sys/vfs/hammer/hammer_ondisk.c

   1 /*
   2  * Copyright (c) 2007 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 1.33 2008/03/18 05:19:16 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);
  52
  53 /*
  54  * Red-Black tree support for various structures
  55  */
  56 static int
  57 hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
  58 {
  59         if (ip1->obj_id < ip2->obj_id)
  60                 return(-1);
  61         if (ip1->obj_id > ip2->obj_id)
  62                 return(1);
  63         if (ip1->obj_asof < ip2->obj_asof)
  64                 return(-1);
  65         if (ip1->obj_asof > ip2->obj_asof)
  66                 return(1);
  67         return(0);
  68 }
  69
  70 static int
  71 hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
  72 {
  73         if (info->obj_id < ip->obj_id)
  74                 return(-1);
  75         if (info->obj_id > ip->obj_id)
  76                 return(1);
  77         if (info->obj_asof < ip->obj_asof)
  78                 return(-1);
  79         if (info->obj_asof > ip->obj_asof)
  80                 return(1);
  81         return(0);
  82 }
  83
  84 static int
  85 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  86 {
  87         if (vol1->vol_no < vol2->vol_no)
  88                 return(-1);
  89         if (vol1->vol_no > vol2->vol_no)
  90                 return(1);
  91         return(0);
  92 }
  93
  94 static int
  95 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  96 {
  97         if (buf1->zone2_offset < buf2->zone2_offset)
  98                 return(-1);
  99         if (buf1->zone2_offset > buf2->zone2_offset)
 100                 return(1);
 101         return(0);
 102 }
 103
 104 static int
 105 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
 106 {
 107         if (node1->node_offset < node2->node_offset)
 108                 return(-1);
 109         if (node1->node_offset > node2->node_offset)
 110                 return(1);
 111         return(0);
 112 }
 113
 114 /*
 115  * Note: The lookup function for hammer_ino_rb_tree winds up being named
 116  * hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).  The other lookup
 117  * functions are normal, e.g. hammer_buf_rb_tree_RB_LOOKUP(root, zone2_offset).
 118  */
 119 RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
 120 RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 121                 hammer_inode_info_cmp, hammer_inode_info_t);
 122 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 123              hammer_vol_rb_compare, int32_t, vol_no);
 124 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 125              hammer_buf_rb_compare, hammer_off_t, zone2_offset);
 126 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
 127              hammer_nod_rb_compare, hammer_off_t, node_offset);
 128
 129 /************************************************************************
 130  *                              VOLUMES                                 *
 131  ************************************************************************
 132  *
 133  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 134  * code on failure.  Volumes must be loaded at mount time, get_volume() will
 135  * not load a new volume.
 136  *
 137  * Calls made to hammer_load_volume() or single-threaded
 138  */
 139 int
 140 hammer_install_volume(struct hammer_mount *hmp, const char *volname)
 141 {
 142         struct mount *mp;
 143         hammer_volume_t volume;
 144         struct hammer_volume_ondisk *ondisk;
 145         struct nlookupdata nd;
 146         struct buf *bp = NULL;
 147         int error;
 148         int ronly;
 149         int setmp = 0;
 150
 151         mp = hmp->mp;
 152         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 153
 154         /*
 155          * Allocate a volume structure
 156          */
 157         ++hammer_count_volumes;
 158         volume = kmalloc(sizeof(*volume), M_HAMMER, M_WAITOK|M_ZERO);
 159         volume->vol_name = kstrdup(volname, M_HAMMER);
 160         volume->hmp = hmp;
 161         hammer_io_init(&volume->io, HAMMER_STRUCTURE_VOLUME);
 162         volume->io.offset = 0LL;
 163
 164         /*
 165          * Get the device vnode
 166          */
 167         error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 168         if (error == 0)
 169                 error = nlookup(&nd);
 170         if (error == 0)
 171                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 172         nlookup_done(&nd);
 173         if (error == 0) {
 174                 if (vn_isdisk(volume->devvp, &error)) {
 175                         error = vfs_mountedon(volume->devvp);
 176                 }
 177         }
 178         if (error == 0 &&
 179             count_udev(volume->devvp->v_umajor, volume->devvp->v_uminor) > 0) {
 180                 error = EBUSY;
 181         }
 182         if (error == 0) {
 183                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 184                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 185                 if (error == 0) {
 186                         error = VOP_OPEN(volume->devvp,
 187                                          (ronly ? FREAD : FREAD|FWRITE),
 188                                          FSCRED, NULL);
 189                 }
 190                 vn_unlock(volume->devvp);
 191         }
 192         if (error) {
 193                 hammer_free_volume(volume);
 194                 return(error);
 195         }
 196         volume->devvp->v_rdev->si_mountpoint = mp;
 197         setmp = 1;
 198
 199         /*
 200          * Extract the volume number from the volume header and do various
 201          * sanity checks.
 202          */
 203         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 204         if (error)
 205                 goto late_failure;
 206         ondisk = (void *)bp->b_data;
 207         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 208                 kprintf("hammer_mount: volume %s has an invalid header\n",
 209                         volume->vol_name);
 210                 error = EFTYPE;
 211                 goto late_failure;
 212         }
 213         volume->vol_no = ondisk->vol_no;
 214         volume->buffer_base = ondisk->vol_buf_beg;
 215         volume->vol_flags = ondisk->vol_flags;
 216         volume->nblocks = ondisk->vol_nblocks;
 217         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 218                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 219         RB_INIT(&volume->rb_bufs_root);
 220
 221         hmp->mp->mnt_stat.f_blocks += volume->nblocks;
 222
 223         if (RB_EMPTY(&hmp->rb_vols_root)) {
 224                 hmp->fsid = ondisk->vol_fsid;
 225         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 226                 kprintf("hammer_mount: volume %s's fsid does not match "
 227                         "other volumes\n", volume->vol_name);
 228                 error = EFTYPE;
 229                 goto late_failure;
 230         }
 231
 232         /*
 233          * Insert the volume structure into the red-black tree.
 234          */
 235         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 236                 kprintf("hammer_mount: volume %s has a duplicate vol_no %d\n",
 237                         volume->vol_name, volume->vol_no);
 238                 error = EEXIST;
 239         }
 240
 241         /*
 242          * Set the root volume .  HAMMER special cases rootvol the structure.
 243          * We do not hold a ref because this would prevent related I/O
 244          * from being flushed.
 245          */
 246         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 247                 hmp->rootvol = volume;
 248                 if (bp) {
 249                         brelse(bp);
 250                         bp = NULL;
 251                 }
 252                 hmp->fsid_udev = dev2udev(vn_todev(volume->devvp));
 253         }
 254 late_failure:
 255         if (bp)
 256                 brelse(bp);
 257         if (error) {
 258                 /*vinvalbuf(volume->devvp, V_SAVE, 0, 0);*/
 259                 if (setmp)
 260                         volume->devvp->v_rdev->si_mountpoint = NULL;
 261                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE);
 262                 hammer_free_volume(volume);
 263         }
 264         return (error);
 265 }
 266
 267 /*
 268  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 269  * so returns -1 on failure.
 270  */
 271 int
 272 hammer_unload_volume(hammer_volume_t volume, void *data __unused)
 273 {
 274         struct hammer_mount *hmp = volume->hmp;
 275         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 276
 277         /*
 278          * Sync clusters, sync volume
 279          */
 280
 281         hmp->mp->mnt_stat.f_blocks -= volume->nblocks;
 282
 283         /*
 284          * Clean up the root volume pointer, which is held unlocked in hmp.
 285          */
 286         if (hmp->rootvol == volume)
 287                 hmp->rootvol = NULL;
 288
 289         /*
 290          * Unload buffers.
 291          */
 292         RB_SCAN(hammer_buf_rb_tree, &volume->rb_bufs_root, NULL,
 293                         hammer_unload_buffer, NULL);
 294         hammer_io_waitdep(&volume->io);
 295
 296         /*
 297          * Release our buffer and flush anything left in the buffer cache.
 298          */
 299         hammer_io_release(&volume->io, 2);
 300
 301         /*
 302          * There should be no references on the volume, no clusters, and
 303          * no super-clusters.
 304          */
 305         KKASSERT(volume->io.lock.refs == 0);
 306         KKASSERT(RB_EMPTY(&volume->rb_bufs_root));
 307
 308         volume->ondisk = NULL;
 309         if (volume->devvp) {
 310                 if (volume->devvp->v_rdev &&
 311                     volume->devvp->v_rdev->si_mountpoint == hmp->mp
 312                 ) {
 313                         volume->devvp->v_rdev->si_mountpoint = NULL;
 314                 }
 315                 if (ronly) {
 316                         vinvalbuf(volume->devvp, 0, 0, 0);
 317                         VOP_CLOSE(volume->devvp, FREAD);
 318                 } else {
 319                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 320                         VOP_CLOSE(volume->devvp, FREAD|FWRITE);
 321                 }
 322         }
 323
 324         /*
 325          * Destroy the structure
 326          */
 327         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 328         hammer_free_volume(volume);
 329         return(0);
 330 }
 331
 332 static
 333 void
 334 hammer_free_volume(hammer_volume_t volume)
 335 {
 336         if (volume->vol_name) {
 337                 kfree(volume->vol_name, M_HAMMER);
 338                 volume->vol_name = NULL;
 339         }
 340         if (volume->devvp) {
 341                 vrele(volume->devvp);
 342                 volume->devvp = NULL;
 343         }
 344         --hammer_count_volumes;
 345         kfree(volume, M_HAMMER);
 346 }
 347
 348 /*
 349  * Get a HAMMER volume.  The volume must already exist.
 350  */
 351 hammer_volume_t
 352 hammer_get_volume(struct hammer_mount *hmp, int32_t vol_no, int *errorp)
 353 {
 354         struct hammer_volume *volume;
 355
 356         /*
 357          * Locate the volume structure
 358          */
 359         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 360         if (volume == NULL) {
 361                 *errorp = ENOENT;
 362                 return(NULL);
 363         }
 364         hammer_ref(&volume->io.lock);
 365
 366         /*
 367          * Deal with on-disk info
 368          */
 369         if (volume->ondisk == NULL || volume->io.loading) {
 370                 *errorp = hammer_load_volume(volume);
 371                 if (*errorp) {
 372                         hammer_rel_volume(volume, 1);
 373                         volume = NULL;
 374                 }
 375         } else {
 376                 *errorp = 0;
 377         }
 378         return(volume);
 379 }
 380
 381 int
 382 hammer_ref_volume(hammer_volume_t volume)
 383 {
 384         int error;
 385
 386         hammer_ref(&volume->io.lock);
 387
 388         /*
 389          * Deal with on-disk info
 390          */
 391         if (volume->ondisk == NULL || volume->io.loading) {
 392                 error = hammer_load_volume(volume);
 393                 if (error)
 394                         hammer_rel_volume(volume, 1);
 395         } else {
 396                 error = 0;
 397         }
 398         return (error);
 399 }
 400
 401 hammer_volume_t
 402 hammer_get_root_volume(struct hammer_mount *hmp, int *errorp)
 403 {
 404         hammer_volume_t volume;
 405
 406         volume = hmp->rootvol;
 407         KKASSERT(volume != NULL);
 408         hammer_ref(&volume->io.lock);
 409
 410         /*
 411          * Deal with on-disk info
 412          */
 413         if (volume->ondisk == NULL || volume->io.loading) {
 414                 *errorp = hammer_load_volume(volume);
 415                 if (*errorp) {
 416                         hammer_rel_volume(volume, 1);
 417                         volume = NULL;
 418                 }
 419         } else {
 420                 *errorp = 0;
 421         }
 422         return (volume);
 423 }
 424
 425 /*
 426  * Load a volume's on-disk information.  The volume must be referenced and
 427  * not locked.  We temporarily acquire an exclusive lock to interlock
 428  * against releases or multiple get's.
 429  */
 430 static int
 431 hammer_load_volume(hammer_volume_t volume)
 432 {
 433         struct hammer_volume_ondisk *ondisk;
 434         int error;
 435
 436         hammer_lock_ex(&volume->io.lock);
 437         KKASSERT(volume->io.loading == 0);
 438         volume->io.loading = 1;
 439
 440         if (volume->ondisk == NULL) {
 441                 error = hammer_io_read(volume->devvp, &volume->io);
 442                 if (error) {
 443                         volume->io.loading = 0;
 444                         hammer_unlock(&volume->io.lock);
 445                         return (error);
 446                 }
 447                 volume->ondisk = ondisk = (void *)volume->io.bp->b_data;
 448         } else {
 449                 error = 0;
 450         }
 451         volume->io.loading = 0;
 452         hammer_unlock(&volume->io.lock);
 453         return(0);
 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         if (volume->io.lock.refs == 1) {
 468                 hammer_lock_ex(&volume->io.lock);
 469                 if (volume->io.lock.refs == 1) {
 470                         volume->ondisk = NULL;
 471                         hammer_io_release(&volume->io, flush);
 472                 } else if (flush) {
 473                         hammer_io_flush(&volume->io);
 474                 }
 475                 hammer_unlock(&volume->io.lock);
 476         }
 477         hammer_unref(&volume->io.lock);
 478 }
 479
 480 /************************************************************************
 481  *                              BUFFERS                                 *
 482  ************************************************************************
 483  *
 484  * Manage buffers.  Currently all blockmap-backed zones are translated
 485  * to zone-2 buffer offsets.
 486  */
 487 hammer_buffer_t
 488 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 489                   int isnew, int *errorp)
 490 {
 491         hammer_buffer_t buffer;
 492         hammer_volume_t volume;
 493         hammer_off_t    zoneX_offset;
 494         int vol_no;
 495         int zone;
 496
 497         zoneX_offset = buf_offset;
 498         zone = HAMMER_ZONE_DECODE(buf_offset);
 499         if (zone >= HAMMER_ZONE_BTREE_INDEX) {
 500                 buf_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 501                 KKASSERT(*errorp == 0);
 502         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 503                 buf_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 504                 KKASSERT(*errorp == 0);
 505         }
 506         buf_offset &= ~HAMMER_BUFMASK64;
 507         KKASSERT((buf_offset & HAMMER_ZONE_RAW_BUFFER) ==
 508                  HAMMER_ZONE_RAW_BUFFER);
 509         vol_no = HAMMER_VOL_DECODE(buf_offset);
 510         volume = hammer_get_volume(hmp, vol_no, errorp);
 511         if (volume == NULL)
 512                 return(NULL);
 513
 514         /*
 515          * NOTE: buf_offset and maxbuf_off are both full offset
 516          * specifications.
 517          */
 518         KKASSERT(buf_offset < volume->maxbuf_off);
 519
 520         /*
 521          * Locate and lock the buffer structure, creating one if necessary.
 522          */
 523 again:
 524         buffer = RB_LOOKUP(hammer_buf_rb_tree, &volume->rb_bufs_root,
 525                            buf_offset);
 526         if (buffer == NULL) {
 527                 ++hammer_count_buffers;
 528                 buffer = kmalloc(sizeof(*buffer), M_HAMMER, M_WAITOK|M_ZERO);
 529                 buffer->zone2_offset = buf_offset;
 530                 buffer->volume = volume;
 531                 hammer_io_init(&buffer->io, HAMMER_STRUCTURE_BUFFER);
 532                 buffer->io.offset = volume->ondisk->vol_buf_beg +
 533                                     (buf_offset & HAMMER_OFF_SHORT_MASK);
 534                 TAILQ_INIT(&buffer->clist);
 535                 hammer_ref(&buffer->io.lock);
 536
 537                 /*
 538                  * Insert the buffer into the RB tree and handle late
 539                  * collisions.
 540                  */
 541                 if (RB_INSERT(hammer_buf_rb_tree, &volume->rb_bufs_root, buffer)) {
 542                         hammer_unref(&buffer->io.lock);
 543                         --hammer_count_buffers;
 544                         kfree(buffer, M_HAMMER);
 545                         goto again;
 546                 }
 547                 hammer_ref(&volume->io.lock);
 548         } else {
 549                 hammer_ref(&buffer->io.lock);
 550         }
 551
 552         /*
 553          * Cache the blockmap translation
 554          */
 555         if ((zoneX_offset & HAMMER_ZONE_RAW_BUFFER) != HAMMER_ZONE_RAW_BUFFER)
 556                 buffer->zoneX_offset = zoneX_offset;
 557
 558         /*
 559          * Deal with on-disk info
 560          */
 561         if (buffer->ondisk == NULL || buffer->io.loading) {
 562                 *errorp = hammer_load_buffer(buffer, isnew);
 563                 if (*errorp) {
 564                         hammer_rel_buffer(buffer, 1);
 565                         buffer = NULL;
 566                 }
 567         } else {
 568                 *errorp = 0;
 569         }
 570         hammer_rel_volume(volume, 0);
 571         return(buffer);
 572 }
 573
 574 static int
 575 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 576 {
 577         hammer_volume_t volume;
 578         void *ondisk;
 579         int error;
 580
 581         /*
 582          * Load the buffer's on-disk info
 583          */
 584         volume = buffer->volume;
 585         hammer_lock_ex(&buffer->io.lock);
 586         KKASSERT(buffer->io.loading == 0);
 587         buffer->io.loading = 1;
 588
 589         if (buffer->ondisk == NULL) {
 590                 if (isnew) {
 591                         error = hammer_io_new(volume->devvp, &buffer->io);
 592                 } else {
 593                         error = hammer_io_read(volume->devvp, &buffer->io);
 594                 }
 595                 if (error) {
 596                         buffer->io.loading = 0;
 597                         hammer_unlock(&buffer->io.lock);
 598                         return (error);
 599                 }
 600                 buffer->ondisk = ondisk = (void *)buffer->io.bp->b_data;
 601         } else if (isnew) {
 602                 error = hammer_io_new(volume->devvp, &buffer->io);
 603         } else {
 604                 error = 0;
 605         }
 606         if (error == 0 && isnew) {
 607                 hammer_modify_buffer(buffer, NULL, 0);
 608                 /* additional initialization goes here */
 609         }
 610         buffer->io.loading = 0;
 611         hammer_unlock(&buffer->io.lock);
 612         return (error);
 613 }
 614
 615 /*
 616  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 617  */
 618 int
 619 hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused)
 620 {
 621         hammer_ref(&buffer->io.lock);
 622         hammer_flush_buffer_nodes(buffer);
 623         KKASSERT(buffer->io.lock.refs == 1);
 624         hammer_rel_buffer(buffer, 2);
 625         return(0);
 626 }
 627
 628 /*
 629  * Reference a buffer that is either already referenced or via a specially
 630  * handled pointer (aka cursor->buffer).
 631  */
 632 int
 633 hammer_ref_buffer(hammer_buffer_t buffer)
 634 {
 635         int error;
 636
 637         hammer_ref(&buffer->io.lock);
 638         if (buffer->ondisk == NULL || buffer->io.loading) {
 639                 error = hammer_load_buffer(buffer, 0);
 640                 if (error) {
 641                         hammer_rel_buffer(buffer, 1);
 642                         /*
 643                          * NOTE: buffer pointer can become stale after
 644                          * the above release.
 645                          */
 646                 }
 647         } else {
 648                 error = 0;
 649         }
 650         return(error);
 651 }
 652
 653 /*
 654  * Release a buffer.  We have to deal with several places where
 655  * another thread can ref the buffer.
 656  *
 657  * Only destroy the structure itself if the related buffer cache buffer
 658  * was disassociated from it.  This ties the management of the structure
 659  * to the buffer cache subsystem.  buffer->ondisk determines whether the
 660  * embedded io is referenced or not.
 661  */
 662 void
 663 hammer_rel_buffer(hammer_buffer_t buffer, int flush)
 664 {
 665         hammer_volume_t volume;
 666
 667         if (buffer->io.lock.refs == 1) {
 668                 hammer_lock_ex(&buffer->io.lock);
 669                 if (buffer->io.lock.refs == 1) {
 670                         hammer_io_release(&buffer->io, flush);
 671
 672                         if (buffer->io.bp == NULL &&
 673                             buffer->io.lock.refs == 1) {
 674                                 hammer_flush_buffer_nodes(buffer);
 675                                 KKASSERT(TAILQ_EMPTY(&buffer->clist));
 676                                 volume = buffer->volume;
 677                                 RB_REMOVE(hammer_buf_rb_tree,
 678                                           &volume->rb_bufs_root, buffer);
 679                                 buffer->volume = NULL; /* sanity */
 680                                 --hammer_count_buffers;
 681                                 kfree(buffer, M_HAMMER);
 682                                 hammer_rel_volume(volume, 0);
 683                                 return;
 684                         }
 685                 } else if (flush) {
 686                         hammer_io_flush(&buffer->io);
 687                 }
 688                 hammer_unlock(&buffer->io.lock);
 689         }
 690         hammer_unref(&buffer->io.lock);
 691 }
 692
 693 /*
 694  * Access the filesystem buffer containing the specified hammer offset.
 695  * buf_offset is a conglomeration of the volume number and vol_buf_beg
 696  * relative buffer offset.  It must also have bit 55 set to be valid.
 697  * (see hammer_off_t in hammer_disk.h).
 698  *
 699  * Any prior buffer in *bufferp will be released and replaced by the
 700  * requested buffer.
 701  */
 702 void *
 703 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 704              struct hammer_buffer **bufferp)
 705 {
 706         hammer_buffer_t buffer;
 707         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 708
 709         buf_offset &= ~HAMMER_BUFMASK64;
 710         KKASSERT((buf_offset & HAMMER_OFF_ZONE_MASK) != 0);
 711
 712         buffer = *bufferp;
 713         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 714                                buffer->zoneX_offset != buf_offset)) {
 715                 if (buffer)
 716                         hammer_rel_buffer(buffer, 0);
 717                 buffer = hammer_get_buffer(hmp, buf_offset, 0, errorp);
 718                 *bufferp = buffer;
 719         } else {
 720                 *errorp = 0;
 721         }
 722
 723         /*
 724          * Return a pointer to the buffer data.
 725          */
 726         if (buffer == NULL)
 727                 return(NULL);
 728         else
 729                 return((char *)buffer->ondisk + xoff);
 730 }
 731
 732 /*
 733  * Access the filesystem buffer containing the specified hammer offset.
 734  * No disk read operation occurs.  The result buffer may contain garbage.
 735  *
 736  * Any prior buffer in *bufferp will be released and replaced by the
 737  * requested buffer.
 738  */
 739 void *
 740 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset, int *errorp,
 741              struct hammer_buffer **bufferp)
 742 {
 743         hammer_buffer_t buffer;
 744         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
 745
 746         buf_offset &= ~HAMMER_BUFMASK64;
 747
 748         buffer = *bufferp;
 749         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
 750                                buffer->zoneX_offset != buf_offset)) {
 751                 if (buffer)
 752                         hammer_rel_buffer(buffer, 0);
 753                 buffer = hammer_get_buffer(hmp, buf_offset, 1, errorp);
 754                 *bufferp = buffer;
 755         } else {
 756                 *errorp = 0;
 757         }
 758
 759         /*
 760          * Return a pointer to the buffer data.
 761          */
 762         if (buffer == NULL)
 763                 return(NULL);
 764         else
 765                 return((char *)buffer->ondisk + xoff);
 766 }
 767
 768 /************************************************************************
 769  *                              NODES                                   *
 770  ************************************************************************
 771  *
 772  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
 773  * method used by the HAMMER filesystem.
 774  *
 775  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
 776  * associated with its buffer, and will only referenced the buffer while
 777  * the node itself is referenced.
 778  *
 779  * A hammer_node can also be passively associated with other HAMMER
 780  * structures, such as inodes, while retaining 0 references.  These
 781  * associations can be cleared backwards using a pointer-to-pointer in
 782  * the hammer_node.
 783  *
 784  * This allows the HAMMER implementation to cache hammer_nodes long-term
 785  * and short-cut a great deal of the infrastructure's complexity.  In
 786  * most cases a cached node can be reacquired without having to dip into
 787  * either the buffer or cluster management code.
 788  *
 789  * The caller must pass a referenced cluster on call and will retain
 790  * ownership of the reference on return.  The node will acquire its own
 791  * additional references, if necessary.
 792  */
 793 hammer_node_t
 794 hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset, int *errorp)
 795 {
 796         hammer_node_t node;
 797
 798         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
 799
 800         /*
 801          * Locate the structure, allocating one if necessary.
 802          */
 803 again:
 804         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
 805         if (node == NULL) {
 806                 ++hammer_count_nodes;
 807                 node = kmalloc(sizeof(*node), M_HAMMER, M_WAITOK|M_ZERO);
 808                 node->node_offset = node_offset;
 809                 node->hmp = hmp;
 810                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
 811                         --hammer_count_nodes;
 812                         kfree(node, M_HAMMER);
 813                         goto again;
 814                 }
 815         }
 816         hammer_ref(&node->lock);
 817         *errorp = hammer_load_node(node);
 818         if (*errorp) {
 819                 hammer_rel_node(node);
 820                 node = NULL;
 821         }
 822         return(node);
 823 }
 824
 825 /*
 826  * Reference an already-referenced node.
 827  */
 828 int
 829 hammer_ref_node(hammer_node_t node)
 830 {
 831         int error;
 832
 833         KKASSERT(node->lock.refs > 0);
 834         hammer_ref(&node->lock);
 835         if ((error = hammer_load_node(node)) != 0)
 836                 hammer_rel_node(node);
 837         return(error);
 838 }
 839
 840 /*
 841  * Load a node's on-disk data reference.
 842  */
 843 static int
 844 hammer_load_node(hammer_node_t node)
 845 {
 846         hammer_buffer_t buffer;
 847         int error;
 848
 849         if (node->ondisk)
 850                 return(0);
 851         error = 0;
 852         hammer_lock_ex(&node->lock);
 853         if (node->ondisk == NULL) {
 854                 /*
 855                  * This is a little confusing but the jist is that
 856                  * node->buffer determines whether the node is on
 857                  * the buffer's clist and node->ondisk determines
 858                  * whether the buffer is referenced.
 859                  */
 860                 if ((buffer = node->buffer) != NULL) {
 861                         error = hammer_ref_buffer(buffer);
 862                 } else {
 863                         buffer = hammer_get_buffer(node->hmp,
 864                                                    node->node_offset, 0,
 865                                                    &error);
 866                         if (buffer) {
 867                                 KKASSERT(error == 0);
 868                                 TAILQ_INSERT_TAIL(&buffer->clist,
 869                                                   node, entry);
 870                                 node->buffer = buffer;
 871                         }
 872                 }
 873                 if (error == 0) {
 874                         node->ondisk = (void *)((char *)buffer->ondisk +
 875                                (node->node_offset & HAMMER_BUFMASK));
 876                 }
 877         }
 878         hammer_unlock(&node->lock);
 879         return (error);
 880 }
 881
 882 /*
 883  * Safely reference a node, interlock against flushes via the IO subsystem.
 884  */
 885 hammer_node_t
 886 hammer_ref_node_safe(struct hammer_mount *hmp, struct hammer_node **cache,
 887                      int *errorp)
 888 {
 889         hammer_node_t node;
 890
 891         if ((node = *cache) != NULL)
 892                 hammer_ref(&node->lock);
 893         if (node) {
 894                 *errorp = hammer_load_node(node);
 895                 if (*errorp) {
 896                         hammer_rel_node(node);
 897                         node = NULL;
 898                 }
 899         } else {
 900                 *errorp = ENOENT;
 901         }
 902         return(node);
 903 }
 904
 905 /*
 906  * Release a hammer_node.  On the last release the node dereferences
 907  * its underlying buffer and may or may not be destroyed.
 908  */
 909 void
 910 hammer_rel_node(hammer_node_t node)
 911 {
 912         hammer_buffer_t buffer;
 913
 914         /*
 915          * If this isn't the last ref just decrement the ref count and
 916          * return.
 917          */
 918         if (node->lock.refs > 1) {
 919                 hammer_unref(&node->lock);
 920                 return;
 921         }
 922
 923         /*
 924          * If there is no ondisk info or no buffer the node failed to load,
 925          * remove the last reference and destroy the node.
 926          */
 927         if (node->ondisk == NULL) {
 928                 hammer_unref(&node->lock);
 929                 hammer_flush_node(node);
 930                 /* node is stale now */
 931                 return;
 932         }
 933
 934         /*
 935          * Do final cleanups and then either destroy the node and leave it
 936          * passively cached.  The buffer reference is removed regardless.
 937          */
 938         buffer = node->buffer;
 939         node->ondisk = NULL;
 940
 941         if ((node->flags & (HAMMER_NODE_DELETED|HAMMER_NODE_FLUSH)) == 0) {
 942                 hammer_unref(&node->lock);
 943                 hammer_rel_buffer(buffer, 0);
 944                 return;
 945         }
 946
 947         /*
 948          * Destroy the node if it has been marked for deletion.  We mark
 949          * it as being free.  Note that the disk space is physically
 950          * freed when the fifo cycles back through the node.
 951          */
 952         if (node->flags & HAMMER_NODE_DELETED) {
 953                 hammer_blockmap_free(node->hmp, node->node_offset,
 954                                      sizeof(*node->ondisk));
 955         }
 956
 957         /*
 958          * Destroy the node.  Record pertainant data because the node
 959          * becomes stale the instant we flush it.
 960          */
 961         hammer_unref(&node->lock);
 962         hammer_flush_node(node);
 963         /* node is stale */
 964         hammer_rel_buffer(buffer, 0);
 965 }
 966
 967 /*
 968  * Passively cache a referenced hammer_node in *cache.  The caller may
 969  * release the node on return.
 970  */
 971 void
 972 hammer_cache_node(hammer_node_t node, struct hammer_node **cache)
 973 {
 974         hammer_node_t old;
 975
 976         /*
 977          * If the node is being deleted, don't cache it!
 978          */
 979         if (node->flags & HAMMER_NODE_DELETED)
 980                 return;
 981
 982         /*
 983          * Cache the node.  If we previously cached a different node we
 984          * have to give HAMMER a chance to destroy it.
 985          */
 986 again:
 987         if (node->cache1 != cache) {
 988                 if (node->cache2 != cache) {
 989                         if ((old = *cache) != NULL) {
 990                                 KKASSERT(node->lock.refs != 0);
 991                                 hammer_uncache_node(cache);
 992                                 goto again;
 993                         }
 994                         if (node->cache2)
 995                                 *node->cache2 = NULL;
 996                         node->cache2 = node->cache1;
 997                         node->cache1 = cache;
 998                         *cache = node;
 999                 } else {
1000                         struct hammer_node **tmp;
1001                         tmp = node->cache1;
1002                         node->cache1 = node->cache2;
1003                         node->cache2 = tmp;
1004                 }
1005         }
1006 }
1007
1008 void
1009 hammer_uncache_node(struct hammer_node **cache)
1010 {
1011         hammer_node_t node;
1012
1013         if ((node = *cache) != NULL) {
1014                 *cache = NULL;
1015                 if (node->cache1 == cache) {
1016                         node->cache1 = node->cache2;
1017                         node->cache2 = NULL;
1018                 } else if (node->cache2 == cache) {
1019                         node->cache2 = NULL;
1020                 } else {
1021                         panic("hammer_uncache_node: missing cache linkage");
1022                 }
1023                 if (node->cache1 == NULL && node->cache2 == NULL)
1024                         hammer_flush_node(node);
1025         }
1026 }
1027
1028 /*
1029  * Remove a node's cache references and destroy the node if it has no
1030  * other references or backing store.
1031  */
1032 void
1033 hammer_flush_node(hammer_node_t node)
1034 {
1035         hammer_buffer_t buffer;
1036
1037         if (node->cache1)
1038                 *node->cache1 = NULL;
1039         if (node->cache2)
1040                 *node->cache2 = NULL;
1041         if (node->lock.refs == 0 && node->ondisk == NULL) {
1042                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1043                 if ((buffer = node->buffer) != NULL) {
1044                         node->buffer = NULL;
1045                         TAILQ_REMOVE(&buffer->clist, node, entry);
1046                         /* buffer is unreferenced because ondisk is NULL */
1047                 }
1048                 --hammer_count_nodes;
1049                 kfree(node, M_HAMMER);
1050         }
1051 }
1052
1053 /*
1054  * Flush passively cached B-Tree nodes associated with this buffer.
1055  * This is only called when the buffer is about to be destroyed, so
1056  * none of the nodes should have any references.
1057  */
1058 void
1059 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1060 {
1061         hammer_node_t node;
1062
1063         while ((node = TAILQ_FIRST(&buffer->clist)) != NULL) {
1064                 KKASSERT(node->lock.refs == 0 && node->ondisk == NULL);
1065                 hammer_ref(&node->lock);
1066                 node->flags |= HAMMER_NODE_FLUSH;
1067                 hammer_rel_node(node);
1068         }
1069 }
1070
1071
1072 /************************************************************************
1073  *                              ALLOCATORS                              *
1074  ************************************************************************/
1075
1076 /*
1077  * Allocate a B-Tree node.
1078  */
1079 hammer_node_t
1080 hammer_alloc_btree(hammer_mount_t hmp, int *errorp)
1081 {
1082         hammer_buffer_t buffer = NULL;
1083         hammer_node_t node = NULL;
1084         hammer_off_t node_offset;
1085
1086         node_offset = hammer_blockmap_alloc(hmp, HAMMER_ZONE_BTREE_INDEX,
1087                                             sizeof(struct hammer_node_ondisk),
1088                                             errorp);
1089         if (*errorp == 0) {
1090                 node = hammer_get_node(hmp, node_offset, errorp);
1091                 hammer_modify_node_noundo(node);
1092                 bzero(node->ondisk, sizeof(*node->ondisk));
1093         }
1094         if (buffer)
1095                 hammer_rel_buffer(buffer, 0);
1096         return(node);
1097 }
1098
1099 /*
1100  * The returned buffers are already appropriately marked as being modified.
1101  * If the caller marks them again unnecessary undo records may be generated.
1102  *
1103  * In-band data is indicated by data_bufferp == NULL.  Pass a data_len of 0
1104  * for zero-fill (caller modifies data_len afterwords).
1105  */
1106 void *
1107 hammer_alloc_record(hammer_mount_t hmp,
1108                     hammer_off_t *rec_offp, u_int16_t rec_type,
1109                     struct hammer_buffer **rec_bufferp,
1110                     int32_t data_len, void **datap,
1111                     struct hammer_buffer **data_bufferp, int *errorp)
1112 {
1113         hammer_record_ondisk_t rec;
1114         hammer_off_t rec_offset;
1115         hammer_off_t data_offset;
1116         int32_t reclen;
1117
1118         if (datap)
1119                 *datap = NULL;
1120
1121         /*
1122          * Allocate the record
1123          */
1124         rec_offset = hammer_blockmap_alloc(hmp, HAMMER_ZONE_RECORD_INDEX,
1125                                            HAMMER_RECORD_SIZE, errorp);
1126         if (*errorp)
1127                 return(NULL);
1128
1129         /*
1130          * Allocate data
1131          */
1132         if (data_len) {
1133                 if (data_bufferp == NULL) {
1134                         switch(rec_type) {
1135                         case HAMMER_RECTYPE_DATA:
1136                                 reclen = offsetof(struct hammer_data_record,
1137                                                   data[0]);
1138                                 break;
1139                         case HAMMER_RECTYPE_DIRENTRY:
1140                                 reclen = offsetof(struct hammer_entry_record,
1141                                                   name[0]);
1142                                 break;
1143                         default:
1144                                 panic("hammer_alloc_record: illegal "
1145                                       "in-band data");
1146                                 /* NOT REACHED */
1147                                 reclen = 0;
1148                                 break;
1149                         }
1150                         KKASSERT(reclen + data_len <= HAMMER_RECORD_SIZE);
1151                         data_offset = rec_offset + reclen;
1152                 } else if (data_len < HAMMER_BUFSIZE) {
1153                         data_offset = hammer_blockmap_alloc(hmp,
1154                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1155                                                 data_len, errorp);
1156                 } else {
1157                         data_offset = hammer_blockmap_alloc(hmp,
1158                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1159                                                 data_len, errorp);
1160                 }
1161         } else {
1162                 data_offset = 0;
1163         }
1164         if (*errorp) {
1165                 hammer_blockmap_free(hmp, rec_offset, HAMMER_RECORD_SIZE);
1166                 return(NULL);
1167         }
1168
1169         /*
1170          * Basic return values.
1171          */
1172         *rec_offp = rec_offset;
1173         rec = hammer_bread(hmp, rec_offset, errorp, rec_bufferp);
1174         hammer_modify_buffer(*rec_bufferp, NULL, 0);
1175         bzero(rec, sizeof(*rec));
1176         KKASSERT(*errorp == 0);
1177         rec->base.data_off = data_offset;
1178         rec->base.data_len = data_len;
1179
1180         if (data_bufferp) {
1181                 if (data_len) {
1182                         *datap = hammer_bread(hmp, data_offset, errorp,
1183                                               data_bufferp);
1184                         KKASSERT(*errorp == 0);
1185                         hammer_modify_buffer(*data_bufferp, NULL, 0);
1186                 } else {
1187                         *datap = NULL;
1188                 }
1189         } else if (data_len) {
1190                 KKASSERT(data_offset + data_len - rec_offset <=
1191                          HAMMER_RECORD_SIZE);
1192                 if (datap) {
1193                         *datap = (void *)((char *)rec +
1194                                           (int32_t)(data_offset - rec_offset));
1195                 }
1196         } else {
1197                 KKASSERT(datap == NULL);
1198         }
1199         KKASSERT(*errorp == 0);
1200         return(rec);
1201 }
1202
1203 void *
1204 hammer_alloc_data(hammer_mount_t hmp, int32_t data_len,
1205                   hammer_off_t *data_offsetp,
1206                   struct hammer_buffer **data_bufferp, int *errorp)
1207 {
1208         void *data;
1209
1210         /*
1211          * Allocate data
1212          */
1213         if (data_len) {
1214                 if (data_len < HAMMER_BUFSIZE) {
1215                         *data_offsetp = hammer_blockmap_alloc(hmp,
1216                                                 HAMMER_ZONE_SMALL_DATA_INDEX,
1217                                                 data_len, errorp);
1218                 } else {
1219                         *data_offsetp = hammer_blockmap_alloc(hmp,
1220                                                 HAMMER_ZONE_LARGE_DATA_INDEX,
1221                                                 data_len, errorp);
1222                 }
1223         } else {
1224                 *data_offsetp = 0;
1225         }
1226         if (*errorp == 0 && data_bufferp) {
1227                 if (data_len) {
1228                         data = hammer_bread(hmp, *data_offsetp, errorp,
1229                                             data_bufferp);
1230                         KKASSERT(*errorp == 0);
1231                         hammer_modify_buffer(*data_bufferp, NULL, 0);
1232                 } else {
1233                         data = NULL;
1234                 }
1235         } else {
1236                 data = NULL;
1237         }
1238         KKASSERT(*errorp == 0);
1239         return(data);
1240 }
1241
1242
1243 /*
1244  * Sync dirty buffers to the media
1245  */
1246
1247 static int hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data);
1248 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1249
1250 int
1251 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1252 {
1253         struct hammer_sync_info info;
1254
1255         info.error = 0;
1256         info.waitfor = waitfor;
1257
1258         vmntvnodescan(hmp->mp, VMSC_GETVP|VMSC_NOWAIT,
1259                       hammer_sync_scan1, hammer_sync_scan2, &info);
1260
1261         RB_SCAN(hammer_vol_rb_tree, &hmp->rb_vols_root, NULL,
1262                 hammer_sync_volume, &info);
1263         return(info.error);
1264 }
1265
1266 static int
1267 hammer_sync_scan1(struct mount *mp, struct vnode *vp, void *data)
1268 {
1269         struct hammer_inode *ip;
1270
1271         ip = VTOI(vp);
1272         if (vp->v_type == VNON || ip == NULL ||
1273             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1274              RB_EMPTY(&vp->v_rbdirty_tree))) {
1275                 return(-1);
1276         }
1277         return(0);
1278 }
1279
1280 static int
1281 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1282 {
1283         struct hammer_sync_info *info = data;
1284         struct hammer_inode *ip;
1285         int error;
1286
1287         ip = VTOI(vp);
1288         if (vp->v_type == VNON || vp->v_type == VBAD ||
1289             ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1290              RB_EMPTY(&vp->v_rbdirty_tree))) {
1291                 return(0);
1292         }
1293         error = VOP_FSYNC(vp, info->waitfor);
1294         if (error)
1295                 info->error = error;
1296         return(0);
1297 }
1298
1299 int
1300 hammer_sync_volume(hammer_volume_t volume, void *data)
1301 {
1302         struct hammer_sync_info *info = data;
1303
1304         hammer_ref(&volume->io.lock);
1305         RB_SCAN(hammer_buf_rb_tree, &volume->rb_bufs_root, NULL,
1306                 hammer_sync_buffer, info);
1307         hammer_rel_volume(volume, 1);
1308         return(0);
1309 }
1310
1311 int
1312 hammer_sync_buffer(hammer_buffer_t buffer, void *data __unused)
1313 {
1314         hammer_ref(&buffer->io.lock);
1315         hammer_rel_buffer(buffer, 1);
1316         return(0);
1317 }
1318