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 1.61 2008/06/20 21:24:53 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_localization < ip2->obj_localization)
  60                 return(-1);
  61         if (ip1->obj_localization > ip2->obj_localization)
  62                 return(1);
  63         if (ip1->obj_id < ip2->obj_id)
  64                 return(-1);
  65         if (ip1->obj_id > ip2->obj_id)
  66                 return(1);
  67         if (ip1->obj_asof < ip2->obj_asof)
  68                 return(-1);
  69         if (ip1->obj_asof > ip2->obj_asof)
  70                 return(1);
  71         return(0);
  72 }
  73
  74 static int
  75 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  76 {
  77         if (info->obj_localization < ip->obj_localization)
  78                 return(-1);
  79         if (info->obj_localization > ip->obj_localization)
  80                 return(1);
  81         if (info->obj_id < ip->obj_id)
  82                 return(-1);
  83         if (info->obj_id > ip->obj_id)
  84                 return(1);
  85         if (info->obj_asof < ip->obj_asof)
  86                 return(-1);
  87         if (info->obj_asof > ip->obj_asof)
  88                 return(1);
  89         return(0);
  90 }
  91
  92 static int
  93 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  94 {
  95         if (vol1->vol_no < vol2->vol_no)
  96                 return(-1);
  97         if (vol1->vol_no > vol2->vol_no)
  98                 return(1);
  99         return(0);
 100 }
 101
 102 static int
 103 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
 104 {
 105         if (buf1->zoneX_offset < buf2->zoneX_offset)
 106                 return(-1);
 107         if (buf1->zoneX_offset > buf2->zoneX_offset)
 108                 return(1);
 109         return(0);
 110 }
 111
 112 static int
 113 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 114 {
 115         if (node1->node_offset < node2->node_offset)
 116                 return(-1);
 117         if (node1->node_offset > node2->node_offset)
 118                 return(1);
 119         return(0);
 120 }
 121
 122 /*
 123  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 124  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 125  * functions are normal, e.g. hammer_buf_rb_tree_RB_LOOKUP(root, zone2_offset).
 126  */
 127 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 128 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 129                 hammer_inode_info_cmp, hammer_inode_info_t);
 130 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 131              hammer_vol_rb_compare, int32_t, vol_no);
 132 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 133              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
 134 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 135              hammer_nod_rb_compare, hammer_off_t, node_offset);
 136
 137 /************************************************************************
 138  *                              VOLUMES                                 *
 139  ************************************************************************
 140  *
 141  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 142  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 143  * not load a new volume.
 144  *
 145  * Calls made to hammer_load_volume() or single-threaded
 146  */
 147 int
 148 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 149 {
 150         struct mount *mp;
 151         hammer_volume_t volume;
 152         struct hammer_volume_ondisk *ondisk;
 153         struct nlookupdata nd;
 154         struct buf *bp = NULL;
 155         int error;
 156         int ronly;
 157         int setmp = 0;
 158
 159         mp = hmp->mp;
 160         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 161
 162         /*
 163          * Allocate a volume structure
 164          */
 165         ++hammer_count_volumes;
 166         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 167         volume->vol_name = kstrdup(volname, M_HAMMER);
 168         hammer_io_init(&volume->io, hmp, HAMMER_STRUCTURE_VOLUME);
 169         volume->io.offset = 0LL;
 170         volume->io.bytes = HAMMER_BUFSIZE;
 171
 172         /*
 173          * Get the device vnode
 174          */
 175         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 176         if (error == 0)
 177                 error = nlookup(&nd);
 178         if (error == 0)
 179                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 180         nlookup_done(&nd);
 181         if (error == 0) {
 182                 if (vn_isdisk(volume->devvp, &error)) {
 183                         error = vfs_mountedon(volume->devvp);
 184                 }
 185         }
 186         if (error == 0 &&
 187             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 188                 error = EBUSY;
 189         }
 190         if (error == 0) {
 191                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 192                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 193                 if (error == 0) {
 194                         error = VOP_OPEN(volume->devvp,
 195                                          (ronly ? FREAD : FREAD|FWRITE),
 196                                          FSCRED, NULL);
 197                 }
 198                 vn_unlock(volume->devvp);
 199         }
 200         if (error) {
 201                 hammer_free_volume(volume);
 202                 return(error);
 203         }
 204         volume->devvp->v_rdev->si_mountpoint = mp;
 205         setmp = 1;
 206
 207         /*
 208          * Extract the volume number from the volume header and do various
 209          * sanity checks.
 210          */
 211         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 212         if (error)
 213                 goto late_failure;
 214         ondisk = (void *)bp->b_data;
 215         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 216                 kprintf("hammer_mount: volume %s has an invalid header\n",
 217                         volume->vol_name);
 218                 error = EFTYPE;
 219                 goto late_failure;
 220         }
 221         volume->vol_no = ondisk->vol_no;
 222         volume->buffer_base = ondisk->vol_buf_beg;
 223         volume->vol_flags = ondisk->vol_flags;
 224         volume->nblocks = ondisk->vol_nblocks;
 225         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 226                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 227         volume->maxraw_off = ondisk->vol_buf_end;
 228
 229         if (RB_EMPTY(&hmp->rb_vols_root)) {
 230                 hmp->fsid = ondisk->vol_fsid;
 231         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 232                 kprintf("hammer_mount: volume %s's fsid does not match "
 233                         "other volumes\n", volume->vol_name);
 234                 error = EFTYPE;
 235                 goto late_failure;
 236         }
 237
 238         /*
 239          * Insert the volume structure into the red-black tree.
 240          */
 241         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 242                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 243                         volume->vol_name, volume->vol_no);
 244                 error = EEXIST;
 245         }
 246
 247         /*
 248          * Set the root volume .  HAMMER special cases rootvol the structure.
 249          * We do not hold a ref because this would prevent related I/O
 250          * from being flushed.
 251          */
 252         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 253                 hmp->rootvol = volume;
 254                 if (bp) {
 255                         brelse(bp);
 256                         bp = NULL;
 257                 }
 258                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 259                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 260                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 261                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 262                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 263         }
 264 late_failure:
 265         if (bp)
 266                 brelse(bp);
 267         if (error) {
 268                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 269                 if (setmp)
 270                         volume->devvp->v_rdev->si_mountpoint = NULL;
 271                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 272                 hammer_free_volume(volume);
 273         }
 274         return (error);
 275 }
 276
 277 /*
 278  * This is called for each volume when updating the mount point from
 279  * read-write to read-only or vise-versa.
 280  */
 281 int
 282 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 283 {
 284         if (volume->devvp) {
 285                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 286                 if (volume->io.hmp->ronly) {
 287                         /* do not call vinvalbuf */
 288                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 289                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 290                 } else {
 291                         /* do not call vinvalbuf */
 292                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 293                         VOP_CLOSE(volume->devvp, FREAD);
 294                 }
 295                 vn_unlock(volume->devvp);
 296         }
 297         return(0);
 298 }
 299
 300 /*
 301  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 302  * so returns -1 on failure.
 303  */
 304 int
 305 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 306 {
 307         struct hammer_mount *hmp = volume->io.hmp;
 308         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 309
 310         /*
 311          * Clean up the root volume pointer, which is held unlocked in hmp.
 312          */
 313         if (hmp->rootvol == volume)
 314                 hmp->rootvol = NULL;
 315
 316         /*
 317          * Release our buffer and flush anything left in the buffer cache.
 318          */
 319         volume->io.waitdep = 1;
 320         hammer_io_release(&volume->io, 1);
 321         hammer_io_clear_modlist(&volume->io);
 322
 323         /*
 324          * There should be no references on the volume, no clusters, and
 325          * no super-clusters.
 326          */
 327         KKASSERT(volume->io.lock.refs == 0);
 328
 329         volume->ondisk = NULL;
 330         if (volume->devvp) {
 331                 if (volume->devvp->v_rdev &&
 332                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 333                 ) {
 334                         volume->devvp->v_rdev->si_mountpoint = NULL;
 335                 }
 336                 if (ronly) {
 337                         vinvalbuf(volume->devvp, 0, 0, 0);
 338                         VOP_CLOSE(volume->devvp, FREAD);
 339                 } else {
 340                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 341                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 342                 }
 343         }
 344
 345         /*
 346          * Destroy the structure
 347          */
 348         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 349         hammer_free_volume(volume);
 350         return(0);
 351 }
 352
 353 static
 354 void
 355 hammer_free_volume(hammer_volume_t volume)
 356 {
 357         if (volume->vol_name) {
 358                 kfree(volume->vol_name, M_HAMMER);
 359                 volume->vol_name = NULL;
 360         }
 361         if (volume->devvp) {
 362                 vrele(volume->devvp);
 363                 volume->devvp = NULL;
 364         }
 365         --hammer_count_volumes;
 366         kfree(volume, M_HAMMER);
 367 }
 368
 369 /*
 370  * Get a HAMMER volume.  The volume must already exist.
 371  */
 372 hammer_volume_t
 373 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 374 {
 375         struct hammer_volume *volume;
 376
 377         /*
 378          * Locate the volume structure
 379          */
 380         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 381         if (volume == NULL) {
 382                 *errorp = ENOENT;
 383                 return(NULL);
 384         }
 385         hammer_ref(&volume->io.lock);
 386
 387         /*
 388          * Deal with on-disk info
 389          */
 390         if (volume->ondisk == NULL || volume->io.loading) {
 391                 *errorp = hammer_load_volume(volume);
 392                 if (*errorp) {
 393                         hammer_rel_volume(volume, 1);
 394                         volume = NULL;
 395                 }
 396         } else {
 397                 *errorp = 0;
 398         }
 399         return(volume);
 400 }
 401
 402 int
 403 hammer_ref_volume(hammer_volume_t volume)
 404 {
 405         int error;
 406
 407         hammer_ref(&volume->io.lock);
 408
 409         /*
 410          * Deal with on-disk info
 411          */
 412         if (volume->ondisk == NULL || volume->io.loading) {
 413                 error = hammer_load_volume(volume);
 414                 if (error)
 415                         hammer_rel_volume(volume, 1);
 416         } else {
 417                 error = 0;
 418         }
 419         return (error);
 420 }
 421
 422 hammer_volume_t
 423 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 424 {
 425         hammer_volume_t volume;
 426
 427         volume = hmp->rootvol;
 428         KKASSERT(volume != NULL);
 429         hammer_ref(&volume->io.lock);
 430
 431         /*
 432          * Deal with on-disk info
 433          */
 434         if (volume->ondisk == NULL || volume->io.loading) {
 435                 *errorp = hammer_load_volume(volume);
 436                 if (*errorp) {
 437                         hammer_rel_volume(volume, 1);
 438                         volume = NULL;
 439                 }
 440         } else {
 441                 *errorp = 0;
 442         }
 443         return (volume);
 444 }
 445
 446 /*
 447  * Load a volume's on-disk information.  The volume must be referenced and
 448  * not locked.  We temporarily acquire an exclusive lock to interlock
 449  * against releases or multiple get's.
 450  */
 451 static int
 452 hammer_load_volume(hammer_volume_t volume)
 453 {
 454         int error;
 455
 456         ++volume->io.loading;
 457         hammer_lock_ex(&volume->io.lock);
 458
 459         if (volume->ondisk == NULL) {
 460                 error = hammer_io_read(volume->devvp, &volume->io,
 461                                        volume->maxraw_off);
 462                 if (error == 0)
 463                         volume->ondisk = (void *)volume->io.bp->b_data;
 464         } else {
 465                 error = 0;
 466         }
 467         --volume->io.loading;
 468         hammer_unlock(&volume->io.lock);
 469         return(error);
 470 }
 471
 472 /*
 473  * Release a volume.  Call hammer_io_release on the last reference.  We have
 474  * to acquire an exclusive lock to interlock against volume->ondisk tests
 475  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 476  * lock to be held.
 477  *
 478  * Volumes are not unloaded from memory during normal operation.
 479  */
 480 void
 481 hammer_rel_volume(hammer_volume_t volume, int flush)
 482 {
 483         crit_enter();
 484         if (volume->io.lock.refs == 1) {
 485                 ++volume->io.loading;
 486                 hammer_lock_ex(&volume->io.lock);
 487                 if (volume->io.lock.refs == 1) {
 488                         volume->ondisk = NULL;
 489                         hammer_io_release(&volume->io, flush);
 490                 }
 491                 --volume->io.loading;
 492                 hammer_unlock(&volume->io.lock);
 493         }
 494         hammer_unref(&volume->io.lock);
 495         crit_exit();
 496 }
 497
 498 /************************************************************************
 499  *                              BUFFERS                                 *
 500  ************************************************************************
 501  *
 502  * Manage buffers.  Currently all blockmap-backed zones are translated
 503  * to zone-2 buffer offsets.
 504  */
 505 hammer_buffer_t
 506 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 507                   int bytes, int isnew, int *errorp)
 508 {
 509         hammer_buffer_t buffer;
 510         hammer_volume_t volume;
 511         hammer_off_t    zone2_offset;
 512         hammer_io_type_t iotype;
 513         int vol_no;
 514         int zone;
 515
 516         buf_offset &= ~HAMMER_BUFMASK64;
 517 again:
 518         /*
 519          * Shortcut if the buffer is already cached
 520          */
 521         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 522         if (buffer) {
 523                 if (buffer->io.lock.refs == 0)
 524                         ++hammer_count_refedbufs;
 525                 hammer_ref(&buffer->io.lock);
 526
 527                 /*
 528                  * Onced refed the ondisk field will not be cleared by
 529                  * any other action.
 530                  */
 531                 if (buffer->ondisk && buffer->io.loading == 0) {
 532                         *errorp = 0;
 533                         return(buffer);
 534                 }
 535
 536                 /*
 537                  * The buffer is no longer loose if it has a ref, and
 538                  * cannot become loose once it gains a ref.  Loose
 539                  * buffers will never be in a modified state.  This should
 540                  * only occur on the 0->1 transition of refs.
 541                  *
 542                  * lose_list can be modified via a biodone() interrupt.
 543                  */
 544                 if (buffer->io.mod_list == &hmp->lose_list) {
 545                         crit_enter();   /* biodone race against list */
 546                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 547                                      mod_entry);
 548                         crit_exit();
 549                         buffer->io.mod_list = NULL;
 550                         KKASSERT(buffer->io.modified == 0);
 551                 }
 552                 goto found;
 553         }
 554
 555         /*
 556          * What is the buffer class?
 557          */
 558         zone = HAMMER_ZONE_DECODE(buf_offset);
 559
 560         switch(zone) {
 561         case HAMMER_ZONE_LARGE_DATA_INDEX:
 562         case HAMMER_ZONE_SMALL_DATA_INDEX:
 563                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 564                 break;
 565         case HAMMER_ZONE_UNDO_INDEX:
 566                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 567                 break;
 568         case HAMMER_ZONE_META_INDEX:
 569         default:
 570                 /*
 571                  * NOTE: inode data and directory entries are placed in this
 572                  * zone.  inode atime/mtime is updated in-place and thus
 573                  * buffers containing inodes must be synchronized as
 574                  * meta-buffers, same as buffers containing B-Tree info.
 575                  */
 576                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 577                 break;
 578         }
 579
 580         /*
 581          * Handle blockmap offset translations
 582          */
 583         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 584                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 585         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 586                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 587         } else {
 588                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 589                 zone2_offset = buf_offset;
 590                 *errorp = 0;
 591         }
 592         if (*errorp)
 593                 return(NULL);
 594
 595         /*
 596          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 597          * specifications.
 598          */
 599         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 600                  HAMMER_ZONE_RAW_BUFFER);
 601         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 602         volume = hammer_get_volume(hmp, vol_no, errorp);
 603         if (volume == NULL)
 604                 return(NULL);
 605
 606         KKASSERT(zone2_offset < volume->maxbuf_off);
 607
 608         /*
 609          * Allocate a new buffer structure.  We will check for races later.
 610          */
 611         ++hammer_count_buffers;
 612         buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 613         buffer->zone2_offset = zone2_offset;
 614         buffer->zoneX_offset = buf_offset;
 615         buffer->volume = volume;
 616
 617         hammer_io_init(&buffer->io, hmp, iotype);
 618         buffer->io.offset = volume->ondisk->vol_buf_beg +
 619                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 620         buffer->io.bytes = bytes;
 621         TAILQ_INIT(&buffer->clist);
 622         hammer_ref(&buffer->io.lock);
 623
 624         /*
 625          * Insert the buffer into the RB tree and handle late collisions.
 626          */
 627         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 628                 hammer_unref(&buffer->io.lock);
 629                 --hammer_count_buffers;
 630                 kfree(buffer, M_HAMMER);
 631                 goto again;
 632         }
 633         ++hammer_count_refedbufs;
 634 found:
 635
 636         /*
 637          * Deal with on-disk info and loading races.
 638          */
 639         if (buffer->ondisk == NULL || buffer->io.loading) {
 640                 *errorp = hammer_load_buffer(buffer, isnew);
 641                 if (*errorp) {
 642                         hammer_rel_buffer(buffer, 1);
 643                         buffer = NULL;
 644                 }
 645         } else {
 646                 *errorp = 0;
 647         }
 648         return(buffer);
 649 }
 650
 651 /*
 652  * Destroy all buffers covering the specified zoneX offset range.  This
 653  * is called when the related blockmap layer2 entry is freed or when
 654  * a direct write bypasses our buffer/buffer-cache subsystem.
 655  *
 656  * The buffers may be referenced by the caller itself.  Setting reclaim
 657  * will cause the buffer to be destroyed when it's ref count reaches zero.
 658  */
 659 void
 660 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 661                    hammer_off_t zone2_offset, int bytes)
 662 {
 663         hammer_buffer_t buffer;
 664         hammer_volume_t volume;
 665         int vol_no;
 666         int error;
 667
 668         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 669         volume = hammer_get_volume(hmp, vol_no, &error);
 670         KKASSERT(error == 0);
 671
 672         while (bytes > 0) {
 673                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 674                                    base_offset);
 675                 if (buffer) {
 676                         KKASSERT(buffer->zone2_offset == zone2_offset);
 677                         hammer_io_clear_modify(&buffer->io, 1);
 678                         buffer->io.reclaim = 1;
 679                         KKASSERT(buffer->volume == volume);
 680                         if (buffer->io.lock.refs == 0)
 681                                 hammer_unload_buffer(buffer, NULL);
 682                 } else {
 683                         hammer_io_inval(volume, zone2_offset);
 684                 }
 685                 base_offset += HAMMER_BUFSIZE;
 686                 zone2_offset += HAMMER_BUFSIZE;
 687                 bytes -= HAMMER_BUFSIZE;
 688         }
 689         hammer_rel_volume(volume, 0);
 690 }
 691
 692 static int
 693 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 694 {
 695         hammer_volume_t volume;
 696         int error;
 697
 698         /*
 699          * Load the buffer's on-disk info
 700          */
 701         volume = buffer->volume;
 702         ++buffer->io.loading;
 703         hammer_lock_ex(&buffer->io.lock);
 704
 705         if (hammer_debug_io & 0x0001) {
 706                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 707                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 708                         buffer->ondisk);
 709         }
 710
 711         if (buffer->ondisk == NULL) {
 712                 if (isnew) {
 713                         error = hammer_io_new(volume->devvp, &buffer->io);
 714                 } else {
 715                         error = hammer_io_read(volume->devvp, &buffer->io,
 716                                                volume->maxraw_off);
 717                 }
 718                 if (error == 0)
 719                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 720         } else if (isnew) {
 721                 error = hammer_io_new(volume->devvp, &buffer->io);
 722         } else {
 723                 error = 0;
 724         }
 725         --buffer->io.loading;
 726         hammer_unlock(&buffer->io.lock);
 727         return (error);
 728 }
 729
 730 /*
 731  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 732  */
 733 int
 734 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 735 {
 736         ++hammer_count_refedbufs;
 737         hammer_ref(&buffer->io.lock);
 738         hammer_flush_buffer_nodes(buffer);
 739         KKASSERT(buffer->io.lock.refs == 1);
 740         hammer_rel_buffer(buffer, 2);
 741         return(0);
 742 }
 743
 744 /*
 745  * Reference a buffer that is either already referenced or via a specially
 746  * handled pointer (aka cursor->buffer).
 747  */
 748 int
 749 hammer_ref_buffer(hammer_buffer_t buffer)
 750 {
 751         int error;
 752
 753         if (buffer->io.lock.refs == 0)
 754                 ++hammer_count_refedbufs;
 755         hammer_ref(&buffer->io.lock);
 756
 757         /*
 758          * At this point a biodone() will not touch the buffer other then
 759          * incidental bits.  However, lose_list can be modified via
 760          * a biodone() interrupt.
 761          *
 762          * No longer loose
 763          */
 764         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 765                 crit_enter();
 766                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 767                 buffer->io.mod_list = NULL;
 768                 crit_exit();
 769         }
 770
 771         if (buffer->ondisk == NULL || buffer->io.loading) {
 772                 error = hammer_load_buffer(buffer, 0);
 773                 if (error) {
 774                         hammer_rel_buffer(buffer, 1);
 775                         /*
 776                          * NOTE: buffer pointer can become stale after
 777                          * the above release.
 778                          */
 779                 }
 780         } else {
 781                 error = 0;
 782         }
 783         return(error);
 784 }
 785
 786 /*
 787  * Release a buffer.  We have to deal with several places where
 788  * another thread can ref the buffer.
 789  *
 790  * Only destroy the structure itself if the related buffer cache buffer
 791  * was disassociated from it.  This ties the management of the structure
 792  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 793  * embedded io is referenced or not.
 794  */
 795 void
 796 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 797 {
 798         hammer_volume_t volume;
 799         int freeme = 0;
 800
 801         crit_enter();
 802         if (buffer->io.lock.refs == 1) {
 803                 ++buffer->io.loading;   /* force interlock check */
 804                 hammer_lock_ex(&buffer->io.lock);
 805                 if (buffer->io.lock.refs == 1) {
 806                         hammer_io_release(&buffer->io, flush);
 807
 808                         if (buffer->io.lock.refs == 1)
 809                                 --hammer_count_refedbufs;
 810
 811                         if (buffer->io.bp == NULL &&
 812                             buffer->io.lock.refs == 1) {
 813                                 /*
 814                                  * Final cleanup
 815                                  *
 816                                  * NOTE: It is impossible for any associated
 817                                  * B-Tree nodes to have refs if the buffer
 818                                  * has no additional refs.
 819                                  */
 820                                 RB_REMOVE(hammer_buf_rb_tree,
 821                                           &buffer->io.hmp->rb_bufs_root,
 822                                           buffer);
 823                                 volume = buffer->volume;
 824                                 buffer->volume = NULL; /* sanity */
 825                                 hammer_rel_volume(volume, 0);
 826                                 hammer_io_clear_modlist(&buffer->io);
 827                                 hammer_flush_buffer_nodes(buffer);
 828                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 829                                 freeme = 1;
 830                         }
 831                 }
 832                 --buffer->io.loading;
 833                 hammer_unlock(&buffer->io.lock);
 834         }
 835         hammer_unref(&buffer->io.lock);
 836         crit_exit();
 837         if (freeme) {
 838                 --hammer_count_buffers;
 839                 kfree(buffer, M_HAMMER);
 840         }
 841 }
 842
 843 /*
 844  * Access the filesystem buffer containing the specified hammer offset.
 845  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 846  * relative buffer offset.  It must also have bit 55 set to be valid.
 847  * (see hammer_off_t in hammer_disk.h).
 848  *
 849  * Any prior buffer in *bufferp will be released and replaced by the
 850  * requested buffer.
 851  */
 852 static __inline
 853 void *
 854 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 855              int *errorp, struct hammer_buffer **bufferp)
 856 {
 857         hammer_buffer_t buffer;
 858         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 859
 860         buf_offset &= ~HAMMER_BUFMASK64;
 861         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 862
 863         buffer = *bufferp;
 864         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 865                                buffer->zoneX_offset != buf_offset)) {
 866                 if (buffer)
 867                         hammer_rel_buffer(buffer, 0);
 868                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 869                 *bufferp = buffer;
 870         } else {
 871                 *errorp = 0;
 872         }
 873
 874         /*
 875          * Return a pointer to the buffer data.
 876          */
 877         if (buffer == NULL)
 878                 return(NULL);
 879         else
 880                 return((char *)buffer->ondisk + xoff);
 881 }
 882
 883 void *
 884 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 885              int *errorp, struct hammer_buffer **bufferp)
 886 {
 887         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 888 }
 889
 890 void *
 891 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 892                  int *errorp, struct hammer_buffer **bufferp)
 893 {
 894         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 895         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 896 }
 897
 898 /*
 899  * Access the filesystem buffer containing the specified hammer offset.
 900  * No disk read operation occurs.  The result buffer may contain garbage.
 901  *
 902  * Any prior buffer in *bufferp will be released and replaced by the
 903  * requested buffer.
 904  *
 905  * This function marks the buffer dirty but does not increment its
 906  * modify_refs count.
 907  */
 908 static __inline
 909 void *
 910 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 911              int *errorp, struct hammer_buffer **bufferp)
 912 {
 913         hammer_buffer_t buffer;
 914         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 915
 916         buf_offset &= ~HAMMER_BUFMASK64;
 917
 918         buffer = *bufferp;
 919         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 920                                buffer->zoneX_offset != buf_offset)) {
 921                 if (buffer)
 922                         hammer_rel_buffer(buffer, 0);
 923                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 924                 *bufferp = buffer;
 925         } else {
 926                 *errorp = 0;
 927         }
 928
 929         /*
 930          * Return a pointer to the buffer data.
 931          */
 932         if (buffer == NULL)
 933                 return(NULL);
 934         else
 935                 return((char *)buffer->ondisk + xoff);
 936 }
 937
 938 void *
 939 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
 940              int *errorp, struct hammer_buffer **bufferp)
 941 {
 942         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 943 }
 944
 945 void *
 946 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 947                 int *errorp, struct hammer_buffer **bufferp)
 948 {
 949         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 950         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
 951 }
 952
 953 /************************************************************************
 954  *                              NODES                                   *
 955  ************************************************************************
 956  *
 957  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 958  * method used by the HAMMER filesystem.
 959  *
 960  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 961  * associated with its buffer, and will only referenced the buffer while
 962  * the node itself is referenced.
 963  *
 964  * A hammer_node can also be passively associated with other HAMMER
 965  * structures, such as inodes, while retaining 0 references.  These
 966  * associations can be cleared backwards using a pointer-to-pointer in
 967  * the hammer_node.
 968  *
 969  * This allows the HAMMER implementation to cache hammer_nodes long-term
 970  * and short-cut a great deal of the infrastructure's complexity.  In
 971  * most cases a cached node can be reacquired without having to dip into
 972  * either the buffer or cluster management code.
 973  *
 974  * The caller must pass a referenced cluster on call and will retain
 975  * ownership of the reference on return.  The node will acquire its own
 976  * additional references, if necessary.
 977  */
 978 hammer_node_t
 979 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 980                 int isnew, int *errorp)
 981 {
 982         hammer_node_t node;
 983
 984         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 985
 986         /*
 987          * Locate the structure, allocating one if necessary.
 988          */
 989 again:
 990         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 991         if (node == NULL) {
 992                 ++hammer_count_nodes;
 993                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 994                 node->node_offset = node_offset;
 995                 node->hmp = hmp;
 996                 TAILQ_INIT(&node->cache_list);
 997                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 998                         --hammer_count_nodes;
 999                         kfree(node, M_HAMMER);
1000                         goto again;
1001                 }
1002         }
1003         hammer_ref(&node->lock);
1004         if (node->ondisk)
1005                 *errorp = 0;
1006         else
1007                 *errorp = hammer_load_node(node, isnew);
1008         if (*errorp) {
1009                 hammer_rel_node(node);
1010                 node = NULL;
1011         }
1012         return(node);
1013 }
1014
1015 /*
1016  * Reference an already-referenced node.
1017  */
1018 void
1019 hammer_ref_node(hammer_node_t node)
1020 {
1021         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1022         hammer_ref(&node->lock);
1023 }
1024
1025 /*
1026  * Load a node's on-disk data reference.
1027  */
1028 static int
1029 hammer_load_node(hammer_node_t node, int isnew)
1030 {
1031         hammer_buffer_t buffer;
1032         hammer_off_t buf_offset;
1033         int error;
1034
1035         error = 0;
1036         ++node->loading;
1037         hammer_lock_ex(&node->lock);
1038         if (node->ondisk == NULL) {
1039                 /*
1040                  * This is a little confusing but the jist is that
1041                  * node->buffer determines whether the node is on
1042                  * the buffer's clist and node->ondisk determines
1043                  * whether the buffer is referenced.
1044                  *
1045                  * We could be racing a buffer release, in which case
1046                  * node->buffer may become NULL while we are blocked
1047                  * referencing the buffer.
1048                  */
1049                 if ((buffer = node->buffer) != NULL) {
1050                         error = hammer_ref_buffer(buffer);
1051                         if (error == 0 && node->buffer == NULL) {
1052                                 TAILQ_INSERT_TAIL(&buffer->clist,
1053                                                   node, entry);
1054                                 node->buffer = buffer;
1055                         }
1056                 } else {
1057                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1058                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1059                                                    HAMMER_BUFSIZE, 0, &error);
1060                         if (buffer) {
1061                                 KKASSERT(error == 0);
1062                                 TAILQ_INSERT_TAIL(&buffer->clist,
1063                                                   node, entry);
1064                                 node->buffer = buffer;
1065                         }
1066                 }
1067                 if (error)
1068                         goto failed;
1069                 node->ondisk = (void *)((char *)buffer->ondisk +
1070                                         (node->node_offset & HAMMER_BUFMASK));
1071                 if (isnew == 0 &&
1072                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1073                         if (hammer_crc_test_btree(node->ondisk) == 0)
1074                                 Debugger("CRC FAILED: B-TREE NODE");
1075                         node->flags |= HAMMER_NODE_CRCGOOD;
1076                 }
1077         }
1078 failed:
1079         --node->loading;
1080         hammer_unlock(&node->lock);
1081         return (error);
1082 }
1083
1084 /*
1085  * Safely reference a node, interlock against flushes via the IO subsystem.
1086  */
1087 hammer_node_t
1088 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1089                      int *errorp)
1090 {
1091         hammer_node_t node;
1092
1093         node = cache->node;
1094         if (node != NULL) {
1095                 hammer_ref(&node->lock);
1096                 if (node->ondisk)
1097                         *errorp = 0;
1098                 else
1099                         *errorp = hammer_load_node(node, 0);
1100                 if (*errorp) {
1101                         hammer_rel_node(node);
1102                         node = NULL;
1103                 }
1104         } else {
1105                 *errorp = ENOENT;
1106         }
1107         return(node);
1108 }
1109
1110 /*
1111  * Release a hammer_node.  On the last release the node dereferences
1112  * its underlying buffer and may or may not be destroyed.
1113  */
1114 void
1115 hammer_rel_node(hammer_node_t node)
1116 {
1117         hammer_buffer_t buffer;
1118
1119         /*
1120          * If this isn't the last ref just decrement the ref count and
1121          * return.
1122          */
1123         if (node->lock.refs > 1) {
1124                 hammer_unref(&node->lock);
1125                 return;
1126         }
1127
1128         /*
1129          * If there is no ondisk info or no buffer the node failed to load,
1130          * remove the last reference and destroy the node.
1131          */
1132         if (node->ondisk == NULL) {
1133                 hammer_unref(&node->lock);
1134                 hammer_flush_node(node);
1135                 /* node is stale now */
1136                 return;
1137         }
1138
1139         /*
1140          * Do not disassociate the node from the buffer if it represents
1141          * a modified B-Tree node that still needs its crc to be generated.
1142          */
1143         if (node->flags & HAMMER_NODE_NEEDSCRC)
1144                 return;
1145
1146         /*
1147          * Do final cleanups and then either destroy the node and leave it
1148          * passively cached.  The buffer reference is removed regardless.
1149          */
1150         buffer = node->buffer;
1151         node->ondisk = NULL;
1152
1153         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1154                 hammer_unref(&node->lock);
1155                 hammer_rel_buffer(buffer, 0);
1156                 return;
1157         }
1158
1159         /*
1160          * Destroy the node.
1161          */
1162         hammer_unref(&node->lock);
1163         hammer_flush_node(node);
1164         /* node is stale */
1165         hammer_rel_buffer(buffer, 0);
1166 }
1167
1168 /*
1169  * Free space on-media associated with a B-Tree node.
1170  */
1171 void
1172 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1173 {
1174         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1175         node->flags |= HAMMER_NODE_DELETED;
1176         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1177 }
1178
1179 /*
1180  * Passively cache a referenced hammer_node.  The caller may release
1181  * the node on return.
1182  */
1183 void
1184 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1185 {
1186         /*
1187          * If the node is being deleted, don't cache it!
1188          */
1189         if (node->flags & HAMMER_NODE_DELETED)
1190                 return;
1191         if (cache->node == node)
1192                 return;
1193         while (cache->node)
1194                 hammer_uncache_node(cache);
1195         if (node->flags & HAMMER_NODE_DELETED)
1196                 return;
1197         cache->node = node;
1198         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1199 }
1200
1201 void
1202 hammer_uncache_node(hammer_node_cache_t cache)
1203 {
1204         hammer_node_t node;
1205
1206         if ((node = cache->node) != NULL) {
1207                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1208                 cache->node = NULL;
1209                 if (TAILQ_EMPTY(&node->cache_list))
1210                         hammer_flush_node(node);
1211         }
1212 }
1213
1214 /*
1215  * Remove a node's cache references and destroy the node if it has no
1216  * other references or backing store.
1217  */
1218 void
1219 hammer_flush_node(hammer_node_t node)
1220 {
1221         hammer_node_cache_t cache;
1222         hammer_buffer_t buffer;
1223
1224         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1225                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1226                 cache->node = NULL;
1227         }
1228         if (node->lock.refs == 0 && node->ondisk == NULL) {
1229                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1230                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1231                 if ((buffer = node->buffer) != NULL) {
1232                         node->buffer = NULL;
1233                         TAILQ_REMOVE(&buffer->clist, node, entry);
1234                         /* buffer is unreferenced because ondisk is NULL */
1235                 }
1236                 --hammer_count_nodes;
1237                 kfree(node, M_HAMMER);
1238         }
1239 }
1240
1241 /*
1242  * Flush passively cached B-Tree nodes associated with this buffer.
1243  * This is only called when the buffer is about to be destroyed, so
1244  * none of the nodes should have any references.  The buffer is locked.
1245  *
1246  * We may be interlocked with the buffer.
1247  */
1248 void
1249 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1250 {
1251         hammer_node_t node;
1252
1253         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1254                 KKASSERT(node->ondisk == NULL);
1255                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1256
1257                 if (node->lock.refs == 0) {
1258                         hammer_ref(&node->lock);
1259                         node->flags |= HAMMER_NODE_FLUSH;
1260                         hammer_rel_node(node);
1261                 } else {
1262                         KKASSERT(node->loading != 0);
1263                         KKASSERT(node->buffer != NULL);
1264                         buffer = node->buffer;
1265                         node->buffer = NULL;
1266                         TAILQ_REMOVE(&buffer->clist, node, entry);
1267                         /* buffer is unreferenced because ondisk is NULL */
1268                 }
1269         }
1270 }
1271
1272
1273 /************************************************************************
1274  *                              ALLOCATORS                              *
1275  ************************************************************************/
1276
1277 /*
1278  * Allocate a B-Tree node.
1279  */
1280 hammer_node_t
1281 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1282 {
1283         hammer_buffer_t buffer = NULL;
1284         hammer_node_t node = NULL;
1285         hammer_off_t node_offset;
1286
1287         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1288                                             sizeof(struct hammer_node_ondisk),
1289                                             errorp);
1290         if (*errorp == 0) {
1291                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1292                 hammer_modify_node_noundo(trans, node);
1293                 bzero(node->ondisk, sizeof(*node->ondisk));
1294                 hammer_modify_node_done(node);
1295         }
1296         if (buffer)
1297                 hammer_rel_buffer(buffer, 0);
1298         return(node);
1299 }
1300
1301 /*
1302  * Allocate data.  If the address of a data buffer is supplied then
1303  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1304  * will be set to the related buffer.  The caller must release it when
1305  * finally done.  The initial *data_bufferp should be set to NULL by
1306  * the caller.
1307  *
1308  * The caller is responsible for making hammer_modify*() calls on the
1309  * *data_bufferp.
1310  */
1311 void *
1312 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1313                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1314                   struct hammer_buffer **data_bufferp, int *errorp)
1315 {
1316         void *data;
1317         int zone;
1318
1319         /*
1320          * Allocate data
1321          */
1322         if (data_len) {
1323                 switch(rec_type) {
1324                 case HAMMER_RECTYPE_INODE:
1325                 case HAMMER_RECTYPE_DIRENTRY:
1326                 case HAMMER_RECTYPE_EXT:
1327                 case HAMMER_RECTYPE_FIX:
1328                         zone = HAMMER_ZONE_META_INDEX;
1329                         break;
1330                 case HAMMER_RECTYPE_DATA:
1331                 case HAMMER_RECTYPE_DB:
1332                         if (data_len <= HAMMER_BUFSIZE / 2) {
1333                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1334                         } else {
1335                                 data_len = (data_len + HAMMER_BUFMASK) &
1336                                            ~HAMMER_BUFMASK;
1337                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1338                         }
1339                         break;
1340                 default:
1341                         panic("hammer_alloc_data: rec_type %04x unknown",
1342                               rec_type);
1343                         zone = 0;       /* NOT REACHED */
1344                         break;
1345                 }
1346                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1347                                                       data_len, errorp);
1348         } else {
1349                 *data_offsetp = 0;
1350         }
1351         if (*errorp == 0 && data_bufferp) {
1352                 if (data_len) {
1353                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1354                                                 data_len, errorp, data_bufferp);
1355                         KKASSERT(*errorp == 0);
1356                 } else {
1357                         data = NULL;
1358                 }
1359         } else {
1360                 data = NULL;
1361         }
1362         KKASSERT(*errorp == 0);
1363         return(data);
1364 }
1365
1366 /*
1367  * Sync dirty buffers to the media and clean-up any loose ends.
1368  */
1369 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1370 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1371
1372 int
1373 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1374 {
1375         struct hammer_sync_info info;
1376
1377         info.error = 0;
1378         info.waitfor = waitfor;
1379         if (waitfor == MNT_WAIT) {
1380                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1381                               hammer_sync_scan1, hammer_sync_scan2, &info);
1382         } else {
1383                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1384                               hammer_sync_scan1, hammer_sync_scan2, &info);
1385         }
1386         return(info.error);
1387 }
1388
1389 /*
1390  * Filesystem sync.  If doing a synchronous sync make a second pass on
1391  * the vnodes in case any were already flushing during the first pass,
1392  * and activate the flusher twice (the second time brings the UNDO FIFO's
1393  * start position up to the end position after the first call).
1394  */
1395 int
1396 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1397 {
1398         struct hammer_sync_info info;
1399
1400         info.error = 0;
1401         info.waitfor = MNT_NOWAIT;
1402         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1403                       hammer_sync_scan1, hammer_sync_scan2, &info);
1404         if (info.error == 0 && waitfor == MNT_WAIT) {
1405                 info.waitfor = waitfor;
1406                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1407                               hammer_sync_scan1, hammer_sync_scan2, &info);
1408         }
1409         if (waitfor == MNT_WAIT) {
1410                 hammer_flusher_sync(hmp);
1411                 hammer_flusher_sync(hmp);
1412         } else {
1413                 hammer_flusher_async(hmp);
1414         }
1415         return(info.error);
1416 }
1417
1418 static int
1419 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1420 {
1421         struct hammer_inode *ip;
1422
1423         ip = VTOI(vp);
1424         if (vp->v_type == VNON || ip == NULL ||
1425             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1426              RB_EMPTY(&vp->v_rbdirty_tree))) {
1427                 return(-1);
1428         }
1429         return(0);
1430 }
1431
1432 static int
1433 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1434 {
1435         struct hammer_sync_info *info = data;
1436         struct hammer_inode *ip;
1437         int error;
1438
1439         ip = VTOI(vp);
1440         if (vp->v_type == VNON || vp->v_type == VBAD ||
1441             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1442              RB_EMPTY(&vp->v_rbdirty_tree))) {
1443                 return(0);
1444         }
1445         error = VOP_FSYNC(vp, info->waitfor);
1446         if (error)
1447                 info->error = error;
1448         return(0);
1449 }
1450