sys/vfs/hammer/hammer_reblock.c

   1 /*
   2  * Copyright (c) 2008-2012 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 /*
  35  * HAMMER reblocker - This code frees up fragmented physical space
  36  *
  37  * HAMMER only keeps track of free space on a big-block basis.  A big-block
  38  * containing holes can only be freed by migrating the remaining data in
  39  * that big-block into a new big-block, then freeing the big-block.
  40  *
  41  * This function is called from an ioctl or via the hammer support thread.
  42  */
  43
  44 #include "hammer.h"
  45
  46 static int hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
  47                                  hammer_cursor_t cursor,
  48                                  hammer_btree_elm_t elm);
  49 static int hammer_reblock_data(struct hammer_ioc_reblock *reblock,
  50                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  51 static int hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
  52                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  53 static int hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
  54                                 hammer_cursor_t cursor, hammer_btree_elm_t elm);
  55 static void hammer_move_node(hammer_cursor_t cursor, hammer_btree_elm_t elm,
  56                                 hammer_node_t onode, hammer_node_t nnode);
  57
  58 int
  59 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
  60                    struct hammer_ioc_reblock *reblock)
  61 {
  62         struct hammer_cursor cursor;
  63         hammer_btree_elm_t elm;
  64         int checkspace_count;
  65         int error;
  66         int seq;
  67         int slop;
  68         uint32_t key_end_localization;
  69
  70         if ((reblock->key_beg.localization | reblock->key_end.localization) &
  71             HAMMER_LOCALIZE_PSEUDOFS_MASK) {
  72                 return(EINVAL);
  73         }
  74         if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
  75                 return(EINVAL);
  76         if (reblock->free_level < 0 ||
  77             reblock->free_level > HAMMER_BIGBLOCK_SIZE)
  78                 return(EINVAL);
  79
  80         /*
  81          * A fill_percentage <= 20% is considered an emergency.  free_level is
  82          * inverted from fill_percentage.
  83          */
  84         if (reblock->free_level >= HAMMER_BIGBLOCK_SIZE * 8 / 10)
  85                 slop = HAMMER_CHKSPC_EMERGENCY;
  86         else
  87                 slop = HAMMER_CHKSPC_REBLOCK;
  88
  89         /*
  90          * Ioctl caller has only set localization type to reblock.
  91          * Initialize cursor key localization with ip localization.
  92          */
  93         reblock->key_cur = reblock->key_beg;
  94         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
  95         if (reblock->allpfs == 0)
  96                 reblock->key_cur.localization |= ip->obj_localization;
  97
  98         key_end_localization = reblock->key_end.localization;
  99         key_end_localization &= HAMMER_LOCALIZE_MASK;
 100         if (reblock->allpfs == 0)
 101                 key_end_localization |= ip->obj_localization;
 102         else
 103                 key_end_localization |= pfs_to_lo(HAMMER_MAX_PFSID);
 104
 105         checkspace_count = 0;
 106         seq = trans->hmp->flusher.done;
 107 retry:
 108         error = hammer_init_cursor(trans, &cursor, NULL, NULL);
 109         if (error) {
 110                 hammer_done_cursor(&cursor);
 111                 goto failed;
 112         }
 113         cursor.key_beg.localization = reblock->key_cur.localization;
 114         cursor.key_beg.obj_id = reblock->key_cur.obj_id;
 115         cursor.key_beg.key = HAMMER_MIN_KEY;
 116         cursor.key_beg.create_tid = 1;
 117         cursor.key_beg.delete_tid = 0;
 118         cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
 119         cursor.key_beg.obj_type = 0;
 120
 121         cursor.key_end.localization = key_end_localization;
 122         cursor.key_end.obj_id = reblock->key_end.obj_id;
 123         cursor.key_end.key = HAMMER_MAX_KEY;
 124         cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
 125         cursor.key_end.delete_tid = 0;
 126         cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
 127         cursor.key_end.obj_type = 0;
 128
 129         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
 130         cursor.flags |= HAMMER_CURSOR_BACKEND;
 131         cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
 132
 133         /*
 134          * This flag allows the btree scan code to return internal nodes,
 135          * so we can reblock them in addition to the leafs.  Only specify it
 136          * if we intend to reblock B-Tree nodes.
 137          */
 138         if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
 139                 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
 140
 141         error = hammer_btree_first(&cursor);
 142         while (error == 0) {
 143                 /*
 144                  * Internal or Leaf node
 145                  */
 146                 KKASSERT(cursor.index < cursor.node->ondisk->count);
 147                 elm = &cursor.node->ondisk->elms[cursor.index];
 148                 reblock->key_cur.obj_id = elm->base.obj_id;
 149                 reblock->key_cur.localization = elm->base.localization;
 150
 151                 /*
 152                  * Yield to more important tasks
 153                  */
 154                 if ((error = hammer_signal_check(trans->hmp)) != 0)
 155                         break;
 156
 157                 /*
 158                  * If there is insufficient free space it may be due to
 159                  * reserved big-blocks, which flushing might fix.
 160                  *
 161                  * We must force a retest in case the unlocked cursor is
 162                  * moved to the end of the leaf, or moved to an internal
 163                  * node.
 164                  *
 165                  * WARNING: See warnings in hammer_unlock_cursor() function.
 166                  */
 167                 if (hammer_checkspace(trans->hmp, slop)) {
 168                         if (++checkspace_count == 10) {
 169                                 error = ENOSPC;
 170                                 break;
 171                         }
 172                         hammer_unlock_cursor(&cursor);
 173                         cursor.flags |= HAMMER_CURSOR_RETEST;
 174                         hammer_flusher_wait(trans->hmp, seq);
 175                         hammer_lock_cursor(&cursor);
 176                         seq = hammer_flusher_async(trans->hmp, NULL);
 177                         goto skip;
 178                 }
 179
 180                 /*
 181                  * Acquiring the sync_lock prevents the operation from
 182                  * crossing a synchronization boundary.
 183                  *
 184                  * NOTE: cursor.node may have changed on return.
 185                  *
 186                  * WARNING: See warnings in hammer_unlock_cursor() function.
 187                  */
 188                 hammer_sync_lock_sh(trans);
 189                 error = hammer_reblock_helper(reblock, &cursor, elm);
 190                 hammer_sync_unlock(trans);
 191
 192                 while (hammer_flusher_meta_halflimit(trans->hmp) ||
 193                        hammer_flusher_undo_exhausted(trans, 2)) {
 194                         hammer_unlock_cursor(&cursor);
 195                         hammer_flusher_wait(trans->hmp, seq);
 196                         hammer_lock_cursor(&cursor);
 197                         seq = hammer_flusher_async_one(trans->hmp);
 198                 }
 199
 200                 /*
 201                  * Setup for iteration, our cursor flags may be modified by
 202                  * other threads while we are unlocked.
 203                  */
 204                 cursor.flags |= HAMMER_CURSOR_ATEDISK;
 205
 206                 /*
 207                  * We allocate data buffers, which atm we don't track
 208                  * dirty levels for because we allow the kernel to write
 209                  * them.  But if we allocate too many we can still deadlock
 210                  * the buffer cache.
 211                  *
 212                  * WARNING: See warnings in hammer_unlock_cursor() function.
 213                  *          (The cursor's node and element may change!)
 214                  */
 215                 if (bd_heatup()) {
 216                         hammer_unlock_cursor(&cursor);
 217                         bwillwrite(HAMMER_XBUFSIZE);
 218                         hammer_lock_cursor(&cursor);
 219                 }
 220                 vm_wait_nominal();
 221 skip:
 222                 if (error == 0) {
 223                         error = hammer_btree_iterate(&cursor);
 224                 }
 225         }
 226         if (error == ENOENT)
 227                 error = 0;
 228         hammer_done_cursor(&cursor);
 229         if (error == EWOULDBLOCK) {
 230                 hammer_flusher_sync(trans->hmp);
 231                 goto retry;
 232         }
 233         if (error == EDEADLK)
 234                 goto retry;
 235         if (error == EINTR) {
 236                 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
 237                 error = 0;
 238         }
 239 failed:
 240         reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
 241         return(error);
 242 }
 243
 244 /*
 245  * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
 246  *
 247  * XXX We have no visibility into internal B-Tree nodes at the moment,
 248  * only leaf nodes.
 249  */
 250 static int
 251 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
 252                       hammer_cursor_t cursor, hammer_btree_elm_t elm)
 253 {
 254         hammer_mount_t hmp;
 255         hammer_off_t tmp_offset;
 256         hammer_node_ondisk_t ondisk;
 257         struct hammer_btree_leaf_elm leaf;
 258         int error;
 259         int bytes;
 260         int cur;
 261         int iocflags;
 262
 263         error = 0;
 264         hmp = cursor->trans->hmp;
 265
 266         /*
 267          * Reblock data.  Note that data embedded in a record is reblocked
 268          * by the record reblock code.  Data processing only occurs at leaf
 269          * nodes and for RECORD element types.
 270          */
 271         if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
 272                 goto skip;
 273         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 274                 return(EINVAL);
 275         tmp_offset = elm->leaf.data_offset;
 276         if (tmp_offset == 0)
 277                 goto skip;
 278
 279         /*
 280          * If reblock->vol_no is specified we only want to reblock data
 281          * in that volume, but ignore everything else.
 282          */
 283         if (reblock->vol_no != -1 &&
 284             reblock->vol_no != HAMMER_VOL_DECODE(tmp_offset))
 285                 goto skip;
 286
 287         /*
 288          * NOTE: Localization restrictions may also have been set-up, we can't
 289          *       just set the match flags willy-nilly here.
 290          */
 291         switch(elm->leaf.base.rec_type) {
 292         case HAMMER_RECTYPE_INODE:
 293         case HAMMER_RECTYPE_SNAPSHOT:
 294         case HAMMER_RECTYPE_CONFIG:
 295                 iocflags = HAMMER_IOC_DO_INODES;
 296                 break;
 297         case HAMMER_RECTYPE_EXT:
 298         case HAMMER_RECTYPE_FIX:
 299         case HAMMER_RECTYPE_PFS:
 300         case HAMMER_RECTYPE_DIRENTRY:
 301                 iocflags = HAMMER_IOC_DO_DIRS;
 302                 break;
 303         case HAMMER_RECTYPE_DATA:
 304         case HAMMER_RECTYPE_DB:
 305                 iocflags = HAMMER_IOC_DO_DATA;
 306                 break;
 307         default:
 308                 iocflags = 0;
 309                 break;
 310         }
 311         if (reblock->head.flags & iocflags) {
 312                 ++reblock->data_count;
 313                 reblock->data_byte_count += elm->leaf.data_len;
 314                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 315                 if (hammer_debug_general & 0x4000)
 316                         hdkprintf("D %6d/%d\n", bytes, reblock->free_level);
 317                 /*
 318                  * Start data reblock if
 319                  * 1. there is no error
 320                  * 2. the data and allocator offset are not in the same
 321                  *    big-block, or free level threshold is 0
 322                  * 3. free bytes in the data's big-block is larger than
 323                  *    free level threshold (means if threshold is 0 then
 324                  *    do reblock no matter what).
 325                  */
 326                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 327                     bytes >= reblock->free_level) {
 328                         /*
 329                          * This is nasty, the uncache code may have to get
 330                          * vnode locks and because of that we can't hold
 331                          * the cursor locked.
 332                          *
 333                          * WARNING: See warnings in hammer_unlock_cursor()
 334                          *          function.
 335                          */
 336                         leaf = elm->leaf;
 337                         hammer_unlock_cursor(cursor);
 338                         hammer_io_direct_uncache(hmp, &leaf);
 339                         hammer_lock_cursor(cursor);
 340
 341                         /*
 342                          * elm may have become stale or invalid, reload it.
 343                          * ondisk variable is temporary only.  Note that
 344                          * cursor->node and thus cursor->node->ondisk may
 345                          * also changed.
 346                          */
 347                         ondisk = cursor->node->ondisk;
 348                         elm = &ondisk->elms[cursor->index];
 349                         if (cursor->flags & HAMMER_CURSOR_RETEST) {
 350                                 hkprintf("debug: retest on reblocker uncache\n");
 351                                 error = EDEADLK;
 352                         } else if (ondisk->type != HAMMER_BTREE_TYPE_LEAF ||
 353                                    cursor->index >= ondisk->count) {
 354                                 hkprintf("debug: shifted on reblocker uncache\n");
 355                                 error = EDEADLK;
 356                         } else if (bcmp(&elm->leaf, &leaf, sizeof(leaf))) {
 357                                 hkprintf("debug: changed on reblocker uncache\n");
 358                                 error = EDEADLK;
 359                         }
 360                         if (error == 0)
 361                                 error = hammer_cursor_upgrade(cursor);
 362                         if (error == 0) {
 363                                 KKASSERT(cursor->index < ondisk->count);
 364                                 error = hammer_reblock_data(reblock,
 365                                                             cursor, elm);
 366                         }
 367                         if (error == 0) {
 368                                 ++reblock->data_moves;
 369                                 reblock->data_byte_moves += elm->leaf.data_len;
 370                         }
 371                 }
 372         }
 373
 374 skip:
 375         /*
 376          * Reblock a B-Tree internal or leaf node.  A leaf node is reblocked
 377          * on initial entry only (element 0).  An internal node is reblocked
 378          * when entered upward from its first leaf node only (also element 0,
 379          * see hammer_btree_iterate() where cursor moves up and may return).
 380          * Further revisits of the internal node (index > 0) are ignored.
 381          */
 382         tmp_offset = cursor->node->node_offset;
 383
 384         /*
 385          * If reblock->vol_no is specified we only want to reblock data
 386          * in that volume, but ignore everything else.
 387          */
 388         if (reblock->vol_no != -1 &&
 389             reblock->vol_no != HAMMER_VOL_DECODE(tmp_offset))
 390                 goto end;
 391
 392         if (cursor->index == 0 &&
 393             error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
 394                 ++reblock->btree_count;
 395                 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
 396                 if (hammer_debug_general & 0x4000)
 397                         hdkprintf("B %6d/%d\n", bytes, reblock->free_level);
 398                 /*
 399                  * Start node reblock if
 400                  * 1. there is no error
 401                  * 2. the node and allocator offset are not in the same
 402                  *    big-block, or free level threshold is 0
 403                  * 3. free bytes in the node's big-block is larger than
 404                  *    free level threshold (means if threshold is 0 then
 405                  *    do reblock no matter what).
 406                  */
 407                 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
 408                     bytes >= reblock->free_level) {
 409                         error = hammer_cursor_upgrade(cursor);
 410                         if (error == 0) {
 411                                 if (cursor->parent) {
 412                                         KKASSERT(cursor->parent_index <
 413                                                  cursor->parent->ondisk->count);
 414                                         elm = &cursor->parent->ondisk->elms[cursor->parent_index];
 415                                 } else {
 416                                         elm = NULL;
 417                                 }
 418                                 switch(cursor->node->ondisk->type) {
 419                                 case HAMMER_BTREE_TYPE_LEAF:
 420                                         error = hammer_reblock_leaf_node(
 421                                                         reblock, cursor, elm);
 422                                         break;
 423                                 case HAMMER_BTREE_TYPE_INTERNAL:
 424                                         error = hammer_reblock_int_node(
 425                                                         reblock, cursor, elm);
 426                                         break;
 427                                 default:
 428                                         hpanic("Illegal B-Tree node type");
 429                                 }
 430                         }
 431                         if (error == 0) {
 432                                 ++reblock->btree_moves;
 433                         }
 434                 }
 435         }
 436 end:
 437         hammer_cursor_downgrade(cursor);
 438         return(error);
 439 }
 440
 441 /*
 442  * Reblock a record's data.  Both the B-Tree element and record pointers
 443  * to the data must be adjusted.
 444  */
 445 static int
 446 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
 447                     hammer_cursor_t cursor, hammer_btree_elm_t elm)
 448 {
 449         struct hammer_buffer *data_buffer = NULL;
 450         hammer_off_t odata_offset;
 451         hammer_off_t ndata_offset;
 452         int error;
 453         void *ndata;
 454
 455         error = hammer_btree_extract_data(cursor);
 456         if (error)
 457                 return (error);
 458         ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
 459                                   elm->leaf.base.rec_type,
 460                                   &ndata_offset, &data_buffer,
 461                                   0, &error);
 462         if (error)
 463                 goto done;
 464         hammer_io_notmeta(data_buffer);
 465
 466         /*
 467          * Move the data.  Note that we must invalidate any cached
 468          * data buffer in the cursor before calling blockmap_free.
 469          * The blockmap_free may free up the entire big-block and
 470          * will not be able to invalidate it if the cursor is holding
 471          * a data buffer cached in that big-block.
 472          */
 473         hammer_modify_buffer_noundo(cursor->trans, data_buffer);
 474         bcopy(cursor->data, ndata, elm->leaf.data_len);
 475         hammer_modify_buffer_done(data_buffer);
 476         hammer_cursor_invalidate_cache(cursor);
 477
 478         hammer_blockmap_free(cursor->trans,
 479                              elm->leaf.data_offset, elm->leaf.data_len);
 480
 481         hammer_modify_node(cursor->trans, cursor->node,
 482                            &elm->leaf.data_offset, sizeof(hammer_off_t));
 483         odata_offset = elm->leaf.data_offset;
 484         elm->leaf.data_offset = ndata_offset;
 485         hammer_modify_node_done(cursor->node);
 486
 487         if (hammer_debug_general & 0x4000) {
 488                 hdkprintf("%08x %016llx -> %016llx\n",
 489                         (elm ? elm->base.localization : -1),
 490                         (long long)odata_offset,
 491                         (long long)ndata_offset);
 492         }
 493 done:
 494         if (data_buffer)
 495                 hammer_rel_buffer(data_buffer, 0);
 496         return (error);
 497 }
 498
 499 /*
 500  * Reblock a B-Tree leaf node.  The parent must be adjusted to point to
 501  * the new copy of the leaf node.
 502  *
 503  * elm is a pointer to the parent element pointing at cursor.node.
 504  */
 505 static int
 506 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
 507                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 508 {
 509         hammer_node_t onode;
 510         hammer_node_t nnode;
 511         int error;
 512
 513         /*
 514          * Don't supply a hint when allocating the leaf.  Fills are done
 515          * from the leaf upwards.
 516          */
 517         onode = cursor->node;
 518         nnode = hammer_alloc_btree(cursor->trans, 0, &error);
 519
 520         if (nnode == NULL)
 521                 return (error);
 522
 523         hammer_lock_ex(&nnode->lock);
 524         hammer_modify_node_noundo(cursor->trans, nnode);
 525
 526         hammer_move_node(cursor, elm, onode, nnode);
 527
 528         /*
 529          * Clean up.
 530          *
 531          * The new node replaces the current node in the cursor.  The cursor
 532          * expects it to be locked so leave it locked.  Discard onode.
 533          */
 534         hammer_cursor_replaced_node(onode, nnode);
 535         hammer_delete_node(cursor->trans, onode);
 536
 537         if (hammer_debug_general & 0x4000) {
 538                 hdkprintf("%08x %016llx -> %016llx\n",
 539                         (elm ? elm->base.localization : -1),
 540                         (long long)onode->node_offset,
 541                         (long long)nnode->node_offset);
 542         }
 543         hammer_modify_node_done(nnode);
 544         cursor->node = nnode;
 545
 546         hammer_unlock(&onode->lock);
 547         hammer_rel_node(onode);
 548
 549         return (error);
 550 }
 551
 552 /*
 553  * Reblock a B-Tree internal node.  The parent must be adjusted to point to
 554  * the new copy of the internal node, and the node's children's parent
 555  * pointers must also be adjusted to point to the new copy.
 556  *
 557  * elm is a pointer to the parent element pointing at cursor.node.
 558  */
 559 static int
 560 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
 561                          hammer_cursor_t cursor, hammer_btree_elm_t elm)
 562 {
 563         struct hammer_node_lock lockroot;
 564         hammer_node_t onode;
 565         hammer_node_t nnode;
 566         int error;
 567
 568         hammer_node_lock_init(&lockroot, cursor->node);
 569         error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
 570         if (error)
 571                 goto done;
 572
 573         /*
 574          * Don't supply a hint when allocating the leaf.  Fills are done
 575          * from the leaf upwards.
 576          */
 577         onode = cursor->node;
 578         nnode = hammer_alloc_btree(cursor->trans, 0, &error);
 579
 580         if (nnode == NULL)
 581                 goto done;
 582
 583         hammer_lock_ex(&nnode->lock);
 584         hammer_modify_node_noundo(cursor->trans, nnode);
 585
 586         hammer_move_node(cursor, elm, onode, nnode);
 587
 588         /*
 589          * Clean up.
 590          *
 591          * The new node replaces the current node in the cursor.  The cursor
 592          * expects it to be locked so leave it locked.  Discard onode.
 593          */
 594         hammer_cursor_replaced_node(onode, nnode);
 595         hammer_delete_node(cursor->trans, onode);
 596
 597         if (hammer_debug_general & 0x4000) {
 598                 hdkprintf("%08x %016llx -> %016llx\n",
 599                         (elm ? elm->base.localization : -1),
 600                         (long long)onode->node_offset,
 601                         (long long)nnode->node_offset);
 602         }
 603         hammer_modify_node_done(nnode);
 604         cursor->node = nnode;
 605
 606         hammer_unlock(&onode->lock);
 607         hammer_rel_node(onode);
 608
 609 done:
 610         hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
 611         return (error);
 612 }
 613
 614 /*
 615  * nnode is a newly allocated node, and now elm becomes the node
 616  * element within nnode's parent that represents a pointer to nnode,
 617  * or nnode becomes the root node if elm does not exist.
 618  */
 619 static void
 620 hammer_move_node(hammer_cursor_t cursor, hammer_btree_elm_t elm,
 621                  hammer_node_t onode, hammer_node_t nnode)
 622 {
 623         int error, i;
 624
 625         bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
 626
 627         /*
 628          * Adjust the parent's pointer to us first.
 629          */
 630         if (elm) {
 631                 /*
 632                  * We are not the root of the B-Tree
 633                  */
 634                 KKASSERT(hammer_is_internal_node_elm(elm));
 635                 hammer_modify_node(cursor->trans, cursor->parent,
 636                                    &elm->internal.subtree_offset,
 637                                    sizeof(elm->internal.subtree_offset));
 638                 elm->internal.subtree_offset = nnode->node_offset;
 639                 hammer_modify_node_done(cursor->parent);
 640         } else {
 641                 /*
 642                  * We are the root of the B-Tree
 643                  */
 644                 hammer_volume_t volume;
 645                 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
 646                 KKASSERT(error == 0);
 647
 648                 hammer_modify_volume_field(cursor->trans, volume,
 649                                            vol0_btree_root);
 650                 volume->ondisk->vol0_btree_root = nnode->node_offset;
 651                 hammer_modify_volume_done(volume);
 652                 hammer_rel_volume(volume, 0);
 653         }
 654
 655         /*
 656          * Now adjust our children's pointers to us
 657          * if we are an internal node.
 658          */
 659         if (nnode->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
 660                 for (i = 0; i < nnode->ondisk->count; ++i) {
 661                         error = btree_set_parent_of_child(cursor->trans, nnode,
 662                                         &nnode->ondisk->elms[i]);
 663                         if (error)
 664                                 hpanic("reblock internal node: fixup problem");
 665                 }
 666         }
 667 }