sys/vfs/hammer/hammer_btree.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_btree.c,v 1.76 2008/08/06 15:38:58 dillon Exp $
  35  */
  36
  37 /*
  38  * HAMMER B-Tree index
  39  *
  40  * HAMMER implements a modified B+Tree.  In documentation this will
  41  * simply be refered to as the HAMMER B-Tree.  Basically a HAMMER B-Tree
  42  * looks like a B+Tree (A B-Tree which stores its records only at the leafs
  43  * of the tree), but adds two additional boundary elements which describe
  44  * the left-most and right-most element a node is able to represent.  In
  45  * otherwords, we have boundary elements at the two ends of a B-Tree node
  46  * instead of sub-tree pointers.
  47  *
  48  * A B-Tree internal node looks like this:
  49  *
  50  *      B N N N N N N B   <-- boundary and internal elements
  51  *       S S S S S S S    <-- subtree pointers
  52  *
  53  * A B-Tree leaf node basically looks like this:
  54  *
  55  *      L L L L L L L L   <-- leaf elemenets
  56  *
  57  * The radix for an internal node is 1 less then a leaf but we get a
  58  * number of significant benefits for our troubles.
  59  *
  60  * The big benefit to using a B-Tree containing boundary information
  61  * is that it is possible to cache pointers into the middle of the tree
  62  * and not have to start searches, insertions, OR deletions at the root
  63  * node.   In particular, searches are able to progress in a definitive
  64  * direction from any point in the tree without revisting nodes.  This
  65  * greatly improves the efficiency of many operations, most especially
  66  * record appends.
  67  *
  68  * B-Trees also make the stacking of trees fairly straightforward.
  69  *
  70  * INSERTIONS:  A search performed with the intention of doing
  71  * an insert will guarantee that the terminal leaf node is not full by
  72  * splitting full nodes.  Splits occur top-down during the dive down the
  73  * B-Tree.
  74  *
  75  * DELETIONS: A deletion makes no attempt to proactively balance the
  76  * tree and will recursively remove nodes that become empty.  If a
  77  * deadlock occurs a deletion may not be able to remove an empty leaf.
  78  * Deletions never allow internal nodes to become empty (that would blow
  79  * up the boundaries).
  80  */
  81 #include "hammer.h"
  82 #include <sys/buf.h>
  83 #include <sys/buf2.h>
  84
  85 static int btree_search(hammer_cursor_t cursor, int flags);
  86 static int btree_split_internal(hammer_cursor_t cursor);
  87 static int btree_split_leaf(hammer_cursor_t cursor);
  88 static int btree_remove(hammer_cursor_t cursor);
  89 static int btree_node_is_full(hammer_node_ondisk_t node);
  90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
  91                         hammer_tid_t mirror_tid);
  92 static void hammer_make_separator(hammer_base_elm_t key1,
  93                         hammer_base_elm_t key2, hammer_base_elm_t dest);
  94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
  95
  96 /*
  97  * Iterate records after a search.  The cursor is iterated forwards past
  98  * the current record until a record matching the key-range requirements
  99  * is found.  ENOENT is returned if the iteration goes past the ending
 100  * key.
 101  *
 102  * The iteration is inclusive of key_beg and can be inclusive or exclusive
 103  * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
 104  *
 105  * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
 106  * may be modified by B-Tree functions.
 107  *
 108  * cursor->key_beg may or may not be modified by this function during
 109  * the iteration.  XXX future - in case of an inverted lock we may have
 110  * to reinitiate the lookup and set key_beg to properly pick up where we
 111  * left off.
 112  *
 113  * NOTE!  EDEADLK *CANNOT* be returned by this procedure.
 114  */
 115 int
 116 hammer_btree_iterate(hammer_cursor_t cursor)
 117 {
 118         hammer_node_ondisk_t node;
 119         hammer_btree_elm_t elm;
 120         int error = 0;
 121         int r;
 122         int s;
 123
 124         /*
 125          * Skip past the current record
 126          */
 127         node = cursor->node->ondisk;
 128         if (node == NULL)
 129                 return(ENOENT);
 130         if (cursor->index < node->count &&
 131             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 132                 ++cursor->index;
 133         }
 134
 135         /*
 136          * Loop until an element is found or we are done.
 137          */
 138         for (;;) {
 139                 /*
 140                  * We iterate up the tree and then index over one element
 141                  * while we are at the last element in the current node.
 142                  *
 143                  * If we are at the root of the filesystem, cursor_up
 144                  * returns ENOENT.
 145                  *
 146                  * XXX this could be optimized by storing the information in
 147                  * the parent reference.
 148                  *
 149                  * XXX we can lose the node lock temporarily, this could mess
 150                  * up our scan.
 151                  */
 152                 ++hammer_stats_btree_iterations;
 153                 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
 154
 155                 if (cursor->index == node->count) {
 156                         if (hammer_debug_btree) {
 157                                 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
 158                                         cursor->node->node_offset,
 159                                         cursor->index,
 160                                         (cursor->parent ? cursor->parent->node_offset : -1),
 161                                         cursor->parent_index,
 162                                         curthread);
 163                         }
 164                         KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
 165                         error = hammer_cursor_up(cursor);
 166                         if (error)
 167                                 break;
 168                         /* reload stale pointer */
 169                         node = cursor->node->ondisk;
 170                         KKASSERT(cursor->index != node->count);
 171
 172                         /*
 173                          * If we are reblocking we want to return internal
 174                          * nodes.  Note that the internal node will be
 175                          * returned multiple times, on each upward recursion
 176                          * from its children.  The caller selects which
 177                          * revisit it cares about (usually first or last only).
 178                          */
 179                         if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
 180                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 181                                 return(0);
 182                         }
 183                         ++cursor->index;
 184                         continue;
 185                 }
 186
 187                 /*
 188                  * Check internal or leaf element.  Determine if the record
 189                  * at the cursor has gone beyond the end of our range.
 190                  *
 191                  * We recurse down through internal nodes.
 192                  */
 193                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 194                         elm = &node->elms[cursor->index];
 195
 196                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 197                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 198                         if (hammer_debug_btree) {
 199                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
 200                                         cursor->node->node_offset,
 201                                         cursor->index,
 202                                         elm[0].internal.base.obj_id,
 203                                         elm[0].internal.base.rec_type,
 204                                         elm[0].internal.base.key,
 205                                         elm[0].internal.base.localization,
 206                                         r,
 207                                         curthread
 208                                 );
 209                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 210                                         cursor->node->node_offset,
 211                                         cursor->index + 1,
 212                                         elm[1].internal.base.obj_id,
 213                                         elm[1].internal.base.rec_type,
 214                                         elm[1].internal.base.key,
 215                                         elm[1].internal.base.localization,
 216                                         s
 217                                 );
 218                         }
 219
 220                         if (r < 0) {
 221                                 error = ENOENT;
 222                                 break;
 223                         }
 224                         if (r == 0 && (cursor->flags &
 225                                        HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 226                                 error = ENOENT;
 227                                 break;
 228                         }
 229                         KKASSERT(s <= 0);
 230
 231                         /*
 232                          * Better not be zero
 233                          */
 234                         KKASSERT(elm->internal.subtree_offset != 0);
 235
 236                         /*
 237                          * If running the mirror filter see if we can skip
 238                          * one or more entire sub-trees.  If we can we
 239                          * return the internal mode and the caller processes
 240                          * the skipped range (see mirror_read)
 241                          */
 242                         if (cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) {
 243                                 if (elm->internal.mirror_tid <
 244                                     cursor->cmirror->mirror_tid) {
 245                                         hammer_cursor_mirror_filter(cursor);
 246                                         return(0);
 247                                 }
 248                         }
 249
 250                         error = hammer_cursor_down(cursor);
 251                         if (error)
 252                                 break;
 253                         KKASSERT(cursor->index == 0);
 254                         /* reload stale pointer */
 255                         node = cursor->node->ondisk;
 256                         continue;
 257                 } else {
 258                         elm = &node->elms[cursor->index];
 259                         r = hammer_btree_cmp(&cursor->key_end, &elm->base);
 260                         if (hammer_debug_btree) {
 261                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
 262                                         cursor->node->node_offset,
 263                                         cursor->index,
 264                                         (elm[0].leaf.base.btype ?
 265                                          elm[0].leaf.base.btype : '?'),
 266                                         elm[0].leaf.base.obj_id,
 267                                         elm[0].leaf.base.rec_type,
 268                                         elm[0].leaf.base.key,
 269                                         elm[0].leaf.base.localization,
 270                                         r
 271                                 );
 272                         }
 273                         if (r < 0) {
 274                                 error = ENOENT;
 275                                 break;
 276                         }
 277
 278                         /*
 279                          * We support both end-inclusive and
 280                          * end-exclusive searches.
 281                          */
 282                         if (r == 0 &&
 283                            (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 284                                 error = ENOENT;
 285                                 break;
 286                         }
 287
 288                         switch(elm->leaf.base.btype) {
 289                         case HAMMER_BTREE_TYPE_RECORD:
 290                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 291                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 292                                         ++cursor->index;
 293                                         continue;
 294                                 }
 295                                 error = 0;
 296                                 break;
 297                         default:
 298                                 error = EINVAL;
 299                                 break;
 300                         }
 301                         if (error)
 302                                 break;
 303                 }
 304                 /*
 305                  * node pointer invalid after loop
 306                  */
 307
 308                 /*
 309                  * Return entry
 310                  */
 311                 if (hammer_debug_btree) {
 312                         int i = cursor->index;
 313                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 314                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
 315                                 cursor->node, i,
 316                                 elm->internal.base.obj_id,
 317                                 elm->internal.base.rec_type,
 318                                 elm->internal.base.key,
 319                                 elm->internal.base.localization
 320                         );
 321                 }
 322                 return(0);
 323         }
 324         return(error);
 325 }
 326
 327 /*
 328  * We hit an internal element that we could skip as part of a mirroring
 329  * scan.  Calculate the entire range being skipped.
 330  *
 331  * It is important to include any gaps between the parent's left_bound
 332  * and the node's left_bound, and same goes for the right side.
 333  */
 334 static void
 335 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
 336 {
 337         struct hammer_cmirror *cmirror;
 338         hammer_node_ondisk_t ondisk;
 339         hammer_btree_elm_t elm;
 340
 341         ondisk = cursor->node->ondisk;
 342         cmirror = cursor->cmirror;
 343
 344         /*
 345          * Calculate the skipped range
 346          */
 347         elm = &ondisk->elms[cursor->index];
 348         if (cursor->index == 0)
 349                 cmirror->skip_beg = *cursor->left_bound;
 350         else
 351                 cmirror->skip_beg = elm->internal.base;
 352         while (cursor->index < ondisk->count) {
 353                 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
 354                         break;
 355                 ++cursor->index;
 356                 ++elm;
 357         }
 358         if (cursor->index == ondisk->count)
 359                 cmirror->skip_end = *cursor->right_bound;
 360         else
 361                 cmirror->skip_end = elm->internal.base;
 362
 363         /*
 364          * clip the returned result.
 365          */
 366         if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
 367                 cmirror->skip_beg = cursor->key_beg;
 368         if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
 369                 cmirror->skip_end = cursor->key_end;
 370 }
 371
 372 /*
 373  * Iterate in the reverse direction.  This is used by the pruning code to
 374  * avoid overlapping records.
 375  */
 376 int
 377 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
 378 {
 379         hammer_node_ondisk_t node;
 380         hammer_btree_elm_t elm;
 381         int error = 0;
 382         int r;
 383         int s;
 384
 385         /* mirror filtering not supported for reverse iteration */
 386         KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
 387
 388         /*
 389          * Skip past the current record.  For various reasons the cursor
 390          * may end up set to -1 or set to point at the end of the current
 391          * node.  These cases must be addressed.
 392          */
 393         node = cursor->node->ondisk;
 394         if (node == NULL)
 395                 return(ENOENT);
 396         if (cursor->index != -1 &&
 397             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 398                 --cursor->index;
 399         }
 400         if (cursor->index == cursor->node->ondisk->count)
 401                 --cursor->index;
 402
 403         /*
 404          * Loop until an element is found or we are done.
 405          */
 406         for (;;) {
 407                 ++hammer_stats_btree_iterations;
 408                 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
 409
 410                 /*
 411                  * We iterate up the tree and then index over one element
 412                  * while we are at the last element in the current node.
 413                  */
 414                 if (cursor->index == -1) {
 415                         error = hammer_cursor_up(cursor);
 416                         if (error) {
 417                                 cursor->index = 0; /* sanity */
 418                                 break;
 419                         }
 420                         /* reload stale pointer */
 421                         node = cursor->node->ondisk;
 422                         KKASSERT(cursor->index != node->count);
 423                         --cursor->index;
 424                         continue;
 425                 }
 426
 427                 /*
 428                  * Check internal or leaf element.  Determine if the record
 429                  * at the cursor has gone beyond the end of our range.
 430                  *
 431                  * We recurse down through internal nodes.
 432                  */
 433                 KKASSERT(cursor->index != node->count);
 434                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 435                         elm = &node->elms[cursor->index];
 436                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 437                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 438                         if (hammer_debug_btree) {
 439                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 440                                         cursor->node->node_offset,
 441                                         cursor->index,
 442                                         elm[0].internal.base.obj_id,
 443                                         elm[0].internal.base.rec_type,
 444                                         elm[0].internal.base.key,
 445                                         elm[0].internal.base.localization,
 446                                         r
 447                                 );
 448                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 449                                         cursor->node->node_offset,
 450                                         cursor->index + 1,
 451                                         elm[1].internal.base.obj_id,
 452                                         elm[1].internal.base.rec_type,
 453                                         elm[1].internal.base.key,
 454                                         elm[1].internal.base.localization,
 455                                         s
 456                                 );
 457                         }
 458
 459                         if (s >= 0) {
 460                                 error = ENOENT;
 461                                 break;
 462                         }
 463                         KKASSERT(r >= 0);
 464
 465                         /*
 466                          * Better not be zero
 467                          */
 468                         KKASSERT(elm->internal.subtree_offset != 0);
 469
 470                         error = hammer_cursor_down(cursor);
 471                         if (error)
 472                                 break;
 473                         KKASSERT(cursor->index == 0);
 474                         /* reload stale pointer */
 475                         node = cursor->node->ondisk;
 476
 477                         /* this can assign -1 if the leaf was empty */
 478                         cursor->index = node->count - 1;
 479                         continue;
 480                 } else {
 481                         elm = &node->elms[cursor->index];
 482                         s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
 483                         if (hammer_debug_btree) {
 484                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
 485                                         cursor->node->node_offset,
 486                                         cursor->index,
 487                                         (elm[0].leaf.base.btype ?
 488                                          elm[0].leaf.base.btype : '?'),
 489                                         elm[0].leaf.base.obj_id,
 490                                         elm[0].leaf.base.rec_type,
 491                                         elm[0].leaf.base.key,
 492                                         elm[0].leaf.base.localization,
 493                                         s
 494                                 );
 495                         }
 496                         if (s > 0) {
 497                                 error = ENOENT;
 498                                 break;
 499                         }
 500
 501                         switch(elm->leaf.base.btype) {
 502                         case HAMMER_BTREE_TYPE_RECORD:
 503                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 504                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 505                                         --cursor->index;
 506                                         continue;
 507                                 }
 508                                 error = 0;
 509                                 break;
 510                         default:
 511                                 error = EINVAL;
 512                                 break;
 513                         }
 514                         if (error)
 515                                 break;
 516                 }
 517                 /*
 518                  * node pointer invalid after loop
 519                  */
 520
 521                 /*
 522                  * Return entry
 523                  */
 524                 if (hammer_debug_btree) {
 525                         int i = cursor->index;
 526                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 527                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
 528                                 cursor->node, i,
 529                                 elm->internal.base.obj_id,
 530                                 elm->internal.base.rec_type,
 531                                 elm->internal.base.key,
 532                                 elm->internal.base.localization
 533                         );
 534                 }
 535                 return(0);
 536         }
 537         return(error);
 538 }
 539
 540 /*
 541  * Lookup cursor->key_beg.  0 is returned on success, ENOENT if the entry
 542  * could not be found, EDEADLK if inserting and a retry is needed, and a
 543  * fatal error otherwise.  When retrying, the caller must terminate the
 544  * cursor and reinitialize it.  EDEADLK cannot be returned if not inserting.
 545  *
 546  * The cursor is suitably positioned for a deletion on success, and suitably
 547  * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
 548  * specified.
 549  *
 550  * The cursor may begin anywhere, the search will traverse the tree in
 551  * either direction to locate the requested element.
 552  *
 553  * Most of the logic implementing historical searches is handled here.  We
 554  * do an initial lookup with create_tid set to the asof TID.  Due to the
 555  * way records are laid out, a backwards iteration may be required if
 556  * ENOENT is returned to locate the historical record.  Here's the
 557  * problem:
 558  *
 559  * create_tid:    10      15       20
 560  *                   LEAF1   LEAF2
 561  * records:         (11)        (18)
 562  *
 563  * Lets say we want to do a lookup AS-OF timestamp 17.  We will traverse
 564  * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
 565  * not visible and thus causes ENOENT to be returned.  We really need
 566  * to check record 11 in LEAF1.  If it also fails then the search fails
 567  * (e.g. it might represent the range 11-16 and thus still not match our
 568  * AS-OF timestamp of 17).  Note that LEAF1 could be empty, requiring
 569  * further iterations.
 570  *
 571  * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
 572  * and the cursor->create_check TID if an iteration might be needed.
 573  * In the above example create_check would be set to 14.
 574  */
 575 int
 576 hammer_btree_lookup(hammer_cursor_t cursor)
 577 {
 578         int error;
 579
 580         KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
 581                   cursor->trans->sync_lock_refs > 0);
 582         ++hammer_stats_btree_lookups;
 583         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 584                 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
 585                 cursor->key_beg.create_tid = cursor->asof;
 586                 for (;;) {
 587                         cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
 588                         error = btree_search(cursor, 0);
 589                         if (error != ENOENT ||
 590                             (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
 591                                 /*
 592                                  * Stop if no error.
 593                                  * Stop if error other then ENOENT.
 594                                  * Stop if ENOENT and not special case.
 595                                  */
 596                                 break;
 597                         }
 598                         if (hammer_debug_btree) {
 599                                 kprintf("CREATE_CHECK %016llx\n",
 600                                         cursor->create_check);
 601                         }
 602                         cursor->key_beg.create_tid = cursor->create_check;
 603                         /* loop */
 604                 }
 605         } else {
 606                 error = btree_search(cursor, 0);
 607         }
 608         if (error == 0)
 609                 error = hammer_btree_extract(cursor, cursor->flags);
 610         return(error);
 611 }
 612
 613 /*
 614  * Execute the logic required to start an iteration.  The first record
 615  * located within the specified range is returned and iteration control
 616  * flags are adjusted for successive hammer_btree_iterate() calls.
 617  */
 618 int
 619 hammer_btree_first(hammer_cursor_t cursor)
 620 {
 621         int error;
 622
 623         error = hammer_btree_lookup(cursor);
 624         if (error == ENOENT) {
 625                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 626                 error = hammer_btree_iterate(cursor);
 627         }
 628         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 629         return(error);
 630 }
 631
 632 /*
 633  * Similarly but for an iteration in the reverse direction.
 634  *
 635  * Set ATEDISK when iterating backwards to skip the current entry,
 636  * which after an ENOENT lookup will be pointing beyond our end point.
 637  */
 638 int
 639 hammer_btree_last(hammer_cursor_t cursor)
 640 {
 641         struct hammer_base_elm save;
 642         int error;
 643
 644         save = cursor->key_beg;
 645         cursor->key_beg = cursor->key_end;
 646         error = hammer_btree_lookup(cursor);
 647         cursor->key_beg = save;
 648         if (error == ENOENT ||
 649             (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 650                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 651                 error = hammer_btree_iterate_reverse(cursor);
 652         }
 653         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 654         return(error);
 655 }
 656
 657 /*
 658  * Extract the record and/or data associated with the cursor's current
 659  * position.  Any prior record or data stored in the cursor is replaced.
 660  * The cursor must be positioned at a leaf node.
 661  *
 662  * NOTE: All extractions occur at the leaf of the B-Tree.
 663  */
 664 int
 665 hammer_btree_extract(hammer_cursor_t cursor, int flags)
 666 {
 667         hammer_node_ondisk_t node;
 668         hammer_btree_elm_t elm;
 669         hammer_off_t data_off;
 670         hammer_mount_t hmp;
 671         int32_t data_len;
 672         int error;
 673
 674         /*
 675          * The case where the data reference resolves to the same buffer
 676          * as the record reference must be handled.
 677          */
 678         node = cursor->node->ondisk;
 679         elm = &node->elms[cursor->index];
 680         cursor->data = NULL;
 681         hmp = cursor->node->hmp;
 682
 683         /*
 684          * There is nothing to extract for an internal element.
 685          */
 686         if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
 687                 return(EINVAL);
 688
 689         /*
 690          * Only record types have data.
 691          */
 692         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 693         cursor->leaf = &elm->leaf;
 694
 695         if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
 696                 return(0);
 697         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 698                 return(0);
 699         data_off = elm->leaf.data_offset;
 700         data_len = elm->leaf.data_len;
 701         if (data_off == 0)
 702                 return(0);
 703
 704         /*
 705          * Load the data
 706          */
 707         KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
 708         cursor->data = hammer_bread_ext(hmp, data_off, data_len,
 709                                         &error, &cursor->data_buffer);
 710         if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
 711                 kprintf("CRC DATA @ %016llx/%d FAILED\n",
 712                         elm->leaf.data_offset, elm->leaf.data_len);
 713                 if (hammer_debug_debug & 0x0001)
 714                         Debugger("CRC FAILED: DATA");
 715                 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
 716                         error = EDOM;   /* less critical (mirroring) */
 717                 else
 718                         error = EIO;    /* critical */
 719         }
 720         return(error);
 721 }
 722
 723
 724 /*
 725  * Insert a leaf element into the B-Tree at the current cursor position.
 726  * The cursor is positioned such that the element at and beyond the cursor
 727  * are shifted to make room for the new record.
 728  *
 729  * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
 730  * flag set and that call must return ENOENT before this function can be
 731  * called.
 732  *
 733  * The caller may depend on the cursor's exclusive lock after return to
 734  * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
 735  *
 736  * ENOSPC is returned if there is no room to insert a new record.
 737  */
 738 int
 739 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
 740                     int *doprop)
 741 {
 742         hammer_node_ondisk_t node;
 743         int i;
 744         int error;
 745
 746         *doprop = 0;
 747         if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
 748                 return(error);
 749         ++hammer_stats_btree_inserts;
 750
 751         /*
 752          * Insert the element at the leaf node and update the count in the
 753          * parent.  It is possible for parent to be NULL, indicating that
 754          * the filesystem's ROOT B-Tree node is a leaf itself, which is
 755          * possible.  The root inode can never be deleted so the leaf should
 756          * never be empty.
 757          *
 758          * Remember that the right-hand boundary is not included in the
 759          * count.
 760          */
 761         hammer_modify_node_all(cursor->trans, cursor->node);
 762         node = cursor->node->ondisk;
 763         i = cursor->index;
 764         KKASSERT(elm->base.btype != 0);
 765         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 766         KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
 767         if (i != node->count) {
 768                 bcopy(&node->elms[i], &node->elms[i+1],
 769                       (node->count - i) * sizeof(*elm));
 770         }
 771         node->elms[i].leaf = *elm;
 772         ++node->count;
 773         hammer_cursor_inserted_element(cursor->node, i);
 774
 775         /*
 776          * Update the leaf node's aggregate mirror_tid for mirroring
 777          * support.
 778          */
 779         if (node->mirror_tid < elm->base.delete_tid) {
 780                 node->mirror_tid = elm->base.delete_tid;
 781                 *doprop = 1;
 782         }
 783         if (node->mirror_tid < elm->base.create_tid) {
 784                 node->mirror_tid = elm->base.create_tid;
 785                 *doprop = 1;
 786         }
 787         hammer_modify_node_done(cursor->node);
 788
 789         /*
 790          * Debugging sanity checks.
 791          */
 792         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
 793         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
 794         if (i) {
 795                 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
 796         }
 797         if (i != node->count - 1)
 798                 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
 799
 800         return(0);
 801 }
 802
 803 /*
 804  * Delete a record from the B-Tree at the current cursor position.
 805  * The cursor is positioned such that the current element is the one
 806  * to be deleted.
 807  *
 808  * On return the cursor will be positioned after the deleted element and
 809  * MAY point to an internal node.  It will be suitable for the continuation
 810  * of an iteration but not for an insertion or deletion.
 811  *
 812  * Deletions will attempt to partially rebalance the B-Tree in an upward
 813  * direction, but will terminate rather then deadlock.  Empty internal nodes
 814  * are never allowed by a deletion which deadlocks may end up giving us an
 815  * empty leaf.  The pruner will clean up and rebalance the tree.
 816  *
 817  * This function can return EDEADLK, requiring the caller to retry the
 818  * operation after clearing the deadlock.
 819  */
 820 int
 821 hammer_btree_delete(hammer_cursor_t cursor)
 822 {
 823         hammer_node_ondisk_t ondisk;
 824         hammer_node_t node;
 825         hammer_node_t parent;
 826         int error;
 827         int i;
 828
 829         KKASSERT (cursor->trans->sync_lock_refs > 0);
 830         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 831                 return(error);
 832         ++hammer_stats_btree_deletes;
 833
 834         /*
 835          * Delete the element from the leaf node.
 836          *
 837          * Remember that leaf nodes do not have boundaries.
 838          */
 839         node = cursor->node;
 840         ondisk = node->ondisk;
 841         i = cursor->index;
 842
 843         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
 844         KKASSERT(i >= 0 && i < ondisk->count);
 845         hammer_modify_node_all(cursor->trans, node);
 846         if (i + 1 != ondisk->count) {
 847                 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
 848                       (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
 849         }
 850         --ondisk->count;
 851         hammer_modify_node_done(node);
 852         hammer_cursor_deleted_element(node, i);
 853
 854         /*
 855          * Validate local parent
 856          */
 857         if (ondisk->parent) {
 858                 parent = cursor->parent;
 859
 860                 KKASSERT(parent != NULL);
 861                 KKASSERT(parent->node_offset == ondisk->parent);
 862         }
 863
 864         /*
 865          * If the leaf becomes empty it must be detached from the parent,
 866          * potentially recursing through to the filesystem root.
 867          *
 868          * This may reposition the cursor at one of the parent's of the
 869          * current node.
 870          *
 871          * Ignore deadlock errors, that simply means that btree_remove
 872          * was unable to recurse and had to leave us with an empty leaf.
 873          */
 874         KKASSERT(cursor->index <= ondisk->count);
 875         if (ondisk->count == 0) {
 876                 error = btree_remove(cursor);
 877                 if (error == EDEADLK)
 878                         error = 0;
 879         } else {
 880                 error = 0;
 881         }
 882         KKASSERT(cursor->parent == NULL ||
 883                  cursor->parent_index < cursor->parent->ondisk->count);
 884         return(error);
 885 }
 886
 887 /*
 888  * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
 889  *
 890  * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
 891  *
 892  * The search can begin ANYWHERE in the B-Tree.  As a first step the search
 893  * iterates up the tree as necessary to properly position itself prior to
 894  * actually doing the sarch.
 895  *
 896  * INSERTIONS: The search will split full nodes and leaves on its way down
 897  * and guarentee that the leaf it ends up on is not full.  If we run out
 898  * of space the search continues to the leaf (to position the cursor for
 899  * the spike), but ENOSPC is returned.
 900  *
 901  * The search is only guarenteed to end up on a leaf if an error code of 0
 902  * is returned, or if inserting and an error code of ENOENT is returned.
 903  * Otherwise it can stop at an internal node.  On success a search returns
 904  * a leaf node.
 905  *
 906  * COMPLEXITY WARNING!  This is the core B-Tree search code for the entire
 907  * filesystem, and it is not simple code.  Please note the following facts:
 908  *
 909  * - Internal node recursions have a boundary on the left AND right.  The
 910  *   right boundary is non-inclusive.  The create_tid is a generic part
 911  *   of the key for internal nodes.
 912  *
 913  * - Leaf nodes contain terminal elements only now.
 914  *
 915  * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
 916  *   historical search.  ASOF and INSERT are mutually exclusive.  When
 917  *   doing an as-of lookup btree_search() checks for a right-edge boundary
 918  *   case.  If while recursing down the left-edge differs from the key
 919  *   by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
 920  *   with cursor->create_check.  This is used by btree_lookup() to iterate.
 921  *   The iteration backwards because as-of searches can wind up going
 922  *   down the wrong branch of the B-Tree.
 923  */
 924 static
 925 int
 926 btree_search(hammer_cursor_t cursor, int flags)
 927 {
 928         hammer_node_ondisk_t node;
 929         hammer_btree_elm_t elm;
 930         int error;
 931         int enospc = 0;
 932         int i;
 933         int r;
 934         int s;
 935
 936         flags |= cursor->flags;
 937         ++hammer_stats_btree_searches;
 938
 939         if (hammer_debug_btree) {
 940                 kprintf("SEARCH   %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
 941                         cursor->node->node_offset,
 942                         cursor->index,
 943                         cursor->key_beg.obj_id,
 944                         cursor->key_beg.rec_type,
 945                         cursor->key_beg.key,
 946                         cursor->key_beg.create_tid,
 947                         cursor->key_beg.localization,
 948                         curthread
 949                 );
 950                 if (cursor->parent)
 951                     kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
 952                         cursor->parent->node_offset, cursor->parent_index,
 953                         cursor->left_bound->obj_id,
 954                         cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
 955                         cursor->right_bound->obj_id,
 956                         cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
 957                         cursor->left_bound,
 958                         &cursor->parent->ondisk->elms[cursor->parent_index],
 959                         cursor->right_bound,
 960                         &cursor->parent->ondisk->elms[cursor->parent_index+1]
 961                     );
 962         }
 963
 964         /*
 965          * Move our cursor up the tree until we find a node whos range covers
 966          * the key we are trying to locate.
 967          *
 968          * The left bound is inclusive, the right bound is non-inclusive.
 969          * It is ok to cursor up too far.
 970          */
 971         for (;;) {
 972                 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
 973                 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
 974                 if (r >= 0 && s < 0)
 975                         break;
 976                 KKASSERT(cursor->parent);
 977                 ++hammer_stats_btree_iterations;
 978                 error = hammer_cursor_up(cursor);
 979                 if (error)
 980                         goto done;
 981         }
 982
 983         /*
 984          * The delete-checks below are based on node, not parent.  Set the
 985          * initial delete-check based on the parent.
 986          */
 987         if (r == 1) {
 988                 KKASSERT(cursor->left_bound->create_tid != 1);
 989                 cursor->create_check = cursor->left_bound->create_tid - 1;
 990                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
 991         }
 992
 993         /*
 994          * We better have ended up with a node somewhere.
 995          */
 996         KKASSERT(cursor->node != NULL);
 997
 998         /*
 999          * If we are inserting we can't start at a full node if the parent
1000          * is also full (because there is no way to split the node),
1001          * continue running up the tree until the requirement is satisfied
1002          * or we hit the root of the filesystem.
1003          *
1004          * (If inserting we aren't doing an as-of search so we don't have
1005          *  to worry about create_check).
1006          */
1007         while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1008                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1009                         if (btree_node_is_full(cursor->node->ondisk) == 0)
1010                                 break;
1011                 } else {
1012                         if (btree_node_is_full(cursor->node->ondisk) ==0)
1013                                 break;
1014                 }
1015                 if (cursor->node->ondisk->parent == 0 ||
1016                     cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1017                         break;
1018                 }
1019                 ++hammer_stats_btree_iterations;
1020                 error = hammer_cursor_up(cursor);
1021                 /* node may have become stale */
1022                 if (error)
1023                         goto done;
1024         }
1025
1026         /*
1027          * Push down through internal nodes to locate the requested key.
1028          */
1029         node = cursor->node->ondisk;
1030         while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1031                 /*
1032                  * Scan the node to find the subtree index to push down into.
1033                  * We go one-past, then back-up.
1034                  *
1035                  * We must proactively remove deleted elements which may
1036                  * have been left over from a deadlocked btree_remove().
1037                  *
1038                  * The left and right boundaries are included in the loop
1039                  * in order to detect edge cases.
1040                  *
1041                  * If the separator only differs by create_tid (r == 1)
1042                  * and we are doing an as-of search, we may end up going
1043                  * down a branch to the left of the one containing the
1044                  * desired key.  This requires numerous special cases.
1045                  */
1046                 ++hammer_stats_btree_iterations;
1047                 if (hammer_debug_btree) {
1048                         kprintf("SEARCH-I %016llx count=%d\n",
1049                                 cursor->node->node_offset,
1050                                 node->count);
1051                 }
1052
1053                 /*
1054                  * Try to shortcut the search before dropping into the
1055                  * linear loop.  Locate the first node where r <= 1.
1056                  */
1057                 i = hammer_btree_search_node(&cursor->key_beg, node);
1058                 while (i <= node->count) {
1059                         ++hammer_stats_btree_elements;
1060                         elm = &node->elms[i];
1061                         r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1062                         if (hammer_debug_btree > 2) {
1063                                 kprintf(" IELM %p %d r=%d\n",
1064                                         &node->elms[i], i, r);
1065                         }
1066                         if (r < 0)
1067                                 break;
1068                         if (r == 1) {
1069                                 KKASSERT(elm->base.create_tid != 1);
1070                                 cursor->create_check = elm->base.create_tid - 1;
1071                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1072                         }
1073                         ++i;
1074                 }
1075                 if (hammer_debug_btree) {
1076                         kprintf("SEARCH-I preI=%d/%d r=%d\n",
1077                                 i, node->count, r);
1078                 }
1079
1080                 /*
1081                  * These cases occur when the parent's idea of the boundary
1082                  * is wider then the child's idea of the boundary, and
1083                  * require special handling.  If not inserting we can
1084                  * terminate the search early for these cases but the
1085                  * child's boundaries cannot be unconditionally modified.
1086                  */
1087                 if (i == 0) {
1088                         /*
1089                          * If i == 0 the search terminated to the LEFT of the
1090                          * left_boundary but to the RIGHT of the parent's left
1091                          * boundary.
1092                          */
1093                         u_int8_t save;
1094
1095                         elm = &node->elms[0];
1096
1097                         /*
1098                          * If we aren't inserting we can stop here.
1099                          */
1100                         if ((flags & (HAMMER_CURSOR_INSERT |
1101                                       HAMMER_CURSOR_PRUNING)) == 0) {
1102                                 cursor->index = 0;
1103                                 return(ENOENT);
1104                         }
1105
1106                         /*
1107                          * Correct a left-hand boundary mismatch.
1108                          *
1109                          * We can only do this if we can upgrade the lock,
1110                          * and synchronized as a background cursor (i.e.
1111                          * inserting or pruning).
1112                          *
1113                          * WARNING: We can only do this if inserting, i.e.
1114                          * we are running on the backend.
1115                          */
1116                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1117                                 return(error);
1118                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1119                         hammer_modify_node_field(cursor->trans, cursor->node,
1120                                                  elms[0]);
1121                         save = node->elms[0].base.btype;
1122                         node->elms[0].base = *cursor->left_bound;
1123                         node->elms[0].base.btype = save;
1124                         hammer_modify_node_done(cursor->node);
1125                 } else if (i == node->count + 1) {
1126                         /*
1127                          * If i == node->count + 1 the search terminated to
1128                          * the RIGHT of the right boundary but to the LEFT
1129                          * of the parent's right boundary.  If we aren't
1130                          * inserting we can stop here.
1131                          *
1132                          * Note that the last element in this case is
1133                          * elms[i-2] prior to adjustments to 'i'.
1134                          */
1135                         --i;
1136                         if ((flags & (HAMMER_CURSOR_INSERT |
1137                                       HAMMER_CURSOR_PRUNING)) == 0) {
1138                                 cursor->index = i;
1139                                 return (ENOENT);
1140                         }
1141
1142                         /*
1143                          * Correct a right-hand boundary mismatch.
1144                          * (actual push-down record is i-2 prior to
1145                          * adjustments to i).
1146                          *
1147                          * We can only do this if we can upgrade the lock,
1148                          * and synchronized as a background cursor (i.e.
1149                          * inserting or pruning).
1150                          *
1151                          * WARNING: We can only do this if inserting, i.e.
1152                          * we are running on the backend.
1153                          */
1154                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1155                                 return(error);
1156                         elm = &node->elms[i];
1157                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1158                         hammer_modify_node(cursor->trans, cursor->node,
1159                                            &elm->base, sizeof(elm->base));
1160                         elm->base = *cursor->right_bound;
1161                         hammer_modify_node_done(cursor->node);
1162                         --i;
1163                 } else {
1164                         /*
1165                          * The push-down index is now i - 1.  If we had
1166                          * terminated on the right boundary this will point
1167                          * us at the last element.
1168                          */
1169                         --i;
1170                 }
1171                 cursor->index = i;
1172                 elm = &node->elms[i];
1173
1174                 if (hammer_debug_btree) {
1175                         kprintf("RESULT-I %016llx[%d] %016llx %02x "
1176                                 "key=%016llx cre=%016llx lo=%02x\n",
1177                                 cursor->node->node_offset,
1178                                 i,
1179                                 elm->internal.base.obj_id,
1180                                 elm->internal.base.rec_type,
1181                                 elm->internal.base.key,
1182                                 elm->internal.base.create_tid,
1183                                 elm->internal.base.localization
1184                         );
1185                 }
1186
1187                 /*
1188                  * We better have a valid subtree offset.
1189                  */
1190                 KKASSERT(elm->internal.subtree_offset != 0);
1191
1192                 /*
1193                  * Handle insertion and deletion requirements.
1194                  *
1195                  * If inserting split full nodes.  The split code will
1196                  * adjust cursor->node and cursor->index if the current
1197                  * index winds up in the new node.
1198                  *
1199                  * If inserting and a left or right edge case was detected,
1200                  * we cannot correct the left or right boundary and must
1201                  * prepend and append an empty leaf node in order to make
1202                  * the boundary correction.
1203                  *
1204                  * If we run out of space we set enospc and continue on
1205                  * to a leaf to provide the spike code with a good point
1206                  * of entry.
1207                  */
1208                 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1209                         if (btree_node_is_full(node)) {
1210                                 error = btree_split_internal(cursor);
1211                                 if (error) {
1212                                         if (error != ENOSPC)
1213                                                 goto done;
1214                                         enospc = 1;
1215                                 }
1216                                 /*
1217                                  * reload stale pointers
1218                                  */
1219                                 i = cursor->index;
1220                                 node = cursor->node->ondisk;
1221                         }
1222                 }
1223
1224                 /*
1225                  * Push down (push into new node, existing node becomes
1226                  * the parent) and continue the search.
1227                  */
1228                 error = hammer_cursor_down(cursor);
1229                 /* node may have become stale */
1230                 if (error)
1231                         goto done;
1232                 node = cursor->node->ondisk;
1233         }
1234
1235         /*
1236          * We are at a leaf, do a linear search of the key array.
1237          *
1238          * On success the index is set to the matching element and 0
1239          * is returned.
1240          *
1241          * On failure the index is set to the insertion point and ENOENT
1242          * is returned.
1243          *
1244          * Boundaries are not stored in leaf nodes, so the index can wind
1245          * up to the left of element 0 (index == 0) or past the end of
1246          * the array (index == node->count).  It is also possible that the
1247          * leaf might be empty.
1248          */
1249         ++hammer_stats_btree_iterations;
1250         KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1251         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1252         if (hammer_debug_btree) {
1253                 kprintf("SEARCH-L %016llx count=%d\n",
1254                         cursor->node->node_offset,
1255                         node->count);
1256         }
1257
1258         /*
1259          * Try to shortcut the search before dropping into the
1260          * linear loop.  Locate the first node where r <= 1.
1261          */
1262         i = hammer_btree_search_node(&cursor->key_beg, node);
1263         while (i < node->count) {
1264                 ++hammer_stats_btree_elements;
1265                 elm = &node->elms[i];
1266
1267                 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1268
1269                 if (hammer_debug_btree > 1)
1270                         kprintf("  ELM %p %d r=%d\n", &node->elms[i], i, r);
1271
1272                 /*
1273                  * We are at a record element.  Stop if we've flipped past
1274                  * key_beg, not counting the create_tid test.  Allow the
1275                  * r == 1 case (key_beg > element but differs only by its
1276                  * create_tid) to fall through to the AS-OF check.
1277                  */
1278                 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1279
1280                 if (r < 0)
1281                         goto failed;
1282                 if (r > 1) {
1283                         ++i;
1284                         continue;
1285                 }
1286
1287                 /*
1288                  * Check our as-of timestamp against the element.
1289                  */
1290                 if (flags & HAMMER_CURSOR_ASOF) {
1291                         if (hammer_btree_chkts(cursor->asof,
1292                                                &node->elms[i].base) != 0) {
1293                                 ++i;
1294                                 continue;
1295                         }
1296                         /* success */
1297                 } else {
1298                         if (r > 0) {    /* can only be +1 */
1299                                 ++i;
1300                                 continue;
1301                         }
1302                         /* success */
1303                 }
1304                 cursor->index = i;
1305                 error = 0;
1306                 if (hammer_debug_btree) {
1307                         kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1308                                 cursor->node->node_offset, i);
1309                 }
1310                 goto done;
1311         }
1312
1313         /*
1314          * The search of the leaf node failed.  i is the insertion point.
1315          */
1316 failed:
1317         if (hammer_debug_btree) {
1318                 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1319                         cursor->node->node_offset, i);
1320         }
1321
1322         /*
1323          * No exact match was found, i is now at the insertion point.
1324          *
1325          * If inserting split a full leaf before returning.  This
1326          * may have the side effect of adjusting cursor->node and
1327          * cursor->index.
1328          */
1329         cursor->index = i;
1330         if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1331              btree_node_is_full(node)) {
1332                 error = btree_split_leaf(cursor);
1333                 if (error) {
1334                         if (error != ENOSPC)
1335                                 goto done;
1336                         enospc = 1;
1337                 }
1338                 /*
1339                  * reload stale pointers
1340                  */
1341                 /* NOT USED
1342                 i = cursor->index;
1343                 node = &cursor->node->internal;
1344                 */
1345         }
1346
1347         /*
1348          * We reached a leaf but did not find the key we were looking for.
1349          * If this is an insert we will be properly positioned for an insert
1350          * (ENOENT) or spike (ENOSPC) operation.
1351          */
1352         error = enospc ? ENOSPC : ENOENT;
1353 done:
1354         return(error);
1355 }
1356
1357 /*
1358  * Heuristical search for the first element whos comparison is <= 1.  May
1359  * return an index whos compare result is > 1 but may only return an index
1360  * whos compare result is <= 1 if it is the first element with that result.
1361  */
1362 int
1363 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1364 {
1365         int b;
1366         int s;
1367         int i;
1368         int r;
1369
1370         /*
1371          * Don't bother if the node does not have very many elements
1372          */
1373         b = 0;
1374         s = node->count;
1375         while (s - b > 4) {
1376                 i = b + (s - b) / 2;
1377                 ++hammer_stats_btree_elements;
1378                 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1379                 if (r <= 1) {
1380                         s = i;
1381                 } else {
1382                         b = i;
1383                 }
1384         }
1385         return(b);
1386 }
1387
1388
1389 /************************************************************************
1390  *                         SPLITTING AND MERGING                        *
1391  ************************************************************************
1392  *
1393  * These routines do all the dirty work required to split and merge nodes.
1394  */
1395
1396 /*
1397  * Split an internal node into two nodes and move the separator at the split
1398  * point to the parent.
1399  *
1400  * (cursor->node, cursor->index) indicates the element the caller intends
1401  * to push into.  We will adjust node and index if that element winds
1402  * up in the split node.
1403  *
1404  * If we are at the root of the filesystem a new root must be created with
1405  * two elements, one pointing to the original root and one pointing to the
1406  * newly allocated split node.
1407  */
1408 static
1409 int
1410 btree_split_internal(hammer_cursor_t cursor)
1411 {
1412         hammer_node_ondisk_t ondisk;
1413         hammer_node_t node;
1414         hammer_node_t parent;
1415         hammer_node_t new_node;
1416         hammer_btree_elm_t elm;
1417         hammer_btree_elm_t parent_elm;
1418         struct hammer_node_lock lockroot;
1419         hammer_mount_t hmp = cursor->trans->hmp;
1420         int parent_index;
1421         int made_root;
1422         int split;
1423         int error;
1424         int i;
1425         const int esize = sizeof(*elm);
1426
1427         hammer_node_lock_init(&lockroot, cursor->node);
1428         error = hammer_btree_lock_children(cursor, 1, &lockroot);
1429         if (error)
1430                 goto done;
1431         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1432                 goto done;
1433         ++hammer_stats_btree_splits;
1434
1435         /*
1436          * We are splitting but elms[split] will be promoted to the parent,
1437          * leaving the right hand node with one less element.  If the
1438          * insertion point will be on the left-hand side adjust the split
1439          * point to give the right hand side one additional node.
1440          */
1441         node = cursor->node;
1442         ondisk = node->ondisk;
1443         split = (ondisk->count + 1) / 2;
1444         if (cursor->index <= split)
1445                 --split;
1446
1447         /*
1448          * If we are at the root of the filesystem, create a new root node
1449          * with 1 element and split normally.  Avoid making major
1450          * modifications until we know the whole operation will work.
1451          */
1452         if (ondisk->parent == 0) {
1453                 parent = hammer_alloc_btree(cursor->trans, &error);
1454                 if (parent == NULL)
1455                         goto done;
1456                 hammer_lock_ex(&parent->lock);
1457                 hammer_modify_node_noundo(cursor->trans, parent);
1458                 ondisk = parent->ondisk;
1459                 ondisk->count = 1;
1460                 ondisk->parent = 0;
1461                 ondisk->mirror_tid = node->ondisk->mirror_tid;
1462                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1463                 ondisk->elms[0].base = hmp->root_btree_beg;
1464                 ondisk->elms[0].base.btype = node->ondisk->type;
1465                 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1466                 ondisk->elms[1].base = hmp->root_btree_end;
1467                 hammer_modify_node_done(parent);
1468                 /* ondisk->elms[1].base.btype - not used */
1469                 made_root = 1;
1470                 parent_index = 0;       /* index of current node in parent */
1471         } else {
1472                 made_root = 0;
1473                 parent = cursor->parent;
1474                 parent_index = cursor->parent_index;
1475         }
1476
1477         /*
1478          * Split node into new_node at the split point.
1479          *
1480          *  B O O O P N N B     <-- P = node->elms[split]
1481          *   0 1 2 3 4 5 6      <-- subtree indices
1482          *
1483          *       x x P x x
1484          *        s S S s
1485          *         /   \
1486          *  B O O O B    B N N B        <--- inner boundary points are 'P'
1487          *   0 1 2 3      4 5 6
1488          *
1489          */
1490         new_node = hammer_alloc_btree(cursor->trans, &error);
1491         if (new_node == NULL) {
1492                 if (made_root) {
1493                         hammer_unlock(&parent->lock);
1494                         hammer_delete_node(cursor->trans, parent);
1495                         hammer_rel_node(parent);
1496                 }
1497                 goto done;
1498         }
1499         hammer_lock_ex(&new_node->lock);
1500
1501         /*
1502          * Create the new node.  P becomes the left-hand boundary in the
1503          * new node.  Copy the right-hand boundary as well.
1504          *
1505          * elm is the new separator.
1506          */
1507         hammer_modify_node_noundo(cursor->trans, new_node);
1508         hammer_modify_node_all(cursor->trans, node);
1509         ondisk = node->ondisk;
1510         elm = &ondisk->elms[split];
1511         bcopy(elm, &new_node->ondisk->elms[0],
1512               (ondisk->count - split + 1) * esize);
1513         new_node->ondisk->count = ondisk->count - split;
1514         new_node->ondisk->parent = parent->node_offset;
1515         new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1516         new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1517         KKASSERT(ondisk->type == new_node->ondisk->type);
1518         hammer_cursor_split_node(node, new_node, split);
1519
1520         /*
1521          * Cleanup the original node.  Elm (P) becomes the new boundary,
1522          * its subtree_offset was moved to the new node.  If we had created
1523          * a new root its parent pointer may have changed.
1524          */
1525         elm->internal.subtree_offset = 0;
1526         ondisk->count = split;
1527
1528         /*
1529          * Insert the separator into the parent, fixup the parent's
1530          * reference to the original node, and reference the new node.
1531          * The separator is P.
1532          *
1533          * Remember that base.count does not include the right-hand boundary.
1534          */
1535         hammer_modify_node_all(cursor->trans, parent);
1536         ondisk = parent->ondisk;
1537         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1538         parent_elm = &ondisk->elms[parent_index+1];
1539         bcopy(parent_elm, parent_elm + 1,
1540               (ondisk->count - parent_index) * esize);
1541         parent_elm->internal.base = elm->base;  /* separator P */
1542         parent_elm->internal.base.btype = new_node->ondisk->type;
1543         parent_elm->internal.subtree_offset = new_node->node_offset;
1544         parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1545         ++ondisk->count;
1546         hammer_modify_node_done(parent);
1547         hammer_cursor_inserted_element(parent, parent_index + 1);
1548
1549         /*
1550          * The children of new_node need their parent pointer set to new_node.
1551          * The children have already been locked by
1552          * hammer_btree_lock_children().
1553          */
1554         for (i = 0; i < new_node->ondisk->count; ++i) {
1555                 elm = &new_node->ondisk->elms[i];
1556                 error = btree_set_parent(cursor->trans, new_node, elm);
1557                 if (error) {
1558                         panic("btree_split_internal: btree-fixup problem");
1559                 }
1560         }
1561         hammer_modify_node_done(new_node);
1562
1563         /*
1564          * The filesystem's root B-Tree pointer may have to be updated.
1565          */
1566         if (made_root) {
1567                 hammer_volume_t volume;
1568
1569                 volume = hammer_get_root_volume(hmp, &error);
1570                 KKASSERT(error == 0);
1571
1572                 hammer_modify_volume_field(cursor->trans, volume,
1573                                            vol0_btree_root);
1574                 volume->ondisk->vol0_btree_root = parent->node_offset;
1575                 hammer_modify_volume_done(volume);
1576                 node->ondisk->parent = parent->node_offset;
1577                 if (cursor->parent) {
1578                         hammer_unlock(&cursor->parent->lock);
1579                         hammer_rel_node(cursor->parent);
1580                 }
1581                 cursor->parent = parent;        /* lock'd and ref'd */
1582                 hammer_rel_volume(volume, 0);
1583         }
1584         hammer_modify_node_done(node);
1585
1586         /*
1587          * Ok, now adjust the cursor depending on which element the original
1588          * index was pointing at.  If we are >= the split point the push node
1589          * is now in the new node.
1590          *
1591          * NOTE: If we are at the split point itself we cannot stay with the
1592          * original node because the push index will point at the right-hand
1593          * boundary, which is illegal.
1594          *
1595          * NOTE: The cursor's parent or parent_index must be adjusted for
1596          * the case where a new parent (new root) was created, and the case
1597          * where the cursor is now pointing at the split node.
1598          */
1599         if (cursor->index >= split) {
1600                 cursor->parent_index = parent_index + 1;
1601                 cursor->index -= split;
1602                 hammer_unlock(&cursor->node->lock);
1603                 hammer_rel_node(cursor->node);
1604                 cursor->node = new_node;        /* locked and ref'd */
1605         } else {
1606                 cursor->parent_index = parent_index;
1607                 hammer_unlock(&new_node->lock);
1608                 hammer_rel_node(new_node);
1609         }
1610
1611         /*
1612          * Fixup left and right bounds
1613          */
1614         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1615         cursor->left_bound = &parent_elm[0].internal.base;
1616         cursor->right_bound = &parent_elm[1].internal.base;
1617         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1618                  &cursor->node->ondisk->elms[0].internal.base) <= 0);
1619         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1620                  &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1621
1622 done:
1623         hammer_btree_unlock_children(cursor, &lockroot);
1624         hammer_cursor_downgrade(cursor);
1625         return (error);
1626 }
1627
1628 /*
1629  * Same as the above, but splits a full leaf node.
1630  *
1631  * This function
1632  */
1633 static
1634 int
1635 btree_split_leaf(hammer_cursor_t cursor)
1636 {
1637         hammer_node_ondisk_t ondisk;
1638         hammer_node_t parent;
1639         hammer_node_t leaf;
1640         hammer_mount_t hmp;
1641         hammer_node_t new_leaf;
1642         hammer_btree_elm_t elm;
1643         hammer_btree_elm_t parent_elm;
1644         hammer_base_elm_t mid_boundary;
1645         int parent_index;
1646         int made_root;
1647         int split;
1648         int error;
1649         const size_t esize = sizeof(*elm);
1650
1651         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1652                 return(error);
1653         ++hammer_stats_btree_splits;
1654
1655         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1656                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1657         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1658                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1659
1660         /*
1661          * Calculate the split point.  If the insertion point will be on
1662          * the left-hand side adjust the split point to give the right
1663          * hand side one additional node.
1664          *
1665          * Spikes are made up of two leaf elements which cannot be
1666          * safely split.
1667          */
1668         leaf = cursor->node;
1669         ondisk = leaf->ondisk;
1670         split = (ondisk->count + 1) / 2;
1671         if (cursor->index <= split)
1672                 --split;
1673         error = 0;
1674         hmp = leaf->hmp;
1675
1676         elm = &ondisk->elms[split];
1677
1678         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1679         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1680         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1681         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1682
1683         /*
1684          * If we are at the root of the tree, create a new root node with
1685          * 1 element and split normally.  Avoid making major modifications
1686          * until we know the whole operation will work.
1687          */
1688         if (ondisk->parent == 0) {
1689                 parent = hammer_alloc_btree(cursor->trans, &error);
1690                 if (parent == NULL)
1691                         goto done;
1692                 hammer_lock_ex(&parent->lock);
1693                 hammer_modify_node_noundo(cursor->trans, parent);
1694                 ondisk = parent->ondisk;
1695                 ondisk->count = 1;
1696                 ondisk->parent = 0;
1697                 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1698                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1699                 ondisk->elms[0].base = hmp->root_btree_beg;
1700                 ondisk->elms[0].base.btype = leaf->ondisk->type;
1701                 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1702                 ondisk->elms[1].base = hmp->root_btree_end;
1703                 /* ondisk->elms[1].base.btype = not used */
1704                 hammer_modify_node_done(parent);
1705                 made_root = 1;
1706                 parent_index = 0;       /* insertion point in parent */
1707         } else {
1708                 made_root = 0;
1709                 parent = cursor->parent;
1710                 parent_index = cursor->parent_index;
1711         }
1712
1713         /*
1714          * Split leaf into new_leaf at the split point.  Select a separator
1715          * value in-between the two leafs but with a bent towards the right
1716          * leaf since comparisons use an 'elm >= separator' inequality.
1717          *
1718          *  L L L L L L L L
1719          *
1720          *       x x P x x
1721          *        s S S s
1722          *         /   \
1723          *  L L L L     L L L L
1724          */
1725         new_leaf = hammer_alloc_btree(cursor->trans, &error);
1726         if (new_leaf == NULL) {
1727                 if (made_root) {
1728                         hammer_unlock(&parent->lock);
1729                         hammer_delete_node(cursor->trans, parent);
1730                         hammer_rel_node(parent);
1731                 }
1732                 goto done;
1733         }
1734         hammer_lock_ex(&new_leaf->lock);
1735
1736         /*
1737          * Create the new node and copy the leaf elements from the split
1738          * point on to the new node.
1739          */
1740         hammer_modify_node_all(cursor->trans, leaf);
1741         hammer_modify_node_noundo(cursor->trans, new_leaf);
1742         ondisk = leaf->ondisk;
1743         elm = &ondisk->elms[split];
1744         bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1745         new_leaf->ondisk->count = ondisk->count - split;
1746         new_leaf->ondisk->parent = parent->node_offset;
1747         new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1748         new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1749         KKASSERT(ondisk->type == new_leaf->ondisk->type);
1750         hammer_modify_node_done(new_leaf);
1751         hammer_cursor_split_node(leaf, new_leaf, split);
1752
1753         /*
1754          * Cleanup the original node.  Because this is a leaf node and
1755          * leaf nodes do not have a right-hand boundary, there
1756          * aren't any special edge cases to clean up.  We just fixup the
1757          * count.
1758          */
1759         ondisk->count = split;
1760
1761         /*
1762          * Insert the separator into the parent, fixup the parent's
1763          * reference to the original node, and reference the new node.
1764          * The separator is P.
1765          *
1766          * Remember that base.count does not include the right-hand boundary.
1767          * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1768          */
1769         hammer_modify_node_all(cursor->trans, parent);
1770         ondisk = parent->ondisk;
1771         KKASSERT(split != 0);
1772         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1773         parent_elm = &ondisk->elms[parent_index+1];
1774         bcopy(parent_elm, parent_elm + 1,
1775               (ondisk->count - parent_index) * esize);
1776
1777         hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1778         parent_elm->internal.base.btype = new_leaf->ondisk->type;
1779         parent_elm->internal.subtree_offset = new_leaf->node_offset;
1780         parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1781         mid_boundary = &parent_elm->base;
1782         ++ondisk->count;
1783         hammer_modify_node_done(parent);
1784         hammer_cursor_inserted_element(parent, parent_index + 1);
1785
1786         /*
1787          * The filesystem's root B-Tree pointer may have to be updated.
1788          */
1789         if (made_root) {
1790                 hammer_volume_t volume;
1791
1792                 volume = hammer_get_root_volume(hmp, &error);
1793                 KKASSERT(error == 0);
1794
1795                 hammer_modify_volume_field(cursor->trans, volume,
1796                                            vol0_btree_root);
1797                 volume->ondisk->vol0_btree_root = parent->node_offset;
1798                 hammer_modify_volume_done(volume);
1799                 leaf->ondisk->parent = parent->node_offset;
1800                 if (cursor->parent) {
1801                         hammer_unlock(&cursor->parent->lock);
1802                         hammer_rel_node(cursor->parent);
1803                 }
1804                 cursor->parent = parent;        /* lock'd and ref'd */
1805                 hammer_rel_volume(volume, 0);
1806         }
1807         hammer_modify_node_done(leaf);
1808
1809         /*
1810          * Ok, now adjust the cursor depending on which element the original
1811          * index was pointing at.  If we are >= the split point the push node
1812          * is now in the new node.
1813          *
1814          * NOTE: If we are at the split point itself we need to select the
1815          * old or new node based on where key_beg's insertion point will be.
1816          * If we pick the wrong side the inserted element will wind up in
1817          * the wrong leaf node and outside that node's bounds.
1818          */
1819         if (cursor->index > split ||
1820             (cursor->index == split &&
1821              hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1822                 cursor->parent_index = parent_index + 1;
1823                 cursor->index -= split;
1824                 hammer_unlock(&cursor->node->lock);
1825                 hammer_rel_node(cursor->node);
1826                 cursor->node = new_leaf;
1827         } else {
1828                 cursor->parent_index = parent_index;
1829                 hammer_unlock(&new_leaf->lock);
1830                 hammer_rel_node(new_leaf);
1831         }
1832
1833         /*
1834          * Fixup left and right bounds
1835          */
1836         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1837         cursor->left_bound = &parent_elm[0].internal.base;
1838         cursor->right_bound = &parent_elm[1].internal.base;
1839
1840         /*
1841          * Assert that the bounds are correct.
1842          */
1843         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1844                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1845         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1846                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1847         KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1848         KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1849
1850 done:
1851         hammer_cursor_downgrade(cursor);
1852         return (error);
1853 }
1854
1855 #if 0
1856
1857 /*
1858  * Recursively correct the right-hand boundary's create_tid to (tid) as
1859  * long as the rest of the key matches.  We have to recurse upward in
1860  * the tree as well as down the left side of each parent's right node.
1861  *
1862  * Return EDEADLK if we were only partially successful, forcing the caller
1863  * to try again.  The original cursor is not modified.  This routine can
1864  * also fail with EDEADLK if it is forced to throw away a portion of its
1865  * record history.
1866  *
1867  * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1868  */
1869 struct hammer_rhb {
1870         TAILQ_ENTRY(hammer_rhb) entry;
1871         hammer_node_t   node;
1872         int             index;
1873 };
1874
1875 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1876
1877 int
1878 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1879 {
1880         struct hammer_mount *hmp;
1881         struct hammer_rhb_list rhb_list;
1882         hammer_base_elm_t elm;
1883         hammer_node_t orig_node;
1884         struct hammer_rhb *rhb;
1885         int orig_index;
1886         int error;
1887
1888         TAILQ_INIT(&rhb_list);
1889         hmp = cursor->trans->hmp;
1890
1891         /*
1892          * Save our position so we can restore it on return.  This also
1893          * gives us a stable 'elm'.
1894          */
1895         orig_node = cursor->node;
1896         hammer_ref_node(orig_node);
1897         hammer_lock_sh(&orig_node->lock);
1898         orig_index = cursor->index;
1899         elm = &orig_node->ondisk->elms[orig_index].base;
1900
1901         /*
1902          * Now build a list of parents going up, allocating a rhb
1903          * structure for each one.
1904          */
1905         while (cursor->parent) {
1906                 /*
1907                  * Stop if we no longer have any right-bounds to fix up
1908                  */
1909                 if (elm->obj_id != cursor->right_bound->obj_id ||
1910                     elm->rec_type != cursor->right_bound->rec_type ||
1911                     elm->key != cursor->right_bound->key) {
1912                         break;
1913                 }
1914
1915                 /*
1916                  * Stop if the right-hand bound's create_tid does not
1917                  * need to be corrected.
1918                  */
1919                 if (cursor->right_bound->create_tid >= tid)
1920                         break;
1921
1922                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
1923                 rhb->node = cursor->parent;
1924                 rhb->index = cursor->parent_index;
1925                 hammer_ref_node(rhb->node);
1926                 hammer_lock_sh(&rhb->node->lock);
1927                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1928
1929                 hammer_cursor_up(cursor);
1930         }
1931
1932         /*
1933          * now safely adjust the right hand bound for each rhb.  This may
1934          * also require taking the right side of the tree and iterating down
1935          * ITS left side.
1936          */
1937         error = 0;
1938         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1939                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1940                 if (error)
1941                         break;
1942                 TAILQ_REMOVE(&rhb_list, rhb, entry);
1943                 hammer_unlock(&rhb->node->lock);
1944                 hammer_rel_node(rhb->node);
1945                 kfree(rhb, hmp->m_misc);
1946
1947                 switch (cursor->node->ondisk->type) {
1948                 case HAMMER_BTREE_TYPE_INTERNAL:
1949                         /*
1950                          * Right-boundary for parent at internal node
1951                          * is one element to the right of the element whos
1952                          * right boundary needs adjusting.  We must then
1953                          * traverse down the left side correcting any left
1954                          * bounds (which may now be too far to the left).
1955                          */
1956                         ++cursor->index;
1957                         error = hammer_btree_correct_lhb(cursor, tid);
1958                         break;
1959                 default:
1960                         panic("hammer_btree_correct_rhb(): Bad node type");
1961                         error = EINVAL;
1962                         break;
1963                 }
1964         }
1965
1966         /*
1967          * Cleanup
1968          */
1969         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1970                 TAILQ_REMOVE(&rhb_list, rhb, entry);
1971                 hammer_unlock(&rhb->node->lock);
1972                 hammer_rel_node(rhb->node);
1973                 kfree(rhb, hmp->m_misc);
1974         }
1975         error = hammer_cursor_seek(cursor, orig_node, orig_index);
1976         hammer_unlock(&orig_node->lock);
1977         hammer_rel_node(orig_node);
1978         return (error);
1979 }
1980
1981 /*
1982  * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1983  * bound going downward starting at the current cursor position.
1984  *
1985  * This function does not restore the cursor after use.
1986  */
1987 int
1988 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1989 {
1990         struct hammer_rhb_list rhb_list;
1991         hammer_base_elm_t elm;
1992         hammer_base_elm_t cmp;
1993         struct hammer_rhb *rhb;
1994         struct hammer_mount *hmp;
1995         int error;
1996
1997         TAILQ_INIT(&rhb_list);
1998         hmp = cursor->trans->hmp;
1999
2000         cmp = &cursor->node->ondisk->elms[cursor->index].base;
2001
2002         /*
2003          * Record the node and traverse down the left-hand side for all
2004          * matching records needing a boundary correction.
2005          */
2006         error = 0;
2007         for (;;) {
2008                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2009                 rhb->node = cursor->node;
2010                 rhb->index = cursor->index;
2011                 hammer_ref_node(rhb->node);
2012                 hammer_lock_sh(&rhb->node->lock);
2013                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2014
2015                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2016                         /*
2017                          * Nothing to traverse down if we are at the right
2018                          * boundary of an internal node.
2019                          */
2020                         if (cursor->index == cursor->node->ondisk->count)
2021                                 break;
2022                 } else {
2023                         elm = &cursor->node->ondisk->elms[cursor->index].base;
2024                         if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2025                                 break;
2026                         panic("Illegal leaf record type %02x", elm->btype);
2027                 }
2028                 error = hammer_cursor_down(cursor);
2029                 if (error)
2030                         break;
2031
2032                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2033                 if (elm->obj_id != cmp->obj_id ||
2034                     elm->rec_type != cmp->rec_type ||
2035                     elm->key != cmp->key) {
2036                         break;
2037                 }
2038                 if (elm->create_tid >= tid)
2039                         break;
2040
2041         }
2042
2043         /*
2044          * Now we can safely adjust the left-hand boundary from the bottom-up.
2045          * The last element we remove from the list is the caller's right hand
2046          * boundary, which must also be adjusted.
2047          */
2048         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2049                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2050                 if (error)
2051                         break;
2052                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2053                 hammer_unlock(&rhb->node->lock);
2054                 hammer_rel_node(rhb->node);
2055                 kfree(rhb, hmp->m_misc);
2056
2057                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2058                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2059                         hammer_modify_node(cursor->trans, cursor->node,
2060                                            &elm->create_tid,
2061                                            sizeof(elm->create_tid));
2062                         elm->create_tid = tid;
2063                         hammer_modify_node_done(cursor->node);
2064                 } else {
2065                         panic("hammer_btree_correct_lhb(): Bad element type");
2066                 }
2067         }
2068
2069         /*
2070          * Cleanup
2071          */
2072         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2073                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2074                 hammer_unlock(&rhb->node->lock);
2075                 hammer_rel_node(rhb->node);
2076                 kfree(rhb, hmp->m_misc);
2077         }
2078         return (error);
2079 }
2080
2081 #endif
2082
2083 /*
2084  * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2085  * (cursor->node).  Returns 0 on success, EDEADLK if we could not complete
2086  * the operation due to a deadlock, or some other error.
2087  *
2088  * This routine is initially called with an empty leaf and may be
2089  * recursively called with single-element internal nodes.
2090  *
2091  * It should also be noted that when removing empty leaves we must be sure
2092  * to test and update mirror_tid because another thread may have deadlocked
2093  * against us (or someone) trying to propagate it up and cannot retry once
2094  * the node has been deleted.
2095  *
2096  * On return the cursor may end up pointing to an internal node, suitable
2097  * for further iteration but not for an immediate insertion or deletion.
2098  */
2099 static int
2100 btree_remove(hammer_cursor_t cursor)
2101 {
2102         hammer_node_ondisk_t ondisk;
2103         hammer_btree_elm_t elm;
2104         hammer_node_t node;
2105         hammer_node_t parent;
2106         const int esize = sizeof(*elm);
2107         int error;
2108
2109         node = cursor->node;
2110
2111         /*
2112          * When deleting the root of the filesystem convert it to
2113          * an empty leaf node.  Internal nodes cannot be empty.
2114          */
2115         ondisk = node->ondisk;
2116         if (ondisk->parent == 0) {
2117                 KKASSERT(cursor->parent == NULL);
2118                 hammer_modify_node_all(cursor->trans, node);
2119                 KKASSERT(ondisk == node->ondisk);
2120                 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2121                 ondisk->count = 0;
2122                 hammer_modify_node_done(node);
2123                 cursor->index = 0;
2124                 return(0);
2125         }
2126
2127         parent = cursor->parent;
2128         hammer_cursor_removed_node(node, parent, cursor->parent_index);
2129
2130         /*
2131          * Attempt to remove the parent's reference to the child.  If the
2132          * parent would become empty we have to recurse.  If we fail we
2133          * leave the parent pointing to an empty leaf node.
2134          *
2135          * We have to recurse successfully before we can delete the internal
2136          * node as it is illegal to have empty internal nodes.  Even though
2137          * the operation may be aborted we must still fixup any unlocked
2138          * cursors as if we had deleted the element prior to recursing
2139          * (by calling hammer_cursor_deleted_element()) so those cursors
2140          * are properly forced up the chain by the recursion.
2141          */
2142         if (parent->ondisk->count == 1) {
2143                 /*
2144                  * This special cursor_up_locked() call leaves the original
2145                  * node exclusively locked and referenced, leaves the
2146                  * original parent locked (as the new node), and locks the
2147                  * new parent.  It can return EDEADLK.
2148                  */
2149                 error = hammer_cursor_up_locked(cursor);
2150                 if (error == 0) {
2151                         hammer_cursor_deleted_element(cursor->node, 0);
2152                         error = btree_remove(cursor);
2153                         if (error == 0) {
2154                                 hammer_modify_node_all(cursor->trans, node);
2155                                 ondisk = node->ondisk;
2156                                 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2157                                 ondisk->count = 0;
2158                                 hammer_modify_node_done(node);
2159                                 hammer_flush_node(node);
2160                                 hammer_delete_node(cursor->trans, node);
2161                         } else {
2162                                 kprintf("Warning: BTREE_REMOVE: Defering "
2163                                         "parent removal1 @ %016llx, skipping\n",
2164                                         node->node_offset);
2165                         }
2166                         hammer_unlock(&node->lock);
2167                         hammer_rel_node(node);
2168                 } else {
2169                         kprintf("Warning: BTREE_REMOVE: Defering parent "
2170                                 "removal2 @ %016llx, skipping\n",
2171                                 node->node_offset);
2172                 }
2173         } else {
2174                 KKASSERT(parent->ondisk->count > 1);
2175
2176                 hammer_modify_node_all(cursor->trans, parent);
2177                 ondisk = parent->ondisk;
2178                 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2179
2180                 elm = &ondisk->elms[cursor->parent_index];
2181                 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2182                 KKASSERT(ondisk->count > 0);
2183
2184                 /*
2185                  * We must retain the highest mirror_tid.  The deleted
2186                  * range is now encompassed by the element to the left.
2187                  * If we are already at the left edge the new left edge
2188                  * inherits mirror_tid.
2189                  *
2190                  * Note that bounds of the parent to our parent may create
2191                  * a gap to the left of our left-most node or to the right
2192                  * of our right-most node.  The gap is silently included
2193                  * in the mirror_tid's area of effect from the point of view
2194                  * of the scan.
2195                  */
2196                 if (cursor->parent_index) {
2197                         if (elm[-1].internal.mirror_tid <
2198                             elm[0].internal.mirror_tid) {
2199                                 elm[-1].internal.mirror_tid =
2200                                     elm[0].internal.mirror_tid;
2201                         }
2202                 } else {
2203                         if (elm[1].internal.mirror_tid <
2204                             elm[0].internal.mirror_tid) {
2205                                 elm[1].internal.mirror_tid =
2206                                     elm[0].internal.mirror_tid;
2207                         }
2208                 }
2209
2210                 /*
2211                  * Delete the subtree reference in the parent
2212                  */
2213                 bcopy(&elm[1], &elm[0],
2214                       (ondisk->count - cursor->parent_index) * esize);
2215                 --ondisk->count;
2216                 hammer_modify_node_done(parent);
2217                 hammer_cursor_deleted_element(parent, cursor->parent_index);
2218                 hammer_flush_node(node);
2219                 hammer_delete_node(cursor->trans, node);
2220
2221                 /*
2222                  * cursor->node is invalid, cursor up to make the cursor
2223                  * valid again.
2224                  */
2225                 error = hammer_cursor_up(cursor);
2226         }
2227         return (error);
2228 }
2229
2230 /*
2231  * Propagate cursor->trans->tid up the B-Tree starting at the current
2232  * cursor position using pseudofs info gleaned from the passed inode.
2233  *
2234  * The passed inode has no relationship to the cursor position other
2235  * then being in the same pseudofs as the insertion or deletion we
2236  * are propagating the mirror_tid for.
2237  */
2238 void
2239 hammer_btree_do_propagation(hammer_cursor_t cursor,
2240                             hammer_pseudofs_inmem_t pfsm,
2241                             hammer_btree_leaf_elm_t leaf)
2242 {
2243         hammer_cursor_t ncursor;
2244         hammer_tid_t mirror_tid;
2245         int error;
2246
2247         /*
2248          * We do not propagate a mirror_tid if the filesystem was mounted
2249          * in no-mirror mode.
2250          */
2251         if (cursor->trans->hmp->master_id < 0)
2252                 return;
2253
2254         /*
2255          * This is a bit of a hack because we cannot deadlock or return
2256          * EDEADLK here.  The related operation has already completed and
2257          * we must propagate the mirror_tid now regardless.
2258          *
2259          * Generate a new cursor which inherits the original's locks and
2260          * unlock the original.  Use the new cursor to propagate the
2261          * mirror_tid.  Then clean up the new cursor and reacquire locks
2262          * on the original.
2263          *
2264          * hammer_dup_cursor() cannot dup locks.  The dup inherits the
2265          * original's locks and the original is tracked and must be
2266          * re-locked.
2267          */
2268         mirror_tid = cursor->node->ondisk->mirror_tid;
2269         KKASSERT(mirror_tid != 0);
2270         ncursor = hammer_push_cursor(cursor);
2271         error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2272         KKASSERT(error == 0);
2273         hammer_pop_cursor(cursor, ncursor);
2274 }
2275
2276
2277 /*
2278  * Propagate a mirror TID update upwards through the B-Tree to the root.
2279  *
2280  * A locked internal node must be passed in.  The node will remain locked
2281  * on return.
2282  *
2283  * This function syncs mirror_tid at the specified internal node's element,
2284  * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2285  */
2286 static int
2287 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2288 {
2289         hammer_btree_internal_elm_t elm;
2290         hammer_node_t node;
2291         int error;
2292
2293         for (;;) {
2294                 error = hammer_cursor_up(cursor);
2295                 if (error == 0)
2296                         error = hammer_cursor_upgrade(cursor);
2297                 while (error == EDEADLK) {
2298                         hammer_recover_cursor(cursor);
2299                         error = hammer_cursor_upgrade(cursor);
2300                 }
2301                 if (error)
2302                         break;
2303                 node = cursor->node;
2304                 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2305
2306                 /*
2307                  * Adjust the node's element
2308                  */
2309                 elm = &node->ondisk->elms[cursor->index].internal;
2310                 if (elm->mirror_tid >= mirror_tid)
2311                         break;
2312                 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2313                                    sizeof(elm->mirror_tid));
2314                 elm->mirror_tid = mirror_tid;
2315                 hammer_modify_node_done(node);
2316                 if (hammer_debug_general & 0x0002) {
2317                         kprintf("mirror_propagate: propagate "
2318                                 "%016llx @%016llx:%d\n",
2319                                 mirror_tid, node->node_offset, cursor->index);
2320                 }
2321
2322
2323                 /*
2324                  * Adjust the node's mirror_tid aggregator
2325                  */
2326                 if (node->ondisk->mirror_tid >= mirror_tid)
2327                         return(0);
2328                 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2329                 node->ondisk->mirror_tid = mirror_tid;
2330                 hammer_modify_node_done(node);
2331                 if (hammer_debug_general & 0x0002) {
2332                         kprintf("mirror_propagate: propagate "
2333                                 "%016llx @%016llx\n",
2334                                 mirror_tid, node->node_offset);
2335                 }
2336         }
2337         if (error == ENOENT)
2338                 error = 0;
2339         return(error);
2340 }
2341
2342 hammer_node_t
2343 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2344                         int *parent_indexp, int *errorp, int try_exclusive)
2345 {
2346         hammer_node_t parent;
2347         hammer_btree_elm_t elm;
2348         int i;
2349
2350         /*
2351          * Get the node
2352          */
2353         parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2354         if (*errorp) {
2355                 KKASSERT(parent == NULL);
2356                 return(NULL);
2357         }
2358         KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2359
2360         /*
2361          * Lock the node
2362          */
2363         if (try_exclusive) {
2364                 if (hammer_lock_ex_try(&parent->lock)) {
2365                         hammer_rel_node(parent);
2366                         *errorp = EDEADLK;
2367                         return(NULL);
2368                 }
2369         } else {
2370                 hammer_lock_sh(&parent->lock);
2371         }
2372
2373         /*
2374          * Figure out which element in the parent is pointing to the
2375          * child.
2376          */
2377         if (node->ondisk->count) {
2378                 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2379                                              parent->ondisk);
2380         } else {
2381                 i = 0;
2382         }
2383         while (i < parent->ondisk->count) {
2384                 elm = &parent->ondisk->elms[i];
2385                 if (elm->internal.subtree_offset == node->node_offset)
2386                         break;
2387                 ++i;
2388         }
2389         if (i == parent->ondisk->count) {
2390                 hammer_unlock(&parent->lock);
2391                 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2392         }
2393         *parent_indexp = i;
2394         KKASSERT(*errorp == 0);
2395         return(parent);
2396 }
2397
2398 /*
2399  * The element (elm) has been moved to a new internal node (node).
2400  *
2401  * If the element represents a pointer to an internal node that node's
2402  * parent must be adjusted to the element's new location.
2403  *
2404  * XXX deadlock potential here with our exclusive locks
2405  */
2406 int
2407 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2408                  hammer_btree_elm_t elm)
2409 {
2410         hammer_node_t child;
2411         int error;
2412
2413         error = 0;
2414
2415         switch(elm->base.btype) {
2416         case HAMMER_BTREE_TYPE_INTERNAL:
2417         case HAMMER_BTREE_TYPE_LEAF:
2418                 child = hammer_get_node(trans, elm->internal.subtree_offset,
2419                                         0, &error);
2420                 if (error == 0) {
2421                         hammer_modify_node_field(trans, child, parent);
2422                         child->ondisk->parent = node->node_offset;
2423                         hammer_modify_node_done(child);
2424                         hammer_rel_node(child);
2425                 }
2426                 break;
2427         default:
2428                 break;
2429         }
2430         return(error);
2431 }
2432
2433 /*
2434  * Initialize the root of a recursive B-Tree node lock list structure.
2435  */
2436 void
2437 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2438 {
2439         TAILQ_INIT(&parent->list);
2440         parent->parent = NULL;
2441         parent->node = node;
2442         parent->index = -1;
2443         parent->count = node->ondisk->count;
2444         parent->copy = NULL;
2445         parent->flags = 0;
2446 }
2447
2448 /*
2449  * Exclusively lock all the children of node.  This is used by the split
2450  * code to prevent anyone from accessing the children of a cursor node
2451  * while we fix-up its parent offset.
2452  *
2453  * If we don't lock the children we can really mess up cursors which block
2454  * trying to cursor-up into our node.
2455  *
2456  * On failure EDEADLK (or some other error) is returned.  If a deadlock
2457  * error is returned the cursor is adjusted to block on termination.
2458  *
2459  * The caller is responsible for managing parent->node, the root's node
2460  * is usually aliased from a cursor.
2461  */
2462 int
2463 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2464                            hammer_node_lock_t parent)
2465 {
2466         hammer_node_t node;
2467         hammer_node_lock_t item;
2468         hammer_node_ondisk_t ondisk;
2469         hammer_btree_elm_t elm;
2470         hammer_node_t child;
2471         struct hammer_mount *hmp;
2472         int error;
2473         int i;
2474
2475         node = parent->node;
2476         ondisk = node->ondisk;
2477         error = 0;
2478         hmp = cursor->trans->hmp;
2479
2480         /*
2481          * We really do not want to block on I/O with exclusive locks held,
2482          * pre-get the children before trying to lock the mess.  This is
2483          * only done one-level deep for now.
2484          */
2485         for (i = 0; i < ondisk->count; ++i) {
2486                 ++hammer_stats_btree_elements;
2487                 elm = &ondisk->elms[i];
2488                 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2489                     elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2490                         continue;
2491                 }
2492                 child = hammer_get_node(cursor->trans,
2493                                         elm->internal.subtree_offset,
2494                                         0, &error);
2495                 if (child)
2496                         hammer_rel_node(child);
2497         }
2498
2499         /*
2500          * Do it for real
2501          */
2502         for (i = 0; error == 0 && i < ondisk->count; ++i) {
2503                 ++hammer_stats_btree_elements;
2504                 elm = &ondisk->elms[i];
2505
2506                 switch(elm->base.btype) {
2507                 case HAMMER_BTREE_TYPE_INTERNAL:
2508                 case HAMMER_BTREE_TYPE_LEAF:
2509                         KKASSERT(elm->internal.subtree_offset != 0);
2510                         child = hammer_get_node(cursor->trans,
2511                                                 elm->internal.subtree_offset,
2512                                                 0, &error);
2513                         break;
2514                 default:
2515                         child = NULL;
2516                         break;
2517                 }
2518                 if (child) {
2519                         if (hammer_lock_ex_try(&child->lock) != 0) {
2520                                 if (cursor->deadlk_node == NULL) {
2521                                         cursor->deadlk_node = child;
2522                                         hammer_ref_node(cursor->deadlk_node);
2523                                 }
2524                                 error = EDEADLK;
2525                                 hammer_rel_node(child);
2526                         } else {
2527                                 item = kmalloc(sizeof(*item), hmp->m_misc,
2528                                                M_WAITOK|M_ZERO);
2529                                 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2530                                 TAILQ_INIT(&item->list);
2531                                 item->parent = parent;
2532                                 item->node = child;
2533                                 item->index = i;
2534                                 item->count = child->ondisk->count;
2535
2536                                 /*
2537                                  * Recurse (used by the rebalancing code)
2538                                  */
2539                                 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2540                                         error = hammer_btree_lock_children(
2541                                                         cursor,
2542                                                         depth - 1,
2543                                                         item);
2544                                 }
2545                         }
2546                 }
2547         }
2548         if (error)
2549                 hammer_btree_unlock_children(cursor, parent);
2550         return(error);
2551 }
2552
2553 /*
2554  * Create an in-memory copy of all B-Tree nodes listed, recursively,
2555  * including the parent.
2556  */
2557 void
2558 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2559 {
2560         hammer_mount_t hmp = cursor->trans->hmp;
2561         hammer_node_lock_t item;
2562
2563         if (parent->copy == NULL) {
2564                 parent->copy = kmalloc(sizeof(*parent->copy), hmp->m_misc,
2565                                        M_WAITOK);
2566                 *parent->copy = *parent->node->ondisk;
2567         }
2568         TAILQ_FOREACH(item, &parent->list, entry) {
2569                 hammer_btree_lock_copy(cursor, item);
2570         }
2571 }
2572
2573 /*
2574  * Recursively sync modified copies to the media.
2575  */
2576 int
2577 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2578 {
2579         hammer_node_lock_t item;
2580         int count = 0;
2581
2582         if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2583                 ++count;
2584                 hammer_modify_node_all(cursor->trans, parent->node);
2585                 *parent->node->ondisk = *parent->copy;
2586                 hammer_modify_node_done(parent->node);
2587                 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2588                         hammer_flush_node(parent->node);
2589                         hammer_delete_node(cursor->trans, parent->node);
2590                 }
2591         }
2592         TAILQ_FOREACH(item, &parent->list, entry) {
2593                 count += hammer_btree_sync_copy(cursor, item);
2594         }
2595         return(count);
2596 }
2597
2598 /*
2599  * Release previously obtained node locks.  The caller is responsible for
2600  * cleaning up parent->node itself (its usually just aliased from a cursor),
2601  * but this function will take care of the copies.
2602  */
2603 void
2604 hammer_btree_unlock_children(hammer_cursor_t cursor, hammer_node_lock_t parent)
2605 {
2606         hammer_node_lock_t item;
2607
2608         if (parent->copy) {
2609                 kfree(parent->copy, cursor->trans->hmp->m_misc);
2610                 parent->copy = NULL;    /* safety */
2611         }
2612         while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2613                 TAILQ_REMOVE(&parent->list, item, entry);
2614                 hammer_btree_unlock_children(cursor, item);
2615                 hammer_unlock(&item->node->lock);
2616                 hammer_rel_node(item->node);
2617                 kfree(item, cursor->trans->hmp->m_misc);
2618         }
2619 }
2620
2621 /************************************************************************
2622  *                         MISCELLANIOUS SUPPORT                        *
2623  ************************************************************************/
2624
2625 /*
2626  * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2627  *
2628  * Note that for this particular function a return value of -1, 0, or +1
2629  * can denote a match if create_tid is otherwise discounted.  A create_tid
2630  * of zero is considered to be 'infinity' in comparisons.
2631  *
2632  * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2633  */
2634 int
2635 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2636 {
2637         if (key1->localization < key2->localization)
2638                 return(-5);
2639         if (key1->localization > key2->localization)
2640                 return(5);
2641
2642         if (key1->obj_id < key2->obj_id)
2643                 return(-4);
2644         if (key1->obj_id > key2->obj_id)
2645                 return(4);
2646
2647         if (key1->rec_type < key2->rec_type)
2648                 return(-3);
2649         if (key1->rec_type > key2->rec_type)
2650                 return(3);
2651
2652         if (key1->key < key2->key)
2653                 return(-2);
2654         if (key1->key > key2->key)
2655                 return(2);
2656
2657         /*
2658          * A create_tid of zero indicates a record which is undeletable
2659          * and must be considered to have a value of positive infinity.
2660          */
2661         if (key1->create_tid == 0) {
2662                 if (key2->create_tid == 0)
2663                         return(0);
2664                 return(1);
2665         }
2666         if (key2->create_tid == 0)
2667                 return(-1);
2668         if (key1->create_tid < key2->create_tid)
2669                 return(-1);
2670         if (key1->create_tid > key2->create_tid)
2671                 return(1);
2672         return(0);
2673 }
2674
2675 /*
2676  * Test a timestamp against an element to determine whether the
2677  * element is visible.  A timestamp of 0 means 'infinity'.
2678  */
2679 int
2680 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2681 {
2682         if (asof == 0) {
2683                 if (base->delete_tid)
2684                         return(1);
2685                 return(0);
2686         }
2687         if (asof < base->create_tid)
2688                 return(-1);
2689         if (base->delete_tid && asof >= base->delete_tid)
2690                 return(1);
2691         return(0);
2692 }
2693
2694 /*
2695  * Create a separator half way inbetween key1 and key2.  For fields just
2696  * one unit apart, the separator will match key2.  key1 is on the left-hand
2697  * side and key2 is on the right-hand side.
2698  *
2699  * key2 must be >= the separator.  It is ok for the separator to match key2.
2700  *
2701  * NOTE: Even if key1 does not match key2, the separator may wind up matching
2702  * key2.
2703  *
2704  * NOTE: It might be beneficial to just scrap this whole mess and just
2705  * set the separator to key2.
2706  */
2707 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2708         dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2709
2710 static void
2711 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2712                       hammer_base_elm_t dest)
2713 {
2714         bzero(dest, sizeof(*dest));
2715
2716         dest->rec_type = key2->rec_type;
2717         dest->key = key2->key;
2718         dest->obj_id = key2->obj_id;
2719         dest->create_tid = key2->create_tid;
2720
2721         MAKE_SEPARATOR(key1, key2, dest, localization);
2722         if (key1->localization == key2->localization) {
2723                 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2724                 if (key1->obj_id == key2->obj_id) {
2725                         MAKE_SEPARATOR(key1, key2, dest, rec_type);
2726                         if (key1->rec_type == key2->rec_type) {
2727                                 MAKE_SEPARATOR(key1, key2, dest, key);
2728                                 /*
2729                                  * Don't bother creating a separator for
2730                                  * create_tid, which also conveniently avoids
2731                                  * having to handle the create_tid == 0
2732                                  * (infinity) case.  Just leave create_tid
2733                                  * set to key2.
2734                                  *
2735                                  * Worst case, dest matches key2 exactly,
2736                                  * which is acceptable.
2737                                  */
2738                         }
2739                 }
2740         }
2741 }
2742
2743 #undef MAKE_SEPARATOR
2744
2745 /*
2746  * Return whether a generic internal or leaf node is full
2747  */
2748 static int
2749 btree_node_is_full(hammer_node_ondisk_t node)
2750 {
2751         switch(node->type) {
2752         case HAMMER_BTREE_TYPE_INTERNAL:
2753                 if (node->count == HAMMER_BTREE_INT_ELMS)
2754                         return(1);
2755                 break;
2756         case HAMMER_BTREE_TYPE_LEAF:
2757                 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2758                         return(1);
2759                 break;
2760         default:
2761                 panic("illegal btree subtype");
2762         }
2763         return(0);
2764 }
2765
2766 #if 0
2767 static int
2768 btree_max_elements(u_int8_t type)
2769 {
2770         if (type == HAMMER_BTREE_TYPE_LEAF)
2771                 return(HAMMER_BTREE_LEAF_ELMS);
2772         if (type == HAMMER_BTREE_TYPE_INTERNAL)
2773                 return(HAMMER_BTREE_INT_ELMS);
2774         panic("btree_max_elements: bad type %d\n", type);
2775 }
2776 #endif
2777
2778 void
2779 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2780 {
2781         hammer_btree_elm_t elm;
2782         int i;
2783
2784         kprintf("node %p count=%d parent=%016llx type=%c\n",
2785                 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2786
2787         /*
2788          * Dump both boundary elements if an internal node
2789          */
2790         if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2791                 for (i = 0; i <= ondisk->count; ++i) {
2792                         elm = &ondisk->elms[i];
2793                         hammer_print_btree_elm(elm, ondisk->type, i);
2794                 }
2795         } else {
2796                 for (i = 0; i < ondisk->count; ++i) {
2797                         elm = &ondisk->elms[i];
2798                         hammer_print_btree_elm(elm, ondisk->type, i);
2799                 }
2800         }
2801 }
2802
2803 void
2804 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2805 {
2806         kprintf("  %2d", i);
2807         kprintf("\tobj_id       = %016llx\n", elm->base.obj_id);
2808         kprintf("\tkey          = %016llx\n", elm->base.key);
2809         kprintf("\tcreate_tid   = %016llx\n", elm->base.create_tid);
2810         kprintf("\tdelete_tid   = %016llx\n", elm->base.delete_tid);
2811         kprintf("\trec_type     = %04x\n", elm->base.rec_type);
2812         kprintf("\tobj_type     = %02x\n", elm->base.obj_type);
2813         kprintf("\tbtype        = %02x (%c)\n",
2814                 elm->base.btype,
2815                 (elm->base.btype ? elm->base.btype : '?'));
2816         kprintf("\tlocalization = %02x\n", elm->base.localization);
2817
2818         switch(type) {
2819         case HAMMER_BTREE_TYPE_INTERNAL:
2820                 kprintf("\tsubtree_off  = %016llx\n",
2821                         elm->internal.subtree_offset);
2822                 break;
2823         case HAMMER_BTREE_TYPE_RECORD:
2824                 kprintf("\tdata_offset  = %016llx\n", elm->leaf.data_offset);
2825                 kprintf("\tdata_len     = %08x\n", elm->leaf.data_len);
2826                 kprintf("\tdata_crc     = %08x\n", elm->leaf.data_crc);
2827                 break;
2828         }
2829 }