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 <sys/nlookup.h>
  43 #include <sys/buf2.h>
  44
  45 #include "hammer.h"
  46
  47 static void hammer_free_volume(hammer_volume_t volume);
  48 static int hammer_load_volume(hammer_volume_t volume);
  49 static int hammer_load_buffer(hammer_buffer_t buffer, int isnew);
  50 static int hammer_load_node(hammer_transaction_t trans,
  51                                 hammer_node_t node, int isnew);
  52 static void _hammer_rel_node(hammer_node_t node, int locked);
  53
  54 static int
  55 hammer_vol_rb_compare(hammer_volume_t vol1, hammer_volume_t vol2)
  56 {
  57         if (vol1->vol_no < vol2->vol_no)
  58                 return(-1);
  59         if (vol1->vol_no > vol2->vol_no)
  60                 return(1);
  61         return(0);
  62 }
  63
  64 /*
  65  * hammer_buffer structures are indexed via their zoneX_offset, not
  66  * their zone2_offset.
  67  */
  68 static int
  69 hammer_buf_rb_compare(hammer_buffer_t buf1, hammer_buffer_t buf2)
  70 {
  71         if (buf1->zoneX_offset < buf2->zoneX_offset)
  72                 return(-1);
  73         if (buf1->zoneX_offset > buf2->zoneX_offset)
  74                 return(1);
  75         return(0);
  76 }
  77
  78 static int
  79 hammer_nod_rb_compare(hammer_node_t node1, hammer_node_t node2)
  80 {
  81         if (node1->node_offset < node2->node_offset)
  82                 return(-1);
  83         if (node1->node_offset > node2->node_offset)
  84                 return(1);
  85         return(0);
  86 }
  87
  88 RB_GENERATE2(hammer_vol_rb_tree, hammer_volume, rb_node,
  89              hammer_vol_rb_compare, int32_t, vol_no);
  90 RB_GENERATE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
  91              hammer_buf_rb_compare, hammer_off_t, zoneX_offset);
  92 RB_GENERATE2(hammer_nod_rb_tree, hammer_node, rb_node,
  93              hammer_nod_rb_compare, hammer_off_t, node_offset);
  94
  95 /************************************************************************
  96  *                              VOLUMES                                 *
  97  ************************************************************************
  98  *
  99  * Load a HAMMER volume by name.  Returns 0 on success or a positive error
 100  * code on failure.  Volumes must be loaded at mount time or via hammer
 101  * volume-add command, hammer_get_volume() will not load a new volume.
 102  *
 103  * The passed devvp is vref()'d but not locked.  This function consumes the
 104  * ref (typically by associating it with the volume structure).
 105  *
 106  * Calls made to hammer_load_volume() or single-threaded
 107  */
 108 int
 109 hammer_install_volume(hammer_mount_t hmp, const char *volname,
 110                       struct vnode *devvp, void *data)
 111 {
 112         struct mount *mp;
 113         hammer_volume_t volume;
 114         struct hammer_volume_ondisk *ondisk;
 115         struct hammer_volume_ondisk *img;
 116         struct nlookupdata nd;
 117         struct buf *bp = NULL;
 118         int error;
 119         int ronly;
 120         int setmp = 0;
 121         int i;
 122
 123         mp = hmp->mp;
 124         ronly = ((mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 125
 126         /*
 127          * Allocate a volume structure
 128          */
 129         ++hammer_count_volumes;
 130         volume = kmalloc(sizeof(*volume), hmp->m_misc, M_WAITOK|M_ZERO);
 131         volume->vol_name = kstrdup(volname, hmp->m_misc);
 132         volume->io.hmp = hmp;   /* bootstrap */
 133         hammer_io_init(&volume->io, volume, HAMMER_STRUCTURE_VOLUME);
 134         volume->io.offset = 0LL;
 135         volume->io.bytes = HAMMER_BUFSIZE;
 136
 137         /*
 138          * Get the device vnode
 139          */
 140         if (devvp == NULL) {
 141                 error = nlookup_init(&nd, volume->vol_name, UIO_SYSSPACE, NLC_FOLLOW);
 142                 if (error == 0)
 143                         error = nlookup(&nd);
 144                 if (error == 0)
 145                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &volume->devvp);
 146                 nlookup_done(&nd);
 147         } else {
 148                 error = 0;
 149                 volume->devvp = devvp;
 150         }
 151
 152         if (error == 0) {
 153                 if (vn_isdisk(volume->devvp, &error)) {
 154                         error = vfs_mountedon(volume->devvp);
 155                 }
 156         }
 157         if (error == 0 && vcount(volume->devvp) > 0)
 158                 error = EBUSY;
 159         if (error == 0) {
 160                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 161                 error = vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 162                 if (error == 0) {
 163                         error = VOP_OPEN(volume->devvp,
 164                                          (ronly ? FREAD : FREAD|FWRITE),
 165                                          FSCRED, NULL);
 166                 }
 167                 vn_unlock(volume->devvp);
 168         }
 169         if (error) {
 170                 hammer_free_volume(volume);
 171                 return(error);
 172         }
 173         volume->devvp->v_rdev->si_mountpoint = mp;
 174         setmp = 1;
 175
 176         /*
 177          * Extract the volume number from the volume header and do various
 178          * sanity checks.
 179          */
 180         error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 181         if (error)
 182                 goto late_failure;
 183         ondisk = (void *)bp->b_data;
 184
 185         /*
 186          * Initialize the volume header with data if the data is specified.
 187          */
 188         if (ronly == 0 && data) {
 189                 img = (struct hammer_volume_ondisk *)data;
 190                 if (ondisk->vol_signature == HAMMER_FSBUF_VOLUME) {
 191                         hkprintf("Formatting of valid HAMMER volume "
 192                                 "%s denied. Erase with dd!\n", volname);
 193                         error = EFTYPE;
 194                         goto late_failure;
 195                 }
 196                 bcopy(img, ondisk, sizeof(*img));
 197         }
 198
 199         if (ondisk->vol_signature != HAMMER_FSBUF_VOLUME) {
 200                 hkprintf("volume %s has an invalid header\n",
 201                         volume->vol_name);
 202                 for (i = 0; i < (int)sizeof(ondisk->vol_signature); i++) {
 203                         kprintf("%02x", ((char*)&ondisk->vol_signature)[i] & 0xFF);
 204                         if (i != (int)sizeof(ondisk->vol_signature) - 1)
 205                                 kprintf(" ");
 206                 }
 207                 kprintf("\n");
 208                 error = EFTYPE;
 209                 goto late_failure;
 210         }
 211         volume->vol_no = ondisk->vol_no;
 212         volume->vol_flags = ondisk->vol_flags;
 213         volume->maxbuf_off = HAMMER_ENCODE_RAW_BUFFER(volume->vol_no,
 214                                     ondisk->vol_buf_end - ondisk->vol_buf_beg);
 215
 216         if (RB_EMPTY(&hmp->rb_vols_root)) {
 217                 hmp->fsid = ondisk->vol_fsid;
 218         } else if (bcmp(&hmp->fsid, &ondisk->vol_fsid, sizeof(uuid_t))) {
 219                 hkprintf("volume %s's fsid does not match other volumes\n",
 220                         volume->vol_name);
 221                 error = EFTYPE;
 222                 goto late_failure;
 223         }
 224
 225         /*
 226          * Insert the volume structure into the red-black tree.
 227          */
 228         if (RB_INSERT(hammer_vol_rb_tree, &hmp->rb_vols_root, volume)) {
 229                 hkprintf("volume %s has a duplicate vol_no %d\n",
 230                         volume->vol_name, volume->vol_no);
 231                 error = EEXIST;
 232         }
 233
 234         if (error == 0)
 235                 HAMMER_VOLUME_NUMBER_ADD(hmp, volume);
 236
 237         /*
 238          * Set the root volume .  HAMMER special cases rootvol the structure.
 239          * We do not hold a ref because this would prevent related I/O
 240          * from being flushed.
 241          */
 242         if (error == 0 && ondisk->vol_rootvol == ondisk->vol_no) {
 243                 hmp->rootvol = volume;
 244                 hmp->nvolumes = ondisk->vol_count;
 245                 if (bp) {
 246                         brelse(bp);
 247                         bp = NULL;
 248                 }
 249                 hmp->mp->mnt_stat.f_blocks += ondisk->vol0_stat_bigblocks *
 250                                                 HAMMER_BUFFERS_PER_BIGBLOCK;
 251                 hmp->mp->mnt_vstat.f_blocks += ondisk->vol0_stat_bigblocks *
 252                                                 HAMMER_BUFFERS_PER_BIGBLOCK;
 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                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 262                 VOP_CLOSE(volume->devvp, ronly ? FREAD : FREAD|FWRITE, NULL);
 263                 vn_unlock(volume->devvp);
 264                 hammer_free_volume(volume);
 265         }
 266         return (error);
 267 }
 268
 269 /*
 270  * This is called for each volume when updating the mount point from
 271  * read-write to read-only or vise-versa.
 272  */
 273 int
 274 hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused)
 275 {
 276         if (volume->devvp) {
 277                 vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 278                 if (volume->io.hmp->ronly) {
 279                         /* do not call vinvalbuf */
 280                         VOP_OPEN(volume->devvp, FREAD, FSCRED, NULL);
 281                         VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL);
 282                 } else {
 283                         /* do not call vinvalbuf */
 284                         VOP_OPEN(volume->devvp, FREAD|FWRITE, FSCRED, NULL);
 285                         VOP_CLOSE(volume->devvp, FREAD, NULL);
 286                 }
 287                 vn_unlock(volume->devvp);
 288         }
 289         return(0);
 290 }
 291
 292 /*
 293  * Unload and free a HAMMER volume.  Must return >= 0 to continue scan
 294  * so returns -1 on failure.
 295  */
 296 int
 297 hammer_unload_volume(hammer_volume_t volume, void *data)
 298 {
 299         hammer_mount_t hmp = volume->io.hmp;
 300         struct buf *bp = NULL;
 301         struct hammer_volume_ondisk *img;
 302         int ronly = ((hmp->mp->mnt_flag & MNT_RDONLY) ? 1 : 0);
 303         int error;
 304
 305         /*
 306          * Clear the volume header with data if the data is specified.
 307          */
 308         if (ronly == 0 && data && volume->devvp) {
 309                 img = (struct hammer_volume_ondisk *)data;
 310                 error = bread(volume->devvp, 0LL, HAMMER_BUFSIZE, &bp);
 311                 if (error || bp->b_bcount < sizeof(*img)) {
 312                         hmkprintf(hmp, "Failed to read volume header: %d\n", error);
 313                         brelse(bp);
 314                 } else {
 315                         bcopy(img, bp->b_data, sizeof(*img));
 316                         error = bwrite(bp);
 317                         if (error)
 318                                 hmkprintf(hmp, "Failed to clear volume header: %d\n",
 319                                         error);
 320                 }
 321         }
 322
 323         /*
 324          * Clean up the root volume pointer, which is held unlocked in hmp.
 325          */
 326         if (hmp->rootvol == volume)
 327                 hmp->rootvol = NULL;
 328
 329         /*
 330          * We must not flush a dirty buffer to disk on umount.  It should
 331          * have already been dealt with by the flusher, or we may be in
 332          * catastrophic failure.
 333          */
 334         hammer_io_clear_modify(&volume->io, 1);
 335         volume->io.waitdep = 1;
 336
 337         /*
 338          * Clean up the persistent ref ioerror might have on the volume
 339          */
 340         if (volume->io.ioerror)
 341                 hammer_io_clear_error_noassert(&volume->io);
 342
 343         /*
 344          * This should release the bp.  Releasing the volume with flush set
 345          * implies the interlock is set.
 346          */
 347         hammer_ref_interlock_true(&volume->io.lock);
 348         hammer_rel_volume(volume, 1);
 349         KKASSERT(volume->io.bp == NULL);
 350
 351         /*
 352          * There should be no references on the volume.
 353          */
 354         KKASSERT(hammer_norefs(&volume->io.lock));
 355
 356         volume->ondisk = NULL;
 357         if (volume->devvp) {
 358                 if (volume->devvp->v_rdev &&
 359                     volume->devvp->v_rdev->si_mountpoint == hmp->mp) {
 360                         volume->devvp->v_rdev->si_mountpoint = NULL;
 361                 }
 362                 if (ronly) {
 363                         /*
 364                          * Make sure we don't sync anything to disk if we
 365                          * are in read-only mode (1) or critically-errored
 366                          * (2).  Note that there may be dirty buffers in
 367                          * normal read-only mode from crash recovery.
 368                          */
 369                         vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 370                         vinvalbuf(volume->devvp, 0, 0, 0);
 371                         VOP_CLOSE(volume->devvp, FREAD, NULL);
 372                         vn_unlock(volume->devvp);
 373                 } else {
 374                         /*
 375                          * Normal termination, save any dirty buffers
 376                          * (XXX there really shouldn't be any).
 377                          */
 378                         vn_lock(volume->devvp, LK_EXCLUSIVE | LK_RETRY);
 379                         vinvalbuf(volume->devvp, V_SAVE, 0, 0);
 380                         VOP_CLOSE(volume->devvp, FREAD|FWRITE, NULL);
 381                         vn_unlock(volume->devvp);
 382                 }
 383         }
 384
 385         /*
 386          * Destroy the structure
 387          */
 388         RB_REMOVE(hammer_vol_rb_tree, &hmp->rb_vols_root, volume);
 389         HAMMER_VOLUME_NUMBER_DEL(hmp, volume);
 390         hammer_free_volume(volume);
 391         return(0);
 392 }
 393
 394 static
 395 void
 396 hammer_free_volume(hammer_volume_t volume)
 397 {
 398         hammer_mount_t hmp = volume->io.hmp;
 399
 400         if (volume->vol_name) {
 401                 kfree(volume->vol_name, hmp->m_misc);
 402                 volume->vol_name = NULL;
 403         }
 404         if (volume->devvp) {
 405                 vrele(volume->devvp);
 406                 volume->devvp = NULL;
 407         }
 408         --hammer_count_volumes;
 409         kfree(volume, hmp->m_misc);
 410 }
 411
 412 /*
 413  * Get a HAMMER volume.  The volume must already exist.
 414  */
 415 hammer_volume_t
 416 hammer_get_volume(hammer_mount_t hmp, int32_t vol_no, int *errorp)
 417 {
 418         struct hammer_volume *volume;
 419
 420         /*
 421          * Locate the volume structure
 422          */
 423         volume = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, vol_no);
 424         if (volume == NULL) {
 425                 *errorp = ENOENT;
 426                 return(NULL);
 427         }
 428
 429         /*
 430          * Reference the volume, load/check the data on the 0->1 transition.
 431          * hammer_load_volume() will dispose of the interlock on return,
 432          * and also clean up the ref count on error.
 433          */
 434         if (hammer_ref_interlock(&volume->io.lock)) {
 435                 *errorp = hammer_load_volume(volume);
 436                 if (*errorp)
 437                         volume = NULL;
 438         } else {
 439                 KKASSERT(volume->ondisk);
 440                 *errorp = 0;
 441         }
 442         return(volume);
 443 }
 444
 445 int
 446 hammer_ref_volume(hammer_volume_t volume)
 447 {
 448         int error;
 449
 450         /*
 451          * Reference the volume and deal with the check condition used to
 452          * load its ondisk info.
 453          */
 454         if (hammer_ref_interlock(&volume->io.lock)) {
 455                 error = hammer_load_volume(volume);
 456         } else {
 457                 KKASSERT(volume->ondisk);
 458                 error = 0;
 459         }
 460         return (error);
 461 }
 462
 463 /*
 464  * May be called without fs_token
 465  */
 466 hammer_volume_t
 467 hammer_get_root_volume(hammer_mount_t hmp, int *errorp)
 468 {
 469         hammer_volume_t volume;
 470
 471         volume = hmp->rootvol;
 472         KKASSERT(volume != NULL);
 473
 474         /*
 475          * Reference the volume and deal with the check condition used to
 476          * load its ondisk info.
 477          */
 478         if (hammer_ref_interlock(&volume->io.lock)) {
 479                 lwkt_gettoken(&volume->io.hmp->fs_token);
 480                 *errorp = hammer_load_volume(volume);
 481                 lwkt_reltoken(&volume->io.hmp->fs_token);
 482                 if (*errorp)
 483                         volume = NULL;
 484         } else {
 485                 KKASSERT(volume->ondisk);
 486                 *errorp = 0;
 487         }
 488         return (volume);
 489 }
 490
 491 /*
 492  * Load a volume's on-disk information.  The volume must be referenced and
 493  * the interlock is held on call.  The interlock will be released on return.
 494  * The reference will also be released on return if an error occurs.
 495  */
 496 static int
 497 hammer_load_volume(hammer_volume_t volume)
 498 {
 499         int error;
 500
 501         if (volume->ondisk == NULL) {
 502                 error = hammer_io_read(volume->devvp, &volume->io,
 503                                        HAMMER_BUFSIZE);
 504                 if (error == 0) {
 505                         volume->ondisk = (void *)volume->io.bp->b_data;
 506                         hammer_ref_interlock_done(&volume->io.lock);
 507                 } else {
 508                         hammer_rel_volume(volume, 1);
 509                 }
 510         } else {
 511                 error = 0;
 512         }
 513         return(error);
 514 }
 515
 516 /*
 517  * Release a previously acquired reference on the volume.
 518  *
 519  * Volumes are not unloaded from memory during normal operation.
 520  *
 521  * May be called without fs_token
 522  */
 523 void
 524 hammer_rel_volume(hammer_volume_t volume, int locked)
 525 {
 526         struct buf *bp;
 527
 528         if (hammer_rel_interlock(&volume->io.lock, locked)) {
 529                 lwkt_gettoken(&volume->io.hmp->fs_token);
 530                 volume->ondisk = NULL;
 531                 bp = hammer_io_release(&volume->io, locked);
 532                 lwkt_reltoken(&volume->io.hmp->fs_token);
 533                 hammer_rel_interlock_done(&volume->io.lock, locked);
 534                 if (bp)
 535                         brelse(bp);
 536         }
 537 }
 538
 539 int
 540 hammer_mountcheck_volumes(hammer_mount_t hmp)
 541 {
 542         hammer_volume_t vol;
 543         int i;
 544
 545         HAMMER_VOLUME_NUMBER_FOREACH(hmp, i) {
 546                 vol = RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, i);
 547                 if (vol == NULL)
 548                         return(EINVAL);
 549         }
 550         return(0);
 551 }
 552
 553 int
 554 hammer_get_installed_volumes(hammer_mount_t hmp)
 555 {
 556         int i, ret = 0;
 557
 558         HAMMER_VOLUME_NUMBER_FOREACH(hmp, i)
 559                 ret++;
 560         return(ret);
 561 }
 562
 563 /************************************************************************
 564  *                              BUFFERS                                 *
 565  ************************************************************************
 566  *
 567  * Manage buffers.  Currently most blockmap-backed zones are direct-mapped
 568  * to zone-2 buffer offsets, without a translation stage.  However, the
 569  * hammer_buffer structure is indexed by its zoneX_offset, not its
 570  * zone2_offset.
 571  *
 572  * The proper zone must be maintained throughout the code-base all the way
 573  * through to the big-block allocator, or routines like hammer_del_buffers()
 574  * will not be able to locate all potentially conflicting buffers.
 575  */
 576
 577 /*
 578  * Helper function returns whether a zone offset can be directly translated
 579  * to a raw buffer index or not.  Really only the volume and undo zones
 580  * can't be directly translated.  Volumes are special-cased and undo zones
 581  * shouldn't be aliased accessed in read-only mode.
 582  *
 583  * This function is ONLY used to detect aliased zones during a read-only
 584  * mount.
 585  */
 586 static __inline int
 587 hammer_direct_zone(hammer_off_t buf_offset)
 588 {
 589         int zone = HAMMER_ZONE_DECODE(buf_offset);
 590
 591         return(hammer_is_direct_mapped_index(zone));
 592 }
 593
 594 hammer_buffer_t
 595 hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 596                   int bytes, int isnew, int *errorp)
 597 {
 598         hammer_buffer_t buffer;
 599         hammer_volume_t volume;
 600         hammer_off_t    zone2_offset;
 601         hammer_io_type_t iotype;
 602         int vol_no;
 603         int zone;
 604
 605         buf_offset &= ~HAMMER_BUFMASK64;
 606 again:
 607         /*
 608          * Shortcut if the buffer is already cached
 609          */
 610         buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root, buf_offset);
 611         if (buffer) {
 612                 /*
 613                  * Once refed the ondisk field will not be cleared by
 614                  * any other action.  Shortcut the operation if the
 615                  * ondisk structure is valid.
 616                  */
 617 found_aliased:
 618                 if (hammer_ref_interlock(&buffer->io.lock) == 0) {
 619                         hammer_io_advance(&buffer->io);
 620                         KKASSERT(buffer->ondisk);
 621                         *errorp = 0;
 622                         return(buffer);
 623                 }
 624
 625                 /*
 626                  * 0->1 transition or defered 0->1 transition (CHECK),
 627                  * interlock now held.  Shortcut if ondisk is already
 628                  * assigned.
 629                  */
 630                 atomic_add_int(&hammer_count_refedbufs, 1);
 631                 if (buffer->ondisk) {
 632                         hammer_io_advance(&buffer->io);
 633                         hammer_ref_interlock_done(&buffer->io.lock);
 634                         *errorp = 0;
 635                         return(buffer);
 636                 }
 637
 638                 /*
 639                  * The buffer is no longer loose if it has a ref, and
 640                  * cannot become loose once it gains a ref.  Loose
 641                  * buffers will never be in a modified state.  This should
 642                  * only occur on the 0->1 transition of refs.
 643                  *
 644                  * lose_root can be modified via a biodone() interrupt
 645                  * so the io_token must be held.
 646                  */
 647                 if (buffer->io.mod_root == &hmp->lose_root) {
 648                         lwkt_gettoken(&hmp->io_token);
 649                         if (buffer->io.mod_root == &hmp->lose_root) {
 650                                 RB_REMOVE(hammer_mod_rb_tree,
 651                                           buffer->io.mod_root, &buffer->io);
 652                                 buffer->io.mod_root = NULL;
 653                                 KKASSERT(buffer->io.modified == 0);
 654                         }
 655                         lwkt_reltoken(&hmp->io_token);
 656                 }
 657                 goto found;
 658         } else if (hmp->ronly && hammer_direct_zone(buf_offset)) {
 659                 /*
 660                  * If this is a read-only mount there could be an alias
 661                  * in the raw-zone.  If there is we use that buffer instead.
 662                  *
 663                  * rw mounts will not have aliases.  Also note when going
 664                  * from ro -> rw the recovered raw buffers are flushed and
 665                  * reclaimed, so again there will not be any aliases once
 666                  * the mount is rw.
 667                  */
 668                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 669                                    hammer_xlate_to_zone2(buf_offset));
 670                 if (buffer) {
 671                         if (hammer_debug_general & 0x0001) {
 672                                 hkrateprintf(&hmp->kdiag,
 673                                             "recovered aliased %016jx\n",
 674                                             (intmax_t)buf_offset);
 675                         }
 676                         goto found_aliased;
 677                 }
 678         }
 679
 680         /*
 681          * What is the buffer class?
 682          */
 683         zone = HAMMER_ZONE_DECODE(buf_offset);
 684
 685         switch(zone) {
 686         case HAMMER_ZONE_LARGE_DATA_INDEX:
 687         case HAMMER_ZONE_SMALL_DATA_INDEX:
 688                 iotype = HAMMER_STRUCTURE_DATA_BUFFER;
 689                 break;
 690         case HAMMER_ZONE_UNDO_INDEX:
 691                 iotype = HAMMER_STRUCTURE_UNDO_BUFFER;
 692                 break;
 693         case HAMMER_ZONE_META_INDEX:
 694         default:
 695                 /*
 696                  * NOTE: inode data and directory entries are placed in this
 697                  * zone.  inode atime/mtime is updated in-place and thus
 698                  * buffers containing inodes must be synchronized as
 699                  * meta-buffers, same as buffers containing B-Tree info.
 700                  */
 701                 iotype = HAMMER_STRUCTURE_META_BUFFER;
 702                 break;
 703         }
 704
 705         /*
 706          * Handle blockmap offset translations
 707          */
 708         if (hammer_is_zone2_mapped_index(zone)) {
 709                 zone2_offset = hammer_blockmap_lookup(hmp, buf_offset, errorp);
 710         } else if (zone == HAMMER_ZONE_UNDO_INDEX) {
 711                 zone2_offset = hammer_undo_lookup(hmp, buf_offset, errorp);
 712         } else {
 713                 /* Must be zone-2 (not 1 or 4 or 15) */
 714                 KKASSERT(zone == HAMMER_ZONE_RAW_BUFFER_INDEX);
 715                 zone2_offset = buf_offset;
 716                 *errorp = 0;
 717         }
 718         if (*errorp)
 719                 return(NULL);
 720
 721         /*
 722          * NOTE: zone2_offset and maxbuf_off are both full zone-2 offset
 723          * specifications.
 724          */
 725         KKASSERT((zone2_offset & HAMMER_OFF_ZONE_MASK) ==
 726                  HAMMER_ZONE_RAW_BUFFER);
 727         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 728         volume = hammer_get_volume(hmp, vol_no, errorp);
 729         if (volume == NULL)
 730                 return(NULL);
 731
 732         KKASSERT(zone2_offset < volume->maxbuf_off);
 733
 734         /*
 735          * Allocate a new buffer structure.  We will check for races later.
 736          */
 737         ++hammer_count_buffers;
 738         buffer = kmalloc(sizeof(*buffer), hmp->m_misc,
 739                          M_WAITOK|M_ZERO|M_USE_RESERVE);
 740         buffer->zone2_offset = zone2_offset;
 741         buffer->zoneX_offset = buf_offset;
 742
 743         hammer_io_init(&buffer->io, volume, iotype);
 744         buffer->io.offset = hammer_xlate_to_phys(volume->ondisk, zone2_offset);
 745         buffer->io.bytes = bytes;
 746         TAILQ_INIT(&buffer->node_list);
 747         hammer_ref_interlock_true(&buffer->io.lock);
 748
 749         /*
 750          * Insert the buffer into the RB tree and handle late collisions.
 751          */
 752         if (RB_INSERT(hammer_buf_rb_tree, &hmp->rb_bufs_root, buffer)) {
 753                 hammer_rel_volume(volume, 0);
 754                 buffer->io.volume = NULL;                       /* safety */
 755                 if (hammer_rel_interlock(&buffer->io.lock, 1))  /* safety */
 756                         hammer_rel_interlock_done(&buffer->io.lock, 1);
 757                 --hammer_count_buffers;
 758                 kfree(buffer, hmp->m_misc);
 759                 goto again;
 760         }
 761         atomic_add_int(&hammer_count_refedbufs, 1);
 762 found:
 763
 764         /*
 765          * The buffer is referenced and interlocked.  Load the buffer
 766          * if necessary.  hammer_load_buffer() deals with the interlock
 767          * and, if an error is returned, also deals with the ref.
 768          */
 769         if (buffer->ondisk == NULL) {
 770                 *errorp = hammer_load_buffer(buffer, isnew);
 771                 if (*errorp)
 772                         buffer = NULL;
 773         } else {
 774                 hammer_io_advance(&buffer->io);
 775                 hammer_ref_interlock_done(&buffer->io.lock);
 776                 *errorp = 0;
 777         }
 778         return(buffer);
 779 }
 780
 781 /*
 782  * This is used by the direct-read code to deal with large-data buffers
 783  * created by the reblocker and mirror-write code.  The direct-read code
 784  * bypasses the HAMMER buffer subsystem and so any aliased dirty or write-
 785  * running hammer buffers must be fully synced to disk before we can issue
 786  * the direct-read.
 787  *
 788  * This code path is not considered critical as only the rebocker and
 789  * mirror-write code will create large-data buffers via the HAMMER buffer
 790  * subsystem.  They do that because they operate at the B-Tree level and
 791  * do not access the vnode/inode structures.
 792  */
 793 void
 794 hammer_sync_buffers(hammer_mount_t hmp, hammer_off_t base_offset, int bytes)
 795 {
 796         hammer_buffer_t buffer;
 797         int error;
 798
 799         KKASSERT((base_offset & HAMMER_OFF_ZONE_MASK) ==
 800                  HAMMER_ZONE_LARGE_DATA);
 801
 802         while (bytes > 0) {
 803                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 804                                    base_offset);
 805                 if (buffer && (buffer->io.modified || buffer->io.running)) {
 806                         error = hammer_ref_buffer(buffer);
 807                         if (error == 0) {
 808                                 hammer_io_wait(&buffer->io);
 809                                 if (buffer->io.modified) {
 810                                         hammer_io_write_interlock(&buffer->io);
 811                                         hammer_io_flush(&buffer->io, 0);
 812                                         hammer_io_done_interlock(&buffer->io);
 813                                         hammer_io_wait(&buffer->io);
 814                                 }
 815                                 hammer_rel_buffer(buffer, 0);
 816                         }
 817                 }
 818                 base_offset += HAMMER_BUFSIZE;
 819                 bytes -= HAMMER_BUFSIZE;
 820         }
 821 }
 822
 823 /*
 824  * Destroy all buffers covering the specified zoneX offset range.  This
 825  * is called when the related blockmap layer2 entry is freed or when
 826  * a direct write bypasses our buffer/buffer-cache subsystem.
 827  *
 828  * The buffers may be referenced by the caller itself.  Setting reclaim
 829  * will cause the buffer to be destroyed when it's ref count reaches zero.
 830  *
 831  * Return 0 on success, EAGAIN if some buffers could not be destroyed due
 832  * to additional references held by other threads, or some other (typically
 833  * fatal) error.
 834  */
 835 int
 836 hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 837                    hammer_off_t zone2_offset, int bytes,
 838                    int report_conflicts)
 839 {
 840         hammer_buffer_t buffer;
 841         hammer_volume_t volume;
 842         int vol_no;
 843         int error;
 844         int ret_error;
 845
 846         vol_no = HAMMER_VOL_DECODE(zone2_offset);
 847         volume = hammer_get_volume(hmp, vol_no, &ret_error);
 848         KKASSERT(ret_error == 0);
 849
 850         while (bytes > 0) {
 851                 buffer = RB_LOOKUP(hammer_buf_rb_tree, &hmp->rb_bufs_root,
 852                                    base_offset);
 853                 if (buffer) {
 854                         error = hammer_ref_buffer(buffer);
 855                         if (hammer_debug_general & 0x20000) {
 856                                 hkprintf("delbufr %016jx rerr=%d 1ref=%d\n",
 857                                         (intmax_t)buffer->zoneX_offset,
 858                                         error,
 859                                         hammer_oneref(&buffer->io.lock));
 860                         }
 861                         if (error == 0 && !hammer_oneref(&buffer->io.lock)) {
 862                                 error = EAGAIN;
 863                                 hammer_rel_buffer(buffer, 0);
 864                         }
 865                         if (error == 0) {
 866                                 KKASSERT(buffer->zone2_offset == zone2_offset);
 867                                 hammer_io_clear_modify(&buffer->io, 1);
 868                                 buffer->io.reclaim = 1;
 869                                 buffer->io.waitdep = 1;
 870                                 KKASSERT(buffer->io.volume == volume);
 871                                 hammer_rel_buffer(buffer, 0);
 872                         }
 873                 } else {
 874                         error = hammer_io_inval(volume, zone2_offset);
 875                 }
 876                 if (error) {
 877                         ret_error = error;
 878                         if (report_conflicts ||
 879                             (hammer_debug_general & 0x8000)) {
 880                                 krateprintf(&hmp->kdiag,
 881                                         "hammer_del_buffers: unable to "
 882                                         "invalidate %016llx buffer=%p "
 883                                         "rep=%d lkrefs=%08x\n",
 884                                         (long long)base_offset,
 885                                         buffer, report_conflicts,
 886                                         (buffer ? buffer->io.lock.refs : -1));
 887                         }
 888                 }
 889                 base_offset += HAMMER_BUFSIZE;
 890                 zone2_offset += HAMMER_BUFSIZE;
 891                 bytes -= HAMMER_BUFSIZE;
 892         }
 893         hammer_rel_volume(volume, 0);
 894         return (ret_error);
 895 }
 896
 897 /*
 898  * Given a referenced and interlocked buffer load/validate the data.
 899  *
 900  * The buffer interlock will be released on return.  If an error is
 901  * returned the buffer reference will also be released (and the buffer
 902  * pointer will thus be stale).
 903  */
 904 static int
 905 hammer_load_buffer(hammer_buffer_t buffer, int isnew)
 906 {
 907         hammer_volume_t volume;
 908         int error;
 909
 910         /*
 911          * Load the buffer's on-disk info
 912          */
 913         volume = buffer->io.volume;
 914
 915         if (hammer_debug_io & 0x0004) {
 916                 hdkprintf("load_buffer %016llx %016llx isnew=%d od=%p\n",
 917                         (long long)buffer->zoneX_offset,
 918                         (long long)buffer->zone2_offset,
 919                         isnew, buffer->ondisk);
 920         }
 921
 922         if (buffer->ondisk == NULL) {
 923                 /*
 924                  * Issue the read or generate a new buffer.  When reading
 925                  * the limit argument controls any read-ahead clustering
 926                  * hammer_io_read() is allowed to do.
 927                  *
 928                  * We cannot read-ahead in the large-data zone and we cannot
 929                  * cross a big-block boundary as the next big-block might
 930                  * use a different buffer size.
 931                  */
 932                 if (isnew) {
 933                         error = hammer_io_new(volume->devvp, &buffer->io);
 934                 } else if ((buffer->zoneX_offset & HAMMER_OFF_ZONE_MASK) ==
 935                            HAMMER_ZONE_LARGE_DATA) {
 936                         error = hammer_io_read(volume->devvp, &buffer->io,
 937                                                buffer->io.bytes);
 938                 } else {
 939                         hammer_off_t limit;
 940
 941                         limit = (buffer->zone2_offset +
 942                                  HAMMER_BIGBLOCK_MASK64) &
 943                                 ~HAMMER_BIGBLOCK_MASK64;
 944                         limit -= buffer->zone2_offset;
 945                         error = hammer_io_read(volume->devvp, &buffer->io,
 946                                                limit);
 947                 }
 948                 if (error == 0)
 949                         buffer->ondisk = (void *)buffer->io.bp->b_data;
 950         } else if (isnew) {
 951                 error = hammer_io_new(volume->devvp, &buffer->io);
 952         } else {
 953                 error = 0;
 954         }
 955         if (error == 0) {
 956                 hammer_io_advance(&buffer->io);
 957                 hammer_ref_interlock_done(&buffer->io.lock);
 958         } else {
 959                 hammer_rel_buffer(buffer, 1);
 960         }
 961         return (error);
 962 }
 963
 964 /*
 965  * NOTE: Called from RB_SCAN, must return >= 0 for scan to continue.
 966  * This routine is only called during unmount or when a volume is
 967  * removed.
 968  *
 969  * If data != NULL, it specifies a volume whoose buffers should
 970  * be unloaded.
 971  */
 972 int
 973 hammer_unload_buffer(hammer_buffer_t buffer, void *data)
 974 {
 975         struct hammer_volume *volume = (struct hammer_volume *) data;
 976
 977         /*
 978          * If volume != NULL we are only interested in unloading buffers
 979          * associated with a particular volume.
 980          */
 981         if (volume != NULL && volume != buffer->io.volume)
 982                 return 0;
 983
 984         /*
 985          * Clean up the persistent ref ioerror might have on the buffer
 986          * and acquire a ref.  Expect a 0->1 transition.
 987          */
 988         if (buffer->io.ioerror) {
 989                 hammer_io_clear_error_noassert(&buffer->io);
 990                 atomic_add_int(&hammer_count_refedbufs, -1);
 991         }
 992         hammer_ref_interlock_true(&buffer->io.lock);
 993         atomic_add_int(&hammer_count_refedbufs, 1);
 994
 995         /*
 996          * We must not flush a dirty buffer to disk on umount.  It should
 997          * have already been dealt with by the flusher, or we may be in
 998          * catastrophic failure.
 999          *
1000          * We must set waitdep to ensure that a running buffer is waited
1001          * on and released prior to us trying to unload the volume.
1002          */
1003         hammer_io_clear_modify(&buffer->io, 1);
1004         hammer_flush_buffer_nodes(buffer);
1005         buffer->io.waitdep = 1;
1006         hammer_rel_buffer(buffer, 1);
1007         return(0);
1008 }
1009
1010 /*
1011  * Reference a buffer that is either already referenced or via a specially
1012  * handled pointer (aka cursor->buffer).
1013  */
1014 int
1015 hammer_ref_buffer(hammer_buffer_t buffer)
1016 {
1017         hammer_mount_t hmp;
1018         int error;
1019         int locked;
1020
1021         /*
1022          * Acquire a ref, plus the buffer will be interlocked on the
1023          * 0->1 transition.
1024          */
1025         locked = hammer_ref_interlock(&buffer->io.lock);
1026         hmp = buffer->io.hmp;
1027
1028         /*
1029          * At this point a biodone() will not touch the buffer other then
1030          * incidental bits.  However, lose_root can be modified via
1031          * a biodone() interrupt.
1032          *
1033          * No longer loose.  lose_root requires the io_token.
1034          */
1035         if (buffer->io.mod_root == &hmp->lose_root) {
1036                 lwkt_gettoken(&hmp->io_token);
1037                 if (buffer->io.mod_root == &hmp->lose_root) {
1038                         RB_REMOVE(hammer_mod_rb_tree,
1039                                   buffer->io.mod_root, &buffer->io);
1040                         buffer->io.mod_root = NULL;
1041                 }
1042                 lwkt_reltoken(&hmp->io_token);
1043         }
1044
1045         if (locked) {
1046                 atomic_add_int(&hammer_count_refedbufs, 1);
1047                 error = hammer_load_buffer(buffer, 0);
1048                 /* NOTE: on error the buffer pointer is stale */
1049         } else {
1050                 error = 0;
1051         }
1052         return(error);
1053 }
1054
1055 /*
1056  * Release a reference on the buffer.  On the 1->0 transition the
1057  * underlying IO will be released but the data reference is left
1058  * cached.
1059  *
1060  * Only destroy the structure itself if the related buffer cache buffer
1061  * was disassociated from it.  This ties the management of the structure
1062  * to the buffer cache subsystem.  buffer->ondisk determines whether the
1063  * embedded io is referenced or not.
1064  */
1065 void
1066 hammer_rel_buffer(hammer_buffer_t buffer, int locked)
1067 {
1068         hammer_volume_t volume;
1069         hammer_mount_t hmp;
1070         struct buf *bp = NULL;
1071         int freeme = 0;
1072
1073         hmp = buffer->io.hmp;
1074
1075         if (hammer_rel_interlock(&buffer->io.lock, locked) == 0)
1076                 return;
1077
1078         /*
1079          * hammer_count_refedbufs accounting.  Decrement if we are in
1080          * the error path or if CHECK is clear.
1081          *
1082          * If we are not in the error path and CHECK is set the caller
1083          * probably just did a hammer_ref() and didn't account for it,
1084          * so we don't account for the loss here.
1085          */
1086         if (locked || (buffer->io.lock.refs & HAMMER_REFS_CHECK) == 0)
1087                 atomic_add_int(&hammer_count_refedbufs, -1);
1088
1089         /*
1090          * If the caller locked us or the normal released transitions
1091          * from 1->0 (and acquired the lock) attempt to release the
1092          * io.  If the called locked us we tell hammer_io_release()
1093          * to flush (which would be the unload or failure path).
1094          */
1095         bp = hammer_io_release(&buffer->io, locked);
1096
1097         /*
1098          * If the buffer has no bp association and no refs we can destroy
1099          * it.
1100          *
1101          * NOTE: It is impossible for any associated B-Tree nodes to have
1102          * refs if the buffer has no additional refs.
1103          */
1104         if (buffer->io.bp == NULL && hammer_norefs(&buffer->io.lock)) {
1105                 RB_REMOVE(hammer_buf_rb_tree,
1106                           &buffer->io.hmp->rb_bufs_root,
1107                           buffer);
1108                 volume = buffer->io.volume;
1109                 buffer->io.volume = NULL; /* sanity */
1110                 hammer_rel_volume(volume, 0);
1111                 hammer_io_clear_modlist(&buffer->io);
1112                 hammer_flush_buffer_nodes(buffer);
1113                 KKASSERT(TAILQ_EMPTY(&buffer->node_list));
1114                 freeme = 1;
1115         }
1116
1117         /*
1118          * Cleanup
1119          */
1120         hammer_rel_interlock_done(&buffer->io.lock, locked);
1121         if (bp)
1122                 brelse(bp);
1123         if (freeme) {
1124                 --hammer_count_buffers;
1125                 kfree(buffer, hmp->m_misc);
1126         }
1127 }
1128
1129 /*
1130  * Access the filesystem buffer containing the specified hammer offset.
1131  * buf_offset is a conglomeration of the volume number and vol_buf_beg
1132  * relative buffer offset.  It must also have bit 55 set to be valid.
1133  * (see hammer_off_t in hammer_disk.h).
1134  *
1135  * Any prior buffer in *bufferp will be released and replaced by the
1136  * requested buffer.
1137  *
1138  * NOTE: The buffer is indexed via its zoneX_offset but we allow the
1139  * passed cached *bufferp to match against either zoneX or zone2.
1140  */
1141 static __inline
1142 void *
1143 _hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1144              int isnew, int *errorp, struct hammer_buffer **bufferp)
1145 {
1146         hammer_buffer_t buffer;
1147         int32_t xoff = (int32_t)buf_offset & HAMMER_BUFMASK;
1148
1149         buf_offset &= ~HAMMER_BUFMASK64;
1150         KKASSERT(HAMMER_ZONE(buf_offset) != 0);
1151
1152         buffer = *bufferp;
1153         if (buffer == NULL || (buffer->zone2_offset != buf_offset &&
1154                                buffer->zoneX_offset != buf_offset)) {
1155                 if (buffer)
1156                         hammer_rel_buffer(buffer, 0);
1157                 buffer = hammer_get_buffer(hmp, buf_offset, bytes, isnew, errorp);
1158                 *bufferp = buffer;
1159         } else {
1160                 *errorp = 0;
1161         }
1162
1163         /*
1164          * Return a pointer to the buffer data.
1165          */
1166         if (buffer == NULL)
1167                 return(NULL);
1168         else
1169                 return((char *)buffer->ondisk + xoff);
1170 }
1171
1172 void *
1173 hammer_bread(hammer_mount_t hmp, hammer_off_t buf_offset,
1174              int *errorp, struct hammer_buffer **bufferp)
1175 {
1176         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, 0, errorp, bufferp));
1177 }
1178
1179 void *
1180 hammer_bread_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1181                  int *errorp, struct hammer_buffer **bufferp)
1182 {
1183         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1184         return(_hammer_bread(hmp, buf_offset, bytes, 0, errorp, bufferp));
1185 }
1186
1187 /*
1188  * Access the filesystem buffer containing the specified hammer offset.
1189  * No disk read operation occurs.  The result buffer may contain garbage.
1190  *
1191  * Any prior buffer in *bufferp will be released and replaced by the
1192  * requested buffer.
1193  *
1194  * This function marks the buffer dirty but does not increment its
1195  * modify_refs count.
1196  */
1197 void *
1198 hammer_bnew(hammer_mount_t hmp, hammer_off_t buf_offset,
1199              int *errorp, struct hammer_buffer **bufferp)
1200 {
1201         return(_hammer_bread(hmp, buf_offset, HAMMER_BUFSIZE, 1, errorp, bufferp));
1202 }
1203
1204 void *
1205 hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t buf_offset, int bytes,
1206                 int *errorp, struct hammer_buffer **bufferp)
1207 {
1208         bytes = (bytes + HAMMER_BUFMASK) & ~HAMMER_BUFMASK;
1209         return(_hammer_bread(hmp, buf_offset, bytes, 1, errorp, bufferp));
1210 }
1211
1212 /************************************************************************
1213  *                              NODES                                   *
1214  ************************************************************************
1215  *
1216  * Manage B-Tree nodes.  B-Tree nodes represent the primary indexing
1217  * method used by the HAMMER filesystem.
1218  *
1219  * Unlike other HAMMER structures, a hammer_node can be PASSIVELY
1220  * associated with its buffer, and will only referenced the buffer while
1221  * the node itself is referenced.
1222  *
1223  * A hammer_node can also be passively associated with other HAMMER
1224  * structures, such as inodes, while retaining 0 references.  These
1225  * associations can be cleared backwards using a pointer-to-pointer in
1226  * the hammer_node.
1227  *
1228  * This allows the HAMMER implementation to cache hammer_nodes long-term
1229  * and short-cut a great deal of the infrastructure's complexity.  In
1230  * most cases a cached node can be reacquired without having to dip into
1231  * the B-Tree.
1232  */
1233 hammer_node_t
1234 hammer_get_node(hammer_transaction_t trans, hammer_off_t node_offset,
1235                 int isnew, int *errorp)
1236 {
1237         hammer_mount_t hmp = trans->hmp;
1238         hammer_node_t node;
1239         int doload;
1240
1241         KKASSERT((node_offset & HAMMER_OFF_ZONE_MASK) == HAMMER_ZONE_BTREE);
1242
1243         /*
1244          * Locate the structure, allocating one if necessary.
1245          */
1246 again:
1247         node = RB_LOOKUP(hammer_nod_rb_tree, &hmp->rb_nods_root, node_offset);
1248         if (node == NULL) {
1249                 ++hammer_count_nodes;
1250                 node = kmalloc(sizeof(*node), hmp->m_misc, M_WAITOK|M_ZERO|M_USE_RESERVE);
1251                 node->node_offset = node_offset;
1252                 node->hmp = hmp;
1253                 TAILQ_INIT(&node->cursor_list);
1254                 TAILQ_INIT(&node->cache_list);
1255                 if (RB_INSERT(hammer_nod_rb_tree, &hmp->rb_nods_root, node)) {
1256                         --hammer_count_nodes;
1257                         kfree(node, hmp->m_misc);
1258                         goto again;
1259                 }
1260                 doload = hammer_ref_interlock_true(&node->lock);
1261         } else {
1262                 doload = hammer_ref_interlock(&node->lock);
1263         }
1264         if (doload) {
1265                 *errorp = hammer_load_node(trans, node, isnew);
1266                 if (*errorp)
1267                         node = NULL;
1268         } else {
1269                 KKASSERT(node->ondisk);
1270                 *errorp = 0;
1271                 hammer_io_advance(&node->buffer->io);
1272         }
1273         return(node);
1274 }
1275
1276 /*
1277  * Reference an already-referenced node.  0->1 transitions should assert
1278  * so we do not have to deal with hammer_ref() setting CHECK.
1279  */
1280 void
1281 hammer_ref_node(hammer_node_t node)
1282 {
1283         KKASSERT(hammer_isactive(&node->lock) && node->ondisk != NULL);
1284         hammer_ref(&node->lock);
1285 }
1286
1287 /*
1288  * Load a node's on-disk data reference.  Called with the node referenced
1289  * and interlocked.
1290  *
1291  * On return the node interlock will be unlocked.  If a non-zero error code
1292  * is returned the node will also be dereferenced (and the caller's pointer
1293  * will be stale).
1294  */
1295 static int
1296 hammer_load_node(hammer_transaction_t trans, hammer_node_t node, int isnew)
1297 {
1298         hammer_buffer_t buffer;
1299         hammer_off_t buf_offset;
1300         int error;
1301
1302         error = 0;
1303         if (node->ondisk == NULL) {
1304                 /*
1305                  * This is a little confusing but the jist is that
1306                  * node->buffer determines whether the node is on
1307                  * the buffer's node_list and node->ondisk determines
1308                  * whether the buffer is referenced.
1309                  *
1310                  * We could be racing a buffer release, in which case
1311                  * node->buffer may become NULL while we are blocked
1312                  * referencing the buffer.
1313                  */
1314                 if ((buffer = node->buffer) != NULL) {
1315                         error = hammer_ref_buffer(buffer);
1316                         if (error == 0 && node->buffer == NULL) {
1317                                 TAILQ_INSERT_TAIL(&buffer->node_list, node, entry);
1318                                 node->buffer = buffer;
1319                         }
1320                 } else {
1321                         buf_offset = node->node_offset & ~HAMMER_BUFMASK64;
1322                         buffer = hammer_get_buffer(node->hmp, buf_offset,
1323                                                    HAMMER_BUFSIZE, 0, &error);
1324                         if (buffer) {
1325                                 KKASSERT(error == 0);
1326                                 TAILQ_INSERT_TAIL(&buffer->node_list, node, entry);
1327                                 node->buffer = buffer;
1328                         }
1329                 }
1330                 if (error)
1331                         goto failed;
1332                 node->ondisk = (void *)((char *)buffer->ondisk +
1333                                         (node->node_offset & HAMMER_BUFMASK));
1334
1335                 /*
1336                  * Check CRC.  NOTE: Neither flag is set and the CRC is not
1337                  * generated on new B-Tree nodes.
1338                  */
1339                 if (isnew == 0 &&
1340                     (node->flags & HAMMER_NODE_CRCANY) == 0) {
1341                         if (hammer_crc_test_btree(node->ondisk) == 0) {
1342                                 hdkprintf("CRC B-TREE NODE @ %016llx/%lu FAILED\n",
1343                                         (long long)node->node_offset,
1344                                         sizeof(*node->ondisk));
1345                                 if (hammer_debug_critical)
1346                                         Debugger("CRC FAILED: B-TREE NODE");
1347                                 node->flags |= HAMMER_NODE_CRCBAD;
1348                         } else {
1349                                 node->flags |= HAMMER_NODE_CRCGOOD;
1350                         }
1351                 }
1352         }
1353         if (node->flags & HAMMER_NODE_CRCBAD) {
1354                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1355                         error = EDOM;
1356                 else
1357                         error = EIO;
1358         }
1359 failed:
1360         if (error) {
1361                 _hammer_rel_node(node, 1);
1362         } else {
1363                 hammer_ref_interlock_done(&node->lock);
1364         }
1365         return (error);
1366 }
1367
1368 /*
1369  * Safely reference a node, interlock against flushes via the IO subsystem.
1370  */
1371 hammer_node_t
1372 hammer_ref_node_safe(hammer_transaction_t trans, hammer_node_cache_t cache,
1373                      int *errorp)
1374 {
1375         hammer_node_t node;
1376         int doload;
1377
1378         node = cache->node;
1379         if (node != NULL) {
1380                 doload = hammer_ref_interlock(&node->lock);
1381                 if (doload) {
1382                         *errorp = hammer_load_node(trans, node, 0);
1383                         if (*errorp)
1384                                 node = NULL;
1385                 } else {
1386                         KKASSERT(node->ondisk);
1387                         if (node->flags & HAMMER_NODE_CRCBAD) {
1388                                 if (trans->flags & HAMMER_TRANSF_CRCDOM)
1389                                         *errorp = EDOM;
1390                                 else
1391                                         *errorp = EIO;
1392                                 _hammer_rel_node(node, 0);
1393                                 node = NULL;
1394                         } else {
1395                                 *errorp = 0;
1396                         }
1397                 }
1398         } else {
1399                 *errorp = ENOENT;
1400         }
1401         return(node);
1402 }
1403
1404 /*
1405  * Release a hammer_node.  On the last release the node dereferences
1406  * its underlying buffer and may or may not be destroyed.
1407  *
1408  * If locked is non-zero the passed node has been interlocked by the
1409  * caller and we are in the failure/unload path, otherwise it has not and
1410  * we are doing a normal release.
1411  *
1412  * This function will dispose of the interlock and the reference.
1413  * On return the node pointer is stale.
1414  */
1415 void
1416 _hammer_rel_node(hammer_node_t node, int locked)
1417 {
1418         hammer_buffer_t buffer;
1419
1420         /*
1421          * Deref the node.  If this isn't the 1->0 transition we're basically
1422          * done.  If locked is non-zero this function will just deref the
1423          * locked node and return 1, otherwise it will deref the locked
1424          * node and either lock and return 1 on the 1->0 transition or
1425          * not lock and return 0.
1426          */
1427         if (hammer_rel_interlock(&node->lock, locked) == 0)
1428                 return;
1429
1430         /*
1431          * Either locked was non-zero and we are interlocked, or the
1432          * hammer_rel_interlock() call returned non-zero and we are
1433          * interlocked.
1434          *
1435          * The ref-count must still be decremented if locked != 0 so
1436          * the cleanup required still varies a bit.
1437          *
1438          * hammer_flush_node() when called with 1 or 2 will dispose of
1439          * the lock and possible ref-count.
1440          */
1441         if (node->ondisk == NULL) {
1442                 hammer_flush_node(node, locked + 1);
1443                 /* node is stale now */
1444                 return;
1445         }
1446
1447         /*
1448          * Do not disassociate the node from the buffer if it represents
1449          * a modified B-Tree node that still needs its crc to be generated.
1450          */
1451         if (node->flags & HAMMER_NODE_NEEDSCRC) {
1452                 hammer_rel_interlock_done(&node->lock, locked);
1453                 return;
1454         }
1455
1456         /*
1457          * Do final cleanups and then either destroy the node and leave it
1458          * passively cached.  The buffer reference is removed regardless.
1459          */
1460         buffer = node->buffer;
1461         node->ondisk = NULL;
1462
1463         if ((node->flags & HAMMER_NODE_FLUSH) == 0) {
1464                 /*
1465                  * Normal release.
1466                  */
1467                 hammer_rel_interlock_done(&node->lock, locked);
1468         } else {
1469                 /*
1470                  * Destroy the node.
1471                  */
1472                 hammer_flush_node(node, locked + 1);
1473                 /* node is stale */
1474
1475         }
1476         hammer_rel_buffer(buffer, 0);
1477 }
1478
1479 void
1480 hammer_rel_node(hammer_node_t node)
1481 {
1482         _hammer_rel_node(node, 0);
1483 }
1484
1485 /*
1486  * Free space on-media associated with a B-Tree node.
1487  */
1488 void
1489 hammer_delete_node(hammer_transaction_t trans, hammer_node_t node)
1490 {
1491         KKASSERT((node->flags & HAMMER_NODE_DELETED) == 0);
1492         node->flags |= HAMMER_NODE_DELETED;
1493         hammer_blockmap_free(trans, node->node_offset, sizeof(*node->ondisk));
1494 }
1495
1496 /*
1497  * Passively cache a referenced hammer_node.  The caller may release
1498  * the node on return.
1499  */
1500 void
1501 hammer_cache_node(hammer_node_cache_t cache, hammer_node_t node)
1502 {
1503         /*
1504          * If the node doesn't exist, or is being deleted, don't cache it!
1505          *
1506          * The node can only ever be NULL in the I/O failure path.
1507          */
1508         if (node == NULL || (node->flags & HAMMER_NODE_DELETED))
1509                 return;
1510         if (cache->node == node)
1511                 return;
1512         while (cache->node)
1513                 hammer_uncache_node(cache);
1514         if (node->flags & HAMMER_NODE_DELETED)
1515                 return;
1516         cache->node = node;
1517         TAILQ_INSERT_TAIL(&node->cache_list, cache, entry);
1518 }
1519
1520 void
1521 hammer_uncache_node(hammer_node_cache_t cache)
1522 {
1523         hammer_node_t node;
1524
1525         if ((node = cache->node) != NULL) {
1526                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1527                 cache->node = NULL;
1528                 if (TAILQ_EMPTY(&node->cache_list))
1529                         hammer_flush_node(node, 0);
1530         }
1531 }
1532
1533 /*
1534  * Remove a node's cache references and destroy the node if it has no
1535  * other references or backing store.
1536  *
1537  * locked == 0  Normal unlocked operation
1538  * locked == 1  Call hammer_rel_interlock_done(..., 0);
1539  * locked == 2  Call hammer_rel_interlock_done(..., 1);
1540  *
1541  * XXX for now this isn't even close to being MPSAFE so the refs check
1542  *     is sufficient.
1543  */
1544 void
1545 hammer_flush_node(hammer_node_t node, int locked)
1546 {
1547         hammer_node_cache_t cache;
1548         hammer_buffer_t buffer;
1549         hammer_mount_t hmp = node->hmp;
1550         int dofree;
1551
1552         while ((cache = TAILQ_FIRST(&node->cache_list)) != NULL) {
1553                 TAILQ_REMOVE(&node->cache_list, cache, entry);
1554                 cache->node = NULL;
1555         }
1556
1557         /*
1558          * NOTE: refs is predisposed if another thread is blocking and
1559          *       will be larger than 0 in that case.  We aren't MPSAFE
1560          *       here.
1561          */
1562         if (node->ondisk == NULL && hammer_norefs(&node->lock)) {
1563                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1564                 RB_REMOVE(hammer_nod_rb_tree, &node->hmp->rb_nods_root, node);
1565                 if ((buffer = node->buffer) != NULL) {
1566                         node->buffer = NULL;
1567                         TAILQ_REMOVE(&buffer->node_list, node, entry);
1568                         /* buffer is unreferenced because ondisk is NULL */
1569                 }
1570                 dofree = 1;
1571         } else {
1572                 dofree = 0;
1573         }
1574
1575         /*
1576          * Deal with the interlock if locked == 1 or locked == 2.
1577          */
1578         if (locked)
1579                 hammer_rel_interlock_done(&node->lock, locked - 1);
1580
1581         /*
1582          * Destroy if requested
1583          */
1584         if (dofree) {
1585                 --hammer_count_nodes;
1586                 kfree(node, hmp->m_misc);
1587         }
1588 }
1589
1590 /*
1591  * Flush passively cached B-Tree nodes associated with this buffer.
1592  * This is only called when the buffer is about to be destroyed, so
1593  * none of the nodes should have any references.  The buffer is locked.
1594  *
1595  * We may be interlocked with the buffer.
1596  */
1597 void
1598 hammer_flush_buffer_nodes(hammer_buffer_t buffer)
1599 {
1600         hammer_node_t node;
1601
1602         while ((node = TAILQ_FIRST(&buffer->node_list)) != NULL) {
1603                 KKASSERT(node->ondisk == NULL);
1604                 KKASSERT((node->flags & HAMMER_NODE_NEEDSCRC) == 0);
1605
1606                 if (hammer_try_interlock_norefs(&node->lock)) {
1607                         hammer_ref(&node->lock);
1608                         node->flags |= HAMMER_NODE_FLUSH;
1609                         _hammer_rel_node(node, 1);
1610                 } else {
1611                         KKASSERT(node->buffer != NULL);
1612                         buffer = node->buffer;
1613                         node->buffer = NULL;
1614                         TAILQ_REMOVE(&buffer->node_list, node, entry);
1615                         /* buffer is unreferenced because ondisk is NULL */
1616                 }
1617         }
1618 }
1619
1620
1621 /************************************************************************
1622  *                              ALLOCATORS                              *
1623  ************************************************************************/
1624
1625 /*
1626  * Allocate a B-Tree node.
1627  */
1628 hammer_node_t
1629 hammer_alloc_btree(hammer_transaction_t trans, hammer_off_t hint, int *errorp)
1630 {
1631         hammer_buffer_t buffer = NULL;
1632         hammer_node_t node = NULL;
1633         hammer_off_t node_offset;
1634
1635         node_offset = hammer_blockmap_alloc(trans, HAMMER_ZONE_BTREE_INDEX,
1636                                             sizeof(struct hammer_node_ondisk),
1637                                             hint, errorp);
1638         if (*errorp == 0) {
1639                 node = hammer_get_node(trans, node_offset, 1, errorp);
1640                 hammer_modify_node_noundo(trans, node);
1641                 bzero(node->ondisk, sizeof(*node->ondisk));
1642                 hammer_modify_node_done(node);
1643         }
1644         if (buffer)
1645                 hammer_rel_buffer(buffer, 0);
1646         return(node);
1647 }
1648
1649 /*
1650  * Allocate data.  If the address of a data buffer is supplied then
1651  * any prior non-NULL *data_bufferp will be released and *data_bufferp
1652  * will be set to the related buffer.  The caller must release it when
1653  * finally done.  The initial *data_bufferp should be set to NULL by
1654  * the caller.
1655  *
1656  * The caller is responsible for making hammer_modify*() calls on the
1657  * *data_bufferp.
1658  */
1659 void *
1660 hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1661                   uint16_t rec_type, hammer_off_t *data_offsetp,
1662                   struct hammer_buffer **data_bufferp,
1663                   hammer_off_t hint, int *errorp)
1664 {
1665         void *data;
1666         int zone;
1667
1668         /*
1669          * Allocate data directly from blockmap.
1670          */
1671         if (data_len) {
1672                 switch(rec_type) {
1673                 case HAMMER_RECTYPE_INODE:
1674                 case HAMMER_RECTYPE_DIRENTRY:
1675                 case HAMMER_RECTYPE_EXT:
1676                 case HAMMER_RECTYPE_FIX:
1677                 case HAMMER_RECTYPE_PFS:
1678                 case HAMMER_RECTYPE_SNAPSHOT:
1679                 case HAMMER_RECTYPE_CONFIG:
1680                         zone = HAMMER_ZONE_META_INDEX;
1681                         break;
1682                 case HAMMER_RECTYPE_DATA:
1683                 case HAMMER_RECTYPE_DB:
1684                         /*
1685                          * Only mirror-write comes here.
1686                          * Regular allocation path uses blockmap reservation.
1687                          */
1688                         zone = hammer_data_zone_index(data_len);
1689                         if (zone == HAMMER_ZONE_LARGE_DATA_INDEX) {
1690                                 /* round up */
1691                                 data_len = (data_len + HAMMER_BUFMASK) &
1692                                            ~HAMMER_BUFMASK;
1693                         }
1694                         break;
1695                 default:
1696                         hpanic("rec_type %04x unknown", rec_type);
1697                         zone = HAMMER_ZONE_UNAVAIL_INDEX; /* NOT REACHED */
1698                         break;
1699                 }
1700                 *data_offsetp = hammer_blockmap_alloc(trans, zone, data_len,
1701                                                       hint, errorp);
1702         } else {
1703                 *data_offsetp = 0;
1704         }
1705
1706         data = NULL;
1707         if (*errorp == 0 && data_bufferp && data_len)
1708                 data = hammer_bread_ext(trans->hmp, *data_offsetp, data_len,
1709                                         errorp, data_bufferp);
1710         return(data);
1711 }
1712
1713 /*
1714  * Sync dirty buffers to the media and clean-up any loose ends.
1715  *
1716  * These functions do not start the flusher going, they simply
1717  * queue everything up to the flusher.
1718  */
1719 static int hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
1720
1721 struct hammer_sync_info {
1722         int error;
1723 };
1724
1725 int
1726 hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor)
1727 {
1728         struct hammer_sync_info info;
1729
1730         info.error = 0;
1731         if (waitfor == MNT_WAIT) {
1732                 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS,
1733                           hammer_sync_scan2, &info);
1734         } else {
1735                 vsyncscan(hmp->mp, VMSC_GETVP | VMSC_ONEPASS | VMSC_NOWAIT,
1736                           hammer_sync_scan2, &info);
1737         }
1738         return(info.error);
1739 }
1740
1741 /*
1742  * Filesystem sync.  If doing a synchronous sync make a second pass on
1743  * the vnodes in case any were already flushing during the first pass,
1744  * and activate the flusher twice (the second time brings the UNDO FIFO's
1745  * start position up to the end position after the first call).
1746  *
1747  * If doing a lazy sync make just one pass on the vnode list, ignoring
1748  * any new vnodes added to the list while the sync is in progress.
1749  */
1750 int
1751 hammer_sync_hmp(hammer_mount_t hmp, int waitfor)
1752 {
1753         struct hammer_sync_info info;
1754         int flags;
1755
1756         flags = VMSC_GETVP;
1757         if (waitfor & MNT_LAZY)
1758                 flags |= VMSC_ONEPASS;
1759
1760         info.error = 0;
1761         vsyncscan(hmp->mp, flags | VMSC_NOWAIT, hammer_sync_scan2, &info);
1762
1763         if (info.error == 0 && (waitfor & MNT_WAIT)) {
1764                 vsyncscan(hmp->mp, flags, hammer_sync_scan2, &info);
1765         }
1766         if (waitfor == MNT_WAIT) {
1767                 hammer_flusher_sync(hmp);
1768                 hammer_flusher_sync(hmp);
1769         } else {
1770                 hammer_flusher_async(hmp, NULL);
1771                 hammer_flusher_async(hmp, NULL);
1772         }
1773         return(info.error);
1774 }
1775
1776 static int
1777 hammer_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1778 {
1779         struct hammer_sync_info *info = data;
1780         struct hammer_inode *ip;
1781         int error;
1782
1783         ip = VTOI(vp);
1784         if (ip == NULL)
1785                 return(0);
1786         if (vp->v_type == VNON || vp->v_type == VBAD) {
1787                 vclrisdirty(vp);
1788                 return(0);
1789         }
1790         if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1791             RB_EMPTY(&vp->v_rbdirty_tree)) {
1792                 vclrisdirty(vp);
1793                 return(0);
1794         }
1795         error = VOP_FSYNC(vp, MNT_NOWAIT, 0);
1796         if (error)
1797                 info->error = error;
1798         return(0);
1799 }