffer, 0);
 862                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 863                 *bufferp = buffer;
 864         } else {
 865                 *errorp = 0;
 866         }
 867
 868         /*
 869          * Return a pointer to the buffer data.
 870          */
 871         if (buffer == NULL)
 872                 return(NULL);
 873         else
 874                 return((char *)buffer->ondisk + xoff);
 875 }
 876
 877 /************************************************************************
 878  *                              NODES                                   *
 879  ************************************************************************
 880  *
 881  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 882  * method used by the HAMMER filesystem.
 883  *
 884  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 885  * associated with its buffer, and will only referenced the buffer while
 886  * the node itself is referenced.
 887  *
 888  * A hammer_node can also be passively associated with other HAMMER
 889  * structures, such as inodes, while retaining 0 references.  These
 890  * associations can be cleared backwards using a pointer-to-pointer in
 891  * the hammer_node.
 892  *
 893  * This allows the HAMMER implementation to cache hammer_nodes long-term
 894  * and short-cut a great deal of the infrastructure's complexity.  In
 895  * most cases a cached node can be reacquired without having to dip into
 896  * either the buffer or cluster management code.
 897  *
 898  * The caller must pass a referenced cluster on call and will retain
 899  * ownership of the reference on return.  The node will acquire its own
 900  * additional references, if necessary.
 901  */
 902 hammer_node_t
 903 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 904                 int isnew, int *errorp)
 905 {
 906         hammer_node_t node;
 907
 908         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 909
 910         /*
 911          * Locate the structure, allocating one if necessary.
 912          */
 913 again:
 914         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 915         if (node == NULL) {
 916                 ++hammer_count_nodes;
 917                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 918                 node->node_offset = node_offset;
 919                 node->hmp = hmp;
 920                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 921                         --hammer_count_nodes;
 922                         kfree(node, M_HAMMER);
 923                         goto again;
 924                 }
 925         }
 926         hammer_ref(&node->lock);
 927         if (node->ondisk)
 928                 *errorp = 0;
 929         else
 930                 *errorp = hammer_load_node(node, isnew);
 931         if (*errorp) {
 932                 hammer_rel_node(node);
 933                 node = NULL;
 934         }
 935         return(node);
 936 }
 937
 938 /*
 939  * Reference an already-referenced node.
 940  */
 941 void
 942 hammer_ref_node(hammer_node_t node)
 943 {
 944         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 945         hammer_ref(&node->lock);
 946 }
 947
 948 /*
 949  * Load a node's on-disk data reference.
 950  */
 951 static int
 952 hammer_load_node(hammer_node_t node, int isnew)
 953 {
 954         hammer_buffer_t buffer;
 955         hammer_off_t buf_offset;
 956         int error;
 957
 958         error = 0;
 959         ++node->loading;
 960         hammer_lock_ex(&node->lock);
 961         if (node->ondisk == NULL) {
 962                 /*
 963                  * This is a little confusing but the jist is that
 964                  * node->buffer determines whether the node is on
 965                  * the buffer's clist and node->ondisk determines
 966                  * whether the buffer is referenced.
 967                  *
 968                  * We could be racing a buffer release, in which case
 969                  * node->buffer may become NULL while we are blocked
 970                  * referencing the buffer.
 971                  */
 972                 if ((buffer = node->buffer) != NULL) {
 973                         error = hammer_ref_buffer(buffer);
 974                         if (error == 0 && node->buffer == NULL) {
 975                                 TAILQ_INSERT_TAIL(&buffer->clist,
 976                                                   node, entry);
 977                                 node->buffer = buffer;
 978                         }
 979                 } else {
 980                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
 981                         buffer = hammer_get_buffer(node->hmp, buf_offset,
 982                                                    0, &error);
 983                         if (buffer) {
 984                                 KKASSERT(error == 0);
 985                                 TAILQ_INSERT_TAIL(&buffer->clist,
 986                                                   node, entry);
 987                                 node->buffer = buffer;
 988                         }
 989                 }
 990                 if (error == 0) {
 991                         node->ondisk = (void *)((char *)buffer->ondisk +
 992                                (node->node_offset & HAMMER_BUFMASK));
 993                         if (isnew == 0 &&
 994                             hammer_crc_test_btree(node->ondisk) == 0) {
 995                                 Debugger("CRC FAILED: B-TREE NODE");
 996                         }
 997                 }
 998         }
 999         --node->loading;
1000         hammer_unlock(&node->lock);
1001         return (error);
1002 }
1003
1004 /*
1005  * Safely reference a node, interlock against flushes via the IO subsystem.
1006  */
1007 hammer_node_t
1008 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
1009                      int *errorp)
1010 {
1011         hammer_node_t node;
1012
1013         node = *cache;
1014         if (node != NULL) {
1015                 hammer_ref(&node->lock);
1016                 if (node->ondisk)
1017                         *errorp = 0;
1018                 else
1019                         *errorp = hammer_load_node(node, 0);
1020                 if (*errorp) {
1021                         hammer_rel_node(node);
1022                         node = NULL;
1023                 }
1024         } else {
1025                 *errorp = ENOENT;
1026         }
1027         return(node);
1028 }
1029
1030 /*
1031  * Release a hammer_node.  On the last release the node dereferences
1032  * its underlying buffer and may or may not be destroyed.
1033  */
1034 void
1035 hammer_rel_node(hammer_node_t node)
1036 {
1037         hammer_buffer_t buffer;
1038
1039         /*
1040          * If this isn't the last ref just decrement the ref count and
1041          * return.
1042          */
1043         if (node->lock.refs > 1) {
1044                 hammer_unref(&node->lock);
1045                 return;
1046         }
1047
1048         /*
1049          * If there is no ondisk info or no buffer the node failed to load,
1050          * remove the last reference and destroy the node.
1051          */
1052         if (node->ondisk == NULL) {
1053                 hammer_unref(&node->lock);
1054                 hammer_flush_node(node);
1055                 /* node is stale now */
1056                 return;
1057         }
1058
1059         /*
1060          * Do final cleanups and then either destroy the node and leave it
1061          * passively cached.  The buffer reference is removed regardless.
1062          */
1063         buffer = node->buffer;
1064         node->ondisk = NULL;
1065
1066         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1067                 hammer_unref(&node->lock);
1068                 hammer_rel_buffer(buffer, 0);
1069                 return;
1070         }
1071
1072         /*
1073          * Destroy the node.
1074          */
1075         hammer_unref(&node->lock);
1076         hammer_flush_node(node);
1077         /* node is stale */
1078         hammer_rel_buffer(buffer, 0);
1079 }
1080
1081 /*
1082  * Free space on-media associated with a B-Tree node.
1083  */
1084 void
1085 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1086 {
1087         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1088         node->flags |= HAMMER_NODE_DELETED;
1089         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1090 }
1091
1092 /*
1093  * Passively cache a referenced hammer_node in *cache.  The caller may
1094  * release the node on return.
1095  */
1096 void
1097 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1098 {
1099         hammer_node_t old;
1100
1101         /*
1102          * If the node is being deleted, don't cache it!
1103          */
1104         if (node->flags & HAMMER_NODE_DELETED)
1105                 return;
1106
1107         /*
1108          * Cache the node.  If we previously cached a different node we
1109          * have to give HAMMER a chance to destroy it.
1110          */
1111 again:
1112         if (node->cache1 != cache) {
1113                 if (node->cache2 != cache) {
1114                         if ((old = *cache) != NULL) {
1115                                 KKASSERT(node->lock.refs != 0);
1116                                 hammer_uncache_node(cache);
1117                                 goto again;
1118                         }
1119                         if (node->cache2)
1120                                 *node->cache2 = NULL;
1121                         node->cache2 = node->cache1;
1122                         node->cache1 = cache;
1123                         *cache = node;
1124                 } else {
1125                         struct hammer_node **tmp;
1126                         tmp = node->cache1;
1127                         node->cache1 = node->cache2;
1128                         node->cache2 = tmp;
1129                 }
1130         }
1131 }
1132
1133 void
1134 hammer_uncache_node(struct hammer_node **cache)
1135 {
1136         hammer_node_t node;
1137
1138         if ((node = *cache) != NULL) {
1139                 *cache = NULL;
1140                 if (node->cache1 == cache) {
1141                         node->cache1 = node->cache2;
1142                         node->cache2 = NULL;
1143                 } else if (node->cache2 == cache) {
1144                         node->cache2 = NULL;
1145                 } else {
1146                         panic("hammer_uncache_node: missing cache linkage");
1147                 }
1148                 if (node->cache1 == NULL && node->cache2 == NULL)
1149                         hammer_flush_node(node);
1150         }
1151 }
1152
1153 /*
1154  * Remove a node's cache references and destroy the node if it has no
1155  * other references or backing store.
1156  */
1157 void
1158 hammer_flush_node(hammer_node_t node)
1159 {
1160         hammer_buffer_t buffer;
1161
1162         if (node->cache1)
1163                 *node->cache1 = NULL;
1164         if (node->cache2)
1165                 *node->cache2 = NULL;
1166         if (node->lock.refs == 0 && node->ondisk == NULL) {
1167                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1168                 if ((buffer = node->buffer) != NULL) {
1169                         node->buffer = NULL;
1170                         TAILQ_REMOVE(&buffer->clist, node, entry);
1171                         /* buffer is unreferenced because ondisk is NULL */
1172                 }
1173                 --hammer_count_nodes;
1174                 kfree(node, M_HAMMER);
1175         }
1176 }
1177
1178 /*
1179  * Flush passively cached B-Tree nodes associated with this buffer.
1180  * This is only called when the buffer is about to be destroyed, so
1181  * none of the nodes should have any references.  The buffer is locked.
1182  *
1183  * We may be interlocked with the buffer.
1184  */
1185 void
1186 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1187 {
1188         hammer_node_t node;
1189
1190         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1191                 KKASSERT(node->ondisk == NULL);
1192
1193                 if (node->lock.refs == 0) {
1194                         hammer_ref(&node->lock);
1195                         node->flags |= HAMMER_NODE_FLUSH;
1196                         hammer_rel_node(node);
1197                 } else {
1198                         KKASSERT(node->loading != 0);
1199                         KKASSERT(node->buffer != NULL);
1200                         buffer = node->buffer;
1201                         node->buffer = NULL;
1202                         TAILQ_REMOVE(&buffer->clist, node, entry);
1203                         /* buffer is unreferenced because ondisk is NULL */
1204                 }
1205         }
1206 }
1207
1208
1209 /************************************************************************
1210  *                              ALLOCATORS                              *
1211  ************************************************************************/
1212
1213 /*
1214  * Allocate a B-Tree node.
1215  */
1216 hammer_node_t
1217 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1218 {
1219         hammer_buffer_t buffer = NULL;
1220         hammer_node_t node = NULL;
1221         hammer_off_t node_offset;
1222
1223         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1224                                             sizeof(struct hammer_node_ondisk),
1225                                             errorp);
1226         if (*errorp == 0) {
1227                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1228                 hammer_modify_node_noundo(trans, node);
1229                 bzero(node->ondisk, sizeof(*node->ondisk));
1230                 hammer_modify_node_done(node);
1231         }
1232         if (buffer)
1233                 hammer_rel_buffer(buffer, 0);
1234         return(node);
1235 }
1236
1237 /*
1238  * Allocate data.  If the address of a data buffer is supplied then
1239  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1240  * will be set to the related buffer.  The caller must release it when
1241  * finally done.  The initial *data_bufferp should be set to NULL by
1242  * the caller.
1243  *
1244  * The caller is responsible for making hammer_modify*() calls on the
1245  * *data_bufferp.
1246  */
1247 void *
1248 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1249                   hammer_off_t *data_offsetp,
1250                   struct hammer_buffer **data_bufferp, int *errorp)
1251 {
1252         void *data;
1253
1254         /*
1255          * Allocate data
1256          */
1257         if (data_len) {
1258                 if (data_len < HAMMER_BUFSIZE) {
1259                         *data_offsetp = hammer_blockmap_alloc(trans,
1260                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1261                                                 data_len, errorp);
1262                 } else {
1263                         *data_offsetp = hammer_blockmap_alloc(trans,
1264                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1265                                                 data_len, errorp);
1266                 }
1267         } else {
1268                 *data_offsetp = 0;
1269         }
1270         if (*errorp == 0 && data_bufferp) {
1271                 if (data_len) {
1272                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1273                                             data_bufferp);
1274                         KKASSERT(*errorp == 0);
1275                 } else {
1276                         data = NULL;
1277                 }
1278         } else {
1279                 data = NULL;
1280         }
1281         KKASSERT(*errorp == 0);
1282         return(data);
1283 }
1284
1285 /*
1286  * Sync dirty buffers to the media and clean-up any loose ends.
1287  */
1288 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1289 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1290
1291 int
1292 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1293 {
1294         struct hammer_sync_info info;
1295
1296         info.error = 0;
1297         info.waitfor = waitfor;
1298         if (waitfor == MNT_WAIT) {
1299                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1300                               hammer_sync_scan1, hammer_sync_scan2, &info);
1301         } else {
1302                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1303                               hammer_sync_scan1, hammer_sync_scan2, &info);
1304         }
1305         return(info.error);
1306 }
1307
1308 int
1309 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1310 {
1311         struct hammer_sync_info info;
1312
1313         info.error = 0;
1314         info.waitfor = waitfor;
1315
1316         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1317                       hammer_sync_scan1, hammer_sync_scan2, &info);
1318         if (waitfor == MNT_WAIT)
1319                 hammer_flusher_sync(hmp);
1320         else
1321                 hammer_flusher_async(hmp);
1322
1323         return(info.error);
1324 }
1325
1326 static int
1327 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1328 {
1329         struct hammer_inode *ip;
1330
1331         ip = VTOI(vp);
1332         if (vp->v_type == VNON || ip == NULL ||
1333             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1334              RB_EMPTY(&vp->v_rbdirty_tree))) {
1335                 return(-1);
1336         }
1337         return(0);
1338 }
1339
1340 static int
1341 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1342 {
1343         struct hammer_sync_info *info = data;
1344         struct hammer_inode *ip;
1345         int error;
1346
1347         ip = VTOI(vp);
1348         if (vp->v_type == VNON || vp->v_type == VBAD ||
1349             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1350              RB_EMPTY(&vp->v_rbdirty_tree))) {
1351                 return(0);
1352         }
1353         error = VOP_FSYNC(vp, info->waitfor);
1354         if (error)
1355                 info->error = error;
1356         return(0);
1357 }
1358