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 1.45 2008/05/15 03:36:40 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 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         if (RB_EMPTY(&hmp->rb_vols_root)) {
 221                 hmp->fsid = ondisk->vol_fsid;
 222         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 223                 kprintf("hammer_mount: volume %s's fsid does not match "
 224                         "other volumes\n", volume->vol_name);
 225                 error = EFTYPE;
 226                 goto late_failure;
 227         }
 228
 229         /*
 230          * Insert the volume structure into the red-black tree.
 231          */
 232         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 233                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 234                         volume->vol_name, volume->vol_no);
 235                 error = EEXIST;
 236         }
 237
 238         /*
 239          * Set the root volume .  HAMMER special cases rootvol the structure.
 240          * We do not hold a ref because this would prevent related I/O
 241          * from being flushed.
 242          */
 243         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 244                 hmp->rootvol = volume;
 245                 if (bp) {
 246                         brelse(bp);
 247                         bp = NULL;
 248                 }
 249                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 250                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 251                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 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          * Clean up the root volume pointer, which is held unlocked in hmp.
 278          */
 279         if (hmp->rootvol == volume)
 280                 hmp->rootvol = NULL;
 281
 282         /*
 283          * Unload buffers.
 284          */
 285         RB_SCAN(hammer_buf_rb_tree, &volume->rb_bufs_root, NULL,
 286                         hammer_unload_buffer, NULL);
 287
 288         /*
 289          * Release our buffer and flush anything left in the buffer cache.
 290          */
 291         volume->io.waitdep = 1;
 292         hammer_io_release(&volume->io, 1);
 293
 294         /*
 295          * There should be no references on the volume, no clusters, and
 296          * no super-clusters.
 297          */
 298         KKASSERT(volume->io.lock.refs == 0);
 299         KKASSERT(RB_EMPTY(&volume->rb_bufs_root));
 300
 301         volume->ondisk = NULL;
 302         if (volume->devvp) {
 303                 if (volume->devvp->v_rdev &&
 304                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 305                 ) {
 306                         volume->devvp->v_rdev->si_mountpoint = NULL;
 307                 }
 308                 if (ronly) {
 309                         vinvalbuf(volume->devvp, 0, 0, 0);
 310                         VOP_CLOSE(volume->devvp, FREAD);
 311                 } else {
 312                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 313                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 314                 }
 315         }
 316
 317         /*
 318          * Destroy the structure
 319          */
 320         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 321         hammer_free_volume(volume);
 322         return(0);
 323 }
 324
 325 static
 326 void
 327 hammer_free_volume(hammer_volume_t volume)
 328 {
 329         if (volume->vol_name) {
 330                 kfree(volume->vol_name, M_HAMMER);
 331                 volume->vol_name = NULL;
 332         }
 333         if (volume->devvp) {
 334                 vrele(volume->devvp);
 335                 volume->devvp = NULL;
 336         }
 337         --hammer_count_volumes;
 338         kfree(volume, M_HAMMER);
 339 }
 340
 341 /*
 342  * Get a HAMMER volume.  The volume must already exist.
 343  */
 344 hammer_volume_t
 345 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 346 {
 347         struct hammer_volume *volume;
 348
 349         /*
 350          * Locate the volume structure
 351          */
 352         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 353         if (volume == NULL) {
 354                 *errorp = ENOENT;
 355                 return(NULL);
 356         }
 357         hammer_ref(&volume->io.lock);
 358
 359         /*
 360          * Deal with on-disk info
 361          */
 362         if (volume->ondisk == NULL || volume->io.loading) {
 363                 *errorp = hammer_load_volume(volume);
 364                 if (*errorp) {
 365                         hammer_rel_volume(volume, 1);
 366                         volume = NULL;
 367                 }
 368         } else {
 369                 *errorp = 0;
 370         }
 371         return(volume);
 372 }
 373
 374 int
 375 hammer_ref_volume(hammer_volume_t volume)
 376 {
 377         int error;
 378
 379         hammer_ref(&volume->io.lock);
 380
 381         /*
 382          * Deal with on-disk info
 383          */
 384         if (volume->ondisk == NULL || volume->io.loading) {
 385                 error = hammer_load_volume(volume);
 386                 if (error)
 387                         hammer_rel_volume(volume, 1);
 388         } else {
 389                 error = 0;
 390         }
 391         return (error);
 392 }
 393
 394 hammer_volume_t
 395 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 396 {
 397         hammer_volume_t volume;
 398
 399         volume = hmp->rootvol;
 400         KKASSERT(volume != NULL);
 401         hammer_ref(&volume->io.lock);
 402
 403         /*
 404          * Deal with on-disk info
 405          */
 406         if (volume->ondisk == NULL || volume->io.loading) {
 407                 *errorp = hammer_load_volume(volume);
 408                 if (*errorp) {
 409                         hammer_rel_volume(volume, 1);
 410                         volume = NULL;
 411                 }
 412         } else {
 413                 *errorp = 0;
 414         }
 415         return (volume);
 416 }
 417
 418 /*
 419  * Load a volume's on-disk information.  The volume must be referenced and
 420  * not locked.  We temporarily acquire an exclusive lock to interlock
 421  * against releases or multiple get's.
 422  */
 423 static int
 424 hammer_load_volume(hammer_volume_t volume)
 425 {
 426         int error;
 427
 428         ++volume->io.loading;
 429         hammer_lock_ex(&volume->io.lock);
 430
 431         if (volume->ondisk == NULL) {
 432                 error = hammer_io_read(volume->devvp, &volume->io);
 433                 if (error == 0)
 434                         volume->ondisk = (void *)volume->io.bp->b_data;
 435         } else {
 436                 error = 0;
 437         }
 438         --volume->io.loading;
 439         hammer_unlock(&volume->io.lock);
 440         return(error);
 441 }
 442
 443 /*
 444  * Release a volume.  Call hammer_io_release on the last reference.  We have
 445  * to acquire an exclusive lock to interlock against volume->ondisk tests
 446  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 447  * lock to be held.
 448  *
 449  * Volumes are not unloaded from memory during normal operation.
 450  */
 451 void
 452 hammer_rel_volume(hammer_volume_t volume, int flush)
 453 {
 454         crit_enter();
 455         if (volume->io.lock.refs == 1) {
 456                 ++volume->io.loading;
 457                 hammer_lock_ex(&volume->io.lock);
 458                 if (volume->io.lock.refs == 1) {
 459                         volume->ondisk = NULL;
 460                         hammer_io_release(&volume->io, flush);
 461                 }
 462                 --volume->io.loading;
 463                 hammer_unlock(&volume->io.lock);
 464         }
 465         hammer_unref(&volume->io.lock);
 466         crit_exit();
 467 }
 468
 469 /************************************************************************
 470  *                              BUFFERS                                 *
 471  ************************************************************************
 472  *
 473  * Manage buffers.  Currently all blockmap-backed zones are translated
 474  * to zone-2 buffer offsets.
 475  */
 476 hammer_buffer_t
 477 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 478                   int isnew, int *errorp)
 479 {
 480         hammer_buffer_t buffer;
 481         hammer_volume_t volume;
 482         hammer_off_t    zoneX_offset;
 483         hammer_io_type_t iotype;
 484         int vol_no;
 485         int zone;
 486
 487         zoneX_offset = buf_offset;
 488         zone = HAMMER_ZONE_DECODE(buf_offset);
 489
 490         /*
 491          * What is the buffer class?
 492          */
 493         switch(zone) {
 494         case HAMMER_ZONE_LARGE_DATA_INDEX:
 495         case HAMMER_ZONE_SMALL_DATA_INDEX:
 496                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 497                 break;
 498         case HAMMER_ZONE_UNDO_INDEX:
 499                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 500                 break;
 501         default:
 502                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 503                 break;
 504         }
 505
 506         /*
 507          * Handle blockmap offset translations
 508          */
 509         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 510                 buf_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 511                 KKASSERT(*errorp == 0);
 512         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 513                 buf_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 514                 KKASSERT(*errorp == 0);
 515         }
 516
 517         /*
 518          * Locate the buffer given its zone-2 offset.
 519          */
 520         buf_offset &= ~HAMMER_BUFMASK64;
 521         KKASSERT((buf_offset & HAMMER_ZONE_RAW_BUFFER) ==
 522                  HAMMER_ZONE_RAW_BUFFER);
 523         vol_no = HAMMER_VOL_DECODE(buf_offset);
 524         volume = hammer_get_volume(hmp, vol_no, errorp);
 525         if (volume == NULL)
 526                 return(NULL);
 527
 528         /*
 529          * NOTE: buf_offset and maxbuf_off are both full offset
 530          * specifications.
 531          */
 532         KKASSERT(buf_offset < volume->maxbuf_off);
 533
 534         /*
 535          * Locate and lock the buffer structure, creating one if necessary.
 536          */
 537 again:
 538         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 539                            buf_offset);
 540         if (buffer == NULL) {
 541                 ++hammer_count_buffers;
 542                 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 543                 buffer->zone2_offset = buf_offset;
 544                 buffer->volume = volume;
 545
 546                 hammer_io_init(&buffer->io, hmp, iotype);
 547                 buffer->io.offset = volume->ondisk->vol_buf_beg +
 548                                     (buf_offset & HAMMER_OFF_SHORT_MASK);
 549                 TAILQ_INIT(&buffer->clist);
 550                 hammer_ref(&buffer->io.lock);
 551
 552                 /*
 553                  * Insert the buffer into the RB tree and handle late
 554                  * collisions.
 555                  */
 556                 if (RB_INSERT(hammer_buf_rb_tree, &volume->rb_bufs_root, buffer)) {
 557                         hammer_unref(&buffer->io.lock);
 558                         --hammer_count_buffers;
 559                         kfree(buffer, M_HAMMER);
 560                         goto again;
 561                 }
 562                 hammer_ref(&volume->io.lock);
 563         } else {
 564                 hammer_ref(&buffer->io.lock);
 565
 566                 /*
 567                  * The buffer is no longer loose if it has a ref.
 568                  */
 569                 if (buffer->io.mod_list == &hmp->lose_list) {
 570                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 571                                      mod_entry);
 572                         buffer->io.mod_list = NULL;
 573                 }
 574                 if (buffer->io.lock.refs == 1)
 575                         hammer_io_reinit(&buffer->io, iotype);
 576                 else
 577                         KKASSERT(buffer->io.type == iotype);
 578         }
 579
 580         /*
 581          * Cache the blockmap translation
 582          */
 583         if ((zoneX_offset & HAMMER_ZONE_RAW_BUFFER) != HAMMER_ZONE_RAW_BUFFER)
 584                 buffer->zoneX_offset = zoneX_offset;
 585
 586         /*
 587          * Deal with on-disk info
 588          */
 589         if (buffer->ondisk == NULL || buffer->io.loading) {
 590                 *errorp = hammer_load_buffer(buffer, isnew);
 591                 if (*errorp) {
 592                         hammer_rel_buffer(buffer, 1);
 593                         buffer = NULL;
 594                 }
 595         } else {
 596                 *errorp = 0;
 597         }
 598         hammer_rel_volume(volume, 0);
 599         return(buffer);
 600 }
 601
 602 static int
 603 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 604 {
 605         hammer_volume_t volume;
 606         int error;
 607
 608         /*
 609          * Load the buffer's on-disk info
 610          */
 611         volume = buffer->volume;
 612         ++buffer->io.loading;
 613         hammer_lock_ex(&buffer->io.lock);
 614
 615         if (buffer->ondisk == NULL) {
 616                 if (isnew) {
 617                         error = hammer_io_new(volume->devvp, &buffer->io);
 618                 } else {
 619                         error = hammer_io_read(volume->devvp, &buffer->io);
 620                 }
 621                 if (error == 0)
 622                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 623         } else if (isnew) {
 624                 error = hammer_io_new(volume->devvp, &buffer->io);
 625         } else {
 626                 error = 0;
 627         }
 628         --buffer->io.loading;
 629         hammer_unlock(&buffer->io.lock);
 630         return (error);
 631 }
 632
 633 /*
 634  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 635  */
 636 int
 637 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 638 {
 639         hammer_ref(&buffer->io.lock);
 640         hammer_flush_buffer_nodes(buffer);
 641         KKASSERT(buffer->io.lock.refs == 1);
 642         hammer_rel_buffer(buffer, 2);
 643         return(0);
 644 }
 645
 646 /*
 647  * Reference a buffer that is either already referenced or via a specially
 648  * handled pointer (aka cursor->buffer).
 649  */
 650 int
 651 hammer_ref_buffer(hammer_buffer_t buffer)
 652 {
 653         int error;
 654
 655         hammer_ref(&buffer->io.lock);
 656
 657         /*
 658          * No longer loose
 659          */
 660         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 661                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 662                 buffer->io.mod_list = NULL;
 663         }
 664
 665         if (buffer->ondisk == NULL || buffer->io.loading) {
 666                 error = hammer_load_buffer(buffer, 0);
 667                 if (error) {
 668                         hammer_rel_buffer(buffer, 1);
 669                         /*
 670                          * NOTE: buffer pointer can become stale after
 671                          * the above release.
 672                          */
 673                 }
 674         } else {
 675                 error = 0;
 676         }
 677         return(error);
 678 }
 679
 680 /*
 681  * Release a buffer.  We have to deal with several places where
 682  * another thread can ref the buffer.
 683  *
 684  * Only destroy the structure itself if the related buffer cache buffer
 685  * was disassociated from it.  This ties the management of the structure
 686  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 687  * embedded io is referenced or not.
 688  */
 689 void
 690 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 691 {
 692         hammer_volume_t volume;
 693         int freeme = 0;
 694
 695         crit_enter();
 696         if (buffer->io.lock.refs == 1) {
 697                 ++buffer->io.loading;   /* force interlock check */
 698                 hammer_lock_ex(&buffer->io.lock);
 699                 if (buffer->io.lock.refs == 1) {
 700                         hammer_io_release(&buffer->io, flush);
 701                         hammer_flush_buffer_nodes(buffer);
 702                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 703
 704                         if (buffer->io.bp == NULL &&
 705                             buffer->io.lock.refs == 1) {
 706                                 /*
 707                                  * Final cleanup
 708                                  */
 709                                 volume = buffer->volume;
 710                                 RB_REMOVE(hammer_buf_rb_tree,
 711                                           &volume->rb_bufs_root, buffer);
 712                                 buffer->volume = NULL; /* sanity */
 713                                 hammer_rel_volume(volume, 0);
 714                                 freeme = 1;
 715                         }
 716                 }
 717                 --buffer->io.loading;
 718                 hammer_unlock(&buffer->io.lock);
 719         }
 720         hammer_unref(&buffer->io.lock);
 721         crit_exit();
 722         if (freeme) {
 723                 KKASSERT(buffer->io.mod_list == NULL);
 724                 --hammer_count_buffers;
 725                 kfree(buffer, M_HAMMER);
 726         }
 727 }
 728
 729 /*
 730  * Remove the zoneX translation cache for a buffer given its zone-2 offset.
 731  */
 732 void
 733 hammer_uncache_buffer(hammer_mount_t hmp, hammer_off_t buf_offset)
 734 {
 735         hammer_volume_t volume;
 736         hammer_buffer_t buffer;
 737         int vol_no;
 738         int error;
 739
 740         buf_offset &= ~HAMMER_BUFMASK64;
 741         KKASSERT((buf_offset & HAMMER_ZONE_RAW_BUFFER) ==
 742                  HAMMER_ZONE_RAW_BUFFER);
 743         vol_no = HAMMER_VOL_DECODE(buf_offset);
 744         volume = hammer_get_volume(hmp, vol_no, &error);
 745         KKASSERT(volume != 0);
 746         KKASSERT(buf_offset < volume->maxbuf_off);
 747
 748         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 749                            buf_offset);
 750         if (buffer)
 751                 buffer->zoneX_offset = 0;
 752         hammer_rel_volume(volume, 0);
 753 }
 754
 755 /*
 756  * Access the filesystem buffer containing the specified hammer offset.
 757  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 758  * relative buffer offset.  It must also have bit 55 set to be valid.
 759  * (see hammer_off_t in hammer_disk.h).
 760  *
 761  * Any prior buffer in *bufferp will be released and replaced by the
 762  * requested buffer.
 763  */
 764 void *
 765 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 766              struct hammer_buffer **bufferp)
 767 {
 768         hammer_buffer_t buffer;
 769         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 770
 771         buf_offset &= ~HAMMER_BUFMASK64;
 772         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 773
 774         buffer = *bufferp;
 775         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 776                                buffer->zoneX_offset != buf_offset)) {
 777                 if (buffer)
 778                         hammer_rel_buffer(buffer, 0);
 779                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 780                 *bufferp = buffer;
 781         } else {
 782                 *errorp = 0;
 783         }
 784
 785         /*
 786          * Return a pointer to the buffer data.
 787          */
 788         if (buffer == NULL)
 789                 return(NULL);
 790         else
 791                 return((char *)buffer->ondisk + xoff);
 792 }
 793
 794 /*
 795  * Access the filesystem buffer containing the specified hammer offset.
 796  * No disk read operation occurs.  The result buffer may contain garbage.
 797  *
 798  * Any prior buffer in *bufferp will be released and replaced by the
 799  * requested buffer.
 800  *
 801  * This function marks the buffer dirty but does not increment its
 802  * modify_refs count.
 803  */
 804 void *
 805 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 806              struct hammer_buffer **bufferp)
 807 {
 808         hammer_buffer_t buffer;
 809         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 810
 811         buf_offset &= ~HAMMER_BUFMASK64;
 812
 813         buffer = *bufferp;
 814         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 815                                buffer->zoneX_offset != buf_offset)) {
 816                 if (buffer)
 817                         hammer_rel_buffer(buffer, 0);
 818                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 819                 *bufferp = buffer;
 820         } else {
 821                 *errorp = 0;
 822         }
 823
 824         /*
 825          * Return a pointer to the buffer data.
 826          */
 827         if (buffer == NULL)
 828                 return(NULL);
 829         else
 830                 return((char *)buffer->ondisk + xoff);
 831 }
 832
 833 /************************************************************************
 834  *                              NODES                                   *
 835  ************************************************************************
 836  *
 837  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 838  * method used by the HAMMER filesystem.
 839  *
 840  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 841  * associated with its buffer, and will only referenced the buffer while
 842  * the node itself is referenced.
 843  *
 844  * A hammer_node can also be passively associated with other HAMMER
 845  * structures, such as inodes, while retaining 0 references.  These
 846  * associations can be cleared backwards using a pointer-to-pointer in
 847  * the hammer_node.
 848  *
 849  * This allows the HAMMER implementation to cache hammer_nodes long-term
 850  * and short-cut a great deal of the infrastructure's complexity.  In
 851  * most cases a cached node can be reacquired without having to dip into
 852  * either the buffer or cluster management code.
 853  *
 854  * The caller must pass a referenced cluster on call and will retain
 855  * ownership of the reference on return.  The node will acquire its own
 856  * additional references, if necessary.
 857  */
 858 hammer_node_t
 859 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 860                 int isnew, int *errorp)
 861 {
 862         hammer_node_t node;
 863
 864         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 865
 866         /*
 867          * Locate the structure, allocating one if necessary.
 868          */
 869 again:
 870         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 871         if (node == NULL) {
 872                 ++hammer_count_nodes;
 873                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 874                 node->node_offset = node_offset;
 875                 node->hmp = hmp;
 876                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 877                         --hammer_count_nodes;
 878                         kfree(node, M_HAMMER);
 879                         goto again;
 880                 }
 881         }
 882         hammer_ref(&node->lock);
 883         if (node->ondisk)
 884                 *errorp = 0;
 885         else
 886                 *errorp = hammer_load_node(node, isnew);
 887         if (*errorp) {
 888                 hammer_rel_node(node);
 889                 node = NULL;
 890         }
 891         return(node);
 892 }
 893
 894 /*
 895  * Reference an already-referenced node.
 896  */
 897 void
 898 hammer_ref_node(hammer_node_t node)
 899 {
 900         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 901         hammer_ref(&node->lock);
 902 }
 903
 904 /*
 905  * Load a node's on-disk data reference.
 906  */
 907 static int
 908 hammer_load_node(hammer_node_t node, int isnew)
 909 {
 910         hammer_buffer_t buffer;
 911         int error;
 912
 913         error = 0;
 914         ++node->loading;
 915         hammer_lock_ex(&node->lock);
 916         if (node->ondisk == NULL) {
 917                 /*
 918                  * This is a little confusing but the jist is that
 919                  * node->buffer determines whether the node is on
 920                  * the buffer's clist and node->ondisk determines
 921                  * whether the buffer is referenced.
 922                  *
 923                  * We could be racing a buffer release, in which case
 924                  * node->buffer may become NULL while we are blocked
 925                  * referencing the buffer.
 926                  */
 927                 if ((buffer = node->buffer) != NULL) {
 928                         error = hammer_ref_buffer(buffer);
 929                         if (error == 0 && node->buffer == NULL) {
 930                                 TAILQ_INSERT_TAIL(&buffer->clist,
 931                                                   node, entry);
 932                                 node->buffer = buffer;
 933                         }
 934                 } else {
 935                         buffer = hammer_get_buffer(node->hmp,
 936                                                    node->node_offset, 0,
 937                                                    &error);
 938                         if (buffer) {
 939                                 KKASSERT(error == 0);
 940                                 TAILQ_INSERT_TAIL(&buffer->clist,
 941                                                   node, entry);
 942                                 node->buffer = buffer;
 943                         }
 944                 }
 945                 if (error == 0) {
 946                         node->ondisk = (void *)((char *)buffer->ondisk +
 947                                (node->node_offset & HAMMER_BUFMASK));
 948                         if (isnew == 0 &&
 949                             hammer_crc_test_btree(node->ondisk) == 0) {
 950                                 Debugger("CRC FAILED: B-TREE NODE");
 951                         }
 952                 }
 953         }
 954         --node->loading;
 955         hammer_unlock(&node->lock);
 956         return (error);
 957 }
 958
 959 /*
 960  * Safely reference a node, interlock against flushes via the IO subsystem.
 961  */
 962 hammer_node_t
 963 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
 964                      int *errorp)
 965 {
 966         hammer_node_t node;
 967
 968         node = *cache;
 969         if (node != NULL) {
 970                 hammer_ref(&node->lock);
 971                 if (node->ondisk)
 972                         *errorp = 0;
 973                 else
 974                         *errorp = hammer_load_node(node, 0);
 975                 if (*errorp) {
 976                         hammer_rel_node(node);
 977                         node = NULL;
 978                 }
 979         } else {
 980                 *errorp = ENOENT;
 981         }
 982         return(node);
 983 }
 984
 985 /*
 986  * Release a hammer_node.  On the last release the node dereferences
 987  * its underlying buffer and may or may not be destroyed.
 988  */
 989 void
 990 hammer_rel_node(hammer_node_t node)
 991 {
 992         hammer_buffer_t buffer;
 993
 994         /*
 995          * If this isn't the last ref just decrement the ref count and
 996          * return.
 997          */
 998         if (node->lock.refs > 1) {
 999                 hammer_unref(&node->lock);
1000                 return;
1001         }
1002
1003         /*
1004          * If there is no ondisk info or no buffer the node failed to load,
1005          * remove the last reference and destroy the node.
1006          */
1007         if (node->ondisk == NULL) {
1008                 hammer_unref(&node->lock);
1009                 hammer_flush_node(node);
1010                 /* node is stale now */
1011                 return;
1012         }
1013
1014         /*
1015          * Do final cleanups and then either destroy the node and leave it
1016          * passively cached.  The buffer reference is removed regardless.
1017          */
1018         buffer = node->buffer;
1019         node->ondisk = NULL;
1020
1021         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1022                 hammer_unref(&node->lock);
1023                 hammer_rel_buffer(buffer, 0);
1024                 return;
1025         }
1026
1027         /*
1028          * Destroy the node.
1029          */
1030         hammer_unref(&node->lock);
1031         hammer_flush_node(node);
1032         /* node is stale */
1033         hammer_rel_buffer(buffer, 0);
1034 }
1035
1036 /*
1037  * Free space on-media associated with a B-Tree node.
1038  */
1039 void
1040 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1041 {
1042         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1043         node->flags |= HAMMER_NODE_DELETED;
1044         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1045 }
1046
1047 /*
1048  * Passively cache a referenced hammer_node in *cache.  The caller may
1049  * release the node on return.
1050  */
1051 void
1052 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1053 {
1054         hammer_node_t old;
1055
1056         /*
1057          * If the node is being deleted, don't cache it!
1058          */
1059         if (node->flags & HAMMER_NODE_DELETED)
1060                 return;
1061
1062         /*
1063          * Cache the node.  If we previously cached a different node we
1064          * have to give HAMMER a chance to destroy it.
1065          */
1066 again:
1067         if (node->cache1 != cache) {
1068                 if (node->cache2 != cache) {
1069                         if ((old = *cache) != NULL) {
1070                                 KKASSERT(node->lock.refs != 0);
1071                                 hammer_uncache_node(cache);
1072                                 goto again;
1073                         }
1074                         if (node->cache2)
1075                                 *node->cache2 = NULL;
1076                         node->cache2 = node->cache1;
1077                         node->cache1 = cache;
1078                         *cache = node;
1079                 } else {
1080                         struct hammer_node **tmp;
1081                         tmp = node->cache1;
1082                         node->cache1 = node->cache2;
1083                         node->cache2 = tmp;
1084                 }
1085         }
1086 }
1087
1088 void
1089 hammer_uncache_node(struct hammer_node **cache)
1090 {
1091         hammer_node_t node;
1092
1093         if ((node = *cache) != NULL) {
1094                 *cache = NULL;
1095                 if (node->cache1 == cache) {
1096                         node->cache1 = node->cache2;
1097                         node->cache2 = NULL;
1098                 } else if (node->cache2 == cache) {
1099                         node->cache2 = NULL;
1100                 } else {
1101                         panic("hammer_uncache_node: missing cache linkage");
1102                 }
1103                 if (node->cache1 == NULL && node->cache2 == NULL)
1104                         hammer_flush_node(node);
1105         }
1106 }
1107
1108 /*
1109  * Remove a node's cache references and destroy the node if it has no
1110  * other references or backing store.
1111  */
1112 void
1113 hammer_flush_node(hammer_node_t node)
1114 {
1115         hammer_buffer_t buffer;
1116
1117         if (node->cache1)
1118                 *node->cache1 = NULL;
1119         if (node->cache2)
1120                 *node->cache2 = NULL;
1121         if (node->lock.refs == 0 && node->ondisk == NULL) {
1122                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1123                 if ((buffer = node->buffer) != NULL) {
1124                         node->buffer = NULL;
1125                         TAILQ_REMOVE(&buffer->clist, node, entry);
1126                         /* buffer is unreferenced because ondisk is NULL */
1127                 }
1128                 --hammer_count_nodes;
1129                 kfree(node, M_HAMMER);
1130         }
1131 }
1132
1133 /*
1134  * Flush passively cached B-Tree nodes associated with this buffer.
1135  * This is only called when the buffer is about to be destroyed, so
1136  * none of the nodes should have any references.  The buffer is locked.
1137  *
1138  * We may be interlocked with the buffer.
1139  */
1140 void
1141 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1142 {
1143         hammer_node_t node;
1144
1145         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1146                 KKASSERT(node->ondisk == NULL);
1147
1148                 if (node->lock.refs == 0) {
1149                         hammer_ref(&node->lock);
1150                         node->flags |= HAMMER_NODE_FLUSH;
1151                         hammer_rel_node(node);
1152                 } else {
1153                         KKASSERT(node->loading != 0);
1154                         KKASSERT(node->buffer != NULL);
1155                         buffer = node->buffer;
1156                         node->buffer = NULL;
1157                         TAILQ_REMOVE(&buffer->clist, node, entry);
1158                         /* buffer is unreferenced because ondisk is NULL */
1159                 }
1160         }
1161 }
1162
1163
1164 /************************************************************************
1165  *                              ALLOCATORS                              *
1166  ************************************************************************/
1167
1168 /*
1169  * Allocate a B-Tree node.
1170  */
1171 hammer_node_t
1172 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1173 {
1174         hammer_buffer_t buffer = NULL;
1175         hammer_node_t node = NULL;
1176         hammer_off_t node_offset;
1177
1178         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1179                                             sizeof(struct hammer_node_ondisk),
1180                                             errorp);
1181         if (*errorp == 0) {
1182                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1183                 hammer_modify_node_noundo(trans, node);
1184                 bzero(node->ondisk, sizeof(*node->ondisk));
1185                 hammer_modify_node_done(node);
1186         }
1187         if (buffer)
1188                 hammer_rel_buffer(buffer, 0);
1189         return(node);
1190 }
1191
1192 #if 0
1193
1194 /*
1195  * The returned buffers are already appropriately marked as being modified.
1196  * If the caller marks them again unnecessary undo records may be generated.
1197  *
1198  * In-band data is indicated by data_bufferp == NULL.  Pass a data_len of 0
1199  * for zero-fill (caller modifies data_len afterwords).
1200  *
1201  * If the caller is responsible for calling hammer_modify_*() prior to making
1202  * any additional modifications to either the returned record buffer or the
1203  * returned data buffer.
1204  */
1205 void *
1206 hammer_alloc_record(hammer_transaction_t trans,
1207                     hammer_off_t *rec_offp, u_int16_t rec_type,
1208                     struct hammer_buffer **rec_bufferp,
1209                     int32_t data_len, void **datap,
1210                     hammer_off_t *data_offp,
1211                     struct hammer_buffer **data_bufferp, int *errorp)
1212 {
1213         hammer_record_ondisk_t rec;
1214         hammer_off_t rec_offset;
1215         hammer_off_t data_offset;
1216         int32_t reclen;
1217
1218         if (datap)
1219                 *datap = NULL;
1220
1221         /*
1222          * Allocate the record
1223          */
1224         rec_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_RECORD_INDEX,
1225                                            HAMMER_RECORD_SIZE, errorp);
1226         if (*errorp)
1227                 return(NULL);
1228         if (data_offp)
1229                 *data_offp = 0;
1230
1231         /*
1232          * Allocate data
1233          */
1234         if (data_len) {
1235                 if (data_bufferp == NULL) {
1236                         switch(rec_type) {
1237                         case HAMMER_RECTYPE_DATA:
1238                                 reclen = offsetof(struct hammer_data_record,
1239                                                   data[0]);
1240                                 break;
1241                         case HAMMER_RECTYPE_DIRENTRY:
1242                                 reclen = offsetof(struct hammer_entry_record,
1243                                                   name[0]);
1244                                 break;
1245                         default:
1246                                 panic("hammer_alloc_record: illegal "
1247                                       "in-band data");
1248                                 /* NOT REACHED */
1249                                 reclen = 0;
1250                                 break;
1251                         }
1252                         KKASSERT(reclen + data_len <= HAMMER_RECORD_SIZE);
1253                         data_offset = rec_offset + reclen;
1254                 } else if (data_len < HAMMER_BUFSIZE) {
1255                         data_offset = hammer_blockmap_alloc(trans,
1256                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1257                                                 data_len, errorp);
1258                         *data_offp = data_offset;
1259                 } else {
1260                         data_offset = hammer_blockmap_alloc(trans,
1261                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1262                                                 data_len, errorp);
1263                         *data_offp = data_offset;
1264                 }
1265         } else {
1266                 data_offset = 0;
1267         }
1268         if (*errorp) {
1269                 hammer_blockmap_free(trans, rec_offset, HAMMER_RECORD_SIZE);
1270                 return(NULL);
1271         }
1272
1273         /*
1274          * Basic return values.
1275          *
1276          * Note that because this is a 'new' buffer, there is no need to
1277          * generate UNDO records for it.
1278          */
1279         *rec_offp = rec_offset;
1280         rec = hammer_bread(trans->hmp, rec_offset, errorp, rec_bufferp);
1281         hammer_modify_buffer(trans, *rec_bufferp, NULL, 0);
1282         bzero(rec, sizeof(*rec));
1283         KKASSERT(*errorp == 0);
1284         rec->base.data_off = data_offset;
1285         rec->base.data_len = data_len;
1286         hammer_modify_buffer_done(*rec_bufferp);
1287
1288         if (data_bufferp) {
1289                 if (data_len) {
1290                         *datap = hammer_bread(trans->hmp, data_offset, errorp,
1291                                               data_bufferp);
1292                         KKASSERT(*errorp == 0);
1293                 } else {
1294                         *datap = NULL;
1295                 }
1296         } else if (data_len) {
1297                 KKASSERT(data_offset + data_len - rec_offset <=
1298                          HAMMER_RECORD_SIZE);
1299                 if (datap) {
1300                         *datap = (void *)((char *)rec +
1301                                           (int32_t)(data_offset - rec_offset));
1302                 }
1303         } else {
1304                 KKASSERT(datap == NULL);
1305         }
1306         KKASSERT(*errorp == 0);
1307         return(rec);
1308 }
1309
1310 #endif
1311
1312 /*
1313  * Allocate data.  If the address of a data buffer is supplied then
1314  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1315  * will be set to the related buffer.  The caller must release it when
1316  * finally done.  The initial *data_bufferp should be set to NULL by
1317  * the caller.
1318  *
1319  * The caller is responsible for making hammer_modify*() calls on the
1320  * *data_bufferp.
1321  */
1322 void *
1323 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1324                   hammer_off_t *data_offsetp,
1325                   struct hammer_buffer **data_bufferp, int *errorp)
1326 {
1327         void *data;
1328
1329         /*
1330          * Allocate data
1331          */
1332         if (data_len) {
1333                 if (data_len < HAMMER_BUFSIZE) {
1334                         *data_offsetp = hammer_blockmap_alloc(trans,
1335                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1336                                                 data_len, errorp);
1337                 } else {
1338                         *data_offsetp = hammer_blockmap_alloc(trans,
1339                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1340                                                 data_len, errorp);
1341                 }
1342         } else {
1343                 *data_offsetp = 0;
1344         }
1345         if (*errorp == 0 && data_bufferp) {
1346                 if (data_len) {
1347                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1348                                             data_bufferp);
1349                         KKASSERT(*errorp == 0);
1350                 } else {
1351                         data = NULL;
1352                 }
1353         } else {
1354                 data = NULL;
1355         }
1356         KKASSERT(*errorp == 0);
1357         return(data);
1358 }
1359
1360 /*
1361  * Sync dirty buffers to the media and clean-up any loose ends.
1362  */
1363 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1364 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1365
1366 int
1367 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1368 {
1369         struct hammer_sync_info info;
1370
1371         info.error = 0;
1372         info.waitfor = waitfor;
1373         if (waitfor == MNT_WAIT) {
1374                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1375                               hammer_sync_scan1, hammer_sync_scan2, &info);
1376         } else {
1377                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1378                               hammer_sync_scan1, hammer_sync_scan2, &info);
1379         }
1380         return(info.error);
1381 }
1382
1383 int
1384 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1385 {
1386         struct hammer_sync_info info;
1387
1388         info.error = 0;
1389         info.waitfor = waitfor;
1390
1391         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1392                       hammer_sync_scan1, hammer_sync_scan2, &info);
1393         if (waitfor == MNT_WAIT)
1394                 hammer_flusher_sync(hmp);
1395         else
1396                 hammer_flusher_async(hmp);
1397
1398         return(info.error);
1399 }
1400
1401 static int
1402 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1403 {
1404         struct hammer_inode *ip;
1405
1406         ip = VTOI(vp);
1407         if (vp->v_type == VNON || ip == NULL ||
1408             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1409              RB_EMPTY(&vp->v_rbdirty_tree))) {
1410                 return(-1);
1411         }
1412         return(0);
1413 }
1414
1415 static int
1416 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1417 {
1418         struct hammer_sync_info *info = data;
1419         struct hammer_inode *ip;
1420         int error;
1421
1422         ip = VTOI(vp);
1423         if (vp->v_type == VNON || vp->v_type == VBAD ||
1424             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1425              RB_EMPTY(&vp->v_rbdirty_tree))) {
1426                 return(0);
1427         }
1428         error = VOP_FSYNC(vp, info->waitfor);
1429         if (error)
1430                 info->error = error;
1431         return(0);
1432 }
1433