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 1.49 2008/06/03 18:47:25 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         volume->maxraw_off = ondisk->vol_buf_end;
 219         RB_INIT(&volume->rb_bufs_root);
 220
 221         if (RB_EMPTY(&hmp->rb_vols_root)) {
 222                 hmp->fsid = ondisk->vol_fsid;
 223         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 224                 kprintf("hammer_mount: volume %s's fsid does not match "
 225                         "other volumes\n", volume->vol_name);
 226                 error = EFTYPE;
 227                 goto late_failure;
 228         }
 229
 230         /*
 231          * Insert the volume structure into the red-black tree.
 232          */
 233         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 234                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 235                         volume->vol_name, volume->vol_no);
 236                 error = EEXIST;
 237         }
 238
 239         /*
 240          * Set the root volume .  HAMMER special cases rootvol the structure.
 241          * We do not hold a ref because this would prevent related I/O
 242          * from being flushed.
 243          */
 244         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 245                 hmp->rootvol = volume;
 246                 if (bp) {
 247                         brelse(bp);
 248                         bp = NULL;
 249                 }
 250                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 251                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 252                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 253                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 254                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 255         }
 256 late_failure:
 257         if (bp)
 258                 brelse(bp);
 259         if (error) {
 260                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 261                 if (setmp)
 262                         volume->devvp->v_rdev->si_mountpoint = NULL;
 263                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 264                 hammer_free_volume(volume);
 265         }
 266         return (error);
 267 }
 268
 269 /*
 270  * This is called for each volume when updating the mount point from
 271  * read-write to read-only or vise-versa.
 272  */
 273 int
 274 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 275 {
 276         if (volume->devvp) {
 277                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 278                 if (volume->io.hmp->ronly) {
 279                         /* do not call vinvalbuf */
 280                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 281                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 282                 } else {
 283                         /* do not call vinvalbuf */
 284                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 285                         VOP_CLOSE(volume->devvp, FREAD);
 286                 }
 287                 vn_unlock(volume->devvp);
 288         }
 289         return(0);
 290 }
 291
 292 /*
 293  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 294  * so returns -1 on failure.
 295  */
 296 int
 297 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 298 {
 299         struct hammer_mount *hmp = volume->io.hmp;
 300         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 301
 302         /*
 303          * Clean up the root volume pointer, which is held unlocked in hmp.
 304          */
 305         if (hmp->rootvol == volume)
 306                 hmp->rootvol = NULL;
 307
 308         /*
 309          * Unload buffers.
 310          */
 311         RB_SCAN(hammer_buf_rb_tree, &volume->rb_bufs_root, NULL,
 312                         hammer_unload_buffer, NULL);
 313
 314         /*
 315          * Release our buffer and flush anything left in the buffer cache.
 316          */
 317         volume->io.waitdep = 1;
 318         hammer_io_release(&volume->io, 1);
 319
 320         /*
 321          * There should be no references on the volume, no clusters, and
 322          * no super-clusters.
 323          */
 324         KKASSERT(volume->io.lock.refs == 0);
 325         KKASSERT(RB_EMPTY(&volume->rb_bufs_root));
 326
 327         volume->ondisk = NULL;
 328         if (volume->devvp) {
 329                 if (volume->devvp->v_rdev &&
 330                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 331                 ) {
 332                         volume->devvp->v_rdev->si_mountpoint = NULL;
 333                 }
 334                 if (ronly) {
 335                         vinvalbuf(volume->devvp, 0, 0, 0);
 336                         VOP_CLOSE(volume->devvp, FREAD);
 337                 } else {
 338                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 339                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 340                 }
 341         }
 342
 343         /*
 344          * Destroy the structure
 345          */
 346         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 347         hammer_free_volume(volume);
 348         return(0);
 349 }
 350
 351 static
 352 void
 353 hammer_free_volume(hammer_volume_t volume)
 354 {
 355         if (volume->vol_name) {
 356                 kfree(volume->vol_name, M_HAMMER);
 357                 volume->vol_name = NULL;
 358         }
 359         if (volume->devvp) {
 360                 vrele(volume->devvp);
 361                 volume->devvp = NULL;
 362         }
 363         --hammer_count_volumes;
 364         kfree(volume, M_HAMMER);
 365 }
 366
 367 /*
 368  * Get a HAMMER volume.  The volume must already exist.
 369  */
 370 hammer_volume_t
 371 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 372 {
 373         struct hammer_volume *volume;
 374
 375         /*
 376          * Locate the volume structure
 377          */
 378         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 379         if (volume == NULL) {
 380                 *errorp = ENOENT;
 381                 return(NULL);
 382         }
 383         hammer_ref(&volume->io.lock);
 384
 385         /*
 386          * Deal with on-disk info
 387          */
 388         if (volume->ondisk == NULL || volume->io.loading) {
 389                 *errorp = hammer_load_volume(volume);
 390                 if (*errorp) {
 391                         hammer_rel_volume(volume, 1);
 392                         volume = NULL;
 393                 }
 394         } else {
 395                 *errorp = 0;
 396         }
 397         return(volume);
 398 }
 399
 400 int
 401 hammer_ref_volume(hammer_volume_t volume)
 402 {
 403         int error;
 404
 405         hammer_ref(&volume->io.lock);
 406
 407         /*
 408          * Deal with on-disk info
 409          */
 410         if (volume->ondisk == NULL || volume->io.loading) {
 411                 error = hammer_load_volume(volume);
 412                 if (error)
 413                         hammer_rel_volume(volume, 1);
 414         } else {
 415                 error = 0;
 416         }
 417         return (error);
 418 }
 419
 420 hammer_volume_t
 421 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 422 {
 423         hammer_volume_t volume;
 424
 425         volume = hmp->rootvol;
 426         KKASSERT(volume != NULL);
 427         hammer_ref(&volume->io.lock);
 428
 429         /*
 430          * Deal with on-disk info
 431          */
 432         if (volume->ondisk == NULL || volume->io.loading) {
 433                 *errorp = hammer_load_volume(volume);
 434                 if (*errorp) {
 435                         hammer_rel_volume(volume, 1);
 436                         volume = NULL;
 437                 }
 438         } else {
 439                 *errorp = 0;
 440         }
 441         return (volume);
 442 }
 443
 444 /*
 445  * Load a volume's on-disk information.  The volume must be referenced and
 446  * not locked.  We temporarily acquire an exclusive lock to interlock
 447  * against releases or multiple get's.
 448  */
 449 static int
 450 hammer_load_volume(hammer_volume_t volume)
 451 {
 452         int error;
 453
 454         ++volume->io.loading;
 455         hammer_lock_ex(&volume->io.lock);
 456
 457         if (volume->ondisk == NULL) {
 458                 error = hammer_io_read(volume->devvp, &volume->io,
 459                                        volume->maxraw_off);
 460                 if (error == 0)
 461                         volume->ondisk = (void *)volume->io.bp->b_data;
 462         } else {
 463                 error = 0;
 464         }
 465         --volume->io.loading;
 466         hammer_unlock(&volume->io.lock);
 467         return(error);
 468 }
 469
 470 /*
 471  * Release a volume.  Call hammer_io_release on the last reference.  We have
 472  * to acquire an exclusive lock to interlock against volume->ondisk tests
 473  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 474  * lock to be held.
 475  *
 476  * Volumes are not unloaded from memory during normal operation.
 477  */
 478 void
 479 hammer_rel_volume(hammer_volume_t volume, int flush)
 480 {
 481         crit_enter();
 482         if (volume->io.lock.refs == 1) {
 483                 ++volume->io.loading;
 484                 hammer_lock_ex(&volume->io.lock);
 485                 if (volume->io.lock.refs == 1) {
 486                         volume->ondisk = NULL;
 487                         hammer_io_release(&volume->io, flush);
 488                 }
 489                 --volume->io.loading;
 490                 hammer_unlock(&volume->io.lock);
 491         }
 492         hammer_unref(&volume->io.lock);
 493         crit_exit();
 494 }
 495
 496 /************************************************************************
 497  *                              BUFFERS                                 *
 498  ************************************************************************
 499  *
 500  * Manage buffers.  Currently all blockmap-backed zones are translated
 501  * to zone-2 buffer offsets.
 502  */
 503 hammer_buffer_t
 504 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 505                   int isnew, int *errorp)
 506 {
 507         hammer_buffer_t buffer;
 508         hammer_volume_t volume;
 509         hammer_off_t    zoneX_offset;
 510         hammer_io_type_t iotype;
 511         int vol_no;
 512         int zone;
 513
 514         zoneX_offset = buf_offset;
 515         zone = HAMMER_ZONE_DECODE(buf_offset);
 516
 517         /*
 518          * What is the buffer class?
 519          */
 520         switch(zone) {
 521         case HAMMER_ZONE_LARGE_DATA_INDEX:
 522         case HAMMER_ZONE_SMALL_DATA_INDEX:
 523                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 524                 break;
 525         case HAMMER_ZONE_UNDO_INDEX:
 526                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 527                 break;
 528         default:
 529                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 530                 break;
 531         }
 532
 533         /*
 534          * Handle blockmap offset translations
 535          */
 536         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 537                 buf_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 538                 KKASSERT(*errorp == 0);
 539         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 540                 buf_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 541                 KKASSERT(*errorp == 0);
 542         }
 543
 544         /*
 545          * Locate the buffer given its zone-2 offset.
 546          */
 547         buf_offset &= ~HAMMER_BUFMASK64;
 548         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 549         vol_no = HAMMER_VOL_DECODE(buf_offset);
 550         volume = hammer_get_volume(hmp, vol_no, errorp);
 551         if (volume == NULL)
 552                 return(NULL);
 553
 554         /*
 555          * NOTE: buf_offset and maxbuf_off are both full zone-2 offset
 556          * specifications.
 557          */
 558         KKASSERT(buf_offset < volume->maxbuf_off);
 559
 560         /*
 561          * Locate and lock the buffer structure, creating one if necessary.
 562          */
 563 again:
 564         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 565                            buf_offset);
 566         if (buffer == NULL) {
 567                 ++hammer_count_buffers;
 568                 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 569                 buffer->zone2_offset = buf_offset;
 570                 buffer->volume = volume;
 571
 572                 hammer_io_init(&buffer->io, hmp, iotype);
 573                 buffer->io.offset = volume->ondisk->vol_buf_beg +
 574                                     (buf_offset & HAMMER_OFF_SHORT_MASK);
 575                 TAILQ_INIT(&buffer->clist);
 576                 hammer_ref(&buffer->io.lock);
 577
 578                 /*
 579                  * Insert the buffer into the RB tree and handle late
 580                  * collisions.
 581                  */
 582                 if (RB_INSERT(hammer_buf_rb_tree, &volume->rb_bufs_root, buffer)) {
 583                         hammer_unref(&buffer->io.lock);
 584                         --hammer_count_buffers;
 585                         kfree(buffer, M_HAMMER);
 586                         goto again;
 587                 }
 588                 hammer_ref(&volume->io.lock);
 589         } else {
 590                 hammer_ref(&buffer->io.lock);
 591
 592                 /*
 593                  * The buffer is no longer loose if it has a ref.
 594                  */
 595                 if (buffer->io.mod_list == &hmp->lose_list) {
 596                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 597                                      mod_entry);
 598                         buffer->io.mod_list = NULL;
 599                 }
 600                 if (buffer->io.lock.refs == 1)
 601                         hammer_io_reinit(&buffer->io, iotype);
 602                 else
 603                         KKASSERT(buffer->io.type == iotype);
 604         }
 605
 606         /*
 607          * Cache the blockmap translation
 608          */
 609         if ((zoneX_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_RAW_BUFFER)
 610                 buffer->zoneX_offset = zoneX_offset;
 611
 612         /*
 613          * Deal with on-disk info
 614          */
 615         if (buffer->ondisk == NULL || buffer->io.loading) {
 616                 *errorp = hammer_load_buffer(buffer, isnew);
 617                 if (*errorp) {
 618                         hammer_rel_buffer(buffer, 1);
 619                         buffer = NULL;
 620                 }
 621         } else {
 622                 *errorp = 0;
 623         }
 624         hammer_rel_volume(volume, 0);
 625         return(buffer);
 626 }
 627
 628 /*
 629  * Clear the cached zone-X translation for a buffer.
 630  */
 631 void
 632 hammer_clrxlate_buffer(hammer_mount_t hmp, hammer_off_t buf_offset)
 633 {
 634         hammer_buffer_t buffer;
 635         hammer_volume_t volume;
 636         int vol_no;
 637         int error;
 638
 639         buf_offset &= ~HAMMER_BUFMASK64;
 640         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 641         vol_no = HAMMER_VOL_DECODE(buf_offset);
 642         volume = hammer_get_volume(hmp, vol_no, &error);
 643         if (volume == NULL)
 644                 return;
 645         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 646                            buf_offset);
 647         if (buffer)
 648                 buffer->zoneX_offset = 0;
 649         hammer_rel_volume(volume, 0);
 650 }
 651
 652 static int
 653 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 654 {
 655         hammer_volume_t volume;
 656         int error;
 657
 658         /*
 659          * Load the buffer's on-disk info
 660          */
 661         volume = buffer->volume;
 662         ++buffer->io.loading;
 663         hammer_lock_ex(&buffer->io.lock);
 664
 665         if (hammer_debug_io & 0x0001) {
 666                 kprintf("load_buffer %016llx %016llx\n",
 667                         buffer->zoneX_offset, buffer->zone2_offset);
 668         }
 669
 670         if (buffer->ondisk == NULL) {
 671                 if (isnew) {
 672                         error = hammer_io_new(volume->devvp, &buffer->io);
 673                 } else {
 674                         error = hammer_io_read(volume->devvp, &buffer->io,
 675                                                volume->maxraw_off);
 676                 }
 677                 if (error == 0)
 678                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 679         } else if (isnew) {
 680                 error = hammer_io_new(volume->devvp, &buffer->io);
 681         } else {
 682                 error = 0;
 683         }
 684         --buffer->io.loading;
 685         hammer_unlock(&buffer->io.lock);
 686         return (error);
 687 }
 688
 689 /*
 690  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 691  */
 692 int
 693 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 694 {
 695         hammer_ref(&buffer->io.lock);
 696         hammer_flush_buffer_nodes(buffer);
 697         KKASSERT(buffer->io.lock.refs == 1);
 698         hammer_rel_buffer(buffer, 2);
 699         return(0);
 700 }
 701
 702 /*
 703  * Reference a buffer that is either already referenced or via a specially
 704  * handled pointer (aka cursor->buffer).
 705  */
 706 int
 707 hammer_ref_buffer(hammer_buffer_t buffer)
 708 {
 709         int error;
 710
 711         hammer_ref(&buffer->io.lock);
 712
 713         /*
 714          * No longer loose
 715          */
 716         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 717                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 718                 buffer->io.mod_list = NULL;
 719         }
 720
 721         if (buffer->ondisk == NULL || buffer->io.loading) {
 722                 error = hammer_load_buffer(buffer, 0);
 723                 if (error) {
 724                         hammer_rel_buffer(buffer, 1);
 725                         /*
 726                          * NOTE: buffer pointer can become stale after
 727                          * the above release.
 728                          */
 729                 }
 730         } else {
 731                 error = 0;
 732         }
 733         return(error);
 734 }
 735
 736 /*
 737  * Release a buffer.  We have to deal with several places where
 738  * another thread can ref the buffer.
 739  *
 740  * Only destroy the structure itself if the related buffer cache buffer
 741  * was disassociated from it.  This ties the management of the structure
 742  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 743  * embedded io is referenced or not.
 744  */
 745 void
 746 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 747 {
 748         hammer_volume_t volume;
 749         int freeme = 0;
 750
 751         crit_enter();
 752         if (buffer->io.lock.refs == 1) {
 753                 ++buffer->io.loading;   /* force interlock check */
 754                 hammer_lock_ex(&buffer->io.lock);
 755                 if (buffer->io.lock.refs == 1) {
 756                         hammer_io_release(&buffer->io, flush);
 757                         hammer_flush_buffer_nodes(buffer);
 758                         KKASSERT(TAILQ_EMPTY(&buffer->clist));
 759
 760                         if (buffer->io.bp == NULL &&
 761                             buffer->io.lock.refs == 1) {
 762                                 /*
 763                                  * Final cleanup
 764                                  */
 765                                 volume = buffer->volume;
 766                                 RB_REMOVE(hammer_buf_rb_tree,
 767                                           &volume->rb_bufs_root, buffer);
 768                                 buffer->volume = NULL; /* sanity */
 769                                 hammer_rel_volume(volume, 0);
 770                                 freeme = 1;
 771                         }
 772                 }
 773                 --buffer->io.loading;
 774                 hammer_unlock(&buffer->io.lock);
 775         }
 776         hammer_unref(&buffer->io.lock);
 777         crit_exit();
 778         if (freeme) {
 779                 KKASSERT(buffer->io.mod_list == NULL);
 780                 --hammer_count_buffers;
 781                 kfree(buffer, M_HAMMER);
 782         }
 783 }
 784
 785 /*
 786  * Remove the zoneX translation cache for a buffer given its zone-2 offset.
 787  */
 788 void
 789 hammer_uncache_buffer(hammer_mount_t hmp, hammer_off_t buf_offset)
 790 {
 791         hammer_volume_t volume;
 792         hammer_buffer_t buffer;
 793         int vol_no;
 794         int error;
 795
 796         buf_offset &= ~HAMMER_BUFMASK64;
 797         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_RAW_BUFFER);
 798         vol_no = HAMMER_VOL_DECODE(buf_offset);
 799         volume = hammer_get_volume(hmp, vol_no, &error);
 800         KKASSERT(volume != 0);
 801         KKASSERT(buf_offset < volume->maxbuf_off);
 802
 803         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 804                            buf_offset);
 805         if (buffer)
 806                 buffer->zoneX_offset = 0;
 807         hammer_rel_volume(volume, 0);
 808 }
 809
 810 /*
 811  * Access the filesystem buffer containing the specified hammer offset.
 812  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 813  * relative buffer offset.  It must also have bit 55 set to be valid.
 814  * (see hammer_off_t in hammer_disk.h).
 815  *
 816  * Any prior buffer in *bufferp will be released and replaced by the
 817  * requested buffer.
 818  */
 819 void *
 820 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 821              struct hammer_buffer **bufferp)
 822 {
 823         hammer_buffer_t buffer;
 824         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 825
 826         buf_offset &= ~HAMMER_BUFMASK64;
 827         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 828
 829         buffer = *bufferp;
 830         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 831                                buffer->zoneX_offset != buf_offset)) {
 832                 if (buffer)
 833                         hammer_rel_buffer(buffer, 0);
 834                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 835                 *bufferp = buffer;
 836         } else {
 837                 *errorp = 0;
 838         }
 839
 840         /*
 841          * Return a pointer to the buffer data.
 842          */
 843         if (buffer == NULL)
 844                 return(NULL);
 845         else
 846                 return((char *)buffer->ondisk + xoff);
 847 }
 848
 849 /*
 850  * Access the filesystem buffer containing the specified hammer offset.
 851  * No disk read operation occurs.  The result buffer may contain garbage.
 852  *
 853  * Any prior buffer in *bufferp will be released and replaced by the
 854  * requested buffer.
 855  *
 856  * This function marks the buffer dirty but does not increment its
 857  * modify_refs count.
 858  */
 859 void *
 860 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 861              struct hammer_buffer **bufferp)
 862 {
 863         hammer_buffer_t buffer;
 864         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 865
 866         buf_offset &= ~HAMMER_BUFMASK64;
 867
 868         buffer = *bufferp;
 869         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 870                                buffer->zoneX_offset != buf_offset)) {
 871                 if (buffer)
 872                         hammer_rel_buffer(buffer, 0);
 873                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 874                 *bufferp = buffer;
 875         } else {
 876                 *errorp = 0;
 877         }
 878
 879         /*
 880          * Return a pointer to the buffer data.
 881          */
 882         if (buffer == NULL)
 883                 return(NULL);
 884         else
 885                 return((char *)buffer->ondisk + xoff);
 886 }
 887
 888 /************************************************************************
 889  *                              NODES                                   *
 890  ************************************************************************
 891  *
 892  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 893  * method used by the HAMMER filesystem.
 894  *
 895  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 896  * associated with its buffer, and will only referenced the buffer while
 897  * the node itself is referenced.
 898  *
 899  * A hammer_node can also be passively associated with other HAMMER
 900  * structures, such as inodes, while retaining 0 references.  These
 901  * associations can be cleared backwards using a pointer-to-pointer in
 902  * the hammer_node.
 903  *
 904  * This allows the HAMMER implementation to cache hammer_nodes long-term
 905  * and short-cut a great deal of the infrastructure's complexity.  In
 906  * most cases a cached node can be reacquired without having to dip into
 907  * either the buffer or cluster management code.
 908  *
 909  * The caller must pass a referenced cluster on call and will retain
 910  * ownership of the reference on return.  The node will acquire its own
 911  * additional references, if necessary.
 912  */
 913 hammer_node_t
 914 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 915                 int isnew, int *errorp)
 916 {
 917         hammer_node_t node;
 918
 919         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 920
 921         /*
 922          * Locate the structure, allocating one if necessary.
 923          */
 924 again:
 925         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 926         if (node == NULL) {
 927                 ++hammer_count_nodes;
 928                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 929                 node->node_offset = node_offset;
 930                 node->hmp = hmp;
 931                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 932                         --hammer_count_nodes;
 933                         kfree(node, M_HAMMER);
 934                         goto again;
 935                 }
 936         }
 937         hammer_ref(&node->lock);
 938         if (node->ondisk)
 939                 *errorp = 0;
 940         else
 941                 *errorp = hammer_load_node(node, isnew);
 942         if (*errorp) {
 943                 hammer_rel_node(node);
 944                 node = NULL;
 945         }
 946         return(node);
 947 }
 948
 949 /*
 950  * Reference an already-referenced node.
 951  */
 952 void
 953 hammer_ref_node(hammer_node_t node)
 954 {
 955         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
 956         hammer_ref(&node->lock);
 957 }
 958
 959 /*
 960  * Load a node's on-disk data reference.
 961  */
 962 static int
 963 hammer_load_node(hammer_node_t node, int isnew)
 964 {
 965         hammer_buffer_t buffer;
 966         hammer_off_t buf_offset;
 967         int error;
 968
 969         error = 0;
 970         ++node->loading;
 971         hammer_lock_ex(&node->lock);
 972         if (node->ondisk == NULL) {
 973                 /*
 974                  * This is a little confusing but the jist is that
 975                  * node->buffer determines whether the node is on
 976                  * the buffer's clist and node->ondisk determines
 977                  * whether the buffer is referenced.
 978                  *
 979                  * We could be racing a buffer release, in which case
 980                  * node->buffer may become NULL while we are blocked
 981                  * referencing the buffer.
 982                  */
 983                 if ((buffer = node->buffer) != NULL) {
 984                         error = hammer_ref_buffer(buffer);
 985                         if (error == 0 && node->buffer == NULL) {
 986                                 TAILQ_INSERT_TAIL(&buffer->clist,
 987                                                   node, entry);
 988                                 node->buffer = buffer;
 989                         }
 990                 } else {
 991                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
 992                         buffer = hammer_get_buffer(node->hmp, buf_offset,
 993                                                    0, &error);
 994                         if (buffer) {
 995                                 KKASSERT(error == 0);
 996                                 TAILQ_INSERT_TAIL(&buffer->clist,
 997                                                   node, entry);
 998                                 node->buffer = buffer;
 999                         }
1000                 }
1001                 if (error == 0) {
1002                         node->ondisk = (void *)((char *)buffer->ondisk +
1003                                (node->node_offset & HAMMER_BUFMASK));
1004                         if (isnew == 0 &&
1005                             hammer_crc_test_btree(node->ondisk) == 0) {
1006                                 Debugger("CRC FAILED: B-TREE NODE");
1007                         }
1008                 }
1009         }
1010         --node->loading;
1011         hammer_unlock(&node->lock);
1012         return (error);
1013 }
1014
1015 /*
1016  * Safely reference a node, interlock against flushes via the IO subsystem.
1017  */
1018 hammer_node_t
1019 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
1020                      int *errorp)
1021 {
1022         hammer_node_t node;
1023
1024         node = *cache;
1025         if (node != NULL) {
1026                 hammer_ref(&node->lock);
1027                 if (node->ondisk)
1028                         *errorp = 0;
1029                 else
1030                         *errorp = hammer_load_node(node, 0);
1031                 if (*errorp) {
1032                         hammer_rel_node(node);
1033                         node = NULL;
1034                 }
1035         } else {
1036                 *errorp = ENOENT;
1037         }
1038         return(node);
1039 }
1040
1041 /*
1042  * Release a hammer_node.  On the last release the node dereferences
1043  * its underlying buffer and may or may not be destroyed.
1044  */
1045 void
1046 hammer_rel_node(hammer_node_t node)
1047 {
1048         hammer_buffer_t buffer;
1049
1050         /*
1051          * If this isn't the last ref just decrement the ref count and
1052          * return.
1053          */
1054         if (node->lock.refs > 1) {
1055                 hammer_unref(&node->lock);
1056                 return;
1057         }
1058
1059         /*
1060          * If there is no ondisk info or no buffer the node failed to load,
1061          * remove the last reference and destroy the node.
1062          */
1063         if (node->ondisk == NULL) {
1064                 hammer_unref(&node->lock);
1065                 hammer_flush_node(node);
1066                 /* node is stale now */
1067                 return;
1068         }
1069
1070         /*
1071          * Do final cleanups and then either destroy the node and leave it
1072          * passively cached.  The buffer reference is removed regardless.
1073          */
1074         buffer = node->buffer;
1075         node->ondisk = NULL;
1076
1077         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1078                 hammer_unref(&node->lock);
1079                 hammer_rel_buffer(buffer, 0);
1080                 return;
1081         }
1082
1083         /*
1084          * Destroy the node.
1085          */
1086         hammer_unref(&node->lock);
1087         hammer_flush_node(node);
1088         /* node is stale */
1089         hammer_rel_buffer(buffer, 0);
1090 }
1091
1092 /*
1093  * Free space on-media associated with a B-Tree node.
1094  */
1095 void
1096 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1097 {
1098         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1099         node->flags |= HAMMER_NODE_DELETED;
1100         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1101 }
1102
1103 /*
1104  * Passively cache a referenced hammer_node in *cache.  The caller may
1105  * release the node on return.
1106  */
1107 void
1108 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
1109 {
1110         hammer_node_t old;
1111
1112         /*
1113          * If the node is being deleted, don't cache it!
1114          */
1115         if (node->flags & HAMMER_NODE_DELETED)
1116                 return;
1117
1118         /*
1119          * Cache the node.  If we previously cached a different node we
1120          * have to give HAMMER a chance to destroy it.
1121          */
1122 again:
1123         if (node->cache1 != cache) {
1124                 if (node->cache2 != cache) {
1125                         if ((old = *cache) != NULL) {
1126                                 KKASSERT(node->lock.refs != 0);
1127                                 hammer_uncache_node(cache);
1128                                 goto again;
1129                         }
1130                         if (node->cache2)
1131                                 *node->cache2 = NULL;
1132                         node->cache2 = node->cache1;
1133                         node->cache1 = cache;
1134                         *cache = node;
1135                 } else {
1136                         struct hammer_node **tmp;
1137                         tmp = node->cache1;
1138                         node->cache1 = node->cache2;
1139                         node->cache2 = tmp;
1140                 }
1141         }
1142 }
1143
1144 void
1145 hammer_uncache_node(struct hammer_node **cache)
1146 {
1147         hammer_node_t node;
1148
1149         if ((node = *cache) != NULL) {
1150                 *cache = NULL;
1151                 if (node->cache1 == cache) {
1152                         node->cache1 = node->cache2;
1153                         node->cache2 = NULL;
1154                 } else if (node->cache2 == cache) {
1155                         node->cache2 = NULL;
1156                 } else {
1157                         panic("hammer_uncache_node: missing cache linkage");
1158                 }
1159                 if (node->cache1 == NULL && node->cache2 == NULL)
1160                         hammer_flush_node(node);
1161         }
1162 }
1163
1164 /*
1165  * Remove a node's cache references and destroy the node if it has no
1166  * other references or backing store.
1167  */
1168 void
1169 hammer_flush_node(hammer_node_t node)
1170 {
1171         hammer_buffer_t buffer;
1172
1173         if (node->cache1)
1174                 *node->cache1 = NULL;
1175         if (node->cache2)
1176                 *node->cache2 = NULL;
1177         if (node->lock.refs == 0 && node->ondisk == NULL) {
1178                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1179                 if ((buffer = node->buffer) != NULL) {
1180                         node->buffer = NULL;
1181                         TAILQ_REMOVE(&buffer->clist, node, entry);
1182                         /* buffer is unreferenced because ondisk is NULL */
1183                 }
1184                 --hammer_count_nodes;
1185                 kfree(node, M_HAMMER);
1186         }
1187 }
1188
1189 /*
1190  * Flush passively cached B-Tree nodes associated with this buffer.
1191  * This is only called when the buffer is about to be destroyed, so
1192  * none of the nodes should have any references.  The buffer is locked.
1193  *
1194  * We may be interlocked with the buffer.
1195  */
1196 void
1197 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1198 {
1199         hammer_node_t node;
1200
1201         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1202                 KKASSERT(node->ondisk == NULL);
1203
1204                 if (node->lock.refs == 0) {
1205                         hammer_ref(&node->lock);
1206                         node->flags |= HAMMER_NODE_FLUSH;
1207                         hammer_rel_node(node);
1208                 } else {
1209                         KKASSERT(node->loading != 0);
1210                         KKASSERT(node->buffer != NULL);
1211                         buffer = node->buffer;
1212                         node->buffer = NULL;
1213                         TAILQ_REMOVE(&buffer->clist, node, entry);
1214                         /* buffer is unreferenced because ondisk is NULL */
1215                 }
1216         }
1217 }
1218
1219
1220 /************************************************************************
1221  *                              ALLOCATORS                              *
1222  ************************************************************************/
1223
1224 /*
1225  * Allocate a B-Tree node.
1226  */
1227 hammer_node_t
1228 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1229 {
1230         hammer_buffer_t buffer = NULL;
1231         hammer_node_t node = NULL;
1232         hammer_off_t node_offset;
1233
1234         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1235                                             sizeof(struct hammer_node_ondisk),
1236                                             errorp);
1237         if (*errorp == 0) {
1238                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1239                 hammer_modify_node_noundo(trans, node);
1240                 bzero(node->ondisk, sizeof(*node->ondisk));
1241                 hammer_modify_node_done(node);
1242         }
1243         if (buffer)
1244                 hammer_rel_buffer(buffer, 0);
1245         return(node);
1246 }
1247
1248 #if 0
1249
1250 /*
1251  * The returned buffers are already appropriately marked as being modified.
1252  * If the caller marks them again unnecessary undo records may be generated.
1253  *
1254  * In-band data is indicated by data_bufferp == NULL.  Pass a data_len of 0
1255  * for zero-fill (caller modifies data_len afterwords).
1256  *
1257  * If the caller is responsible for calling hammer_modify_*() prior to making
1258  * any additional modifications to either the returned record buffer or the
1259  * returned data buffer.
1260  */
1261 void *
1262 hammer_alloc_record(hammer_transaction_t trans,
1263                     hammer_off_t *rec_offp, u_int16_t rec_type,
1264                     struct hammer_buffer **rec_bufferp,
1265                     int32_t data_len, void **datap,
1266                     hammer_off_t *data_offp,
1267                     struct hammer_buffer **data_bufferp, int *errorp)
1268 {
1269         hammer_record_ondisk_t rec;
1270         hammer_off_t rec_offset;
1271         hammer_off_t data_offset;
1272         int32_t reclen;
1273
1274         if (datap)
1275                 *datap = NULL;
1276
1277         /*
1278          * Allocate the record
1279          */
1280         rec_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_RECORD_INDEX,
1281                                            HAMMER_RECORD_SIZE, errorp);
1282         if (*errorp)
1283                 return(NULL);
1284         if (data_offp)
1285                 *data_offp = 0;
1286
1287         /*
1288          * Allocate data
1289          */
1290         if (data_len) {
1291                 if (data_bufferp == NULL) {
1292                         switch(rec_type) {
1293                         case HAMMER_RECTYPE_DATA:
1294                                 reclen = offsetof(struct hammer_data_record,
1295                                                   data[0]);
1296                                 break;
1297                         case HAMMER_RECTYPE_DIRENTRY:
1298                                 reclen = offsetof(struct hammer_entry_record,
1299                                                   name[0]);
1300                                 break;
1301                         default:
1302                                 panic("hammer_alloc_record: illegal "
1303                                       "in-band data");
1304                                 /* NOT REACHED */
1305                                 reclen = 0;
1306                                 break;
1307                         }
1308                         KKASSERT(reclen + data_len <= HAMMER_RECORD_SIZE);
1309                         data_offset = rec_offset + reclen;
1310                 } else if (data_len < HAMMER_BUFSIZE) {
1311                         data_offset = hammer_blockmap_alloc(trans,
1312                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1313                                                 data_len, errorp);
1314                         *data_offp = data_offset;
1315                 } else {
1316                         data_offset = hammer_blockmap_alloc(trans,
1317                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1318                                                 data_len, errorp);
1319                         *data_offp = data_offset;
1320                 }
1321         } else {
1322                 data_offset = 0;
1323         }
1324         if (*errorp) {
1325                 hammer_blockmap_free(trans, rec_offset, HAMMER_RECORD_SIZE);
1326                 return(NULL);
1327         }
1328
1329         /*
1330          * Basic return values.
1331          *
1332          * Note that because this is a 'new' buffer, there is no need to
1333          * generate UNDO records for it.
1334          */
1335         *rec_offp = rec_offset;
1336         rec = hammer_bread(trans->hmp, rec_offset, errorp, rec_bufferp);
1337         hammer_modify_buffer(trans, *rec_bufferp, NULL, 0);
1338         bzero(rec, sizeof(*rec));
1339         KKASSERT(*errorp == 0);
1340         rec->base.data_off = data_offset;
1341         rec->base.data_len = data_len;
1342         hammer_modify_buffer_done(*rec_bufferp);
1343
1344         if (data_bufferp) {
1345                 if (data_len) {
1346                         *datap = hammer_bread(trans->hmp, data_offset, errorp,
1347                                               data_bufferp);
1348                         KKASSERT(*errorp == 0);
1349                 } else {
1350                         *datap = NULL;
1351                 }
1352         } else if (data_len) {
1353                 KKASSERT(data_offset + data_len - rec_offset <=
1354                          HAMMER_RECORD_SIZE);
1355                 if (datap) {
1356                         *datap = (void *)((char *)rec +
1357                                           (int32_t)(data_offset - rec_offset));
1358                 }
1359         } else {
1360                 KKASSERT(datap == NULL);
1361         }
1362         KKASSERT(*errorp == 0);
1363         return(rec);
1364 }
1365
1366 #endif
1367
1368 /*
1369  * Allocate data.  If the address of a data buffer is supplied then
1370  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1371  * will be set to the related buffer.  The caller must release it when
1372  * finally done.  The initial *data_bufferp should be set to NULL by
1373  * the caller.
1374  *
1375  * The caller is responsible for making hammer_modify*() calls on the
1376  * *data_bufferp.
1377  */
1378 void *
1379 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1380                   hammer_off_t *data_offsetp,
1381                   struct hammer_buffer **data_bufferp, int *errorp)
1382 {
1383         void *data;
1384
1385         /*
1386          * Allocate data
1387          */
1388         if (data_len) {
1389                 if (data_len < HAMMER_BUFSIZE) {
1390                         *data_offsetp = hammer_blockmap_alloc(trans,
1391                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1392                                                 data_len, errorp);
1393                 } else {
1394                         *data_offsetp = hammer_blockmap_alloc(trans,
1395                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1396                                                 data_len, errorp);
1397                 }
1398         } else {
1399                 *data_offsetp = 0;
1400         }
1401         if (*errorp == 0 && data_bufferp) {
1402                 if (data_len) {
1403                         data = hammer_bread(trans->hmp, *data_offsetp, errorp,
1404                                             data_bufferp);
1405                         KKASSERT(*errorp == 0);
1406                 } else {
1407                         data = NULL;
1408                 }
1409         } else {
1410                 data = NULL;
1411         }
1412         KKASSERT(*errorp == 0);
1413         return(data);
1414 }
1415
1416 /*
1417  * Sync dirty buffers to the media and clean-up any loose ends.
1418  */
1419 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1420 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1421
1422 int
1423 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1424 {
1425         struct hammer_sync_info info;
1426
1427         info.error = 0;
1428         info.waitfor = waitfor;
1429         if (waitfor == MNT_WAIT) {
1430                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1431                               hammer_sync_scan1, hammer_sync_scan2, &info);
1432         } else {
1433                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1434                               hammer_sync_scan1, hammer_sync_scan2, &info);
1435         }
1436         return(info.error);
1437 }
1438
1439 int
1440 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1441 {
1442         struct hammer_sync_info info;
1443
1444         info.error = 0;
1445         info.waitfor = waitfor;
1446
1447         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1448                       hammer_sync_scan1, hammer_sync_scan2, &info);
1449         if (waitfor == MNT_WAIT)
1450                 hammer_flusher_sync(hmp);
1451         else
1452                 hammer_flusher_async(hmp);
1453
1454         return(info.error);
1455 }
1456
1457 static int
1458 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1459 {
1460         struct hammer_inode *ip;
1461
1462         ip = VTOI(vp);
1463         if (vp->v_type == VNON || ip == NULL ||
1464             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1465              RB_EMPTY(&vp->v_rbdirty_tree))) {
1466                 return(-1);
1467         }
1468         return(0);
1469 }
1470
1471 static int
1472 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1473 {
1474         struct hammer_sync_info *info = data;
1475         struct hammer_inode *ip;
1476         int error;
1477
1478         ip = VTOI(vp);
1479         if (vp->v_type == VNON || vp->v_type == VBAD ||
1480             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1481              RB_EMPTY(&vp->v_rbdirty_tree))) {
1482                 return(0);
1483         }
1484         error = VOP_FSYNC(vp, info->waitfor);
1485         if (error)
1486                 info->error = error;
1487         return(0);
1488 }
1489