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), hmp->m_misc, M_WAITOK|M_ZERO);
 121         volume->vol_name = kstrdup(volname, hmp->m_misc);
 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         hammer_mount_t 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         hammer_mount_t hmp = volume->io.hmp;
 342
 343         if (volume->vol_name) {
 344                 kfree(volume->vol_name, hmp->m_misc);
 345                 volume->vol_name = NULL;
 346         }
 347         if (volume->devvp) {
 348                 vrele(volume->devvp);
 349                 volume->devvp = NULL;
 350         }
 351         --hammer_count_volumes;
 352         kfree(volume, hmp->m_misc);
 353 }
 354
 355 /*
 356  * Get a HAMMER volume.  The volume must already exist.
 357  */
 358 hammer_volume_t
 359 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 360 {
 361         struct hammer_volume *volume;
 362
 363         /*
 364          * Locate the volume structure
 365          */
 366         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 367         if (volume == NULL) {
 368                 *errorp = ENOENT;
 369                 return(NULL);
 370         }
 371         hammer_ref(&volume->io.lock);
 372
 373         /*
 374          * Deal with on-disk info
 375          */
 376         if (volume->ondisk == NULL || volume->io.loading) {
 377                 *errorp = hammer_load_volume(volume);
 378                 if (*errorp) {
 379                         hammer_rel_volume(volume, 1);
 380                         volume = NULL;
 381                 }
 382         } else {
 383                 *errorp = 0;
 384         }
 385         return(volume);
 386 }
 387
 388 int
 389 hammer_ref_volume(hammer_volume_t volume)
 390 {
 391         int error;
 392
 393         hammer_ref(&volume->io.lock);
 394
 395         /*
 396          * Deal with on-disk info
 397          */
 398         if (volume->ondisk == NULL || volume->io.loading) {
 399                 error = hammer_load_volume(volume);
 400                 if (error)
 401                         hammer_rel_volume(volume, 1);
 402         } else {
 403                 error = 0;
 404         }
 405         return (error);
 406 }
 407
 408 hammer_volume_t
 409 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 410 {
 411         hammer_volume_t volume;
 412
 413         volume = hmp->rootvol;
 414         KKASSERT(volume != NULL);
 415         hammer_ref(&volume->io.lock);
 416
 417         /*
 418          * Deal with on-disk info
 419          */
 420         if (volume->ondisk == NULL || volume->io.loading) {
 421                 *errorp = hammer_load_volume(volume);
 422                 if (*errorp) {
 423                         hammer_rel_volume(volume, 1);
 424                         volume = NULL;
 425                 }
 426         } else {
 427                 *errorp = 0;
 428         }
 429         return (volume);
 430 }
 431
 432 /*
 433  * Load a volume's on-disk information.  The volume must be referenced and
 434  * not locked.  We temporarily acquire an exclusive lock to interlock
 435  * against releases or multiple get's.
 436  */
 437 static int
 438 hammer_load_volume(hammer_volume_t volume)
 439 {
 440         int error;
 441
 442         ++volume->io.loading;
 443         hammer_lock_ex(&volume->io.lock);
 444
 445         if (volume->ondisk == NULL) {
 446                 error = hammer_io_read(volume->devvp, &volume->io,
 447                                        volume->maxraw_off);
 448                 if (error == 0)
 449                         volume->ondisk = (void *)volume->io.bp->b_data;
 450         } else {
 451                 error = 0;
 452         }
 453         --volume->io.loading;
 454         hammer_unlock(&volume->io.lock);
 455         return(error);
 456 }
 457
 458 /*
 459  * Release a volume.  Call hammer_io_release on the last reference.  We have
 460  * to acquire an exclusive lock to interlock against volume->ondisk tests
 461  * in hammer_load_volume(), and hammer_io_release() also expects an exclusive
 462  * lock to be held.
 463  *
 464  * Volumes are not unloaded from memory during normal operation.
 465  */
 466 void
 467 hammer_rel_volume(hammer_volume_t volume, int flush)
 468 {
 469         struct buf *bp = NULL;
 470
 471         crit_enter();
 472         if (volume->io.lock.refs == 1) {
 473                 ++volume->io.loading;
 474                 hammer_lock_ex(&volume->io.lock);
 475                 if (volume->io.lock.refs == 1) {
 476                         volume->ondisk = NULL;
 477                         bp = hammer_io_release(&volume->io, flush);
 478                 }
 479                 --volume->io.loading;
 480                 hammer_unlock(&volume->io.lock);
 481         }
 482         hammer_unref(&volume->io.lock);
 483         if (bp)
 484                 brelse(bp);
 485         crit_exit();
 486 }
 487
 488 int
 489 hammer_mountcheck_volumes(struct hammer_mount *hmp)
 490 {
 491         hammer_volume_t vol;
 492         int i;
 493
 494         for (i = 0; i < hmp->nvolumes; ++i) {
 495                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 496                 if (vol == NULL)
 497                         return(EINVAL);
 498         }
 499         return(0);
 500 }
 501
 502 /************************************************************************
 503  *                              BUFFERS                                 *
 504  ************************************************************************
 505  *
 506  * Manage buffers.  Currently all blockmap-backed zones are translated
 507  * to zone-2 buffer offsets.
 508  */
 509 hammer_buffer_t
 510 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 511                   int bytes, int isnew, int *errorp)
 512 {
 513         hammer_buffer_t buffer;
 514         hammer_volume_t volume;
 515         hammer_off_t    zone2_offset;
 516         hammer_io_type_t iotype;
 517         int vol_no;
 518         int zone;
 519
 520         buf_offset &= ~HAMMER_BUFMASK64;
 521 again:
 522         /*
 523          * Shortcut if the buffer is already cached
 524          */
 525         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 526         if (buffer) {
 527                 if (buffer->io.lock.refs == 0)
 528                         ++hammer_count_refedbufs;
 529                 hammer_ref(&buffer->io.lock);
 530
 531                 /*
 532                  * Once refed the ondisk field will not be cleared by
 533                  * any other action.
 534                  */
 535                 if (buffer->ondisk && buffer->io.loading == 0) {
 536                         *errorp = 0;
 537                         return(buffer);
 538                 }
 539
 540                 /*
 541                  * The buffer is no longer loose if it has a ref, and
 542                  * cannot become loose once it gains a ref.  Loose
 543                  * buffers will never be in a modified state.  This should
 544                  * only occur on the 0->1 transition of refs.
 545                  *
 546                  * lose_list can be modified via a biodone() interrupt.
 547                  */
 548                 if (buffer->io.mod_list == &hmp->lose_list) {
 549                         crit_enter();   /* biodone race against list */
 550                         TAILQ_REMOVE(buffer->io.mod_list, &buffer->io,
 551                                      mod_entry);
 552                         crit_exit();
 553                         buffer->io.mod_list = NULL;
 554                         KKASSERT(buffer->io.modified == 0);
 555                 }
 556                 goto found;
 557         }
 558
 559         /*
 560          * What is the buffer class?
 561          */
 562         zone = HAMMER_ZONE_DECODE(buf_offset);
 563
 564         switch(zone) {
 565         case HAMMER_ZONE_LARGE_DATA_INDEX:
 566         case HAMMER_ZONE_SMALL_DATA_INDEX:
 567                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 568                 break;
 569         case HAMMER_ZONE_UNDO_INDEX:
 570                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 571                 break;
 572         case HAMMER_ZONE_META_INDEX:
 573         default:
 574                 /*
 575                  * NOTE: inode data and directory entries are placed in this
 576                  * zone.  inode atime/mtime is updated in-place and thus
 577                  * buffers containing inodes must be synchronized as
 578                  * meta-buffers, same as buffers containing B-Tree info.
 579                  */
 580                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 581                 break;
 582         }
 583
 584         /*
 585          * Handle blockmap offset translations
 586          */
 587         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 588                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 589         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 590                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 591         } else {
 592                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 593                 zone2_offset = buf_offset;
 594                 *errorp = 0;
 595         }
 596         if (*errorp)
 597                 return(NULL);
 598
 599         /*
 600          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 601          * specifications.
 602          */
 603         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 604                  HAMMER_ZONE_RAW_BUFFER);
 605         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 606         volume = hammer_get_volume(hmp, vol_no, errorp);
 607         if (volume == NULL)
 608                 return(NULL);
 609
 610         KKASSERT(zone2_offset < volume->maxbuf_off);
 611
 612         /*
 613          * Allocate a new buffer structure.  We will check for races later.
 614          */
 615         ++hammer_count_buffers;
 616         buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
 617                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 618         buffer->zone2_offset = zone2_offset;
 619         buffer->zoneX_offset = buf_offset;
 620
 621         hammer_io_init(&buffer->io, volume, iotype);
 622         buffer->io.offset = volume->ondisk->vol_buf_beg +
 623                             (zone2_offset & HAMMER_OFF_SHORT_MASK);
 624         buffer->io.bytes = bytes;
 625         TAILQ_INIT(&buffer->clist);
 626         hammer_ref(&buffer->io.lock);
 627
 628         /*
 629          * Insert the buffer into the RB tree and handle late collisions.
 630          */
 631         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 632                 hammer_unref(&buffer->io.lock);
 633                 --hammer_count_buffers;
 634                 kfree(buffer, hmp->m_misc);
 635                 goto again;
 636         }
 637         ++hammer_count_refedbufs;
 638 found:
 639
 640         /*
 641          * Deal with on-disk info and loading races.
 642          */
 643         if (buffer->ondisk == NULL || buffer->io.loading) {
 644                 *errorp = hammer_load_buffer(buffer, isnew);
 645                 if (*errorp) {
 646                         hammer_rel_buffer(buffer, 1);
 647                         buffer = NULL;
 648                 }
 649         } else {
 650                 *errorp = 0;
 651         }
 652         return(buffer);
 653 }
 654
 655 /*
 656  * This is used by the direct-read code to deal with large-data buffers
 657  * created by the reblocker and mirror-write code.  The direct-read code
 658  * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
 659  * running hammer buffers must be fully synced to disk before we can issue
 660  * the direct-read.
 661  *
 662  * This code path is not considered critical as only the rebocker and
 663  * mirror-write code will create large-data buffers via the HAMMER buffer
 664  * subsystem.  They do that because they operate at the B-Tree level and
 665  * do not access the vnode/inode structures.
 666  */
 667 void
 668 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 669 {
 670         hammer_buffer_t buffer;
 671         int error;
 672
 673         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 674                  HAMMER_ZONE_LARGE_DATA);
 675
 676         while (bytes > 0) {
 677                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 678                                    base_offset);
 679                 if (buffer && (buffer->io.modified || buffer->io.running)) {
 680                         error = hammer_ref_buffer(buffer);
 681                         if (error == 0) {
 682                                 hammer_io_wait(&buffer->io);
 683                                 if (buffer->io.modified) {
 684                                         hammer_io_write_interlock(&buffer->io);
 685                                         hammer_io_flush(&buffer->io);
 686                                         hammer_io_done_interlock(&buffer->io);
 687                                         hammer_io_wait(&buffer->io);
 688                                 }
 689                                 hammer_rel_buffer(buffer, 0);
 690                         }
 691                 }
 692                 base_offset += HAMMER_BUFSIZE;
 693                 bytes -= HAMMER_BUFSIZE;
 694         }
 695 }
 696
 697 /*
 698  * Destroy all buffers covering the specified zoneX offset range.  This
 699  * is called when the related blockmap layer2 entry is freed or when
 700  * a direct write bypasses our buffer/buffer-cache subsystem.
 701  *
 702  * The buffers may be referenced by the caller itself.  Setting reclaim
 703  * will cause the buffer to be destroyed when it's ref count reaches zero.
 704  */
 705 void
 706 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 707                    hammer_off_t zone2_offset, int bytes)
 708 {
 709         hammer_buffer_t buffer;
 710         hammer_volume_t volume;
 711         int vol_no;
 712         int error;
 713
 714         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 715         volume = hammer_get_volume(hmp, vol_no, &error);
 716         KKASSERT(error == 0);
 717
 718         while (bytes > 0) {
 719                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 720                                    base_offset);
 721                 if (buffer) {
 722                         error = hammer_ref_buffer(buffer);
 723                         if (error == 0) {
 724                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 725                                 hammer_io_clear_modify(&buffer->io, 1);
 726                                 buffer->io.reclaim = 1;
 727                                 buffer->io.waitdep = 1;
 728                                 KKASSERT(buffer->io.volume == volume);
 729                                 hammer_rel_buffer(buffer, 0);
 730                         }
 731                 } else {
 732                         hammer_io_inval(volume, zone2_offset);
 733                 }
 734                 base_offset += HAMMER_BUFSIZE;
 735                 zone2_offset += HAMMER_BUFSIZE;
 736                 bytes -= HAMMER_BUFSIZE;
 737         }
 738         hammer_rel_volume(volume, 0);
 739 }
 740
 741 static int
 742 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 743 {
 744         hammer_volume_t volume;
 745         int error;
 746
 747         /*
 748          * Load the buffer's on-disk info
 749          */
 750         volume = buffer->io.volume;
 751         ++buffer->io.loading;
 752         hammer_lock_ex(&buffer->io.lock);
 753
 754         if (hammer_debug_io & 0x0001) {
 755                 kprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 756                         buffer->zoneX_offset, buffer->zone2_offset, isnew,
 757                         buffer->ondisk);
 758         }
 759
 760         if (buffer->ondisk == NULL) {
 761                 if (isnew) {
 762                         error = hammer_io_new(volume->devvp, &buffer->io);
 763                 } else {
 764                         error = hammer_io_read(volume->devvp, &buffer->io,
 765                                                volume->maxraw_off);
 766                 }
 767                 if (error == 0)
 768                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 769         } else if (isnew) {
 770                 error = hammer_io_new(volume->devvp, &buffer->io);
 771         } else {
 772                 error = 0;
 773         }
 774         --buffer->io.loading;
 775         hammer_unlock(&buffer->io.lock);
 776         return (error);
 777 }
 778
 779 /*
 780  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 781  * This routine is only called during unmount.
 782  */
 783 int
 784 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 785 {
 786         /*
 787          * Clean up the persistent ref ioerror might have on the buffer
 788          * and acquire a ref (steal ioerror's if we can).
 789          */
 790         if (buffer->io.ioerror) {
 791                 buffer->io.ioerror = 0;
 792         } else {
 793                 if (buffer->io.lock.refs == 0)
 794                         ++hammer_count_refedbufs;
 795                 hammer_ref(&buffer->io.lock);
 796         }
 797
 798         /*
 799          * We must not flush a dirty buffer to disk on umount.  It should
 800          * have already been dealt with by the flusher, or we may be in
 801          * catastrophic failure.
 802          */
 803         hammer_io_clear_modify(&buffer->io, 1);
 804         hammer_flush_buffer_nodes(buffer);
 805         KKASSERT(buffer->io.lock.refs == 1);
 806         hammer_rel_buffer(buffer, 2);
 807         return(0);
 808 }
 809
 810 /*
 811  * Reference a buffer that is either already referenced or via a specially
 812  * handled pointer (aka cursor->buffer).
 813  */
 814 int
 815 hammer_ref_buffer(hammer_buffer_t buffer)
 816 {
 817         int error;
 818
 819         if (buffer->io.lock.refs == 0)
 820                 ++hammer_count_refedbufs;
 821         hammer_ref(&buffer->io.lock);
 822
 823         /*
 824          * At this point a biodone() will not touch the buffer other then
 825          * incidental bits.  However, lose_list can be modified via
 826          * a biodone() interrupt.
 827          *
 828          * No longer loose
 829          */
 830         if (buffer->io.mod_list == &buffer->io.hmp->lose_list) {
 831                 crit_enter();
 832                 TAILQ_REMOVE(buffer->io.mod_list, &buffer->io, mod_entry);
 833                 buffer->io.mod_list = NULL;
 834                 crit_exit();
 835         }
 836
 837         if (buffer->ondisk == NULL || buffer->io.loading) {
 838                 error = hammer_load_buffer(buffer, 0);
 839                 if (error) {
 840                         hammer_rel_buffer(buffer, 1);
 841                         /*
 842                          * NOTE: buffer pointer can become stale after
 843                          * the above release.
 844                          */
 845                 }
 846         } else {
 847                 error = 0;
 848         }
 849         return(error);
 850 }
 851
 852 /*
 853  * Release a buffer.  We have to deal with several places where
 854  * another thread can ref the buffer.
 855  *
 856  * Only destroy the structure itself if the related buffer cache buffer
 857  * was disassociated from it.  This ties the management of the structure
 858  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 859  * embedded io is referenced or not.
 860  */
 861 void
 862 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 863 {
 864         hammer_volume_t volume;
 865         hammer_mount_t hmp;
 866         struct buf *bp = NULL;
 867         int freeme = 0;
 868
 869         hmp = buffer->io.hmp;
 870
 871         crit_enter();
 872         if (buffer->io.lock.refs == 1) {
 873                 ++buffer->io.loading;   /* force interlock check */
 874                 hammer_lock_ex(&buffer->io.lock);
 875                 if (buffer->io.lock.refs == 1) {
 876                         bp = hammer_io_release(&buffer->io, flush);
 877
 878                         if (buffer->io.lock.refs == 1)
 879                                 --hammer_count_refedbufs;
 880
 881                         if (buffer->io.bp == NULL &&
 882                             buffer->io.lock.refs == 1) {
 883                                 /*
 884                                  * Final cleanup
 885                                  *
 886                                  * NOTE: It is impossible for any associated
 887                                  * B-Tree nodes to have refs if the buffer
 888                                  * has no additional refs.
 889                                  */
 890                                 RB_REMOVE(hammer_buf_rb_tree,
 891                                           &buffer->io.hmp->rb_bufs_root,
 892                                           buffer);
 893                                 volume = buffer->io.volume;
 894                                 buffer->io.volume = NULL; /* sanity */
 895                                 hammer_rel_volume(volume, 0);
 896                                 hammer_io_clear_modlist(&buffer->io);
 897                                 hammer_flush_buffer_nodes(buffer);
 898                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 899                                 freeme = 1;
 900                         }
 901                 }
 902                 --buffer->io.loading;
 903                 hammer_unlock(&buffer->io.lock);
 904         }
 905         hammer_unref(&buffer->io.lock);
 906         crit_exit();
 907         if (bp)
 908                 brelse(bp);
 909         if (freeme) {
 910                 --hammer_count_buffers;
 911                 kfree(buffer, hmp->m_misc);
 912         }
 913 }
 914
 915 /*
 916  * Access the filesystem buffer containing the specified hammer offset.
 917  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 918  * relative buffer offset.  It must also have bit 55 set to be valid.
 919  * (see hammer_off_t in hammer_disk.h).
 920  *
 921  * Any prior buffer in *bufferp will be released and replaced by the
 922  * requested buffer.
 923  */
 924 static __inline
 925 void *
 926 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 927              int *errorp, struct hammer_buffer **bufferp)
 928 {
 929         hammer_buffer_t buffer;
 930         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 931
 932         buf_offset &= ~HAMMER_BUFMASK64;
 933         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 934
 935         buffer = *bufferp;
 936         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 937                                buffer->zoneX_offset != buf_offset)) {
 938                 if (buffer)
 939                         hammer_rel_buffer(buffer, 0);
 940                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 0, errorp);
 941                 *bufferp = buffer;
 942         } else {
 943                 *errorp = 0;
 944         }
 945
 946         /*
 947          * Return a pointer to the buffer data.
 948          */
 949         if (buffer == NULL)
 950                 return(NULL);
 951         else
 952                 return((char *)buffer->ondisk + xoff);
 953 }
 954
 955 void *
 956 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
 957              int *errorp, struct hammer_buffer **bufferp)
 958 {
 959         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
 960 }
 961
 962 void *
 963 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 964                  int *errorp, struct hammer_buffer **bufferp)
 965 {
 966         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
 967         return(_hammer_bread(hmp, buf_offset, bytes, errorp, bufferp));
 968 }
 969
 970 /*
 971  * Access the filesystem buffer containing the specified hammer offset.
 972  * No disk read operation occurs.  The result buffer may contain garbage.
 973  *
 974  * Any prior buffer in *bufferp will be released and replaced by the
 975  * requested buffer.
 976  *
 977  * This function marks the buffer dirty but does not increment its
 978  * modify_refs count.
 979  */
 980 static __inline
 981 void *
 982 _hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
 983              int *errorp, struct hammer_buffer **bufferp)
 984 {
 985         hammer_buffer_t buffer;
 986         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 987
 988         buf_offset &= ~HAMMER_BUFMASK64;
 989
 990         buffer = *bufferp;
 991         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 992                                buffer->zoneX_offset != buf_offset)) {
 993                 if (buffer)
 994                         hammer_rel_buffer(buffer, 0);
 995                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, 1, errorp);
 996                 *bufferp = buffer;
 997         } else {
 998                 *errorp = 0;
 999         }
1000
1001         /*
1002          * Return a pointer to the buffer data.
1003          */
1004         if (buffer == NULL)
1005                 return(NULL);
1006         else
1007                 return((char *)buffer->ondisk + xoff);
1008 }
1009
1010 void *
1011 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1012              int *errorp, struct hammer_buffer **bufferp)
1013 {
1014         return(_hammer_bnew(hmp, buf_offset, HAMMER_BUFSIZE, errorp, bufferp));
1015 }
1016
1017 void *
1018 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1019                 int *errorp, struct hammer_buffer **bufferp)
1020 {
1021         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1022         return(_hammer_bnew(hmp, buf_offset, bytes, errorp, bufferp));
1023 }
1024
1025 /************************************************************************
1026  *                              NODES                                   *
1027  ************************************************************************
1028  *
1029  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1030  * method used by the HAMMER filesystem.
1031  *
1032  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1033  * associated with its buffer, and will only referenced the buffer while
1034  * the node itself is referenced.
1035  *
1036  * A hammer_node can also be passively associated with other HAMMER
1037  * structures, such as inodes, while retaining 0 references.  These
1038  * associations can be cleared backwards using a pointer-to-pointer in
1039  * the hammer_node.
1040  *
1041  * This allows the HAMMER implementation to cache hammer_nodes long-term
1042  * and short-cut a great deal of the infrastructure's complexity.  In
1043  * most cases a cached node can be reacquired without having to dip into
1044  * either the buffer or cluster management code.
1045  *
1046  * The caller must pass a referenced cluster on call and will retain
1047  * ownership of the reference on return.  The node will acquire its own
1048  * additional references, if necessary.
1049  */
1050 hammer_node_t
1051 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1052                 int isnew, int *errorp)
1053 {
1054         hammer_mount_t hmp = trans->hmp;
1055         hammer_node_t node;
1056
1057         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1058
1059         /*
1060          * Locate the structure, allocating one if necessary.
1061          */
1062 again:
1063         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1064         if (node == NULL) {
1065                 ++hammer_count_nodes;
1066                 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1067                 node->node_offset = node_offset;
1068                 node->hmp = hmp;
1069                 TAILQ_INIT(&node->cursor_list);
1070                 TAILQ_INIT(&node->cache_list);
1071                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1072                         --hammer_count_nodes;
1073                         kfree(node, hmp->m_misc);
1074                         goto again;
1075                 }
1076         }
1077         hammer_ref(&node->lock);
1078         if (node->ondisk) {
1079                 *errorp = 0;
1080         } else {
1081                 *errorp = hammer_load_node(node, isnew);
1082                 trans->flags |= HAMMER_TRANSF_DIDIO;
1083         }
1084         if (*errorp) {
1085                 hammer_rel_node(node);
1086                 node = NULL;
1087         }
1088         return(node);
1089 }
1090
1091 /*
1092  * Reference an already-referenced node.
1093  */
1094 void
1095 hammer_ref_node(hammer_node_t node)
1096 {
1097         KKASSERT(node->lock.refs > 0 && node->ondisk != NULL);
1098         hammer_ref(&node->lock);
1099 }
1100
1101 /*
1102  * Load a node's on-disk data reference.
1103  */
1104 static int
1105 hammer_load_node(hammer_node_t node, int isnew)
1106 {
1107         hammer_buffer_t buffer;
1108         hammer_off_t buf_offset;
1109         int error;
1110
1111         error = 0;
1112         ++node->loading;
1113         hammer_lock_ex(&node->lock);
1114         if (node->ondisk == NULL) {
1115                 /*
1116                  * This is a little confusing but the jist is that
1117                  * node->buffer determines whether the node is on
1118                  * the buffer's clist and node->ondisk determines
1119                  * whether the buffer is referenced.
1120                  *
1121                  * We could be racing a buffer release, in which case
1122                  * node->buffer may become NULL while we are blocked
1123                  * referencing the buffer.
1124                  */
1125                 if ((buffer = node->buffer) != NULL) {
1126                         error = hammer_ref_buffer(buffer);
1127                         if (error == 0 && node->buffer == NULL) {
1128                                 TAILQ_INSERT_TAIL(&buffer->clist,
1129                                                   node, entry);
1130                                 node->buffer = buffer;
1131                         }
1132                 } else {
1133                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1134                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1135                                                    HAMMER_BUFSIZE, 0, &error);
1136                         if (buffer) {
1137                                 KKASSERT(error == 0);
1138                                 TAILQ_INSERT_TAIL(&buffer->clist,
1139                                                   node, entry);
1140                                 node->buffer = buffer;
1141                         }
1142                 }
1143                 if (error)
1144                         goto failed;
1145                 node->ondisk = (void *)((char *)buffer->ondisk +
1146                                         (node->node_offset & HAMMER_BUFMASK));
1147                 if (isnew == 0 &&
1148                     (node->flags & HAMMER_NODE_CRCGOOD) == 0) {
1149                         if (hammer_crc_test_btree(node->ondisk) == 0)
1150                                 Debugger("CRC FAILED: B-TREE NODE");
1151                         node->flags |= HAMMER_NODE_CRCGOOD;
1152                 }
1153         }
1154 failed:
1155         --node->loading;
1156         hammer_unlock(&node->lock);
1157         return (error);
1158 }
1159
1160 /*
1161  * Safely reference a node, interlock against flushes via the IO subsystem.
1162  */
1163 hammer_node_t
1164 hammer_ref_node_safe(struct hammer_mount *hmp, hammer_node_cache_t cache,
1165                      int *errorp)
1166 {
1167         hammer_node_t node;
1168
1169         node = cache->node;
1170         if (node != NULL) {
1171                 hammer_ref(&node->lock);
1172                 if (node->ondisk)
1173                         *errorp = 0;
1174                 else
1175                         *errorp = hammer_load_node(node, 0);
1176                 if (*errorp) {
1177                         hammer_rel_node(node);
1178                         node = NULL;
1179                 }
1180         } else {
1181                 *errorp = ENOENT;
1182         }
1183         return(node);
1184 }
1185
1186 /*
1187  * Release a hammer_node.  On the last release the node dereferences
1188  * its underlying buffer and may or may not be destroyed.
1189  */
1190 void
1191 hammer_rel_node(hammer_node_t node)
1192 {
1193         hammer_buffer_t buffer;
1194
1195         /*
1196          * If this isn't the last ref just decrement the ref count and
1197          * return.
1198          */
1199         if (node->lock.refs > 1) {
1200                 hammer_unref(&node->lock);
1201                 return;
1202         }
1203
1204         /*
1205          * If there is no ondisk info or no buffer the node failed to load,
1206          * remove the last reference and destroy the node.
1207          */
1208         if (node->ondisk == NULL) {
1209                 hammer_unref(&node->lock);
1210                 hammer_flush_node(node);
1211                 /* node is stale now */
1212                 return;
1213         }
1214
1215         /*
1216          * Do not disassociate the node from the buffer if it represents
1217          * a modified B-Tree node that still needs its crc to be generated.
1218          */
1219         if (node->flags & HAMMER_NODE_NEEDSCRC)
1220                 return;
1221
1222         /*
1223          * Do final cleanups and then either destroy the node and leave it
1224          * passively cached.  The buffer reference is removed regardless.
1225          */
1226         buffer = node->buffer;
1227         node->ondisk = NULL;
1228
1229         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1230                 hammer_unref(&node->lock);
1231                 hammer_rel_buffer(buffer, 0);
1232                 return;
1233         }
1234
1235         /*
1236          * Destroy the node.
1237          */
1238         hammer_unref(&node->lock);
1239         hammer_flush_node(node);
1240         /* node is stale */
1241         hammer_rel_buffer(buffer, 0);
1242 }
1243
1244 /*
1245  * Free space on-media associated with a B-Tree node.
1246  */
1247 void
1248 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1249 {
1250         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1251         node->flags |= HAMMER_NODE_DELETED;
1252         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1253 }
1254
1255 /*
1256  * Passively cache a referenced hammer_node.  The caller may release
1257  * the node on return.
1258  */
1259 void
1260 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1261 {
1262         /*
1263          * If the node doesn't exist, or is being deleted, don't cache it!
1264          *
1265          * The node can only ever be NULL in the I/O failure path.
1266          */
1267         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1268                 return;
1269         if (cache->node == node)
1270                 return;
1271         while (cache->node)
1272                 hammer_uncache_node(cache);
1273         if (node->flags & HAMMER_NODE_DELETED)
1274                 return;
1275         cache->node = node;
1276         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1277 }
1278
1279 void
1280 hammer_uncache_node(hammer_node_cache_t cache)
1281 {
1282         hammer_node_t node;
1283
1284         if ((node = cache->node) != NULL) {
1285                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1286                 cache->node = NULL;
1287                 if (TAILQ_EMPTY(&node->cache_list))
1288                         hammer_flush_node(node);
1289         }
1290 }
1291
1292 /*
1293  * Remove a node's cache references and destroy the node if it has no
1294  * other references or backing store.
1295  */
1296 void
1297 hammer_flush_node(hammer_node_t node)
1298 {
1299         hammer_node_cache_t cache;
1300         hammer_buffer_t buffer;
1301         hammer_mount_t hmp = node->hmp;
1302
1303         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1304                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1305                 cache->node = NULL;
1306         }
1307         if (node->lock.refs == 0 && node->ondisk == NULL) {
1308                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1309                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1310                 if ((buffer = node->buffer) != NULL) {
1311                         node->buffer = NULL;
1312                         TAILQ_REMOVE(&buffer->clist, node, entry);
1313                         /* buffer is unreferenced because ondisk is NULL */
1314                 }
1315                 --hammer_count_nodes;
1316                 kfree(node, hmp->m_misc);
1317         }
1318 }
1319
1320 /*
1321  * Flush passively cached B-Tree nodes associated with this buffer.
1322  * This is only called when the buffer is about to be destroyed, so
1323  * none of the nodes should have any references.  The buffer is locked.
1324  *
1325  * We may be interlocked with the buffer.
1326  */
1327 void
1328 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1329 {
1330         hammer_node_t node;
1331
1332         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1333                 KKASSERT(node->ondisk == NULL);
1334                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1335
1336                 if (node->lock.refs == 0) {
1337                         hammer_ref(&node->lock);
1338                         node->flags |= HAMMER_NODE_FLUSH;
1339                         hammer_rel_node(node);
1340                 } else {
1341                         KKASSERT(node->loading != 0);
1342                         KKASSERT(node->buffer != NULL);
1343                         buffer = node->buffer;
1344                         node->buffer = NULL;
1345                         TAILQ_REMOVE(&buffer->clist, node, entry);
1346                         /* buffer is unreferenced because ondisk is NULL */
1347                 }
1348         }
1349 }
1350
1351
1352 /************************************************************************
1353  *                              ALLOCATORS                              *
1354  ************************************************************************/
1355
1356 /*
1357  * Allocate a B-Tree node.
1358  */
1359 hammer_node_t
1360 hammer_alloc_btree(hammer_transaction_t trans, int *errorp)
1361 {
1362         hammer_buffer_t buffer = NULL;
1363         hammer_node_t node = NULL;
1364         hammer_off_t node_offset;
1365
1366         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1367                                             sizeof(struct hammer_node_ondisk),
1368                                             errorp);
1369         if (*errorp == 0) {
1370                 node = hammer_get_node(trans, node_offset, 1, errorp);
1371                 hammer_modify_node_noundo(trans, node);
1372                 bzero(node->ondisk, sizeof(*node->ondisk));
1373                 hammer_modify_node_done(node);
1374         }
1375         if (buffer)
1376                 hammer_rel_buffer(buffer, 0);
1377         return(node);
1378 }
1379
1380 /*
1381  * Allocate data.  If the address of a data buffer is supplied then
1382  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1383  * will be set to the related buffer.  The caller must release it when
1384  * finally done.  The initial *data_bufferp should be set to NULL by
1385  * the caller.
1386  *
1387  * The caller is responsible for making hammer_modify*() calls on the
1388  * *data_bufferp.
1389  */
1390 void *
1391 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1392                   u_int16_t rec_type, hammer_off_t *data_offsetp,
1393                   struct hammer_buffer **data_bufferp, int *errorp)
1394 {
1395         void *data;
1396         int zone;
1397
1398         /*
1399          * Allocate data
1400          */
1401         if (data_len) {
1402                 switch(rec_type) {
1403                 case HAMMER_RECTYPE_INODE:
1404                 case HAMMER_RECTYPE_DIRENTRY:
1405                 case HAMMER_RECTYPE_EXT:
1406                 case HAMMER_RECTYPE_FIX:
1407                 case HAMMER_RECTYPE_PFS:
1408                         zone = HAMMER_ZONE_META_INDEX;
1409                         break;
1410                 case HAMMER_RECTYPE_DATA:
1411                 case HAMMER_RECTYPE_DB:
1412                         if (data_len <= HAMMER_BUFSIZE / 2) {
1413                                 zone = HAMMER_ZONE_SMALL_DATA_INDEX;
1414                         } else {
1415                                 data_len = (data_len + HAMMER_BUFMASK) &
1416                                            ~HAMMER_BUFMASK;
1417                                 zone = HAMMER_ZONE_LARGE_DATA_INDEX;
1418                         }
1419                         break;
1420                 default:
1421                         panic("hammer_alloc_data: rec_type %04x unknown",
1422                               rec_type);
1423                         zone = 0;       /* NOT REACHED */
1424                         break;
1425                 }
1426                 *data_offsetp = hammer_blockmap_alloc(trans, zone,
1427                                                       data_len, errorp);
1428         } else {
1429                 *data_offsetp = 0;
1430         }
1431         if (*errorp == 0 && data_bufferp) {
1432                 if (data_len) {
1433                         data = hammer_bread_ext(trans->hmp, *data_offsetp,
1434                                                 data_len, errorp, data_bufferp);
1435                 } else {
1436                         data = NULL;
1437                 }
1438         } else {
1439                 data = NULL;
1440         }
1441         return(data);
1442 }
1443
1444 /*
1445  * Sync dirty buffers to the media and clean-up any loose ends.
1446  *
1447  * These functions do not start the flusher going, they simply
1448  * queue everything up to the flusher.
1449  */
1450 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1451 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1452
1453 int
1454 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1455 {
1456         struct hammer_sync_info info;
1457
1458         info.error = 0;
1459         info.waitfor = waitfor;
1460         if (waitfor == MNT_WAIT) {
1461                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS,
1462                               hammer_sync_scan1, hammer_sync_scan2, &info);
1463         } else {
1464                 vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_ONEPASS|VMSC_NOWAIT,
1465                               hammer_sync_scan1, hammer_sync_scan2, &info);
1466         }
1467         return(info.error);
1468 }
1469
1470 /*
1471  * Filesystem sync.  If doing a synchronous sync make a second pass on
1472  * the vnodes in case any were already flushing during the first pass,
1473  * and activate the flusher twice (the second time brings the UNDO FIFO's
1474  * start position up to the end position after the first call).
1475  */
1476 int
1477 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1478 {
1479         struct hammer_sync_info info;
1480
1481         info.error = 0;
1482         info.waitfor = MNT_NOWAIT;
1483         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1484                       hammer_sync_scan1, hammer_sync_scan2, &info);
1485         if (info.error == 0 && waitfor == MNT_WAIT) {
1486                 info.waitfor = waitfor;
1487                 vmntvnodescan(hmp->mp, VMSC_GETVP,
1488                               hammer_sync_scan1, hammer_sync_scan2, &info);
1489         }
1490         if (waitfor == MNT_WAIT) {
1491                 hammer_flusher_sync(hmp);
1492                 hammer_flusher_sync(hmp);
1493         } else {
1494                 hammer_flusher_async(hmp, NULL);
1495                 hammer_flusher_async(hmp, NULL);
1496         }
1497         return(info.error);
1498 }
1499
1500 static int
1501 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1502 {
1503         struct hammer_inode *ip;
1504
1505         ip = VTOI(vp);
1506         if (vp->v_type == VNON || ip == NULL ||
1507             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1508              RB_EMPTY(&vp->v_rbdirty_tree))) {
1509                 return(-1);
1510         }
1511         return(0);
1512 }
1513
1514 static int
1515 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1516 {
1517         struct hammer_sync_info *info = data;
1518         struct hammer_inode *ip;
1519         int error;
1520
1521         ip = VTOI(vp);
1522         if (vp->v_type == VNON || vp->v_type == VBAD ||
1523             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1524              RB_EMPTY(&vp->v_rbdirty_tree))) {
1525                 return(0);
1526         }
1527         error = VOP_FSYNC(vp, MNT_NOWAIT);
1528         if (error)
1529                 info->error = error;
1530         return(0);
1531 }
1532