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 1.76 2008/08/29 20:19:08 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_transaction_t trans,
  52                                 hammer_node_t node, int isnew);
  53
  54 static int
  55 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  56 {
  57         if (vol1->vol_no < vol2->vol_no)
  58                 return(-1);
  59         if (vol1->vol_no > vol2->vol_no)
  60                 return(1);
  61         return(0);
  62 }
  63
  64 /*
  65  * hammer_buffer structures are indexed via their zoneX_offset, not
  66  * their zone2_offset.
  67  */
  68 static int
  69 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  70 {
  71         if (buf1->zoneX_offset < buf2->zoneX_offset)
  72                 return(-1);
  73         if (buf1->zoneX_offset > buf2->zoneX_offset)
  74                 return(1);
  75         return(0);
  76 }
  77
  78 static int
  79 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  80 {
  81         if (node1->node_offset < node2->node_offset)
  82                 return(-1);
  83         if (node1->node_offset > node2->node_offset)
  84                 return(1);
  85         return(0);
  86 }
  87
  88 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  89              hammer_vol_rb_compare, int32_t, vol_no);
  90 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  91              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  92 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  93              hammer_nod_rb_compare, hammer_off_t, node_offset);
  94
  95 /************************************************************************
  96  *                              VOLUMES                                 *
  97  ************************************************************************
  98  *
  99  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 100  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 101  * not load a new volume.
 102  *
 103  * Calls made to hammer_load_volume() or single-threaded
 104  */
 105 int
 106 hammer_install_volume(struct hammer_mount *hmp, const char *volname,
 107                       struct vnode *devvp)
 108 {
 109         struct mount *mp;
 110         hammer_volume_t volume;
 111         struct hammer_volume_ondisk *ondisk;
 112         struct nlookupdata nd;
 113         struct buf *bp = NULL;
 114         int error;
 115         int ronly;
 116         int setmp = 0;
 117
 118         mp = hmp->mp;
 119         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 120
 121         /*
 122          * Allocate a volume structure
 123          */
 124         ++hammer_count_volumes;
 125         volume = kmalloc(sizeof(*volume), hmp->m_misc, M_WAITOK|M_ZERO);
 126         volume->vol_name = kstrdup(volname, hmp->m_misc);
 127         volume->io.hmp = hmp;   /* bootstrap */
 128         hammer_io_init(&volume->io, volume, HAMMER_STRUCTURE_VOLUME);
 129         volume->io.offset = 0LL;
 130         volume->io.bytes = HAMMER_BUFSIZE;
 131
 132         /*
 133          * Get the device vnode
 134          */
 135         if (devvp == NULL) {
 136                 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 137                 if (error == 0)
 138                         error = nlookup(&nd);
 139                 if (error == 0)
 140                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 141                 nlookup_done(&nd);
 142         } else {
 143                 error = 0;
 144                 volume->devvp = devvp;
 145         }
 146
 147         if (error == 0) {
 148                 if (vn_isdisk(volume->devvp, &error)) {
 149                         error = vfs_mountedon(volume->devvp);
 150                 }
 151         }
 152         if (error == 0 &&
 153             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 154                 error = EBUSY;
 155         }
 156         if (error == 0) {
 157                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 158                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 159                 if (error == 0) {
 160                         error = VOP_OPEN(volume->devvp,
 161                                          (ronly ? FREAD : FREAD|FWRITE),
 162                                          FSCRED, NULL);
 163                 }
 164                 vn_unlock(volume->devvp);
 165         }
 166         if (error) {
 167                 hammer_free_volume(volume);
 168                 return(error);
 169         }
 170         volume->devvp->v_rdev->si_mountpoint = mp;
 171         setmp = 1;
 172
 173         /*
 174          * Extract the volume number from the volume header and do various
 175          * sanity checks.
 176          */
 177         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 178         if (error)
 179                 goto late_failure;
 180         ondisk = (void *)bp->b_data;
 181         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 182                 kprintf("hammer_mount: volume %s has an invalid header\n",
 183                         volume->vol_name);
 184                 error = EFTYPE;
 185                 goto late_failure;
 186         }
 187         volume->vol_no = ondisk->vol_no;
 188         volume->buffer_base = ondisk->vol_buf_beg;
 189         volume->vol_flags = ondisk->vol_flags;
 190         volume->nblocks = ondisk->vol_nblocks;
 191         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 192                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 193         volume->maxraw_off = ondisk->vol_buf_end;
 194
 195         if (RB_EMPTY(&hmp->rb_vols_root)) {
 196                 hmp->fsid = ondisk->vol_fsid;
 197         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 198                 kprintf("hammer_mount: volume %s's fsid does not match "
 199                         "other volumes\n", volume->vol_name);
 200                 error = EFTYPE;
 201                 goto late_failure;
 202         }
 203
 204         /*
 205          * Insert the volume structure into the red-black tree.
 206          */
 207         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 208                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 209                         volume->vol_name, volume->vol_no);
 210                 error = EEXIST;
 211         }
 212
 213         /*
 214          * Set the root volume .  HAMMER special cases rootvol the structure.
 215          * We do not hold a ref because this would prevent related I/O
 216          * from being flushed.
 217          */
 218         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 219                 hmp->rootvol = volume;
 220                 hmp->nvolumes = ondisk->vol_count;
 221                 if (bp) {
 222                         brelse(bp);
 223                         bp = NULL;
 224                 }
 225                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 226                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 227                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 228                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 229         }
 230 late_failure:
 231         if (bp)
 232                 brelse(bp);
 233         if (error) {
 234                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 235                 if (setmp)
 236                         volume->devvp->v_rdev->si_mountpoint = NULL;
 237                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 238                 hammer_free_volume(volume);
 239         }
 240         return (error);
 241 }
 242
 243 /*
 244  * This is called for each volume when updating the mount point from
 245  * read-write to read-only or vise-versa.
 246  */
 247 int
 248 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 249 {
 250         if (volume->devvp) {
 251                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 252                 if (volume->io.hmp->ronly) {
 253                         /* do not call vinvalbuf */
 254                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 255                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 256                 } else {
 257                         /* do not call vinvalbuf */
 258                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 259                         VOP_CLOSE(volume->devvp, FREAD);
 260                 }
 261                 vn_unlock(volume->devvp);
 262         }
 263         return(0);
 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         hammer_mount_t hmp = volume->io.hmp;
 274         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 275         struct buf *bp;
 276
 277         /*
 278          * Clean up the root volume pointer, which is held unlocked in hmp.
 279          */
 280         if (hmp->rootvol == volume)
 281                 hmp->rootvol = NULL;
 282
 283         /*
 284          * We must not flush a dirty buffer to disk on umount.  It should
 285          * have already been dealt with by the flusher, or we may be in
 286          * catastrophic failure.
 287          */
 288         hammer_io_clear_modify(&volume->io, 1);
 289         volume->io.waitdep = 1;
 290         bp = hammer_io_release(&volume->io, 1);
 291
 292         /*
 293          * Clean up the persistent ref ioerror might have on the volume
 294          */
 295         if (volume->io.ioerror) {
 296                 volume->io.ioerror = 0;
 297                 hammer_unref(&volume->io.lock);
 298         }
 299
 300         /*
 301          * There should be no references on the volume, no clusters, and
 302          * no super-clusters.
 303          */
 304         KKASSERT(volume->io.lock.refs == 0);
 305         if (bp)
 306                 brelse(bp);
 307
 308         volume->ondisk = NULL;
 309         if (volume->devvp) {
 310                 if (volume->devvp->v_rdev &&
 311                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 312                 ) {
 313                         volume->devvp->v_rdev->si_mountpoint = NULL;
 314                 }
 315                 if (ronly) {
 316                         /*
 317                          * Make sure we don't sync anything to disk if we
 318                          * are in read-only mode (1) or critically-errored
 319                          * (2).  Note that there may be dirty buffers in
 320                          * normal read-only mode from crash recovery.
 321                          */
 322                         vinvalbuf(volume->devvp, 0, 0, 0);
 323                         VOP_CLOSE(volume->devvp, FREAD);
 324                 } else {
 325                         /*
 326                          * Normal termination, save any dirty buffers
 327                          * (XXX there really shouldn't be any).
 328                          */
 329                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 330                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 331                 }
 332         }
 333
 334         /*
 335          * Destroy the structure
 336          */
 337         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 338         hammer_free_volume(volume);
 339         return(0);
 340 }
 341
 342 static
 343 void
 344 hammer_free_volume(hammer_volume_t volume)
 345 {
 346         hammer_mount_t hmp = volume->io.hmp;
 347
 348         if (volume->vol_name) {
 349                 kfree(volume->vol_name, hmp->m_misc);
 350                 volume->vol_name = NULL;
 351         }
 352         if (volume->devvp) {
 353                 vrele(volume->devvp);
 354                 volume->devvp = NULL;
 355         }
 356         --hammer_count_volumes;
 357         kfree(volume, hmp->m_misc);
 358 }
 359
 360 /*
 361  * Get a HAMMER volume.  The volume must already exist.
 362  */
 363 hammer_volume_t
 364 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 365 {
 366         struct hammer_volume *volume;
 367
 368         /*
 369          * Locate the volume structure
 370          */
 371         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 372         if (volume == NULL) {
 373                 *errorp = ENOENT;
 374                 return(NULL);
 375         }
 376         hammer_ref(&volume->io.lock);
 377
 378         /*
 379          * Deal with on-disk info
 380          */
 381         if (volume->ondisk == NULL || volume->io.loading) {
 382                 *errorp = hammer_load_volume(volume);
 383                 if (*errorp) {
 384                         hammer_rel_volume(volume, 1);
 385                         volume = NULL;
 386                 }
 387         } else {
 388                 *errorp = 0;
 389         }
 390         return(volume);
 391 }
 392
 393 int
 394 hammer_ref_volume(hammer_volume_t volume)
 395 {
 396         int error;
 397
 398         hammer_ref(&volume->io.lock);
 399
 400         /*
 401          * Deal with on-disk info
 402          */
 403         if (volume->ondisk == NULL || volume->io.loading) {
 404                 error = hammer_load_volume(volume);
 405                 if (error)
 406                         hammer_rel_volume(volume, 1);
 407         } else {
 408                 error = 0;
 409         }
 410         return (error);
 411 }
 412
 413 hammer_volume_t
 414 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 415 {
 416         hammer_volume_t volume;
 417
 418         volume = hmp->rootvol;
 419         KKASSERT(volume != NULL);
 420         hammer_ref(&volume->io.lock);
 421
 422         /*
 423          * Deal with on-disk info
 424          */
 425         if (volume->ondisk == NULL || volume->io.loading) {
 426                 *errorp = hammer_load_volume(volume);
 427                 if (*errorp) {
 428                         hammer_rel_volume(volume, 1);
 429                         volume = NULL;
 430                 }
 431         } else {
 432                 *errorp = 0;
 433         }
 434         return (volume);
 435 }
 436
 437 /*
 438  * Load a volume's on-disk information.  The volume must be referenced and
 439  * not locked.  We temporarily acquire an exclusive lock to interlock
 440  * against releases or multiple get's.
 441  */
 442 static int
 443 hammer_load_volume(hammer_volume_t volume)
 444 {
 445         int error;
 446
 447         ++volume->io.loading;
 448         hammer_lock_ex(&volume->io.lock);
 449
 450         if (volume->ondisk == NULL) {
 451                 error = hammer_io_read(volume->devvp, &volume->io,
 452                                        volume->maxraw_off);
 453                 if (error == 0)
 454                         volume->ondisk = (void *)volume->io.bp->b_data;
 455         } else {
 456                 error = 0;
 457         }
 458         --volume->io.loading;
 459         hammer_unlock(&volume->io.lock);
 460         return(error);
 461 }
 462
 463 /*
 464  * Release a volume.  Call hammer_io_release on the last reference.  We have
 465  * to acquire an exclusive lock to interlock against volume->ondisk tests
 466  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 467  * lock to be held.
 468  *
 469  * Volumes are not unloaded from memory during normal operation.
 470  */
 471 void
 472 hammer_rel_volume(hammer_volume_t volume, int flush)
 473 {
 474         struct buf *bp = NULL;
 475
 476         crit_enter();
 477         if (volume->io.lock.refs == 1) {
 478                 ++volume->io.loading;
 479                 hammer_lock_ex(&volume->io.lock);
 480                 if (volume->io.lock.refs == 1) {
 481                         volume->ondisk = NULL;
 482                         bp = hammer_io_release(&volume->io, flush);
 483                 }
 484                 --volume->io.loading;
 485                 hammer_unlock(&volume->io.lock);
 486         }
 487         hammer_unref(&volume->io.lock);
 488         if (bp)
 489                 brelse(bp);
 490         crit_exit();
 491 }
 492
 493 int
 494 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 495 {
 496         hammer_volume_t vol;
 497         int i;
 498
 499         for (i = 0; i < hmp->nvolumes; ++i) {
 500                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 501                 if (vol == NULL)
 502                         return(EINVAL);
 503         }
 504         return(0);
 505 }
 506
 507 /************************************************************************
 508  *                              BUFFERS                                 *
 509  ************************************************************************
 510  *
 511  * Manage buffers.  Currently all blockmap-backed zones are direct-mapped
 512  * to zone-2 buffer offsets, without a translation stage.  However, the
 513  * hammer_buffer structure is indexed by its zoneX_offset, not its
 514  * zone2_offset.
 515  *
 516  * The proper zone must be maintained throughout the code-base all the way
 517  * through to the big-block allocator, or routines like hammer_del_buffers()
 518  * will not be able to locate all potentially conflicting buffers.
 519  */
 520 hammer_buffer_t
 521 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 522                   int bytes, int isnew, int *errorp)
 523 {
 524         hammer_buffer_t buffer;
 525         hammer_volume_t volume;
 526         hammer_off_t    zone2_offset;
 527         hammer_io_type_t iotype;
 528         int vol_no;
 529         int zone;
 530
 531         buf_offset &= ~HAMMER_BUFMASK64;
 532 again:
 533         /*
 534          * Shortcut if the buffer is already cached
 535          */
 536         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 537         if (buffer) {
 538                 if (buffer->io.lock.refs == 0)
 539                         ++hammer_count_refedbufs;
 540                 hammer_ref(&buffer->io.lock);
 541
 542                 /*
 543                  * Once refed the ondisk field will not be cleared by
 544                  * any other action.
 545                  */
 546                 if (buffer->ondisk && buffer->io.loading == 0) {
 547                         *errorp = 0;
 548                         return(buffer);
 549                 }
 550
 551                 /*
 552                  * The buffer is no longer loose if it has a ref, and
 553                  * cannot become loose once it gains a ref.  Loose
 554                  * buffers will never be in a modified state.  This should
 555                  * only occur on the 0->1 transition of refs.
 556                  *
 557                  * lose_list can be modified via a biodone() interrupt.
 558                  */
 559                 if (buffer->io.mod_list == &hmp->lose_list) {
 560                         crit_enter();   /* biodone race against list */
 561                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 562                                      mod_entry);
 563                         crit_exit();
 564                         buffer->io.mod_list = NULL;
 565                         KKASSERT(buffer->io.modified == 0);
 566                 }
 567                 goto found;
 568         }
 569
 570         /*
 571          * What is the buffer class?
 572          */
 573         zone = HAMMER_ZONE_DECODE(buf_offset);
 574
 575         switch(zone) {
 576         case HAMMER_ZONE_LARGE_DATA_INDEX:
 577         case HAMMER_ZONE_SMALL_DATA_INDEX:
 578                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 579                 break;
 580         case HAMMER_ZONE_UNDO_INDEX:
 581                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 582                 break;
 583         case HAMMER_ZONE_META_INDEX:
 584         default:
 585                 /*
 586                  * NOTE: inode data and directory entries are placed in this
 587                  * zone.  inode atime/mtime is updated in-place and thus
 588                  * buffers containing inodes must be synchronized as
 589                  * meta-buffers, same as buffers containing B-Tree info.
 590                  */
 591                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 592                 break;
 593         }
 594
 595         /*
 596          * Handle blockmap offset translations
 597          */
 598         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 599                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 600         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 601                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 602         } else {
 603                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 604                 zone2_offset = buf_offset;
 605                 *errorp = 0;
 606         }
 607         if (*errorp)
 608                 return(NULL);
 609
 610         /*
 611          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 612          * specifications.
 613          */
 614         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 615                  HAMMER_ZONE_RAW_BUFFER);
 616         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 617         volume = hammer_get_volume(hmp, vol_no, errorp);
 618         if (volume == NULL)
 619                 return(NULL);
 620
 621         KKASSERT(zone2_offset < volume->maxbuf_off);
 622
 623         /*
 624          * Allocate a new buffer structure.  We will check for races later.
 625          */
 626         ++hammer_count_buffers;
 627         buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
 628                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 629         buffer->zone2_offset = zone2_offset;
 630         buffer->zoneX_offset = buf_offset;
 631
 632         hammer_io_init(&buffer->io, volume, iotype);
 633         buffer->io.offset = volume->ondisk->vol_buf_beg +
 634                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 635         buffer->io.bytes = bytes;
 636         TAILQ_INIT(&buffer->clist);
 637         hammer_ref(&buffer->io.lock);
 638
 639         /*
 640          * Insert the buffer into the RB tree and handle late collisions.
 641          */
 642         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 643                 hammer_unref(&buffer->io.lock); /* safety */
 644                 --hammer_count_buffers;
 645                 hammer_rel_volume(volume, 0);
 646                 buffer->io.volume = NULL;       /* safety */
 647                 kfree(buffer, hmp->m_misc);
 648                 goto again;
 649         }
 650         ++hammer_count_refedbufs;
 651 found:
 652
 653         /*
 654          * Deal with on-disk info and loading races.
 655          */
 656         if (buffer->ondisk == NULL || buffer->io.loading) {
 657                 *errorp = hammer_load_buffer(buffer, isnew);
 658                 if (*errorp) {
 659                         hammer_rel_buffer(buffer, 1);
 660                         buffer = NULL;
 661                 }
 662         } else {
 663                 *errorp = 0;
 664         }
 665         return(buffer);
 666 }
 667
 668 /*
 669  * This is used by the direct-read code to deal with large-data buffers
 670  * created by the reblocker and mirror-write code.  The direct-read code
 671  * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
 672  * running hammer buffers must be fully synced to disk before we can issue
 673  * the direct-read.
 674  *
 675  * This code path is not considered critical as only the rebocker and
 676  * mirror-write code will create large-data buffers via the HAMMER buffer
 677  * subsystem.  They do that because they operate at the B-Tree level and
 678  * do not access the vnode/inode structures.
 679  */
 680 void
 681 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 682 {
 683         hammer_buffer_t buffer;
 684         int error;
 685
 686         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 687                  HAMMER_ZONE_LARGE_DATA);
 688
 689         while (bytes > 0) {
 690                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 691                                    base_offset);
 692                 if (buffer && (buffer->io.modified || buffer->io.running)) {
 693                         error = hammer_ref_buffer(buffer);
 694                         if (error == 0) {
 695                                 hammer_io_wait(&buffer->io);
 696                                 if (buffer->io.modified) {
 697                                         hammer_io_write_interlock(&buffer->io);
 698                                         hammer_io_flush(&buffer->io);
 699                                         hammer_io_done_interlock(&buffer->io);
 700                                         hammer_io_wait(&buffer->io);
 701                                 }
 702                                 hammer_rel_buffer(buffer, 0);
 703                         }
 704                 }
 705                 base_offset += HAMMER_BUFSIZE;
 706                 bytes -= HAMMER_BUFSIZE;
 707         }
 708 }
 709
 710 /*
 711  * Destroy all buffers covering the specified zoneX offset range.  This
 712  * is called when the related blockmap layer2 entry is freed or when
 713  * a direct write bypasses our buffer/buffer-cache subsystem.
 714  *
 715  * The buffers may be referenced by the caller itself.  Setting reclaim
 716  * will cause the buffer to be destroyed when it's ref count reaches zero.
 717  *
 718  * Return 0 on success, EAGAIN if some buffers could not be destroyed due
 719  * to additional references held by other threads, or some other (typically
 720  * fatal) error.
 721  */
 722 int
 723 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 724                    hammer_off_t zone2_offset, int bytes,
 725                    int report_conflicts)
 726 {
 727         hammer_buffer_t buffer;
 728         hammer_volume_t volume;
 729         int vol_no;
 730         int error;
 731         int ret_error;
 732
 733         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 734         volume = hammer_get_volume(hmp, vol_no, &ret_error);
 735         KKASSERT(ret_error == 0);
 736
 737         while (bytes > 0) {
 738                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 739                                    base_offset);
 740                 if (buffer) {
 741                         error = hammer_ref_buffer(buffer);
 742                         if (error == 0 && buffer->io.lock.refs != 1) {
 743                                 error = EAGAIN;
 744                                 hammer_rel_buffer(buffer, 0);
 745                         }
 746                         if (error == 0) {
 747                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 748                                 hammer_io_clear_modify(&buffer->io, 1);
 749                                 buffer->io.reclaim = 1;
 750                                 buffer->io.waitdep = 1;
 751                                 KKASSERT(buffer->io.volume == volume);
 752                                 hammer_rel_buffer(buffer, 0);
 753                         }
 754                 } else {
 755                         error = hammer_io_inval(volume, zone2_offset);
 756                 }
 757                 if (error) {
 758                         ret_error = error;
 759                         if (report_conflicts || (hammer_debug_general & 0x8000))
 760                                 kprintf("hammer_del_buffers: unable to invalidate %016llx buffer=%p rep=%d\n", base_offset, buffer, report_conflicts);
 761                 }
 762                 base_offset += HAMMER_BUFSIZE;
 763                 zone2_offset += HAMMER_BUFSIZE;
 764                 bytes -= HAMMER_BUFSIZE;
 765         }
 766         hammer_rel_volume(volume, 0);
 767         return (ret_error);
 768 }
 769
 770 static int
 771 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 772 {
 773         hammer_volume_t volume;
 774         int error;
 775
 776         /*
 777          * Load the buffer's on-disk info
 778          */
 779         volume = buffer->io.volume;
 780         ++buffer->io.loading;
 781         hammer_lock_ex(&buffer->io.lock);
 782
 783         if (hammer_debug_io & 0x0001) {
 784                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 785                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 786                         buffer->ondisk);
 787         }
 788
 789         if (buffer->ondisk == NULL) {
 790                 if (isnew) {
 791                         error = hammer_io_new(volume->devvp, &buffer->io);
 792                 } else {
 793                         error = hammer_io_read(volume->devvp, &buffer->io,
 794                                                volume->maxraw_off);
 795                 }
 796                 if (error == 0)
 797                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 798         } else if (isnew) {
 799                 error = hammer_io_new(volume->devvp, &buffer->io);
 800         } else {
 801                 error = 0;
 802         }
 803         --buffer->io.loading;
 804         hammer_unlock(&buffer->io.lock);
 805         return (error);
 806 }
 807
 808 /*
 809  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 810  * This routine is only called during unmount.
 811  */
 812 int
 813 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 814 {
 815         /*
 816          * Clean up the persistent ref ioerror might have on the buffer
 817          * and acquire a ref (steal ioerror's if we can).
 818          */
 819         if (buffer->io.ioerror) {
 820                 buffer->io.ioerror = 0;
 821         } else {
 822                 if (buffer->io.lock.refs == 0)
 823                         ++hammer_count_refedbufs;
 824                 hammer_ref(&buffer->io.lock);
 825         }
 826
 827         /*
 828          * We must not flush a dirty buffer to disk on umount.  It should
 829          * have already been dealt with by the flusher, or we may be in
 830          * catastrophic failure.
 831          */
 832         hammer_io_clear_modify(&buffer->io, 1);
 833         hammer_flush_buffer_nodes(buffer);
 834         KKASSERT(buffer->io.lock.refs == 1);
 835         hammer_rel_buffer(buffer, 2);
 836         return(0);
 837 }
 838
 839 /*
 840  * Reference a buffer that is either already referenced or via a specially
 841  * handled pointer (aka cursor->buffer).
 842  */
 843 int
 844 hammer_ref_buffer(hammer_buffer_t buffer)
 845 {
 846         int error;
 847
 848         if (buffer->io.lock.refs == 0)
 849                 ++hammer_count_refedbufs;
 850         hammer_ref(&buffer->io.lock);
 851
 852         /*
 853          * At this point a biodone() will not touch the buffer other then
 854          * incidental bits.  However, lose_list can be modified via
 855          * a biodone() interrupt.
 856          *
 857          * No longer loose
 858          */
 859         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 860                 crit_enter();
 861                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 862                 buffer->io.mod_list = NULL;
 863                 crit_exit();
 864         }
 865
 866         if (buffer->ondisk == NULL || buffer->io.loading) {
 867                 error = hammer_load_buffer(buffer, 0);
 868                 if (error) {
 869                         hammer_rel_buffer(buffer, 1);
 870                         /*
 871                          * NOTE: buffer pointer can become stale after
 872                          * the above release.
 873                          */
 874                 }
 875         } else {
 876                 error = 0;
 877         }
 878         return(error);
 879 }
 880
 881 /*
 882  * Release a buffer.  We have to deal with several places where
 883  * another thread can ref the buffer.
 884  *
 885  * Only destroy the structure itself if the related buffer cache buffer
 886  * was disassociated from it.  This ties the management of the structure
 887  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 888  * embedded io is referenced or not.
 889  */
 890 void
 891 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 892 {
 893         hammer_volume_t volume;
 894         hammer_mount_t hmp;
 895         struct buf *bp = NULL;
 896         int freeme = 0;
 897
 898         hmp = buffer->io.hmp;
 899
 900         crit_enter();
 901         if (buffer->io.lock.refs == 1) {
 902                 ++buffer->io.loading;   /* force interlock check */
 903                 hammer_lock_ex(&buffer->io.lock);
 904                 if (buffer->io.lock.refs == 1) {
 905                         bp = hammer_io_release(&buffer->io, flush);
 906
 907                         if (buffer->io.lock.refs == 1)
 908                                 --hammer_count_refedbufs;
 909
 910                         if (buffer->io.bp == NULL &&
 911                             buffer->io.lock.refs == 1) {
 912                                 /*
 913                                  * Final cleanup
 914                                  *
 915                                  * NOTE: It is impossible for any associated
 916                                  * B-Tree nodes to have refs if the buffer
 917                                  * has no additional refs.
 918                                  */
 919                                 RB_REMOVE(hammer_buf_rb_tree,
 920                                           &buffer->io.hmp->rb_bufs_root,
 921                                           buffer);
 922                                 volume = buffer->io.volume;
 923                                 buffer->io.volume = NULL; /* sanity */
 924                                 hammer_rel_volume(volume, 0);
 925                                 hammer_io_clear_modlist(&buffer->io);
 926                                 hammer_flush_buffer_nodes(buffer);
 927                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 928                                 freeme = 1;
 929                         }
 930                 }
 931                 --buffer->io.loading;
 932                 hammer_unlock(&buffer->io.lock);
 933         }
 934         hammer_unref(&buffer->io.lock);
 935         crit_exit();
 936         if (bp)
 937                 brelse(bp);
 938         if (freeme) {
 939                 --hammer_count_buffers;
 940                 kfree(buffer, hmp->m_misc);
 941         }
 942 }
 943
 944 /*
 945  * Access the filesystem buffer containing the specified hammer offset.
 946  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 947  * relative buffer offset.  It must also have bit 55 set to be valid.
 948  * (see hammer_off_t in hammer_disk.h).
 949  *
 950  * Any prior buffer in *bufferp will be released and replaced by the
 951  * requested buffer.
 952  *
 953  * NOTE: The buffer is indexed via its zoneX_offset but we allow the
 954  * passed cached *bufferp to match against either zoneX or zone2.
 955  */
 956 static __inline
 957 void *
 958 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 959              int *errorp, struct hammer_buffer **bufferp)
 960 {
 961         hammer_buffer_t buffer;
 962         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 963
 964         buf_offset &= ~HAMMER_BUFMASK64;
 965         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 966
 967         buffer = *bufferp;
 968         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 969                                buffer->zoneX_offset != buf_offset)) {
 970                 if (buffer)
 971                         hammer_rel_buffer(buffer, 0);
 972                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 973                 *bufferp = buffer;
 974         } else {
 975                 *errorp = 0;
 976         }
 977
 978         /*
 979          * Return a pointer to the buffer data.
 980          */
 981         if (buffer == NULL)
 982                 return(NULL);
 983         else
 984                 return((char *)buffer->ondisk + xoff);
 985 }
 986
 987 void *
 988 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 989              int *errorp, struct hammer_buffer **bufferp)
 990 {
 991         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 992 }
 993
 994 void *
 995 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 996                  int *errorp, struct hammer_buffer **bufferp)
 997 {
 998         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 999         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
1000 }
1001
1002 /*
1003  * Access the filesystem buffer containing the specified hammer offset.
1004  * No disk read operation occurs.  The result buffer may contain garbage.
1005  *
1006  * Any prior buffer in *bufferp will be released and replaced by the
1007  * requested buffer.
1008  *
1009  * This function marks the buffer dirty but does not increment its
1010  * modify_refs count.
1011  */
1012 static __inline
1013 void *
1014 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1015              int *errorp, struct hammer_buffer **bufferp)
1016 {
1017         hammer_buffer_t buffer;
1018         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1019
1020         buf_offset &= ~HAMMER_BUFMASK64;
1021
1022         buffer = *bufferp;
1023         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1024                                buffer->zoneX_offset != buf_offset)) {
1025                 if (buffer)
1026                         hammer_rel_buffer(buffer, 0);
1027                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
1028                 *bufferp = buffer;
1029         } else {
1030                 *errorp = 0;
1031         }
1032
1033         /*
1034          * Return a pointer to the buffer data.
1035          */
1036         if (buffer == NULL)
1037                 return(NULL);
1038         else
1039                 return((char *)buffer->ondisk + xoff);
1040 }
1041
1042 void *
1043 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1044              int *errorp, struct hammer_buffer **bufferp)
1045 {
1046         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1047 }
1048
1049 void *
1050 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1051                 int *errorp, struct hammer_buffer **bufferp)
1052 {
1053         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1054         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1055 }
1056
1057 /************************************************************************
1058  *                              NODES                                   *
1059  ************************************************************************
1060  *
1061  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1062  * method used by the HAMMER filesystem.
1063  *
1064  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1065  * associated with its buffer, and will only referenced the buffer while
1066  * the node itself is referenced.
1067  *
1068  * A hammer_node can also be passively associated with other HAMMER
1069  * structures, such as inodes, while retaining 0 references.  These
1070  * associations can be cleared backwards using a pointer-to-pointer in
1071  * the hammer_node.
1072  *
1073  * This allows the HAMMER implementation to cache hammer_nodes long-term
1074  * and short-cut a great deal of the infrastructure's complexity.  In
1075  * most cases a cached node can be reacquired without having to dip into
1076  * either the buffer or cluster management code.
1077  *
1078  * The caller must pass a referenced cluster on call and will retain
1079  * ownership of the reference on return.  The node will acquire its own
1080  * additional references, if necessary.
1081  */
1082 hammer_node_t
1083 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1084                 int isnew, int *errorp)
1085 {
1086         hammer_mount_t hmp = trans->hmp;
1087         hammer_node_t node;
1088
1089         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1090
1091         /*
1092          * Locate the structure, allocating one if necessary.
1093          */
1094 again:
1095         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1096         if (node == NULL) {
1097                 ++hammer_count_nodes;
1098                 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1099                 node->node_offset = node_offset;
1100                 node->hmp = hmp;
1101                 TAILQ_INIT(&node->cursor_list);
1102                 TAILQ_INIT(&node->cache_list);
1103                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1104                         --hammer_count_nodes;
1105                         kfree(node, hmp->m_misc);
1106                         goto again;
1107                 }
1108         }
1109         hammer_ref(&node->lock);
1110         if (node->ondisk) {
1111                 *errorp = 0;
1112         } else {
1113                 *errorp = hammer_load_node(trans, node, isnew);
1114                 trans->flags |= HAMMER_TRANSF_DIDIO;
1115         }
1116         if (*errorp) {
1117                 hammer_rel_node(node);
1118                 node = NULL;
1119         }
1120         return(node);
1121 }
1122
1123 /*
1124  * Reference an already-referenced node.
1125  */
1126 void
1127 hammer_ref_node(hammer_node_t node)
1128 {
1129         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1130         hammer_ref(&node->lock);
1131 }
1132
1133 /*
1134  * Load a node's on-disk data reference.
1135  */
1136 static int
1137 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew)
1138 {
1139         hammer_buffer_t buffer;
1140         hammer_off_t buf_offset;
1141         int error;
1142
1143         error = 0;
1144         ++node->loading;
1145         hammer_lock_ex(&node->lock);
1146         if (node->ondisk == NULL) {
1147                 /*
1148                  * This is a little confusing but the jist is that
1149                  * node->buffer determines whether the node is on
1150                  * the buffer's clist and node->ondisk determines
1151                  * whether the buffer is referenced.
1152                  *
1153                  * We could be racing a buffer release, in which case
1154                  * node->buffer may become NULL while we are blocked
1155                  * referencing the buffer.
1156                  */
1157                 if ((buffer = node->buffer) != NULL) {
1158                         error = hammer_ref_buffer(buffer);
1159                         if (error == 0 && node->buffer == NULL) {
1160                                 TAILQ_INSERT_TAIL(&buffer->clist,
1161                                                   node, entry);
1162                                 node->buffer = buffer;
1163                         }
1164                 } else {
1165                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1166                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1167                                                    HAMMER_BUFSIZE, 0, &error);
1168                         if (buffer) {
1169                                 KKASSERT(error == 0);
1170                                 TAILQ_INSERT_TAIL(&buffer->clist,
1171                                                   node, entry);
1172                                 node->buffer = buffer;
1173                         }
1174                 }
1175                 if (error)
1176                         goto failed;
1177                 node->ondisk = (void *)((char *)buffer->ondisk +
1178                                         (node->node_offset & HAMMER_BUFMASK));
1179
1180                 /*
1181                  * Check CRC.  NOTE: Neither flag is set and the CRC is not
1182                  * generated on new B-Tree nodes.
1183                  */
1184                 if (isnew == 0 &&
1185                     (node->flags & HAMMER_NODE_CRCANY) == 0) {
1186                         if (hammer_crc_test_btree(node->ondisk) == 0) {
1187                                 if (hammer_debug_debug & 0x0002)
1188                                         Debugger("CRC FAILED: B-TREE NODE");
1189                                 node->flags |= HAMMER_NODE_CRCBAD;
1190                         } else {
1191                                 node->flags |= HAMMER_NODE_CRCGOOD;
1192                         }
1193                 }
1194         }
1195         if (node->flags & HAMMER_NODE_CRCBAD) {
1196                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1197                         error = EDOM;
1198                 else
1199                         error = EIO;
1200         }
1201 failed:
1202         --node->loading;
1203         hammer_unlock(&node->lock);
1204         return (error);
1205 }
1206
1207 /*
1208  * Safely reference a node, interlock against flushes via the IO subsystem.
1209  */
1210 hammer_node_t
1211 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache,
1212                      int *errorp)
1213 {
1214         hammer_node_t node;
1215
1216         node = cache->node;
1217         if (node != NULL) {
1218                 hammer_ref(&node->lock);
1219                 if (node->ondisk) {
1220                         if (node->flags & HAMMER_NODE_CRCBAD) {
1221                                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1222                                         *errorp = EDOM;
1223                                 else
1224                                         *errorp = EIO;
1225                         } else {
1226                                 *errorp = 0;
1227                         }
1228                 } else {
1229                         *errorp = hammer_load_node(trans, node, 0);
1230                 }
1231                 if (*errorp) {
1232                         hammer_rel_node(node);
1233                         node = NULL;
1234                 }
1235         } else {
1236                 *errorp = ENOENT;
1237         }
1238         return(node);
1239 }
1240
1241 /*
1242  * Release a hammer_node.  On the last release the node dereferences
1243  * its underlying buffer and may or may not be destroyed.
1244  */
1245 void
1246 hammer_rel_node(hammer_node_t node)
1247 {
1248         hammer_buffer_t buffer;
1249
1250         /*
1251          * If this isn't the last ref just decrement the ref count and
1252          * return.
1253          */
1254         if (node->lock.refs > 1) {
1255                 hammer_unref(&node->lock);
1256                 return;
1257         }
1258
1259         /*
1260          * If there is no ondisk info or no buffer the node failed to load,
1261          * remove the last reference and destroy the node.
1262          */
1263         if (node->ondisk == NULL) {
1264                 hammer_unref(&node->lock);
1265                 hammer_flush_node(node);
1266                 /* node is stale now */
1267                 return;
1268         }
1269
1270         /*
1271          * Do not disassociate the node from the buffer if it represents
1272          * a modified B-Tree node that still needs its crc to be generated.
1273          */
1274         if (node->flags & HAMMER_NODE_NEEDSCRC)
1275                 return;
1276
1277         /*
1278          * Do final cleanups and then either destroy the node and leave it
1279          * passively cached.  The buffer reference is removed regardless.
1280          */
1281         buffer = node->buffer;
1282         node->ondisk = NULL;
1283
1284         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1285                 hammer_unref(&node->lock);
1286                 hammer_rel_buffer(buffer, 0);
1287                 return;
1288         }
1289
1290         /*
1291          * Destroy the node.
1292          */
1293         hammer_unref(&node->lock);
1294         hammer_flush_node(node);
1295         /* node is stale */
1296         hammer_rel_buffer(buffer, 0);
1297 }
1298
1299 /*
1300  * Free space on-media associated with a B-Tree node.
1301  */
1302 void
1303 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1304 {
1305         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1306         node->flags |= HAMMER_NODE_DELETED;
1307         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1308 }
1309
1310 /*
1311  * Passively cache a referenced hammer_node.  The caller may release
1312  * the node on return.
1313  */
1314 void
1315 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1316 {
1317         /*
1318          * If the node doesn't exist, or is being deleted, don't cache it!
1319          *
1320          * The node can only ever be NULL in the I/O failure path.
1321          */
1322         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1323                 return;
1324         if (cache->node == node)
1325                 return;
1326         while (cache->node)
1327                 hammer_uncache_node(cache);
1328         if (node->flags & HAMMER_NODE_DELETED)
1329                 return;
1330         cache->node = node;
1331         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1332 }
1333
1334 void
1335 hammer_uncache_node(hammer_node_cache_t cache)
1336 {
1337         hammer_node_t node;
1338
1339         if ((node = cache->node) != NULL) {
1340                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1341                 cache->node = NULL;
1342                 if (TAILQ_EMPTY(&node->cache_list))
1343                         hammer_flush_node(node);
1344         }
1345 }
1346
1347 /*
1348  * Remove a node's cache references and destroy the node if it has no
1349  * other references or backing store.
1350  */
1351 void
1352 hammer_flush_node(hammer_node_t node)
1353 {
1354         hammer_node_cache_t cache;
1355         hammer_buffer_t buffer;
1356         hammer_mount_t hmp = node->hmp;
1357
1358         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1359                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1360                 cache->node = NULL;
1361         }
1362         if (node->lock.refs == 0 && node->ondisk == NULL) {
1363                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1364                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1365                 if ((buffer = node->buffer) != NULL) {
1366                         node->buffer = NULL;
1367                         TAILQ_REMOVE(&buffer->clist, node, entry);
1368                         /* buffer is unreferenced because ondisk is NULL */
1369                 }
1370                 --hammer_count_nodes;
1371                 kfree(node, hmp->m_misc);
1372         }
1373 }
1374
1375 /*
1376  * Flush passively cached B-Tree nodes associated with this buffer.
1377  * This is only called when the buffer is about to be destroyed, so
1378  * none of the nodes should have any references.  The buffer is locked.
1379  *
1380  * We may be interlocked with the buffer.
1381  */
1382 void
1383 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1384 {
1385         hammer_node_t node;
1386
1387         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1388                 KKASSERT(node->ondisk == NULL);
1389                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1390
1391                 if (node->lock.refs == 0) {
1392                         hammer_ref(&node->lock);
1393                         node->flags |= HAMMER_NODE_FLUSH;
1394                         hammer_rel_node(node);
1395                 } else {
1396                         KKASSERT(node->loading != 0);
1397                         KKASSERT(node->buffer != NULL);
1398                         buffer = node->buffer;
1399                         node->buffer = NULL;
1400                         TAILQ_REMOVE(&buffer->clist, node, entry);
1401                         /* buffer is unreferenced because ondisk is NULL */
1402                 }
1403         }
1404 }
1405
1406
1407 /************************************************************************
1408  *                              ALLOCATORS                              *
1409  ************************************************************************/
1410
1411 /*
1412  * Allocate a B-Tree node.
1413  */
1414 hammer_node_t
1415 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1416 {
1417         hammer_buffer_t buffer = NULL;
1418         hammer_node_t node = NULL;
1419         hammer_off_t node_offset;
1420
1421         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1422                                             sizeof(struct hammer_node_ondisk),
1423                                             errorp);
1424         if (*errorp == 0) {
1425                 node = hammer_get_node(trans, node_offset, 1, errorp);
1426                 hammer_modify_node_noundo(trans, node);
1427                 bzero(node->ondisk, sizeof(*node->ondisk));
1428                 hammer_modify_node_done(node);
1429         }
1430         if (buffer)
1431                 hammer_rel_buffer(buffer, 0);
1432         return(node);
1433 }
1434
1435 /*
1436  * Allocate data.  If the address of a data buffer is supplied then
1437  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1438  * will be set to the related buffer.  The caller must release it when
1439  * finally done.  The initial *data_bufferp should be set to NULL by
1440  * the caller.
1441  *
1442  * The caller is responsible for making hammer_modify*() calls on the
1443  * *data_bufferp.
1444  */
1445 void *
1446 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1447                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1448                   struct hammer_buffer **data_bufferp, int *errorp)
1449 {
1450         void *data;
1451         int zone;
1452
1453         /*
1454          * Allocate data
1455          */
1456         if (data_len) {
1457                 switch(rec_type) {
1458                 case HAMMER_RECTYPE_INODE:
1459                 case HAMMER_RECTYPE_DIRENTRY:
1460                 case HAMMER_RECTYPE_EXT:
1461                 case HAMMER_RECTYPE_FIX:
1462                 case HAMMER_RECTYPE_PFS:
1463                         zone = HAMMER_ZONE_META_INDEX;
1464                         break;
1465                 case HAMMER_RECTYPE_DATA:
1466                 case HAMMER_RECTYPE_DB:
1467                         if (data_len <= HAMMER_BUFSIZE / 2) {
1468                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1469                         } else {
1470                                 data_len = (data_len + HAMMER_BUFMASK) &
1471                                            ~HAMMER_BUFMASK;
1472                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1473                         }
1474                         break;
1475                 default:
1476                         panic("hammer_alloc_data: rec_type %04x unknown",
1477                               rec_type);
1478                         zone = 0;       /* NOT REACHED */
1479                         break;
1480                 }
1481                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1482                                                       data_len, errorp);
1483         } else {
1484                 *data_offsetp = 0;
1485         }
1486         if (*errorp == 0 && data_bufferp) {
1487                 if (data_len) {
1488                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1489                                                 data_len, errorp, data_bufferp);
1490                 } else {
1491                         data = NULL;
1492                 }
1493         } else {
1494                 data = NULL;
1495         }
1496         return(data);
1497 }
1498
1499 /*
1500  * Sync dirty buffers to the media and clean-up any loose ends.
1501  *
1502  * These functions do not start the flusher going, they simply
1503  * queue everything up to the flusher.
1504  */
1505 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1506 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1507
1508 int
1509 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1510 {
1511         struct hammer_sync_info info;
1512
1513         info.error = 0;
1514         info.waitfor = waitfor;
1515         if (waitfor == MNT_WAIT) {
1516                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1517                               hammer_sync_scan1, hammer_sync_scan2, &info);
1518         } else {
1519                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1520                               hammer_sync_scan1, hammer_sync_scan2, &info);
1521         }
1522         return(info.error);
1523 }
1524
1525 /*
1526  * Filesystem sync.  If doing a synchronous sync make a second pass on
1527  * the vnodes in case any were already flushing during the first pass,
1528  * and activate the flusher twice (the second time brings the UNDO FIFO's
1529  * start position up to the end position after the first call).
1530  */
1531 int
1532 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1533 {
1534         struct hammer_sync_info info;
1535
1536         info.error = 0;
1537         info.waitfor = MNT_NOWAIT;
1538         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1539                       hammer_sync_scan1, hammer_sync_scan2, &info);
1540         if (info.error == 0 && waitfor == MNT_WAIT) {
1541                 info.waitfor = waitfor;
1542                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1543                               hammer_sync_scan1, hammer_sync_scan2, &info);
1544         }
1545         if (waitfor == MNT_WAIT) {
1546                 hammer_flusher_sync(hmp);
1547                 hammer_flusher_sync(hmp);
1548         } else {
1549                 hammer_flusher_async(hmp, NULL);
1550                 hammer_flusher_async(hmp, NULL);
1551         }
1552         return(info.error);
1553 }
1554
1555 static int
1556 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1557 {
1558         struct hammer_inode *ip;
1559
1560         ip = VTOI(vp);
1561         if (vp->v_type == VNON || ip == NULL ||
1562             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1563              RB_EMPTY(&vp->v_rbdirty_tree))) {
1564                 return(-1);
1565         }
1566         return(0);
1567 }
1568
1569 static int
1570 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1571 {
1572         struct hammer_sync_info *info = data;
1573         struct hammer_inode *ip;
1574         int error;
1575
1576         ip = VTOI(vp);
1577         if (vp->v_type == VNON || vp->v_type == VBAD ||
1578             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1579              RB_EMPTY(&vp->v_rbdirty_tree))) {
1580                 return(0);
1581         }
1582         error = VOP_FSYNC(vp, MNT_NOWAIT);
1583         if (error)
1584                 info->error = error;
1585         return(0);
1586 }
1587