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 1.63 2008/06/29 07:50:40 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->fsid_udev = dev2udev(vn_todev(volume->devvp));
 212                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 213                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 214                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 215                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 216         }
 217 late_failure:
 218         if (bp)
 219                 brelse(bp);
 220         if (error) {
 221                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 222                 if (setmp)
 223                         volume->devvp->v_rdev->si_mountpoint = NULL;
 224                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 225                 hammer_free_volume(volume);
 226         }
 227         return (error);
 228 }
 229
 230 /*
 231  * This is called for each volume when updating the mount point from
 232  * read-write to read-only or vise-versa.
 233  */
 234 int
 235 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 236 {
 237         if (volume->devvp) {
 238                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 239                 if (volume->io.hmp->ronly) {
 240                         /* do not call vinvalbuf */
 241                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 242                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 243                 } else {
 244                         /* do not call vinvalbuf */
 245                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 246                         VOP_CLOSE(volume->devvp, FREAD);
 247                 }
 248                 vn_unlock(volume->devvp);
 249         }
 250         return(0);
 251 }
 252
 253 /*
 254  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 255  * so returns -1 on failure.
 256  */
 257 int
 258 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 259 {
 260         struct hammer_mount *hmp = volume->io.hmp;
 261         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 262         struct buf *bp;
 263
 264         /*
 265          * Clean up the root volume pointer, which is held unlocked in hmp.
 266          */
 267         if (hmp->rootvol == volume)
 268                 hmp->rootvol = NULL;
 269
 270         /*
 271          * Release our buffer and flush anything left in the buffer cache.
 272          */
 273         volume->io.waitdep = 1;
 274         bp = hammer_io_release(&volume->io, 1);
 275         hammer_io_clear_modlist(&volume->io);
 276
 277         /*
 278          * There should be no references on the volume, no clusters, and
 279          * no super-clusters.
 280          */
 281         KKASSERT(volume->io.lock.refs == 0);
 282         if (bp)
 283                 brelse(bp);
 284
 285         volume->ondisk = NULL;
 286         if (volume->devvp) {
 287                 if (volume->devvp->v_rdev &&
 288                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 289                 ) {
 290                         volume->devvp->v_rdev->si_mountpoint = NULL;
 291                 }
 292                 if (ronly) {
 293                         vinvalbuf(volume->devvp, 0, 0, 0);
 294                         VOP_CLOSE(volume->devvp, FREAD);
 295                 } else {
 296                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 297                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 298                 }
 299         }
 300
 301         /*
 302          * Destroy the structure
 303          */
 304         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 305         hammer_free_volume(volume);
 306         return(0);
 307 }
 308
 309 static
 310 void
 311 hammer_free_volume(hammer_volume_t volume)
 312 {
 313         if (volume->vol_name) {
 314                 kfree(volume->vol_name, M_HAMMER);
 315                 volume->vol_name = NULL;
 316         }
 317         if (volume->devvp) {
 318                 vrele(volume->devvp);
 319                 volume->devvp = NULL;
 320         }
 321         --hammer_count_volumes;
 322         kfree(volume, M_HAMMER);
 323 }
 324
 325 /*
 326  * Get a HAMMER volume.  The volume must already exist.
 327  */
 328 hammer_volume_t
 329 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 330 {
 331         struct hammer_volume *volume;
 332
 333         /*
 334          * Locate the volume structure
 335          */
 336         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 337         if (volume == NULL) {
 338                 *errorp = ENOENT;
 339                 return(NULL);
 340         }
 341         hammer_ref(&volume->io.lock);
 342
 343         /*
 344          * Deal with on-disk info
 345          */
 346         if (volume->ondisk == NULL || volume->io.loading) {
 347                 *errorp = hammer_load_volume(volume);
 348                 if (*errorp) {
 349                         hammer_rel_volume(volume, 1);
 350                         volume = NULL;
 351                 }
 352         } else {
 353                 *errorp = 0;
 354         }
 355         return(volume);
 356 }
 357
 358 int
 359 hammer_ref_volume(hammer_volume_t volume)
 360 {
 361         int error;
 362
 363         hammer_ref(&volume->io.lock);
 364
 365         /*
 366          * Deal with on-disk info
 367          */
 368         if (volume->ondisk == NULL || volume->io.loading) {
 369                 error = hammer_load_volume(volume);
 370                 if (error)
 371                         hammer_rel_volume(volume, 1);
 372         } else {
 373                 error = 0;
 374         }
 375         return (error);
 376 }
 377
 378 hammer_volume_t
 379 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 380 {
 381         hammer_volume_t volume;
 382
 383         volume = hmp->rootvol;
 384         KKASSERT(volume != NULL);
 385         hammer_ref(&volume->io.lock);
 386
 387         /*
 388          * Deal with on-disk info
 389          */
 390         if (volume->ondisk == NULL || volume->io.loading) {
 391                 *errorp = hammer_load_volume(volume);
 392                 if (*errorp) {
 393                         hammer_rel_volume(volume, 1);
 394                         volume = NULL;
 395                 }
 396         } else {
 397                 *errorp = 0;
 398         }
 399         return (volume);
 400 }
 401
 402 /*
 403  * Load a volume's on-disk information.  The volume must be referenced and
 404  * not locked.  We temporarily acquire an exclusive lock to interlock
 405  * against releases or multiple get's.
 406  */
 407 static int
 408 hammer_load_volume(hammer_volume_t volume)
 409 {
 410         int error;
 411
 412         ++volume->io.loading;
 413         hammer_lock_ex(&volume->io.lock);
 414
 415         if (volume->ondisk == NULL) {
 416                 error = hammer_io_read(volume->devvp, &volume->io,
 417                                        volume->maxraw_off);
 418                 if (error == 0)
 419                         volume->ondisk = (void *)volume->io.bp->b_data;
 420         } else {
 421                 error = 0;
 422         }
 423         --volume->io.loading;
 424         hammer_unlock(&volume->io.lock);
 425         return(error);
 426 }
 427
 428 /*
 429  * Release a volume.  Call hammer_io_release on the last reference.  We have
 430  * to acquire an exclusive lock to interlock against volume->ondisk tests
 431  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 432  * lock to be held.
 433  *
 434  * Volumes are not unloaded from memory during normal operation.
 435  */
 436 void
 437 hammer_rel_volume(hammer_volume_t volume, int flush)
 438 {
 439         struct buf *bp = NULL;
 440
 441         crit_enter();
 442         if (volume->io.lock.refs == 1) {
 443                 ++volume->io.loading;
 444                 hammer_lock_ex(&volume->io.lock);
 445                 if (volume->io.lock.refs == 1) {
 446                         volume->ondisk = NULL;
 447                         bp = hammer_io_release(&volume->io, flush);
 448                 }
 449                 --volume->io.loading;
 450                 hammer_unlock(&volume->io.lock);
 451         }
 452         hammer_unref(&volume->io.lock);
 453         if (bp)
 454                 brelse(bp);
 455         crit_exit();
 456 }
 457
 458 /************************************************************************
 459  *                              BUFFERS                                 *
 460  ************************************************************************
 461  *
 462  * Manage buffers.  Currently all blockmap-backed zones are translated
 463  * to zone-2 buffer offsets.
 464  */
 465 hammer_buffer_t
 466 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 467                   int bytes, int isnew, int *errorp)
 468 {
 469         hammer_buffer_t buffer;
 470         hammer_volume_t volume;
 471         hammer_off_t    zone2_offset;
 472         hammer_io_type_t iotype;
 473         int vol_no;
 474         int zone;
 475
 476         buf_offset &= ~HAMMER_BUFMASK64;
 477 again:
 478         /*
 479          * Shortcut if the buffer is already cached
 480          */
 481         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 482         if (buffer) {
 483                 if (buffer->io.lock.refs == 0)
 484                         ++hammer_count_refedbufs;
 485                 hammer_ref(&buffer->io.lock);
 486
 487                 /*
 488                  * Onced refed the ondisk field will not be cleared by
 489                  * any other action.
 490                  */
 491                 if (buffer->ondisk && buffer->io.loading == 0) {
 492                         *errorp = 0;
 493                         return(buffer);
 494                 }
 495
 496                 /*
 497                  * The buffer is no longer loose if it has a ref, and
 498                  * cannot become loose once it gains a ref.  Loose
 499                  * buffers will never be in a modified state.  This should
 500                  * only occur on the 0->1 transition of refs.
 501                  *
 502                  * lose_list can be modified via a biodone() interrupt.
 503                  */
 504                 if (buffer->io.mod_list == &hmp->lose_list) {
 505                         crit_enter();   /* biodone race against list */
 506                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 507                                      mod_entry);
 508                         crit_exit();
 509                         buffer->io.mod_list = NULL;
 510                         KKASSERT(buffer->io.modified == 0);
 511                 }
 512                 goto found;
 513         }
 514
 515         /*
 516          * What is the buffer class?
 517          */
 518         zone = HAMMER_ZONE_DECODE(buf_offset);
 519
 520         switch(zone) {
 521         case HAMMER_ZONE_LARGE_DATA_INDEX:
 522         case HAMMER_ZONE_SMALL_DATA_INDEX:
 523                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 524                 break;
 525         case HAMMER_ZONE_UNDO_INDEX:
 526                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 527                 break;
 528         case HAMMER_ZONE_META_INDEX:
 529         default:
 530                 /*
 531                  * NOTE: inode data and directory entries are placed in this
 532                  * zone.  inode atime/mtime is updated in-place and thus
 533                  * buffers containing inodes must be synchronized as
 534                  * meta-buffers, same as buffers containing B-Tree info.
 535                  */
 536                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 537                 break;
 538         }
 539
 540         /*
 541          * Handle blockmap offset translations
 542          */
 543         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 544                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 545         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 546                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 547         } else {
 548                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 549                 zone2_offset = buf_offset;
 550                 *errorp = 0;
 551         }
 552         if (*errorp)
 553                 return(NULL);
 554
 555         /*
 556          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 557          * specifications.
 558          */
 559         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 560                  HAMMER_ZONE_RAW_BUFFER);
 561         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 562         volume = hammer_get_volume(hmp, vol_no, errorp);
 563         if (volume == NULL)
 564                 return(NULL);
 565
 566         KKASSERT(zone2_offset < volume->maxbuf_off);
 567
 568         /*
 569          * Allocate a new buffer structure.  We will check for races later.
 570          */
 571         ++hammer_count_buffers;
 572         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 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);
 957                 node->node_offset = node_offset;
 958                 node->hmp = hmp;
 959                 TAILQ_INIT(&node->cache_list);
 960                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 961                         --hammer_count_nodes;
 962                         kfree(node, M_HAMMER);
 963                         goto again;
 964                 }
 965         }
 966         hammer_ref(&node->lock);
 967         if (node->ondisk)
 968                 *errorp = 0;
 969         else
 970                 *errorp = hammer_load_node(node, isnew);
 971         if (*errorp) {
 972                 hammer_rel_node(node);
 973                 node = NULL;
 974         }
 975         return(node);
 976 }
 977
 978 /*
 979  * Reference an already-referenced node.
 980  */
 981 void
 982 hammer_ref_node(hammer_node_t node)
 983 {
 984         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 985         hammer_ref(&node->lock);
 986 }
 987
 988 /*
 989  * Load a node's on-disk data reference.
 990  */
 991 static int
 992 hammer_load_node(hammer_node_t node, int isnew)
 993 {
 994         hammer_buffer_t buffer;
 995         hammer_off_t buf_offset;
 996         int error;
 997
 998         error = 0;
 999         ++node->loading;
1000         hammer_lock_ex(&node->lock);
1001         if (node->ondisk == NULL) {
1002                 /*
1003                  * This is a little confusing but the jist is that
1004                  * node->buffer determines whether the node is on
1005                  * the buffer's clist and node->ondisk determines
1006                  * whether the buffer is referenced.
1007                  *
1008                  * We could be racing a buffer release, in which case
1009                  * node->buffer may become NULL while we are blocked
1010                  * referencing the buffer.
1011                  */
1012                 if ((buffer = node->buffer) != NULL) {
1013                         error = hammer_ref_buffer(buffer);
1014                         if (error == 0 && node->buffer == NULL) {
1015                                 TAILQ_INSERT_TAIL(&buffer->clist,
1016                                                   node, entry);
1017                                 node->buffer = buffer;
1018                         }
1019                 } else {
1020                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1021                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1022                                                    HAMMER_BUFSIZE, 0, &error);
1023                         if (buffer) {
1024                                 KKASSERT(error == 0);
1025                                 TAILQ_INSERT_TAIL(&buffer->clist,
1026                                                   node, entry);
1027                                 node->buffer = buffer;
1028                         }
1029                 }
1030                 if (error)
1031                         goto failed;
1032                 node->ondisk = (void *)((char *)buffer->ondisk +
1033                                         (node->node_offset & HAMMER_BUFMASK));
1034                 if (isnew == 0 &&
1035                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1036                         if (hammer_crc_test_btree(node->ondisk) == 0)
1037                                 Debugger("CRC FAILED: B-TREE NODE");
1038                         node->flags |= HAMMER_NODE_CRCGOOD;
1039                 }
1040         }
1041 failed:
1042         --node->loading;
1043         hammer_unlock(&node->lock);
1044         return (error);
1045 }
1046
1047 /*
1048  * Safely reference a node, interlock against flushes via the IO subsystem.
1049  */
1050 hammer_node_t
1051 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1052                      int *errorp)
1053 {
1054         hammer_node_t node;
1055
1056         node = cache->node;
1057         if (node != NULL) {
1058                 hammer_ref(&node->lock);
1059                 if (node->ondisk)
1060                         *errorp = 0;
1061                 else
1062                         *errorp = hammer_load_node(node, 0);
1063                 if (*errorp) {
1064                         hammer_rel_node(node);
1065                         node = NULL;
1066                 }
1067         } else {
1068                 *errorp = ENOENT;
1069         }
1070         return(node);
1071 }
1072
1073 /*
1074  * Release a hammer_node.  On the last release the node dereferences
1075  * its underlying buffer and may or may not be destroyed.
1076  */
1077 void
1078 hammer_rel_node(hammer_node_t node)
1079 {
1080         hammer_buffer_t buffer;
1081
1082         /*
1083          * If this isn't the last ref just decrement the ref count and
1084          * return.
1085          */
1086         if (node->lock.refs > 1) {
1087                 hammer_unref(&node->lock);
1088                 return;
1089         }
1090
1091         /*
1092          * If there is no ondisk info or no buffer the node failed to load,
1093          * remove the last reference and destroy the node.
1094          */
1095         if (node->ondisk == NULL) {
1096                 hammer_unref(&node->lock);
1097                 hammer_flush_node(node);
1098                 /* node is stale now */
1099                 return;
1100         }
1101
1102         /*
1103          * Do not disassociate the node from the buffer if it represents
1104          * a modified B-Tree node that still needs its crc to be generated.
1105          */
1106         if (node->flags & HAMMER_NODE_NEEDSCRC)
1107                 return;
1108
1109         /*
1110          * Do final cleanups and then either destroy the node and leave it
1111          * passively cached.  The buffer reference is removed regardless.
1112          */
1113         buffer = node->buffer;
1114         node->ondisk = NULL;
1115
1116         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1117                 hammer_unref(&node->lock);
1118                 hammer_rel_buffer(buffer, 0);
1119                 return;
1120         }
1121
1122         /*
1123          * Destroy the node.
1124          */
1125         hammer_unref(&node->lock);
1126         hammer_flush_node(node);
1127         /* node is stale */
1128         hammer_rel_buffer(buffer, 0);
1129 }
1130
1131 /*
1132  * Free space on-media associated with a B-Tree node.
1133  */
1134 void
1135 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1136 {
1137         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1138         node->flags |= HAMMER_NODE_DELETED;
1139         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1140 }
1141
1142 /*
1143  * Passively cache a referenced hammer_node.  The caller may release
1144  * the node on return.
1145  */
1146 void
1147 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1148 {
1149         /*
1150          * If the node is being deleted, don't cache it!
1151          */
1152         if (node->flags & HAMMER_NODE_DELETED)
1153                 return;
1154         if (cache->node == node)
1155                 return;
1156         while (cache->node)
1157                 hammer_uncache_node(cache);
1158         if (node->flags & HAMMER_NODE_DELETED)
1159                 return;
1160         cache->node = node;
1161         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1162 }
1163
1164 void
1165 hammer_uncache_node(hammer_node_cache_t cache)
1166 {
1167         hammer_node_t node;
1168
1169         if ((node = cache->node) != NULL) {
1170                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1171                 cache->node = NULL;
1172                 if (TAILQ_EMPTY(&node->cache_list))
1173                         hammer_flush_node(node);
1174         }
1175 }
1176
1177 /*
1178  * Remove a node's cache references and destroy the node if it has no
1179  * other references or backing store.
1180  */
1181 void
1182 hammer_flush_node(hammer_node_t node)
1183 {
1184         hammer_node_cache_t cache;
1185         hammer_buffer_t buffer;
1186
1187         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1188                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1189                 cache->node = NULL;
1190         }
1191         if (node->lock.refs == 0 && node->ondisk == NULL) {
1192                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1193                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1194                 if ((buffer = node->buffer) != NULL) {
1195                         node->buffer = NULL;
1196                         TAILQ_REMOVE(&buffer->clist, node, entry);
1197                         /* buffer is unreferenced because ondisk is NULL */
1198                 }
1199                 --hammer_count_nodes;
1200                 kfree(node, M_HAMMER);
1201         }
1202 }
1203
1204 /*
1205  * Flush passively cached B-Tree nodes associated with this buffer.
1206  * This is only called when the buffer is about to be destroyed, so
1207  * none of the nodes should have any references.  The buffer is locked.
1208  *
1209  * We may be interlocked with the buffer.
1210  */
1211 void
1212 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1213 {
1214         hammer_node_t node;
1215
1216         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1217                 KKASSERT(node->ondisk == NULL);
1218                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1219
1220                 if (node->lock.refs == 0) {
1221                         hammer_ref(&node->lock);
1222                         node->flags |= HAMMER_NODE_FLUSH;
1223                         hammer_rel_node(node);
1224                 } else {
1225                         KKASSERT(node->loading != 0);
1226                         KKASSERT(node->buffer != NULL);
1227                         buffer = node->buffer;
1228                         node->buffer = NULL;
1229                         TAILQ_REMOVE(&buffer->clist, node, entry);
1230                         /* buffer is unreferenced because ondisk is NULL */
1231                 }
1232         }
1233 }
1234
1235
1236 /************************************************************************
1237  *                              ALLOCATORS                              *
1238  ************************************************************************/
1239
1240 /*
1241  * Allocate a B-Tree node.
1242  */
1243 hammer_node_t
1244 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1245 {
1246         hammer_buffer_t buffer = NULL;
1247         hammer_node_t node = NULL;
1248         hammer_off_t node_offset;
1249
1250         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1251                                             sizeof(struct hammer_node_ondisk),
1252                                             errorp);
1253         if (*errorp == 0) {
1254                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1255                 hammer_modify_node_noundo(trans, node);
1256                 bzero(node->ondisk, sizeof(*node->ondisk));
1257                 hammer_modify_node_done(node);
1258         }
1259         if (buffer)
1260                 hammer_rel_buffer(buffer, 0);
1261         return(node);
1262 }
1263
1264 /*
1265  * Allocate data.  If the address of a data buffer is supplied then
1266  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1267  * will be set to the related buffer.  The caller must release it when
1268  * finally done.  The initial *data_bufferp should be set to NULL by
1269  * the caller.
1270  *
1271  * The caller is responsible for making hammer_modify*() calls on the
1272  * *data_bufferp.
1273  */
1274 void *
1275 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1276                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1277                   struct hammer_buffer **data_bufferp, int *errorp)
1278 {
1279         void *data;
1280         int zone;
1281
1282         /*
1283          * Allocate data
1284          */
1285         if (data_len) {
1286                 switch(rec_type) {
1287                 case HAMMER_RECTYPE_INODE:
1288                 case HAMMER_RECTYPE_DIRENTRY:
1289                 case HAMMER_RECTYPE_EXT:
1290                 case HAMMER_RECTYPE_FIX:
1291                         zone = HAMMER_ZONE_META_INDEX;
1292                         break;
1293                 case HAMMER_RECTYPE_DATA:
1294                 case HAMMER_RECTYPE_DB:
1295                         if (data_len <= HAMMER_BUFSIZE / 2) {
1296                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1297                         } else {
1298                                 data_len = (data_len + HAMMER_BUFMASK) &
1299                                            ~HAMMER_BUFMASK;
1300                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1301                         }
1302                         break;
1303                 default:
1304                         panic("hammer_alloc_data: rec_type %04x unknown",
1305                               rec_type);
1306                         zone = 0;       /* NOT REACHED */
1307                         break;
1308                 }
1309                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1310                                                       data_len, errorp);
1311         } else {
1312                 *data_offsetp = 0;
1313         }
1314         if (*errorp == 0 && data_bufferp) {
1315                 if (data_len) {
1316                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1317                                                 data_len, errorp, data_bufferp);
1318                         KKASSERT(*errorp == 0);
1319                 } else {
1320                         data = NULL;
1321                 }
1322         } else {
1323                 data = NULL;
1324         }
1325         KKASSERT(*errorp == 0);
1326         return(data);
1327 }
1328
1329 /*
1330  * Sync dirty buffers to the media and clean-up any loose ends.
1331  */
1332 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1333 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1334
1335 int
1336 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1337 {
1338         struct hammer_sync_info info;
1339
1340         info.error = 0;
1341         info.waitfor = waitfor;
1342         if (waitfor == MNT_WAIT) {
1343                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1344                               hammer_sync_scan1, hammer_sync_scan2, &info);
1345         } else {
1346                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1347                               hammer_sync_scan1, hammer_sync_scan2, &info);
1348         }
1349         return(info.error);
1350 }
1351
1352 /*
1353  * Filesystem sync.  If doing a synchronous sync make a second pass on
1354  * the vnodes in case any were already flushing during the first pass,
1355  * and activate the flusher twice (the second time brings the UNDO FIFO's
1356  * start position up to the end position after the first call).
1357  */
1358 int
1359 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1360 {
1361         struct hammer_sync_info info;
1362
1363         info.error = 0;
1364         info.waitfor = MNT_NOWAIT;
1365         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1366                       hammer_sync_scan1, hammer_sync_scan2, &info);
1367         if (info.error == 0 && waitfor == MNT_WAIT) {
1368                 info.waitfor = waitfor;
1369                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1370                               hammer_sync_scan1, hammer_sync_scan2, &info);
1371         }
1372         if (waitfor == MNT_WAIT) {
1373                 hammer_flusher_sync(hmp);
1374                 hammer_flusher_sync(hmp);
1375         } else {
1376                 hammer_flusher_async(hmp);
1377         }
1378         return(info.error);
1379 }
1380
1381 static int
1382 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1383 {
1384         struct hammer_inode *ip;
1385
1386         ip = VTOI(vp);
1387         if (vp->v_type == VNON || ip == NULL ||
1388             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1389              RB_EMPTY(&vp->v_rbdirty_tree))) {
1390                 return(-1);
1391         }
1392         return(0);
1393 }
1394
1395 static int
1396 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1397 {
1398         struct hammer_sync_info *info = data;
1399         struct hammer_inode *ip;
1400         int error;
1401
1402         ip = VTOI(vp);
1403         if (vp->v_type == VNON || vp->v_type == VBAD ||
1404             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1405              RB_EMPTY(&vp->v_rbdirty_tree))) {
1406                 return(0);
1407         }
1408         error = VOP_FSYNC(vp, info->waitfor);
1409         if (error)
1410                 info->error = error;
1411         return(0);
1412 }
1413