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_node_t node, int isnew);
  52
  53 static int
  54 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  55 {
  56         if (vol1->vol_no < vol2->vol_no)
  57                 return(-1);
  58         if (vol1->vol_no > vol2->vol_no)
  59                 return(1);
  60         return(0);
  61 }
  62
  63 static int
  64 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  65 {
  66         if (buf1->zoneX_offset < buf2->zoneX_offset)
  67                 return(-1);
  68         if (buf1->zoneX_offset > buf2->zoneX_offset)
  69                 return(1);
  70         return(0);
  71 }
  72
  73 static int
  74 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  75 {
  76         if (node1->node_offset < node2->node_offset)
  77                 return(-1);
  78         if (node1->node_offset > node2->node_offset)
  79                 return(1);
  80         return(0);
  81 }
  82
  83 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  84              hammer_vol_rb_compare, int32_t, vol_no);
  85 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  86              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  87 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  88              hammer_nod_rb_compare, hammer_off_t, node_offset);
  89
  90 /************************************************************************
  91  *                              VOLUMES                                 *
  92  ************************************************************************
  93  *
  94  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
  95  * code on failure.  Volumes must be loaded at mount time, get_volume() will
  96  * not load a new volume.
  97  *
  98  * Calls made to hammer_load_volume() or single-threaded
  99  */
 100 int
 101 hammer_install_volume(struct hammer_mount *hmp, const char *volname,
 102                       struct vnode *devvp)
 103 {
 104         struct mount *mp;
 105         hammer_volume_t volume;
 106         struct hammer_volume_ondisk *ondisk;
 107         struct nlookupdata nd;
 108         struct buf *bp = NULL;
 109         int error;
 110         int ronly;
 111         int setmp = 0;
 112
 113         mp = hmp->mp;
 114         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 115
 116         /*
 117          * Allocate a volume structure
 118          */
 119         ++hammer_count_volumes;
 120         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 121         volume->vol_name = kstrdup(volname, M_HAMMER);
 122         volume->io.hmp = hmp;   /* bootstrap */
 123         hammer_io_init(&volume->io, volume, HAMMER_STRUCTURE_VOLUME);
 124         volume->io.offset = 0LL;
 125         volume->io.bytes = HAMMER_BUFSIZE;
 126
 127         /*
 128          * Get the device vnode
 129          */
 130         if (devvp == NULL) {
 131                 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 132                 if (error == 0)
 133                         error = nlookup(&nd);
 134                 if (error == 0)
 135                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 136                 nlookup_done(&nd);
 137         } else {
 138                 error = 0;
 139                 volume->devvp = devvp;
 140         }
 141
 142         if (error == 0) {
 143                 if (vn_isdisk(volume->devvp, &error)) {
 144                         error = vfs_mountedon(volume->devvp);
 145                 }
 146         }
 147         if (error == 0 &&
 148             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 149                 error = EBUSY;
 150         }
 151         if (error == 0) {
 152                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 153                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 154                 if (error == 0) {
 155                         error = VOP_OPEN(volume->devvp,
 156                                          (ronly ? FREAD : FREAD|FWRITE),
 157                                          FSCRED, NULL);
 158                 }
 159                 vn_unlock(volume->devvp);
 160         }
 161         if (error) {
 162                 hammer_free_volume(volume);
 163                 return(error);
 164         }
 165         volume->devvp->v_rdev->si_mountpoint = mp;
 166         setmp = 1;
 167
 168         /*
 169          * Extract the volume number from the volume header and do various
 170          * sanity checks.
 171          */
 172         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 173         if (error)
 174                 goto late_failure;
 175         ondisk = (void *)bp->b_data;
 176         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 177                 kprintf("hammer_mount: volume %s has an invalid header\n",
 178                         volume->vol_name);
 179                 error = EFTYPE;
 180                 goto late_failure;
 181         }
 182         volume->vol_no = ondisk->vol_no;
 183         volume->buffer_base = ondisk->vol_buf_beg;
 184         volume->vol_flags = ondisk->vol_flags;
 185         volume->nblocks = ondisk->vol_nblocks;
 186         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 187                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 188         volume->maxraw_off = ondisk->vol_buf_end;
 189
 190         if (RB_EMPTY(&hmp->rb_vols_root)) {
 191                 hmp->fsid = ondisk->vol_fsid;
 192         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 193                 kprintf("hammer_mount: volume %s's fsid does not match "
 194                         "other volumes\n", volume->vol_name);
 195                 error = EFTYPE;
 196                 goto late_failure;
 197         }
 198
 199         /*
 200          * Insert the volume structure into the red-black tree.
 201          */
 202         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 203                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 204                         volume->vol_name, volume->vol_no);
 205                 error = EEXIST;
 206         }
 207
 208         /*
 209          * Set the root volume .  HAMMER special cases rootvol the structure.
 210          * We do not hold a ref because this would prevent related I/O
 211          * from being flushed.
 212          */
 213         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 214                 hmp->rootvol = volume;
 215                 hmp->nvolumes = ondisk->vol_count;
 216                 if (bp) {
 217                         brelse(bp);
 218                         bp = NULL;
 219                 }
 220                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 221                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 222                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 223                         (HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
 224         }
 225 late_failure:
 226         if (bp)
 227                 brelse(bp);
 228         if (error) {
 229                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 230                 if (setmp)
 231                         volume->devvp->v_rdev->si_mountpoint = NULL;
 232                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 233                 hammer_free_volume(volume);
 234         }
 235         return (error);
 236 }
 237
 238 /*
 239  * This is called for each volume when updating the mount point from
 240  * read-write to read-only or vise-versa.
 241  */
 242 int
 243 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 244 {
 245         if (volume->devvp) {
 246                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 247                 if (volume->io.hmp->ronly) {
 248                         /* do not call vinvalbuf */
 249                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 250                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 251                 } else {
 252                         /* do not call vinvalbuf */
 253                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 254                         VOP_CLOSE(volume->devvp, FREAD);
 255                 }
 256                 vn_unlock(volume->devvp);
 257         }
 258         return(0);
 259 }
 260
 261 /*
 262  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 263  * so returns -1 on failure.
 264  */
 265 int
 266 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 267 {
 268         struct hammer_mount *hmp = volume->io.hmp;
 269         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 270         struct buf *bp;
 271
 272         /*
 273          * Clean up the root volume pointer, which is held unlocked in hmp.
 274          */
 275         if (hmp->rootvol == volume)
 276                 hmp->rootvol = NULL;
 277
 278         /*
 279          * We must not flush a dirty buffer to disk on umount.  It should
 280          * have already been dealt with by the flusher, or we may be in
 281          * catastrophic failure.
 282          */
 283         hammer_io_clear_modify(&volume->io, 1);
 284         volume->io.waitdep = 1;
 285         bp = hammer_io_release(&volume->io, 1);
 286
 287         /*
 288          * Clean up the persistent ref ioerror might have on the volume
 289          */
 290         if (volume->io.ioerror) {
 291                 volume->io.ioerror = 0;
 292                 hammer_unref(&volume->io.lock);
 293         }
 294
 295         /*
 296          * There should be no references on the volume, no clusters, and
 297          * no super-clusters.
 298          */
 299         KKASSERT(volume->io.lock.refs == 0);
 300         if (bp)
 301                 brelse(bp);
 302
 303         volume->ondisk = NULL;
 304         if (volume->devvp) {
 305                 if (volume->devvp->v_rdev &&
 306                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 307                 ) {
 308                         volume->devvp->v_rdev->si_mountpoint = NULL;
 309                 }
 310                 if (ronly) {
 311                         /*
 312                          * Make sure we don't sync anything to disk if we
 313                          * are in read-only mode (1) or critically-errored
 314                          * (2).  Note that there may be dirty buffers in
 315                          * normal read-only mode from crash recovery.
 316                          */
 317                         vinvalbuf(volume->devvp, 0, 0, 0);
 318                         VOP_CLOSE(volume->devvp, FREAD);
 319                 } else {
 320                         /*
 321                          * Normal termination, save any dirty buffers
 322                          * (XXX there really shouldn't be any).
 323                          */
 324                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 325                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 326                 }
 327         }
 328
 329         /*
 330          * Destroy the structure
 331          */
 332         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 333         hammer_free_volume(volume);
 334         return(0);
 335 }
 336
 337 static
 338 void
 339 hammer_free_volume(hammer_volume_t volume)
 340 {
 341         if (volume->vol_name) {
 342                 kfree(volume->vol_name, M_HAMMER);
 343                 volume->vol_name = NULL;
 344         }
 345         if (volume->devvp) {
 346                 vrele(volume->devvp);
 347                 volume->devvp = NULL;
 348         }
 349         --hammer_count_volumes;
 350         kfree(volume, M_HAMMER);
 351 }
 352
 353 /*
 354  * Get a HAMMER volume.  The volume must already exist.
 355  */
 356 hammer_volume_t
 357 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 358 {
 359         struct hammer_volume *volume;
 360
 361         /*
 362          * Locate the volume structure
 363          */
 364         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 365         if (volume == NULL) {
 366                 *errorp = ENOENT;
 367                 return(NULL);
 368         }
 369         hammer_ref(&volume->io.lock);
 370
 371         /*
 372          * Deal with on-disk info
 373          */
 374         if (volume->ondisk == NULL || volume->io.loading) {
 375                 *errorp = hammer_load_volume(volume);
 376                 if (*errorp) {
 377                         hammer_rel_volume(volume, 1);
 378                         volume = NULL;
 379                 }
 380         } else {
 381                 *errorp = 0;
 382         }
 383         return(volume);
 384 }
 385
 386 int
 387 hammer_ref_volume(hammer_volume_t volume)
 388 {
 389         int error;
 390
 391         hammer_ref(&volume->io.lock);
 392
 393         /*
 394          * Deal with on-disk info
 395          */
 396         if (volume->ondisk == NULL || volume->io.loading) {
 397                 error = hammer_load_volume(volume);
 398                 if (error)
 399                         hammer_rel_volume(volume, 1);
 400         } else {
 401                 error = 0;
 402         }
 403         return (error);
 404 }
 405
 406 hammer_volume_t
 407 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 408 {
 409         hammer_volume_t volume;
 410
 411         volume = hmp->rootvol;
 412         KKASSERT(volume != NULL);
 413         hammer_ref(&volume->io.lock);
 414
 415         /*
 416          * Deal with on-disk info
 417          */
 418         if (volume->ondisk == NULL || volume->io.loading) {
 419                 *errorp = hammer_load_volume(volume);
 420                 if (*errorp) {
 421                         hammer_rel_volume(volume, 1);
 422                         volume = NULL;
 423                 }
 424         } else {
 425                 *errorp = 0;
 426         }
 427         return (volume);
 428 }
 429
 430 /*
 431  * Load a volume's on-disk information.  The volume must be referenced and
 432  * not locked.  We temporarily acquire an exclusive lock to interlock
 433  * against releases or multiple get's.
 434  */
 435 static int
 436 hammer_load_volume(hammer_volume_t volume)
 437 {
 438         int error;
 439
 440         ++volume->io.loading;
 441         hammer_lock_ex(&volume->io.lock);
 442
 443         if (volume->ondisk == NULL) {
 444                 error = hammer_io_read(volume->devvp, &volume->io,
 445                                        volume->maxraw_off);
 446                 if (error == 0)
 447                         volume->ondisk = (void *)volume->io.bp->b_data;
 448         } else {
 449                 error = 0;
 450         }
 451         --volume->io.loading;
 452         hammer_unlock(&volume->io.lock);
 453         return(error);
 454 }
 455
 456 /*
 457  * Release a volume.  Call hammer_io_release on the last reference.  We have
 458  * to acquire an exclusive lock to interlock against volume->ondisk tests
 459  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 460  * lock to be held.
 461  *
 462  * Volumes are not unloaded from memory during normal operation.
 463  */
 464 void
 465 hammer_rel_volume(hammer_volume_t volume, int flush)
 466 {
 467         struct buf *bp = NULL;
 468
 469         crit_enter();
 470         if (volume->io.lock.refs == 1) {
 471                 ++volume->io.loading;
 472                 hammer_lock_ex(&volume->io.lock);
 473                 if (volume->io.lock.refs == 1) {
 474                         volume->ondisk = NULL;
 475                         bp = hammer_io_release(&volume->io, flush);
 476                 }
 477                 --volume->io.loading;
 478                 hammer_unlock(&volume->io.lock);
 479         }
 480         hammer_unref(&volume->io.lock);
 481         if (bp)
 482                 brelse(bp);
 483         crit_exit();
 484 }
 485
 486 int
 487 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 488 {
 489         hammer_volume_t vol;
 490         int i;
 491
 492         for (i = 0; i < hmp->nvolumes; ++i) {
 493                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 494                 if (vol == NULL)
 495                         return(EINVAL);
 496         }
 497         return(0);
 498 }
 499
 500 /************************************************************************
 501  *                              BUFFERS                                 *
 502  ************************************************************************
 503  *
 504  * Manage buffers.  Currently all blockmap-backed zones are translated
 505  * to zone-2 buffer offsets.
 506  */
 507 hammer_buffer_t
 508 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 509                   int bytes, int isnew, int *errorp)
 510 {
 511         hammer_buffer_t buffer;
 512         hammer_volume_t volume;
 513         hammer_off_t    zone2_offset;
 514         hammer_io_type_t iotype;
 515         int vol_no;
 516         int zone;
 517
 518         buf_offset &= ~HAMMER_BUFMASK64;
 519 again:
 520         /*
 521          * Shortcut if the buffer is already cached
 522          */
 523         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 524         if (buffer) {
 525                 if (buffer->io.lock.refs == 0)
 526                         ++hammer_count_refedbufs;
 527                 hammer_ref(&buffer->io.lock);
 528
 529                 /*
 530                  * Once refed the ondisk field will not be cleared by
 531                  * any other action.
 532                  */
 533                 if (buffer->ondisk && buffer->io.loading == 0) {
 534                         *errorp = 0;
 535                         return(buffer);
 536                 }
 537
 538                 /*
 539                  * The buffer is no longer loose if it has a ref, and
 540                  * cannot become loose once it gains a ref.  Loose
 541                  * buffers will never be in a modified state.  This should
 542                  * only occur on the 0->1 transition of refs.
 543                  *
 544                  * lose_list can be modified via a biodone() interrupt.
 545                  */
 546                 if (buffer->io.mod_list == &hmp->lose_list) {
 547                         crit_enter();   /* biodone race against list */
 548                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 549                                      mod_entry);
 550                         crit_exit();
 551                         buffer->io.mod_list = NULL;
 552                         KKASSERT(buffer->io.modified == 0);
 553                 }
 554                 goto found;
 555         }
 556
 557         /*
 558          * What is the buffer class?
 559          */
 560         zone = HAMMER_ZONE_DECODE(buf_offset);
 561
 562         switch(zone) {
 563         case HAMMER_ZONE_LARGE_DATA_INDEX:
 564         case HAMMER_ZONE_SMALL_DATA_INDEX:
 565                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 566                 break;
 567         case HAMMER_ZONE_UNDO_INDEX:
 568                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 569                 break;
 570         case HAMMER_ZONE_META_INDEX:
 571         default:
 572                 /*
 573                  * NOTE: inode data and directory entries are placed in this
 574                  * zone.  inode atime/mtime is updated in-place and thus
 575                  * buffers containing inodes must be synchronized as
 576                  * meta-buffers, same as buffers containing B-Tree info.
 577                  */
 578                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 579                 break;
 580         }
 581
 582         /*
 583          * Handle blockmap offset translations
 584          */
 585         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 586                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 587         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 588                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 589         } else {
 590                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 591                 zone2_offset = buf_offset;
 592                 *errorp = 0;
 593         }
 594         if (*errorp)
 595                 return(NULL);
 596
 597         /*
 598          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 599          * specifications.
 600          */
 601         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 602                  HAMMER_ZONE_RAW_BUFFER);
 603         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 604         volume = hammer_get_volume(hmp, vol_no, errorp);
 605         if (volume == NULL)
 606                 return(NULL);
 607
 608         KKASSERT(zone2_offset < volume->maxbuf_off);
 609
 610         /*
 611          * Allocate a new buffer structure.  We will check for races later.
 612          */
 613         ++hammer_count_buffers;
 614         buffer = kmalloc(sizeof(*buffer), M_HAMMER,
 615                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 616         buffer->zone2_offset = zone2_offset;
 617         buffer->zoneX_offset = buf_offset;
 618
 619         hammer_io_init(&buffer->io, volume, iotype);
 620         buffer->io.offset = volume->ondisk->vol_buf_beg +
 621                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 622         buffer->io.bytes = bytes;
 623         TAILQ_INIT(&buffer->clist);
 624         hammer_ref(&buffer->io.lock);
 625
 626         /*
 627          * Insert the buffer into the RB tree and handle late collisions.
 628          */
 629         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 630                 hammer_unref(&buffer->io.lock);
 631                 --hammer_count_buffers;
 632                 kfree(buffer, M_HAMMER);
 633                 goto again;
 634         }
 635         ++hammer_count_refedbufs;
 636 found:
 637
 638         /*
 639          * Deal with on-disk info and loading races.
 640          */
 641         if (buffer->ondisk == NULL || buffer->io.loading) {
 642                 *errorp = hammer_load_buffer(buffer, isnew);
 643                 if (*errorp) {
 644                         hammer_rel_buffer(buffer, 1);
 645                         buffer = NULL;
 646                 }
 647         } else {
 648                 *errorp = 0;
 649         }
 650         return(buffer);
 651 }
 652
 653 /*
 654  * This is used by the direct-read code to deal with large-data buffers
 655  * created by the reblocker and mirror-write code.  The direct-read code
 656  * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
 657  * running hammer buffers must be fully synced to disk before we can issue
 658  * the direct-read.
 659  *
 660  * This code path is not considered critical as only the rebocker and
 661  * mirror-write code will create large-data buffers via the HAMMER buffer
 662  * subsystem.  They do that because they operate at the B-Tree level and
 663  * do not access the vnode/inode structures.
 664  */
 665 void
 666 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 667 {
 668         hammer_buffer_t buffer;
 669         int error;
 670
 671         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 672                  HAMMER_ZONE_LARGE_DATA);
 673
 674         while (bytes > 0) {
 675                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 676                                    base_offset);
 677                 if (buffer && (buffer->io.modified || buffer->io.running)) {
 678                         error = hammer_ref_buffer(buffer);
 679                         if (error == 0) {
 680                                 hammer_io_wait(&buffer->io);
 681                                 if (buffer->io.modified) {
 682                                         hammer_io_write_interlock(&buffer->io);
 683                                         hammer_io_flush(&buffer->io);
 684                                         hammer_io_done_interlock(&buffer->io);
 685                                         hammer_io_wait(&buffer->io);
 686                                 }
 687                                 hammer_rel_buffer(buffer, 0);
 688                         }
 689                 }
 690                 base_offset += HAMMER_BUFSIZE;
 691                 bytes -= HAMMER_BUFSIZE;
 692         }
 693 }
 694
 695 /*
 696  * Destroy all buffers covering the specified zoneX offset range.  This
 697  * is called when the related blockmap layer2 entry is freed or when
 698  * a direct write bypasses our buffer/buffer-cache subsystem.
 699  *
 700  * The buffers may be referenced by the caller itself.  Setting reclaim
 701  * will cause the buffer to be destroyed when it's ref count reaches zero.
 702  */
 703 void
 704 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 705                    hammer_off_t zone2_offset, int bytes)
 706 {
 707         hammer_buffer_t buffer;
 708         hammer_volume_t volume;
 709         int vol_no;
 710         int error;
 711
 712         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 713         volume = hammer_get_volume(hmp, vol_no, &error);
 714         KKASSERT(error == 0);
 715
 716         while (bytes > 0) {
 717                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 718                                    base_offset);
 719                 if (buffer) {
 720                         error = hammer_ref_buffer(buffer);
 721                         if (error == 0) {
 722                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 723                                 hammer_io_clear_modify(&buffer->io, 1);
 724                                 buffer->io.reclaim = 1;
 725                                 buffer->io.waitdep = 1;
 726                                 KKASSERT(buffer->io.volume == volume);
 727                                 hammer_rel_buffer(buffer, 0);
 728                         }
 729                 } else {
 730                         hammer_io_inval(volume, zone2_offset);
 731                 }
 732                 base_offset += HAMMER_BUFSIZE;
 733                 zone2_offset += HAMMER_BUFSIZE;
 734                 bytes -= HAMMER_BUFSIZE;
 735         }
 736         hammer_rel_volume(volume, 0);
 737 }
 738
 739 static int
 740 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 741 {
 742         hammer_volume_t volume;
 743         int error;
 744
 745         /*
 746          * Load the buffer's on-disk info
 747          */
 748         volume = buffer->io.volume;
 749         ++buffer->io.loading;
 750         hammer_lock_ex(&buffer->io.lock);
 751
 752         if (hammer_debug_io & 0x0001) {
 753                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 754                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 755                         buffer->ondisk);
 756         }
 757
 758         if (buffer->ondisk == NULL) {
 759                 if (isnew) {
 760                         error = hammer_io_new(volume->devvp, &buffer->io);
 761                 } else {
 762                         error = hammer_io_read(volume->devvp, &buffer->io,
 763                                                volume->maxraw_off);
 764                 }
 765                 if (error == 0)
 766                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 767         } else if (isnew) {
 768                 error = hammer_io_new(volume->devvp, &buffer->io);
 769         } else {
 770                 error = 0;
 771         }
 772         --buffer->io.loading;
 773         hammer_unlock(&buffer->io.lock);
 774         return (error);
 775 }
 776
 777 /*
 778  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 779  * This routine is only called during unmount.
 780  */
 781 int
 782 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 783 {
 784         /*
 785          * Clean up the persistent ref ioerror might have on the buffer
 786          * and acquire a ref (steal ioerror's if we can).
 787          */
 788         if (buffer->io.ioerror) {
 789                 buffer->io.ioerror = 0;
 790         } else {
 791                 if (buffer->io.lock.refs == 0)
 792                         ++hammer_count_refedbufs;
 793                 hammer_ref(&buffer->io.lock);
 794         }
 795
 796         /*
 797          * We must not flush a dirty buffer to disk on umount.  It should
 798          * have already been dealt with by the flusher, or we may be in
 799          * catastrophic failure.
 800          */
 801         hammer_io_clear_modify(&buffer->io, 1);
 802         hammer_flush_buffer_nodes(buffer);
 803         KKASSERT(buffer->io.lock.refs == 1);
 804         hammer_rel_buffer(buffer, 2);
 805         return(0);
 806 }
 807
 808 /*
 809  * Reference a buffer that is either already referenced or via a specially
 810  * handled pointer (aka cursor->buffer).
 811  */
 812 int
 813 hammer_ref_buffer(hammer_buffer_t buffer)
 814 {
 815         int error;
 816
 817         if (buffer->io.lock.refs == 0)
 818                 ++hammer_count_refedbufs;
 819         hammer_ref(&buffer->io.lock);
 820
 821         /*
 822          * At this point a biodone() will not touch the buffer other then
 823          * incidental bits.  However, lose_list can be modified via
 824          * a biodone() interrupt.
 825          *
 826          * No longer loose
 827          */
 828         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 829                 crit_enter();
 830                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 831                 buffer->io.mod_list = NULL;
 832                 crit_exit();
 833         }
 834
 835         if (buffer->ondisk == NULL || buffer->io.loading) {
 836                 error = hammer_load_buffer(buffer, 0);
 837                 if (error) {
 838                         hammer_rel_buffer(buffer, 1);
 839                         /*
 840                          * NOTE: buffer pointer can become stale after
 841                          * the above release.
 842                          */
 843                 }
 844         } else {
 845                 error = 0;
 846         }
 847         return(error);
 848 }
 849
 850 /*
 851  * Release a buffer.  We have to deal with several places where
 852  * another thread can ref the buffer.
 853  *
 854  * Only destroy the structure itself if the related buffer cache buffer
 855  * was disassociated from it.  This ties the management of the structure
 856  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 857  * embedded io is referenced or not.
 858  */
 859 void
 860 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 861 {
 862         hammer_volume_t volume;
 863         struct buf *bp = NULL;
 864         int freeme = 0;
 865
 866         crit_enter();
 867         if (buffer->io.lock.refs == 1) {
 868                 ++buffer->io.loading;   /* force interlock check */
 869                 hammer_lock_ex(&buffer->io.lock);
 870                 if (buffer->io.lock.refs == 1) {
 871                         bp = hammer_io_release(&buffer->io, flush);
 872
 873                         if (buffer->io.lock.refs == 1)
 874                                 --hammer_count_refedbufs;
 875
 876                         if (buffer->io.bp == NULL &&
 877                             buffer->io.lock.refs == 1) {
 878                                 /*
 879                                  * Final cleanup
 880                                  *
 881                                  * NOTE: It is impossible for any associated
 882                                  * B-Tree nodes to have refs if the buffer
 883                                  * has no additional refs.
 884                                  */
 885                                 RB_REMOVE(hammer_buf_rb_tree,
 886                                           &buffer->io.hmp->rb_bufs_root,
 887                                           buffer);
 888                                 volume = buffer->io.volume;
 889                                 buffer->io.volume = NULL; /* sanity */
 890                                 hammer_rel_volume(volume, 0);
 891                                 hammer_io_clear_modlist(&buffer->io);
 892                                 hammer_flush_buffer_nodes(buffer);
 893                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 894                                 freeme = 1;
 895                         }
 896                 }
 897                 --buffer->io.loading;
 898                 hammer_unlock(&buffer->io.lock);
 899         }
 900         hammer_unref(&buffer->io.lock);
 901         crit_exit();
 902         if (bp)
 903                 brelse(bp);
 904         if (freeme) {
 905                 --hammer_count_buffers;
 906                 kfree(buffer, M_HAMMER);
 907         }
 908 }
 909
 910 /*
 911  * Access the filesystem buffer containing the specified hammer offset.
 912  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 913  * relative buffer offset.  It must also have bit 55 set to be valid.
 914  * (see hammer_off_t in hammer_disk.h).
 915  *
 916  * Any prior buffer in *bufferp will be released and replaced by the
 917  * requested buffer.
 918  */
 919 static __inline
 920 void *
 921 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 922              int *errorp, struct hammer_buffer **bufferp)
 923 {
 924         hammer_buffer_t buffer;
 925         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 926
 927         buf_offset &= ~HAMMER_BUFMASK64;
 928         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 929
 930         buffer = *bufferp;
 931         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 932                                buffer->zoneX_offset != buf_offset)) {
 933                 if (buffer)
 934                         hammer_rel_buffer(buffer, 0);
 935                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 936                 *bufferp = buffer;
 937         } else {
 938                 *errorp = 0;
 939         }
 940
 941         /*
 942          * Return a pointer to the buffer data.
 943          */
 944         if (buffer == NULL)
 945                 return(NULL);
 946         else
 947                 return((char *)buffer->ondisk + xoff);
 948 }
 949
 950 void *
 951 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 952              int *errorp, struct hammer_buffer **bufferp)
 953 {
 954         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 955 }
 956
 957 void *
 958 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 959                  int *errorp, struct hammer_buffer **bufferp)
 960 {
 961         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 962         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 963 }
 964
 965 /*
 966  * Access the filesystem buffer containing the specified hammer offset.
 967  * No disk read operation occurs.  The result buffer may contain garbage.
 968  *
 969  * Any prior buffer in *bufferp will be released and replaced by the
 970  * requested buffer.
 971  *
 972  * This function marks the buffer dirty but does not increment its
 973  * modify_refs count.
 974  */
 975 static __inline
 976 void *
 977 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 978              int *errorp, struct hammer_buffer **bufferp)
 979 {
 980         hammer_buffer_t buffer;
 981         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 982
 983         buf_offset &= ~HAMMER_BUFMASK64;
 984
 985         buffer = *bufferp;
 986         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 987                                buffer->zoneX_offset != buf_offset)) {
 988                 if (buffer)
 989                         hammer_rel_buffer(buffer, 0);
 990                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 991                 *bufferp = buffer;
 992         } else {
 993                 *errorp = 0;
 994         }
 995
 996         /*
 997          * Return a pointer to the buffer data.
 998          */
 999         if (buffer == NULL)
1000                 return(NULL);
1001         else
1002                 return((char *)buffer->ondisk + xoff);
1003 }
1004
1005 void *
1006 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1007              int *errorp, struct hammer_buffer **bufferp)
1008 {
1009         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1010 }
1011
1012 void *
1013 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1014                 int *errorp, struct hammer_buffer **bufferp)
1015 {
1016         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1017         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1018 }
1019
1020 /************************************************************************
1021  *                              NODES                                   *
1022  ************************************************************************
1023  *
1024  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1025  * method used by the HAMMER filesystem.
1026  *
1027  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1028  * associated with its buffer, and will only referenced the buffer while
1029  * the node itself is referenced.
1030  *
1031  * A hammer_node can also be passively associated with other HAMMER
1032  * structures, such as inodes, while retaining 0 references.  These
1033  * associations can be cleared backwards using a pointer-to-pointer in
1034  * the hammer_node.
1035  *
1036  * This allows the HAMMER implementation to cache hammer_nodes long-term
1037  * and short-cut a great deal of the infrastructure's complexity.  In
1038  * most cases a cached node can be reacquired without having to dip into
1039  * either the buffer or cluster management code.
1040  *
1041  * The caller must pass a referenced cluster on call and will retain
1042  * ownership of the reference on return.  The node will acquire its own
1043  * additional references, if necessary.
1044  */
1045 hammer_node_t
1046 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
1047                 int isnew, int *errorp)
1048 {
1049         hammer_node_t node;
1050
1051         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1052
1053         /*
1054          * Locate the structure, allocating one if necessary.
1055          */
1056 again:
1057         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1058         if (node == NULL) {
1059                 ++hammer_count_nodes;
1060                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO|M_USE_RESERVE);
1061                 node->node_offset = node_offset;
1062                 node->hmp = hmp;
1063                 TAILQ_INIT(&node->cursor_list);
1064                 TAILQ_INIT(&node->cache_list);
1065                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1066                         --hammer_count_nodes;
1067                         kfree(node, M_HAMMER);
1068                         goto again;
1069                 }
1070         }
1071         hammer_ref(&node->lock);
1072         if (node->ondisk)
1073                 *errorp = 0;
1074         else
1075                 *errorp = hammer_load_node(node, isnew);
1076         if (*errorp) {
1077                 hammer_rel_node(node);
1078                 node = NULL;
1079         }
1080         return(node);
1081 }
1082
1083 /*
1084  * Reference an already-referenced node.
1085  */
1086 void
1087 hammer_ref_node(hammer_node_t node)
1088 {
1089         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1090         hammer_ref(&node->lock);
1091 }
1092
1093 /*
1094  * Load a node's on-disk data reference.
1095  */
1096 static int
1097 hammer_load_node(hammer_node_t node, int isnew)
1098 {
1099         hammer_buffer_t buffer;
1100         hammer_off_t buf_offset;
1101         int error;
1102
1103         error = 0;
1104         ++node->loading;
1105         hammer_lock_ex(&node->lock);
1106         if (node->ondisk == NULL) {
1107                 /*
1108                  * This is a little confusing but the jist is that
1109                  * node->buffer determines whether the node is on
1110                  * the buffer's clist and node->ondisk determines
1111                  * whether the buffer is referenced.
1112                  *
1113                  * We could be racing a buffer release, in which case
1114                  * node->buffer may become NULL while we are blocked
1115                  * referencing the buffer.
1116                  */
1117                 if ((buffer = node->buffer) != NULL) {
1118                         error = hammer_ref_buffer(buffer);
1119                         if (error == 0 && node->buffer == NULL) {
1120                                 TAILQ_INSERT_TAIL(&buffer->clist,
1121                                                   node, entry);
1122                                 node->buffer = buffer;
1123                         }
1124                 } else {
1125                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1126                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1127                                                    HAMMER_BUFSIZE, 0, &error);
1128                         if (buffer) {
1129                                 KKASSERT(error == 0);
1130                                 TAILQ_INSERT_TAIL(&buffer->clist,
1131                                                   node, entry);
1132                                 node->buffer = buffer;
1133                         }
1134                 }
1135                 if (error)
1136                         goto failed;
1137                 node->ondisk = (void *)((char *)buffer->ondisk +
1138                                         (node->node_offset & HAMMER_BUFMASK));
1139                 if (isnew == 0 &&
1140                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1141                         if (hammer_crc_test_btree(node->ondisk) == 0)
1142                                 Debugger("CRC FAILED: B-TREE NODE");
1143                         node->flags |= HAMMER_NODE_CRCGOOD;
1144                 }
1145         }
1146 failed:
1147         --node->loading;
1148         hammer_unlock(&node->lock);
1149         return (error);
1150 }
1151
1152 /*
1153  * Safely reference a node, interlock against flushes via the IO subsystem.
1154  */
1155 hammer_node_t
1156 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1157                      int *errorp)
1158 {
1159         hammer_node_t node;
1160
1161         node = cache->node;
1162         if (node != NULL) {
1163                 hammer_ref(&node->lock);
1164                 if (node->ondisk)
1165                         *errorp = 0;
1166                 else
1167                         *errorp = hammer_load_node(node, 0);
1168                 if (*errorp) {
1169                         hammer_rel_node(node);
1170                         node = NULL;
1171                 }
1172         } else {
1173                 *errorp = ENOENT;
1174         }
1175         return(node);
1176 }
1177
1178 /*
1179  * Release a hammer_node.  On the last release the node dereferences
1180  * its underlying buffer and may or may not be destroyed.
1181  */
1182 void
1183 hammer_rel_node(hammer_node_t node)
1184 {
1185         hammer_buffer_t buffer;
1186
1187         /*
1188          * If this isn't the last ref just decrement the ref count and
1189          * return.
1190          */
1191         if (node->lock.refs > 1) {
1192                 hammer_unref(&node->lock);
1193                 return;
1194         }
1195
1196         /*
1197          * If there is no ondisk info or no buffer the node failed to load,
1198          * remove the last reference and destroy the node.
1199          */
1200         if (node->ondisk == NULL) {
1201                 hammer_unref(&node->lock);
1202                 hammer_flush_node(node);
1203                 /* node is stale now */
1204                 return;
1205         }
1206
1207         /*
1208          * Do not disassociate the node from the buffer if it represents
1209          * a modified B-Tree node that still needs its crc to be generated.
1210          */
1211         if (node->flags & HAMMER_NODE_NEEDSCRC)
1212                 return;
1213
1214         /*
1215          * Do final cleanups and then either destroy the node and leave it
1216          * passively cached.  The buffer reference is removed regardless.
1217          */
1218         buffer = node->buffer;
1219         node->ondisk = NULL;
1220
1221         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1222                 hammer_unref(&node->lock);
1223                 hammer_rel_buffer(buffer, 0);
1224                 return;
1225         }
1226
1227         /*
1228          * Destroy the node.
1229          */
1230         hammer_unref(&node->lock);
1231         hammer_flush_node(node);
1232         /* node is stale */
1233         hammer_rel_buffer(buffer, 0);
1234 }
1235
1236 /*
1237  * Free space on-media associated with a B-Tree node.
1238  */
1239 void
1240 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1241 {
1242         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1243         node->flags |= HAMMER_NODE_DELETED;
1244         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1245 }
1246
1247 /*
1248  * Passively cache a referenced hammer_node.  The caller may release
1249  * the node on return.
1250  */
1251 void
1252 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1253 {
1254         /*
1255          * If the node doesn't exist, or is being deleted, don't cache it!
1256          *
1257          * The node can only ever be NULL in the I/O failure path.
1258          */
1259         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1260                 return;
1261         if (cache->node == node)
1262                 return;
1263         while (cache->node)
1264                 hammer_uncache_node(cache);
1265         if (node->flags & HAMMER_NODE_DELETED)
1266                 return;
1267         cache->node = node;
1268         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1269 }
1270
1271 void
1272 hammer_uncache_node(hammer_node_cache_t cache)
1273 {
1274         hammer_node_t node;
1275
1276         if ((node = cache->node) != NULL) {
1277                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1278                 cache->node = NULL;
1279                 if (TAILQ_EMPTY(&node->cache_list))
1280                         hammer_flush_node(node);
1281         }
1282 }
1283
1284 /*
1285  * Remove a node's cache references and destroy the node if it has no
1286  * other references or backing store.
1287  */
1288 void
1289 hammer_flush_node(hammer_node_t node)
1290 {
1291         hammer_node_cache_t cache;
1292         hammer_buffer_t buffer;
1293
1294         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1295                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1296                 cache->node = NULL;
1297         }
1298         if (node->lock.refs == 0 && node->ondisk == NULL) {
1299                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1300                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1301                 if ((buffer = node->buffer) != NULL) {
1302                         node->buffer = NULL;
1303                         TAILQ_REMOVE(&buffer->clist, node, entry);
1304                         /* buffer is unreferenced because ondisk is NULL */
1305                 }
1306                 --hammer_count_nodes;
1307                 kfree(node, M_HAMMER);
1308         }
1309 }
1310
1311 /*
1312  * Flush passively cached B-Tree nodes associated with this buffer.
1313  * This is only called when the buffer is about to be destroyed, so
1314  * none of the nodes should have any references.  The buffer is locked.
1315  *
1316  * We may be interlocked with the buffer.
1317  */
1318 void
1319 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1320 {
1321         hammer_node_t node;
1322
1323         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1324                 KKASSERT(node->ondisk == NULL);
1325                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1326
1327                 if (node->lock.refs == 0) {
1328                         hammer_ref(&node->lock);
1329                         node->flags |= HAMMER_NODE_FLUSH;
1330                         hammer_rel_node(node);
1331                 } else {
1332                         KKASSERT(node->loading != 0);
1333                         KKASSERT(node->buffer != NULL);
1334                         buffer = node->buffer;
1335                         node->buffer = NULL;
1336                         TAILQ_REMOVE(&buffer->clist, node, entry);
1337                         /* buffer is unreferenced because ondisk is NULL */
1338                 }
1339         }
1340 }
1341
1342
1343 /************************************************************************
1344  *                              ALLOCATORS                              *
1345  ************************************************************************/
1346
1347 /*
1348  * Allocate a B-Tree node.
1349  */
1350 hammer_node_t
1351 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1352 {
1353         hammer_buffer_t buffer = NULL;
1354         hammer_node_t node = NULL;
1355         hammer_off_t node_offset;
1356
1357         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1358                                             sizeof(struct hammer_node_ondisk),
1359                                             errorp);
1360         if (*errorp == 0) {
1361                 node = hammer_get_node(trans->hmp, node_offset, 1, errorp);
1362                 hammer_modify_node_noundo(trans, node);
1363                 bzero(node->ondisk, sizeof(*node->ondisk));
1364                 hammer_modify_node_done(node);
1365         }
1366         if (buffer)
1367                 hammer_rel_buffer(buffer, 0);
1368         return(node);
1369 }
1370
1371 /*
1372  * Allocate data.  If the address of a data buffer is supplied then
1373  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1374  * will be set to the related buffer.  The caller must release it when
1375  * finally done.  The initial *data_bufferp should be set to NULL by
1376  * the caller.
1377  *
1378  * The caller is responsible for making hammer_modify*() calls on the
1379  * *data_bufferp.
1380  */
1381 void *
1382 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1383                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1384                   struct hammer_buffer **data_bufferp, int *errorp)
1385 {
1386         void *data;
1387         int zone;
1388
1389         /*
1390          * Allocate data
1391          */
1392         if (data_len) {
1393                 switch(rec_type) {
1394                 case HAMMER_RECTYPE_INODE:
1395                 case HAMMER_RECTYPE_DIRENTRY:
1396                 case HAMMER_RECTYPE_EXT:
1397                 case HAMMER_RECTYPE_FIX:
1398                 case HAMMER_RECTYPE_PFS:
1399                         zone = HAMMER_ZONE_META_INDEX;
1400                         break;
1401                 case HAMMER_RECTYPE_DATA:
1402                 case HAMMER_RECTYPE_DB:
1403                         if (data_len <= HAMMER_BUFSIZE / 2) {
1404                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1405                         } else {
1406                                 data_len = (data_len + HAMMER_BUFMASK) &
1407                                            ~HAMMER_BUFMASK;
1408                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1409                         }
1410                         break;
1411                 default:
1412                         panic("hammer_alloc_data: rec_type %04x unknown",
1413                               rec_type);
1414                         zone = 0;       /* NOT REACHED */
1415                         break;
1416                 }
1417                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1418                                                       data_len, errorp);
1419         } else {
1420                 *data_offsetp = 0;
1421         }
1422         if (*errorp == 0 && data_bufferp) {
1423                 if (data_len) {
1424                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1425                                                 data_len, errorp, data_bufferp);
1426                 } else {
1427                         data = NULL;
1428                 }
1429         } else {
1430                 data = NULL;
1431         }
1432         return(data);
1433 }
1434
1435 /*
1436  * Sync dirty buffers to the media and clean-up any loose ends.
1437  *
1438  * These functions do not start the flusher going, they simply
1439  * queue everything up to the flusher.
1440  */
1441 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1442 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1443
1444 int
1445 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1446 {
1447         struct hammer_sync_info info;
1448
1449         info.error = 0;
1450         info.waitfor = waitfor;
1451         if (waitfor == MNT_WAIT) {
1452                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1453                               hammer_sync_scan1, hammer_sync_scan2, &info);
1454         } else {
1455                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1456                               hammer_sync_scan1, hammer_sync_scan2, &info);
1457         }
1458         return(info.error);
1459 }
1460
1461 /*
1462  * Filesystem sync.  If doing a synchronous sync make a second pass on
1463  * the vnodes in case any were already flushing during the first pass,
1464  * and activate the flusher twice (the second time brings the UNDO FIFO's
1465  * start position up to the end position after the first call).
1466  */
1467 int
1468 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1469 {
1470         struct hammer_sync_info info;
1471
1472         info.error = 0;
1473         info.waitfor = MNT_NOWAIT;
1474         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1475                       hammer_sync_scan1, hammer_sync_scan2, &info);
1476         if (info.error == 0 && waitfor == MNT_WAIT) {
1477                 info.waitfor = waitfor;
1478                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1479                               hammer_sync_scan1, hammer_sync_scan2, &info);
1480         }
1481         if (waitfor == MNT_WAIT) {
1482                 hammer_flusher_sync(hmp);
1483                 hammer_flusher_sync(hmp);
1484         } else {
1485                 hammer_flusher_async(hmp, NULL);
1486                 hammer_flusher_async(hmp, NULL);
1487         }
1488         return(info.error);
1489 }
1490
1491 static int
1492 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1493 {
1494         struct hammer_inode *ip;
1495
1496         ip = VTOI(vp);
1497         if (vp->v_type == VNON || ip == NULL ||
1498             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1499              RB_EMPTY(&vp->v_rbdirty_tree))) {
1500                 return(-1);
1501         }
1502         return(0);
1503 }
1504
1505 static int
1506 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1507 {
1508         struct hammer_sync_info *info = data;
1509         struct hammer_inode *ip;
1510         int error;
1511
1512         ip = VTOI(vp);
1513         if (vp->v_type == VNON || vp->v_type == VBAD ||
1514             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1515              RB_EMPTY(&vp->v_rbdirty_tree))) {
1516                 return(0);
1517         }
1518         error = VOP_FSYNC(vp, MNT_NOWAIT);
1519         if (error)
1520                 info->error = error;
1521         return(0);
1522 }
1523