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 1.72 2008/07/27 21:34:04 mneumann 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                       struct vnode *devvp)
 103 {
 104         struct mount *mp;
 105         hammer_volume_t volume;
 106         struct hammer_volume_ondisk *ondisk;
 107         struct nlookupdata nd;
 108         struct buf *bp = NULL;
 109         int error;
 110         int ronly;
 111         int setmp = 0;
 112
 113         mp = hmp->mp;
 114         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 115
 116         /*
 117          * Allocate a volume structure
 118          */
 119         ++hammer_count_volumes;
 120         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 121         volume->vol_name = kstrdup(volname, M_HAMMER);
 122         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 123         volume->io.offset = 0LL;
 124         volume->io.bytes = HAMMER_BUFSIZE;
 125
 126         /*
 127          * Get the device vnode
 128          */
 129         if (devvp == NULL) {
 130                 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 131                 if (error == 0)
 132                         error = nlookup(&nd);
 133                 if (error == 0)
 134                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 135                 nlookup_done(&nd);
 136         } else {
 137                 error = 0;
 138                 volume->devvp = devvp;
 139         }
 140
 141         if (error == 0) {
 142                 if (vn_isdisk(volume->devvp, &error)) {
 143                         error = vfs_mountedon(volume->devvp);
 144                 }
 145         }
 146         if (error == 0 &&
 147             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 148                 error = EBUSY;
 149         }
 150         if (error == 0) {
 151                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 152                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 153                 if (error == 0) {
 154                         error = VOP_OPEN(volume->devvp,
 155                                          (ronly ? FREAD : FREAD|FWRITE),
 156                                          FSCRED, NULL);
 157                 }
 158                 vn_unlock(volume->devvp);
 159         }
 160         if (error) {
 161                 hammer_free_volume(volume);
 162                 return(error);
 163         }
 164         volume->devvp->v_rdev->si_mountpoint = mp;
 165         setmp = 1;
 166
 167         /*
 168          * Extract the volume number from the volume header and do various
 169          * sanity checks.
 170          */
 171         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 172         if (error)
 173                 goto late_failure;
 174         ondisk = (void *)bp->b_data;
 175         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 176                 kprintf("hammer_mount: volume %s has an invalid header\n",
 177                         volume->vol_name);
 178                 error = EFTYPE;
 179                 goto late_failure;
 180         }
 181         volume->vol_no = ondisk->vol_no;
 182         volume->buffer_base = ondisk->vol_buf_beg;
 183         volume->vol_flags = ondisk->vol_flags;
 184         volume->nblocks = ondisk->vol_nblocks;
 185         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 186                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 187         volume->maxraw_off = ondisk->vol_buf_end;
 188
 189         if (RB_EMPTY(&hmp->rb_vols_root)) {
 190                 hmp->fsid = ondisk->vol_fsid;
 191         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 192                 kprintf("hammer_mount: volume %s's fsid does not match "
 193                         "other volumes\n", volume->vol_name);
 194                 error = EFTYPE;
 195                 goto late_failure;
 196         }
 197
 198         /*
 199          * Insert the volume structure into the red-black tree.
 200          */
 201         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 202                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 203                         volume->vol_name, volume->vol_no);
 204                 error = EEXIST;
 205         }
 206
 207         /*
 208          * Set the root volume .  HAMMER special cases rootvol the structure.
 209          * We do not hold a ref because this would prevent related I/O
 210          * from being flushed.
 211          */
 212         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 213                 hmp->rootvol = volume;
 214                 hmp->nvolumes = ondisk->vol_count;
 215                 if (bp) {
 216                         brelse(bp);
 217                         bp = NULL;
 218                 }
 219                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 220                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 221                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 222                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 223         }
 224 late_failure:
 225         if (bp)
 226                 brelse(bp);
 227         if (error) {
 228                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 229                 if (setmp)
 230                         volume->devvp->v_rdev->si_mountpoint = NULL;
 231                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 232                 hammer_free_volume(volume);
 233         }
 234         return (error);
 235 }
 236
 237 /*
 238  * This is called for each volume when updating the mount point from
 239  * read-write to read-only or vise-versa.
 240  */
 241 int
 242 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 243 {
 244         if (volume->devvp) {
 245                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 246                 if (volume->io.hmp->ronly) {
 247                         /* do not call vinvalbuf */
 248                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 249                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 250                 } else {
 251                         /* do not call vinvalbuf */
 252                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 253                         VOP_CLOSE(volume->devvp, FREAD);
 254                 }
 255                 vn_unlock(volume->devvp);
 256         }
 257         return(0);
 258 }
 259
 260 /*
 261  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 262  * so returns -1 on failure.
 263  */
 264 int
 265 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 266 {
 267         struct hammer_mount *hmp = volume->io.hmp;
 268         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 269         struct buf *bp;
 270
 271         /*
 272          * Clean up the root volume pointer, which is held unlocked in hmp.
 273          */
 274         if (hmp->rootvol == volume)
 275                 hmp->rootvol = NULL;
 276
 277         /*
 278          * We must not flush a dirty buffer to disk on umount.  It should
 279          * have already been dealt with by the flusher, or we may be in
 280          * catastrophic failure.
 281          */
 282         hammer_io_clear_modify(&volume->io, 1);
 283         volume->io.waitdep = 1;
 284         bp = hammer_io_release(&volume->io, 1);
 285
 286         /*
 287          * Clean up the persistent ref ioerror might have on the volume
 288          */
 289         if (volume->io.ioerror) {
 290                 volume->io.ioerror = 0;
 291                 hammer_unref(&volume->io.lock);
 292         }
 293
 294         /*
 295          * There should be no references on the volume, no clusters, and
 296          * no super-clusters.
 297          */
 298         KKASSERT(volume->io.lock.refs == 0);
 299         if (bp)
 300                 brelse(bp);
 301
 302         volume->ondisk = NULL;
 303         if (volume->devvp) {
 304                 if (volume->devvp->v_rdev &&
 305                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 306                 ) {
 307                         volume->devvp->v_rdev->si_mountpoint = NULL;
 308                 }
 309                 if (ronly) {
 310                         /*
 311                          * Make sure we don't sync anything to disk if we
 312                          * are in read-only mode (1) or critically-errored
 313                          * (2).  Note that there may be dirty buffers in
 314                          * normal read-only mode from crash recovery.
 315                          */
 316                         vinvalbuf(volume->devvp, 0, 0, 0);
 317                         VOP_CLOSE(volume->devvp, FREAD);
 318                 } else {
 319                         /*
 320                          * Normal termination, save any dirty buffers
 321                          * (XXX there really shouldn't be any).
 322                          */
 323                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 324                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 325                 }
 326         }
 327
 328         /*
 329          * Destroy the structure
 330          */
 331         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 332         hammer_free_volume(volume);
 333         return(0);
 334 }
 335
 336 static
 337 void
 338 hammer_free_volume(hammer_volume_t volume)
 339 {
 340         if (volume->vol_name) {
 341                 kfree(volume->vol_name, M_HAMMER);
 342                 volume->vol_name = NULL;
 343         }
 344         if (volume->devvp) {
 345                 vrele(volume->devvp);
 346                 volume->devvp = NULL;
 347         }
 348         --hammer_count_volumes;
 349         kfree(volume, M_HAMMER);
 350 }
 351
 352 /*
 353  * Get a HAMMER volume.  The volume must already exist.
 354  */
 355 hammer_volume_t
 356 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 357 {
 358         struct hammer_volume *volume;
 359
 360         /*
 361          * Locate the volume structure
 362          */
 363         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 364         if (volume == NULL) {
 365                 *errorp = ENOENT;
 366                 return(NULL);
 367         }
 368         hammer_ref(&volume->io.lock);
 369
 370         /*
 371          * Deal with on-disk info
 372          */
 373         if (volume->ondisk == NULL || volume->io.loading) {
 374                 *errorp = hammer_load_volume(volume);
 375                 if (*errorp) {
 376                         hammer_rel_volume(volume, 1);
 377                         volume = NULL;
 378                 }
 379         } else {
 380                 *errorp = 0;
 381         }
 382         return(volume);
 383 }
 384
 385 int
 386 hammer_ref_volume(hammer_volume_t volume)
 387 {
 388         int error;
 389
 390         hammer_ref(&volume->io.lock);
 391
 392         /*
 393          * Deal with on-disk info
 394          */
 395         if (volume->ondisk == NULL || volume->io.loading) {
 396                 error = hammer_load_volume(volume);
 397                 if (error)
 398                         hammer_rel_volume(volume, 1);
 399         } else {
 400                 error = 0;
 401         }
 402         return (error);
 403 }
 404
 405 hammer_volume_t
 406 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 407 {
 408         hammer_volume_t volume;
 409
 410         volume = hmp->rootvol;
 411         KKASSERT(volume != NULL);
 412         hammer_ref(&volume->io.lock);
 413
 414         /*
 415          * Deal with on-disk info
 416          */
 417         if (volume->ondisk == NULL || volume->io.loading) {
 418                 *errorp = hammer_load_volume(volume);
 419                 if (*errorp) {
 420                         hammer_rel_volume(volume, 1);
 421                         volume = NULL;
 422                 }
 423         } else {
 424                 *errorp = 0;
 425         }
 426         return (volume);
 427 }
 428
 429 /*
 430  * Load a volume's on-disk information.  The volume must be referenced and
 431  * not locked.  We temporarily acquire an exclusive lock to interlock
 432  * against releases or multiple get's.
 433  */
 434 static int
 435 hammer_load_volume(hammer_volume_t volume)
 436 {
 437         int error;
 438
 439         ++volume->io.loading;
 440         hammer_lock_ex(&volume->io.lock);
 441
 442         if (volume->ondisk == NULL) {
 443                 error = hammer_io_read(volume->devvp, &volume->io,
 444                                        volume->maxraw_off);
 445                 if (error == 0)
 446                         volume->ondisk = (void *)volume->io.bp->b_data;
 447         } else {
 448                 error = 0;
 449         }
 450         --volume->io.loading;
 451         hammer_unlock(&volume->io.lock);
 452         return(error);
 453 }
 454
 455 /*
 456  * Release a volume.  Call hammer_io_release on the last reference.  We have
 457  * to acquire an exclusive lock to interlock against volume->ondisk tests
 458  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 459  * lock to be held.
 460  *
 461  * Volumes are not unloaded from memory during normal operation.
 462  */
 463 void
 464 hammer_rel_volume(hammer_volume_t volume, int flush)
 465 {
 466         struct buf *bp = NULL;
 467
 468         crit_enter();
 469         if (volume->io.lock.refs == 1) {
 470                 ++volume->io.loading;
 471                 hammer_lock_ex(&volume->io.lock);
 472                 if (volume->io.lock.refs == 1) {
 473                         volume->ondisk = NULL;
 474                         bp = hammer_io_release(&volume->io, flush);
 475                 }
 476                 --volume->io.loading;
 477                 hammer_unlock(&volume->io.lock);
 478         }
 479         hammer_unref(&volume->io.lock);
 480         if (bp)
 481                 brelse(bp);
 482         crit_exit();
 483 }
 484
 485 int
 486 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 487 {
 488         hammer_volume_t vol;
 489         int i;
 490
 491         for (i = 0; i < hmp->nvolumes; ++i) {
 492                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 493                 if (vol == NULL)
 494                         return(EINVAL);
 495         }
 496         return(0);
 497 }
 498
 499 /************************************************************************
 500  *                              BUFFERS                                 *
 501  ************************************************************************
 502  *
 503  * Manage buffers.  Currently all blockmap-backed zones are translated
 504  * to zone-2 buffer offsets.
 505  */
 506 hammer_buffer_t
 507 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 508                   int bytes, int isnew, int *errorp)
 509 {
 510         hammer_buffer_t buffer;
 511         hammer_volume_t volume;
 512         hammer_off_t    zone2_offset;
 513         hammer_io_type_t iotype;
 514         int vol_no;
 515         int zone;
 516
 517         buf_offset &= ~HAMMER_BUFMASK64;
 518 again:
 519         /*
 520          * Shortcut if the buffer is already cached
 521          */
 522         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 523         if (buffer) {
 524                 if (buffer->io.lock.refs == 0)
 525                         ++hammer_count_refedbufs;
 526                 hammer_ref(&buffer->io.lock);
 527
 528                 /*
 529                  * Onced refed the ondisk field will not be cleared by
 530                  * any other action.
 531                  */
 532                 if (buffer->ondisk && buffer->io.loading == 0) {
 533                         *errorp = 0;
 534                         return(buffer);
 535                 }
 536
 537                 /*
 538                  * The buffer is no longer loose if it has a ref, and
 539                  * cannot become loose once it gains a ref.  Loose
 540                  * buffers will never be in a modified state.  This should
 541                  * only occur on the 0->1 transition of refs.
 542                  *
 543                  * lose_list can be modified via a biodone() interrupt.
 544                  */
 545                 if (buffer->io.mod_list == &hmp->lose_list) {
 546                         crit_enter();   /* biodone race against list */
 547                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 548                                      mod_entry);
 549                         crit_exit();
 550                         buffer->io.mod_list = NULL;
 551                         KKASSERT(buffer->io.modified == 0);
 552                 }
 553                 goto found;
 554         }
 555
 556         /*
 557          * What is the buffer class?
 558          */
 559         zone = HAMMER_ZONE_DECODE(buf_offset);
 560
 561         switch(zone) {
 562         case HAMMER_ZONE_LARGE_DATA_INDEX:
 563         case HAMMER_ZONE_SMALL_DATA_INDEX:
 564                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 565                 break;
 566         case HAMMER_ZONE_UNDO_INDEX:
 567                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 568                 break;
 569         case HAMMER_ZONE_META_INDEX:
 570         default:
 571                 /*
 572                  * NOTE: inode data and directory entries are placed in this
 573                  * zone.  inode atime/mtime is updated in-place and thus
 574                  * buffers containing inodes must be synchronized as
 575                  * meta-buffers, same as buffers containing B-Tree info.
 576                  */
 577                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 578                 break;
 579         }
 580
 581         /*
 582          * Handle blockmap offset translations
 583          */
 584         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 585                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 586         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 587                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 588         } else {
 589                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 590                 zone2_offset = buf_offset;
 591                 *errorp = 0;
 592         }
 593         if (*errorp)
 594                 return(NULL);
 595
 596         /*
 597          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 598          * specifications.
 599          */
 600         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 601                  HAMMER_ZONE_RAW_BUFFER);
 602         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 603         volume = hammer_get_volume(hmp, vol_no, errorp);
 604         if (volume == NULL)
 605                 return(NULL);
 606
 607         KKASSERT(zone2_offset < volume->maxbuf_off);
 608
 609         /*
 610          * Allocate a new buffer structure.  We will check for races later.
 611          */
 612         ++hammer_count_buffers;
 613         buffer = kmalloc(sizeof(*buffer), M_HAMMER,
 614                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 615         buffer->zone2_offset = zone2_offset;
 616         buffer->zoneX_offset = buf_offset;
 617         buffer->volume = volume;
 618
 619         hammer_io_init(&buffer->io, hmp, iotype);
 620         buffer->io.offset = volume->ondisk->vol_buf_beg +
 621                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 622         buffer->io.bytes = bytes;
 623         TAILQ_INIT(&buffer->clist);
 624         hammer_ref(&buffer->io.lock);
 625
 626         /*
 627          * Insert the buffer into the RB tree and handle late collisions.
 628          */
 629         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 630                 hammer_unref(&buffer->io.lock);
 631                 --hammer_count_buffers;
 632                 kfree(buffer, M_HAMMER);
 633                 goto again;
 634         }
 635         ++hammer_count_refedbufs;
 636 found:
 637
 638         /*
 639          * Deal with on-disk info and loading races.
 640          */
 641         if (buffer->ondisk == NULL || buffer->io.loading) {
 642                 *errorp = hammer_load_buffer(buffer, isnew);
 643                 if (*errorp) {
 644                         hammer_rel_buffer(buffer, 1);
 645                         buffer = NULL;
 646                 }
 647         } else {
 648                 *errorp = 0;
 649         }
 650         return(buffer);
 651 }
 652
 653 /*
 654  * This is used by the direct-read code to deal with large-data buffers
 655  * created by the reblocker and mirror-write code.  The direct-read code
 656  * bypasses the HAMMER buffer subsystem and so any aliased dirty hammer
 657  * buffers must be fully synced to disk before we can issue the direct-read.
 658  *
 659  * This code path is not considered critical as only the rebocker and
 660  * mirror-write code will create large-data buffers via the HAMMER buffer
 661  * subsystem.  They do that because they operate at the B-Tree level and
 662  * do not access the vnode/inode structures.
 663  */
 664 void
 665 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 666 {
 667         hammer_buffer_t buffer;
 668         int error;
 669
 670         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 671                  HAMMER_ZONE_LARGE_DATA);
 672
 673         while (bytes > 0) {
 674                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 675                                    base_offset);
 676                 if (buffer && buffer->io.modified) {
 677                         error = hammer_ref_buffer(buffer);
 678                         if (error == 0 && buffer->io.modified) {
 679                                 hammer_io_write_interlock(&buffer->io);
 680                                 hammer_io_flush(&buffer->io);
 681                                 hammer_io_done_interlock(&buffer->io);
 682                                 hammer_io_wait(&buffer->io);
 683                                 hammer_rel_buffer(buffer, 0);
 684                         }
 685                 }
 686                 base_offset += HAMMER_BUFSIZE;
 687                 bytes -= HAMMER_BUFSIZE;
 688         }
 689 }
 690
 691 /*
 692  * Destroy all buffers covering the specified zoneX offset range.  This
 693  * is called when the related blockmap layer2 entry is freed or when
 694  * a direct write bypasses our buffer/buffer-cache subsystem.
 695  *
 696  * The buffers may be referenced by the caller itself.  Setting reclaim
 697  * will cause the buffer to be destroyed when it's ref count reaches zero.
 698  */
 699 void
 700 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 701                    hammer_off_t zone2_offset, int bytes)
 702 {
 703         hammer_buffer_t buffer;
 704         hammer_volume_t volume;
 705         int vol_no;
 706         int error;
 707
 708         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 709         volume = hammer_get_volume(hmp, vol_no, &error);
 710         KKASSERT(error == 0);
 711
 712         while (bytes > 0) {
 713                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 714                                    base_offset);
 715                 if (buffer) {
 716                         error = hammer_ref_buffer(buffer);
 717                         if (error == 0) {
 718                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 719                                 hammer_io_clear_modify(&buffer->io, 1);
 720                                 buffer->io.reclaim = 1;
 721                                 KKASSERT(buffer->volume == volume);
 722                                 hammer_rel_buffer(buffer, 0);
 723                         }
 724                 } else {
 725                         hammer_io_inval(volume, zone2_offset);
 726                 }
 727                 base_offset += HAMMER_BUFSIZE;
 728                 zone2_offset += HAMMER_BUFSIZE;
 729                 bytes -= HAMMER_BUFSIZE;
 730         }
 731         hammer_rel_volume(volume, 0);
 732 }
 733
 734 static int
 735 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 736 {
 737         hammer_volume_t volume;
 738         int error;
 739
 740         /*
 741          * Load the buffer's on-disk info
 742          */
 743         volume = buffer->volume;
 744         ++buffer->io.loading;
 745         hammer_lock_ex(&buffer->io.lock);
 746
 747         if (hammer_debug_io & 0x0001) {
 748                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 749                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 750                         buffer->ondisk);
 751         }
 752
 753         if (buffer->ondisk == NULL) {
 754                 if (isnew) {
 755                         error = hammer_io_new(volume->devvp, &buffer->io);
 756                 } else {
 757                         error = hammer_io_read(volume->devvp, &buffer->io,
 758                                                volume->maxraw_off);
 759                 }
 760                 if (error == 0)
 761                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 762         } else if (isnew) {
 763                 error = hammer_io_new(volume->devvp, &buffer->io);
 764         } else {
 765                 error = 0;
 766         }
 767         --buffer->io.loading;
 768         hammer_unlock(&buffer->io.lock);
 769         return (error);
 770 }
 771
 772 /*
 773  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 774  * This routine is only called during unmount.
 775  */
 776 int
 777 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 778 {
 779         /*
 780          * Clean up the persistent ref ioerror might have on the buffer
 781          * and acquire a ref (steal ioerror's if we can).
 782          */
 783         if (buffer->io.ioerror) {
 784                 buffer->io.ioerror = 0;
 785         } else {
 786                 if (buffer->io.lock.refs == 0)
 787                         ++hammer_count_refedbufs;
 788                 hammer_ref(&buffer->io.lock);
 789         }
 790
 791         /*
 792          * We must not flush a dirty buffer to disk on umount.  It should
 793          * have already been dealt with by the flusher, or we may be in
 794          * catastrophic failure.
 795          */
 796         hammer_io_clear_modify(&buffer->io, 1);
 797         hammer_flush_buffer_nodes(buffer);
 798         KKASSERT(buffer->io.lock.refs == 1);
 799         hammer_rel_buffer(buffer, 2);
 800         return(0);
 801 }
 802
 803 /*
 804  * Reference a buffer that is either already referenced or via a specially
 805  * handled pointer (aka cursor->buffer).
 806  */
 807 int
 808 hammer_ref_buffer(hammer_buffer_t buffer)
 809 {
 810         int error;
 811
 812         if (buffer->io.lock.refs == 0)
 813                 ++hammer_count_refedbufs;
 814         hammer_ref(&buffer->io.lock);
 815
 816         /*
 817          * At this point a biodone() will not touch the buffer other then
 818          * incidental bits.  However, lose_list can be modified via
 819          * a biodone() interrupt.
 820          *
 821          * No longer loose
 822          */
 823         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 824                 crit_enter();
 825                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 826                 buffer->io.mod_list = NULL;
 827                 crit_exit();
 828         }
 829
 830         if (buffer->ondisk == NULL || buffer->io.loading) {
 831                 error = hammer_load_buffer(buffer, 0);
 832                 if (error) {
 833                         hammer_rel_buffer(buffer, 1);
 834                         /*
 835                          * NOTE: buffer pointer can become stale after
 836                          * the above release.
 837                          */
 838                 }
 839         } else {
 840                 error = 0;
 841         }
 842         return(error);
 843 }
 844
 845 /*
 846  * Release a buffer.  We have to deal with several places where
 847  * another thread can ref the buffer.
 848  *
 849  * Only destroy the structure itself if the related buffer cache buffer
 850  * was disassociated from it.  This ties the management of the structure
 851  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 852  * embedded io is referenced or not.
 853  */
 854 void
 855 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 856 {
 857         hammer_volume_t volume;
 858         struct buf *bp = NULL;
 859         int freeme = 0;
 860
 861         crit_enter();
 862         if (buffer->io.lock.refs == 1) {
 863                 ++buffer->io.loading;   /* force interlock check */
 864                 hammer_lock_ex(&buffer->io.lock);
 865                 if (buffer->io.lock.refs == 1) {
 866                         bp = hammer_io_release(&buffer->io, flush);
 867
 868                         if (buffer->io.lock.refs == 1)
 869                                 --hammer_count_refedbufs;
 870
 871                         if (buffer->io.bp == NULL &&
 872                             buffer->io.lock.refs == 1) {
 873                                 /*
 874                                  * Final cleanup
 875                                  *
 876                                  * NOTE: It is impossible for any associated
 877                                  * B-Tree nodes to have refs if the buffer
 878                                  * has no additional refs.
 879                                  */
 880                                 RB_REMOVE(hammer_buf_rb_tree,
 881                                           &buffer->io.hmp->rb_bufs_root,
 882                                           buffer);
 883                                 volume = buffer->volume;
 884                                 buffer->volume = NULL; /* sanity */
 885                                 hammer_rel_volume(volume, 0);
 886                                 hammer_io_clear_modlist(&buffer->io);
 887                                 hammer_flush_buffer_nodes(buffer);
 888                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 889                                 freeme = 1;
 890                         }
 891                 }
 892                 --buffer->io.loading;
 893                 hammer_unlock(&buffer->io.lock);
 894         }
 895         hammer_unref(&buffer->io.lock);
 896         crit_exit();
 897         if (bp)
 898                 brelse(bp);
 899         if (freeme) {
 900                 --hammer_count_buffers;
 901                 kfree(buffer, M_HAMMER);
 902         }
 903 }
 904
 905 /*
 906  * Access the filesystem buffer containing the specified hammer offset.
 907  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 908  * relative buffer offset.  It must also have bit 55 set to be valid.
 909  * (see hammer_off_t in hammer_disk.h).
 910  *
 911  * Any prior buffer in *bufferp will be released and replaced by the
 912  * requested buffer.
 913  */
 914 static __inline
 915 void *
 916 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 917              int *errorp, struct hammer_buffer **bufferp)
 918 {
 919         hammer_buffer_t buffer;
 920         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 921
 922         buf_offset &= ~HAMMER_BUFMASK64;
 923         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 924
 925         buffer = *bufferp;
 926         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 927                                buffer->zoneX_offset != buf_offset)) {
 928                 if (buffer)
 929                         hammer_rel_buffer(buffer, 0);
 930                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 931                 *bufferp = buffer;
 932         } else {
 933                 *errorp = 0;
 934         }
 935
 936         /*
 937          * Return a pointer to the buffer data.
 938          */
 939         if (buffer == NULL)
 940                 return(NULL);
 941         else
 942                 return((char *)buffer->ondisk + xoff);
 943 }
 944
 945 void *
 946 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 947              int *errorp, struct hammer_buffer **bufferp)
 948 {
 949         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 950 }
 951
 952 void *
 953 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 954                  int *errorp, struct hammer_buffer **bufferp)
 955 {
 956         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 957         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 958 }
 959
 960 /*
 961  * Access the filesystem buffer containing the specified hammer offset.
 962  * No disk read operation occurs.  The result buffer may contain garbage.
 963  *
 964  * Any prior buffer in *bufferp will be released and replaced by the
 965  * requested buffer.
 966  *
 967  * This function marks the buffer dirty but does not increment its
 968  * modify_refs count.
 969  */
 970 static __inline
 971 void *
 972 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 973              int *errorp, struct hammer_buffer **bufferp)
 974 {
 975         hammer_buffer_t buffer;
 976         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 977
 978         buf_offset &= ~HAMMER_BUFMASK64;
 979
 980         buffer = *bufferp;
 981         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 982                                buffer->zoneX_offset != buf_offset)) {
 983                 if (buffer)
 984                         hammer_rel_buffer(buffer, 0);
 985                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 986                 *bufferp = buffer;
 987         } else {
 988                 *errorp = 0;
 989         }
 990
 991         /*
 992          * Return a pointer to the buffer data.
 993          */
 994         if (buffer == NULL)
 995                 return(NULL);
 996         else
 997                 return((char *)buffer->ondisk + xoff);
 998 }
 999
1000 void *
1001 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1002              int *errorp, struct hammer_buffer **bufferp)
1003 {
1004         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1005 }
1006
1007 void *
1008 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1009                 int *errorp, struct hammer_buffer **bufferp)
1010 {
1011         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1012         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1013 }
1014
1015 /************************************************************************
1016  *                              NODES                                   *
1017  ************************************************************************
1018  *
1019  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1020  * method used by the HAMMER filesystem.
1021  *
1022  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1023  * associated with its buffer, and will only referenced the buffer while
1024  * the node itself is referenced.
1025  *
1026  * A hammer_node can also be passively associated with other HAMMER
1027  * structures, such as inodes, while retaining 0 references.  These
1028  * associations can be cleared backwards using a pointer-to-pointer in
1029  * the hammer_node.
1030  *
1031  * This allows the HAMMER implementation to cache hammer_nodes long-term
1032  * and short-cut a great deal of the infrastructure's complexity.  In
1033  * most cases a cached node can be reacquired without having to dip into
1034  * either the buffer or cluster management code.
1035  *
1036  * The caller must pass a referenced cluster on call and will retain
1037  * ownership of the reference on return.  The node will acquire its own
1038  * additional references, if necessary.
1039  */
1040 hammer_node_t
1041 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
1042                 int isnew, int *errorp)
1043 {
1044         hammer_node_t node;
1045
1046         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1047
1048         /*
1049          * Locate the structure, allocating one if necessary.
1050          */
1051 again:
1052         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1053         if (node == NULL) {
1054                 ++hammer_count_nodes;
1055                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
1056                 node->node_offset = node_offset;
1057                 node->hmp = hmp;
1058                 TAILQ_INIT(&node->cursor_list);
1059                 TAILQ_INIT(&node->cache_list);
1060                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1061                         --hammer_count_nodes;
1062                         kfree(node, M_HAMMER);
1063                         goto again;
1064                 }
1065         }
1066         hammer_ref(&node->lock);
1067         if (node->ondisk)
1068                 *errorp = 0;
1069         else
1070                 *errorp = hammer_load_node(node, isnew);
1071         if (*errorp) {
1072                 hammer_rel_node(node);
1073                 node = NULL;
1074         }
1075         return(node);
1076 }
1077
1078 /*
1079  * Reference an already-referenced node.
1080  */
1081 void
1082 hammer_ref_node(hammer_node_t node)
1083 {
1084         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1085         hammer_ref(&node->lock);
1086 }
1087
1088 /*
1089  * Load a node's on-disk data reference.
1090  */
1091 static int
1092 hammer_load_node(hammer_node_t node, int isnew)
1093 {
1094         hammer_buffer_t buffer;
1095         hammer_off_t buf_offset;
1096         int error;
1097
1098         error = 0;
1099         ++node->loading;
1100         hammer_lock_ex(&node->lock);
1101         if (node->ondisk == NULL) {
1102                 /*
1103                  * This is a little confusing but the jist is that
1104                  * node->buffer determines whether the node is on
1105                  * the buffer's clist and node->ondisk determines
1106                  * whether the buffer is referenced.
1107                  *
1108                  * We could be racing a buffer release, in which case
1109                  * node->buffer may become NULL while we are blocked
1110                  * referencing the buffer.
1111                  */
1112                 if ((buffer = node->buffer) != NULL) {
1113                         error = hammer_ref_buffer(buffer);
1114                         if (error == 0 && node->buffer == NULL) {
1115                                 TAILQ_INSERT_TAIL(&buffer->clist,
1116                                                   node, entry);
1117                                 node->buffer = buffer;
1118                         }
1119                 } else {
1120                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1121                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1122                                                    HAMMER_BUFSIZE, 0, &error);
1123                         if (buffer) {
1124                                 KKASSERT(error == 0);
1125                                 TAILQ_INSERT_TAIL(&buffer->clist,
1126                                                   node, entry);
1127                                 node->buffer = buffer;
1128                         }
1129                 }
1130                 if (error)
1131                         goto failed;
1132                 node->ondisk = (void *)((char *)buffer->ondisk +
1133                                         (node->node_offset & HAMMER_BUFMASK));
1134                 if (isnew == 0 &&
1135                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1136                         if (hammer_crc_test_btree(node->ondisk) == 0)
1137                                 Debugger("CRC FAILED: B-TREE NODE");
1138                         node->flags |= HAMMER_NODE_CRCGOOD;
1139                 }
1140         }
1141 failed:
1142         --node->loading;
1143         hammer_unlock(&node->lock);
1144         return (error);
1145 }
1146
1147 /*
1148  * Safely reference a node, interlock against flushes via the IO subsystem.
1149  */
1150 hammer_node_t
1151 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1152                      int *errorp)
1153 {
1154         hammer_node_t node;
1155
1156         node = cache->node;
1157         if (node != NULL) {
1158                 hammer_ref(&node->lock);
1159                 if (node->ondisk)
1160                         *errorp = 0;
1161                 else
1162                         *errorp = hammer_load_node(node, 0);
1163                 if (*errorp) {
1164                         hammer_rel_node(node);
1165                         node = NULL;
1166                 }
1167         } else {
1168                 *errorp = ENOENT;
1169         }
1170         return(node);
1171 }
1172
1173 /*
1174  * Release a hammer_node.  On the last release the node dereferences
1175  * its underlying buffer and may or may not be destroyed.
1176  */
1177 void
1178 hammer_rel_node(hammer_node_t node)
1179 {
1180         hammer_buffer_t buffer;
1181
1182         /*
1183          * If this isn't the last ref just decrement the ref count and
1184          * return.
1185          */
1186         if (node->lock.refs > 1) {
1187                 hammer_unref(&node->lock);
1188                 return;
1189         }
1190
1191         /*
1192          * If there is no ondisk info or no buffer the node failed to load,
1193          * remove the last reference and destroy the node.
1194          */
1195         if (node->ondisk == NULL) {
1196                 hammer_unref(&node->lock);
1197                 hammer_flush_node(node);
1198                 /* node is stale now */
1199                 return;
1200         }
1201
1202         /*
1203          * Do not disassociate the node from the buffer if it represents
1204          * a modified B-Tree node that still needs its crc to be generated.
1205          */
1206         if (node->flags & HAMMER_NODE_NEEDSCRC)
1207                 return;
1208
1209         /*
1210          * Do final cleanups and then either destroy the node and leave it
1211          * passively cached.  The buffer reference is removed regardless.
1212          */
1213         buffer = node->buffer;
1214         node->ondisk = NULL;
1215
1216         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1217                 hammer_unref(&node->lock);
1218                 hammer_rel_buffer(buffer, 0);
1219                 return;
1220         }
1221
1222         /*
1223          * Destroy the node.
1224          */
1225         hammer_unref(&node->lock);
1226         hammer_flush_node(node);
1227         /* node is stale */
1228         hammer_rel_buffer(buffer, 0);
1229 }
1230
1231 /*
1232  * Free space on-media associated with a B-Tree node.
1233  */
1234 void
1235 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1236 {
1237         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1238         node->flags |= HAMMER_NODE_DELETED;
1239         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1240 }
1241
1242 /*
1243  * Passively cache a referenced hammer_node.  The caller may release
1244  * the node on return.
1245  */
1246 void
1247 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1248 {
1249         /*
1250          * If the node doesn't exist, or is being deleted, don't cache it!
1251          *
1252          * The node can only ever be NULL in the I/O failure path.
1253          */
1254         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1255                 return;
1256         if (cache->node == node)
1257                 return;
1258         while (cache->node)
1259                 hammer_uncache_node(cache);
1260         if (node->flags & HAMMER_NODE_DELETED)
1261                 return;
1262         cache->node = node;
1263         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1264 }
1265
1266 void
1267 hammer_uncache_node(hammer_node_cache_t cache)
1268 {
1269         hammer_node_t node;
1270
1271         if ((node = cache->node) != NULL) {
1272                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1273                 cache->node = NULL;
1274                 if (TAILQ_EMPTY(&node->cache_list))
1275                         hammer_flush_node(node);
1276         }
1277 }
1278
1279 /*
1280  * Remove a node's cache references and destroy the node if it has no
1281  * other references or backing store.
1282  */
1283 void
1284 hammer_flush_node(hammer_node_t node)
1285 {
1286         hammer_node_cache_t cache;
1287         hammer_buffer_t buffer;
1288
1289         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1290                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1291                 cache->node = NULL;
1292         }
1293         if (node->lock.refs == 0 && node->ondisk == NULL) {
1294                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1295                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1296                 if ((buffer = node->buffer) != NULL) {
1297                         node->buffer = NULL;
1298                         TAILQ_REMOVE(&buffer->clist, node, entry);
1299                         /* buffer is unreferenced because ondisk is NULL */
1300                 }
1301                 --hammer_count_nodes;
1302                 kfree(node, M_HAMMER);
1303         }
1304 }
1305
1306 /*
1307  * Flush passively cached B-Tree nodes associated with this buffer.
1308  * This is only called when the buffer is about to be destroyed, so
1309  * none of the nodes should have any references.  The buffer is locked.
1310  *
1311  * We may be interlocked with the buffer.
1312  */
1313 void
1314 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1315 {
1316         hammer_node_t node;
1317
1318         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1319                 KKASSERT(node->ondisk == NULL);
1320                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1321
1322                 if (node->lock.refs == 0) {
1323                         hammer_ref(&node->lock);
1324                         node->flags |= HAMMER_NODE_FLUSH;
1325                         hammer_rel_node(node);
1326                 } else {
1327                         KKASSERT(node->loading != 0);
1328                         KKASSERT(node->buffer != NULL);
1329                         buffer = node->buffer;
1330                         node->buffer = NULL;
1331                         TAILQ_REMOVE(&buffer->clist, node, entry);
1332                         /* buffer is unreferenced because ondisk is NULL */
1333                 }
1334         }
1335 }
1336
1337
1338 /************************************************************************
1339  *                              ALLOCATORS                              *
1340  ************************************************************************/
1341
1342 /*
1343  * Allocate a B-Tree node.
1344  */
1345 hammer_node_t
1346 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1347 {
1348         hammer_buffer_t buffer = NULL;
1349         hammer_node_t node = NULL;
1350         hammer_off_t node_offset;
1351
1352         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1353                                             sizeof(struct hammer_node_ondisk),
1354                                             errorp);
1355         if (*errorp == 0) {
1356                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1357                 hammer_modify_node_noundo(trans, node);
1358                 bzero(node->ondisk, sizeof(*node->ondisk));
1359                 hammer_modify_node_done(node);
1360         }
1361         if (buffer)
1362                 hammer_rel_buffer(buffer, 0);
1363         return(node);
1364 }
1365
1366 /*
1367  * Allocate data.  If the address of a data buffer is supplied then
1368  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1369  * will be set to the related buffer.  The caller must release it when
1370  * finally done.  The initial *data_bufferp should be set to NULL by
1371  * the caller.
1372  *
1373  * The caller is responsible for making hammer_modify*() calls on the
1374  * *data_bufferp.
1375  */
1376 void *
1377 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1378                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1379                   struct hammer_buffer **data_bufferp, int *errorp)
1380 {
1381         void *data;
1382         int zone;
1383
1384         /*
1385          * Allocate data
1386          */
1387         if (data_len) {
1388                 switch(rec_type) {
1389                 case HAMMER_RECTYPE_INODE:
1390                 case HAMMER_RECTYPE_DIRENTRY:
1391                 case HAMMER_RECTYPE_EXT:
1392                 case HAMMER_RECTYPE_FIX:
1393                 case HAMMER_RECTYPE_PFS:
1394                         zone = HAMMER_ZONE_META_INDEX;
1395                         break;
1396                 case HAMMER_RECTYPE_DATA:
1397                 case HAMMER_RECTYPE_DB:
1398                         if (data_len <= HAMMER_BUFSIZE / 2) {
1399                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1400                         } else {
1401                                 data_len = (data_len + HAMMER_BUFMASK) &
1402                                            ~HAMMER_BUFMASK;
1403                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1404                         }
1405                         break;
1406                 default:
1407                         panic("hammer_alloc_data: rec_type %04x unknown",
1408                               rec_type);
1409                         zone = 0;       /* NOT REACHED */
1410                         break;
1411                 }
1412                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1413                                                       data_len, errorp);
1414         } else {
1415                 *data_offsetp = 0;
1416         }
1417         if (*errorp == 0 && data_bufferp) {
1418                 if (data_len) {
1419                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1420                                                 data_len, errorp, data_bufferp);
1421                 } else {
1422                         data = NULL;
1423                 }
1424         } else {
1425                 data = NULL;
1426         }
1427         return(data);
1428 }
1429
1430 /*
1431  * Sync dirty buffers to the media and clean-up any loose ends.
1432  *
1433  * These functions do not start the flusher going, they simply
1434  * queue everything up to the flusher.
1435  */
1436 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1437 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1438
1439 int
1440 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1441 {
1442         struct hammer_sync_info info;
1443
1444         info.error = 0;
1445         info.waitfor = waitfor;
1446         if (waitfor == MNT_WAIT) {
1447                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1448                               hammer_sync_scan1, hammer_sync_scan2, &info);
1449         } else {
1450                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1451                               hammer_sync_scan1, hammer_sync_scan2, &info);
1452         }
1453         return(info.error);
1454 }
1455
1456 /*
1457  * Filesystem sync.  If doing a synchronous sync make a second pass on
1458  * the vnodes in case any were already flushing during the first pass,
1459  * and activate the flusher twice (the second time brings the UNDO FIFO's
1460  * start position up to the end position after the first call).
1461  */
1462 int
1463 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1464 {
1465         struct hammer_sync_info info;
1466
1467         info.error = 0;
1468         info.waitfor = MNT_NOWAIT;
1469         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1470                       hammer_sync_scan1, hammer_sync_scan2, &info);
1471         if (info.error == 0 && waitfor == MNT_WAIT) {
1472                 info.waitfor = waitfor;
1473                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1474                               hammer_sync_scan1, hammer_sync_scan2, &info);
1475         }
1476         if (waitfor == MNT_WAIT) {
1477                 hammer_flusher_sync(hmp);
1478                 hammer_flusher_sync(hmp);
1479         } else {
1480                 hammer_flusher_async(hmp, NULL);
1481         }
1482         return(info.error);
1483 }
1484
1485 static int
1486 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1487 {
1488         struct hammer_inode *ip;
1489
1490         ip = VTOI(vp);
1491         if (vp->v_type == VNON || ip == NULL ||
1492             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1493              RB_EMPTY(&vp->v_rbdirty_tree))) {
1494                 return(-1);
1495         }
1496         return(0);
1497 }
1498
1499 static int
1500 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1501 {
1502         struct hammer_sync_info *info = data;
1503         struct hammer_inode *ip;
1504         int error;
1505
1506         ip = VTOI(vp);
1507         if (vp->v_type == VNON || vp->v_type == VBAD ||
1508             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1509              RB_EMPTY(&vp->v_rbdirty_tree))) {
1510                 return(0);
1511         }
1512         error = VOP_FSYNC(vp, MNT_NOWAIT);
1513         if (error)
1514                 info->error = error;
1515         return(0);
1516 }
1517