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 1.51 2008/06/08 18:16:26 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 static 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
 313         /*
 314          * There should be no references on the volume, no clusters, and
 315          * no super-clusters.
 316          */
 317         KKASSERT(volume->io.lock.refs == 0);
 318
 319         volume->ondisk = NULL;
 320         if (volume->devvp) {
 321                 if (volume->devvp->v_rdev &&
 322                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 323                 ) {
 324                         volume->devvp->v_rdev->si_mountpoint = NULL;
 325                 }
 326                 if (ronly) {
 327                         vinvalbuf(volume->devvp, 0, 0, 0);
 328                         VOP_CLOSE(volume->devvp, FREAD);
 329                 } else {
 330                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 331                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 332                 }
 333         }
 334
 335         /*
 336          * Destroy the structure
 337          */
 338         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 339         hammer_free_volume(volume);
 340         return(0);
 341 }
 342
 343 static
 344 void
 345 hammer_free_volume(hammer_volume_t volume)
 346 {
 347         if (volume->vol_name) {
 348                 kfree(volume->vol_name, M_HAMMER);
 349                 volume->vol_name = NULL;
 350         }
 351         if (volume->devvp) {
 352                 vrele(volume->devvp);
 353                 volume->devvp = NULL;
 354         }
 355         --hammer_count_volumes;
 356         kfree(volume, M_HAMMER);
 357 }
 358
 359 /*
 360  * Get a HAMMER volume.  The volume must already exist.
 361  */
 362 hammer_volume_t
 363 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 364 {
 365         struct hammer_volume *volume;
 366
 367         /*
 368          * Locate the volume structure
 369          */
 370         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 371         if (volume == NULL) {
 372                 *errorp = ENOENT;
 373                 return(NULL);
 374         }
 375         hammer_ref(&volume->io.lock);
 376
 377         /*
 378          * Deal with on-disk info
 379          */
 380         if (volume->ondisk == NULL || volume->io.loading) {
 381                 *errorp = hammer_load_volume(volume);
 382                 if (*errorp) {
 383                         hammer_rel_volume(volume, 1);
 384                         volume = NULL;
 385                 }
 386         } else {
 387                 *errorp = 0;
 388         }
 389         return(volume);
 390 }
 391
 392 int
 393 hammer_ref_volume(hammer_volume_t volume)
 394 {
 395         int error;
 396
 397         hammer_ref(&volume->io.lock);
 398
 399         /*
 400          * Deal with on-disk info
 401          */
 402         if (volume->ondisk == NULL || volume->io.loading) {
 403                 error = hammer_load_volume(volume);
 404                 if (error)
 405                         hammer_rel_volume(volume, 1);
 406         } else {
 407                 error = 0;
 408         }
 409         return (error);
 410 }
 411
 412 hammer_volume_t
 413 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 414 {
 415         hammer_volume_t volume;
 416
 417         volume = hmp->rootvol;
 418         KKASSERT(volume != NULL);
 419         hammer_ref(&volume->io.lock);
 420
 421         /*
 422          * Deal with on-disk info
 423          */
 424         if (volume->ondisk == NULL || volume->io.loading) {
 425                 *errorp = hammer_load_volume(volume);
 426                 if (*errorp) {
 427                         hammer_rel_volume(volume, 1);
 428                         volume = NULL;
 429                 }
 430         } else {
 431                 *errorp = 0;
 432         }
 433         return (volume);
 434 }
 435
 436 /*
 437  * Load a volume's on-disk information.  The volume must be referenced and
 438  * not locked.  We temporarily acquire an exclusive lock to interlock
 439  * against releases or multiple get's.
 440  */
 441 static int
 442 hammer_load_volume(hammer_volume_t volume)
 443 {
 444         int error;
 445
 446         ++volume->io.loading;
 447         hammer_lock_ex(&volume->io.lock);
 448
 449         if (volume->ondisk == NULL) {
 450                 error = hammer_io_read(volume->devvp, &volume->io,
 451                                        volume->maxraw_off);
 452                 if (error == 0)
 453                         volume->ondisk = (void *)volume->io.bp->b_data;
 454         } else {
 455                 error = 0;
 456         }
 457         --volume->io.loading;
 458         hammer_unlock(&volume->io.lock);
 459         return(error);
 460 }
 461
 462 /*
 463  * Release a volume.  Call hammer_io_release on the last reference.  We have
 464  * to acquire an exclusive lock to interlock against volume->ondisk tests
 465  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 466  * lock to be held.
 467  *
 468  * Volumes are not unloaded from memory during normal operation.
 469  */
 470 void
 471 hammer_rel_volume(hammer_volume_t volume, int flush)
 472 {
 473         crit_enter();
 474         if (volume->io.lock.refs == 1) {
 475                 ++volume->io.loading;
 476                 hammer_lock_ex(&volume->io.lock);
 477                 if (volume->io.lock.refs == 1) {
 478                         volume->ondisk = NULL;
 479                         hammer_io_release(&volume->io, flush);
 480                 }
 481                 --volume->io.loading;
 482                 hammer_unlock(&volume->io.lock);
 483         }
 484         hammer_unref(&volume->io.lock);
 485         crit_exit();
 486 }
 487
 488 /************************************************************************
 489  *                              BUFFERS                                 *
 490  ************************************************************************
 491  *
 492  * Manage buffers.  Currently all blockmap-backed zones are translated
 493  * to zone-2 buffer offsets.
 494  */
 495 hammer_buffer_t
 496 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 497                   int isnew, int *errorp)
 498 {
 499         hammer_buffer_t buffer;
 500         hammer_volume_t volume;
 501         hammer_off_t    zone2_offset;
 502         hammer_io_type_t iotype;
 503         int vol_no;
 504         int zone;
 505
 506 again:
 507         /*
 508          * Shortcut if the buffer is already cached
 509          */
 510         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 511                            buf_offset & ~HAMMER_BUFMASK64);
 512         if (buffer) {
 513                 hammer_ref(&buffer->io.lock);
 514                 if (buffer->ondisk && buffer->io.loading == 0) {
 515                         *errorp = 0;
 516                         return(buffer);
 517                 }
 518
 519                 /*
 520                  * The buffer is no longer loose if it has a ref.  Loose
 521                  * buffers will never be in a modified state.  This should
 522                  * only occur on the 0->1 transition of refs.
 523                  */
 524                 if (buffer->io.mod_list == &hmp->lose_list) {
 525                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 526                                      mod_entry);
 527                         buffer->io.mod_list = NULL;
 528                         KKASSERT(buffer->io.modified == 0);
 529                 }
 530                 goto found;
 531         }
 532
 533         /*
 534          * What is the buffer class?
 535          */
 536         zone = HAMMER_ZONE_DECODE(buf_offset);
 537
 538         switch(zone) {
 539         case HAMMER_ZONE_LARGE_DATA_INDEX:
 540         case HAMMER_ZONE_SMALL_DATA_INDEX:
 541                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 542                 break;
 543         case HAMMER_ZONE_UNDO_INDEX:
 544                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 545                 break;
 546         default:
 547                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 548                 break;
 549         }
 550
 551         /*
 552          * Handle blockmap offset translations
 553          */
 554         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 555                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 556         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 557                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 558         } else {
 559                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 560                 zone2_offset = buf_offset;
 561                 *errorp = 0;
 562         }
 563         if (*errorp)
 564                 return(NULL);
 565
 566         /*
 567          * Calculate the base zone2-offset and acquire the volume
 568          *
 569          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 570          * specifications.
 571          */
 572         zone2_offset &= ~HAMMER_BUFMASK64;
 573         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 574                  HAMMER_ZONE_RAW_BUFFER);
 575         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 576         volume = hammer_get_volume(hmp, vol_no, errorp);
 577         if (volume == NULL)
 578                 return(NULL);
 579
 580         KKASSERT(zone2_offset < volume->maxbuf_off);
 581
 582         /*
 583          * Allocate a new buffer structure.  We will check for races later.
 584          */
 585         ++hammer_count_buffers;
 586         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 587         buffer->zone2_offset = zone2_offset;
 588         buffer->zoneX_offset = buf_offset;
 589         buffer->volume = volume;
 590
 591         hammer_io_init(&buffer->io, hmp, iotype);
 592         buffer->io.offset = volume->ondisk->vol_buf_beg +
 593                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 594         TAILQ_INIT(&buffer->clist);
 595         hammer_ref(&buffer->io.lock);
 596
 597         /*
 598          * Insert the buffer into the RB tree and handle late collisions.
 599          */
 600         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 601                 hammer_unref(&buffer->io.lock);
 602                 --hammer_count_buffers;
 603                 kfree(buffer, M_HAMMER);
 604                 goto again;
 605         }
 606 found:
 607
 608         /*
 609          * Deal with on-disk info and loading races.
 610          */
 611         if (buffer->ondisk == NULL || buffer->io.loading) {
 612                 *errorp = hammer_load_buffer(buffer, isnew);
 613                 if (*errorp) {
 614                         hammer_rel_buffer(buffer, 1);
 615                         buffer = NULL;
 616                 }
 617         } else {
 618                 *errorp = 0;
 619         }
 620         return(buffer);
 621 }
 622
 623 /*
 624  * Destroy all buffers covering the specified zoneX offset range.  This
 625  * is called when the related blockmap layer2 entry is freed.  The buffers
 626  * must not be in use or modified.
 627  */
 628 void
 629 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 630                    hammer_off_t zone2_offset, int bytes)
 631 {
 632         hammer_buffer_t buffer;
 633         hammer_volume_t volume;
 634         int vol_no;
 635         int error;
 636
 637         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 638         volume = hammer_get_volume(hmp, vol_no, &error);
 639         KKASSERT(error == 0);
 640
 641         while (bytes > 0) {
 642                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 643                                    base_offset);
 644                 if (buffer) {
 645                         KKASSERT(buffer->io.lock.refs == 0);
 646                         KKASSERT(buffer->io.modified == 0);
 647                         KKASSERT(buffer->zone2_offset == zone2_offset);
 648                         KKASSERT(buffer->volume == volume);
 649                         hammer_unload_buffer(buffer, NULL);
 650                 }
 651                 hammer_io_inval(volume, zone2_offset);
 652                 base_offset += HAMMER_BUFSIZE;
 653                 zone2_offset += HAMMER_BUFSIZE;
 654                 bytes -= HAMMER_BUFSIZE;
 655         }
 656         hammer_rel_volume(volume, 0);
 657 }
 658
 659 static int
 660 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 661 {
 662         hammer_volume_t volume;
 663         int error;
 664
 665         /*
 666          * Load the buffer's on-disk info
 667          */
 668         volume = buffer->volume;
 669         ++buffer->io.loading;
 670         hammer_lock_ex(&buffer->io.lock);
 671
 672         if (hammer_debug_io & 0x0001) {
 673                 kprintf("load_buffer %016llx %016llx\n",
 674                         buffer->zoneX_offset, buffer->zone2_offset);
 675         }
 676
 677         if (buffer->ondisk == NULL) {
 678                 if (isnew) {
 679                         error = hammer_io_new(volume->devvp, &buffer->io);
 680                 } else {
 681                         error = hammer_io_read(volume->devvp, &buffer->io,
 682                                                volume->maxraw_off);
 683                 }
 684                 if (error == 0)
 685                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 686         } else if (isnew) {
 687                 error = hammer_io_new(volume->devvp, &buffer->io);
 688         } else {
 689                 error = 0;
 690         }
 691         --buffer->io.loading;
 692         hammer_unlock(&buffer->io.lock);
 693         return (error);
 694 }
 695
 696 /*
 697  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 698  */
 699 int
 700 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 701 {
 702         hammer_ref(&buffer->io.lock);
 703         hammer_flush_buffer_nodes(buffer);
 704         KKASSERT(buffer->io.lock.refs == 1);
 705         hammer_rel_buffer(buffer, 2);
 706         return(0);
 707 }
 708
 709 /*
 710  * Reference a buffer that is either already referenced or via a specially
 711  * handled pointer (aka cursor->buffer).
 712  */
 713 int
 714 hammer_ref_buffer(hammer_buffer_t buffer)
 715 {
 716         int error;
 717
 718         hammer_ref(&buffer->io.lock);
 719
 720         /*
 721          * No longer loose
 722          */
 723         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 724                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 725                 buffer->io.mod_list = NULL;
 726         }
 727
 728         if (buffer->ondisk == NULL || buffer->io.loading) {
 729                 error = hammer_load_buffer(buffer, 0);
 730                 if (error) {
 731                         hammer_rel_buffer(buffer, 1);
 732                         /*
 733                          * NOTE: buffer pointer can become stale after
 734                          * the above release.
 735                          */
 736                 }
 737         } else {
 738                 error = 0;
 739         }
 740         return(error);
 741 }
 742
 743 /*
 744  * Release a buffer.  We have to deal with several places where
 745  * another thread can ref the buffer.
 746  *
 747  * Only destroy the structure itself if the related buffer cache buffer
 748  * was disassociated from it.  This ties the management of the structure
 749  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 750  * embedded io is referenced or not.
 751  */
 752 void
 753 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 754 {
 755         hammer_volume_t volume;
 756         int freeme = 0;
 757
 758         crit_enter();
 759         if (buffer->io.lock.refs == 1) {
 760                 ++buffer->io.loading;   /* force interlock check */
 761                 hammer_lock_ex(&buffer->io.lock);
 762                 if (buffer->io.lock.refs == 1) {
 763                         hammer_io_release(&buffer->io, flush);
 764                         hammer_flush_buffer_nodes(buffer);
 765                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 766
 767                         if (buffer->io.bp == NULL &&
 768                             buffer->io.lock.refs == 1) {
 769                                 /*
 770                                  * Final cleanup
 771                                  */
 772                                 RB_REMOVE(hammer_buf_rb_tree,
 773                                           &buffer->io.hmp->rb_bufs_root,
 774                                           buffer);
 775                                 volume = buffer->volume;
 776                                 buffer->volume = NULL; /* sanity */
 777                                 hammer_rel_volume(volume, 0);
 778                                 freeme = 1;
 779                         }
 780                 }
 781                 --buffer->io.loading;
 782                 hammer_unlock(&buffer->io.lock);
 783         }
 784         hammer_unref(&buffer->io.lock);
 785         crit_exit();
 786         if (freeme) {
 787                 KKASSERT(buffer->io.mod_list == NULL);
 788                 --hammer_count_buffers;
 789                 kfree(buffer, M_HAMMER);
 790         }
 791 }
 792
 793 /*
 794  * Access the filesystem buffer containing the specified hammer offset.
 795  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 796  * relative buffer offset.  It must also have bit 55 set to be valid.
 797  * (see hammer_off_t in hammer_disk.h).
 798  *
 799  * Any prior buffer in *bufferp will be released and replaced by the
 800  * requested buffer.
 801  */
 802 void *
 803 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 804              struct hammer_buffer **bufferp)
 805 {
 806         hammer_buffer_t buffer;
 807         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 808
 809         buf_offset &= ~HAMMER_BUFMASK64;
 810         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 811
 812         buffer = *bufferp;
 813         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 814                                buffer->zoneX_offset != buf_offset)) {
 815                 if (buffer)
 816                         hammer_rel_buffer(buffer, 0);
 817                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 818                 *bufferp = buffer;
 819         } else {
 820                 *errorp = 0;
 821         }
 822
 823         /*
 824          * Return a pointer to the buffer data.
 825          */
 826         if (buffer == NULL)
 827                 return(NULL);
 828         else
 829                 return((char *)buffer->ondisk + xoff);
 830 }
 831
 832 /*
 833  * Access the filesystem buffer containing the specified hammer offset.
 834  * No disk read operation occurs.  The result buffer may contain garbage.
 835  *
 836  * Any prior buffer in *bufferp will be released and replaced by the
 837  * requested buffer.
 838  *
 839  * This function marks the buffer dirty but does not increment its
 840  * modify_refs count.
 841  */
 842 void *
 843 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 844              struct hammer_buffer **bufferp)
 845 {
 846         hammer_buffer_t buffer;
 847         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 848
 849         buf_offset &= ~HAMMER_BUFMASK64;
 850
 851         buffer = *bufferp;
 852         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 853                                buffer->zoneX_offset != buf_offset)) {
 854                 if (buffer)
 855                         hammer_rel_buffer(buffer, 0);
 856                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 857                 *bufferp = buffer;
 858         } else {
 859                 *errorp = 0;
 860         }
 861
 862         /*
 863          * Return a pointer to the buffer data.
 864          */
 865         if (buffer == NULL)
 866                 return(NULL);
 867         else
 868                 return((char *)buffer->ondisk + xoff);
 869 }
 870
 871 /************************************************************************
 872  *                              NODES                                   *
 873  ************************************************************************
 874  *
 875  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 876  * method used by the HAMMER filesystem.
 877  *
 878  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 879  * associated with its buffer, and will only referenced the buffer while
 880  * the node itself is referenced.
 881  *
 882  * A hammer_node can also be passively associated with other HAMMER
 883  * structures, such as inodes, while retaining 0 references.  These
 884  * associations can be cleared backwards using a pointer-to-pointer in
 885  * the hammer_node.
 886  *
 887  * This allows the HAMMER implementation to cache hammer_nodes long-term
 888  * and short-cut a great deal of the infrastructure's complexity.  In
 889  * most cases a cached node can be reacquired without having to dip into
 890  * either the buffer or cluster management code.
 891  *
 892  * The caller must pass a referenced cluster on call and will retain
 893  * ownership of the reference on return.  The node will acquire its own
 894  * additional references, if necessary.
 895  */
 896 hammer_node_t
 897 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 898                 int isnew, int *errorp)
 899 {
 900         hammer_node_t node;
 901
 902         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 903
 904         /*
 905          * Locate the structure, allocating one if necessary.
 906          */
 907 again:
 908         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 909         if (node == NULL) {
 910                 ++hammer_count_nodes;
 911                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 912                 node->node_offset = node_offset;
 913                 node->hmp = hmp;
 914                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 915                         --hammer_count_nodes;
 916                         kfree(node, M_HAMMER);
 917                         goto again;
 918                 }
 919         }
 920         hammer_ref(&node->lock);
 921         if (node->ondisk)
 922                 *errorp = 0;
 923         else
 924                 *errorp = hammer_load_node(node, isnew);
 925         if (*errorp) {
 926                 hammer_rel_node(node);
 927                 node = NULL;
 928         }
 929         return(node);
 930 }
 931
 932 /*
 933  * Reference an already-referenced node.
 934  */
 935 void
 936 hammer_ref_node(hammer_node_t node)
 937 {
 938         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 939         hammer_ref(&node->lock);
 940 }
 941
 942 /*
 943  * Load a node's on-disk data reference.
 944  */
 945 static int
 946 hammer_load_node(hammer_node_t node, int isnew)
 947 {
 948         hammer_buffer_t buffer;
 949         hammer_off_t buf_offset;
 950         int error;
 951
 952         error = 0;
 953         ++node->loading;
 954         hammer_lock_ex(&node->lock);
 955         if (node->ondisk == NULL) {
 956                 /*
 957                  * This is a little confusing but the jist is that
 958                  * node->buffer determines whether the node is on
 959                  * the buffer's clist and node->ondisk determines
 960                  * whether the buffer is referenced.
 961                  *
 962                  * We could be racing a buffer release, in which case
 963                  * node->buffer may become NULL while we are blocked
 964                  * referencing the buffer.
 965                  */
 966                 if ((buffer = node->buffer) != NULL) {
 967                         error = hammer_ref_buffer(buffer);
 968                         if (error == 0 && node->buffer == NULL) {
 969                                 TAILQ_INSERT_TAIL(&buffer->clist,
 970                                                   node, entry);
 971                                 node->buffer = buffer;
 972                         }
 973                 } else {
 974                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
 975                         buffer = hammer_get_buffer(node->hmp, buf_offset,
 976                                                    0, &error);
 977                         if (buffer) {
 978                                 KKASSERT(error == 0);
 979                                 TAILQ_INSERT_TAIL(&buffer->clist,
 980                                                   node, entry);
 981                                 node->buffer = buffer;
 982                         }
 983                 }
 984                 if (error == 0) {
 985                         node->ondisk = (void *)((char *)buffer->ondisk +
 986                                (node->node_offset & HAMMER_BUFMASK));
 987                         if (isnew == 0 &&
 988                             hammer_crc_test_btree(node->ondisk) == 0) {
 989                                 Debugger("CRC FAILED: B-TREE NODE");
 990                         }
 991                 }
 992         }
 993         --node->loading;
 994         hammer_unlock(&node->lock);
 995         return (error);
 996 }
 997
 998 /*
 999  * Safely reference a node, interlock against flushes via the IO subsystem.
1000  */
1001 hammer_node_t
1002 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
1003                      int *errorp)
1004 {
1005         hammer_node_t node;
1006
1007         node = *cache;
1008         if (node != NULL) {
1009                 hammer_ref(&node->lock);
1010                 if (node->ondisk)
1011                         *errorp = 0;
1012                 else
1013                         *errorp = hammer_load_node(node, 0);
1014                 if (*errorp) {
1015                         hammer_rel_node(node);
1016                         node = NULL;
1017                 }
1018         } else {
1019                 *errorp = ENOENT;
1020         }
1021         return(node);
1022 }
1023
1024 /*
1025  * Release a hammer_node.  On the last release the node dereferences
1026  * its underlying buffer and may or may not be destroyed.
1027  */
1028 void
1029 hammer_rel_node(hammer_node_t node)
1030 {
1031         hammer_buffer_t buffer;
1032
1033         /*
1034          * If this isn't the last ref just decrement the ref count and
1035          * return.
1036          */
1037         if (node->lock.refs > 1) {
1038                 hammer_unref(&node->lock);
1039                 return;
1040         }
1041
1042         /*
1043          * If there is no ondisk info or no buffer the node failed to load,
1044          * remove the last reference and destroy the node.
1045          */
1046         if (node->ondisk == NULL) {
1047                 hammer_unref(&node->lock);
1048                 hammer_flush_node(node);
1049                 /* node is stale now */
1050                 return;
1051         }
1052
1053         /*
1054          * Do final cleanups and then either destroy the node and leave it
1055          * passively cached.  The buffer reference is removed regardless.
1056          */
1057         buffer = node->buffer;
1058         node->ondisk = NULL;
1059
1060         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1061                 hammer_unref(&node->lock);
1062                 hammer_rel_buffer(buffer, 0);
1063                 return;
1064         }
1065
1066         /*
1067          * Destroy the node.
1068          */
1069         hammer_unref(&node->lock);
1070         hammer_flush_node(node);
1071         /* node is stale */
1072         hammer_rel_buffer(buffer, 0);
1073 }
1074
1075 /*
1076  * Free space on-media associated with a B-Tree node.
1077  */
1078 void
1079 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1080 {
1081         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1082         node->flags |= HAMMER_NODE_DELETED;
1083         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1084 }
1085
1086 /*
1087  * Passively cache a referenced hammer_node in *cache.  The caller may
1088  * release the node on return.
1089  */
1090 void
1091 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1092 {
1093         hammer_node_t old;
1094
1095         /*
1096          * If the node is being deleted, don't cache it!
1097          */
1098         if (node->flags & HAMMER_NODE_DELETED)
1099                 return;
1100
1101         /*
1102          * Cache the node.  If we previously cached a different node we
1103          * have to give HAMMER a chance to destroy it.
1104          */
1105 again:
1106         if (node->cache1 != cache) {
1107                 if (node->cache2 != cache) {
1108                         if ((old = *cache) != NULL) {
1109                                 KKASSERT(node->lock.refs != 0);
1110                                 hammer_uncache_node(cache);
1111                                 goto again;
1112                         }
1113                         if (node->cache2)
1114                                 *node->cache2 = NULL;
1115                         node->cache2 = node->cache1;
1116                         node->cache1 = cache;
1117                         *cache = node;
1118                 } else {
1119                         struct hammer_node **tmp;
1120                         tmp = node->cache1;
1121                         node->cache1 = node->cache2;
1122                         node->cache2 = tmp;
1123                 }
1124         }
1125 }
1126
1127 void
1128 hammer_uncache_node(struct hammer_node **cache)
1129 {
1130         hammer_node_t node;
1131
1132         if ((node = *cache) != NULL) {
1133                 *cache = NULL;
1134                 if (node->cache1 == cache) {
1135                         node->cache1 = node->cache2;
1136                         node->cache2 = NULL;
1137                 } else if (node->cache2 == cache) {
1138                         node->cache2 = NULL;
1139                 } else {
1140                         panic("hammer_uncache_node: missing cache linkage");
1141                 }
1142                 if (node->cache1 == NULL && node->cache2 == NULL)
1143                         hammer_flush_node(node);
1144         }
1145 }
1146
1147 /*
1148  * Remove a node's cache references and destroy the node if it has no
1149  * other references or backing store.
1150  */
1151 void
1152 hammer_flush_node(hammer_node_t node)
1153 {
1154         hammer_buffer_t buffer;
1155
1156         if (node->cache1)
1157                 *node->cache1 = NULL;
1158         if (node->cache2)
1159                 *node->cache2 = NULL;
1160         if (node->lock.refs == 0 && node->ondisk == NULL) {
1161                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1162                 if ((buffer = node->buffer) != NULL) {
1163                         node->buffer = NULL;
1164                         TAILQ_REMOVE(&buffer->clist, node, entry);
1165                         /* buffer is unreferenced because ondisk is NULL */
1166                 }
1167                 --hammer_count_nodes;
1168                 kfree(node, M_HAMMER);
1169         }
1170 }
1171
1172 /*
1173  * Flush passively cached B-Tree nodes associated with this buffer.
1174  * This is only called when the buffer is about to be destroyed, so
1175  * none of the nodes should have any references.  The buffer is locked.
1176  *
1177  * We may be interlocked with the buffer.
1178  */
1179 void
1180 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1181 {
1182         hammer_node_t node;
1183
1184         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1185                 KKASSERT(node->ondisk == NULL);
1186
1187                 if (node->lock.refs == 0) {
1188                         hammer_ref(&node->lock);
1189                         node->flags |= HAMMER_NODE_FLUSH;
1190                         hammer_rel_node(node);
1191                 } else {
1192                         KKASSERT(node->loading != 0);
1193                         KKASSERT(node->buffer != NULL);
1194                         buffer = node->buffer;
1195                         node->buffer = NULL;
1196                         TAILQ_REMOVE(&buffer->clist, node, entry);
1197                         /* buffer is unreferenced because ondisk is NULL */
1198                 }
1199         }
1200 }
1201
1202
1203 /************************************************************************
1204  *                              ALLOCATORS                              *
1205  ************************************************************************/
1206
1207 /*
1208  * Allocate a B-Tree node.
1209  */
1210 hammer_node_t
1211 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1212 {
1213         hammer_buffer_t buffer = NULL;
1214         hammer_node_t node = NULL;
1215         hammer_off_t node_offset;
1216
1217         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1218                                             sizeof(struct hammer_node_ondisk),
1219                                             errorp);
1220         if (*errorp == 0) {
1221                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1222                 hammer_modify_node_noundo(trans, node);
1223                 bzero(node->ondisk, sizeof(*node->ondisk));
1224                 hammer_modify_node_done(node);
1225         }
1226         if (buffer)
1227                 hammer_rel_buffer(buffer, 0);
1228         return(node);
1229 }
1230
1231 /*
1232  * Allocate data.  If the address of a data buffer is supplied then
1233  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1234  * will be set to the related buffer.  The caller must release it when
1235  * finally done.  The initial *data_bufferp should be set to NULL by
1236  * the caller.
1237  *
1238  * The caller is responsible for making hammer_modify*() calls on the
1239  * *data_bufferp.
1240  */
1241 void *
1242 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1243                   hammer_off_t *data_offsetp,
1244                   struct hammer_buffer **data_bufferp, int *errorp)
1245 {
1246         void *data;
1247
1248         /*
1249          * Allocate data
1250          */
1251         if (data_len) {
1252                 if (data_len < HAMMER_BUFSIZE) {
1253                         *data_offsetp = hammer_blockmap_alloc(trans,
1254                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1255                                                 data_len, errorp);
1256                 } else {
1257                         *data_offsetp = hammer_blockmap_alloc(trans,
1258                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1259                                                 data_len, errorp);
1260                 }
1261         } else {
1262                 *data_offsetp = 0;
1263         }
1264         if (*errorp == 0 && data_bufferp) {
1265                 if (data_len) {
1266                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1267                                             data_bufferp);
1268                         KKASSERT(*errorp == 0);
1269                 } else {
1270                         data = NULL;
1271                 }
1272         } else {
1273                 data = NULL;
1274         }
1275         KKASSERT(*errorp == 0);
1276         return(data);
1277 }
1278
1279 /*
1280  * Sync dirty buffers to the media and clean-up any loose ends.
1281  */
1282 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1283 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1284
1285 int
1286 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1287 {
1288         struct hammer_sync_info info;
1289
1290         info.error = 0;
1291         info.waitfor = waitfor;
1292         if (waitfor == MNT_WAIT) {
1293                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1294                               hammer_sync_scan1, hammer_sync_scan2, &info);
1295         } else {
1296                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1297                               hammer_sync_scan1, hammer_sync_scan2, &info);
1298         }
1299         return(info.error);
1300 }
1301
1302 int
1303 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1304 {
1305         struct hammer_sync_info info;
1306
1307         info.error = 0;
1308         info.waitfor = waitfor;
1309
1310         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1311                       hammer_sync_scan1, hammer_sync_scan2, &info);
1312         if (waitfor == MNT_WAIT)
1313                 hammer_flusher_sync(hmp);
1314         else
1315                 hammer_flusher_async(hmp);
1316
1317         return(info.error);
1318 }
1319
1320 static int
1321 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1322 {
1323         struct hammer_inode *ip;
1324
1325         ip = VTOI(vp);
1326         if (vp->v_type == VNON || ip == NULL ||
1327             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1328              RB_EMPTY(&vp->v_rbdirty_tree))) {
1329                 return(-1);
1330         }
1331         return(0);
1332 }
1333
1334 static int
1335 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1336 {
1337         struct hammer_sync_info *info = data;
1338         struct hammer_inode *ip;
1339         int error;
1340
1341         ip = VTOI(vp);
1342         if (vp->v_type == VNON || vp->v_type == VBAD ||
1343             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1344              RB_EMPTY(&vp->v_rbdirty_tree))) {
1345                 return(0);
1346         }
1347         error = VOP_FSYNC(vp, info->waitfor);
1348         if (error)
1349                 info->error = error;
1350         return(0);
1351 }
1352