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 1.76 2008/08/29 20:19:08 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_transaction_t trans,
  52                                 hammer_node_t node, int isnew);
  53
  54 static int
  55 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  56 {
  57         if (vol1->vol_no < vol2->vol_no)
  58                 return(-1);
  59         if (vol1->vol_no > vol2->vol_no)
  60                 return(1);
  61         return(0);
  62 }
  63
  64 /*
  65  * hammer_buffer structures are indexed via their zoneX_offset, not
  66  * their zone2_offset.
  67  */
  68 static int
  69 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  70 {
  71         if (buf1->zoneX_offset < buf2->zoneX_offset)
  72                 return(-1);
  73         if (buf1->zoneX_offset > buf2->zoneX_offset)
  74                 return(1);
  75         return(0);
  76 }
  77
  78 static int
  79 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  80 {
  81         if (node1->node_offset < node2->node_offset)
  82                 return(-1);
  83         if (node1->node_offset > node2->node_offset)
  84                 return(1);
  85         return(0);
  86 }
  87
  88 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  89              hammer_vol_rb_compare, int32_t, vol_no);
  90 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  91              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  92 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  93              hammer_nod_rb_compare, hammer_off_t, node_offset);
  94
  95 /************************************************************************
  96  *                              VOLUMES                                 *
  97  ************************************************************************
  98  *
  99  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 100  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 101  * not load a new volume.
 102  *
 103  * Calls made to hammer_load_volume() or single-threaded
 104  */
 105 int
 106 hammer_install_volume(struct hammer_mount *hmp, const char *volname,
 107                       struct vnode *devvp)
 108 {
 109         struct mount *mp;
 110         hammer_volume_t volume;
 111         struct hammer_volume_ondisk *ondisk;
 112         struct nlookupdata nd;
 113         struct buf *bp = NULL;
 114         int error;
 115         int ronly;
 116         int setmp = 0;
 117
 118         mp = hmp->mp;
 119         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 120
 121         /*
 122          * Allocate a volume structure
 123          */
 124         ++hammer_count_volumes;
 125         volume = kmalloc(sizeof(*volume), hmp->m_misc, M_WAITOK|M_ZERO);
 126         volume->vol_name = kstrdup(volname, hmp->m_misc);
 127         volume->io.hmp = hmp;   /* bootstrap */
 128         hammer_io_init(&volume->io, volume, HAMMER_STRUCTURE_VOLUME);
 129         volume->io.offset = 0LL;
 130         volume->io.bytes = HAMMER_BUFSIZE;
 131
 132         /*
 133          * Get the device vnode
 134          */
 135         if (devvp == NULL) {
 136                 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 137                 if (error == 0)
 138                         error = nlookup(&nd);
 139                 if (error == 0)
 140                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 141                 nlookup_done(&nd);
 142         } else {
 143                 error = 0;
 144                 volume->devvp = devvp;
 145         }
 146
 147         if (error == 0) {
 148                 if (vn_isdisk(volume->devvp, &error)) {
 149                         error = vfs_mountedon(volume->devvp);
 150                 }
 151         }
 152         if (error == 0 && vcount(volume->devvp) > 0)
 153                 error = EBUSY;
 154         if (error == 0) {
 155                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 156                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 157                 if (error == 0) {
 158                         error = VOP_OPEN(volume->devvp,
 159                                          (ronly ? FREAD : FREAD|FWRITE),
 160                                          FSCRED, NULL);
 161                 }
 162                 vn_unlock(volume->devvp);
 163         }
 164         if (error) {
 165                 hammer_free_volume(volume);
 166                 return(error);
 167         }
 168         volume->devvp->v_rdev->si_mountpoint = mp;
 169         setmp = 1;
 170
 171         /*
 172          * Extract the volume number from the volume header and do various
 173          * sanity checks.
 174          */
 175         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 176         if (error)
 177                 goto late_failure;
 178         ondisk = (void *)bp->b_data;
 179         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 180                 kprintf("hammer_mount: volume %s has an invalid header\n",
 181                         volume->vol_name);
 182                 error = EFTYPE;
 183                 goto late_failure;
 184         }
 185         volume->vol_no = ondisk->vol_no;
 186         volume->buffer_base = ondisk->vol_buf_beg;
 187         volume->vol_flags = ondisk->vol_flags;
 188         volume->nblocks = ondisk->vol_nblocks;
 189         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 190                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 191         volume->maxraw_off = ondisk->vol_buf_end;
 192
 193         if (RB_EMPTY(&hmp->rb_vols_root)) {
 194                 hmp->fsid = ondisk->vol_fsid;
 195         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 196                 kprintf("hammer_mount: volume %s's fsid does not match "
 197                         "other volumes\n", volume->vol_name);
 198                 error = EFTYPE;
 199                 goto late_failure;
 200         }
 201
 202         /*
 203          * Insert the volume structure into the red-black tree.
 204          */
 205         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 206                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 207                         volume->vol_name, volume->vol_no);
 208                 error = EEXIST;
 209         }
 210
 211         /*
 212          * Set the root volume .  HAMMER special cases rootvol the structure.
 213          * We do not hold a ref because this would prevent related I/O
 214          * from being flushed.
 215          */
 216         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 217                 hmp->rootvol = volume;
 218                 hmp->nvolumes = ondisk->vol_count;
 219                 if (bp) {
 220                         brelse(bp);
 221                         bp = NULL;
 222                 }
 223                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 224                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 225                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 226                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 227         }
 228 late_failure:
 229         if (bp)
 230                 brelse(bp);
 231         if (error) {
 232                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 233                 if (setmp)
 234                         volume->devvp->v_rdev->si_mountpoint = NULL;
 235                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 236                 hammer_free_volume(volume);
 237         }
 238         return (error);
 239 }
 240
 241 /*
 242  * This is called for each volume when updating the mount point from
 243  * read-write to read-only or vise-versa.
 244  */
 245 int
 246 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 247 {
 248         if (volume->devvp) {
 249                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 250                 if (volume->io.hmp->ronly) {
 251                         /* do not call vinvalbuf */
 252                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 253                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 254                 } else {
 255                         /* do not call vinvalbuf */
 256                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 257                         VOP_CLOSE(volume->devvp, FREAD);
 258                 }
 259                 vn_unlock(volume->devvp);
 260         }
 261         return(0);
 262 }
 263
 264 /*
 265  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 266  * so returns -1 on failure.
 267  */
 268 int
 269 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 270 {
 271         hammer_mount_t hmp = volume->io.hmp;
 272         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 273         struct buf *bp;
 274
 275         /*
 276          * Clean up the root volume pointer, which is held unlocked in hmp.
 277          */
 278         if (hmp->rootvol == volume)
 279                 hmp->rootvol = NULL;
 280
 281         /*
 282          * We must not flush a dirty buffer to disk on umount.  It should
 283          * have already been dealt with by the flusher, or we may be in
 284          * catastrophic failure.
 285          */
 286         hammer_io_clear_modify(&volume->io, 1);
 287         volume->io.waitdep = 1;
 288         bp = hammer_io_release(&volume->io, 1);
 289
 290         /*
 291          * Clean up the persistent ref ioerror might have on the volume
 292          */
 293         if (volume->io.ioerror) {
 294                 volume->io.ioerror = 0;
 295                 hammer_unref(&volume->io.lock);
 296         }
 297
 298         /*
 299          * There should be no references on the volume, no clusters, and
 300          * no super-clusters.
 301          */
 302         KKASSERT(volume->io.lock.refs == 0);
 303         if (bp)
 304                 brelse(bp);
 305
 306         volume->ondisk = NULL;
 307         if (volume->devvp) {
 308                 if (volume->devvp->v_rdev &&
 309                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 310                 ) {
 311                         volume->devvp->v_rdev->si_mountpoint = NULL;
 312                 }
 313                 if (ronly) {
 314                         /*
 315                          * Make sure we don't sync anything to disk if we
 316                          * are in read-only mode (1) or critically-errored
 317                          * (2).  Note that there may be dirty buffers in
 318                          * normal read-only mode from crash recovery.
 319                          */
 320                         vinvalbuf(volume->devvp, 0, 0, 0);
 321                         VOP_CLOSE(volume->devvp, FREAD);
 322                 } else {
 323                         /*
 324                          * Normal termination, save any dirty buffers
 325                          * (XXX there really shouldn't be any).
 326                          */
 327                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 328                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 329                 }
 330         }
 331
 332         /*
 333          * Destroy the structure
 334          */
 335         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 336         hammer_free_volume(volume);
 337         return(0);
 338 }
 339
 340 static
 341 void
 342 hammer_free_volume(hammer_volume_t volume)
 343 {
 344         hammer_mount_t hmp = volume->io.hmp;
 345
 346         if (volume->vol_name) {
 347                 kfree(volume->vol_name, hmp->m_misc);
 348                 volume->vol_name = NULL;
 349         }
 350         if (volume->devvp) {
 351                 vrele(volume->devvp);
 352                 volume->devvp = NULL;
 353         }
 354         --hammer_count_volumes;
 355         kfree(volume, hmp->m_misc);
 356 }
 357
 358 /*
 359  * Get a HAMMER volume.  The volume must already exist.
 360  */
 361 hammer_volume_t
 362 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 363 {
 364         struct hammer_volume *volume;
 365
 366         /*
 367          * Locate the volume structure
 368          */
 369         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 370         if (volume == NULL) {
 371                 *errorp = ENOENT;
 372                 return(NULL);
 373         }
 374         hammer_ref(&volume->io.lock);
 375
 376         /*
 377          * Deal with on-disk info
 378          */
 379         if (volume->ondisk == NULL || volume->io.loading) {
 380                 *errorp = hammer_load_volume(volume);
 381                 if (*errorp) {
 382                         hammer_rel_volume(volume, 1);
 383                         volume = NULL;
 384                 }
 385         } else {
 386                 *errorp = 0;
 387         }
 388         return(volume);
 389 }
 390
 391 int
 392 hammer_ref_volume(hammer_volume_t volume)
 393 {
 394         int error;
 395
 396         hammer_ref(&volume->io.lock);
 397
 398         /*
 399          * Deal with on-disk info
 400          */
 401         if (volume->ondisk == NULL || volume->io.loading) {
 402                 error = hammer_load_volume(volume);
 403                 if (error)
 404                         hammer_rel_volume(volume, 1);
 405         } else {
 406                 error = 0;
 407         }
 408         return (error);
 409 }
 410
 411 hammer_volume_t
 412 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 413 {
 414         hammer_volume_t volume;
 415
 416         volume = hmp->rootvol;
 417         KKASSERT(volume != NULL);
 418         hammer_ref(&volume->io.lock);
 419
 420         /*
 421          * Deal with on-disk info
 422          */
 423         if (volume->ondisk == NULL || volume->io.loading) {
 424                 *errorp = hammer_load_volume(volume);
 425                 if (*errorp) {
 426                         hammer_rel_volume(volume, 1);
 427                         volume = NULL;
 428                 }
 429         } else {
 430                 *errorp = 0;
 431         }
 432         return (volume);
 433 }
 434
 435 /*
 436  * Load a volume's on-disk information.  The volume must be referenced and
 437  * not locked.  We temporarily acquire an exclusive lock to interlock
 438  * against releases or multiple get's.
 439  */
 440 static int
 441 hammer_load_volume(hammer_volume_t volume)
 442 {
 443         int error;
 444
 445         ++volume->io.loading;
 446         hammer_lock_ex(&volume->io.lock);
 447
 448         if (volume->ondisk == NULL) {
 449                 error = hammer_io_read(volume->devvp, &volume->io,
 450                                        volume->maxraw_off);
 451                 if (error == 0)
 452                         volume->ondisk = (void *)volume->io.bp->b_data;
 453         } else {
 454                 error = 0;
 455         }
 456         --volume->io.loading;
 457         hammer_unlock(&volume->io.lock);
 458         return(error);
 459 }
 460
 461 /*
 462  * Release a volume.  Call hammer_io_release on the last reference.  We have
 463  * to acquire an exclusive lock to interlock against volume->ondisk tests
 464  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 465  * lock to be held.
 466  *
 467  * Volumes are not unloaded from memory during normal operation.
 468  */
 469 void
 470 hammer_rel_volume(hammer_volume_t volume, int flush)
 471 {
 472         struct buf *bp = NULL;
 473
 474         crit_enter();
 475         if (volume->io.lock.refs == 1) {
 476                 ++volume->io.loading;
 477                 hammer_lock_ex(&volume->io.lock);
 478                 if (volume->io.lock.refs == 1) {
 479                         volume->ondisk = NULL;
 480                         bp = hammer_io_release(&volume->io, flush);
 481                 }
 482                 --volume->io.loading;
 483                 hammer_unlock(&volume->io.lock);
 484         }
 485         hammer_unref(&volume->io.lock);
 486         if (bp)
 487                 brelse(bp);
 488         crit_exit();
 489 }
 490
 491 int
 492 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 493 {
 494         hammer_volume_t vol;
 495         int i;
 496
 497         for (i = 0; i < hmp->nvolumes; ++i) {
 498                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 499                 if (vol == NULL)
 500                         return(EINVAL);
 501         }
 502         return(0);
 503 }
 504
 505 /************************************************************************
 506  *                              BUFFERS                                 *
 507  ************************************************************************
 508  *
 509  * Manage buffers.  Currently all blockmap-backed zones are direct-mapped
 510  * to zone-2 buffer offsets, without a translation stage.  However, the
 511  * hammer_buffer structure is indexed by its zoneX_offset, not its
 512  * zone2_offset.
 513  *
 514  * The proper zone must be maintained throughout the code-base all the way
 515  * through to the big-block allocator, or routines like hammer_del_buffers()
 516  * will not be able to locate all potentially conflicting buffers.
 517  */
 518 hammer_buffer_t
 519 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 520                   int bytes, int isnew, int *errorp)
 521 {
 522         hammer_buffer_t buffer;
 523         hammer_volume_t volume;
 524         hammer_off_t    zone2_offset;
 525         hammer_io_type_t iotype;
 526         int vol_no;
 527         int zone;
 528
 529         buf_offset &= ~HAMMER_BUFMASK64;
 530 again:
 531         /*
 532          * Shortcut if the buffer is already cached
 533          */
 534         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 535         if (buffer) {
 536                 if (buffer->io.lock.refs == 0)
 537                         ++hammer_count_refedbufs;
 538                 hammer_ref(&buffer->io.lock);
 539
 540                 /*
 541                  * Once refed the ondisk field will not be cleared by
 542                  * any other action.
 543                  */
 544                 if (buffer->ondisk && buffer->io.loading == 0) {
 545                         *errorp = 0;
 546                         hammer_io_advance(&buffer->io);
 547                         return(buffer);
 548                 }
 549
 550                 /*
 551                  * The buffer is no longer loose if it has a ref, and
 552                  * cannot become loose once it gains a ref.  Loose
 553                  * buffers will never be in a modified state.  This should
 554                  * only occur on the 0->1 transition of refs.
 555                  *
 556                  * lose_list can be modified via a biodone() interrupt.
 557                  */
 558                 if (buffer->io.mod_list == &hmp->lose_list) {
 559                         crit_enter();   /* biodone race against list */
 560                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 561                                      mod_entry);
 562                         crit_exit();
 563                         buffer->io.mod_list = NULL;
 564                         KKASSERT(buffer->io.modified == 0);
 565                 }
 566                 goto found;
 567         }
 568
 569         /*
 570          * What is the buffer class?
 571          */
 572         zone = HAMMER_ZONE_DECODE(buf_offset);
 573
 574         switch(zone) {
 575         case HAMMER_ZONE_LARGE_DATA_INDEX:
 576         case HAMMER_ZONE_SMALL_DATA_INDEX:
 577                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 578                 break;
 579         case HAMMER_ZONE_UNDO_INDEX:
 580                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 581                 break;
 582         case HAMMER_ZONE_META_INDEX:
 583         default:
 584                 /*
 585                  * NOTE: inode data and directory entries are placed in this
 586                  * zone.  inode atime/mtime is updated in-place and thus
 587                  * buffers containing inodes must be synchronized as
 588                  * meta-buffers, same as buffers containing B-Tree info.
 589                  */
 590                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 591                 break;
 592         }
 593
 594         /*
 595          * Handle blockmap offset translations
 596          */
 597         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 598                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 599         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 600                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 601         } else {
 602                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 603                 zone2_offset = buf_offset;
 604                 *errorp = 0;
 605         }
 606         if (*errorp)
 607                 return(NULL);
 608
 609         /*
 610          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 611          * specifications.
 612          */
 613         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 614                  HAMMER_ZONE_RAW_BUFFER);
 615         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 616         volume = hammer_get_volume(hmp, vol_no, errorp);
 617         if (volume == NULL)
 618                 return(NULL);
 619
 620         KKASSERT(zone2_offset < volume->maxbuf_off);
 621
 622         /*
 623          * Allocate a new buffer structure.  We will check for races later.
 624          */
 625         ++hammer_count_buffers;
 626         buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
 627                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 628         buffer->zone2_offset = zone2_offset;
 629         buffer->zoneX_offset = buf_offset;
 630
 631         hammer_io_init(&buffer->io, volume, iotype);
 632         buffer->io.offset = volume->ondisk->vol_buf_beg +
 633                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 634         buffer->io.bytes = bytes;
 635         TAILQ_INIT(&buffer->clist);
 636         hammer_ref(&buffer->io.lock);
 637
 638         /*
 639          * Insert the buffer into the RB tree and handle late collisions.
 640          */
 641         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 642                 hammer_unref(&buffer->io.lock); /* safety */
 643                 --hammer_count_buffers;
 644                 hammer_rel_volume(volume, 0);
 645                 buffer->io.volume = NULL;       /* safety */
 646                 kfree(buffer, hmp->m_misc);
 647                 goto again;
 648         }
 649         ++hammer_count_refedbufs;
 650 found:
 651
 652         /*
 653          * Deal with on-disk info and loading races.
 654          */
 655         if (buffer->ondisk == NULL || buffer->io.loading) {
 656                 *errorp = hammer_load_buffer(buffer, isnew);
 657                 if (*errorp) {
 658                         hammer_rel_buffer(buffer, 1);
 659                         buffer = NULL;
 660                 }
 661         } else {
 662                 *errorp = 0;
 663         }
 664         hammer_io_advance(&buffer->io);
 665         return(buffer);
 666 }
 667
 668 /*
 669  * This is used by the direct-read code to deal with large-data buffers
 670  * created by the reblocker and mirror-write code.  The direct-read code
 671  * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
 672  * running hammer buffers must be fully synced to disk before we can issue
 673  * the direct-read.
 674  *
 675  * This code path is not considered critical as only the rebocker and
 676  * mirror-write code will create large-data buffers via the HAMMER buffer
 677  * subsystem.  They do that because they operate at the B-Tree level and
 678  * do not access the vnode/inode structures.
 679  */
 680 void
 681 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 682 {
 683         hammer_buffer_t buffer;
 684         int error;
 685
 686         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 687                  HAMMER_ZONE_LARGE_DATA);
 688
 689         while (bytes > 0) {
 690                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 691                                    base_offset);
 692                 if (buffer && (buffer->io.modified || buffer->io.running)) {
 693                         error = hammer_ref_buffer(buffer);
 694                         if (error == 0) {
 695                                 hammer_io_wait(&buffer->io);
 696                                 if (buffer->io.modified) {
 697                                         hammer_io_write_interlock(&buffer->io);
 698                                         hammer_io_flush(&buffer->io, 0);
 699                                         hammer_io_done_interlock(&buffer->io);
 700                                         hammer_io_wait(&buffer->io);
 701                                 }
 702                                 hammer_rel_buffer(buffer, 0);
 703                         }
 704                 }
 705                 base_offset += HAMMER_BUFSIZE;
 706                 bytes -= HAMMER_BUFSIZE;
 707         }
 708 }
 709
 710 /*
 711  * Destroy all buffers covering the specified zoneX offset range.  This
 712  * is called when the related blockmap layer2 entry is freed or when
 713  * a direct write bypasses our buffer/buffer-cache subsystem.
 714  *
 715  * The buffers may be referenced by the caller itself.  Setting reclaim
 716  * will cause the buffer to be destroyed when it's ref count reaches zero.
 717  *
 718  * Return 0 on success, EAGAIN if some buffers could not be destroyed due
 719  * to additional references held by other threads, or some other (typically
 720  * fatal) error.
 721  */
 722 int
 723 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 724                    hammer_off_t zone2_offset, int bytes,
 725                    int report_conflicts)
 726 {
 727         hammer_buffer_t buffer;
 728         hammer_volume_t volume;
 729         int vol_no;
 730         int error;
 731         int ret_error;
 732
 733         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 734         volume = hammer_get_volume(hmp, vol_no, &ret_error);
 735         KKASSERT(ret_error == 0);
 736
 737         while (bytes > 0) {
 738                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 739                                    base_offset);
 740                 if (buffer) {
 741                         error = hammer_ref_buffer(buffer);
 742                         if (error == 0 && buffer->io.lock.refs != 1) {
 743                                 error = EAGAIN;
 744                                 hammer_rel_buffer(buffer, 0);
 745                         }
 746                         if (error == 0) {
 747                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 748                                 hammer_io_clear_modify(&buffer->io, 1);
 749                                 buffer->io.reclaim = 1;
 750                                 buffer->io.waitdep = 1;
 751                                 KKASSERT(buffer->io.volume == volume);
 752                                 hammer_rel_buffer(buffer, 0);
 753                         }
 754                 } else {
 755                         error = hammer_io_inval(volume, zone2_offset);
 756                 }
 757                 if (error) {
 758                         ret_error = error;
 759                         if (report_conflicts ||
 760                             (hammer_debug_general & 0x8000)) {
 761                                 kprintf("hammer_del_buffers: unable to "
 762                                         "invalidate %016llx buffer=%p rep=%d\n",
 763                                         (long long)base_offset,
 764                                         buffer, report_conflicts);
 765                         }
 766                 }
 767                 base_offset += HAMMER_BUFSIZE;
 768                 zone2_offset += HAMMER_BUFSIZE;
 769                 bytes -= HAMMER_BUFSIZE;
 770         }
 771         hammer_rel_volume(volume, 0);
 772         return (ret_error);
 773 }
 774
 775 static int
 776 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 777 {
 778         hammer_volume_t volume;
 779         int error;
 780
 781         /*
 782          * Load the buffer's on-disk info
 783          */
 784         volume = buffer->io.volume;
 785         ++buffer->io.loading;
 786         hammer_lock_ex(&buffer->io.lock);
 787
 788         if (hammer_debug_io & 0x0001) {
 789                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 790                         (long long)buffer->zoneX_offset,
 791                         (long long)buffer->zone2_offset,
 792                         isnew, buffer->ondisk);
 793         }
 794
 795         if (buffer->ondisk == NULL) {
 796                 if (isnew) {
 797                         error = hammer_io_new(volume->devvp, &buffer->io);
 798                 } else {
 799                         error = hammer_io_read(volume->devvp, &buffer->io,
 800                                                volume->maxraw_off);
 801                 }
 802                 if (error == 0)
 803                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 804         } else if (isnew) {
 805                 error = hammer_io_new(volume->devvp, &buffer->io);
 806         } else {
 807                 error = 0;
 808         }
 809         --buffer->io.loading;
 810         hammer_unlock(&buffer->io.lock);
 811         return (error);
 812 }
 813
 814 /*
 815  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 816  * This routine is only called during unmount.
 817  */
 818 int
 819 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 820 {
 821         /*
 822          * Clean up the persistent ref ioerror might have on the buffer
 823          * and acquire a ref (steal ioerror's if we can).
 824          */
 825         if (buffer->io.ioerror) {
 826                 buffer->io.ioerror = 0;
 827         } else {
 828                 if (buffer->io.lock.refs == 0)
 829                         ++hammer_count_refedbufs;
 830                 hammer_ref(&buffer->io.lock);
 831         }
 832
 833         /*
 834          * We must not flush a dirty buffer to disk on umount.  It should
 835          * have already been dealt with by the flusher, or we may be in
 836          * catastrophic failure.
 837          *
 838          * We must set waitdep to ensure that a running buffer is waited
 839          * on and released prior to us trying to unload the volume.
 840          */
 841         hammer_io_clear_modify(&buffer->io, 1);
 842         hammer_flush_buffer_nodes(buffer);
 843         KKASSERT(buffer->io.lock.refs == 1);
 844         buffer->io.waitdep = 1;
 845         hammer_rel_buffer(buffer, 2);
 846         return(0);
 847 }
 848
 849 /*
 850  * Reference a buffer that is either already referenced or via a specially
 851  * handled pointer (aka cursor->buffer).
 852  */
 853 int
 854 hammer_ref_buffer(hammer_buffer_t buffer)
 855 {
 856         int error;
 857
 858         if (buffer->io.lock.refs == 0)
 859                 ++hammer_count_refedbufs;
 860         hammer_ref(&buffer->io.lock);
 861
 862         /*
 863          * At this point a biodone() will not touch the buffer other then
 864          * incidental bits.  However, lose_list can be modified via
 865          * a biodone() interrupt.
 866          *
 867          * No longer loose
 868          */
 869         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 870                 crit_enter();
 871                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 872                 buffer->io.mod_list = NULL;
 873                 crit_exit();
 874         }
 875
 876         if (buffer->ondisk == NULL || buffer->io.loading) {
 877                 error = hammer_load_buffer(buffer, 0);
 878                 if (error) {
 879                         hammer_rel_buffer(buffer, 1);
 880                         /*
 881                          * NOTE: buffer pointer can become stale after
 882                          * the above release.
 883                          */
 884                 }
 885         } else {
 886                 error = 0;
 887         }
 888         return(error);
 889 }
 890
 891 /*
 892  * Release a buffer.  We have to deal with several places where
 893  * another thread can ref the buffer.
 894  *
 895  * Only destroy the structure itself if the related buffer cache buffer
 896  * was disassociated from it.  This ties the management of the structure
 897  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 898  * embedded io is referenced or not.
 899  */
 900 void
 901 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 902 {
 903         hammer_volume_t volume;
 904         hammer_mount_t hmp;
 905         struct buf *bp = NULL;
 906         int freeme = 0;
 907
 908         hmp = buffer->io.hmp;
 909
 910         crit_enter();
 911         if (buffer->io.lock.refs == 1) {
 912                 ++buffer->io.loading;   /* force interlock check */
 913                 hammer_lock_ex(&buffer->io.lock);
 914                 if (buffer->io.lock.refs == 1) {
 915                         bp = hammer_io_release(&buffer->io, flush);
 916
 917                         if (buffer->io.lock.refs == 1)
 918                                 --hammer_count_refedbufs;
 919
 920                         if (buffer->io.bp == NULL &&
 921                             buffer->io.lock.refs == 1) {
 922                                 /*
 923                                  * Final cleanup
 924                                  *
 925                                  * NOTE: It is impossible for any associated
 926                                  * B-Tree nodes to have refs if the buffer
 927                                  * has no additional refs.
 928                                  */
 929                                 RB_REMOVE(hammer_buf_rb_tree,
 930                                           &buffer->io.hmp->rb_bufs_root,
 931                                           buffer);
 932                                 volume = buffer->io.volume;
 933                                 buffer->io.volume = NULL; /* sanity */
 934                                 hammer_rel_volume(volume, 0);
 935                                 hammer_io_clear_modlist(&buffer->io);
 936                                 hammer_flush_buffer_nodes(buffer);
 937                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 938                                 freeme = 1;
 939                         }
 940                 }
 941                 --buffer->io.loading;
 942                 hammer_unlock(&buffer->io.lock);
 943         }
 944         hammer_unref(&buffer->io.lock);
 945         crit_exit();
 946         if (bp)
 947                 brelse(bp);
 948         if (freeme) {
 949                 --hammer_count_buffers;
 950                 kfree(buffer, hmp->m_misc);
 951         }
 952 }
 953
 954 /*
 955  * Access the filesystem buffer containing the specified hammer offset.
 956  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 957  * relative buffer offset.  It must also have bit 55 set to be valid.
 958  * (see hammer_off_t in hammer_disk.h).
 959  *
 960  * Any prior buffer in *bufferp will be released and replaced by the
 961  * requested buffer.
 962  *
 963  * NOTE: The buffer is indexed via its zoneX_offset but we allow the
 964  * passed cached *bufferp to match against either zoneX or zone2.
 965  */
 966 static __inline
 967 void *
 968 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 969              int *errorp, struct hammer_buffer **bufferp)
 970 {
 971         hammer_buffer_t buffer;
 972         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 973
 974         buf_offset &= ~HAMMER_BUFMASK64;
 975         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 976
 977         buffer = *bufferp;
 978         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 979                                buffer->zoneX_offset != buf_offset)) {
 980                 if (buffer)
 981                         hammer_rel_buffer(buffer, 0);
 982                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 983                 *bufferp = buffer;
 984         } else {
 985                 *errorp = 0;
 986         }
 987
 988         /*
 989          * Return a pointer to the buffer data.
 990          */
 991         if (buffer == NULL)
 992                 return(NULL);
 993         else
 994                 return((char *)buffer->ondisk + xoff);
 995 }
 996
 997 void *
 998 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 999              int *errorp, struct hammer_buffer **bufferp)
1000 {
1001         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1002 }
1003
1004 void *
1005 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1006                  int *errorp, struct hammer_buffer **bufferp)
1007 {
1008         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1009         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
1010 }
1011
1012 /*
1013  * Access the filesystem buffer containing the specified hammer offset.
1014  * No disk read operation occurs.  The result buffer may contain garbage.
1015  *
1016  * Any prior buffer in *bufferp will be released and replaced by the
1017  * requested buffer.
1018  *
1019  * This function marks the buffer dirty but does not increment its
1020  * modify_refs count.
1021  */
1022 static __inline
1023 void *
1024 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1025              int *errorp, struct hammer_buffer **bufferp)
1026 {
1027         hammer_buffer_t buffer;
1028         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1029
1030         buf_offset &= ~HAMMER_BUFMASK64;
1031
1032         buffer = *bufferp;
1033         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1034                                buffer->zoneX_offset != buf_offset)) {
1035                 if (buffer)
1036                         hammer_rel_buffer(buffer, 0);
1037                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
1038                 *bufferp = buffer;
1039         } else {
1040                 *errorp = 0;
1041         }
1042
1043         /*
1044          * Return a pointer to the buffer data.
1045          */
1046         if (buffer == NULL)
1047                 return(NULL);
1048         else
1049                 return((char *)buffer->ondisk + xoff);
1050 }
1051
1052 void *
1053 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1054              int *errorp, struct hammer_buffer **bufferp)
1055 {
1056         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1057 }
1058
1059 void *
1060 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1061                 int *errorp, struct hammer_buffer **bufferp)
1062 {
1063         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1064         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1065 }
1066
1067 /************************************************************************
1068  *                              NODES                                   *
1069  ************************************************************************
1070  *
1071  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1072  * method used by the HAMMER filesystem.
1073  *
1074  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1075  * associated with its buffer, and will only referenced the buffer while
1076  * the node itself is referenced.
1077  *
1078  * A hammer_node can also be passively associated with other HAMMER
1079  * structures, such as inodes, while retaining 0 references.  These
1080  * associations can be cleared backwards using a pointer-to-pointer in
1081  * the hammer_node.
1082  *
1083  * This allows the HAMMER implementation to cache hammer_nodes long-term
1084  * and short-cut a great deal of the infrastructure's complexity.  In
1085  * most cases a cached node can be reacquired without having to dip into
1086  * either the buffer or cluster management code.
1087  *
1088  * The caller must pass a referenced cluster on call and will retain
1089  * ownership of the reference on return.  The node will acquire its own
1090  * additional references, if necessary.
1091  */
1092 hammer_node_t
1093 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1094                 int isnew, int *errorp)
1095 {
1096         hammer_mount_t hmp = trans->hmp;
1097         hammer_node_t node;
1098
1099         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1100
1101         /*
1102          * Locate the structure, allocating one if necessary.
1103          */
1104 again:
1105         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1106         if (node == NULL) {
1107                 ++hammer_count_nodes;
1108                 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1109                 node->node_offset = node_offset;
1110                 node->hmp = hmp;
1111                 TAILQ_INIT(&node->cursor_list);
1112                 TAILQ_INIT(&node->cache_list);
1113                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1114                         --hammer_count_nodes;
1115                         kfree(node, hmp->m_misc);
1116                         goto again;
1117                 }
1118         }
1119         hammer_ref(&node->lock);
1120         if (node->ondisk) {
1121                 *errorp = 0;
1122                 hammer_io_advance(&node->buffer->io);
1123         } else {
1124                 *errorp = hammer_load_node(trans, node, isnew);
1125                 trans->flags |= HAMMER_TRANSF_DIDIO;
1126         }
1127         if (*errorp) {
1128                 hammer_rel_node(node);
1129                 node = NULL;
1130         }
1131         return(node);
1132 }
1133
1134 /*
1135  * Reference an already-referenced node.
1136  */
1137 void
1138 hammer_ref_node(hammer_node_t node)
1139 {
1140         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1141         hammer_ref(&node->lock);
1142 }
1143
1144 /*
1145  * Load a node's on-disk data reference.
1146  */
1147 static int
1148 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew)
1149 {
1150         hammer_buffer_t buffer;
1151         hammer_off_t buf_offset;
1152         int error;
1153
1154         error = 0;
1155         ++node->loading;
1156         hammer_lock_ex(&node->lock);
1157         if (node->ondisk == NULL) {
1158                 /*
1159                  * This is a little confusing but the jist is that
1160                  * node->buffer determines whether the node is on
1161                  * the buffer's clist and node->ondisk determines
1162                  * whether the buffer is referenced.
1163                  *
1164                  * We could be racing a buffer release, in which case
1165                  * node->buffer may become NULL while we are blocked
1166                  * referencing the buffer.
1167                  */
1168                 if ((buffer = node->buffer) != NULL) {
1169                         error = hammer_ref_buffer(buffer);
1170                         if (error == 0 && node->buffer == NULL) {
1171                                 TAILQ_INSERT_TAIL(&buffer->clist,
1172                                                   node, entry);
1173                                 node->buffer = buffer;
1174                         }
1175                 } else {
1176                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1177                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1178                                                    HAMMER_BUFSIZE, 0, &error);
1179                         if (buffer) {
1180                                 KKASSERT(error == 0);
1181                                 TAILQ_INSERT_TAIL(&buffer->clist,
1182                                                   node, entry);
1183                                 node->buffer = buffer;
1184                         }
1185                 }
1186                 if (error)
1187                         goto failed;
1188                 node->ondisk = (void *)((char *)buffer->ondisk +
1189                                         (node->node_offset & HAMMER_BUFMASK));
1190
1191                 /*
1192                  * Check CRC.  NOTE: Neither flag is set and the CRC is not
1193                  * generated on new B-Tree nodes.
1194                  */
1195                 if (isnew == 0 &&
1196                     (node->flags & HAMMER_NODE_CRCANY) == 0) {
1197                         if (hammer_crc_test_btree(node->ondisk) == 0) {
1198                                 if (hammer_debug_debug & 0x0002)
1199                                         Debugger("CRC FAILED: B-TREE NODE");
1200                                 node->flags |= HAMMER_NODE_CRCBAD;
1201                         } else {
1202                                 node->flags |= HAMMER_NODE_CRCGOOD;
1203                         }
1204                 }
1205         }
1206         if (node->flags & HAMMER_NODE_CRCBAD) {
1207                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1208                         error = EDOM;
1209                 else
1210                         error = EIO;
1211         }
1212 failed:
1213         --node->loading;
1214         hammer_unlock(&node->lock);
1215         return (error);
1216 }
1217
1218 /*
1219  * Safely reference a node, interlock against flushes via the IO subsystem.
1220  */
1221 hammer_node_t
1222 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache,
1223                      int *errorp)
1224 {
1225         hammer_node_t node;
1226
1227         node = cache->node;
1228         if (node != NULL) {
1229                 hammer_ref(&node->lock);
1230                 if (node->ondisk) {
1231                         if (node->flags & HAMMER_NODE_CRCBAD) {
1232                                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1233                                         *errorp = EDOM;
1234                                 else
1235                                         *errorp = EIO;
1236                         } else {
1237                                 *errorp = 0;
1238                         }
1239                 } else {
1240                         *errorp = hammer_load_node(trans, node, 0);
1241                 }
1242                 if (*errorp) {
1243                         hammer_rel_node(node);
1244                         node = NULL;
1245                 }
1246         } else {
1247                 *errorp = ENOENT;
1248         }
1249         return(node);
1250 }
1251
1252 /*
1253  * Release a hammer_node.  On the last release the node dereferences
1254  * its underlying buffer and may or may not be destroyed.
1255  */
1256 void
1257 hammer_rel_node(hammer_node_t node)
1258 {
1259         hammer_buffer_t buffer;
1260
1261         /*
1262          * If this isn't the last ref just decrement the ref count and
1263          * return.
1264          */
1265         if (node->lock.refs > 1) {
1266                 hammer_unref(&node->lock);
1267                 return;
1268         }
1269
1270         /*
1271          * If there is no ondisk info or no buffer the node failed to load,
1272          * remove the last reference and destroy the node.
1273          */
1274         if (node->ondisk == NULL) {
1275                 hammer_unref(&node->lock);
1276                 hammer_flush_node(node);
1277                 /* node is stale now */
1278                 return;
1279         }
1280
1281         /*
1282          * Do not disassociate the node from the buffer if it represents
1283          * a modified B-Tree node that still needs its crc to be generated.
1284          */
1285         if (node->flags & HAMMER_NODE_NEEDSCRC)
1286                 return;
1287
1288         /*
1289          * Do final cleanups and then either destroy the node and leave it
1290          * passively cached.  The buffer reference is removed regardless.
1291          */
1292         buffer = node->buffer;
1293         node->ondisk = NULL;
1294
1295         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1296                 hammer_unref(&node->lock);
1297                 hammer_rel_buffer(buffer, 0);
1298                 return;
1299         }
1300
1301         /*
1302          * Destroy the node.
1303          */
1304         hammer_unref(&node->lock);
1305         hammer_flush_node(node);
1306         /* node is stale */
1307         hammer_rel_buffer(buffer, 0);
1308 }
1309
1310 /*
1311  * Free space on-media associated with a B-Tree node.
1312  */
1313 void
1314 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1315 {
1316         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1317         node->flags |= HAMMER_NODE_DELETED;
1318         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1319 }
1320
1321 /*
1322  * Passively cache a referenced hammer_node.  The caller may release
1323  * the node on return.
1324  */
1325 void
1326 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1327 {
1328         /*
1329          * If the node doesn't exist, or is being deleted, don't cache it!
1330          *
1331          * The node can only ever be NULL in the I/O failure path.
1332          */
1333         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1334                 return;
1335         if (cache->node == node)
1336                 return;
1337         while (cache->node)
1338                 hammer_uncache_node(cache);
1339         if (node->flags & HAMMER_NODE_DELETED)
1340                 return;
1341         cache->node = node;
1342         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1343 }
1344
1345 void
1346 hammer_uncache_node(hammer_node_cache_t cache)
1347 {
1348         hammer_node_t node;
1349
1350         if ((node = cache->node) != NULL) {
1351                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1352                 cache->node = NULL;
1353                 if (TAILQ_EMPTY(&node->cache_list))
1354                         hammer_flush_node(node);
1355         }
1356 }
1357
1358 /*
1359  * Remove a node's cache references and destroy the node if it has no
1360  * other references or backing store.
1361  */
1362 void
1363 hammer_flush_node(hammer_node_t node)
1364 {
1365         hammer_node_cache_t cache;
1366         hammer_buffer_t buffer;
1367         hammer_mount_t hmp = node->hmp;
1368
1369         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1370                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1371                 cache->node = NULL;
1372         }
1373         if (node->lock.refs == 0 && node->ondisk == NULL) {
1374                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1375                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1376                 if ((buffer = node->buffer) != NULL) {
1377                         node->buffer = NULL;
1378                         TAILQ_REMOVE(&buffer->clist, node, entry);
1379                         /* buffer is unreferenced because ondisk is NULL */
1380                 }
1381                 --hammer_count_nodes;
1382                 kfree(node, hmp->m_misc);
1383         }
1384 }
1385
1386 /*
1387  * Flush passively cached B-Tree nodes associated with this buffer.
1388  * This is only called when the buffer is about to be destroyed, so
1389  * none of the nodes should have any references.  The buffer is locked.
1390  *
1391  * We may be interlocked with the buffer.
1392  */
1393 void
1394 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1395 {
1396         hammer_node_t node;
1397
1398         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1399                 KKASSERT(node->ondisk == NULL);
1400                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1401
1402                 if (node->lock.refs == 0) {
1403                         hammer_ref(&node->lock);
1404                         node->flags |= HAMMER_NODE_FLUSH;
1405                         hammer_rel_node(node);
1406                 } else {
1407                         KKASSERT(node->loading != 0);
1408                         KKASSERT(node->buffer != NULL);
1409                         buffer = node->buffer;
1410                         node->buffer = NULL;
1411                         TAILQ_REMOVE(&buffer->clist, node, entry);
1412                         /* buffer is unreferenced because ondisk is NULL */
1413                 }
1414         }
1415 }
1416
1417
1418 /************************************************************************
1419  *                              ALLOCATORS                              *
1420  ************************************************************************/
1421
1422 /*
1423  * Allocate a B-Tree node.
1424  */
1425 hammer_node_t
1426 hammer_alloc_btree(hammer_transaction_t trans, hammer_off_t hint, int *errorp)
1427 {
1428         hammer_buffer_t buffer = NULL;
1429         hammer_node_t node = NULL;
1430         hammer_off_t node_offset;
1431
1432         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1433                                             sizeof(struct hammer_node_ondisk),
1434                                             hint, errorp);
1435         if (*errorp == 0) {
1436                 node = hammer_get_node(trans, node_offset, 1, errorp);
1437                 hammer_modify_node_noundo(trans, node);
1438                 bzero(node->ondisk, sizeof(*node->ondisk));
1439                 hammer_modify_node_done(node);
1440         }
1441         if (buffer)
1442                 hammer_rel_buffer(buffer, 0);
1443         return(node);
1444 }
1445
1446 /*
1447  * Allocate data.  If the address of a data buffer is supplied then
1448  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1449  * will be set to the related buffer.  The caller must release it when
1450  * finally done.  The initial *data_bufferp should be set to NULL by
1451  * the caller.
1452  *
1453  * The caller is responsible for making hammer_modify*() calls on the
1454  * *data_bufferp.
1455  */
1456 void *
1457 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1458                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1459                   struct hammer_buffer **data_bufferp,
1460                   hammer_off_t hint, int *errorp)
1461 {
1462         void *data;
1463         int zone;
1464
1465         /*
1466          * Allocate data
1467          */
1468         if (data_len) {
1469                 switch(rec_type) {
1470                 case HAMMER_RECTYPE_INODE:
1471                 case HAMMER_RECTYPE_DIRENTRY:
1472                 case HAMMER_RECTYPE_EXT:
1473                 case HAMMER_RECTYPE_FIX:
1474                 case HAMMER_RECTYPE_PFS:
1475                         zone = HAMMER_ZONE_META_INDEX;
1476                         break;
1477                 case HAMMER_RECTYPE_DATA:
1478                 case HAMMER_RECTYPE_DB:
1479                         if (data_len <= HAMMER_BUFSIZE / 2) {
1480                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1481                         } else {
1482                                 data_len = (data_len + HAMMER_BUFMASK) &
1483                                            ~HAMMER_BUFMASK;
1484                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1485                         }
1486                         break;
1487                 default:
1488                         panic("hammer_alloc_data: rec_type %04x unknown",
1489                               rec_type);
1490                         zone = 0;       /* NOT REACHED */
1491                         break;
1492                 }
1493                 *data_offsetp = hammer_blockmap_alloc(trans, zone, data_len,
1494                                                       hint, errorp);
1495         } else {
1496                 *data_offsetp = 0;
1497         }
1498         if (*errorp == 0 && data_bufferp) {
1499                 if (data_len) {
1500                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1501                                                 data_len, errorp, data_bufferp);
1502                 } else {
1503                         data = NULL;
1504                 }
1505         } else {
1506                 data = NULL;
1507         }
1508         return(data);
1509 }
1510
1511 /*
1512  * Sync dirty buffers to the media and clean-up any loose ends.
1513  *
1514  * These functions do not start the flusher going, they simply
1515  * queue everything up to the flusher.
1516  */
1517 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1518 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1519
1520 int
1521 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1522 {
1523         struct hammer_sync_info info;
1524
1525         info.error = 0;
1526         info.waitfor = waitfor;
1527         if (waitfor == MNT_WAIT) {
1528                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1529                               hammer_sync_scan1, hammer_sync_scan2, &info);
1530         } else {
1531                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1532                               hammer_sync_scan1, hammer_sync_scan2, &info);
1533         }
1534         return(info.error);
1535 }
1536
1537 /*
1538  * Filesystem sync.  If doing a synchronous sync make a second pass on
1539  * the vnodes in case any were already flushing during the first pass,
1540  * and activate the flusher twice (the second time brings the UNDO FIFO's
1541  * start position up to the end position after the first call).
1542  */
1543 int
1544 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1545 {
1546         struct hammer_sync_info info;
1547
1548         info.error = 0;
1549         info.waitfor = MNT_NOWAIT;
1550         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1551                       hammer_sync_scan1, hammer_sync_scan2, &info);
1552         if (info.error == 0 && waitfor == MNT_WAIT) {
1553                 info.waitfor = waitfor;
1554                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1555                               hammer_sync_scan1, hammer_sync_scan2, &info);
1556         }
1557         if (waitfor == MNT_WAIT) {
1558                 hammer_flusher_sync(hmp);
1559                 hammer_flusher_sync(hmp);
1560         } else {
1561                 hammer_flusher_async(hmp, NULL);
1562                 hammer_flusher_async(hmp, NULL);
1563         }
1564         return(info.error);
1565 }
1566
1567 static int
1568 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1569 {
1570         struct hammer_inode *ip;
1571
1572         ip = VTOI(vp);
1573         if (vp->v_type == VNON || ip == NULL ||
1574             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1575              RB_EMPTY(&vp->v_rbdirty_tree))) {
1576                 return(-1);
1577         }
1578         return(0);
1579 }
1580
1581 static int
1582 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1583 {
1584         struct hammer_sync_info *info = data;
1585         struct hammer_inode *ip;
1586         int error;
1587
1588         ip = VTOI(vp);
1589         if (vp->v_type == VNON || vp->v_type == VBAD ||
1590             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1591              RB_EMPTY(&vp->v_rbdirty_tree))) {
1592                 return(0);
1593         }
1594         error = VOP_FSYNC(vp, MNT_NOWAIT, 0);
1595         if (error)
1596                 info->error = error;
1597         return(0);
1598 }
1599