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 1.39 2008/04/26 02:54:00 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);
  52
  53 /*
  54  * Red-Black tree support for various structures
  55  */
  56 static 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->zone2_offset < buf2->zone2_offset)
  98                 return(-1);
  99         if (buf1->zone2_offset > buf2->zone2_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, zone2_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
 163         /*
 164          * Get the device vnode
 165          */
 166         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 167         if (error == 0)
 168                 error = nlookup(&nd);
 169         if (error == 0)
 170                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 171         nlookup_done(&nd);
 172         if (error == 0) {
 173                 if (vn_isdisk(volume->devvp, &error)) {
 174                         error = vfs_mountedon(volume->devvp);
 175                 }
 176         }
 177         if (error == 0 &&
 178             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 179                 error = EBUSY;
 180         }
 181         if (error == 0) {
 182                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 183                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 184                 if (error == 0) {
 185                         error = VOP_OPEN(volume->devvp,
 186                                          (ronly ? FREAD : FREAD|FWRITE),
 187                                          FSCRED, NULL);
 188                 }
 189                 vn_unlock(volume->devvp);
 190         }
 191         if (error) {
 192                 hammer_free_volume(volume);
 193                 return(error);
 194         }
 195         volume->devvp->v_rdev->si_mountpoint = mp;
 196         setmp = 1;
 197
 198         /*
 199          * Extract the volume number from the volume header and do various
 200          * sanity checks.
 201          */
 202         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 203         if (error)
 204                 goto late_failure;
 205         ondisk = (void *)bp->b_data;
 206         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 207                 kprintf("hammer_mount: volume %s has an invalid header\n",
 208                         volume->vol_name);
 209                 error = EFTYPE;
 210                 goto late_failure;
 211         }
 212         volume->vol_no = ondisk->vol_no;
 213         volume->buffer_base = ondisk->vol_buf_beg;
 214         volume->vol_flags = ondisk->vol_flags;
 215         volume->nblocks = ondisk->vol_nblocks;
 216         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 217                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 218         RB_INIT(&volume->rb_bufs_root);
 219
 220         hmp->mp->mnt_stat.f_blocks += volume->nblocks;
 221
 222         if (RB_EMPTY(&hmp->rb_vols_root)) {
 223                 hmp->fsid = ondisk->vol_fsid;
 224         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 225                 kprintf("hammer_mount: volume %s's fsid does not match "
 226                         "other volumes\n", volume->vol_name);
 227                 error = EFTYPE;
 228                 goto late_failure;
 229         }
 230
 231         /*
 232          * Insert the volume structure into the red-black tree.
 233          */
 234         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 235                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 236                         volume->vol_name, volume->vol_no);
 237                 error = EEXIST;
 238         }
 239
 240         /*
 241          * Set the root volume .  HAMMER special cases rootvol the structure.
 242          * We do not hold a ref because this would prevent related I/O
 243          * from being flushed.
 244          */
 245         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 246                 hmp->rootvol = volume;
 247                 if (bp) {
 248                         brelse(bp);
 249                         bp = NULL;
 250                 }
 251                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 252         }
 253 late_failure:
 254         if (bp)
 255                 brelse(bp);
 256         if (error) {
 257                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 258                 if (setmp)
 259                         volume->devvp->v_rdev->si_mountpoint = NULL;
 260                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 261                 hammer_free_volume(volume);
 262         }
 263         return (error);
 264 }
 265
 266 /*
 267  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 268  * so returns -1 on failure.
 269  */
 270 int
 271 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 272 {
 273         struct hammer_mount *hmp = volume->io.hmp;
 274         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 275
 276         /*
 277          * Sync clusters, sync volume
 278          */
 279
 280         hmp->mp->mnt_stat.f_blocks -= volume->nblocks;
 281
 282         /*
 283          * Clean up the root volume pointer, which is held unlocked in hmp.
 284          */
 285         if (hmp->rootvol == volume)
 286                 hmp->rootvol = NULL;
 287
 288         /*
 289          * Unload buffers.
 290          */
 291         RB_SCAN(hammer_buf_rb_tree, &volume->rb_bufs_root, NULL,
 292                         hammer_unload_buffer, NULL);
 293
 294         /*
 295          * Release our buffer and flush anything left in the buffer cache.
 296          */
 297         volume->io.flush = 1;
 298         volume->io.waitdep = 1;
 299         hammer_io_release(&volume->io);
 300
 301         /*
 302          * There should be no references on the volume, no clusters, and
 303          * no super-clusters.
 304          */
 305         KKASSERT(volume->io.lock.refs == 0);
 306         KKASSERT(RB_EMPTY(&volume->rb_bufs_root));
 307
 308         volume->ondisk = NULL;
 309         if (volume->devvp) {
 310                 if (volume->devvp->v_rdev &&
 311                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 312                 ) {
 313                         volume->devvp->v_rdev->si_mountpoint = NULL;
 314                 }
 315                 if (ronly) {
 316                         vinvalbuf(volume->devvp, 0, 0, 0);
 317                         VOP_CLOSE(volume->devvp, FREAD);
 318                 } else {
 319                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 320                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 321                 }
 322         }
 323
 324         /*
 325          * Destroy the structure
 326          */
 327         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 328         hammer_free_volume(volume);
 329         return(0);
 330 }
 331
 332 static
 333 void
 334 hammer_free_volume(hammer_volume_t volume)
 335 {
 336         if (volume->vol_name) {
 337                 kfree(volume->vol_name, M_HAMMER);
 338                 volume->vol_name = NULL;
 339         }
 340         if (volume->devvp) {
 341                 vrele(volume->devvp);
 342                 volume->devvp = NULL;
 343         }
 344         --hammer_count_volumes;
 345         kfree(volume, M_HAMMER);
 346 }
 347
 348 /*
 349  * Get a HAMMER volume.  The volume must already exist.
 350  */
 351 hammer_volume_t
 352 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 353 {
 354         struct hammer_volume *volume;
 355
 356         /*
 357          * Locate the volume structure
 358          */
 359         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 360         if (volume == NULL) {
 361                 *errorp = ENOENT;
 362                 return(NULL);
 363         }
 364         hammer_ref(&volume->io.lock);
 365
 366         /*
 367          * Deal with on-disk info
 368          */
 369         if (volume->ondisk == NULL || volume->io.loading) {
 370                 *errorp = hammer_load_volume(volume);
 371                 if (*errorp) {
 372                         hammer_rel_volume(volume, 1);
 373                         volume = NULL;
 374                 }
 375         } else {
 376                 *errorp = 0;
 377         }
 378         return(volume);
 379 }
 380
 381 int
 382 hammer_ref_volume(hammer_volume_t volume)
 383 {
 384         int error;
 385
 386         hammer_ref(&volume->io.lock);
 387
 388         /*
 389          * Deal with on-disk info
 390          */
 391         if (volume->ondisk == NULL || volume->io.loading) {
 392                 error = hammer_load_volume(volume);
 393                 if (error)
 394                         hammer_rel_volume(volume, 1);
 395         } else {
 396                 error = 0;
 397         }
 398         return (error);
 399 }
 400
 401 hammer_volume_t
 402 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 403 {
 404         hammer_volume_t volume;
 405
 406         volume = hmp->rootvol;
 407         KKASSERT(volume != NULL);
 408         hammer_ref(&volume->io.lock);
 409
 410         /*
 411          * Deal with on-disk info
 412          */
 413         if (volume->ondisk == NULL || volume->io.loading) {
 414                 *errorp = hammer_load_volume(volume);
 415                 if (*errorp) {
 416                         hammer_rel_volume(volume, 1);
 417                         volume = NULL;
 418                 }
 419         } else {
 420                 *errorp = 0;
 421         }
 422         return (volume);
 423 }
 424
 425 /*
 426  * Load a volume's on-disk information.  The volume must be referenced and
 427  * not locked.  We temporarily acquire an exclusive lock to interlock
 428  * against releases or multiple get's.
 429  */
 430 static int
 431 hammer_load_volume(hammer_volume_t volume)
 432 {
 433         int error;
 434
 435         ++volume->io.loading;
 436         hammer_lock_ex(&volume->io.lock);
 437
 438         if (volume->ondisk == NULL) {
 439                 error = hammer_io_read(volume->devvp, &volume->io);
 440                 if (error == 0)
 441                         volume->ondisk = (void *)volume->io.bp->b_data;
 442         } else {
 443                 error = 0;
 444         }
 445         --volume->io.loading;
 446         hammer_unlock(&volume->io.lock);
 447         return(error);
 448 }
 449
 450 /*
 451  * Release a volume.  Call hammer_io_release on the last reference.  We have
 452  * to acquire an exclusive lock to interlock against volume->ondisk tests
 453  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 454  * lock to be held.
 455  *
 456  * Volumes are not unloaded from memory during normal operation.
 457  */
 458 void
 459 hammer_rel_volume(hammer_volume_t volume, int flush)
 460 {
 461         if (flush)
 462                 volume->io.flush = 1;
 463         crit_enter();
 464         if (volume->io.lock.refs == 1) {
 465                 ++volume->io.loading;
 466                 hammer_lock_ex(&volume->io.lock);
 467                 if (volume->io.lock.refs == 1) {
 468                         volume->ondisk = NULL;
 469                         hammer_io_release(&volume->io);
 470                 }
 471                 --volume->io.loading;
 472                 hammer_unlock(&volume->io.lock);
 473         }
 474         hammer_unref(&volume->io.lock);
 475         crit_exit();
 476 }
 477
 478 /************************************************************************
 479  *                              BUFFERS                                 *
 480  ************************************************************************
 481  *
 482  * Manage buffers.  Currently all blockmap-backed zones are translated
 483  * to zone-2 buffer offsets.
 484  */
 485 hammer_buffer_t
 486 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 487                   int isnew, int *errorp)
 488 {
 489         hammer_buffer_t buffer;
 490         hammer_volume_t volume;
 491         hammer_off_t    zoneX_offset;
 492         hammer_io_type_t iotype;
 493         int vol_no;
 494         int zone;
 495
 496         zoneX_offset = buf_offset;
 497         zone = HAMMER_ZONE_DECODE(buf_offset);
 498
 499         /*
 500          * What is the buffer class?
 501          */
 502         switch(zone) {
 503         case HAMMER_ZONE_LARGE_DATA_INDEX:
 504         case HAMMER_ZONE_SMALL_DATA_INDEX:
 505                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 506                 break;
 507         case HAMMER_ZONE_UNDO_INDEX:
 508                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 509                 break;
 510         default:
 511                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 512                 break;
 513         }
 514
 515         /*
 516          * Handle blockmap offset translations
 517          */
 518         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 519                 buf_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 520                 KKASSERT(*errorp == 0);
 521         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 522                 buf_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 523                 KKASSERT(*errorp == 0);
 524         }
 525
 526         /*
 527          * Locate the buffer given its zone-2 offset.
 528          */
 529         buf_offset &= ~HAMMER_BUFMASK64;
 530         KKASSERT((buf_offset & HAMMER_ZONE_RAW_BUFFER) ==
 531                  HAMMER_ZONE_RAW_BUFFER);
 532         vol_no = HAMMER_VOL_DECODE(buf_offset);
 533         volume = hammer_get_volume(hmp, vol_no, errorp);
 534         if (volume == NULL)
 535                 return(NULL);
 536
 537         /*
 538          * NOTE: buf_offset and maxbuf_off are both full offset
 539          * specifications.
 540          */
 541         KKASSERT(buf_offset < volume->maxbuf_off);
 542
 543         /*
 544          * Locate and lock the buffer structure, creating one if necessary.
 545          */
 546 again:
 547         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 548                            buf_offset);
 549         if (buffer == NULL) {
 550                 ++hammer_count_buffers;
 551                 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 552                 buffer->zone2_offset = buf_offset;
 553                 buffer->volume = volume;
 554
 555                 hammer_io_init(&buffer->io, hmp, iotype);
 556                 buffer->io.offset = volume->ondisk->vol_buf_beg +
 557                                     (buf_offset & HAMMER_OFF_SHORT_MASK);
 558                 TAILQ_INIT(&buffer->clist);
 559                 hammer_ref(&buffer->io.lock);
 560
 561                 /*
 562                  * Insert the buffer into the RB tree and handle late
 563                  * collisions.
 564                  */
 565                 if (RB_INSERT(hammer_buf_rb_tree, &volume->rb_bufs_root, buffer)) {
 566                         hammer_unref(&buffer->io.lock);
 567                         --hammer_count_buffers;
 568                         kfree(buffer, M_HAMMER);
 569                         goto again;
 570                 }
 571                 hammer_ref(&volume->io.lock);
 572         } else {
 573                 hammer_ref(&buffer->io.lock);
 574
 575                 /*
 576                  * The buffer is no longer loose if it has a ref.
 577                  */
 578                 if (buffer->io.mod_list == &hmp->lose_list) {
 579                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 580                                      mod_entry);
 581                         buffer->io.mod_list = NULL;
 582                 }
 583                 if (buffer->io.lock.refs == 1)
 584                         hammer_io_reinit(&buffer->io, iotype);
 585                 else
 586                         KKASSERT(buffer->io.type == iotype);
 587         }
 588
 589         /*
 590          * Cache the blockmap translation
 591          */
 592         if ((zoneX_offset & HAMMER_ZONE_RAW_BUFFER) != HAMMER_ZONE_RAW_BUFFER)
 593                 buffer->zoneX_offset = zoneX_offset;
 594
 595         /*
 596          * Deal with on-disk info
 597          */
 598         if (buffer->ondisk == NULL || buffer->io.loading) {
 599                 *errorp = hammer_load_buffer(buffer, isnew);
 600                 if (*errorp) {
 601                         hammer_rel_buffer(buffer, 1);
 602                         buffer = NULL;
 603                 }
 604         } else {
 605                 *errorp = 0;
 606         }
 607         hammer_rel_volume(volume, 0);
 608         return(buffer);
 609 }
 610
 611 static int
 612 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 613 {
 614         hammer_volume_t volume;
 615         int error;
 616
 617         /*
 618          * Load the buffer's on-disk info
 619          */
 620         volume = buffer->volume;
 621         ++buffer->io.loading;
 622         hammer_lock_ex(&buffer->io.lock);
 623
 624         if (buffer->ondisk == NULL) {
 625                 if (isnew) {
 626                         error = hammer_io_new(volume->devvp, &buffer->io);
 627                 } else {
 628                         error = hammer_io_read(volume->devvp, &buffer->io);
 629                 }
 630                 if (error == 0)
 631                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 632         } else if (isnew) {
 633                 error = hammer_io_new(volume->devvp, &buffer->io);
 634         } else {
 635                 error = 0;
 636         }
 637         --buffer->io.loading;
 638         hammer_unlock(&buffer->io.lock);
 639         return (error);
 640 }
 641
 642 /*
 643  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 644  */
 645 int
 646 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 647 {
 648         hammer_ref(&buffer->io.lock);
 649         hammer_flush_buffer_nodes(buffer);
 650         KKASSERT(buffer->io.lock.refs == 1);
 651         hammer_rel_buffer(buffer, 2);
 652         return(0);
 653 }
 654
 655 /*
 656  * Reference a buffer that is either already referenced or via a specially
 657  * handled pointer (aka cursor->buffer).
 658  */
 659 int
 660 hammer_ref_buffer(hammer_buffer_t buffer)
 661 {
 662         int error;
 663
 664         hammer_ref(&buffer->io.lock);
 665
 666         /*
 667          * No longer loose
 668          */
 669         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 670                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 671                 buffer->io.mod_list = NULL;
 672         }
 673
 674         if (buffer->ondisk == NULL || buffer->io.loading) {
 675                 error = hammer_load_buffer(buffer, 0);
 676                 if (error) {
 677                         hammer_rel_buffer(buffer, 1);
 678                         /*
 679                          * NOTE: buffer pointer can become stale after
 680                          * the above release.
 681                          */
 682                 }
 683         } else {
 684                 error = 0;
 685         }
 686         return(error);
 687 }
 688
 689 /*
 690  * Release a buffer.  We have to deal with several places where
 691  * another thread can ref the buffer.
 692  *
 693  * Only destroy the structure itself if the related buffer cache buffer
 694  * was disassociated from it.  This ties the management of the structure
 695  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 696  * embedded io is referenced or not.
 697  */
 698 void
 699 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 700 {
 701         hammer_volume_t volume;
 702         int freeme = 0;
 703
 704         if (flush)
 705                 buffer->io.flush = 1;
 706         crit_enter();
 707         if (buffer->io.lock.refs == 1) {
 708                 ++buffer->io.loading;   /* force interlock check */
 709                 hammer_lock_ex(&buffer->io.lock);
 710                 if (buffer->io.lock.refs == 1) {
 711                         hammer_io_release(&buffer->io);
 712                         hammer_flush_buffer_nodes(buffer);
 713                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 714
 715                         if (buffer->io.bp == NULL &&
 716                             buffer->io.lock.refs == 1) {
 717                                 /*
 718                                  * Final cleanup
 719                                  */
 720                                 volume = buffer->volume;
 721                                 RB_REMOVE(hammer_buf_rb_tree,
 722                                           &volume->rb_bufs_root, buffer);
 723                                 buffer->volume = NULL; /* sanity */
 724                                 hammer_rel_volume(volume, 0);
 725                                 freeme = 1;
 726                         }
 727                 }
 728                 --buffer->io.loading;
 729                 hammer_unlock(&buffer->io.lock);
 730         }
 731         hammer_unref(&buffer->io.lock);
 732         crit_exit();
 733         if (freeme) {
 734                 KKASSERT(buffer->io.mod_list == NULL);
 735                 --hammer_count_buffers;
 736                 kfree(buffer, M_HAMMER);
 737         }
 738 }
 739
 740 /*
 741  * Remove the zoneX translation cache for a buffer given its zone-2 offset.
 742  */
 743 void
 744 hammer_uncache_buffer(hammer_mount_t hmp, hammer_off_t buf_offset)
 745 {
 746         hammer_volume_t volume;
 747         hammer_buffer_t buffer;
 748         int vol_no;
 749         int error;
 750
 751         buf_offset &= ~HAMMER_BUFMASK64;
 752         KKASSERT((buf_offset & HAMMER_ZONE_RAW_BUFFER) ==
 753                  HAMMER_ZONE_RAW_BUFFER);
 754         vol_no = HAMMER_VOL_DECODE(buf_offset);
 755         volume = hammer_get_volume(hmp, vol_no, &error);
 756         KKASSERT(volume != 0);
 757         KKASSERT(buf_offset < volume->maxbuf_off);
 758
 759         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 760                            buf_offset);
 761         if (buffer)
 762                 buffer->zoneX_offset = 0;
 763         hammer_rel_volume(volume, 0);
 764 }
 765
 766 /*
 767  * Access the filesystem buffer containing the specified hammer offset.
 768  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 769  * relative buffer offset.  It must also have bit 55 set to be valid.
 770  * (see hammer_off_t in hammer_disk.h).
 771  *
 772  * Any prior buffer in *bufferp will be released and replaced by the
 773  * requested buffer.
 774  */
 775 void *
 776 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 777              struct hammer_buffer **bufferp)
 778 {
 779         hammer_buffer_t buffer;
 780         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 781
 782         buf_offset &= ~HAMMER_BUFMASK64;
 783         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 784
 785         buffer = *bufferp;
 786         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 787                                buffer->zoneX_offset != buf_offset)) {
 788                 if (buffer)
 789                         hammer_rel_buffer(buffer, 0);
 790                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 791                 *bufferp = buffer;
 792         } else {
 793                 *errorp = 0;
 794         }
 795
 796         /*
 797          * Return a pointer to the buffer data.
 798          */
 799         if (buffer == NULL)
 800                 return(NULL);
 801         else
 802                 return((char *)buffer->ondisk + xoff);
 803 }
 804
 805 /*
 806  * Access the filesystem buffer containing the specified hammer offset.
 807  * No disk read operation occurs.  The result buffer may contain garbage.
 808  *
 809  * Any prior buffer in *bufferp will be released and replaced by the
 810  * requested buffer.
 811  *
 812  * This function marks the buffer dirty but does not increment its
 813  * modify_refs count.
 814  */
 815 void *
 816 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 817              struct hammer_buffer **bufferp)
 818 {
 819         hammer_buffer_t buffer;
 820         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 821
 822         buf_offset &= ~HAMMER_BUFMASK64;
 823
 824         buffer = *bufferp;
 825         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 826                                buffer->zoneX_offset != buf_offset)) {
 827                 if (buffer)
 828                         hammer_rel_buffer(buffer, 0);
 829                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 830                 *bufferp = buffer;
 831         } else {
 832                 *errorp = 0;
 833         }
 834
 835         /*
 836          * Return a pointer to the buffer data.
 837          */
 838         if (buffer == NULL)
 839                 return(NULL);
 840         else
 841                 return((char *)buffer->ondisk + xoff);
 842 }
 843
 844 /************************************************************************
 845  *                              NODES                                   *
 846  ************************************************************************
 847  *
 848  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 849  * method used by the HAMMER filesystem.
 850  *
 851  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 852  * associated with its buffer, and will only referenced the buffer while
 853  * the node itself is referenced.
 854  *
 855  * A hammer_node can also be passively associated with other HAMMER
 856  * structures, such as inodes, while retaining 0 references.  These
 857  * associations can be cleared backwards using a pointer-to-pointer in
 858  * the hammer_node.
 859  *
 860  * This allows the HAMMER implementation to cache hammer_nodes long-term
 861  * and short-cut a great deal of the infrastructure's complexity.  In
 862  * most cases a cached node can be reacquired without having to dip into
 863  * either the buffer or cluster management code.
 864  *
 865  * The caller must pass a referenced cluster on call and will retain
 866  * ownership of the reference on return.  The node will acquire its own
 867  * additional references, if necessary.
 868  */
 869 hammer_node_t
 870 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset, int *errorp)
 871 {
 872         hammer_node_t node;
 873
 874         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 875
 876         /*
 877          * Locate the structure, allocating one if necessary.
 878          */
 879 again:
 880         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 881         if (node == NULL) {
 882                 ++hammer_count_nodes;
 883                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 884                 node->node_offset = node_offset;
 885                 node->hmp = hmp;
 886                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 887                         --hammer_count_nodes;
 888                         kfree(node, M_HAMMER);
 889                         goto again;
 890                 }
 891         }
 892         hammer_ref(&node->lock);
 893         if (node->ondisk)
 894                 *errorp = 0;
 895         else
 896                 *errorp = hammer_load_node(node);
 897         if (*errorp) {
 898                 hammer_rel_node(node);
 899                 node = NULL;
 900         }
 901         return(node);
 902 }
 903
 904 /*
 905  * Reference an already-referenced node.
 906  */
 907 void
 908 hammer_ref_node(hammer_node_t node)
 909 {
 910         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 911         hammer_ref(&node->lock);
 912 }
 913
 914 /*
 915  * Load a node's on-disk data reference.
 916  */
 917 static int
 918 hammer_load_node(hammer_node_t node)
 919 {
 920         hammer_buffer_t buffer;
 921         int error;
 922
 923         error = 0;
 924         ++node->loading;
 925         hammer_lock_ex(&node->lock);
 926         if (node->ondisk == NULL) {
 927                 /*
 928                  * This is a little confusing but the jist is that
 929                  * node->buffer determines whether the node is on
 930                  * the buffer's clist and node->ondisk determines
 931                  * whether the buffer is referenced.
 932                  *
 933                  * We could be racing a buffer release, in which case
 934                  * node->buffer may become NULL while we are blocked
 935                  * referencing the buffer.
 936                  */
 937                 if ((buffer = node->buffer) != NULL) {
 938                         error = hammer_ref_buffer(buffer);
 939                         if (error == 0 && node->buffer == NULL) {
 940                                 TAILQ_INSERT_TAIL(&buffer->clist,
 941                                                   node, entry);
 942                                 node->buffer = buffer;
 943                         }
 944                 } else {
 945                         buffer = hammer_get_buffer(node->hmp,
 946                                                    node->node_offset, 0,
 947                                                    &error);
 948                         if (buffer) {
 949                                 KKASSERT(error == 0);
 950                                 TAILQ_INSERT_TAIL(&buffer->clist,
 951                                                   node, entry);
 952                                 node->buffer = buffer;
 953                         }
 954                 }
 955                 if (error == 0) {
 956                         node->ondisk = (void *)((char *)buffer->ondisk +
 957                                (node->node_offset & HAMMER_BUFMASK));
 958                 }
 959         }
 960         --node->loading;
 961         hammer_unlock(&node->lock);
 962         return (error);
 963 }
 964
 965 /*
 966  * Safely reference a node, interlock against flushes via the IO subsystem.
 967  */
 968 hammer_node_t
 969 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
 970                      int *errorp)
 971 {
 972         hammer_node_t node;
 973
 974         node = *cache;
 975         if (node != NULL) {
 976                 hammer_ref(&node->lock);
 977                 if (node->ondisk)
 978                         *errorp = 0;
 979                 else
 980                         *errorp = hammer_load_node(node);
 981                 if (*errorp) {
 982                         hammer_rel_node(node);
 983                         node = NULL;
 984                 }
 985         } else {
 986                 *errorp = ENOENT;
 987         }
 988         return(node);
 989 }
 990
 991 /*
 992  * Release a hammer_node.  On the last release the node dereferences
 993  * its underlying buffer and may or may not be destroyed.
 994  */
 995 void
 996 hammer_rel_node(hammer_node_t node)
 997 {
 998         hammer_buffer_t buffer;
 999
1000         /*
1001          * If this isn't the last ref just decrement the ref count and
1002          * return.
1003          */
1004         if (node->lock.refs > 1) {
1005                 hammer_unref(&node->lock);
1006                 return;
1007         }
1008
1009         /*
1010          * If there is no ondisk info or no buffer the node failed to load,
1011          * remove the last reference and destroy the node.
1012          */
1013         if (node->ondisk == NULL) {
1014                 hammer_unref(&node->lock);
1015                 hammer_flush_node(node);
1016                 /* node is stale now */
1017                 return;
1018         }
1019
1020         /*
1021          * Do final cleanups and then either destroy the node and leave it
1022          * passively cached.  The buffer reference is removed regardless.
1023          */
1024         buffer = node->buffer;
1025         node->ondisk = NULL;
1026
1027         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1028                 hammer_unref(&node->lock);
1029                 hammer_rel_buffer(buffer, 0);
1030                 return;
1031         }
1032
1033         /*
1034          * Destroy the node.
1035          */
1036         hammer_unref(&node->lock);
1037         hammer_flush_node(node);
1038         /* node is stale */
1039         hammer_rel_buffer(buffer, 0);
1040 }
1041
1042 /*
1043  *
1044  *
1045  */
1046 void
1047 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1048 {
1049         node->flags |= HAMMER_NODE_DELETED;
1050         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1051 }
1052
1053 /*
1054  * Passively cache a referenced hammer_node in *cache.  The caller may
1055  * release the node on return.
1056  */
1057 void
1058 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1059 {
1060         hammer_node_t old;
1061
1062         /*
1063          * If the node is being deleted, don't cache it!
1064          */
1065         if (node->flags & HAMMER_NODE_DELETED)
1066                 return;
1067
1068         /*
1069          * Cache the node.  If we previously cached a different node we
1070          * have to give HAMMER a chance to destroy it.
1071          */
1072 again:
1073         if (node->cache1 != cache) {
1074                 if (node->cache2 != cache) {
1075                         if ((old = *cache) != NULL) {
1076                                 KKASSERT(node->lock.refs != 0);
1077                                 hammer_uncache_node(cache);
1078                                 goto again;
1079                         }
1080                         if (node->cache2)
1081                                 *node->cache2 = NULL;
1082                         node->cache2 = node->cache1;
1083                         node->cache1 = cache;
1084                         *cache = node;
1085                 } else {
1086                         struct hammer_node **tmp;
1087                         tmp = node->cache1;
1088                         node->cache1 = node->cache2;
1089                         node->cache2 = tmp;
1090                 }
1091         }
1092 }
1093
1094 void
1095 hammer_uncache_node(struct hammer_node **cache)
1096 {
1097         hammer_node_t node;
1098
1099         if ((node = *cache) != NULL) {
1100                 *cache = NULL;
1101                 if (node->cache1 == cache) {
1102                         node->cache1 = node->cache2;
1103                         node->cache2 = NULL;
1104                 } else if (node->cache2 == cache) {
1105                         node->cache2 = NULL;
1106                 } else {
1107                         panic("hammer_uncache_node: missing cache linkage");
1108                 }
1109                 if (node->cache1 == NULL && node->cache2 == NULL)
1110                         hammer_flush_node(node);
1111         }
1112 }
1113
1114 /*
1115  * Remove a node's cache references and destroy the node if it has no
1116  * other references or backing store.
1117  */
1118 void
1119 hammer_flush_node(hammer_node_t node)
1120 {
1121         hammer_buffer_t buffer;
1122
1123         if (node->cache1)
1124                 *node->cache1 = NULL;
1125         if (node->cache2)
1126                 *node->cache2 = NULL;
1127         if (node->lock.refs == 0 && node->ondisk == NULL) {
1128                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1129                 if ((buffer = node->buffer) != NULL) {
1130                         node->buffer = NULL;
1131                         TAILQ_REMOVE(&buffer->clist, node, entry);
1132                         /* buffer is unreferenced because ondisk is NULL */
1133                 }
1134                 --hammer_count_nodes;
1135                 kfree(node, M_HAMMER);
1136         }
1137 }
1138
1139 /*
1140  * Flush passively cached B-Tree nodes associated with this buffer.
1141  * This is only called when the buffer is about to be destroyed, so
1142  * none of the nodes should have any references.  The buffer is locked.
1143  *
1144  * We may be interlocked with the buffer.
1145  */
1146 void
1147 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1148 {
1149         hammer_node_t node;
1150
1151         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1152                 KKASSERT(node->ondisk == NULL);
1153
1154                 if (node->lock.refs == 0) {
1155                         hammer_ref(&node->lock);
1156                         node->flags |= HAMMER_NODE_FLUSH;
1157                         hammer_rel_node(node);
1158                 } else {
1159                         KKASSERT(node->loading != 0);
1160                         KKASSERT(node->buffer != NULL);
1161                         buffer = node->buffer;
1162                         node->buffer = NULL;
1163                         TAILQ_REMOVE(&buffer->clist, node, entry);
1164                         /* buffer is unreferenced because ondisk is NULL */
1165                 }
1166         }
1167 }
1168
1169
1170 /************************************************************************
1171  *                              ALLOCATORS                              *
1172  ************************************************************************/
1173
1174 /*
1175  * Allocate a B-Tree node.
1176  */
1177 hammer_node_t
1178 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1179 {
1180         hammer_buffer_t buffer = NULL;
1181         hammer_node_t node = NULL;
1182         hammer_off_t node_offset;
1183
1184         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1185                                             sizeof(struct hammer_node_ondisk),
1186                                             errorp);
1187         if (*errorp == 0) {
1188                 node = hammer_get_node(trans->hmp, node_offset, errorp);
1189                 hammer_modify_node_noundo(trans, node);
1190                 bzero(node->ondisk, sizeof(*node->ondisk));
1191                 hammer_modify_node_done(node);
1192         }
1193         if (buffer)
1194                 hammer_rel_buffer(buffer, 0);
1195         return(node);
1196 }
1197
1198 /*
1199  * The returned buffers are already appropriately marked as being modified.
1200  * If the caller marks them again unnecessary undo records may be generated.
1201  *
1202  * In-band data is indicated by data_bufferp == NULL.  Pass a data_len of 0
1203  * for zero-fill (caller modifies data_len afterwords).
1204  *
1205  * If the caller is responsible for calling hammer_modify_*() prior to making
1206  * any additional modifications to either the returned record buffer or the
1207  * returned data buffer.
1208  */
1209 void *
1210 hammer_alloc_record(hammer_transaction_t trans,
1211                     hammer_off_t *rec_offp, u_int16_t rec_type,
1212                     struct hammer_buffer **rec_bufferp,
1213                     int32_t data_len, void **datap,
1214                     struct hammer_buffer **data_bufferp, int *errorp)
1215 {
1216         hammer_record_ondisk_t rec;
1217         hammer_off_t rec_offset;
1218         hammer_off_t data_offset;
1219         int32_t reclen;
1220
1221         if (datap)
1222                 *datap = NULL;
1223
1224         /*
1225          * Allocate the record
1226          */
1227         rec_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_RECORD_INDEX,
1228                                            HAMMER_RECORD_SIZE, errorp);
1229         if (*errorp)
1230                 return(NULL);
1231
1232         /*
1233          * Allocate data
1234          */
1235         if (data_len) {
1236                 if (data_bufferp == NULL) {
1237                         switch(rec_type) {
1238                         case HAMMER_RECTYPE_DATA:
1239                                 reclen = offsetof(struct hammer_data_record,
1240                                                   data[0]);
1241                                 break;
1242                         case HAMMER_RECTYPE_DIRENTRY:
1243                                 reclen = offsetof(struct hammer_entry_record,
1244                                                   name[0]);
1245                                 break;
1246                         default:
1247                                 panic("hammer_alloc_record: illegal "
1248                                       "in-band data");
1249                                 /* NOT REACHED */
1250                                 reclen = 0;
1251                                 break;
1252                         }
1253                         KKASSERT(reclen + data_len <= HAMMER_RECORD_SIZE);
1254                         data_offset = rec_offset + reclen;
1255                 } else if (data_len < HAMMER_BUFSIZE) {
1256                         data_offset = hammer_blockmap_alloc(trans,
1257                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1258                                                 data_len, errorp);
1259                 } else {
1260                         data_offset = hammer_blockmap_alloc(trans,
1261                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1262                                                 data_len, errorp);
1263                 }
1264         } else {
1265                 data_offset = 0;
1266         }
1267         if (*errorp) {
1268                 hammer_blockmap_free(trans, rec_offset, HAMMER_RECORD_SIZE);
1269                 return(NULL);
1270         }
1271
1272         /*
1273          * Basic return values.
1274          *
1275          * Note that because this is a 'new' buffer, there is no need to
1276          * generate UNDO records for it.
1277          */
1278         *rec_offp = rec_offset;
1279         rec = hammer_bread(trans->hmp, rec_offset, errorp, rec_bufferp);
1280         hammer_modify_buffer(trans, *rec_bufferp, NULL, 0);
1281         bzero(rec, sizeof(*rec));
1282         KKASSERT(*errorp == 0);
1283         rec->base.data_off = data_offset;
1284         rec->base.data_len = data_len;
1285         hammer_modify_buffer_done(*rec_bufferp);
1286
1287         if (data_bufferp) {
1288                 if (data_len) {
1289                         *datap = hammer_bread(trans->hmp, data_offset, errorp,
1290                                               data_bufferp);
1291                         KKASSERT(*errorp == 0);
1292                 } else {
1293                         *datap = NULL;
1294                 }
1295         } else if (data_len) {
1296                 KKASSERT(data_offset + data_len - rec_offset <=
1297                          HAMMER_RECORD_SIZE);
1298                 if (datap) {
1299                         *datap = (void *)((char *)rec +
1300                                           (int32_t)(data_offset - rec_offset));
1301                 }
1302         } else {
1303                 KKASSERT(datap == NULL);
1304         }
1305         KKASSERT(*errorp == 0);
1306         return(rec);
1307 }
1308
1309 /*
1310  * Allocate data.  If the address of a data buffer is supplied then
1311  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1312  * will be set to the related buffer.  The caller must release it when
1313  * finally done.  The initial *data_bufferp should be set to NULL by
1314  * the caller.
1315  *
1316  * The caller is responsible for making hammer_modify*() calls on the
1317  * *data_bufferp.
1318  */
1319 void *
1320 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1321                   hammer_off_t *data_offsetp,
1322                   struct hammer_buffer **data_bufferp, int *errorp)
1323 {
1324         void *data;
1325
1326         /*
1327          * Allocate data
1328          */
1329         if (data_len) {
1330                 if (data_len < HAMMER_BUFSIZE) {
1331                         *data_offsetp = hammer_blockmap_alloc(trans,
1332                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1333                                                 data_len, errorp);
1334                 } else {
1335                         *data_offsetp = hammer_blockmap_alloc(trans,
1336                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1337                                                 data_len, errorp);
1338                 }
1339         } else {
1340                 *data_offsetp = 0;
1341         }
1342         if (*errorp == 0 && data_bufferp) {
1343                 if (data_len) {
1344                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1345                                             data_bufferp);
1346                         KKASSERT(*errorp == 0);
1347                 } else {
1348                         data = NULL;
1349                 }
1350         } else {
1351                 data = NULL;
1352         }
1353         KKASSERT(*errorp == 0);
1354         return(data);
1355 }
1356
1357 /*
1358  * Sync dirty buffers to the media and clean-up any loose ends.
1359  */
1360 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1361 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1362
1363 int
1364 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1365 {
1366         struct hammer_sync_info info;
1367
1368         info.error = 0;
1369         info.waitfor = waitfor;
1370
1371         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1372                       hammer_sync_scan1, hammer_sync_scan2, &info);
1373         if (waitfor == MNT_WAIT)
1374                 hammer_flusher_sync(hmp);
1375         else
1376                 hammer_flusher_async(hmp);
1377
1378         return(info.error);
1379 }
1380
1381 static int
1382 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1383 {
1384         struct hammer_inode *ip;
1385
1386         ip = VTOI(vp);
1387         if (vp->v_type == VNON || ip == NULL ||
1388             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1389              RB_EMPTY(&vp->v_rbdirty_tree))) {
1390                 return(-1);
1391         }
1392         return(0);
1393 }
1394
1395 static int
1396 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1397 {
1398         struct hammer_sync_info *info = data;
1399         struct hammer_inode *ip;
1400         int error;
1401
1402         ip = VTOI(vp);
1403         if (vp->v_type == VNON || vp->v_type == VBAD ||
1404             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1405              RB_EMPTY(&vp->v_rbdirty_tree))) {
1406                 return(0);
1407         }
1408         error = VOP_FSYNC(vp, info->waitfor);
1409         if (error)
1410                 info->error = error;
1411         return(0);
1412 }
1413