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 1.68 2008/07/12 23:04:50 dillon Exp $
  35  */
  36 /*
  37  * Manage HAMMER's on-disk structures.  These routines are primarily
  38  * responsible for interfacing with the kernel's I/O subsystem and for
  39  * managing in-memory structures.
  40  */
  41
  42 #include "hammer.h"
  43 #include <sys/fcntl.h>
  44 #include <sys/nlookup.h>
  45 #include <sys/buf.h>
  46 #include <sys/buf2.h>
  47
  48 static void hammer_free_volume(hammer_volume_t volume);
  49 static int hammer_load_volume(hammer_volume_t volume);
  50 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
  51 static int hammer_load_node(hammer_node_t node, int isnew);
  52
  53 static int
  54 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  55 {
  56         if (vol1->vol_no < vol2->vol_no)
  57                 return(-1);
  58         if (vol1->vol_no > vol2->vol_no)
  59                 return(1);
  60         return(0);
  61 }
  62
  63 static int
  64 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  65 {
  66         if (buf1->zoneX_offset < buf2->zoneX_offset)
  67                 return(-1);
  68         if (buf1->zoneX_offset > buf2->zoneX_offset)
  69                 return(1);
  70         return(0);
  71 }
  72
  73 static int
  74 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  75 {
  76         if (node1->node_offset < node2->node_offset)
  77                 return(-1);
  78         if (node1->node_offset > node2->node_offset)
  79                 return(1);
  80         return(0);
  81 }
  82
  83 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  84              hammer_vol_rb_compare, int32_t, vol_no);
  85 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  86              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  87 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  88              hammer_nod_rb_compare, hammer_off_t, node_offset);
  89
  90 /************************************************************************
  91  *                              VOLUMES                                 *
  92  ************************************************************************
  93  *
  94  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
  95  * code on failure.  Volumes must be loaded at mount time, get_volume() will
  96  * not load a new volume.
  97  *
  98  * Calls made to hammer_load_volume() or single-threaded
  99  */
 100 int
 101 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 102 {
 103         struct mount *mp;
 104         hammer_volume_t volume;
 105         struct hammer_volume_ondisk *ondisk;
 106         struct nlookupdata nd;
 107         struct buf *bp = NULL;
 108         int error;
 109         int ronly;
 110         int setmp = 0;
 111
 112         mp = hmp->mp;
 113         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 114
 115         /*
 116          * Allocate a volume structure
 117          */
 118         ++hammer_count_volumes;
 119         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 120         volume->vol_name = kstrdup(volname, M_HAMMER);
 121         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 122         volume->io.offset = 0LL;
 123         volume->io.bytes = HAMMER_BUFSIZE;
 124
 125         /*
 126          * Get the device vnode
 127          */
 128         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 129         if (error == 0)
 130                 error = nlookup(&nd);
 131         if (error == 0)
 132                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 133         nlookup_done(&nd);
 134         if (error == 0) {
 135                 if (vn_isdisk(volume->devvp, &error)) {
 136                         error = vfs_mountedon(volume->devvp);
 137                 }
 138         }
 139         if (error == 0 &&
 140             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 141                 error = EBUSY;
 142         }
 143         if (error == 0) {
 144                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 145                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 146                 if (error == 0) {
 147                         error = VOP_OPEN(volume->devvp,
 148                                          (ronly ? FREAD : FREAD|FWRITE),
 149                                          FSCRED, NULL);
 150                 }
 151                 vn_unlock(volume->devvp);
 152         }
 153         if (error) {
 154                 hammer_free_volume(volume);
 155                 return(error);
 156         }
 157         volume->devvp->v_rdev->si_mountpoint = mp;
 158         setmp = 1;
 159
 160         /*
 161          * Extract the volume number from the volume header and do various
 162          * sanity checks.
 163          */
 164         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 165         if (error)
 166                 goto late_failure;
 167         ondisk = (void *)bp->b_data;
 168         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 169                 kprintf("hammer_mount: volume %s has an invalid header\n",
 170                         volume->vol_name);
 171                 error = EFTYPE;
 172                 goto late_failure;
 173         }
 174         volume->vol_no = ondisk->vol_no;
 175         volume->buffer_base = ondisk->vol_buf_beg;
 176         volume->vol_flags = ondisk->vol_flags;
 177         volume->nblocks = ondisk->vol_nblocks;
 178         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 179                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 180         volume->maxraw_off = ondisk->vol_buf_end;
 181
 182         if (RB_EMPTY(&hmp->rb_vols_root)) {
 183                 hmp->fsid = ondisk->vol_fsid;
 184         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 185                 kprintf("hammer_mount: volume %s's fsid does not match "
 186                         "other volumes\n", volume->vol_name);
 187                 error = EFTYPE;
 188                 goto late_failure;
 189         }
 190
 191         /*
 192          * Insert the volume structure into the red-black tree.
 193          */
 194         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 195                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 196                         volume->vol_name, volume->vol_no);
 197                 error = EEXIST;
 198         }
 199
 200         /*
 201          * Set the root volume .  HAMMER special cases rootvol the structure.
 202          * We do not hold a ref because this would prevent related I/O
 203          * from being flushed.
 204          */
 205         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 206                 hmp->rootvol = volume;
 207                 if (bp) {
 208                         brelse(bp);
 209                         bp = NULL;
 210                 }
 211                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 212                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 213                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 214                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 215         }
 216 late_failure:
 217         if (bp)
 218                 brelse(bp);
 219         if (error) {
 220                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 221                 if (setmp)
 222                         volume->devvp->v_rdev->si_mountpoint = NULL;
 223                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 224                 hammer_free_volume(volume);
 225         }
 226         return (error);
 227 }
 228
 229 /*
 230  * This is called for each volume when updating the mount point from
 231  * read-write to read-only or vise-versa.
 232  */
 233 int
 234 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 235 {
 236         if (volume->devvp) {
 237                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 238                 if (volume->io.hmp->ronly) {
 239                         /* do not call vinvalbuf */
 240                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 241                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 242                 } else {
 243                         /* do not call vinvalbuf */
 244                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 245                         VOP_CLOSE(volume->devvp, FREAD);
 246                 }
 247                 vn_unlock(volume->devvp);
 248         }
 249         return(0);
 250 }
 251
 252 /*
 253  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 254  * so returns -1 on failure.
 255  */
 256 int
 257 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 258 {
 259         struct hammer_mount *hmp = volume->io.hmp;
 260         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 261         struct buf *bp;
 262
 263         /*
 264          * Clean up the root volume pointer, which is held unlocked in hmp.
 265          */
 266         if (hmp->rootvol == volume)
 267                 hmp->rootvol = NULL;
 268
 269         /*
 270          * Release our buffer and flush anything left in the buffer cache.
 271          */
 272         volume->io.waitdep = 1;
 273         bp = hammer_io_release(&volume->io, 1);
 274         hammer_io_clear_modlist(&volume->io);
 275
 276         /*
 277          * There should be no references on the volume, no clusters, and
 278          * no super-clusters.
 279          */
 280         KKASSERT(volume->io.lock.refs == 0);
 281         if (bp)
 282                 brelse(bp);
 283
 284         volume->ondisk = NULL;
 285         if (volume->devvp) {
 286                 if (volume->devvp->v_rdev &&
 287                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 288                 ) {
 289                         volume->devvp->v_rdev->si_mountpoint = NULL;
 290                 }
 291                 if (ronly) {
 292                         vinvalbuf(volume->devvp, 0, 0, 0);
 293                         VOP_CLOSE(volume->devvp, FREAD);
 294                 } else {
 295                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 296                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 297                 }
 298         }
 299
 300         /*
 301          * Destroy the structure
 302          */
 303         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 304         hammer_free_volume(volume);
 305         return(0);
 306 }
 307
 308 static
 309 void
 310 hammer_free_volume(hammer_volume_t volume)
 311 {
 312         if (volume->vol_name) {
 313                 kfree(volume->vol_name, M_HAMMER);
 314                 volume->vol_name = NULL;
 315         }
 316         if (volume->devvp) {
 317                 vrele(volume->devvp);
 318                 volume->devvp = NULL;
 319         }
 320         --hammer_count_volumes;
 321         kfree(volume, M_HAMMER);
 322 }
 323
 324 /*
 325  * Get a HAMMER volume.  The volume must already exist.
 326  */
 327 hammer_volume_t
 328 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 329 {
 330         struct hammer_volume *volume;
 331
 332         /*
 333          * Locate the volume structure
 334          */
 335         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 336         if (volume == NULL) {
 337                 *errorp = ENOENT;
 338                 return(NULL);
 339         }
 340         hammer_ref(&volume->io.lock);
 341
 342         /*
 343          * Deal with on-disk info
 344          */
 345         if (volume->ondisk == NULL || volume->io.loading) {
 346                 *errorp = hammer_load_volume(volume);
 347                 if (*errorp) {
 348                         hammer_rel_volume(volume, 1);
 349                         volume = NULL;
 350                 }
 351         } else {
 352                 *errorp = 0;
 353         }
 354         return(volume);
 355 }
 356
 357 int
 358 hammer_ref_volume(hammer_volume_t volume)
 359 {
 360         int error;
 361
 362         hammer_ref(&volume->io.lock);
 363
 364         /*
 365          * Deal with on-disk info
 366          */
 367         if (volume->ondisk == NULL || volume->io.loading) {
 368                 error = hammer_load_volume(volume);
 369                 if (error)
 370                         hammer_rel_volume(volume, 1);
 371         } else {
 372                 error = 0;
 373         }
 374         return (error);
 375 }
 376
 377 hammer_volume_t
 378 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 379 {
 380         hammer_volume_t volume;
 381
 382         volume = hmp->rootvol;
 383         KKASSERT(volume != NULL);
 384         hammer_ref(&volume->io.lock);
 385
 386         /*
 387          * Deal with on-disk info
 388          */
 389         if (volume->ondisk == NULL || volume->io.loading) {
 390                 *errorp = hammer_load_volume(volume);
 391                 if (*errorp) {
 392                         hammer_rel_volume(volume, 1);
 393                         volume = NULL;
 394                 }
 395         } else {
 396                 *errorp = 0;
 397         }
 398         return (volume);
 399 }
 400
 401 /*
 402  * Load a volume's on-disk information.  The volume must be referenced and
 403  * not locked.  We temporarily acquire an exclusive lock to interlock
 404  * against releases or multiple get's.
 405  */
 406 static int
 407 hammer_load_volume(hammer_volume_t volume)
 408 {
 409         int error;
 410
 411         ++volume->io.loading;
 412         hammer_lock_ex(&volume->io.lock);
 413
 414         if (volume->ondisk == NULL) {
 415                 error = hammer_io_read(volume->devvp, &volume->io,
 416                                        volume->maxraw_off);
 417                 if (error == 0)
 418                         volume->ondisk = (void *)volume->io.bp->b_data;
 419         } else {
 420                 error = 0;
 421         }
 422         --volume->io.loading;
 423         hammer_unlock(&volume->io.lock);
 424         return(error);
 425 }
 426
 427 /*
 428  * Release a volume.  Call hammer_io_release on the last reference.  We have
 429  * to acquire an exclusive lock to interlock against volume->ondisk tests
 430  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 431  * lock to be held.
 432  *
 433  * Volumes are not unloaded from memory during normal operation.
 434  */
 435 void
 436 hammer_rel_volume(hammer_volume_t volume, int flush)
 437 {
 438         struct buf *bp = NULL;
 439
 440         crit_enter();
 441         if (volume->io.lock.refs == 1) {
 442                 ++volume->io.loading;
 443                 hammer_lock_ex(&volume->io.lock);
 444                 if (volume->io.lock.refs == 1) {
 445                         volume->ondisk = NULL;
 446                         bp = hammer_io_release(&volume->io, flush);
 447                 }
 448                 --volume->io.loading;
 449                 hammer_unlock(&volume->io.lock);
 450         }
 451         hammer_unref(&volume->io.lock);
 452         if (bp)
 453                 brelse(bp);
 454         crit_exit();
 455 }
 456
 457 /************************************************************************
 458  *                              BUFFERS                                 *
 459  ************************************************************************
 460  *
 461  * Manage buffers.  Currently all blockmap-backed zones are translated
 462  * to zone-2 buffer offsets.
 463  */
 464 hammer_buffer_t
 465 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 466                   int bytes, int isnew, int *errorp)
 467 {
 468         hammer_buffer_t buffer;
 469         hammer_volume_t volume;
 470         hammer_off_t    zone2_offset;
 471         hammer_io_type_t iotype;
 472         int vol_no;
 473         int zone;
 474
 475         buf_offset &= ~HAMMER_BUFMASK64;
 476 again:
 477         /*
 478          * Shortcut if the buffer is already cached
 479          */
 480         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 481         if (buffer) {
 482                 if (buffer->io.lock.refs == 0)
 483                         ++hammer_count_refedbufs;
 484                 hammer_ref(&buffer->io.lock);
 485
 486                 /*
 487                  * Onced refed the ondisk field will not be cleared by
 488                  * any other action.
 489                  */
 490                 if (buffer->ondisk && buffer->io.loading == 0) {
 491                         *errorp = 0;
 492                         return(buffer);
 493                 }
 494
 495                 /*
 496                  * The buffer is no longer loose if it has a ref, and
 497                  * cannot become loose once it gains a ref.  Loose
 498                  * buffers will never be in a modified state.  This should
 499                  * only occur on the 0->1 transition of refs.
 500                  *
 501                  * lose_list can be modified via a biodone() interrupt.
 502                  */
 503                 if (buffer->io.mod_list == &hmp->lose_list) {
 504                         crit_enter();   /* biodone race against list */
 505                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 506                                      mod_entry);
 507                         crit_exit();
 508                         buffer->io.mod_list = NULL;
 509                         KKASSERT(buffer->io.modified == 0);
 510                 }
 511                 goto found;
 512         }
 513
 514         /*
 515          * What is the buffer class?
 516          */
 517         zone = HAMMER_ZONE_DECODE(buf_offset);
 518
 519         switch(zone) {
 520         case HAMMER_ZONE_LARGE_DATA_INDEX:
 521         case HAMMER_ZONE_SMALL_DATA_INDEX:
 522                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 523                 break;
 524         case HAMMER_ZONE_UNDO_INDEX:
 525                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 526                 break;
 527         case HAMMER_ZONE_META_INDEX:
 528         default:
 529                 /*
 530                  * NOTE: inode data and directory entries are placed in this
 531                  * zone.  inode atime/mtime is updated in-place and thus
 532                  * buffers containing inodes must be synchronized as
 533                  * meta-buffers, same as buffers containing B-Tree info.
 534                  */
 535                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 536                 break;
 537         }
 538
 539         /*
 540          * Handle blockmap offset translations
 541          */
 542         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 543                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 544         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 545                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 546         } else {
 547                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 548                 zone2_offset = buf_offset;
 549                 *errorp = 0;
 550         }
 551         if (*errorp)
 552                 return(NULL);
 553
 554         /*
 555          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 556          * specifications.
 557          */
 558         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 559                  HAMMER_ZONE_RAW_BUFFER);
 560         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 561         volume = hammer_get_volume(hmp, vol_no, errorp);
 562         if (volume == NULL)
 563                 return(NULL);
 564
 565         KKASSERT(zone2_offset < volume->maxbuf_off);
 566
 567         /*
 568          * Allocate a new buffer structure.  We will check for races later.
 569          */
 570         ++hammer_count_buffers;
 571         buffer = kmalloc(sizeof(*buffer), M_HAMMER,
 572                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 573         buffer->zone2_offset = zone2_offset;
 574         buffer->zoneX_offset = buf_offset;
 575         buffer->volume = volume;
 576
 577         hammer_io_init(&buffer->io, hmp, iotype);
 578         buffer->io.offset = volume->ondisk->vol_buf_beg +
 579                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 580         buffer->io.bytes = bytes;
 581         TAILQ_INIT(&buffer->clist);
 582         hammer_ref(&buffer->io.lock);
 583
 584         /*
 585          * Insert the buffer into the RB tree and handle late collisions.
 586          */
 587         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 588                 hammer_unref(&buffer->io.lock);
 589                 --hammer_count_buffers;
 590                 kfree(buffer, M_HAMMER);
 591                 goto again;
 592         }
 593         ++hammer_count_refedbufs;
 594 found:
 595
 596         /*
 597          * Deal with on-disk info and loading races.
 598          */
 599         if (buffer->ondisk == NULL || buffer->io.loading) {
 600                 *errorp = hammer_load_buffer(buffer, isnew);
 601                 if (*errorp) {
 602                         hammer_rel_buffer(buffer, 1);
 603                         buffer = NULL;
 604                 }
 605         } else {
 606                 *errorp = 0;
 607         }
 608         return(buffer);
 609 }
 610
 611 /*
 612  * Destroy all buffers covering the specified zoneX offset range.  This
 613  * is called when the related blockmap layer2 entry is freed or when
 614  * a direct write bypasses our buffer/buffer-cache subsystem.
 615  *
 616  * The buffers may be referenced by the caller itself.  Setting reclaim
 617  * will cause the buffer to be destroyed when it's ref count reaches zero.
 618  */
 619 void
 620 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 621                    hammer_off_t zone2_offset, int bytes)
 622 {
 623         hammer_buffer_t buffer;
 624         hammer_volume_t volume;
 625         int vol_no;
 626         int error;
 627
 628         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 629         volume = hammer_get_volume(hmp, vol_no, &error);
 630         KKASSERT(error == 0);
 631
 632         while (bytes > 0) {
 633                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 634                                    base_offset);
 635                 if (buffer) {
 636                         KKASSERT(buffer->zone2_offset == zone2_offset);
 637                         hammer_io_clear_modify(&buffer->io, 1);
 638                         buffer->io.reclaim = 1;
 639                         KKASSERT(buffer->volume == volume);
 640                         if (buffer->io.lock.refs == 0)
 641                                 hammer_unload_buffer(buffer, NULL);
 642                 } else {
 643                         hammer_io_inval(volume, zone2_offset);
 644                 }
 645                 base_offset += HAMMER_BUFSIZE;
 646                 zone2_offset += HAMMER_BUFSIZE;
 647                 bytes -= HAMMER_BUFSIZE;
 648         }
 649         hammer_rel_volume(volume, 0);
 650 }
 651
 652 static int
 653 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 654 {
 655         hammer_volume_t volume;
 656         int error;
 657
 658         /*
 659          * Load the buffer's on-disk info
 660          */
 661         volume = buffer->volume;
 662         ++buffer->io.loading;
 663         hammer_lock_ex(&buffer->io.lock);
 664
 665         if (hammer_debug_io & 0x0001) {
 666                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 667                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 668                         buffer->ondisk);
 669         }
 670
 671         if (buffer->ondisk == NULL) {
 672                 if (isnew) {
 673                         error = hammer_io_new(volume->devvp, &buffer->io);
 674                 } else {
 675                         error = hammer_io_read(volume->devvp, &buffer->io,
 676                                                volume->maxraw_off);
 677                 }
 678                 if (error == 0)
 679                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 680         } else if (isnew) {
 681                 error = hammer_io_new(volume->devvp, &buffer->io);
 682         } else {
 683                 error = 0;
 684         }
 685         --buffer->io.loading;
 686         hammer_unlock(&buffer->io.lock);
 687         return (error);
 688 }
 689
 690 /*
 691  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 692  */
 693 int
 694 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 695 {
 696         ++hammer_count_refedbufs;
 697         hammer_ref(&buffer->io.lock);
 698         hammer_flush_buffer_nodes(buffer);
 699         KKASSERT(buffer->io.lock.refs == 1);
 700         hammer_rel_buffer(buffer, 2);
 701         return(0);
 702 }
 703
 704 /*
 705  * Reference a buffer that is either already referenced or via a specially
 706  * handled pointer (aka cursor->buffer).
 707  */
 708 int
 709 hammer_ref_buffer(hammer_buffer_t buffer)
 710 {
 711         int error;
 712
 713         if (buffer->io.lock.refs == 0)
 714                 ++hammer_count_refedbufs;
 715         hammer_ref(&buffer->io.lock);
 716
 717         /*
 718          * At this point a biodone() will not touch the buffer other then
 719          * incidental bits.  However, lose_list can be modified via
 720          * a biodone() interrupt.
 721          *
 722          * No longer loose
 723          */
 724         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 725                 crit_enter();
 726                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 727                 buffer->io.mod_list = NULL;
 728                 crit_exit();
 729         }
 730
 731         if (buffer->ondisk == NULL || buffer->io.loading) {
 732                 error = hammer_load_buffer(buffer, 0);
 733                 if (error) {
 734                         hammer_rel_buffer(buffer, 1);
 735                         /*
 736                          * NOTE: buffer pointer can become stale after
 737                          * the above release.
 738                          */
 739                 }
 740         } else {
 741                 error = 0;
 742         }
 743         return(error);
 744 }
 745
 746 /*
 747  * Release a buffer.  We have to deal with several places where
 748  * another thread can ref the buffer.
 749  *
 750  * Only destroy the structure itself if the related buffer cache buffer
 751  * was disassociated from it.  This ties the management of the structure
 752  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 753  * embedded io is referenced or not.
 754  */
 755 void
 756 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 757 {
 758         hammer_volume_t volume;
 759         struct buf *bp = NULL;
 760         int freeme = 0;
 761
 762         crit_enter();
 763         if (buffer->io.lock.refs == 1) {
 764                 ++buffer->io.loading;   /* force interlock check */
 765                 hammer_lock_ex(&buffer->io.lock);
 766                 if (buffer->io.lock.refs == 1) {
 767                         bp = hammer_io_release(&buffer->io, flush);
 768
 769                         if (buffer->io.lock.refs == 1)
 770                                 --hammer_count_refedbufs;
 771
 772                         if (buffer->io.bp == NULL &&
 773                             buffer->io.lock.refs == 1) {
 774                                 /*
 775                                  * Final cleanup
 776                                  *
 777                                  * NOTE: It is impossible for any associated
 778                                  * B-Tree nodes to have refs if the buffer
 779                                  * has no additional refs.
 780                                  */
 781                                 RB_REMOVE(hammer_buf_rb_tree,
 782                                           &buffer->io.hmp->rb_bufs_root,
 783                                           buffer);
 784                                 volume = buffer->volume;
 785                                 buffer->volume = NULL; /* sanity */
 786                                 hammer_rel_volume(volume, 0);
 787                                 hammer_io_clear_modlist(&buffer->io);
 788                                 hammer_flush_buffer_nodes(buffer);
 789                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 790                                 freeme = 1;
 791                         }
 792                 }
 793                 --buffer->io.loading;
 794                 hammer_unlock(&buffer->io.lock);
 795         }
 796         hammer_unref(&buffer->io.lock);
 797         crit_exit();
 798         if (bp)
 799                 brelse(bp);
 800         if (freeme) {
 801                 --hammer_count_buffers;
 802                 kfree(buffer, M_HAMMER);
 803         }
 804 }
 805
 806 /*
 807  * Access the filesystem buffer containing the specified hammer offset.
 808  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 809  * relative buffer offset.  It must also have bit 55 set to be valid.
 810  * (see hammer_off_t in hammer_disk.h).
 811  *
 812  * Any prior buffer in *bufferp will be released and replaced by the
 813  * requested buffer.
 814  */
 815 static __inline
 816 void *
 817 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 818              int *errorp, struct hammer_buffer **bufferp)
 819 {
 820         hammer_buffer_t buffer;
 821         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 822
 823         buf_offset &= ~HAMMER_BUFMASK64;
 824         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 825
 826         buffer = *bufferp;
 827         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 828                                buffer->zoneX_offset != buf_offset)) {
 829                 if (buffer)
 830                         hammer_rel_buffer(buffer, 0);
 831                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 832                 *bufferp = buffer;
 833         } else {
 834                 *errorp = 0;
 835         }
 836
 837         /*
 838          * Return a pointer to the buffer data.
 839          */
 840         if (buffer == NULL)
 841                 return(NULL);
 842         else
 843                 return((char *)buffer->ondisk + xoff);
 844 }
 845
 846 void *
 847 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 848              int *errorp, struct hammer_buffer **bufferp)
 849 {
 850         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 851 }
 852
 853 void *
 854 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 855                  int *errorp, struct hammer_buffer **bufferp)
 856 {
 857         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 858         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 859 }
 860
 861 /*
 862  * Access the filesystem buffer containing the specified hammer offset.
 863  * No disk read operation occurs.  The result buffer may contain garbage.
 864  *
 865  * Any prior buffer in *bufferp will be released and replaced by the
 866  * requested buffer.
 867  *
 868  * This function marks the buffer dirty but does not increment its
 869  * modify_refs count.
 870  */
 871 static __inline
 872 void *
 873 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 874              int *errorp, struct hammer_buffer **bufferp)
 875 {
 876         hammer_buffer_t buffer;
 877         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 878
 879         buf_offset &= ~HAMMER_BUFMASK64;
 880
 881         buffer = *bufferp;
 882         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 883                                buffer->zoneX_offset != buf_offset)) {
 884                 if (buffer)
 885                         hammer_rel_buffer(buffer, 0);
 886                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 887                 *bufferp = buffer;
 888         } else {
 889                 *errorp = 0;
 890         }
 891
 892         /*
 893          * Return a pointer to the buffer data.
 894          */
 895         if (buffer == NULL)
 896                 return(NULL);
 897         else
 898                 return((char *)buffer->ondisk + xoff);
 899 }
 900
 901 void *
 902 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 903              int *errorp, struct hammer_buffer **bufferp)
 904 {
 905         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 906 }
 907
 908 void *
 909 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 910                 int *errorp, struct hammer_buffer **bufferp)
 911 {
 912         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 913         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
 914 }
 915
 916 /************************************************************************
 917  *                              NODES                                   *
 918  ************************************************************************
 919  *
 920  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 921  * method used by the HAMMER filesystem.
 922  *
 923  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 924  * associated with its buffer, and will only referenced the buffer while
 925  * the node itself is referenced.
 926  *
 927  * A hammer_node can also be passively associated with other HAMMER
 928  * structures, such as inodes, while retaining 0 references.  These
 929  * associations can be cleared backwards using a pointer-to-pointer in
 930  * the hammer_node.
 931  *
 932  * This allows the HAMMER implementation to cache hammer_nodes long-term
 933  * and short-cut a great deal of the infrastructure's complexity.  In
 934  * most cases a cached node can be reacquired without having to dip into
 935  * either the buffer or cluster management code.
 936  *
 937  * The caller must pass a referenced cluster on call and will retain
 938  * ownership of the reference on return.  The node will acquire its own
 939  * additional references, if necessary.
 940  */
 941 hammer_node_t
 942 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 943                 int isnew, int *errorp)
 944 {
 945         hammer_node_t node;
 946
 947         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 948
 949         /*
 950          * Locate the structure, allocating one if necessary.
 951          */
 952 again:
 953         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 954         if (node == NULL) {
 955                 ++hammer_count_nodes;
 956                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
 957                 node->node_offset = node_offset;
 958                 node->hmp = hmp;
 959                 TAILQ_INIT(&node->cursor_list);
 960                 TAILQ_INIT(&node->cache_list);
 961                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 962                         --hammer_count_nodes;
 963                         kfree(node, M_HAMMER);
 964                         goto again;
 965                 }
 966         }
 967         hammer_ref(&node->lock);
 968         if (node->ondisk)
 969                 *errorp = 0;
 970         else
 971                 *errorp = hammer_load_node(node, isnew);
 972         if (*errorp) {
 973                 hammer_rel_node(node);
 974                 node = NULL;
 975         }
 976         return(node);
 977 }
 978
 979 /*
 980  * Reference an already-referenced node.
 981  */
 982 void
 983 hammer_ref_node(hammer_node_t node)
 984 {
 985         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 986         hammer_ref(&node->lock);
 987 }
 988
 989 /*
 990  * Load a node's on-disk data reference.
 991  */
 992 static int
 993 hammer_load_node(hammer_node_t node, int isnew)
 994 {
 995         hammer_buffer_t buffer;
 996         hammer_off_t buf_offset;
 997         int error;
 998
 999         error = 0;
1000         ++node->loading;
1001         hammer_lock_ex(&node->lock);
1002         if (node->ondisk == NULL) {
1003                 /*
1004                  * This is a little confusing but the jist is that
1005                  * node->buffer determines whether the node is on
1006                  * the buffer's clist and node->ondisk determines
1007                  * whether the buffer is referenced.
1008                  *
1009                  * We could be racing a buffer release, in which case
1010                  * node->buffer may become NULL while we are blocked
1011                  * referencing the buffer.
1012                  */
1013                 if ((buffer = node->buffer) != NULL) {
1014                         error = hammer_ref_buffer(buffer);
1015                         if (error == 0 && node->buffer == NULL) {
1016                                 TAILQ_INSERT_TAIL(&buffer->clist,
1017                                                   node, entry);
1018                                 node->buffer = buffer;
1019                         }
1020                 } else {
1021                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1022                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1023                                                    HAMMER_BUFSIZE, 0, &error);
1024                         if (buffer) {
1025                                 KKASSERT(error == 0);
1026                                 TAILQ_INSERT_TAIL(&buffer->clist,
1027                                                   node, entry);
1028                                 node->buffer = buffer;
1029                         }
1030                 }
1031                 if (error)
1032                         goto failed;
1033                 node->ondisk = (void *)((char *)buffer->ondisk +
1034                                         (node->node_offset & HAMMER_BUFMASK));
1035                 if (isnew == 0 &&
1036                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1037                         if (hammer_crc_test_btree(node->ondisk) == 0)
1038                                 Debugger("CRC FAILED: B-TREE NODE");
1039                         node->flags |= HAMMER_NODE_CRCGOOD;
1040                 }
1041         }
1042 failed:
1043         --node->loading;
1044         hammer_unlock(&node->lock);
1045         return (error);
1046 }
1047
1048 /*
1049  * Safely reference a node, interlock against flushes via the IO subsystem.
1050  */
1051 hammer_node_t
1052 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1053                      int *errorp)
1054 {
1055         hammer_node_t node;
1056
1057         node = cache->node;
1058         if (node != NULL) {
1059                 hammer_ref(&node->lock);
1060                 if (node->ondisk)
1061                         *errorp = 0;
1062                 else
1063                         *errorp = hammer_load_node(node, 0);
1064                 if (*errorp) {
1065                         hammer_rel_node(node);
1066                         node = NULL;
1067                 }
1068         } else {
1069                 *errorp = ENOENT;
1070         }
1071         return(node);
1072 }
1073
1074 /*
1075  * Release a hammer_node.  On the last release the node dereferences
1076  * its underlying buffer and may or may not be destroyed.
1077  */
1078 void
1079 hammer_rel_node(hammer_node_t node)
1080 {
1081         hammer_buffer_t buffer;
1082
1083         /*
1084          * If this isn't the last ref just decrement the ref count and
1085          * return.
1086          */
1087         if (node->lock.refs > 1) {
1088                 hammer_unref(&node->lock);
1089                 return;
1090         }
1091
1092         /*
1093          * If there is no ondisk info or no buffer the node failed to load,
1094          * remove the last reference and destroy the node.
1095          */
1096         if (node->ondisk == NULL) {
1097                 hammer_unref(&node->lock);
1098                 hammer_flush_node(node);
1099                 /* node is stale now */
1100                 return;
1101         }
1102
1103         /*
1104          * Do not disassociate the node from the buffer if it represents
1105          * a modified B-Tree node that still needs its crc to be generated.
1106          */
1107         if (node->flags & HAMMER_NODE_NEEDSCRC)
1108                 return;
1109
1110         /*
1111          * Do final cleanups and then either destroy the node and leave it
1112          * passively cached.  The buffer reference is removed regardless.
1113          */
1114         buffer = node->buffer;
1115         node->ondisk = NULL;
1116
1117         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1118                 hammer_unref(&node->lock);
1119                 hammer_rel_buffer(buffer, 0);
1120                 return;
1121         }
1122
1123         /*
1124          * Destroy the node.
1125          */
1126         hammer_unref(&node->lock);
1127         hammer_flush_node(node);
1128         /* node is stale */
1129         hammer_rel_buffer(buffer, 0);
1130 }
1131
1132 /*
1133  * Free space on-media associated with a B-Tree node.
1134  */
1135 void
1136 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1137 {
1138         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1139         node->flags |= HAMMER_NODE_DELETED;
1140         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1141 }
1142
1143 /*
1144  * Passively cache a referenced hammer_node.  The caller may release
1145  * the node on return.
1146  */
1147 void
1148 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1149 {
1150         /*
1151          * If the node is being deleted, don't cache it!
1152          */
1153         if (node->flags & HAMMER_NODE_DELETED)
1154                 return;
1155         if (cache->node == node)
1156                 return;
1157         while (cache->node)
1158                 hammer_uncache_node(cache);
1159         if (node->flags & HAMMER_NODE_DELETED)
1160                 return;
1161         cache->node = node;
1162         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1163 }
1164
1165 void
1166 hammer_uncache_node(hammer_node_cache_t cache)
1167 {
1168         hammer_node_t node;
1169
1170         if ((node = cache->node) != NULL) {
1171                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1172                 cache->node = NULL;
1173                 if (TAILQ_EMPTY(&node->cache_list))
1174                         hammer_flush_node(node);
1175         }
1176 }
1177
1178 /*
1179  * Remove a node's cache references and destroy the node if it has no
1180  * other references or backing store.
1181  */
1182 void
1183 hammer_flush_node(hammer_node_t node)
1184 {
1185         hammer_node_cache_t cache;
1186         hammer_buffer_t buffer;
1187
1188         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1189                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1190                 cache->node = NULL;
1191         }
1192         if (node->lock.refs == 0 && node->ondisk == NULL) {
1193                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1194                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1195                 if ((buffer = node->buffer) != NULL) {
1196                         node->buffer = NULL;
1197                         TAILQ_REMOVE(&buffer->clist, node, entry);
1198                         /* buffer is unreferenced because ondisk is NULL */
1199                 }
1200                 --hammer_count_nodes;
1201                 kfree(node, M_HAMMER);
1202         }
1203 }
1204
1205 /*
1206  * Flush passively cached B-Tree nodes associated with this buffer.
1207  * This is only called when the buffer is about to be destroyed, so
1208  * none of the nodes should have any references.  The buffer is locked.
1209  *
1210  * We may be interlocked with the buffer.
1211  */
1212 void
1213 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1214 {
1215         hammer_node_t node;
1216
1217         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1218                 KKASSERT(node->ondisk == NULL);
1219                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1220
1221                 if (node->lock.refs == 0) {
1222                         hammer_ref(&node->lock);
1223                         node->flags |= HAMMER_NODE_FLUSH;
1224                         hammer_rel_node(node);
1225                 } else {
1226                         KKASSERT(node->loading != 0);
1227                         KKASSERT(node->buffer != NULL);
1228                         buffer = node->buffer;
1229                         node->buffer = NULL;
1230                         TAILQ_REMOVE(&buffer->clist, node, entry);
1231                         /* buffer is unreferenced because ondisk is NULL */
1232                 }
1233         }
1234 }
1235
1236
1237 /************************************************************************
1238  *                              ALLOCATORS                              *
1239  ************************************************************************/
1240
1241 /*
1242  * Allocate a B-Tree node.
1243  */
1244 hammer_node_t
1245 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1246 {
1247         hammer_buffer_t buffer = NULL;
1248         hammer_node_t node = NULL;
1249         hammer_off_t node_offset;
1250
1251         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1252                                             sizeof(struct hammer_node_ondisk),
1253                                             errorp);
1254         if (*errorp == 0) {
1255                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1256                 hammer_modify_node_noundo(trans, node);
1257                 bzero(node->ondisk, sizeof(*node->ondisk));
1258                 hammer_modify_node_done(node);
1259         }
1260         if (buffer)
1261                 hammer_rel_buffer(buffer, 0);
1262         return(node);
1263 }
1264
1265 /*
1266  * Allocate data.  If the address of a data buffer is supplied then
1267  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1268  * will be set to the related buffer.  The caller must release it when
1269  * finally done.  The initial *data_bufferp should be set to NULL by
1270  * the caller.
1271  *
1272  * The caller is responsible for making hammer_modify*() calls on the
1273  * *data_bufferp.
1274  */
1275 void *
1276 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1277                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1278                   struct hammer_buffer **data_bufferp, int *errorp)
1279 {
1280         void *data;
1281         int zone;
1282
1283         /*
1284          * Allocate data
1285          */
1286         if (data_len) {
1287                 switch(rec_type) {
1288                 case HAMMER_RECTYPE_INODE:
1289                 case HAMMER_RECTYPE_DIRENTRY:
1290                 case HAMMER_RECTYPE_EXT:
1291                 case HAMMER_RECTYPE_FIX:
1292                 case HAMMER_RECTYPE_PFS:
1293                         zone = HAMMER_ZONE_META_INDEX;
1294                         break;
1295                 case HAMMER_RECTYPE_DATA:
1296                 case HAMMER_RECTYPE_DB:
1297                         if (data_len <= HAMMER_BUFSIZE / 2) {
1298                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1299                         } else {
1300                                 data_len = (data_len + HAMMER_BUFMASK) &
1301                                            ~HAMMER_BUFMASK;
1302                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1303                         }
1304                         break;
1305                 default:
1306                         panic("hammer_alloc_data: rec_type %04x unknown",
1307                               rec_type);
1308                         zone = 0;       /* NOT REACHED */
1309                         break;
1310                 }
1311                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1312                                                       data_len, errorp);
1313         } else {
1314                 *data_offsetp = 0;
1315         }
1316         if (*errorp == 0 && data_bufferp) {
1317                 if (data_len) {
1318                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1319                                                 data_len, errorp, data_bufferp);
1320                         KKASSERT(*errorp == 0);
1321                 } else {
1322                         data = NULL;
1323                 }
1324         } else {
1325                 data = NULL;
1326         }
1327         KKASSERT(*errorp == 0);
1328         return(data);
1329 }
1330
1331 /*
1332  * Sync dirty buffers to the media and clean-up any loose ends.
1333  */
1334 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1335 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1336
1337 int
1338 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1339 {
1340         struct hammer_sync_info info;
1341
1342         info.error = 0;
1343         info.waitfor = waitfor;
1344         if (waitfor == MNT_WAIT) {
1345                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1346                               hammer_sync_scan1, hammer_sync_scan2, &info);
1347         } else {
1348                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1349                               hammer_sync_scan1, hammer_sync_scan2, &info);
1350         }
1351         return(info.error);
1352 }
1353
1354 /*
1355  * Filesystem sync.  If doing a synchronous sync make a second pass on
1356  * the vnodes in case any were already flushing during the first pass,
1357  * and activate the flusher twice (the second time brings the UNDO FIFO's
1358  * start position up to the end position after the first call).
1359  */
1360 int
1361 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1362 {
1363         struct hammer_sync_info info;
1364
1365         info.error = 0;
1366         info.waitfor = MNT_NOWAIT;
1367         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1368                       hammer_sync_scan1, hammer_sync_scan2, &info);
1369         if (info.error == 0 && waitfor == MNT_WAIT) {
1370                 info.waitfor = waitfor;
1371                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1372                               hammer_sync_scan1, hammer_sync_scan2, &info);
1373         }
1374         if (waitfor == MNT_WAIT) {
1375                 hammer_flusher_sync(hmp);
1376                 hammer_flusher_sync(hmp);
1377         } else {
1378                 hammer_flusher_async(hmp, NULL);
1379         }
1380         return(info.error);
1381 }
1382
1383 static int
1384 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1385 {
1386         struct hammer_inode *ip;
1387
1388         ip = VTOI(vp);
1389         if (vp->v_type == VNON || ip == NULL ||
1390             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1391              RB_EMPTY(&vp->v_rbdirty_tree))) {
1392                 return(-1);
1393         }
1394         return(0);
1395 }
1396
1397 static int
1398 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1399 {
1400         struct hammer_sync_info *info = data;
1401         struct hammer_inode *ip;
1402         int error;
1403
1404         ip = VTOI(vp);
1405         if (vp->v_type == VNON || vp->v_type == VBAD ||
1406             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1407              RB_EMPTY(&vp->v_rbdirty_tree))) {
1408                 return(0);
1409         }
1410         error = VOP_FSYNC(vp, info->waitfor);
1411         if (error)
1412                 info->error = error;
1413         return(0);
1414 }
1415