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 1.54 2008/06/11 22:33:21 dillon Exp $
  35  */
  36 /*
  37  * Manage HAMMER's on-disk structures.  These routines are primarily
  38  * responsible for interfacing with the kernel's I/O subsystem and for
  39  * managing in-memory structures.
  40  */
  41
  42 #include "hammer.h"
  43 #include <sys/fcntl.h>
  44 #include <sys/nlookup.h>
  45 #include <sys/buf.h>
  46 #include <sys/buf2.h>
  47
  48 static void hammer_free_volume(hammer_volume_t volume);
  49 static int hammer_load_volume(hammer_volume_t volume);
  50 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
  51 static int hammer_load_node(hammer_node_t node, int isnew);
  52
  53 /*
  54  * Red-Black tree support for various structures
  55  */
  56 int
  57 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  58 {
  59         if (ip1->obj_id < ip2->obj_id)
  60                 return(-1);
  61         if (ip1->obj_id > ip2->obj_id)
  62                 return(1);
  63         if (ip1->obj_asof < ip2->obj_asof)
  64                 return(-1);
  65         if (ip1->obj_asof > ip2->obj_asof)
  66                 return(1);
  67         return(0);
  68 }
  69
  70 static int
  71 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  72 {
  73         if (info->obj_id < ip->obj_id)
  74                 return(-1);
  75         if (info->obj_id > ip->obj_id)
  76                 return(1);
  77         if (info->obj_asof < ip->obj_asof)
  78                 return(-1);
  79         if (info->obj_asof > ip->obj_asof)
  80                 return(1);
  81         return(0);
  82 }
  83
  84 static int
  85 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  86 {
  87         if (vol1->vol_no < vol2->vol_no)
  88                 return(-1);
  89         if (vol1->vol_no > vol2->vol_no)
  90                 return(1);
  91         return(0);
  92 }
  93
  94 static int
  95 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  96 {
  97         if (buf1->zoneX_offset < buf2->zoneX_offset)
  98                 return(-1);
  99         if (buf1->zoneX_offset > buf2->zoneX_offset)
 100                 return(1);
 101         return(0);
 102 }
 103
 104 static int
 105 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 106 {
 107         if (node1->node_offset < node2->node_offset)
 108                 return(-1);
 109         if (node1->node_offset > node2->node_offset)
 110                 return(1);
 111         return(0);
 112 }
 113
 114 /*
 115  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 116  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 117  * functions are normal, e.g. hammer_buf_rb_tree_RB_LOOKUP(root, zone2_offset).
 118  */
 119 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 120 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 121                 hammer_inode_info_cmp, hammer_inode_info_t);
 122 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 123              hammer_vol_rb_compare, int32_t, vol_no);
 124 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 125              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
 126 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 127              hammer_nod_rb_compare, hammer_off_t, node_offset);
 128
 129 /************************************************************************
 130  *                              VOLUMES                                 *
 131  ************************************************************************
 132  *
 133  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 134  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 135  * not load a new volume.
 136  *
 137  * Calls made to hammer_load_volume() or single-threaded
 138  */
 139 int
 140 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 141 {
 142         struct mount *mp;
 143         hammer_volume_t volume;
 144         struct hammer_volume_ondisk *ondisk;
 145         struct nlookupdata nd;
 146         struct buf *bp = NULL;
 147         int error;
 148         int ronly;
 149         int setmp = 0;
 150
 151         mp = hmp->mp;
 152         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 153
 154         /*
 155          * Allocate a volume structure
 156          */
 157         ++hammer_count_volumes;
 158         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 159         volume->vol_name = kstrdup(volname, M_HAMMER);
 160         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 161         volume->io.offset = 0LL;
 162
 163         /*
 164          * Get the device vnode
 165          */
 166         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 167         if (error == 0)
 168                 error = nlookup(&nd);
 169         if (error == 0)
 170                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 171         nlookup_done(&nd);
 172         if (error == 0) {
 173                 if (vn_isdisk(volume->devvp, &error)) {
 174                         error = vfs_mountedon(volume->devvp);
 175                 }
 176         }
 177         if (error == 0 &&
 178             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 179                 error = EBUSY;
 180         }
 181         if (error == 0) {
 182                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 183                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 184                 if (error == 0) {
 185                         error = VOP_OPEN(volume->devvp,
 186                                          (ronly ? FREAD : FREAD|FWRITE),
 187                                          FSCRED, NULL);
 188                 }
 189                 vn_unlock(volume->devvp);
 190         }
 191         if (error) {
 192                 hammer_free_volume(volume);
 193                 return(error);
 194         }
 195         volume->devvp->v_rdev->si_mountpoint = mp;
 196         setmp = 1;
 197
 198         /*
 199          * Extract the volume number from the volume header and do various
 200          * sanity checks.
 201          */
 202         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 203         if (error)
 204                 goto late_failure;
 205         ondisk = (void *)bp->b_data;
 206         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 207                 kprintf("hammer_mount: volume %s has an invalid header\n",
 208                         volume->vol_name);
 209                 error = EFTYPE;
 210                 goto late_failure;
 211         }
 212         volume->vol_no = ondisk->vol_no;
 213         volume->buffer_base = ondisk->vol_buf_beg;
 214         volume->vol_flags = ondisk->vol_flags;
 215         volume->nblocks = ondisk->vol_nblocks;
 216         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 217                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 218         volume->maxraw_off = ondisk->vol_buf_end;
 219
 220         if (RB_EMPTY(&hmp->rb_vols_root)) {
 221                 hmp->fsid = ondisk->vol_fsid;
 222         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 223                 kprintf("hammer_mount: volume %s's fsid does not match "
 224                         "other volumes\n", volume->vol_name);
 225                 error = EFTYPE;
 226                 goto late_failure;
 227         }
 228
 229         /*
 230          * Insert the volume structure into the red-black tree.
 231          */
 232         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 233                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 234                         volume->vol_name, volume->vol_no);
 235                 error = EEXIST;
 236         }
 237
 238         /*
 239          * Set the root volume .  HAMMER special cases rootvol the structure.
 240          * We do not hold a ref because this would prevent related I/O
 241          * from being flushed.
 242          */
 243         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 244                 hmp->rootvol = volume;
 245                 if (bp) {
 246                         brelse(bp);
 247                         bp = NULL;
 248                 }
 249                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 250                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 251                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 252                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 253                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 254         }
 255 late_failure:
 256         if (bp)
 257                 brelse(bp);
 258         if (error) {
 259                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 260                 if (setmp)
 261                         volume->devvp->v_rdev->si_mountpoint = NULL;
 262                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 263                 hammer_free_volume(volume);
 264         }
 265         return (error);
 266 }
 267
 268 /*
 269  * This is called for each volume when updating the mount point from
 270  * read-write to read-only or vise-versa.
 271  */
 272 int
 273 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 274 {
 275         if (volume->devvp) {
 276                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 277                 if (volume->io.hmp->ronly) {
 278                         /* do not call vinvalbuf */
 279                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 280                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 281                 } else {
 282                         /* do not call vinvalbuf */
 283                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 284                         VOP_CLOSE(volume->devvp, FREAD);
 285                 }
 286                 vn_unlock(volume->devvp);
 287         }
 288         return(0);
 289 }
 290
 291 /*
 292  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 293  * so returns -1 on failure.
 294  */
 295 int
 296 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 297 {
 298         struct hammer_mount *hmp = volume->io.hmp;
 299         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 300
 301         /*
 302          * Clean up the root volume pointer, which is held unlocked in hmp.
 303          */
 304         if (hmp->rootvol == volume)
 305                 hmp->rootvol = NULL;
 306
 307         /*
 308          * Release our buffer and flush anything left in the buffer cache.
 309          */
 310         volume->io.waitdep = 1;
 311         hammer_io_release(&volume->io, 1);
 312         hammer_io_clear_modlist(&volume->io);
 313
 314         /*
 315          * There should be no references on the volume, no clusters, and
 316          * no super-clusters.
 317          */
 318         KKASSERT(volume->io.lock.refs == 0);
 319
 320         volume->ondisk = NULL;
 321         if (volume->devvp) {
 322                 if (volume->devvp->v_rdev &&
 323                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 324                 ) {
 325                         volume->devvp->v_rdev->si_mountpoint = NULL;
 326                 }
 327                 if (ronly) {
 328                         vinvalbuf(volume->devvp, 0, 0, 0);
 329                         VOP_CLOSE(volume->devvp, FREAD);
 330                 } else {
 331                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 332                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 333                 }
 334         }
 335
 336         /*
 337          * Destroy the structure
 338          */
 339         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 340         hammer_free_volume(volume);
 341         return(0);
 342 }
 343
 344 static
 345 void
 346 hammer_free_volume(hammer_volume_t volume)
 347 {
 348         if (volume->vol_name) {
 349                 kfree(volume->vol_name, M_HAMMER);
 350                 volume->vol_name = NULL;
 351         }
 352         if (volume->devvp) {
 353                 vrele(volume->devvp);
 354                 volume->devvp = NULL;
 355         }
 356         --hammer_count_volumes;
 357         kfree(volume, M_HAMMER);
 358 }
 359
 360 /*
 361  * Get a HAMMER volume.  The volume must already exist.
 362  */
 363 hammer_volume_t
 364 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 365 {
 366         struct hammer_volume *volume;
 367
 368         /*
 369          * Locate the volume structure
 370          */
 371         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 372         if (volume == NULL) {
 373                 *errorp = ENOENT;
 374                 return(NULL);
 375         }
 376         hammer_ref(&volume->io.lock);
 377
 378         /*
 379          * Deal with on-disk info
 380          */
 381         if (volume->ondisk == NULL || volume->io.loading) {
 382                 *errorp = hammer_load_volume(volume);
 383                 if (*errorp) {
 384                         hammer_rel_volume(volume, 1);
 385                         volume = NULL;
 386                 }
 387         } else {
 388                 *errorp = 0;
 389         }
 390         return(volume);
 391 }
 392
 393 int
 394 hammer_ref_volume(hammer_volume_t volume)
 395 {
 396         int error;
 397
 398         hammer_ref(&volume->io.lock);
 399
 400         /*
 401          * Deal with on-disk info
 402          */
 403         if (volume->ondisk == NULL || volume->io.loading) {
 404                 error = hammer_load_volume(volume);
 405                 if (error)
 406                         hammer_rel_volume(volume, 1);
 407         } else {
 408                 error = 0;
 409         }
 410         return (error);
 411 }
 412
 413 hammer_volume_t
 414 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 415 {
 416         hammer_volume_t volume;
 417
 418         volume = hmp->rootvol;
 419         KKASSERT(volume != NULL);
 420         hammer_ref(&volume->io.lock);
 421
 422         /*
 423          * Deal with on-disk info
 424          */
 425         if (volume->ondisk == NULL || volume->io.loading) {
 426                 *errorp = hammer_load_volume(volume);
 427                 if (*errorp) {
 428                         hammer_rel_volume(volume, 1);
 429                         volume = NULL;
 430                 }
 431         } else {
 432                 *errorp = 0;
 433         }
 434         return (volume);
 435 }
 436
 437 /*
 438  * Load a volume's on-disk information.  The volume must be referenced and
 439  * not locked.  We temporarily acquire an exclusive lock to interlock
 440  * against releases or multiple get's.
 441  */
 442 static int
 443 hammer_load_volume(hammer_volume_t volume)
 444 {
 445         int error;
 446
 447         ++volume->io.loading;
 448         hammer_lock_ex(&volume->io.lock);
 449
 450         if (volume->ondisk == NULL) {
 451                 error = hammer_io_read(volume->devvp, &volume->io,
 452                                        volume->maxraw_off);
 453                 if (error == 0)
 454                         volume->ondisk = (void *)volume->io.bp->b_data;
 455         } else {
 456                 error = 0;
 457         }
 458         --volume->io.loading;
 459         hammer_unlock(&volume->io.lock);
 460         return(error);
 461 }
 462
 463 /*
 464  * Release a volume.  Call hammer_io_release on the last reference.  We have
 465  * to acquire an exclusive lock to interlock against volume->ondisk tests
 466  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 467  * lock to be held.
 468  *
 469  * Volumes are not unloaded from memory during normal operation.
 470  */
 471 void
 472 hammer_rel_volume(hammer_volume_t volume, int flush)
 473 {
 474         crit_enter();
 475         if (volume->io.lock.refs == 1) {
 476                 ++volume->io.loading;
 477                 hammer_lock_ex(&volume->io.lock);
 478                 if (volume->io.lock.refs == 1) {
 479                         volume->ondisk = NULL;
 480                         hammer_io_release(&volume->io, flush);
 481                 }
 482                 --volume->io.loading;
 483                 hammer_unlock(&volume->io.lock);
 484         }
 485         hammer_unref(&volume->io.lock);
 486         crit_exit();
 487 }
 488
 489 /************************************************************************
 490  *                              BUFFERS                                 *
 491  ************************************************************************
 492  *
 493  * Manage buffers.  Currently all blockmap-backed zones are translated
 494  * to zone-2 buffer offsets.
 495  */
 496 hammer_buffer_t
 497 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 498                   int isnew, int *errorp)
 499 {
 500         hammer_buffer_t buffer;
 501         hammer_volume_t volume;
 502         hammer_off_t    zone2_offset;
 503         hammer_io_type_t iotype;
 504         int vol_no;
 505         int zone;
 506
 507 again:
 508         /*
 509          * Shortcut if the buffer is already cached
 510          */
 511         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 512                            buf_offset & ~HAMMER_BUFMASK64);
 513         if (buffer) {
 514                 if (buffer->io.lock.refs == 0)
 515                         ++hammer_count_refedbufs;
 516                 hammer_ref(&buffer->io.lock);
 517                 if (buffer->ondisk && buffer->io.loading == 0) {
 518                         *errorp = 0;
 519                         return(buffer);
 520                 }
 521
 522                 /*
 523                  * The buffer is no longer loose if it has a ref.  Loose
 524                  * buffers will never be in a modified state.  This should
 525                  * only occur on the 0->1 transition of refs.
 526                  */
 527                 if (buffer->io.mod_list == &hmp->lose_list) {
 528                         crit_enter();   /* biodone race against list */
 529                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 530                                      mod_entry);
 531                         crit_exit();
 532                         buffer->io.mod_list = NULL;
 533                         KKASSERT(buffer->io.modified == 0);
 534                 }
 535                 goto found;
 536         }
 537
 538         /*
 539          * What is the buffer class?
 540          */
 541         zone = HAMMER_ZONE_DECODE(buf_offset);
 542
 543         switch(zone) {
 544         case HAMMER_ZONE_LARGE_DATA_INDEX:
 545         case HAMMER_ZONE_SMALL_DATA_INDEX:
 546                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 547                 break;
 548         case HAMMER_ZONE_UNDO_INDEX:
 549                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 550                 break;
 551         default:
 552                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 553                 break;
 554         }
 555
 556         /*
 557          * Handle blockmap offset translations
 558          */
 559         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 560                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 561         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 562                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 563         } else {
 564                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 565                 zone2_offset = buf_offset;
 566                 *errorp = 0;
 567         }
 568         if (*errorp)
 569                 return(NULL);
 570
 571         /*
 572          * Calculate the base zone2-offset and acquire the volume
 573          *
 574          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 575          * specifications.
 576          */
 577         zone2_offset &= ~HAMMER_BUFMASK64;
 578         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 579                  HAMMER_ZONE_RAW_BUFFER);
 580         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 581         volume = hammer_get_volume(hmp, vol_no, errorp);
 582         if (volume == NULL)
 583                 return(NULL);
 584
 585         KKASSERT(zone2_offset < volume->maxbuf_off);
 586
 587         /*
 588          * Allocate a new buffer structure.  We will check for races later.
 589          */
 590         ++hammer_count_buffers;
 591         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 592         buffer->zone2_offset = zone2_offset;
 593         buffer->zoneX_offset = buf_offset;
 594         buffer->volume = volume;
 595
 596         hammer_io_init(&buffer->io, hmp, iotype);
 597         buffer->io.offset = volume->ondisk->vol_buf_beg +
 598                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 599         TAILQ_INIT(&buffer->clist);
 600         hammer_ref(&buffer->io.lock);
 601
 602         /*
 603          * Insert the buffer into the RB tree and handle late collisions.
 604          */
 605         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 606                 hammer_unref(&buffer->io.lock);
 607                 --hammer_count_buffers;
 608                 kfree(buffer, M_HAMMER);
 609                 goto again;
 610         }
 611         ++hammer_count_refedbufs;
 612 found:
 613
 614         /*
 615          * Deal with on-disk info and loading races.
 616          */
 617         if (buffer->ondisk == NULL || buffer->io.loading) {
 618                 *errorp = hammer_load_buffer(buffer, isnew);
 619                 if (*errorp) {
 620                         hammer_rel_buffer(buffer, 1);
 621                         buffer = NULL;
 622                 }
 623         } else {
 624                 *errorp = 0;
 625         }
 626         return(buffer);
 627 }
 628
 629 /*
 630  * Destroy all buffers covering the specified zoneX offset range.  This
 631  * is called when the related blockmap layer2 entry is freed or when
 632  * a direct write bypasses our buffer/buffer-cache subsystem.
 633  *
 634  * The buffers may be referenced by the caller itself.  Setting reclaim
 635  * will cause the buffer to be destroyed when it's ref count reaches zero.
 636  */
 637 void
 638 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 639                    hammer_off_t zone2_offset, int bytes)
 640 {
 641         hammer_buffer_t buffer;
 642         hammer_volume_t volume;
 643         int vol_no;
 644         int error;
 645
 646         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 647         volume = hammer_get_volume(hmp, vol_no, &error);
 648         KKASSERT(error == 0);
 649
 650         while (bytes > 0) {
 651                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 652                                    base_offset);
 653                 if (buffer) {
 654                         KKASSERT(buffer->zone2_offset == zone2_offset);
 655                         hammer_io_clear_modify(&buffer->io);
 656                         buffer->io.reclaim = 1;
 657                         KKASSERT(buffer->volume == volume);
 658                         if (buffer->io.lock.refs == 0)
 659                                 hammer_unload_buffer(buffer, NULL);
 660                 } else {
 661                         hammer_io_inval(volume, zone2_offset);
 662                 }
 663                 base_offset += HAMMER_BUFSIZE;
 664                 zone2_offset += HAMMER_BUFSIZE;
 665                 bytes -= HAMMER_BUFSIZE;
 666         }
 667         hammer_rel_volume(volume, 0);
 668 }
 669
 670 static int
 671 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 672 {
 673         hammer_volume_t volume;
 674         int error;
 675
 676         /*
 677          * Load the buffer's on-disk info
 678          */
 679         volume = buffer->volume;
 680         ++buffer->io.loading;
 681         hammer_lock_ex(&buffer->io.lock);
 682
 683         if (hammer_debug_io & 0x0001) {
 684                 kprintf("load_buffer %016llx %016llx isnew=%d\n",
 685                         buffer->zoneX_offset, buffer->zone2_offset, isnew);
 686         }
 687
 688         if (buffer->ondisk == NULL) {
 689                 if (isnew) {
 690                         error = hammer_io_new(volume->devvp, &buffer->io);
 691                 } else {
 692                         error = hammer_io_read(volume->devvp, &buffer->io,
 693                                                volume->maxraw_off);
 694                 }
 695                 if (error == 0)
 696                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 697         } else if (isnew) {
 698                 error = hammer_io_new(volume->devvp, &buffer->io);
 699         } else {
 700                 error = 0;
 701         }
 702         --buffer->io.loading;
 703         hammer_unlock(&buffer->io.lock);
 704         return (error);
 705 }
 706
 707 /*
 708  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 709  */
 710 int
 711 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 712 {
 713         ++hammer_count_refedbufs;
 714         hammer_ref(&buffer->io.lock);
 715         hammer_flush_buffer_nodes(buffer);
 716         KKASSERT(buffer->io.lock.refs == 1);
 717         hammer_rel_buffer(buffer, 2);
 718         return(0);
 719 }
 720
 721 /*
 722  * Reference a buffer that is either already referenced or via a specially
 723  * handled pointer (aka cursor->buffer).
 724  */
 725 int
 726 hammer_ref_buffer(hammer_buffer_t buffer)
 727 {
 728         int error;
 729
 730         if (buffer->io.lock.refs == 0)
 731                 ++hammer_count_refedbufs;
 732         hammer_ref(&buffer->io.lock);
 733
 734         /*
 735          * No longer loose
 736          */
 737         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 738                 crit_enter();   /* biodone race against list */
 739                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 740                 buffer->io.mod_list = NULL;
 741                 crit_exit();
 742         }
 743
 744         if (buffer->ondisk == NULL || buffer->io.loading) {
 745                 error = hammer_load_buffer(buffer, 0);
 746                 if (error) {
 747                         hammer_rel_buffer(buffer, 1);
 748                         /*
 749                          * NOTE: buffer pointer can become stale after
 750                          * the above release.
 751                          */
 752                 }
 753         } else {
 754                 error = 0;
 755         }
 756         return(error);
 757 }
 758
 759 /*
 760  * Release a buffer.  We have to deal with several places where
 761  * another thread can ref the buffer.
 762  *
 763  * Only destroy the structure itself if the related buffer cache buffer
 764  * was disassociated from it.  This ties the management of the structure
 765  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 766  * embedded io is referenced or not.
 767  */
 768 void
 769 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 770 {
 771         hammer_volume_t volume;
 772         int freeme = 0;
 773
 774         crit_enter();
 775         if (buffer->io.lock.refs == 1) {
 776                 ++buffer->io.loading;   /* force interlock check */
 777                 hammer_lock_ex(&buffer->io.lock);
 778                 if (buffer->io.lock.refs == 1) {
 779                         hammer_io_release(&buffer->io, flush);
 780                         hammer_flush_buffer_nodes(buffer);
 781                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 782                         --hammer_count_refedbufs;
 783
 784                         if (buffer->io.bp == NULL &&
 785                             buffer->io.lock.refs == 1) {
 786                                 /*
 787                                  * Final cleanup
 788                                  */
 789                                 RB_REMOVE(hammer_buf_rb_tree,
 790                                           &buffer->io.hmp->rb_bufs_root,
 791                                           buffer);
 792                                 volume = buffer->volume;
 793                                 buffer->volume = NULL; /* sanity */
 794                                 hammer_rel_volume(volume, 0);
 795                                 hammer_io_clear_modlist(&buffer->io);
 796                                 freeme = 1;
 797                         }
 798                 }
 799                 --buffer->io.loading;
 800                 hammer_unlock(&buffer->io.lock);
 801         }
 802         hammer_unref(&buffer->io.lock);
 803         crit_exit();
 804         if (freeme) {
 805                 --hammer_count_buffers;
 806                 kfree(buffer, M_HAMMER);
 807         }
 808 }
 809
 810 /*
 811  * Access the filesystem buffer containing the specified hammer offset.
 812  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 813  * relative buffer offset.  It must also have bit 55 set to be valid.
 814  * (see hammer_off_t in hammer_disk.h).
 815  *
 816  * Any prior buffer in *bufferp will be released and replaced by the
 817  * requested buffer.
 818  */
 819 void *
 820 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 821              struct hammer_buffer **bufferp)
 822 {
 823         hammer_buffer_t buffer;
 824         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 825
 826         buf_offset &= ~HAMMER_BUFMASK64;
 827         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 828
 829         buffer = *bufferp;
 830         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 831                                buffer->zoneX_offset != buf_offset)) {
 832                 if (buffer)
 833                         hammer_rel_buffer(buffer, 0);
 834                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 835                 *bufferp = buffer;
 836         } else {
 837                 *errorp = 0;
 838         }
 839
 840         /*
 841          * Return a pointer to the buffer data.
 842          */
 843         if (buffer == NULL)
 844                 return(NULL);
 845         else
 846                 return((char *)buffer->ondisk + xoff);
 847 }
 848
 849 /*
 850  * Access the filesystem buffer containing the specified hammer offset.
 851  * No disk read operation occurs.  The result buffer may contain garbage.
 852  *
 853  * Any prior buffer in *bufferp will be released and replaced by the
 854  * requested buffer.
 855  *
 856  * This function marks the buffer dirty but does not increment its
 857  * modify_refs count.
 858  */
 859 void *
 860 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 861              struct hammer_buffer **bufferp)
 862 {
 863         hammer_buffer_t buffer;
 864         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 865
 866         buf_offset &= ~HAMMER_BUFMASK64;
 867
 868         buffer = *bufferp;
 869         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 870                                buffer->zoneX_offset != buf_offset)) {
 871                 if (buffer)
 872                         hammer_rel_buffer(buffer, 0);
 873                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 874                 *bufferp = buffer;
 875         } else {
 876                 *errorp = 0;
 877         }
 878
 879         /*
 880          * Return a pointer to the buffer data.
 881          */
 882         if (buffer == NULL)
 883                 return(NULL);
 884         else
 885                 return((char *)buffer->ondisk + xoff);
 886 }
 887
 888 /************************************************************************
 889  *                              NODES                                   *
 890  ************************************************************************
 891  *
 892  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 893  * method used by the HAMMER filesystem.
 894  *
 895  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 896  * associated with its buffer, and will only referenced the buffer while
 897  * the node itself is referenced.
 898  *
 899  * A hammer_node can also be passively associated with other HAMMER
 900  * structures, such as inodes, while retaining 0 references.  These
 901  * associations can be cleared backwards using a pointer-to-pointer in
 902  * the hammer_node.
 903  *
 904  * This allows the HAMMER implementation to cache hammer_nodes long-term
 905  * and short-cut a great deal of the infrastructure's complexity.  In
 906  * most cases a cached node can be reacquired without having to dip into
 907  * either the buffer or cluster management code.
 908  *
 909  * The caller must pass a referenced cluster on call and will retain
 910  * ownership of the reference on return.  The node will acquire its own
 911  * additional references, if necessary.
 912  */
 913 hammer_node_t
 914 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 915                 int isnew, int *errorp)
 916 {
 917         hammer_node_t node;
 918
 919         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 920
 921         /*
 922          * Locate the structure, allocating one if necessary.
 923          */
 924 again:
 925         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 926         if (node == NULL) {
 927                 ++hammer_count_nodes;
 928                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 929                 node->node_offset = node_offset;
 930                 node->hmp = hmp;
 931                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 932                         --hammer_count_nodes;
 933                         kfree(node, M_HAMMER);
 934                         goto again;
 935                 }
 936         }
 937         hammer_ref(&node->lock);
 938         if (node->ondisk)
 939                 *errorp = 0;
 940         else
 941                 *errorp = hammer_load_node(node, isnew);
 942         if (*errorp) {
 943                 hammer_rel_node(node);
 944                 node = NULL;
 945         }
 946         return(node);
 947 }
 948
 949 /*
 950  * Reference an already-referenced node.
 951  */
 952 void
 953 hammer_ref_node(hammer_node_t node)
 954 {
 955         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 956         hammer_ref(&node->lock);
 957 }
 958
 959 /*
 960  * Load a node's on-disk data reference.
 961  */
 962 static int
 963 hammer_load_node(hammer_node_t node, int isnew)
 964 {
 965         hammer_buffer_t buffer;
 966         hammer_off_t buf_offset;
 967         int error;
 968
 969         error = 0;
 970         ++node->loading;
 971         hammer_lock_ex(&node->lock);
 972         if (node->ondisk == NULL) {
 973                 /*
 974                  * This is a little confusing but the jist is that
 975                  * node->buffer determines whether the node is on
 976                  * the buffer's clist and node->ondisk determines
 977                  * whether the buffer is referenced.
 978                  *
 979                  * We could be racing a buffer release, in which case
 980                  * node->buffer may become NULL while we are blocked
 981                  * referencing the buffer.
 982                  */
 983                 if ((buffer = node->buffer) != NULL) {
 984                         error = hammer_ref_buffer(buffer);
 985                         if (error == 0 && node->buffer == NULL) {
 986                                 TAILQ_INSERT_TAIL(&buffer->clist,
 987                                                   node, entry);
 988                                 node->buffer = buffer;
 989                         }
 990                 } else {
 991                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
 992                         buffer = hammer_get_buffer(node->hmp, buf_offset,
 993                                                    0, &error);
 994                         if (buffer) {
 995                                 KKASSERT(error == 0);
 996                                 TAILQ_INSERT_TAIL(&buffer->clist,
 997                                                   node, entry);
 998                                 node->buffer = buffer;
 999                         }
1000                 }
1001                 if (error == 0) {
1002                         node->ondisk = (void *)((char *)buffer->ondisk +
1003                                (node->node_offset & HAMMER_BUFMASK));
1004                         if (isnew == 0 &&
1005                             hammer_crc_test_btree(node->ondisk) == 0) {
1006                                 Debugger("CRC FAILED: B-TREE NODE");
1007                         }
1008                 }
1009         }
1010         --node->loading;
1011         hammer_unlock(&node->lock);
1012         return (error);
1013 }
1014
1015 /*
1016  * Safely reference a node, interlock against flushes via the IO subsystem.
1017  */
1018 hammer_node_t
1019 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
1020                      int *errorp)
1021 {
1022         hammer_node_t node;
1023
1024         node = *cache;
1025         if (node != NULL) {
1026                 hammer_ref(&node->lock);
1027                 if (node->ondisk)
1028                         *errorp = 0;
1029                 else
1030                         *errorp = hammer_load_node(node, 0);
1031                 if (*errorp) {
1032                         hammer_rel_node(node);
1033                         node = NULL;
1034                 }
1035         } else {
1036                 *errorp = ENOENT;
1037         }
1038         return(node);
1039 }
1040
1041 /*
1042  * Release a hammer_node.  On the last release the node dereferences
1043  * its underlying buffer and may or may not be destroyed.
1044  */
1045 void
1046 hammer_rel_node(hammer_node_t node)
1047 {
1048         hammer_buffer_t buffer;
1049
1050         /*
1051          * If this isn't the last ref just decrement the ref count and
1052          * return.
1053          */
1054         if (node->lock.refs > 1) {
1055                 hammer_unref(&node->lock);
1056                 return;
1057         }
1058
1059         /*
1060          * If there is no ondisk info or no buffer the node failed to load,
1061          * remove the last reference and destroy the node.
1062          */
1063         if (node->ondisk == NULL) {
1064                 hammer_unref(&node->lock);
1065                 hammer_flush_node(node);
1066                 /* node is stale now */
1067                 return;
1068         }
1069
1070         /*
1071          * Do final cleanups and then either destroy the node and leave it
1072          * passively cached.  The buffer reference is removed regardless.
1073          */
1074         buffer = node->buffer;
1075         node->ondisk = NULL;
1076
1077         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1078                 hammer_unref(&node->lock);
1079                 hammer_rel_buffer(buffer, 0);
1080                 return;
1081         }
1082
1083         /*
1084          * Destroy the node.
1085          */
1086         hammer_unref(&node->lock);
1087         hammer_flush_node(node);
1088         /* node is stale */
1089         hammer_rel_buffer(buffer, 0);
1090 }
1091
1092 /*
1093  * Free space on-media associated with a B-Tree node.
1094  */
1095 void
1096 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1097 {
1098         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1099         node->flags |= HAMMER_NODE_DELETED;
1100         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1101 }
1102
1103 /*
1104  * Passively cache a referenced hammer_node in *cache.  The caller may
1105  * release the node on return.
1106  */
1107 void
1108 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1109 {
1110         hammer_node_t old;
1111
1112         /*
1113          * If the node is being deleted, don't cache it!
1114          */
1115         if (node->flags & HAMMER_NODE_DELETED)
1116                 return;
1117
1118         /*
1119          * Cache the node.  If we previously cached a different node we
1120          * have to give HAMMER a chance to destroy it.
1121          */
1122 again:
1123         if (node->cache1 != cache) {
1124                 if (node->cache2 != cache) {
1125                         if ((old = *cache) != NULL) {
1126                                 KKASSERT(node->lock.refs != 0);
1127                                 hammer_uncache_node(cache);
1128                                 goto again;
1129                         }
1130                         if (node->cache2)
1131                                 *node->cache2 = NULL;
1132                         node->cache2 = node->cache1;
1133                         node->cache1 = cache;
1134                         *cache = node;
1135                 } else {
1136                         struct hammer_node **tmp;
1137                         tmp = node->cache1;
1138                         node->cache1 = node->cache2;
1139                         node->cache2 = tmp;
1140                 }
1141         }
1142 }
1143
1144 void
1145 hammer_uncache_node(struct hammer_node **cache)
1146 {
1147         hammer_node_t node;
1148
1149         if ((node = *cache) != NULL) {
1150                 *cache = NULL;
1151                 if (node->cache1 == cache) {
1152                         node->cache1 = node->cache2;
1153                         node->cache2 = NULL;
1154                 } else if (node->cache2 == cache) {
1155                         node->cache2 = NULL;
1156                 } else {
1157                         panic("hammer_uncache_node: missing cache linkage");
1158                 }
1159                 if (node->cache1 == NULL && node->cache2 == NULL)
1160                         hammer_flush_node(node);
1161         }
1162 }
1163
1164 /*
1165  * Remove a node's cache references and destroy the node if it has no
1166  * other references or backing store.
1167  */
1168 void
1169 hammer_flush_node(hammer_node_t node)
1170 {
1171         hammer_buffer_t buffer;
1172
1173         if (node->cache1)
1174                 *node->cache1 = NULL;
1175         if (node->cache2)
1176                 *node->cache2 = NULL;
1177         if (node->lock.refs == 0 && node->ondisk == NULL) {
1178                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1179                 if ((buffer = node->buffer) != NULL) {
1180                         node->buffer = NULL;
1181                         TAILQ_REMOVE(&buffer->clist, node, entry);
1182                         /* buffer is unreferenced because ondisk is NULL */
1183                 }
1184                 --hammer_count_nodes;
1185                 kfree(node, M_HAMMER);
1186         }
1187 }
1188
1189 /*
1190  * Flush passively cached B-Tree nodes associated with this buffer.
1191  * This is only called when the buffer is about to be destroyed, so
1192  * none of the nodes should have any references.  The buffer is locked.
1193  *
1194  * We may be interlocked with the buffer.
1195  */
1196 void
1197 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1198 {
1199         hammer_node_t node;
1200
1201         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1202                 KKASSERT(node->ondisk == NULL);
1203
1204                 if (node->lock.refs == 0) {
1205                         hammer_ref(&node->lock);
1206                         node->flags |= HAMMER_NODE_FLUSH;
1207                         hammer_rel_node(node);
1208                 } else {
1209                         KKASSERT(node->loading != 0);
1210                         KKASSERT(node->buffer != NULL);
1211                         buffer = node->buffer;
1212                         node->buffer = NULL;
1213                         TAILQ_REMOVE(&buffer->clist, node, entry);
1214                         /* buffer is unreferenced because ondisk is NULL */
1215                 }
1216         }
1217 }
1218
1219
1220 /************************************************************************
1221  *                              ALLOCATORS                              *
1222  ************************************************************************/
1223
1224 /*
1225  * Allocate a B-Tree node.
1226  */
1227 hammer_node_t
1228 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1229 {
1230         hammer_buffer_t buffer = NULL;
1231         hammer_node_t node = NULL;
1232         hammer_off_t node_offset;
1233
1234         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1235                                             sizeof(struct hammer_node_ondisk),
1236                                             errorp);
1237         if (*errorp == 0) {
1238                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1239                 hammer_modify_node_noundo(trans, node);
1240                 bzero(node->ondisk, sizeof(*node->ondisk));
1241                 hammer_modify_node_done(node);
1242         }
1243         if (buffer)
1244                 hammer_rel_buffer(buffer, 0);
1245         return(node);
1246 }
1247
1248 /*
1249  * Allocate data.  If the address of a data buffer is supplied then
1250  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1251  * will be set to the related buffer.  The caller must release it when
1252  * finally done.  The initial *data_bufferp should be set to NULL by
1253  * the caller.
1254  *
1255  * The caller is responsible for making hammer_modify*() calls on the
1256  * *data_bufferp.
1257  */
1258 void *
1259 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1260                   hammer_off_t *data_offsetp,
1261                   struct hammer_buffer **data_bufferp, int *errorp)
1262 {
1263         void *data;
1264
1265         /*
1266          * Allocate data
1267          */
1268         if (data_len) {
1269                 if (data_len < HAMMER_BUFSIZE) {
1270                         *data_offsetp = hammer_blockmap_alloc(trans,
1271                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1272                                                 data_len, errorp);
1273                 } else {
1274                         *data_offsetp = hammer_blockmap_alloc(trans,
1275                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1276                                                 data_len, errorp);
1277                 }
1278         } else {
1279                 *data_offsetp = 0;
1280         }
1281         if (*errorp == 0 && data_bufferp) {
1282                 if (data_len) {
1283                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1284                                             data_bufferp);
1285                         KKASSERT(*errorp == 0);
1286                 } else {
1287                         data = NULL;
1288                 }
1289         } else {
1290                 data = NULL;
1291         }
1292         KKASSERT(*errorp == 0);
1293         return(data);
1294 }
1295
1296 /*
1297  * Sync dirty buffers to the media and clean-up any loose ends.
1298  */
1299 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1300 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1301
1302 int
1303 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1304 {
1305         struct hammer_sync_info info;
1306
1307         info.error = 0;
1308         info.waitfor = waitfor;
1309         if (waitfor == MNT_WAIT) {
1310                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1311                               hammer_sync_scan1, hammer_sync_scan2, &info);
1312         } else {
1313                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1314                               hammer_sync_scan1, hammer_sync_scan2, &info);
1315         }
1316         return(info.error);
1317 }
1318
1319 int
1320 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1321 {
1322         struct hammer_sync_info info;
1323
1324         info.error = 0;
1325         info.waitfor = waitfor;
1326
1327         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1328                       hammer_sync_scan1, hammer_sync_scan2, &info);
1329         if (waitfor == MNT_WAIT)
1330                 hammer_flusher_sync(hmp);
1331         else
1332                 hammer_flusher_async(hmp);
1333
1334         return(info.error);
1335 }
1336
1337 static int
1338 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1339 {
1340         struct hammer_inode *ip;
1341
1342         ip = VTOI(vp);
1343         if (vp->v_type == VNON || ip == NULL ||
1344             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1345              RB_EMPTY(&vp->v_rbdirty_tree))) {
1346                 return(-1);
1347         }
1348         return(0);
1349 }
1350
1351 static int
1352 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1353 {
1354         struct hammer_sync_info *info = data;
1355         struct hammer_inode *ip;
1356         int error;
1357
1358         ip = VTOI(vp);
1359         if (vp->v_type == VNON || vp->v_type == VBAD ||
1360             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1361              RB_EMPTY(&vp->v_rbdirty_tree))) {
1362                 return(0);
1363         }
1364         error = VOP_FSYNC(vp, info->waitfor);
1365         if (error)
1366                 info->error = error;
1367         return(0);
1368 }
1369