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 1.56 2008/06/13 00:25:33 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                 if (buffer->io.lock.refs == 0)
 515                         ++hammer_count_refedbufs;
 516                 hammer_ref(&buffer->io.lock);
 517
 518                 /*
 519                  * Onced refed the ondisk field will not be cleared by
 520                  * any other action.
 521                  */
 522                 if (buffer->ondisk && buffer->io.loading == 0) {
 523                         *errorp = 0;
 524                         return(buffer);
 525                 }
 526
 527                 /*
 528                  * The buffer is no longer loose if it has a ref, and
 529                  * cannot become loose once it gains a ref.  Loose
 530                  * buffers will never be in a modified state.  This should
 531                  * only occur on the 0->1 transition of refs.
 532                  */
 533                 if (buffer->io.mod_list == &hmp->lose_list) {
 534                         crit_enter();   /* biodone race against list */
 535                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 536                                      mod_entry);
 537                         crit_exit();
 538                         buffer->io.mod_list = NULL;
 539                         KKASSERT(buffer->io.modified == 0);
 540                 }
 541                 goto found;
 542         }
 543
 544         /*
 545          * What is the buffer class?
 546          */
 547         zone = HAMMER_ZONE_DECODE(buf_offset);
 548
 549         switch(zone) {
 550         case HAMMER_ZONE_LARGE_DATA_INDEX:
 551         case HAMMER_ZONE_SMALL_DATA_INDEX:
 552                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 553                 break;
 554         case HAMMER_ZONE_UNDO_INDEX:
 555                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 556                 break;
 557         default:
 558                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 559                 break;
 560         }
 561
 562         /*
 563          * Handle blockmap offset translations
 564          */
 565         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 566                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 567         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 568                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 569         } else {
 570                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 571                 zone2_offset = buf_offset;
 572                 *errorp = 0;
 573         }
 574         if (*errorp)
 575                 return(NULL);
 576
 577         /*
 578          * Calculate the base zone2-offset and acquire the volume
 579          *
 580          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 581          * specifications.
 582          */
 583         zone2_offset &= ~HAMMER_BUFMASK64;
 584         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 585                  HAMMER_ZONE_RAW_BUFFER);
 586         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 587         volume = hammer_get_volume(hmp, vol_no, errorp);
 588         if (volume == NULL)
 589                 return(NULL);
 590
 591         KKASSERT(zone2_offset < volume->maxbuf_off);
 592
 593         /*
 594          * Allocate a new buffer structure.  We will check for races later.
 595          */
 596         ++hammer_count_buffers;
 597         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 598         buffer->zone2_offset = zone2_offset;
 599         buffer->zoneX_offset = buf_offset;
 600         buffer->volume = volume;
 601
 602         hammer_io_init(&buffer->io, hmp, iotype);
 603         buffer->io.offset = volume->ondisk->vol_buf_beg +
 604                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 605         TAILQ_INIT(&buffer->clist);
 606         hammer_ref(&buffer->io.lock);
 607
 608         /*
 609          * Insert the buffer into the RB tree and handle late collisions.
 610          */
 611         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 612                 hammer_unref(&buffer->io.lock);
 613                 --hammer_count_buffers;
 614                 kfree(buffer, M_HAMMER);
 615                 goto again;
 616         }
 617         ++hammer_count_refedbufs;
 618 found:
 619
 620         /*
 621          * Deal with on-disk info and loading races.
 622          */
 623         if (buffer->ondisk == NULL || buffer->io.loading) {
 624                 *errorp = hammer_load_buffer(buffer, isnew);
 625                 if (*errorp) {
 626                         hammer_rel_buffer(buffer, 1);
 627                         buffer = NULL;
 628                 }
 629         } else {
 630                 *errorp = 0;
 631         }
 632         return(buffer);
 633 }
 634
 635 /*
 636  * Destroy all buffers covering the specified zoneX offset range.  This
 637  * is called when the related blockmap layer2 entry is freed or when
 638  * a direct write bypasses our buffer/buffer-cache subsystem.
 639  *
 640  * The buffers may be referenced by the caller itself.  Setting reclaim
 641  * will cause the buffer to be destroyed when it's ref count reaches zero.
 642  */
 643 void
 644 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 645                    hammer_off_t zone2_offset, int bytes)
 646 {
 647         hammer_buffer_t buffer;
 648         hammer_volume_t volume;
 649         int vol_no;
 650         int error;
 651
 652         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 653         volume = hammer_get_volume(hmp, vol_no, &error);
 654         KKASSERT(error == 0);
 655
 656         while (bytes > 0) {
 657                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 658                                    base_offset);
 659                 if (buffer) {
 660                         KKASSERT(buffer->zone2_offset == zone2_offset);
 661                         hammer_io_clear_modify(&buffer->io);
 662                         buffer->io.reclaim = 1;
 663                         KKASSERT(buffer->volume == volume);
 664                         if (buffer->io.lock.refs == 0)
 665                                 hammer_unload_buffer(buffer, NULL);
 666                 } else {
 667                         hammer_io_inval(volume, zone2_offset);
 668                 }
 669                 base_offset += HAMMER_BUFSIZE;
 670                 zone2_offset += HAMMER_BUFSIZE;
 671                 bytes -= HAMMER_BUFSIZE;
 672         }
 673         hammer_rel_volume(volume, 0);
 674 }
 675
 676 static int
 677 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 678 {
 679         hammer_volume_t volume;
 680         int error;
 681
 682         /*
 683          * Load the buffer's on-disk info
 684          */
 685         volume = buffer->volume;
 686         ++buffer->io.loading;
 687         hammer_lock_ex(&buffer->io.lock);
 688
 689         if (hammer_debug_io & 0x0001) {
 690                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 691                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 692                         buffer->ondisk);
 693         }
 694
 695         if (buffer->ondisk == NULL) {
 696                 if (isnew) {
 697                         error = hammer_io_new(volume->devvp, &buffer->io);
 698                 } else {
 699                         error = hammer_io_read(volume->devvp, &buffer->io,
 700                                                volume->maxraw_off);
 701                 }
 702                 if (error == 0)
 703                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 704         } else if (isnew) {
 705                 error = hammer_io_new(volume->devvp, &buffer->io);
 706         } else {
 707                 error = 0;
 708         }
 709         --buffer->io.loading;
 710         hammer_unlock(&buffer->io.lock);
 711         return (error);
 712 }
 713
 714 /*
 715  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 716  */
 717 int
 718 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 719 {
 720         ++hammer_count_refedbufs;
 721         hammer_ref(&buffer->io.lock);
 722         hammer_flush_buffer_nodes(buffer);
 723         KKASSERT(buffer->io.lock.refs == 1);
 724         hammer_rel_buffer(buffer, 2);
 725         return(0);
 726 }
 727
 728 /*
 729  * Reference a buffer that is either already referenced or via a specially
 730  * handled pointer (aka cursor->buffer).
 731  */
 732 int
 733 hammer_ref_buffer(hammer_buffer_t buffer)
 734 {
 735         int error;
 736
 737         if (buffer->io.lock.refs == 0)
 738                 ++hammer_count_refedbufs;
 739         hammer_ref(&buffer->io.lock);
 740
 741         /*
 742          * No longer loose
 743          */
 744         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 745                 crit_enter();   /* biodone race against list */
 746                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 747                 buffer->io.mod_list = NULL;
 748                 crit_exit();
 749         }
 750
 751         if (buffer->ondisk == NULL || buffer->io.loading) {
 752                 error = hammer_load_buffer(buffer, 0);
 753                 if (error) {
 754                         hammer_rel_buffer(buffer, 1);
 755                         /*
 756                          * NOTE: buffer pointer can become stale after
 757                          * the above release.
 758                          */
 759                 }
 760         } else {
 761                 error = 0;
 762         }
 763         return(error);
 764 }
 765
 766 /*
 767  * Release a buffer.  We have to deal with several places where
 768  * another thread can ref the buffer.
 769  *
 770  * Only destroy the structure itself if the related buffer cache buffer
 771  * was disassociated from it.  This ties the management of the structure
 772  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 773  * embedded io is referenced or not.
 774  */
 775 void
 776 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 777 {
 778         hammer_volume_t volume;
 779         int freeme = 0;
 780
 781         crit_enter();
 782         if (buffer->io.lock.refs == 1) {
 783                 ++buffer->io.loading;   /* force interlock check */
 784                 hammer_lock_ex(&buffer->io.lock);
 785                 if (buffer->io.lock.refs == 1) {
 786                         hammer_io_release(&buffer->io, flush);
 787                         hammer_flush_buffer_nodes(buffer);
 788                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 789                         --hammer_count_refedbufs;
 790
 791                         if (buffer->io.bp == NULL &&
 792                             buffer->io.lock.refs == 1) {
 793                                 /*
 794                                  * Final cleanup
 795                                  */
 796                                 RB_REMOVE(hammer_buf_rb_tree,
 797                                           &buffer->io.hmp->rb_bufs_root,
 798                                           buffer);
 799                                 volume = buffer->volume;
 800                                 buffer->volume = NULL; /* sanity */
 801                                 hammer_rel_volume(volume, 0);
 802                                 hammer_io_clear_modlist(&buffer->io);
 803                                 freeme = 1;
 804                         }
 805                 }
 806                 --buffer->io.loading;
 807                 hammer_unlock(&buffer->io.lock);
 808         }
 809         hammer_unref(&buffer->io.lock);
 810         crit_exit();
 811         if (freeme) {
 812                 --hammer_count_buffers;
 813                 kfree(buffer, M_HAMMER);
 814         }
 815 }
 816
 817 /*
 818  * Access the filesystem buffer containing the specified hammer offset.
 819  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 820  * relative buffer offset.  It must also have bit 55 set to be valid.
 821  * (see hammer_off_t in hammer_disk.h).
 822  *
 823  * Any prior buffer in *bufferp will be released and replaced by the
 824  * requested buffer.
 825  */
 826 void *
 827 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 828              struct hammer_buffer **bufferp)
 829 {
 830         hammer_buffer_t buffer;
 831         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 832
 833         buf_offset &= ~HAMMER_BUFMASK64;
 834         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 835
 836         buffer = *bufferp;
 837         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 838                                buffer->zoneX_offset != buf_offset)) {
 839                 if (buffer)
 840                         hammer_rel_buffer(buffer, 0);
 841                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 842                 *bufferp = buffer;
 843         } else {
 844                 *errorp = 0;
 845         }
 846
 847         /*
 848          * Return a pointer to the buffer data.
 849          */
 850         if (buffer == NULL)
 851                 return(NULL);
 852         else
 853                 return((char *)buffer->ondisk + xoff);
 854 }
 855
 856 /*
 857  * Access the filesystem buffer containing the specified hammer offset.
 858  * No disk read operation occurs.  The result buffer may contain garbage.
 859  *
 860  * Any prior buffer in *bufferp will be released and replaced by the
 861  * requested buffer.
 862  *
 863  * This function marks the buffer dirty but does not increment its
 864  * modify_refs count.
 865  */
 866 void *
 867 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 868              struct hammer_buffer **bufferp)
 869 {
 870         hammer_buffer_t buffer;
 871         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 872
 873         buf_offset &= ~HAMMER_BUFMASK64;
 874
 875         buffer = *bufferp;
 876         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 877                                buffer->zoneX_offset != buf_offset)) {
 878                 if (buffer)
 879                         hammer_rel_buffer(buffer, 0);
 880                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 881                 *bufferp = buffer;
 882         } else {
 883                 *errorp = 0;
 884         }
 885
 886         /*
 887          * Return a pointer to the buffer data.
 888          */
 889         if (buffer == NULL)
 890                 return(NULL);
 891         else
 892                 return((char *)buffer->ondisk + xoff);
 893 }
 894
 895 /************************************************************************
 896  *                              NODES                                   *
 897  ************************************************************************
 898  *
 899  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 900  * method used by the HAMMER filesystem.
 901  *
 902  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 903  * associated with its buffer, and will only referenced the buffer while
 904  * the node itself is referenced.
 905  *
 906  * A hammer_node can also be passively associated with other HAMMER
 907  * structures, such as inodes, while retaining 0 references.  These
 908  * associations can be cleared backwards using a pointer-to-pointer in
 909  * the hammer_node.
 910  *
 911  * This allows the HAMMER implementation to cache hammer_nodes long-term
 912  * and short-cut a great deal of the infrastructure's complexity.  In
 913  * most cases a cached node can be reacquired without having to dip into
 914  * either the buffer or cluster management code.
 915  *
 916  * The caller must pass a referenced cluster on call and will retain
 917  * ownership of the reference on return.  The node will acquire its own
 918  * additional references, if necessary.
 919  */
 920 hammer_node_t
 921 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 922                 int isnew, int *errorp)
 923 {
 924         hammer_node_t node;
 925
 926         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 927
 928         /*
 929          * Locate the structure, allocating one if necessary.
 930          */
 931 again:
 932         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 933         if (node == NULL) {
 934                 ++hammer_count_nodes;
 935                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 936                 node->node_offset = node_offset;
 937                 node->hmp = hmp;
 938                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 939                         --hammer_count_nodes;
 940                         kfree(node, M_HAMMER);
 941                         goto again;
 942                 }
 943         }
 944         hammer_ref(&node->lock);
 945         if (node->ondisk)
 946                 *errorp = 0;
 947         else
 948                 *errorp = hammer_load_node(node, isnew);
 949         if (*errorp) {
 950                 hammer_rel_node(node);
 951                 node = NULL;
 952         }
 953         return(node);
 954 }
 955
 956 /*
 957  * Reference an already-referenced node.
 958  */
 959 void
 960 hammer_ref_node(hammer_node_t node)
 961 {
 962         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 963         hammer_ref(&node->lock);
 964 }
 965
 966 /*
 967  * Load a node's on-disk data reference.
 968  */
 969 static int
 970 hammer_load_node(hammer_node_t node, int isnew)
 971 {
 972         hammer_buffer_t buffer;
 973         hammer_off_t buf_offset;
 974         int error;
 975
 976         error = 0;
 977         ++node->loading;
 978         hammer_lock_ex(&node->lock);
 979         if (node->ondisk == NULL) {
 980                 /*
 981                  * This is a little confusing but the jist is that
 982                  * node->buffer determines whether the node is on
 983                  * the buffer's clist and node->ondisk determines
 984                  * whether the buffer is referenced.
 985                  *
 986                  * We could be racing a buffer release, in which case
 987                  * node->buffer may become NULL while we are blocked
 988                  * referencing the buffer.
 989                  */
 990                 if ((buffer = node->buffer) != NULL) {
 991                         error = hammer_ref_buffer(buffer);
 992                         if (error == 0 && node->buffer == NULL) {
 993                                 TAILQ_INSERT_TAIL(&buffer->clist,
 994                                                   node, entry);
 995                                 node->buffer = buffer;
 996                         }
 997                 } else {
 998                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
 999                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1000                                                    0, &error);
1001                         if (buffer) {
1002                                 KKASSERT(error == 0);
1003                                 TAILQ_INSERT_TAIL(&buffer->clist,
1004                                                   node, entry);
1005                                 node->buffer = buffer;
1006                         }
1007                 }
1008                 if (error == 0) {
1009                         node->ondisk = (void *)((char *)buffer->ondisk +
1010                                (node->node_offset & HAMMER_BUFMASK));
1011                         if (isnew == 0 &&
1012                             hammer_crc_test_btree(node->ondisk) == 0) {
1013                                 Debugger("CRC FAILED: B-TREE NODE");
1014                         }
1015                 }
1016         }
1017         --node->loading;
1018         hammer_unlock(&node->lock);
1019         return (error);
1020 }
1021
1022 /*
1023  * Safely reference a node, interlock against flushes via the IO subsystem.
1024  */
1025 hammer_node_t
1026 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
1027                      int *errorp)
1028 {
1029         hammer_node_t node;
1030
1031         node = *cache;
1032         if (node != NULL) {
1033                 hammer_ref(&node->lock);
1034                 if (node->ondisk)
1035                         *errorp = 0;
1036                 else
1037                         *errorp = hammer_load_node(node, 0);
1038                 if (*errorp) {
1039                         hammer_rel_node(node);
1040                         node = NULL;
1041                 }
1042         } else {
1043                 *errorp = ENOENT;
1044         }
1045         return(node);
1046 }
1047
1048 /*
1049  * Release a hammer_node.  On the last release the node dereferences
1050  * its underlying buffer and may or may not be destroyed.
1051  */
1052 void
1053 hammer_rel_node(hammer_node_t node)
1054 {
1055         hammer_buffer_t buffer;
1056
1057         /*
1058          * If this isn't the last ref just decrement the ref count and
1059          * return.
1060          */
1061         if (node->lock.refs > 1) {
1062                 hammer_unref(&node->lock);
1063                 return;
1064         }
1065
1066         /*
1067          * If there is no ondisk info or no buffer the node failed to load,
1068          * remove the last reference and destroy the node.
1069          */
1070         if (node->ondisk == NULL) {
1071                 hammer_unref(&node->lock);
1072                 hammer_flush_node(node);
1073                 /* node is stale now */
1074                 return;
1075         }
1076
1077         /*
1078          * Do final cleanups and then either destroy the node and leave it
1079          * passively cached.  The buffer reference is removed regardless.
1080          */
1081         buffer = node->buffer;
1082         node->ondisk = NULL;
1083
1084         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1085                 hammer_unref(&node->lock);
1086                 hammer_rel_buffer(buffer, 0);
1087                 return;
1088         }
1089
1090         /*
1091          * Destroy the node.
1092          */
1093         hammer_unref(&node->lock);
1094         hammer_flush_node(node);
1095         /* node is stale */
1096         hammer_rel_buffer(buffer, 0);
1097 }
1098
1099 /*
1100  * Free space on-media associated with a B-Tree node.
1101  */
1102 void
1103 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1104 {
1105         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1106         node->flags |= HAMMER_NODE_DELETED;
1107         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1108 }
1109
1110 /*
1111  * Passively cache a referenced hammer_node in *cache.  The caller may
1112  * release the node on return.
1113  */
1114 void
1115 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1116 {
1117         hammer_node_t old;
1118
1119         /*
1120          * If the node is being deleted, don't cache it!
1121          */
1122         if (node->flags & HAMMER_NODE_DELETED)
1123                 return;
1124
1125         /*
1126          * Cache the node.  If we previously cached a different node we
1127          * have to give HAMMER a chance to destroy it.
1128          */
1129 again:
1130         if (node->cache1 != cache) {
1131                 if (node->cache2 != cache) {
1132                         if ((old = *cache) != NULL) {
1133                                 KKASSERT(node->lock.refs != 0);
1134                                 hammer_uncache_node(cache);
1135                                 goto again;
1136                         }
1137                         if (node->cache2)
1138                                 *node->cache2 = NULL;
1139                         node->cache2 = node->cache1;
1140                         node->cache1 = cache;
1141                         *cache = node;
1142                 } else {
1143                         struct hammer_node **tmp;
1144                         tmp = node->cache1;
1145                         node->cache1 = node->cache2;
1146                         node->cache2 = tmp;
1147                 }
1148         }
1149 }
1150
1151 void
1152 hammer_uncache_node(struct hammer_node **cache)
1153 {
1154         hammer_node_t node;
1155
1156         if ((node = *cache) != NULL) {
1157                 *cache = NULL;
1158                 if (node->cache1 == cache) {
1159                         node->cache1 = node->cache2;
1160                         node->cache2 = NULL;
1161                 } else if (node->cache2 == cache) {
1162                         node->cache2 = NULL;
1163                 } else {
1164                         panic("hammer_uncache_node: missing cache linkage");
1165                 }
1166                 if (node->cache1 == NULL && node->cache2 == NULL)
1167                         hammer_flush_node(node);
1168         }
1169 }
1170
1171 /*
1172  * Remove a node's cache references and destroy the node if it has no
1173  * other references or backing store.
1174  */
1175 void
1176 hammer_flush_node(hammer_node_t node)
1177 {
1178         hammer_buffer_t buffer;
1179
1180         if (node->cache1)
1181                 *node->cache1 = NULL;
1182         if (node->cache2)
1183                 *node->cache2 = NULL;
1184         if (node->lock.refs == 0 && node->ondisk == NULL) {
1185                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1186                 if ((buffer = node->buffer) != NULL) {
1187                         node->buffer = NULL;
1188                         TAILQ_REMOVE(&buffer->clist, node, entry);
1189                         /* buffer is unreferenced because ondisk is NULL */
1190                 }
1191                 --hammer_count_nodes;
1192                 kfree(node, M_HAMMER);
1193         }
1194 }
1195
1196 /*
1197  * Flush passively cached B-Tree nodes associated with this buffer.
1198  * This is only called when the buffer is about to be destroyed, so
1199  * none of the nodes should have any references.  The buffer is locked.
1200  *
1201  * We may be interlocked with the buffer.
1202  */
1203 void
1204 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1205 {
1206         hammer_node_t node;
1207
1208         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1209                 KKASSERT(node->ondisk == NULL);
1210
1211                 if (node->lock.refs == 0) {
1212                         hammer_ref(&node->lock);
1213                         node->flags |= HAMMER_NODE_FLUSH;
1214                         hammer_rel_node(node);
1215                 } else {
1216                         KKASSERT(node->loading != 0);
1217                         KKASSERT(node->buffer != NULL);
1218                         buffer = node->buffer;
1219                         node->buffer = NULL;
1220                         TAILQ_REMOVE(&buffer->clist, node, entry);
1221                         /* buffer is unreferenced because ondisk is NULL */
1222                 }
1223         }
1224 }
1225
1226
1227 /************************************************************************
1228  *                              ALLOCATORS                              *
1229  ************************************************************************/
1230
1231 /*
1232  * Allocate a B-Tree node.
1233  */
1234 hammer_node_t
1235 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1236 {
1237         hammer_buffer_t buffer = NULL;
1238         hammer_node_t node = NULL;
1239         hammer_off_t node_offset;
1240
1241         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1242                                             sizeof(struct hammer_node_ondisk),
1243                                             errorp);
1244         if (*errorp == 0) {
1245                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1246                 hammer_modify_node_noundo(trans, node);
1247                 bzero(node->ondisk, sizeof(*node->ondisk));
1248                 hammer_modify_node_done(node);
1249         }
1250         if (buffer)
1251                 hammer_rel_buffer(buffer, 0);
1252         return(node);
1253 }
1254
1255 /*
1256  * Allocate data.  If the address of a data buffer is supplied then
1257  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1258  * will be set to the related buffer.  The caller must release it when
1259  * finally done.  The initial *data_bufferp should be set to NULL by
1260  * the caller.
1261  *
1262  * The caller is responsible for making hammer_modify*() calls on the
1263  * *data_bufferp.
1264  */
1265 void *
1266 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1267                   hammer_off_t *data_offsetp,
1268                   struct hammer_buffer **data_bufferp, int *errorp)
1269 {
1270         void *data;
1271
1272         /*
1273          * Allocate data
1274          */
1275         if (data_len) {
1276                 if (data_len < HAMMER_BUFSIZE) {
1277                         *data_offsetp = hammer_blockmap_alloc(trans,
1278                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1279                                                 data_len, errorp);
1280                 } else {
1281                         *data_offsetp = hammer_blockmap_alloc(trans,
1282                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1283                                                 data_len, errorp);
1284                 }
1285         } else {
1286                 *data_offsetp = 0;
1287         }
1288         if (*errorp == 0 && data_bufferp) {
1289                 if (data_len) {
1290                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1291                                             data_bufferp);
1292                         KKASSERT(*errorp == 0);
1293                 } else {
1294                         data = NULL;
1295                 }
1296         } else {
1297                 data = NULL;
1298         }
1299         KKASSERT(*errorp == 0);
1300         return(data);
1301 }
1302
1303 /*
1304  * Sync dirty buffers to the media and clean-up any loose ends.
1305  */
1306 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1307 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1308
1309 int
1310 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1311 {
1312         struct hammer_sync_info info;
1313
1314         info.error = 0;
1315         info.waitfor = waitfor;
1316         if (waitfor == MNT_WAIT) {
1317                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1318                               hammer_sync_scan1, hammer_sync_scan2, &info);
1319         } else {
1320                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1321                               hammer_sync_scan1, hammer_sync_scan2, &info);
1322         }
1323         return(info.error);
1324 }
1325
1326 int
1327 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1328 {
1329         struct hammer_sync_info info;
1330
1331         info.error = 0;
1332         info.waitfor = waitfor;
1333
1334         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1335                       hammer_sync_scan1, hammer_sync_scan2, &info);
1336         if (waitfor == MNT_WAIT)
1337                 hammer_flusher_sync(hmp);
1338         else
1339                 hammer_flusher_async(hmp);
1340
1341         return(info.error);
1342 }
1343
1344 static int
1345 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1346 {
1347         struct hammer_inode *ip;
1348
1349         ip = VTOI(vp);
1350         if (vp->v_type == VNON || ip == NULL ||
1351             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1352              RB_EMPTY(&vp->v_rbdirty_tree))) {
1353                 return(-1);
1354         }
1355         return(0);
1356 }
1357
1358 static int
1359 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1360 {
1361         struct hammer_sync_info *info = data;
1362         struct hammer_inode *ip;
1363         int error;
1364
1365         ip = VTOI(vp);
1366         if (vp->v_type == VNON || vp->v_type == VBAD ||
1367             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1368              RB_EMPTY(&vp->v_rbdirty_tree))) {
1369                 return(0);
1370         }
1371         error = VOP_FSYNC(vp, info->waitfor);
1372         if (error)
1373                 info->error = error;
1374         return(0);
1375 }
1376