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 1.60 2008/06/20 05:38:26 dillon Exp $
  35  */
  36 /*
  37  * Manage HAMMER's on-disk structures.  These routines are primarily
  38  * responsible for interfacing with the kernel's I/O subsystem and for
  39  * managing in-memory structures.
  40  */
  41
  42 #include "hammer.h"
  43 #include <sys/fcntl.h>
  44 #include <sys/nlookup.h>
  45 #include <sys/buf.h>
  46 #include <sys/buf2.h>
  47
  48 static void hammer_free_volume(hammer_volume_t volume);
  49 static int hammer_load_volume(hammer_volume_t volume);
  50 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
  51 static int hammer_load_node(hammer_node_t node, int isnew);
  52
  53 /*
  54  * Red-Black tree support for various structures
  55  */
  56 int
  57 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  58 {
  59         if (ip1->obj_id < ip2->obj_id)
  60                 return(-1);
  61         if (ip1->obj_id > ip2->obj_id)
  62                 return(1);
  63         if (ip1->obj_asof < ip2->obj_asof)
  64                 return(-1);
  65         if (ip1->obj_asof > ip2->obj_asof)
  66                 return(1);
  67         return(0);
  68 }
  69
  70 static int
  71 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  72 {
  73         if (info->obj_id < ip->obj_id)
  74                 return(-1);
  75         if (info->obj_id > ip->obj_id)
  76                 return(1);
  77         if (info->obj_asof < ip->obj_asof)
  78                 return(-1);
  79         if (info->obj_asof > ip->obj_asof)
  80                 return(1);
  81         return(0);
  82 }
  83
  84 static int
  85 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  86 {
  87         if (vol1->vol_no < vol2->vol_no)
  88                 return(-1);
  89         if (vol1->vol_no > vol2->vol_no)
  90                 return(1);
  91         return(0);
  92 }
  93
  94 static int
  95 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  96 {
  97         if (buf1->zoneX_offset < buf2->zoneX_offset)
  98                 return(-1);
  99         if (buf1->zoneX_offset > buf2->zoneX_offset)
 100                 return(1);
 101         return(0);
 102 }
 103
 104 static int
 105 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 106 {
 107         if (node1->node_offset < node2->node_offset)
 108                 return(-1);
 109         if (node1->node_offset > node2->node_offset)
 110                 return(1);
 111         return(0);
 112 }
 113
 114 /*
 115  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 116  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 117  * functions are normal, e.g. hammer_buf_rb_tree_RB_LOOKUP(root, zone2_offset).
 118  */
 119 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 120 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 121                 hammer_inode_info_cmp, hammer_inode_info_t);
 122 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 123              hammer_vol_rb_compare, int32_t, vol_no);
 124 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 125              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
 126 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 127              hammer_nod_rb_compare, hammer_off_t, node_offset);
 128
 129 /************************************************************************
 130  *                              VOLUMES                                 *
 131  ************************************************************************
 132  *
 133  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 134  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 135  * not load a new volume.
 136  *
 137  * Calls made to hammer_load_volume() or single-threaded
 138  */
 139 int
 140 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 141 {
 142         struct mount *mp;
 143         hammer_volume_t volume;
 144         struct hammer_volume_ondisk *ondisk;
 145         struct nlookupdata nd;
 146         struct buf *bp = NULL;
 147         int error;
 148         int ronly;
 149         int setmp = 0;
 150
 151         mp = hmp->mp;
 152         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 153
 154         /*
 155          * Allocate a volume structure
 156          */
 157         ++hammer_count_volumes;
 158         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 159         volume->vol_name = kstrdup(volname, M_HAMMER);
 160         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 161         volume->io.offset = 0LL;
 162         volume->io.bytes = HAMMER_BUFSIZE;
 163
 164         /*
 165          * Get the device vnode
 166          */
 167         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 168         if (error == 0)
 169                 error = nlookup(&nd);
 170         if (error == 0)
 171                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 172         nlookup_done(&nd);
 173         if (error == 0) {
 174                 if (vn_isdisk(volume->devvp, &error)) {
 175                         error = vfs_mountedon(volume->devvp);
 176                 }
 177         }
 178         if (error == 0 &&
 179             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 180                 error = EBUSY;
 181         }
 182         if (error == 0) {
 183                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 184                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 185                 if (error == 0) {
 186                         error = VOP_OPEN(volume->devvp,
 187                                          (ronly ? FREAD : FREAD|FWRITE),
 188                                          FSCRED, NULL);
 189                 }
 190                 vn_unlock(volume->devvp);
 191         }
 192         if (error) {
 193                 hammer_free_volume(volume);
 194                 return(error);
 195         }
 196         volume->devvp->v_rdev->si_mountpoint = mp;
 197         setmp = 1;
 198
 199         /*
 200          * Extract the volume number from the volume header and do various
 201          * sanity checks.
 202          */
 203         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 204         if (error)
 205                 goto late_failure;
 206         ondisk = (void *)bp->b_data;
 207         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 208                 kprintf("hammer_mount: volume %s has an invalid header\n",
 209                         volume->vol_name);
 210                 error = EFTYPE;
 211                 goto late_failure;
 212         }
 213         volume->vol_no = ondisk->vol_no;
 214         volume->buffer_base = ondisk->vol_buf_beg;
 215         volume->vol_flags = ondisk->vol_flags;
 216         volume->nblocks = ondisk->vol_nblocks;
 217         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 218                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 219         volume->maxraw_off = ondisk->vol_buf_end;
 220
 221         if (RB_EMPTY(&hmp->rb_vols_root)) {
 222                 hmp->fsid = ondisk->vol_fsid;
 223         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 224                 kprintf("hammer_mount: volume %s's fsid does not match "
 225                         "other volumes\n", volume->vol_name);
 226                 error = EFTYPE;
 227                 goto late_failure;
 228         }
 229
 230         /*
 231          * Insert the volume structure into the red-black tree.
 232          */
 233         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 234                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 235                         volume->vol_name, volume->vol_no);
 236                 error = EEXIST;
 237         }
 238
 239         /*
 240          * Set the root volume .  HAMMER special cases rootvol the structure.
 241          * We do not hold a ref because this would prevent related I/O
 242          * from being flushed.
 243          */
 244         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 245                 hmp->rootvol = volume;
 246                 if (bp) {
 247                         brelse(bp);
 248                         bp = NULL;
 249                 }
 250                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 251                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 252                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 253                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 254                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 255         }
 256 late_failure:
 257         if (bp)
 258                 brelse(bp);
 259         if (error) {
 260                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 261                 if (setmp)
 262                         volume->devvp->v_rdev->si_mountpoint = NULL;
 263                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 264                 hammer_free_volume(volume);
 265         }
 266         return (error);
 267 }
 268
 269 /*
 270  * This is called for each volume when updating the mount point from
 271  * read-write to read-only or vise-versa.
 272  */
 273 int
 274 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 275 {
 276         if (volume->devvp) {
 277                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 278                 if (volume->io.hmp->ronly) {
 279                         /* do not call vinvalbuf */
 280                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 281                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 282                 } else {
 283                         /* do not call vinvalbuf */
 284                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 285                         VOP_CLOSE(volume->devvp, FREAD);
 286                 }
 287                 vn_unlock(volume->devvp);
 288         }
 289         return(0);
 290 }
 291
 292 /*
 293  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 294  * so returns -1 on failure.
 295  */
 296 int
 297 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 298 {
 299         struct hammer_mount *hmp = volume->io.hmp;
 300         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 301
 302         /*
 303          * Clean up the root volume pointer, which is held unlocked in hmp.
 304          */
 305         if (hmp->rootvol == volume)
 306                 hmp->rootvol = NULL;
 307
 308         /*
 309          * Release our buffer and flush anything left in the buffer cache.
 310          */
 311         volume->io.waitdep = 1;
 312         hammer_io_release(&volume->io, 1);
 313         hammer_io_clear_modlist(&volume->io);
 314
 315         /*
 316          * There should be no references on the volume, no clusters, and
 317          * no super-clusters.
 318          */
 319         KKASSERT(volume->io.lock.refs == 0);
 320
 321         volume->ondisk = NULL;
 322         if (volume->devvp) {
 323                 if (volume->devvp->v_rdev &&
 324                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 325                 ) {
 326                         volume->devvp->v_rdev->si_mountpoint = NULL;
 327                 }
 328                 if (ronly) {
 329                         vinvalbuf(volume->devvp, 0, 0, 0);
 330                         VOP_CLOSE(volume->devvp, FREAD);
 331                 } else {
 332                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 333                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 334                 }
 335         }
 336
 337         /*
 338          * Destroy the structure
 339          */
 340         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 341         hammer_free_volume(volume);
 342         return(0);
 343 }
 344
 345 static
 346 void
 347 hammer_free_volume(hammer_volume_t volume)
 348 {
 349         if (volume->vol_name) {
 350                 kfree(volume->vol_name, M_HAMMER);
 351                 volume->vol_name = NULL;
 352         }
 353         if (volume->devvp) {
 354                 vrele(volume->devvp);
 355                 volume->devvp = NULL;
 356         }
 357         --hammer_count_volumes;
 358         kfree(volume, M_HAMMER);
 359 }
 360
 361 /*
 362  * Get a HAMMER volume.  The volume must already exist.
 363  */
 364 hammer_volume_t
 365 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 366 {
 367         struct hammer_volume *volume;
 368
 369         /*
 370          * Locate the volume structure
 371          */
 372         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 373         if (volume == NULL) {
 374                 *errorp = ENOENT;
 375                 return(NULL);
 376         }
 377         hammer_ref(&volume->io.lock);
 378
 379         /*
 380          * Deal with on-disk info
 381          */
 382         if (volume->ondisk == NULL || volume->io.loading) {
 383                 *errorp = hammer_load_volume(volume);
 384                 if (*errorp) {
 385                         hammer_rel_volume(volume, 1);
 386                         volume = NULL;
 387                 }
 388         } else {
 389                 *errorp = 0;
 390         }
 391         return(volume);
 392 }
 393
 394 int
 395 hammer_ref_volume(hammer_volume_t volume)
 396 {
 397         int error;
 398
 399         hammer_ref(&volume->io.lock);
 400
 401         /*
 402          * Deal with on-disk info
 403          */
 404         if (volume->ondisk == NULL || volume->io.loading) {
 405                 error = hammer_load_volume(volume);
 406                 if (error)
 407                         hammer_rel_volume(volume, 1);
 408         } else {
 409                 error = 0;
 410         }
 411         return (error);
 412 }
 413
 414 hammer_volume_t
 415 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 416 {
 417         hammer_volume_t volume;
 418
 419         volume = hmp->rootvol;
 420         KKASSERT(volume != NULL);
 421         hammer_ref(&volume->io.lock);
 422
 423         /*
 424          * Deal with on-disk info
 425          */
 426         if (volume->ondisk == NULL || volume->io.loading) {
 427                 *errorp = hammer_load_volume(volume);
 428                 if (*errorp) {
 429                         hammer_rel_volume(volume, 1);
 430                         volume = NULL;
 431                 }
 432         } else {
 433                 *errorp = 0;
 434         }
 435         return (volume);
 436 }
 437
 438 /*
 439  * Load a volume's on-disk information.  The volume must be referenced and
 440  * not locked.  We temporarily acquire an exclusive lock to interlock
 441  * against releases or multiple get's.
 442  */
 443 static int
 444 hammer_load_volume(hammer_volume_t volume)
 445 {
 446         int error;
 447
 448         ++volume->io.loading;
 449         hammer_lock_ex(&volume->io.lock);
 450
 451         if (volume->ondisk == NULL) {
 452                 error = hammer_io_read(volume->devvp, &volume->io,
 453                                        volume->maxraw_off);
 454                 if (error == 0)
 455                         volume->ondisk = (void *)volume->io.bp->b_data;
 456         } else {
 457                 error = 0;
 458         }
 459         --volume->io.loading;
 460         hammer_unlock(&volume->io.lock);
 461         return(error);
 462 }
 463
 464 /*
 465  * Release a volume.  Call hammer_io_release on the last reference.  We have
 466  * to acquire an exclusive lock to interlock against volume->ondisk tests
 467  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 468  * lock to be held.
 469  *
 470  * Volumes are not unloaded from memory during normal operation.
 471  */
 472 void
 473 hammer_rel_volume(hammer_volume_t volume, int flush)
 474 {
 475         crit_enter();
 476         if (volume->io.lock.refs == 1) {
 477                 ++volume->io.loading;
 478                 hammer_lock_ex(&volume->io.lock);
 479                 if (volume->io.lock.refs == 1) {
 480                         volume->ondisk = NULL;
 481                         hammer_io_release(&volume->io, flush);
 482                 }
 483                 --volume->io.loading;
 484                 hammer_unlock(&volume->io.lock);
 485         }
 486         hammer_unref(&volume->io.lock);
 487         crit_exit();
 488 }
 489
 490 /************************************************************************
 491  *                              BUFFERS                                 *
 492  ************************************************************************
 493  *
 494  * Manage buffers.  Currently all blockmap-backed zones are translated
 495  * to zone-2 buffer offsets.
 496  */
 497 hammer_buffer_t
 498 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 499                   int bytes, int isnew, int *errorp)
 500 {
 501         hammer_buffer_t buffer;
 502         hammer_volume_t volume;
 503         hammer_off_t    zone2_offset;
 504         hammer_io_type_t iotype;
 505         int vol_no;
 506         int zone;
 507
 508         buf_offset &= ~HAMMER_BUFMASK64;
 509 again:
 510         /*
 511          * Shortcut if the buffer is already cached
 512          */
 513         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 514         if (buffer) {
 515                 if (buffer->io.lock.refs == 0)
 516                         ++hammer_count_refedbufs;
 517                 hammer_ref(&buffer->io.lock);
 518
 519                 /*
 520                  * Onced refed the ondisk field will not be cleared by
 521                  * any other action.
 522                  */
 523                 if (buffer->ondisk && buffer->io.loading == 0) {
 524                         *errorp = 0;
 525                         return(buffer);
 526                 }
 527
 528                 /*
 529                  * The buffer is no longer loose if it has a ref, and
 530                  * cannot become loose once it gains a ref.  Loose
 531                  * buffers will never be in a modified state.  This should
 532                  * only occur on the 0->1 transition of refs.
 533                  *
 534                  * lose_list can be modified via a biodone() interrupt.
 535                  */
 536                 if (buffer->io.mod_list == &hmp->lose_list) {
 537                         crit_enter();   /* biodone race against list */
 538                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 539                                      mod_entry);
 540                         crit_exit();
 541                         buffer->io.mod_list = NULL;
 542                         KKASSERT(buffer->io.modified == 0);
 543                 }
 544                 goto found;
 545         }
 546
 547         /*
 548          * What is the buffer class?
 549          */
 550         zone = HAMMER_ZONE_DECODE(buf_offset);
 551
 552         switch(zone) {
 553         case HAMMER_ZONE_LARGE_DATA_INDEX:
 554         case HAMMER_ZONE_SMALL_DATA_INDEX:
 555                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 556                 break;
 557         case HAMMER_ZONE_UNDO_INDEX:
 558                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 559                 break;
 560         case HAMMER_ZONE_META_INDEX:
 561         default:
 562                 /*
 563                  * NOTE: inode data and directory entries are placed in this
 564                  * zone.  inode atime/mtime is updated in-place and thus
 565                  * buffers containing inodes must be synchronized as
 566                  * meta-buffers, same as buffers containing B-Tree info.
 567                  */
 568                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 569                 break;
 570         }
 571
 572         /*
 573          * Handle blockmap offset translations
 574          */
 575         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 576                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 577         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 578                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 579         } else {
 580                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 581                 zone2_offset = buf_offset;
 582                 *errorp = 0;
 583         }
 584         if (*errorp)
 585                 return(NULL);
 586
 587         /*
 588          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 589          * specifications.
 590          */
 591         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 592                  HAMMER_ZONE_RAW_BUFFER);
 593         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 594         volume = hammer_get_volume(hmp, vol_no, errorp);
 595         if (volume == NULL)
 596                 return(NULL);
 597
 598         KKASSERT(zone2_offset < volume->maxbuf_off);
 599
 600         /*
 601          * Allocate a new buffer structure.  We will check for races later.
 602          */
 603         ++hammer_count_buffers;
 604         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 605         buffer->zone2_offset = zone2_offset;
 606         buffer->zoneX_offset = buf_offset;
 607         buffer->volume = volume;
 608
 609         hammer_io_init(&buffer->io, hmp, iotype);
 610         buffer->io.offset = volume->ondisk->vol_buf_beg +
 611                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 612         buffer->io.bytes = bytes;
 613         TAILQ_INIT(&buffer->clist);
 614         hammer_ref(&buffer->io.lock);
 615
 616         /*
 617          * Insert the buffer into the RB tree and handle late collisions.
 618          */
 619         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 620                 hammer_unref(&buffer->io.lock);
 621                 --hammer_count_buffers;
 622                 kfree(buffer, M_HAMMER);
 623                 goto again;
 624         }
 625         ++hammer_count_refedbufs;
 626 found:
 627
 628         /*
 629          * Deal with on-disk info and loading races.
 630          */
 631         if (buffer->ondisk == NULL || buffer->io.loading) {
 632                 *errorp = hammer_load_buffer(buffer, isnew);
 633                 if (*errorp) {
 634                         hammer_rel_buffer(buffer, 1);
 635                         buffer = NULL;
 636                 }
 637         } else {
 638                 *errorp = 0;
 639         }
 640         return(buffer);
 641 }
 642
 643 /*
 644  * Destroy all buffers covering the specified zoneX offset range.  This
 645  * is called when the related blockmap layer2 entry is freed or when
 646  * a direct write bypasses our buffer/buffer-cache subsystem.
 647  *
 648  * The buffers may be referenced by the caller itself.  Setting reclaim
 649  * will cause the buffer to be destroyed when it's ref count reaches zero.
 650  */
 651 void
 652 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 653                    hammer_off_t zone2_offset, int bytes)
 654 {
 655         hammer_buffer_t buffer;
 656         hammer_volume_t volume;
 657         int vol_no;
 658         int error;
 659
 660         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 661         volume = hammer_get_volume(hmp, vol_no, &error);
 662         KKASSERT(error == 0);
 663
 664         while (bytes > 0) {
 665                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 666                                    base_offset);
 667                 if (buffer) {
 668                         KKASSERT(buffer->zone2_offset == zone2_offset);
 669                         hammer_io_clear_modify(&buffer->io, 1);
 670                         buffer->io.reclaim = 1;
 671                         KKASSERT(buffer->volume == volume);
 672                         if (buffer->io.lock.refs == 0)
 673                                 hammer_unload_buffer(buffer, NULL);
 674                 } else {
 675                         hammer_io_inval(volume, zone2_offset);
 676                 }
 677                 base_offset += HAMMER_BUFSIZE;
 678                 zone2_offset += HAMMER_BUFSIZE;
 679                 bytes -= HAMMER_BUFSIZE;
 680         }
 681         hammer_rel_volume(volume, 0);
 682 }
 683
 684 static int
 685 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 686 {
 687         hammer_volume_t volume;
 688         int error;
 689
 690         /*
 691          * Load the buffer's on-disk info
 692          */
 693         volume = buffer->volume;
 694         ++buffer->io.loading;
 695         hammer_lock_ex(&buffer->io.lock);
 696
 697         if (hammer_debug_io & 0x0001) {
 698                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 699                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 700                         buffer->ondisk);
 701         }
 702
 703         if (buffer->ondisk == NULL) {
 704                 if (isnew) {
 705                         error = hammer_io_new(volume->devvp, &buffer->io);
 706                 } else {
 707                         error = hammer_io_read(volume->devvp, &buffer->io,
 708                                                volume->maxraw_off);
 709                 }
 710                 if (error == 0)
 711                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 712         } else if (isnew) {
 713                 error = hammer_io_new(volume->devvp, &buffer->io);
 714         } else {
 715                 error = 0;
 716         }
 717         --buffer->io.loading;
 718         hammer_unlock(&buffer->io.lock);
 719         return (error);
 720 }
 721
 722 /*
 723  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 724  */
 725 int
 726 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 727 {
 728         ++hammer_count_refedbufs;
 729         hammer_ref(&buffer->io.lock);
 730         hammer_flush_buffer_nodes(buffer);
 731         KKASSERT(buffer->io.lock.refs == 1);
 732         hammer_rel_buffer(buffer, 2);
 733         return(0);
 734 }
 735
 736 /*
 737  * Reference a buffer that is either already referenced or via a specially
 738  * handled pointer (aka cursor->buffer).
 739  */
 740 int
 741 hammer_ref_buffer(hammer_buffer_t buffer)
 742 {
 743         int error;
 744
 745         if (buffer->io.lock.refs == 0)
 746                 ++hammer_count_refedbufs;
 747         hammer_ref(&buffer->io.lock);
 748
 749         /*
 750          * At this point a biodone() will not touch the buffer other then
 751          * incidental bits.  However, lose_list can be modified via
 752          * a biodone() interrupt.
 753          *
 754          * No longer loose
 755          */
 756         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 757                 crit_enter();
 758                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 759                 buffer->io.mod_list = NULL;
 760                 crit_exit();
 761         }
 762
 763         if (buffer->ondisk == NULL || buffer->io.loading) {
 764                 error = hammer_load_buffer(buffer, 0);
 765                 if (error) {
 766                         hammer_rel_buffer(buffer, 1);
 767                         /*
 768                          * NOTE: buffer pointer can become stale after
 769                          * the above release.
 770                          */
 771                 }
 772         } else {
 773                 error = 0;
 774         }
 775         return(error);
 776 }
 777
 778 /*
 779  * Release a buffer.  We have to deal with several places where
 780  * another thread can ref the buffer.
 781  *
 782  * Only destroy the structure itself if the related buffer cache buffer
 783  * was disassociated from it.  This ties the management of the structure
 784  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 785  * embedded io is referenced or not.
 786  */
 787 void
 788 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 789 {
 790         hammer_volume_t volume;
 791         int freeme = 0;
 792
 793         crit_enter();
 794         if (buffer->io.lock.refs == 1) {
 795                 ++buffer->io.loading;   /* force interlock check */
 796                 hammer_lock_ex(&buffer->io.lock);
 797                 if (buffer->io.lock.refs == 1) {
 798                         hammer_io_release(&buffer->io, flush);
 799
 800                         if (buffer->io.lock.refs == 1)
 801                                 --hammer_count_refedbufs;
 802
 803                         if (buffer->io.bp == NULL &&
 804                             buffer->io.lock.refs == 1) {
 805                                 /*
 806                                  * Final cleanup
 807                                  *
 808                                  * NOTE: It is impossible for any associated
 809                                  * B-Tree nodes to have refs if the buffer
 810                                  * has no additional refs.
 811                                  */
 812                                 RB_REMOVE(hammer_buf_rb_tree,
 813                                           &buffer->io.hmp->rb_bufs_root,
 814                                           buffer);
 815                                 volume = buffer->volume;
 816                                 buffer->volume = NULL; /* sanity */
 817                                 hammer_rel_volume(volume, 0);
 818                                 hammer_io_clear_modlist(&buffer->io);
 819                                 hammer_flush_buffer_nodes(buffer);
 820                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 821                                 freeme = 1;
 822                         }
 823                 }
 824                 --buffer->io.loading;
 825                 hammer_unlock(&buffer->io.lock);
 826         }
 827         hammer_unref(&buffer->io.lock);
 828         crit_exit();
 829         if (freeme) {
 830                 --hammer_count_buffers;
 831                 kfree(buffer, M_HAMMER);
 832         }
 833 }
 834
 835 /*
 836  * Access the filesystem buffer containing the specified hammer offset.
 837  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 838  * relative buffer offset.  It must also have bit 55 set to be valid.
 839  * (see hammer_off_t in hammer_disk.h).
 840  *
 841  * Any prior buffer in *bufferp will be released and replaced by the
 842  * requested buffer.
 843  */
 844 static __inline
 845 void *
 846 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 847              int *errorp, struct hammer_buffer **bufferp)
 848 {
 849         hammer_buffer_t buffer;
 850         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 851
 852         buf_offset &= ~HAMMER_BUFMASK64;
 853         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 854
 855         buffer = *bufferp;
 856         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 857                                buffer->zoneX_offset != buf_offset)) {
 858                 if (buffer)
 859                         hammer_rel_buffer(buffer, 0);
 860                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 861                 *bufferp = buffer;
 862         } else {
 863                 *errorp = 0;
 864         }
 865
 866         /*
 867          * Return a pointer to the buffer data.
 868          */
 869         if (buffer == NULL)
 870                 return(NULL);
 871         else
 872                 return((char *)buffer->ondisk + xoff);
 873 }
 874
 875 void *
 876 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 877              int *errorp, struct hammer_buffer **bufferp)
 878 {
 879         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 880 }
 881
 882 void *
 883 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 884                  int *errorp, struct hammer_buffer **bufferp)
 885 {
 886         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 887         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 888 }
 889
 890 /*
 891  * Access the filesystem buffer containing the specified hammer offset.
 892  * No disk read operation occurs.  The result buffer may contain garbage.
 893  *
 894  * Any prior buffer in *bufferp will be released and replaced by the
 895  * requested buffer.
 896  *
 897  * This function marks the buffer dirty but does not increment its
 898  * modify_refs count.
 899  */
 900 static __inline
 901 void *
 902 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 903              int *errorp, struct hammer_buffer **bufferp)
 904 {
 905         hammer_buffer_t buffer;
 906         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 907
 908         buf_offset &= ~HAMMER_BUFMASK64;
 909
 910         buffer = *bufferp;
 911         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 912                                buffer->zoneX_offset != buf_offset)) {
 913                 if (buffer)
 914                         hammer_rel_buffer(buffer, 0);
 915                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 916                 *bufferp = buffer;
 917         } else {
 918                 *errorp = 0;
 919         }
 920
 921         /*
 922          * Return a pointer to the buffer data.
 923          */
 924         if (buffer == NULL)
 925                 return(NULL);
 926         else
 927                 return((char *)buffer->ondisk + xoff);
 928 }
 929
 930 void *
 931 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 932              int *errorp, struct hammer_buffer **bufferp)
 933 {
 934         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 935 }
 936
 937 void *
 938 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 939                 int *errorp, struct hammer_buffer **bufferp)
 940 {
 941         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 942         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
 943 }
 944
 945 /************************************************************************
 946  *                              NODES                                   *
 947  ************************************************************************
 948  *
 949  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 950  * method used by the HAMMER filesystem.
 951  *
 952  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 953  * associated with its buffer, and will only referenced the buffer while
 954  * the node itself is referenced.
 955  *
 956  * A hammer_node can also be passively associated with other HAMMER
 957  * structures, such as inodes, while retaining 0 references.  These
 958  * associations can be cleared backwards using a pointer-to-pointer in
 959  * the hammer_node.
 960  *
 961  * This allows the HAMMER implementation to cache hammer_nodes long-term
 962  * and short-cut a great deal of the infrastructure's complexity.  In
 963  * most cases a cached node can be reacquired without having to dip into
 964  * either the buffer or cluster management code.
 965  *
 966  * The caller must pass a referenced cluster on call and will retain
 967  * ownership of the reference on return.  The node will acquire its own
 968  * additional references, if necessary.
 969  */
 970 hammer_node_t
 971 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 972                 int isnew, int *errorp)
 973 {
 974         hammer_node_t node;
 975
 976         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 977
 978         /*
 979          * Locate the structure, allocating one if necessary.
 980          */
 981 again:
 982         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 983         if (node == NULL) {
 984                 ++hammer_count_nodes;
 985                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 986                 node->node_offset = node_offset;
 987                 node->hmp = hmp;
 988                 TAILQ_INIT(&node->cache_list);
 989                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 990                         --hammer_count_nodes;
 991                         kfree(node, M_HAMMER);
 992                         goto again;
 993                 }
 994         }
 995         hammer_ref(&node->lock);
 996         if (node->ondisk)
 997                 *errorp = 0;
 998         else
 999                 *errorp = hammer_load_node(node, isnew);
1000         if (*errorp) {
1001                 hammer_rel_node(node);
1002                 node = NULL;
1003         }
1004         return(node);
1005 }
1006
1007 /*
1008  * Reference an already-referenced node.
1009  */
1010 void
1011 hammer_ref_node(hammer_node_t node)
1012 {
1013         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1014         hammer_ref(&node->lock);
1015 }
1016
1017 /*
1018  * Load a node's on-disk data reference.
1019  */
1020 static int
1021 hammer_load_node(hammer_node_t node, int isnew)
1022 {
1023         hammer_buffer_t buffer;
1024         hammer_off_t buf_offset;
1025         int error;
1026
1027         error = 0;
1028         ++node->loading;
1029         hammer_lock_ex(&node->lock);
1030         if (node->ondisk == NULL) {
1031                 /*
1032                  * This is a little confusing but the jist is that
1033                  * node->buffer determines whether the node is on
1034                  * the buffer's clist and node->ondisk determines
1035                  * whether the buffer is referenced.
1036                  *
1037                  * We could be racing a buffer release, in which case
1038                  * node->buffer may become NULL while we are blocked
1039                  * referencing the buffer.
1040                  */
1041                 if ((buffer = node->buffer) != NULL) {
1042                         error = hammer_ref_buffer(buffer);
1043                         if (error == 0 && node->buffer == NULL) {
1044                                 TAILQ_INSERT_TAIL(&buffer->clist,
1045                                                   node, entry);
1046                                 node->buffer = buffer;
1047                         }
1048                 } else {
1049                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1050                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1051                                                    HAMMER_BUFSIZE, 0, &error);
1052                         if (buffer) {
1053                                 KKASSERT(error == 0);
1054                                 TAILQ_INSERT_TAIL(&buffer->clist,
1055                                                   node, entry);
1056                                 node->buffer = buffer;
1057                         }
1058                 }
1059                 if (error)
1060                         goto failed;
1061                 node->ondisk = (void *)((char *)buffer->ondisk +
1062                                         (node->node_offset & HAMMER_BUFMASK));
1063                 if (isnew == 0 &&
1064                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1065                         if (hammer_crc_test_btree(node->ondisk) == 0)
1066                                 Debugger("CRC FAILED: B-TREE NODE");
1067                         node->flags |= HAMMER_NODE_CRCGOOD;
1068                 }
1069         }
1070 failed:
1071         --node->loading;
1072         hammer_unlock(&node->lock);
1073         return (error);
1074 }
1075
1076 /*
1077  * Safely reference a node, interlock against flushes via the IO subsystem.
1078  */
1079 hammer_node_t
1080 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1081                      int *errorp)
1082 {
1083         hammer_node_t node;
1084
1085         node = cache->node;
1086         if (node != NULL) {
1087                 hammer_ref(&node->lock);
1088                 if (node->ondisk)
1089                         *errorp = 0;
1090                 else
1091                         *errorp = hammer_load_node(node, 0);
1092                 if (*errorp) {
1093                         hammer_rel_node(node);
1094                         node = NULL;
1095                 }
1096         } else {
1097                 *errorp = ENOENT;
1098         }
1099         return(node);
1100 }
1101
1102 /*
1103  * Release a hammer_node.  On the last release the node dereferences
1104  * its underlying buffer and may or may not be destroyed.
1105  */
1106 void
1107 hammer_rel_node(hammer_node_t node)
1108 {
1109         hammer_buffer_t buffer;
1110
1111         /*
1112          * If this isn't the last ref just decrement the ref count and
1113          * return.
1114          */
1115         if (node->lock.refs > 1) {
1116                 hammer_unref(&node->lock);
1117                 return;
1118         }
1119
1120         /*
1121          * If there is no ondisk info or no buffer the node failed to load,
1122          * remove the last reference and destroy the node.
1123          */
1124         if (node->ondisk == NULL) {
1125                 hammer_unref(&node->lock);
1126                 hammer_flush_node(node);
1127                 /* node is stale now */
1128                 return;
1129         }
1130
1131         /*
1132          * Do not disassociate the node from the buffer if it represents
1133          * a modified B-Tree node that still needs its crc to be generated.
1134          */
1135         if (node->flags & HAMMER_NODE_NEEDSCRC)
1136                 return;
1137
1138         /*
1139          * Do final cleanups and then either destroy the node and leave it
1140          * passively cached.  The buffer reference is removed regardless.
1141          */
1142         buffer = node->buffer;
1143         node->ondisk = NULL;
1144
1145         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1146                 hammer_unref(&node->lock);
1147                 hammer_rel_buffer(buffer, 0);
1148                 return;
1149         }
1150
1151         /*
1152          * Destroy the node.
1153          */
1154         hammer_unref(&node->lock);
1155         hammer_flush_node(node);
1156         /* node is stale */
1157         hammer_rel_buffer(buffer, 0);
1158 }
1159
1160 /*
1161  * Free space on-media associated with a B-Tree node.
1162  */
1163 void
1164 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1165 {
1166         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1167         node->flags |= HAMMER_NODE_DELETED;
1168         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1169 }
1170
1171 /*
1172  * Passively cache a referenced hammer_node.  The caller may release
1173  * the node on return.
1174  */
1175 void
1176 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1177 {
1178         /*
1179          * If the node is being deleted, don't cache it!
1180          */
1181         if (node->flags & HAMMER_NODE_DELETED)
1182                 return;
1183         if (cache->node == node)
1184                 return;
1185         while (cache->node)
1186                 hammer_uncache_node(cache);
1187         if (node->flags & HAMMER_NODE_DELETED)
1188                 return;
1189         cache->node = node;
1190         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1191 }
1192
1193 void
1194 hammer_uncache_node(hammer_node_cache_t cache)
1195 {
1196         hammer_node_t node;
1197
1198         if ((node = cache->node) != NULL) {
1199                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1200                 cache->node = NULL;
1201                 if (TAILQ_EMPTY(&node->cache_list))
1202                         hammer_flush_node(node);
1203         }
1204 }
1205
1206 /*
1207  * Remove a node's cache references and destroy the node if it has no
1208  * other references or backing store.
1209  */
1210 void
1211 hammer_flush_node(hammer_node_t node)
1212 {
1213         hammer_node_cache_t cache;
1214         hammer_buffer_t buffer;
1215
1216         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1217                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1218                 cache->node = NULL;
1219         }
1220         if (node->lock.refs == 0 && node->ondisk == NULL) {
1221                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1222                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1223                 if ((buffer = node->buffer) != NULL) {
1224                         node->buffer = NULL;
1225                         TAILQ_REMOVE(&buffer->clist, node, entry);
1226                         /* buffer is unreferenced because ondisk is NULL */
1227                 }
1228                 --hammer_count_nodes;
1229                 kfree(node, M_HAMMER);
1230         }
1231 }
1232
1233 /*
1234  * Flush passively cached B-Tree nodes associated with this buffer.
1235  * This is only called when the buffer is about to be destroyed, so
1236  * none of the nodes should have any references.  The buffer is locked.
1237  *
1238  * We may be interlocked with the buffer.
1239  */
1240 void
1241 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1242 {
1243         hammer_node_t node;
1244
1245         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1246                 KKASSERT(node->ondisk == NULL);
1247                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1248
1249                 if (node->lock.refs == 0) {
1250                         hammer_ref(&node->lock);
1251                         node->flags |= HAMMER_NODE_FLUSH;
1252                         hammer_rel_node(node);
1253                 } else {
1254                         KKASSERT(node->loading != 0);
1255                         KKASSERT(node->buffer != NULL);
1256                         buffer = node->buffer;
1257                         node->buffer = NULL;
1258                         TAILQ_REMOVE(&buffer->clist, node, entry);
1259                         /* buffer is unreferenced because ondisk is NULL */
1260                 }
1261         }
1262 }
1263
1264
1265 /************************************************************************
1266  *                              ALLOCATORS                              *
1267  ************************************************************************/
1268
1269 /*
1270  * Allocate a B-Tree node.
1271  */
1272 hammer_node_t
1273 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1274 {
1275         hammer_buffer_t buffer = NULL;
1276         hammer_node_t node = NULL;
1277         hammer_off_t node_offset;
1278
1279         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1280                                             sizeof(struct hammer_node_ondisk),
1281                                             errorp);
1282         if (*errorp == 0) {
1283                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1284                 hammer_modify_node_noundo(trans, node);
1285                 bzero(node->ondisk, sizeof(*node->ondisk));
1286                 hammer_modify_node_done(node);
1287         }
1288         if (buffer)
1289                 hammer_rel_buffer(buffer, 0);
1290         return(node);
1291 }
1292
1293 /*
1294  * Allocate data.  If the address of a data buffer is supplied then
1295  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1296  * will be set to the related buffer.  The caller must release it when
1297  * finally done.  The initial *data_bufferp should be set to NULL by
1298  * the caller.
1299  *
1300  * The caller is responsible for making hammer_modify*() calls on the
1301  * *data_bufferp.
1302  */
1303 void *
1304 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1305                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1306                   struct hammer_buffer **data_bufferp, int *errorp)
1307 {
1308         void *data;
1309         int zone;
1310
1311         /*
1312          * Allocate data
1313          */
1314         if (data_len) {
1315                 switch(rec_type) {
1316                 case HAMMER_RECTYPE_INODE:
1317                 case HAMMER_RECTYPE_PSEUDO_INODE:
1318                 case HAMMER_RECTYPE_DIRENTRY:
1319                 case HAMMER_RECTYPE_EXT:
1320                 case HAMMER_RECTYPE_FIX:
1321                         zone = HAMMER_ZONE_META_INDEX;
1322                         break;
1323                 case HAMMER_RECTYPE_DATA:
1324                 case HAMMER_RECTYPE_DB:
1325                         if (data_len <= HAMMER_BUFSIZE / 2) {
1326                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1327                         } else {
1328                                 data_len = (data_len + HAMMER_BUFMASK) &
1329                                            ~HAMMER_BUFMASK;
1330                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1331                         }
1332                         break;
1333                 default:
1334                         panic("hammer_alloc_data: rec_type %04x unknown",
1335                               rec_type);
1336                         zone = 0;       /* NOT REACHED */
1337                         break;
1338                 }
1339                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1340                                                       data_len, errorp);
1341         } else {
1342                 *data_offsetp = 0;
1343         }
1344         if (*errorp == 0 && data_bufferp) {
1345                 if (data_len) {
1346                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1347                                                 data_len, errorp, data_bufferp);
1348                         KKASSERT(*errorp == 0);
1349                 } else {
1350                         data = NULL;
1351                 }
1352         } else {
1353                 data = NULL;
1354         }
1355         KKASSERT(*errorp == 0);
1356         return(data);
1357 }
1358
1359 /*
1360  * Sync dirty buffers to the media and clean-up any loose ends.
1361  */
1362 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1363 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1364
1365 int
1366 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1367 {
1368         struct hammer_sync_info info;
1369
1370         info.error = 0;
1371         info.waitfor = waitfor;
1372         if (waitfor == MNT_WAIT) {
1373                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1374                               hammer_sync_scan1, hammer_sync_scan2, &info);
1375         } else {
1376                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1377                               hammer_sync_scan1, hammer_sync_scan2, &info);
1378         }
1379         return(info.error);
1380 }
1381
1382 int
1383 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1384 {
1385         struct hammer_sync_info info;
1386
1387         info.error = 0;
1388         info.waitfor = waitfor;
1389
1390         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1391                       hammer_sync_scan1, hammer_sync_scan2, &info);
1392         if (waitfor == MNT_WAIT)
1393                 hammer_flusher_sync(hmp);
1394         else
1395                 hammer_flusher_async(hmp);
1396
1397         return(info.error);
1398 }
1399
1400 static int
1401 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1402 {
1403         struct hammer_inode *ip;
1404
1405         ip = VTOI(vp);
1406         if (vp->v_type == VNON || ip == NULL ||
1407             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1408              RB_EMPTY(&vp->v_rbdirty_tree))) {
1409                 return(-1);
1410         }
1411         return(0);
1412 }
1413
1414 static int
1415 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1416 {
1417         struct hammer_sync_info *info = data;
1418         struct hammer_inode *ip;
1419         int error;
1420
1421         ip = VTOI(vp);
1422         if (vp->v_type == VNON || vp->v_type == VBAD ||
1423             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1424              RB_EMPTY(&vp->v_rbdirty_tree))) {
1425                 return(0);
1426         }
1427         error = VOP_FSYNC(vp, info->waitfor);
1428         if (error)
1429                 info->error = error;
1430         return(0);
1431 }
1432