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
  35 /*
  36  * HAMMER B-Tree index
  37  *
  38  * HAMMER implements a modified B+Tree.  In documentation this will
  39  * simply be refered to as the HAMMER B-Tree.  Basically a HAMMER B-Tree
  40  * looks like a B+Tree (A B-Tree which stores its records only at the leafs
  41  * of the tree), but adds two additional boundary elements which describe
  42  * the left-most and right-most element a node is able to represent.  In
  43  * otherwords, we have boundary elements at the two ends of a B-Tree node
  44  * with no valid sub-tree pointer for the right-most element.
  45  *
  46  * A B-Tree internal node looks like this:
  47  *
  48  *      B N N N N N N B   <-- boundary and internal elements
  49  *       S S S S S S S    <-- subtree pointers
  50  *
  51  * A B-Tree leaf node basically looks like this:
  52  *
  53  *      L L L L L L L L   <-- leaf elemenets
  54  *
  55  * The radix for an internal node is 1 less then a leaf but we get a
  56  * number of significant benefits for our troubles.
  57  * The left-hand boundary (B in the left) is integrated into the first
  58  * element so it doesn't require 2 elements to accomodate boundaries.
  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  * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor
 114  * was reverse indexed due to being moved to a parent while unlocked,
 115  * and something else might have inserted an element outside the iteration
 116  * range.  When this case occurs the iterator just keeps iterating until
 117  * it gets back into the iteration range (instead of asserting).
 118  *
 119  * NOTE!  EDEADLK *CANNOT* be returned by this procedure.
 120  */
 121 int
 122 hammer_btree_iterate(hammer_cursor_t cursor)
 123 {
 124         hammer_node_ondisk_t node;
 125         hammer_btree_elm_t elm;
 126         hammer_mount_t hmp;
 127         int error = 0;
 128         int r;
 129         int s;
 130
 131         /*
 132          * Skip past the current record
 133          */
 134         hmp = cursor->trans->hmp;
 135         node = cursor->node->ondisk;
 136         if (node == NULL)
 137                 return(ENOENT);
 138         if (cursor->index < node->count &&
 139             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 140                 ++cursor->index;
 141         }
 142
 143         /*
 144          * HAMMER can wind up being cpu-bound.
 145          */
 146         if (++hmp->check_yield > hammer_yield_check) {
 147                 hmp->check_yield = 0;
 148                 lwkt_user_yield();
 149         }
 150
 151
 152         /*
 153          * Loop until an element is found or we are done.
 154          */
 155         for (;;) {
 156                 /*
 157                  * We iterate up the tree and then index over one element
 158                  * while we are at the last element in the current node.
 159                  *
 160                  * If we are at the root of the filesystem, cursor_up
 161                  * returns ENOENT.
 162                  *
 163                  * XXX this could be optimized by storing the information in
 164                  * the parent reference.
 165                  *
 166                  * XXX we can lose the node lock temporarily, this could mess
 167                  * up our scan.
 168                  */
 169                 ++hammer_stats_btree_iterations;
 170                 hammer_flusher_clean_loose_ios(hmp);
 171
 172                 if (cursor->index == node->count) {
 173                         if (hammer_debug_btree) {
 174                                 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
 175                                         (long long)cursor->node->node_offset,
 176                                         cursor->index,
 177                                         (long long)(cursor->parent ? cursor->parent->node_offset : -1),
 178                                         cursor->parent_index,
 179                                         curthread);
 180                         }
 181                         KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
 182                         error = hammer_cursor_up(cursor);
 183                         if (error)
 184                                 break;
 185                         /* reload stale pointer */
 186                         node = cursor->node->ondisk;
 187                         KKASSERT(cursor->index != node->count);
 188
 189                         /*
 190                          * If we are reblocking we want to return internal
 191                          * nodes.  Note that the internal node will be
 192                          * returned multiple times, on each upward recursion
 193                          * from its children.  The caller selects which
 194                          * revisit it cares about (usually first or last only).
 195                          */
 196                         if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
 197                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 198                                 return(0);
 199                         }
 200                         ++cursor->index;
 201                         continue;
 202                 }
 203
 204                 /*
 205                  * Check internal or leaf element.  Determine if the record
 206                  * at the cursor has gone beyond the end of our range.
 207                  *
 208                  * We recurse down through internal nodes.
 209                  */
 210                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 211                         elm = &node->elms[cursor->index];
 212
 213                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 214                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 215                         if (hammer_debug_btree) {
 216                                 kprintf("BRACKETL %016llx[%d] %016llx %02x "
 217                                         "key=%016llx lo=%02x %d (td=%p)\n",
 218                                         (long long)cursor->node->node_offset,
 219                                         cursor->index,
 220                                         (long long)elm[0].internal.base.obj_id,
 221                                         elm[0].internal.base.rec_type,
 222                                         (long long)elm[0].internal.base.key,
 223                                         elm[0].internal.base.localization,
 224                                         r,
 225                                         curthread
 226                                 );
 227                                 kprintf("BRACKETR %016llx[%d] %016llx %02x "
 228                                         "key=%016llx lo=%02x %d\n",
 229                                         (long long)cursor->node->node_offset,
 230                                         cursor->index + 1,
 231                                         (long long)elm[1].internal.base.obj_id,
 232                                         elm[1].internal.base.rec_type,
 233                                         (long long)elm[1].internal.base.key,
 234                                         elm[1].internal.base.localization,
 235                                         s
 236                                 );
 237                         }
 238
 239                         if (r < 0) {
 240                                 error = ENOENT;
 241                                 break;
 242                         }
 243                         if (r == 0 && (cursor->flags &
 244                                        HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 245                                 error = ENOENT;
 246                                 break;
 247                         }
 248
 249                         /*
 250                          * Better not be zero
 251                          */
 252                         KKASSERT(elm->internal.subtree_offset != 0);
 253
 254                         if (s <= 0) {
 255                                 /*
 256                                  * If running the mirror filter see if we
 257                                  * can skip one or more entire sub-trees.
 258                                  * If we can we return the internal node
 259                                  * and the caller processes the skipped
 260                                  * range (see mirror_read).
 261                                  */
 262                                 if (cursor->flags &
 263                                     HAMMER_CURSOR_MIRROR_FILTERED) {
 264                                         if (elm->internal.mirror_tid <
 265                                             cursor->cmirror->mirror_tid) {
 266                                                 hammer_cursor_mirror_filter(cursor);
 267                                                 return(0);
 268                                         }
 269                                 }
 270                         } else {
 271                                 /*
 272                                  * Normally it would be impossible for the
 273                                  * cursor to have gotten back-indexed,
 274                                  * but it can happen if a node is deleted
 275                                  * and the cursor is moved to its parent
 276                                  * internal node.  ITERATE_CHECK will be set.
 277                                  */
 278                                 KKASSERT(cursor->flags &
 279                                          HAMMER_CURSOR_ITERATE_CHECK);
 280                                 kprintf("hammer_btree_iterate: "
 281                                         "DEBUG: Caught parent seek "
 282                                         "in internal iteration\n");
 283                         }
 284
 285                         error = hammer_cursor_down(cursor);
 286                         if (error)
 287                                 break;
 288                         KKASSERT(cursor->index == 0);
 289                         /* reload stale pointer */
 290                         node = cursor->node->ondisk;
 291                         continue;
 292                 } else {
 293                         elm = &node->elms[cursor->index];
 294                         r = hammer_btree_cmp(&cursor->key_end, &elm->base);
 295                         if (hammer_debug_btree) {
 296                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x "
 297                                         "key=%016llx lo=%02x %d\n",
 298                                         (long long)cursor->node->node_offset,
 299                                         cursor->index,
 300                                         (elm[0].leaf.base.btype ?
 301                                          elm[0].leaf.base.btype : '?'),
 302                                         (long long)elm[0].leaf.base.obj_id,
 303                                         elm[0].leaf.base.rec_type,
 304                                         (long long)elm[0].leaf.base.key,
 305                                         elm[0].leaf.base.localization,
 306                                         r
 307                                 );
 308                         }
 309                         if (r < 0) {
 310                                 error = ENOENT;
 311                                 break;
 312                         }
 313
 314                         /*
 315                          * We support both end-inclusive and
 316                          * end-exclusive searches.
 317                          */
 318                         if (r == 0 &&
 319                            (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 320                                 error = ENOENT;
 321                                 break;
 322                         }
 323
 324                         /*
 325                          * If ITERATE_CHECK is set an unlocked cursor may
 326                          * have been moved to a parent and the iterate can
 327                          * happen upon elements that are not in the requested
 328                          * range.
 329                          */
 330                         if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) {
 331                                 s = hammer_btree_cmp(&cursor->key_beg,
 332                                                      &elm->base);
 333                                 if (s > 0) {
 334                                         kprintf("hammer_btree_iterate: "
 335                                                 "DEBUG: Caught parent seek "
 336                                                 "in leaf iteration\n");
 337                                         ++cursor->index;
 338                                         continue;
 339                                 }
 340                         }
 341                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 342
 343                         /*
 344                          * Return the element
 345                          */
 346                         switch(elm->leaf.base.btype) {
 347                         case HAMMER_BTREE_TYPE_RECORD:
 348                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 349                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 350                                         ++cursor->index;
 351                                         continue;
 352                                 }
 353                                 error = 0;
 354                                 break;
 355                         default:
 356                                 error = EINVAL;
 357                                 break;
 358                         }
 359                         if (error)
 360                                 break;
 361                 }
 362
 363                 /*
 364                  * Return entry
 365                  */
 366                 if (hammer_debug_btree) {
 367                         int i = cursor->index;
 368                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 369                         kprintf("ITERATE  %p:%d %016llx %02x "
 370                                 "key=%016llx lo=%02x\n",
 371                                 cursor->node, i,
 372                                 (long long)elm->internal.base.obj_id,
 373                                 elm->internal.base.rec_type,
 374                                 (long long)elm->internal.base.key,
 375                                 elm->internal.base.localization
 376                         );
 377                 }
 378                 return(0);
 379         }
 380         return(error);
 381 }
 382
 383 /*
 384  * We hit an internal element that we could skip as part of a mirroring
 385  * scan.  Calculate the entire range being skipped.
 386  *
 387  * It is important to include any gaps between the parent's left_bound
 388  * and the node's left_bound, and same goes for the right side.
 389  */
 390 static void
 391 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
 392 {
 393         struct hammer_cmirror *cmirror;
 394         hammer_node_ondisk_t ondisk;
 395         hammer_btree_elm_t elm;
 396
 397         ondisk = cursor->node->ondisk;
 398         cmirror = cursor->cmirror;
 399
 400         /*
 401          * Calculate the skipped range
 402          */
 403         elm = &ondisk->elms[cursor->index];
 404         if (cursor->index == 0)
 405                 cmirror->skip_beg = *cursor->left_bound;
 406         else
 407                 cmirror->skip_beg = elm->internal.base;
 408         while (cursor->index < ondisk->count) {
 409                 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
 410                         break;
 411                 ++cursor->index;
 412                 ++elm;
 413         }
 414         if (cursor->index == ondisk->count)
 415                 cmirror->skip_end = *cursor->right_bound;
 416         else
 417                 cmirror->skip_end = elm->internal.base;
 418
 419         /*
 420          * clip the returned result.
 421          */
 422         if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
 423                 cmirror->skip_beg = cursor->key_beg;
 424         if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
 425                 cmirror->skip_end = cursor->key_end;
 426 }
 427
 428 /*
 429  * Iterate in the reverse direction.  This is used by the pruning code to
 430  * avoid overlapping records.
 431  */
 432 int
 433 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
 434 {
 435         hammer_node_ondisk_t node;
 436         hammer_btree_elm_t elm;
 437         hammer_mount_t hmp;
 438         int error = 0;
 439         int r;
 440         int s;
 441
 442         /* mirror filtering not supported for reverse iteration */
 443         KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
 444
 445         /*
 446          * Skip past the current record.  For various reasons the cursor
 447          * may end up set to -1 or set to point at the end of the current
 448          * node.  These cases must be addressed.
 449          */
 450         node = cursor->node->ondisk;
 451         if (node == NULL)
 452                 return(ENOENT);
 453         if (cursor->index != -1 &&
 454             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 455                 --cursor->index;
 456         }
 457         if (cursor->index == cursor->node->ondisk->count)
 458                 --cursor->index;
 459
 460         /*
 461          * HAMMER can wind up being cpu-bound.
 462          */
 463         hmp = cursor->trans->hmp;
 464         if (++hmp->check_yield > hammer_yield_check) {
 465                 hmp->check_yield = 0;
 466                 lwkt_user_yield();
 467         }
 468
 469         /*
 470          * Loop until an element is found or we are done.
 471          */
 472         for (;;) {
 473                 ++hammer_stats_btree_iterations;
 474                 hammer_flusher_clean_loose_ios(hmp);
 475
 476                 /*
 477                  * We iterate up the tree and then index over one element
 478                  * while we are at the last element in the current node.
 479                  */
 480                 if (cursor->index == -1) {
 481                         error = hammer_cursor_up(cursor);
 482                         if (error) {
 483                                 cursor->index = 0; /* sanity */
 484                                 break;
 485                         }
 486                         /* reload stale pointer */
 487                         node = cursor->node->ondisk;
 488                         KKASSERT(cursor->index != node->count);
 489                         --cursor->index;
 490                         continue;
 491                 }
 492
 493                 /*
 494                  * Check internal or leaf element.  Determine if the record
 495                  * at the cursor has gone beyond the end of our range.
 496                  *
 497                  * We recurse down through internal nodes.
 498                  */
 499                 KKASSERT(cursor->index != node->count);
 500                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 501                         elm = &node->elms[cursor->index];
 502
 503                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 504                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 505                         if (hammer_debug_btree) {
 506                                 kprintf("BRACKETL %016llx[%d] %016llx %02x "
 507                                         "key=%016llx lo=%02x %d (td=%p)\n",
 508                                         (long long)cursor->node->node_offset,
 509                                         cursor->index,
 510                                         (long long)elm[0].internal.base.obj_id,
 511                                         elm[0].internal.base.rec_type,
 512                                         (long long)elm[0].internal.base.key,
 513                                         elm[0].internal.base.localization,
 514                                         r,
 515                                         curthread
 516                                 );
 517                                 kprintf("BRACKETR %016llx[%d] %016llx %02x "
 518                                         "key=%016llx lo=%02x %d\n",
 519                                         (long long)cursor->node->node_offset,
 520                                         cursor->index + 1,
 521                                         (long long)elm[1].internal.base.obj_id,
 522                                         elm[1].internal.base.rec_type,
 523                                         (long long)elm[1].internal.base.key,
 524                                         elm[1].internal.base.localization,
 525                                         s
 526                                 );
 527                         }
 528
 529                         if (s >= 0) {
 530                                 error = ENOENT;
 531                                 break;
 532                         }
 533
 534                         /*
 535                          * It shouldn't be possible to be seeked past key_end,
 536                          * even if the cursor got moved to a parent.
 537                          */
 538                         KKASSERT(r >= 0);
 539
 540                         /*
 541                          * Better not be zero
 542                          */
 543                         KKASSERT(elm->internal.subtree_offset != 0);
 544
 545                         error = hammer_cursor_down(cursor);
 546                         if (error)
 547                                 break;
 548                         KKASSERT(cursor->index == 0);
 549                         /* reload stale pointer */
 550                         node = cursor->node->ondisk;
 551
 552                         /* this can assign -1 if the leaf was empty */
 553                         cursor->index = node->count - 1;
 554                         continue;
 555                 } else {
 556                         elm = &node->elms[cursor->index];
 557                         s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
 558                         if (hammer_debug_btree) {
 559                                 kprintf("ELEMENTR %016llx:%d %c %016llx %02x "
 560                                         "key=%016llx lo=%02x %d\n",
 561                                         (long long)cursor->node->node_offset,
 562                                         cursor->index,
 563                                         (elm[0].leaf.base.btype ?
 564                                          elm[0].leaf.base.btype : '?'),
 565                                         (long long)elm[0].leaf.base.obj_id,
 566                                         elm[0].leaf.base.rec_type,
 567                                         (long long)elm[0].leaf.base.key,
 568                                         elm[0].leaf.base.localization,
 569                                         s
 570                                 );
 571                         }
 572                         if (s > 0) {
 573                                 error = ENOENT;
 574                                 break;
 575                         }
 576
 577                         /*
 578                          * It shouldn't be possible to be seeked past key_end,
 579                          * even if the cursor got moved to a parent.
 580                          */
 581                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 582
 583                         /*
 584                          * Return the element
 585                          */
 586                         switch(elm->leaf.base.btype) {
 587                         case HAMMER_BTREE_TYPE_RECORD:
 588                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 589                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 590                                         --cursor->index;
 591                                         continue;
 592                                 }
 593                                 error = 0;
 594                                 break;
 595                         default:
 596                                 error = EINVAL;
 597                                 break;
 598                         }
 599                         if (error)
 600                                 break;
 601                 }
 602
 603                 /*
 604                  * Return entry
 605                  */
 606                 if (hammer_debug_btree) {
 607                         int i = cursor->index;
 608                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 609                         kprintf("ITERATER %p:%d %016llx %02x "
 610                                 "key=%016llx lo=%02x\n",
 611                                 cursor->node, i,
 612                                 (long long)elm->internal.base.obj_id,
 613                                 elm->internal.base.rec_type,
 614                                 (long long)elm->internal.base.key,
 615                                 elm->internal.base.localization
 616                         );
 617                 }
 618                 return(0);
 619         }
 620         return(error);
 621 }
 622
 623 /*
 624  * Lookup cursor->key_beg.  0 is returned on success, ENOENT if the entry
 625  * could not be found, EDEADLK if inserting and a retry is needed, and a
 626  * fatal error otherwise.  When retrying, the caller must terminate the
 627  * cursor and reinitialize it.  EDEADLK cannot be returned if not inserting.
 628  *
 629  * The cursor is suitably positioned for a deletion on success, and suitably
 630  * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
 631  * specified.
 632  *
 633  * The cursor may begin anywhere, the search will traverse the tree in
 634  * either direction to locate the requested element.
 635  *
 636  * Most of the logic implementing historical searches is handled here.  We
 637  * do an initial lookup with create_tid set to the asof TID.  Due to the
 638  * way records are laid out, a backwards iteration may be required if
 639  * ENOENT is returned to locate the historical record.  Here's the
 640  * problem:
 641  *
 642  * create_tid:    10      15       20
 643  *                   LEAF1   LEAF2
 644  * records:         (11)        (18)
 645  *
 646  * Lets say we want to do a lookup AS-OF timestamp 17.  We will traverse
 647  * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
 648  * not visible and thus causes ENOENT to be returned.  We really need
 649  * to check record 11 in LEAF1.  If it also fails then the search fails
 650  * (e.g. it might represent the range 11-16 and thus still not match our
 651  * AS-OF timestamp of 17).  Note that LEAF1 could be empty, requiring
 652  * further iterations.
 653  *
 654  * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
 655  * and the cursor->create_check TID if an iteration might be needed.
 656  * In the above example create_check would be set to 14.
 657  */
 658 int
 659 hammer_btree_lookup(hammer_cursor_t cursor)
 660 {
 661         int error;
 662
 663         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 664         KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
 665                   cursor->trans->sync_lock_refs > 0);
 666         ++hammer_stats_btree_lookups;
 667         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 668                 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
 669                 cursor->key_beg.create_tid = cursor->asof;
 670                 for (;;) {
 671                         cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
 672                         error = btree_search(cursor, 0);
 673                         if (error != ENOENT ||
 674                             (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
 675                                 /*
 676                                  * Stop if no error.
 677                                  * Stop if error other then ENOENT.
 678                                  * Stop if ENOENT and not special case.
 679                                  */
 680                                 break;
 681                         }
 682                         if (hammer_debug_btree) {
 683                                 kprintf("CREATE_CHECK %016llx\n",
 684                                         (long long)cursor->create_check);
 685                         }
 686                         cursor->key_beg.create_tid = cursor->create_check;
 687                         /* loop */
 688                 }
 689         } else {
 690                 error = btree_search(cursor, 0);
 691         }
 692         if (error == 0)
 693                 error = hammer_btree_extract(cursor, cursor->flags);
 694         return(error);
 695 }
 696
 697 /*
 698  * Execute the logic required to start an iteration.  The first record
 699  * located within the specified range is returned and iteration control
 700  * flags are adjusted for successive hammer_btree_iterate() calls.
 701  *
 702  * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
 703  * in a loop without worrying about it.  Higher-level merged searches will
 704  * adjust the flag appropriately.
 705  */
 706 int
 707 hammer_btree_first(hammer_cursor_t cursor)
 708 {
 709         int error;
 710
 711         error = hammer_btree_lookup(cursor);
 712         if (error == ENOENT) {
 713                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 714                 error = hammer_btree_iterate(cursor);
 715         }
 716         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 717         return(error);
 718 }
 719
 720 /*
 721  * Similarly but for an iteration in the reverse direction.
 722  *
 723  * Set ATEDISK when iterating backwards to skip the current entry,
 724  * which after an ENOENT lookup will be pointing beyond our end point.
 725  *
 726  * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
 727  * in a loop without worrying about it.  Higher-level merged searches will
 728  * adjust the flag appropriately.
 729  */
 730 int
 731 hammer_btree_last(hammer_cursor_t cursor)
 732 {
 733         struct hammer_base_elm save;
 734         int error;
 735
 736         save = cursor->key_beg;
 737         cursor->key_beg = cursor->key_end;
 738         error = hammer_btree_lookup(cursor);
 739         cursor->key_beg = save;
 740         if (error == ENOENT ||
 741             (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 742                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 743                 error = hammer_btree_iterate_reverse(cursor);
 744         }
 745         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 746         return(error);
 747 }
 748
 749 /*
 750  * Extract the record and/or data associated with the cursor's current
 751  * position.  Any prior record or data stored in the cursor is replaced.
 752  *
 753  * NOTE: All extractions occur at the leaf of the B-Tree.
 754  */
 755 int
 756 hammer_btree_extract(hammer_cursor_t cursor, int flags)
 757 {
 758         hammer_node_ondisk_t node;
 759         hammer_btree_elm_t elm;
 760         hammer_off_t data_off;
 761         hammer_mount_t hmp;
 762         int32_t data_len;
 763         int error;
 764
 765         /*
 766          * The case where the data reference resolves to the same buffer
 767          * as the record reference must be handled.
 768          */
 769         node = cursor->node->ondisk;
 770         elm = &node->elms[cursor->index];
 771         cursor->data = NULL;
 772         hmp = cursor->node->hmp;
 773
 774         /*
 775          * There is nothing to extract for an internal element.
 776          */
 777         if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
 778                 return(EINVAL);
 779
 780         /*
 781          * Only record types have data.
 782          */
 783         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 784         cursor->leaf = &elm->leaf;
 785
 786         if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
 787                 return(0);
 788         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 789                 return(0);
 790         data_off = elm->leaf.data_offset;
 791         data_len = elm->leaf.data_len;
 792         if (data_off == 0)
 793                 return(0);
 794
 795         /*
 796          * Load the data
 797          */
 798         KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
 799         cursor->data = hammer_bread_ext(hmp, data_off, data_len,
 800                                         &error, &cursor->data_buffer);
 801
 802         /*
 803          * Mark the data buffer as not being meta-data if it isn't
 804          * meta-data (sometimes bulk data is accessed via a volume
 805          * block device).
 806          */
 807         if (error == 0) {
 808                 switch(elm->leaf.base.rec_type) {
 809                 case HAMMER_RECTYPE_DATA:
 810                 case HAMMER_RECTYPE_DB:
 811                         if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0)
 812                                 break;
 813                         if (hammer_double_buffer == 0 ||
 814                             (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) {
 815                                 hammer_io_notmeta(cursor->data_buffer);
 816                         }
 817                         break;
 818                 default:
 819                         break;
 820                 }
 821         }
 822
 823         /*
 824          * Deal with CRC errors on the extracted data.
 825          */
 826         if (error == 0 &&
 827             hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
 828                 kprintf("CRC DATA @ %016llx/%d FAILED\n",
 829                         (long long)elm->leaf.data_offset, elm->leaf.data_len);
 830                 if (hammer_debug_critical)
 831                         Debugger("CRC FAILED: DATA");
 832                 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
 833                         error = EDOM;   /* less critical (mirroring) */
 834                 else
 835                         error = EIO;    /* critical */
 836         }
 837         return(error);
 838 }
 839
 840
 841 /*
 842  * Insert a leaf element into the B-Tree at the current cursor position.
 843  * The cursor is positioned such that the element at and beyond the cursor
 844  * are shifted to make room for the new record.
 845  *
 846  * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
 847  * flag set and that call must return ENOENT before this function can be
 848  * called. ENOSPC is returned if there is no room to insert a new record.
 849  *
 850  * The caller may depend on the cursor's exclusive lock after return to
 851  * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
 852  */
 853 int
 854 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
 855                     int *doprop)
 856 {
 857         hammer_node_ondisk_t node;
 858         int i;
 859         int error;
 860
 861         *doprop = 0;
 862         if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
 863                 return(error);
 864         ++hammer_stats_btree_inserts;
 865
 866         /*
 867          * Insert the element at the leaf node and update the count in the
 868          * parent.  It is possible for parent to be NULL, indicating that
 869          * the filesystem's ROOT B-Tree node is a leaf itself, which is
 870          * possible.  The root inode can never be deleted so the leaf should
 871          * never be empty.
 872          *
 873          * Remember that leaf nodes do not have boundaries.
 874          */
 875         hammer_modify_node_all(cursor->trans, cursor->node);
 876         node = cursor->node->ondisk;
 877         i = cursor->index;
 878         KKASSERT(elm->base.btype != 0);
 879         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 880         KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
 881         if (i != node->count) {
 882                 bcopy(&node->elms[i], &node->elms[i+1],
 883                       (node->count - i) * sizeof(*elm));
 884         }
 885         node->elms[i].leaf = *elm;
 886         ++node->count;
 887         hammer_cursor_inserted_element(cursor->node, i);
 888
 889         /*
 890          * Update the leaf node's aggregate mirror_tid for mirroring
 891          * support.
 892          */
 893         if (node->mirror_tid < elm->base.delete_tid) {
 894                 node->mirror_tid = elm->base.delete_tid;
 895                 *doprop = 1;
 896         }
 897         if (node->mirror_tid < elm->base.create_tid) {
 898                 node->mirror_tid = elm->base.create_tid;
 899                 *doprop = 1;
 900         }
 901         hammer_modify_node_done(cursor->node);
 902
 903         /*
 904          * Debugging sanity checks.
 905          */
 906         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
 907         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
 908         if (i) {
 909                 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
 910         }
 911         if (i != node->count - 1)
 912                 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
 913
 914         return(0);
 915 }
 916
 917 /*
 918  * Delete a record from the B-Tree at the current cursor position.
 919  * The cursor is positioned such that the current element is the one
 920  * to be deleted.
 921  *
 922  * On return the cursor will be positioned after the deleted element and
 923  * MAY point to an internal node.  It will be suitable for the continuation
 924  * of an iteration but not for an insertion or deletion.
 925  *
 926  * Deletions will attempt to partially rebalance the B-Tree in an upward
 927  * direction, but will terminate rather then deadlock.  Empty internal nodes
 928  * are never allowed by a deletion which deadlocks may end up giving us an
 929  * empty leaf.  The pruner will clean up and rebalance the tree.
 930  *
 931  * This function can return EDEADLK, requiring the caller to retry the
 932  * operation after clearing the deadlock.
 933  */
 934 int
 935 hammer_btree_delete(hammer_cursor_t cursor)
 936 {
 937         hammer_node_ondisk_t ondisk;
 938         hammer_node_t node;
 939         hammer_node_t parent __debugvar;
 940         int error;
 941         int i;
 942
 943         KKASSERT (cursor->trans->sync_lock_refs > 0);
 944         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 945                 return(error);
 946         ++hammer_stats_btree_deletes;
 947
 948         /*
 949          * Delete the element from the leaf node.
 950          *
 951          * Remember that leaf nodes do not have boundaries.
 952          */
 953         node = cursor->node;
 954         ondisk = node->ondisk;
 955         i = cursor->index;
 956
 957         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
 958         KKASSERT(i >= 0 && i < ondisk->count);
 959         hammer_modify_node_all(cursor->trans, node);
 960         if (i + 1 != ondisk->count) {
 961                 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
 962                       (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
 963         }
 964         --ondisk->count;
 965         hammer_modify_node_done(node);
 966         hammer_cursor_deleted_element(node, i);
 967
 968         /*
 969          * Validate local parent
 970          */
 971         if (ondisk->parent) {
 972                 parent = cursor->parent;
 973
 974                 KKASSERT(parent != NULL);
 975                 KKASSERT(parent->node_offset == ondisk->parent);
 976         }
 977
 978         /*
 979          * If the leaf becomes empty it must be detached from the parent,
 980          * potentially recursing through to the filesystem root.
 981          *
 982          * This may reposition the cursor at one of the parent's of the
 983          * current node.
 984          *
 985          * Ignore deadlock errors, that simply means that btree_remove
 986          * was unable to recurse and had to leave us with an empty leaf.
 987          */
 988         KKASSERT(cursor->index <= ondisk->count);
 989         if (ondisk->count == 0) {
 990                 error = btree_remove(cursor);
 991                 if (error == EDEADLK)
 992                         error = 0;
 993         } else {
 994                 error = 0;
 995         }
 996         KKASSERT(cursor->parent == NULL ||
 997                  cursor->parent_index < cursor->parent->ondisk->count);
 998         return(error);
 999 }
1000
1001 /*
1002  * PRIMARY B-TREE SEARCH SUPPORT PROCEDURE
1003  *
1004  * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
1005  *
1006  * The search can begin ANYWHERE in the B-Tree.  As a first step the search
1007  * iterates up the tree as necessary to properly position itself prior to
1008  * actually doing the sarch.
1009  *
1010  * INSERTIONS: The search will split full nodes and leaves on its way down
1011  * and guarentee that the leaf it ends up on is not full.  If we run out
1012  * of space the search continues to the leaf, but ENOSPC is returned.
1013  *
1014  * The search is only guarenteed to end up on a leaf if an error code of 0
1015  * is returned, or if inserting and an error code of ENOENT is returned.
1016  * Otherwise it can stop at an internal node.  On success a search returns
1017  * a leaf node.
1018  *
1019  * COMPLEXITY WARNING!  This is the core B-Tree search code for the entire
1020  * filesystem, and it is not simple code.  Please note the following facts:
1021  *
1022  * - Internal node recursions have a boundary on the left AND right.  The
1023  *   right boundary is non-inclusive.  The create_tid is a generic part
1024  *   of the key for internal nodes.
1025  *
1026  * - Leaf nodes contain terminal elements only now.
1027  *
1028  * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1029  *   historical search.  ASOF and INSERT are mutually exclusive.  When
1030  *   doing an as-of lookup btree_search() checks for a right-edge boundary
1031  *   case.  If while recursing down the left-edge differs from the key
1032  *   by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1033  *   with cursor->create_check.  This is used by btree_lookup() to iterate.
1034  *   The iteration backwards because as-of searches can wind up going
1035  *   down the wrong branch of the B-Tree.
1036  */
1037 static
1038 int
1039 btree_search(hammer_cursor_t cursor, int flags)
1040 {
1041         hammer_node_ondisk_t node;
1042         hammer_btree_elm_t elm;
1043         int error;
1044         int enospc = 0;
1045         int i;
1046         int r;
1047         int s;
1048
1049         flags |= cursor->flags;
1050         ++hammer_stats_btree_searches;
1051
1052         if (hammer_debug_btree) {
1053                 kprintf("SEARCH   %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td=%p)\n",
1054                         (long long)cursor->node->node_offset,
1055                         cursor->index,
1056                         (long long)cursor->key_beg.obj_id,
1057                         cursor->key_beg.rec_type,
1058                         (long long)cursor->key_beg.key,
1059                         (long long)cursor->key_beg.create_tid,
1060                         cursor->key_beg.localization,
1061                         curthread
1062                 );
1063                 if (cursor->parent)
1064                     kprintf("SEARCHP  %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
1065                         (long long)cursor->parent->node_offset,
1066                         cursor->parent_index,
1067                         (long long)cursor->left_bound->obj_id,
1068                         (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
1069                         (long long)cursor->right_bound->obj_id,
1070                         (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
1071                         cursor->left_bound,
1072                         &cursor->parent->ondisk->elms[cursor->parent_index],
1073                         cursor->right_bound,
1074                         &cursor->parent->ondisk->elms[cursor->parent_index+1]
1075                     );
1076         }
1077
1078         /*
1079          * Move our cursor up the tree until we find a node whos range covers
1080          * the key we are trying to locate.
1081          *
1082          * The left bound is inclusive, the right bound is non-inclusive.
1083          * It is ok to cursor up too far.
1084          */
1085         for (;;) {
1086                 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1087                 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1088                 if (r >= 0 && s < 0)
1089                         break;
1090                 KKASSERT(cursor->parent);
1091                 ++hammer_stats_btree_iterations;
1092                 error = hammer_cursor_up(cursor);
1093                 if (error)
1094                         goto done;
1095         }
1096
1097         /*
1098          * The delete-checks below are based on node, not parent.  Set the
1099          * initial delete-check based on the parent.
1100          */
1101         if (r == 1) {
1102                 KKASSERT(cursor->left_bound->create_tid != 1);
1103                 cursor->create_check = cursor->left_bound->create_tid - 1;
1104                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1105         }
1106
1107         /*
1108          * We better have ended up with a node somewhere.
1109          */
1110         KKASSERT(cursor->node != NULL);
1111
1112         /*
1113          * If we are inserting we can't start at a full node if the parent
1114          * is also full (because there is no way to split the node),
1115          * continue running up the tree until the requirement is satisfied
1116          * or we hit the root of the filesystem.
1117          *
1118          * (If inserting we aren't doing an as-of search so we don't have
1119          *  to worry about create_check).
1120          */
1121         while (flags & HAMMER_CURSOR_INSERT) {
1122                 if (btree_node_is_full(cursor->node->ondisk) == 0)
1123                         break;
1124                 if (cursor->node->ondisk->parent == 0 ||
1125                     cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1126                         break;
1127                 }
1128                 ++hammer_stats_btree_iterations;
1129                 error = hammer_cursor_up(cursor);
1130                 /* node may have become stale */
1131                 if (error)
1132                         goto done;
1133         }
1134
1135         /*
1136          * Push down through internal nodes to locate the requested key.
1137          */
1138         node = cursor->node->ondisk;
1139         while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1140                 /*
1141                  * Scan the node to find the subtree index to push down into.
1142                  * We go one-past, then back-up.
1143                  *
1144                  * We must proactively remove deleted elements which may
1145                  * have been left over from a deadlocked btree_remove().
1146                  *
1147                  * The left and right boundaries are included in the loop
1148                  * in order to detect edge cases.
1149                  *
1150                  * If the separator only differs by create_tid (r == 1)
1151                  * and we are doing an as-of search, we may end up going
1152                  * down a branch to the left of the one containing the
1153                  * desired key.  This requires numerous special cases.
1154                  */
1155                 ++hammer_stats_btree_iterations;
1156                 if (hammer_debug_btree) {
1157                         kprintf("SEARCH-I %016llx count=%d\n",
1158                                 (long long)cursor->node->node_offset,
1159                                 node->count);
1160                 }
1161
1162                 /*
1163                  * Try to shortcut the search before dropping into the
1164                  * linear loop.  Locate the first node where r <= 1.
1165                  */
1166                 i = hammer_btree_search_node(&cursor->key_beg, node);
1167                 while (i <= node->count) {
1168                         ++hammer_stats_btree_elements;
1169                         elm = &node->elms[i];
1170                         r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1171                         if (hammer_debug_btree > 2) {
1172                                 kprintf(" IELM %p %d r=%d\n",
1173                                         &node->elms[i], i, r);
1174                         }
1175                         if (r < 0)
1176                                 break;
1177                         if (r == 1) {
1178                                 KKASSERT(elm->base.create_tid != 1);
1179                                 cursor->create_check = elm->base.create_tid - 1;
1180                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1181                         }
1182                         ++i;
1183                 }
1184                 if (hammer_debug_btree) {
1185                         kprintf("SEARCH-I preI=%d/%d r=%d\n",
1186                                 i, node->count, r);
1187                 }
1188
1189                 /*
1190                  * These cases occur when the parent's idea of the boundary
1191                  * is wider then the child's idea of the boundary, and
1192                  * require special handling.  If not inserting we can
1193                  * terminate the search early for these cases but the
1194                  * child's boundaries cannot be unconditionally modified.
1195                  */
1196                 if (i == 0) {
1197                         /*
1198                          * If i == 0 the search terminated to the LEFT of the
1199                          * left_boundary but to the RIGHT of the parent's left
1200                          * boundary.
1201                          */
1202                         u_int8_t save;
1203
1204                         elm = &node->elms[0];
1205
1206                         /*
1207                          * If we aren't inserting we can stop here.
1208                          */
1209                         if ((flags & (HAMMER_CURSOR_INSERT |
1210                                       HAMMER_CURSOR_PRUNING)) == 0) {
1211                                 cursor->index = 0;
1212                                 return(ENOENT);
1213                         }
1214
1215                         /*
1216                          * Correct a left-hand boundary mismatch.
1217                          *
1218                          * We can only do this if we can upgrade the lock,
1219                          * and synchronized as a background cursor (i.e.
1220                          * inserting or pruning).
1221                          *
1222                          * WARNING: We can only do this if inserting, i.e.
1223                          * we are running on the backend.
1224                          */
1225                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1226                                 return(error);
1227                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1228                         hammer_modify_node_field(cursor->trans, cursor->node,
1229                                                  elms[0]);
1230                         save = node->elms[0].base.btype;
1231                         node->elms[0].base = *cursor->left_bound;
1232                         node->elms[0].base.btype = save;
1233                         hammer_modify_node_done(cursor->node);
1234                 } else if (i == node->count + 1) {
1235                         /*
1236                          * If i == node->count + 1 the search terminated to
1237                          * the RIGHT of the right boundary but to the LEFT
1238                          * of the parent's right boundary.  If we aren't
1239                          * inserting we can stop here.
1240                          *
1241                          * Note that the last element in this case is
1242                          * elms[i-2] prior to adjustments to 'i'.
1243                          */
1244                         --i;
1245                         if ((flags & (HAMMER_CURSOR_INSERT |
1246                                       HAMMER_CURSOR_PRUNING)) == 0) {
1247                                 cursor->index = i;
1248                                 return (ENOENT);
1249                         }
1250
1251                         /*
1252                          * Correct a right-hand boundary mismatch.
1253                          * (actual push-down record is i-2 prior to
1254                          * adjustments to i).
1255                          *
1256                          * We can only do this if we can upgrade the lock,
1257                          * and synchronized as a background cursor (i.e.
1258                          * inserting or pruning).
1259                          *
1260                          * WARNING: We can only do this if inserting, i.e.
1261                          * we are running on the backend.
1262                          */
1263                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1264                                 return(error);
1265                         elm = &node->elms[i];
1266                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1267                         hammer_modify_node(cursor->trans, cursor->node,
1268                                            &elm->base, sizeof(elm->base));
1269                         elm->base = *cursor->right_bound;
1270                         hammer_modify_node_done(cursor->node);
1271                         --i;
1272                 } else {
1273                         /*
1274                          * The push-down index is now i - 1.  If we had
1275                          * terminated on the right boundary this will point
1276                          * us at the last element.
1277                          */
1278                         --i;
1279                 }
1280                 cursor->index = i;
1281                 elm = &node->elms[i];
1282
1283                 if (hammer_debug_btree) {
1284                         kprintf("RESULT-I %016llx[%d] %016llx %02x "
1285                                 "key=%016llx cre=%016llx lo=%02x\n",
1286                                 (long long)cursor->node->node_offset,
1287                                 i,
1288                                 (long long)elm->internal.base.obj_id,
1289                                 elm->internal.base.rec_type,
1290                                 (long long)elm->internal.base.key,
1291                                 (long long)elm->internal.base.create_tid,
1292                                 elm->internal.base.localization
1293                         );
1294                 }
1295
1296                 /*
1297                  * We better have a valid subtree offset.
1298                  */
1299                 KKASSERT(elm->internal.subtree_offset != 0);
1300
1301                 /*
1302                  * Handle insertion and deletion requirements.
1303                  *
1304                  * If inserting split full nodes.  The split code will
1305                  * adjust cursor->node and cursor->index if the current
1306                  * index winds up in the new node.
1307                  *
1308                  * If inserting and a left or right edge case was detected,
1309                  * we cannot correct the left or right boundary and must
1310                  * prepend and append an empty leaf node in order to make
1311                  * the boundary correction.
1312                  *
1313                  * If we run out of space we set enospc but continue on
1314                  * to a leaf.
1315                  */
1316                 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1317                         if (btree_node_is_full(node)) {
1318                                 error = btree_split_internal(cursor);
1319                                 if (error) {
1320                                         if (error != ENOSPC)
1321                                                 goto done;
1322                                         enospc = 1;
1323                                 }
1324                                 /*
1325                                  * reload stale pointers
1326                                  */
1327                                 i = cursor->index;
1328                                 node = cursor->node->ondisk;
1329                         }
1330                 }
1331
1332                 /*
1333                  * Push down (push into new node, existing node becomes
1334                  * the parent) and continue the search.
1335                  */
1336                 error = hammer_cursor_down(cursor);
1337                 /* node may have become stale */
1338                 if (error)
1339                         goto done;
1340                 node = cursor->node->ondisk;
1341         }
1342
1343         /*
1344          * We are at a leaf, do a linear search of the key array.
1345          *
1346          * On success the index is set to the matching element and 0
1347          * is returned.
1348          *
1349          * On failure the index is set to the insertion point and ENOENT
1350          * is returned.
1351          *
1352          * Boundaries are not stored in leaf nodes, so the index can wind
1353          * up to the left of element 0 (index == 0) or past the end of
1354          * the array (index == node->count).  It is also possible that the
1355          * leaf might be empty.
1356          */
1357         ++hammer_stats_btree_iterations;
1358         KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1359         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1360         if (hammer_debug_btree) {
1361                 kprintf("SEARCH-L %016llx count=%d\n",
1362                         (long long)cursor->node->node_offset,
1363                         node->count);
1364         }
1365
1366         /*
1367          * Try to shortcut the search before dropping into the
1368          * linear loop.  Locate the first node where r <= 1.
1369          */
1370         i = hammer_btree_search_node(&cursor->key_beg, node);
1371         while (i < node->count) {
1372                 ++hammer_stats_btree_elements;
1373                 elm = &node->elms[i];
1374
1375                 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1376
1377                 if (hammer_debug_btree > 1)
1378                         kprintf("  ELM %p %d r=%d\n", &node->elms[i], i, r);
1379
1380                 /*
1381                  * We are at a record element.  Stop if we've flipped past
1382                  * key_beg, not counting the create_tid test.  Allow the
1383                  * r == 1 case (key_beg > element but differs only by its
1384                  * create_tid) to fall through to the AS-OF check.
1385                  */
1386                 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1387
1388                 if (r < 0)
1389                         goto failed;
1390                 if (r > 1) {
1391                         ++i;
1392                         continue;
1393                 }
1394
1395                 /*
1396                  * Check our as-of timestamp against the element.
1397                  */
1398                 if (flags & HAMMER_CURSOR_ASOF) {
1399                         if (hammer_btree_chkts(cursor->asof,
1400                                                &node->elms[i].base) != 0) {
1401                                 ++i;
1402                                 continue;
1403                         }
1404                         /* success */
1405                 } else {
1406                         if (r > 0) {    /* can only be +1 */
1407                                 ++i;
1408                                 continue;
1409                         }
1410                         /* success */
1411                 }
1412                 cursor->index = i;
1413                 error = 0;
1414                 if (hammer_debug_btree) {
1415                         kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1416                                 (long long)cursor->node->node_offset, i);
1417                 }
1418                 goto done;
1419         }
1420
1421         /*
1422          * The search of the leaf node failed.  i is the insertion point.
1423          */
1424 failed:
1425         if (hammer_debug_btree) {
1426                 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1427                         (long long)cursor->node->node_offset, i);
1428         }
1429
1430         /*
1431          * No exact match was found, i is now at the insertion point.
1432          *
1433          * If inserting split a full leaf before returning.  This
1434          * may have the side effect of adjusting cursor->node and
1435          * cursor->index.
1436          */
1437         cursor->index = i;
1438         if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1439              btree_node_is_full(node)) {
1440                 error = btree_split_leaf(cursor);
1441                 if (error) {
1442                         if (error != ENOSPC)
1443                                 goto done;
1444                         enospc = 1;
1445                 }
1446                 /*
1447                  * reload stale pointers
1448                  */
1449                 /* NOT USED
1450                 i = cursor->index;
1451                 node = &cursor->node->internal;
1452                 */
1453         }
1454
1455         /*
1456          * We reached a leaf but did not find the key we were looking for.
1457          * If this is an insert we will be properly positioned for an insert
1458          * (ENOENT) or unable to insert (ENOSPC).
1459          */
1460         error = enospc ? ENOSPC : ENOENT;
1461 done:
1462         return(error);
1463 }
1464
1465 /*
1466  * Heuristical search for the first element whos comparison is <= 1.  May
1467  * return an index whos compare result is > 1 but may only return an index
1468  * whos compare result is <= 1 if it is the first element with that result.
1469  */
1470 int
1471 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1472 {
1473         int b;
1474         int s;
1475         int i;
1476         int r;
1477
1478         /*
1479          * Don't bother if the node does not have very many elements
1480          */
1481         b = 0;
1482         s = node->count;
1483         while (s - b > 4) {
1484                 i = b + (s - b) / 2;
1485                 ++hammer_stats_btree_elements;
1486                 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1487                 if (r <= 1) {
1488                         s = i;
1489                 } else {
1490                         b = i;
1491                 }
1492         }
1493         return(b);
1494 }
1495
1496
1497 /************************************************************************
1498  *                         SPLITTING AND MERGING                        *
1499  ************************************************************************
1500  *
1501  * These routines do all the dirty work required to split and merge nodes.
1502  */
1503
1504 /*
1505  * Split an internal node into two nodes and move the separator at the split
1506  * point to the parent.
1507  *
1508  * (cursor->node, cursor->index) indicates the element the caller intends
1509  * to push into.  We will adjust node and index if that element winds
1510  * up in the split node.
1511  *
1512  * If we are at the root of the filesystem a new root must be created with
1513  * two elements, one pointing to the original root and one pointing to the
1514  * newly allocated split node.
1515  */
1516 static
1517 int
1518 btree_split_internal(hammer_cursor_t cursor)
1519 {
1520         hammer_node_ondisk_t ondisk;
1521         hammer_node_t node;
1522         hammer_node_t parent;
1523         hammer_node_t new_node;
1524         hammer_btree_elm_t elm;
1525         hammer_btree_elm_t parent_elm;
1526         struct hammer_node_lock lockroot;
1527         hammer_mount_t hmp = cursor->trans->hmp;
1528         int parent_index;
1529         int made_root;
1530         int split;
1531         int error;
1532         int i;
1533         const int esize = sizeof(*elm);
1534
1535         hammer_node_lock_init(&lockroot, cursor->node);
1536         error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
1537         if (error)
1538                 goto done;
1539         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1540                 goto done;
1541         ++hammer_stats_btree_splits;
1542
1543         /*
1544          * Calculate the split point.  If the insertion point is at the
1545          * end of the leaf we adjust the split point significantly to the
1546          * right to try to optimize node fill and flag it.  If we hit
1547          * that same leaf again our heuristic failed and we don't try
1548          * to optimize node fill (it could lead to a degenerate case).
1549          */
1550         node = cursor->node;
1551         ondisk = node->ondisk;
1552         KKASSERT(ondisk->count > 4);
1553         if (cursor->index == ondisk->count &&
1554             (node->flags & HAMMER_NODE_NONLINEAR) == 0) {
1555                 split = (ondisk->count + 1) * 3 / 4;
1556                 node->flags |= HAMMER_NODE_NONLINEAR;
1557         } else {
1558                 /*
1559                  * We are splitting but elms[split] will be promoted to
1560                  * the parent, leaving the right hand node with one less
1561                  * element.  If the insertion point will be on the
1562                  * left-hand side adjust the split point to give the
1563                  * right hand side one additional node.
1564                  */
1565                 split = (ondisk->count + 1) / 2;
1566                 if (cursor->index <= split)
1567                         --split;
1568         }
1569
1570         /*
1571          * If we are at the root of the filesystem, create a new root node
1572          * with 1 element and split normally.  Avoid making major
1573          * modifications until we know the whole operation will work.
1574          */
1575         if (ondisk->parent == 0) {
1576                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1577                 if (parent == NULL)
1578                         goto done;
1579                 hammer_lock_ex(&parent->lock);
1580                 hammer_modify_node_noundo(cursor->trans, parent);
1581                 ondisk = parent->ondisk;
1582                 ondisk->count = 1;
1583                 ondisk->parent = 0;
1584                 ondisk->mirror_tid = node->ondisk->mirror_tid;
1585                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1586                 ondisk->elms[0].base = hmp->root_btree_beg;
1587                 ondisk->elms[0].base.btype = node->ondisk->type;
1588                 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1589                 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid;
1590                 ondisk->elms[1].base = hmp->root_btree_end;
1591                 hammer_modify_node_done(parent);
1592                 /* ondisk->elms[1].base.btype - not used */
1593                 made_root = 1;
1594                 parent_index = 0;       /* index of current node in parent */
1595         } else {
1596                 made_root = 0;
1597                 parent = cursor->parent;
1598                 parent_index = cursor->parent_index;
1599         }
1600
1601         /*
1602          * Split node into new_node at the split point.
1603          *
1604          *  B O O O P N N B     <-- P = node->elms[split] (index 4)
1605          *   0 1 2 3 4 5 6      <-- subtree indices
1606          *
1607          *       x x P x x
1608          *        s S S s
1609          *         /   \
1610          *  B O O O B    B N N B        <--- inner boundary points are 'P'
1611          *   0 1 2 3      4 5 6
1612          */
1613         new_node = hammer_alloc_btree(cursor->trans, 0, &error);
1614         if (new_node == NULL) {
1615                 if (made_root) {
1616                         hammer_unlock(&parent->lock);
1617                         hammer_delete_node(cursor->trans, parent);
1618                         hammer_rel_node(parent);
1619                 }
1620                 goto done;
1621         }
1622         hammer_lock_ex(&new_node->lock);
1623
1624         /*
1625          * Create the new node.  P becomes the left-hand boundary in the
1626          * new node.  Copy the right-hand boundary as well.
1627          *
1628          * elm is the new separator.
1629          */
1630         hammer_modify_node_noundo(cursor->trans, new_node);
1631         hammer_modify_node_all(cursor->trans, node);
1632         ondisk = node->ondisk;
1633         elm = &ondisk->elms[split];
1634         bcopy(elm, &new_node->ondisk->elms[0],
1635               (ondisk->count - split + 1) * esize);
1636         new_node->ondisk->count = ondisk->count - split;
1637         new_node->ondisk->parent = parent->node_offset;
1638         new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1639         new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1640         KKASSERT(ondisk->type == new_node->ondisk->type);
1641         hammer_cursor_split_node(node, new_node, split);
1642
1643         /*
1644          * Cleanup the original node.  Elm (P) becomes the new boundary,
1645          * its subtree_offset was moved to the new node.  If we had created
1646          * a new root its parent pointer may have changed.
1647          */
1648         elm->internal.subtree_offset = 0;
1649         ondisk->count = split;
1650
1651         /*
1652          * Insert the separator into the parent, fixup the parent's
1653          * reference to the original node, and reference the new node.
1654          * The separator is P.
1655          *
1656          * Remember that base.count does not include the right-hand boundary.
1657          */
1658         hammer_modify_node_all(cursor->trans, parent);
1659         ondisk = parent->ondisk;
1660         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1661         parent_elm = &ondisk->elms[parent_index+1];
1662         bcopy(parent_elm, parent_elm + 1,
1663               (ondisk->count - parent_index) * esize);
1664         parent_elm->internal.base = elm->base;  /* separator P */
1665         parent_elm->internal.base.btype = new_node->ondisk->type;
1666         parent_elm->internal.subtree_offset = new_node->node_offset;
1667         parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1668         ++ondisk->count;
1669         hammer_modify_node_done(parent);
1670         hammer_cursor_inserted_element(parent, parent_index + 1);
1671
1672         /*
1673          * The children of new_node need their parent pointer set to new_node.
1674          * The children have already been locked by
1675          * hammer_btree_lock_children().
1676          */
1677         for (i = 0; i < new_node->ondisk->count; ++i) {
1678                 elm = &new_node->ondisk->elms[i];
1679                 error = btree_set_parent(cursor->trans, new_node, elm);
1680                 if (error) {
1681                         panic("btree_split_internal: btree-fixup problem");
1682                 }
1683         }
1684         hammer_modify_node_done(new_node);
1685
1686         /*
1687          * The filesystem's root B-Tree pointer may have to be updated.
1688          */
1689         if (made_root) {
1690                 hammer_volume_t volume;
1691
1692                 volume = hammer_get_root_volume(hmp, &error);
1693                 KKASSERT(error == 0);
1694
1695                 hammer_modify_volume_field(cursor->trans, volume,
1696                                            vol0_btree_root);
1697                 volume->ondisk->vol0_btree_root = parent->node_offset;
1698                 hammer_modify_volume_done(volume);
1699                 node->ondisk->parent = parent->node_offset;
1700                 if (cursor->parent) {
1701                         hammer_unlock(&cursor->parent->lock);
1702                         hammer_rel_node(cursor->parent);
1703                 }
1704                 cursor->parent = parent;        /* lock'd and ref'd */
1705                 hammer_rel_volume(volume, 0);
1706         }
1707         hammer_modify_node_done(node);
1708
1709         /*
1710          * Ok, now adjust the cursor depending on which element the original
1711          * index was pointing at.  If we are >= the split point the push node
1712          * is now in the new node.
1713          *
1714          * NOTE: If we are at the split point itself we cannot stay with the
1715          * original node because the push index will point at the right-hand
1716          * boundary, which is illegal.
1717          *
1718          * NOTE: The cursor's parent or parent_index must be adjusted for
1719          * the case where a new parent (new root) was created, and the case
1720          * where the cursor is now pointing at the split node.
1721          */
1722         if (cursor->index >= split) {
1723                 cursor->parent_index = parent_index + 1;
1724                 cursor->index -= split;
1725                 hammer_unlock(&cursor->node->lock);
1726                 hammer_rel_node(cursor->node);
1727                 cursor->node = new_node;        /* locked and ref'd */
1728         } else {
1729                 cursor->parent_index = parent_index;
1730                 hammer_unlock(&new_node->lock);
1731                 hammer_rel_node(new_node);
1732         }
1733
1734         /*
1735          * Fixup left and right bounds
1736          */
1737         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1738         cursor->left_bound = &parent_elm[0].internal.base;
1739         cursor->right_bound = &parent_elm[1].internal.base;
1740         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1741                  &cursor->node->ondisk->elms[0].internal.base) <= 0);
1742         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1743                  &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1744
1745 done:
1746         hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
1747         hammer_cursor_downgrade(cursor);
1748         return (error);
1749 }
1750
1751 /*
1752  * Same as the above, but splits a full leaf node.
1753  */
1754 static
1755 int
1756 btree_split_leaf(hammer_cursor_t cursor)
1757 {
1758         hammer_node_ondisk_t ondisk;
1759         hammer_node_t parent;
1760         hammer_node_t leaf;
1761         hammer_mount_t hmp;
1762         hammer_node_t new_leaf;
1763         hammer_btree_elm_t elm;
1764         hammer_btree_elm_t parent_elm;
1765         hammer_base_elm_t mid_boundary;
1766         int parent_index;
1767         int made_root;
1768         int split;
1769         int error;
1770         const size_t esize = sizeof(*elm);
1771
1772         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1773                 return(error);
1774         ++hammer_stats_btree_splits;
1775
1776         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1777                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1778         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1779                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1780
1781         /*
1782          * Calculate the split point.  If the insertion point is at the
1783          * end of the leaf we adjust the split point significantly to the
1784          * right to try to optimize node fill and flag it.  If we hit
1785          * that same leaf again our heuristic failed and we don't try
1786          * to optimize node fill (it could lead to a degenerate case).
1787          */
1788         leaf = cursor->node;
1789         ondisk = leaf->ondisk;
1790         KKASSERT(ondisk->count > 4);
1791         if (cursor->index == ondisk->count &&
1792             (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) {
1793                 split = (ondisk->count + 1) * 3 / 4;
1794                 leaf->flags |= HAMMER_NODE_NONLINEAR;
1795         } else {
1796                 split = (ondisk->count + 1) / 2;
1797         }
1798
1799 #if 0
1800         /*
1801          * If the insertion point is at the split point shift the
1802          * split point left so we don't have to worry about
1803          */
1804         if (cursor->index == split)
1805                 --split;
1806 #endif
1807         KKASSERT(split > 0 && split < ondisk->count);
1808
1809         error = 0;
1810         hmp = leaf->hmp;
1811
1812         elm = &ondisk->elms[split];
1813
1814         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1815         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1816         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1817         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1818
1819         /*
1820          * If we are at the root of the tree, create a new root node with
1821          * 1 element and split normally.  Avoid making major modifications
1822          * until we know the whole operation will work.
1823          */
1824         if (ondisk->parent == 0) {
1825                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1826                 if (parent == NULL)
1827                         goto done;
1828                 hammer_lock_ex(&parent->lock);
1829                 hammer_modify_node_noundo(cursor->trans, parent);
1830                 ondisk = parent->ondisk;
1831                 ondisk->count = 1;
1832                 ondisk->parent = 0;
1833                 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1834                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1835                 ondisk->elms[0].base = hmp->root_btree_beg;
1836                 ondisk->elms[0].base.btype = leaf->ondisk->type;
1837                 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1838                 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid;
1839                 ondisk->elms[1].base = hmp->root_btree_end;
1840                 /* ondisk->elms[1].base.btype = not used */
1841                 hammer_modify_node_done(parent);
1842                 made_root = 1;
1843                 parent_index = 0;       /* insertion point in parent */
1844         } else {
1845                 made_root = 0;
1846                 parent = cursor->parent;
1847                 parent_index = cursor->parent_index;
1848         }
1849
1850         /*
1851          * Split leaf into new_leaf at the split point.  Select a separator
1852          * value in-between the two leafs but with a bent towards the right
1853          * leaf since comparisons use an 'elm >= separator' inequality.
1854          *
1855          *  L L L L L L L L
1856          *
1857          *       x x P x x
1858          *        s S S s
1859          *         /   \
1860          *  L L L L     L L L L
1861          */
1862         new_leaf = hammer_alloc_btree(cursor->trans, 0, &error);
1863         if (new_leaf == NULL) {
1864                 if (made_root) {
1865                         hammer_unlock(&parent->lock);
1866                         hammer_delete_node(cursor->trans, parent);
1867                         hammer_rel_node(parent);
1868                 }
1869                 goto done;
1870         }
1871         hammer_lock_ex(&new_leaf->lock);
1872
1873         /*
1874          * Create the new node and copy the leaf elements from the split
1875          * point on to the new node.
1876          */
1877         hammer_modify_node_all(cursor->trans, leaf);
1878         hammer_modify_node_noundo(cursor->trans, new_leaf);
1879         ondisk = leaf->ondisk;
1880         elm = &ondisk->elms[split];
1881         bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1882         new_leaf->ondisk->count = ondisk->count - split;
1883         new_leaf->ondisk->parent = parent->node_offset;
1884         new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1885         new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1886         KKASSERT(ondisk->type == new_leaf->ondisk->type);
1887         hammer_modify_node_done(new_leaf);
1888         hammer_cursor_split_node(leaf, new_leaf, split);
1889
1890         /*
1891          * Cleanup the original node.  Because this is a leaf node and
1892          * leaf nodes do not have a right-hand boundary, there
1893          * aren't any special edge cases to clean up.  We just fixup the
1894          * count.
1895          */
1896         ondisk->count = split;
1897
1898         /*
1899          * Insert the separator into the parent, fixup the parent's
1900          * reference to the original node, and reference the new node.
1901          * The separator is P.
1902          *
1903          * Remember that base.count does not include the right-hand boundary.
1904          * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1905          */
1906         hammer_modify_node_all(cursor->trans, parent);
1907         ondisk = parent->ondisk;
1908         KKASSERT(split != 0);
1909         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1910         parent_elm = &ondisk->elms[parent_index+1];
1911         bcopy(parent_elm, parent_elm + 1,
1912               (ondisk->count - parent_index) * esize);
1913
1914         hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1915         parent_elm->internal.base.btype = new_leaf->ondisk->type;
1916         parent_elm->internal.subtree_offset = new_leaf->node_offset;
1917         parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1918         mid_boundary = &parent_elm->base;
1919         ++ondisk->count;
1920         hammer_modify_node_done(parent);
1921         hammer_cursor_inserted_element(parent, parent_index + 1);
1922
1923         /*
1924          * The filesystem's root B-Tree pointer may have to be updated.
1925          */
1926         if (made_root) {
1927                 hammer_volume_t volume;
1928
1929                 volume = hammer_get_root_volume(hmp, &error);
1930                 KKASSERT(error == 0);
1931
1932                 hammer_modify_volume_field(cursor->trans, volume,
1933                                            vol0_btree_root);
1934                 volume->ondisk->vol0_btree_root = parent->node_offset;
1935                 hammer_modify_volume_done(volume);
1936                 leaf->ondisk->parent = parent->node_offset;
1937                 if (cursor->parent) {
1938                         hammer_unlock(&cursor->parent->lock);
1939                         hammer_rel_node(cursor->parent);
1940                 }
1941                 cursor->parent = parent;        /* lock'd and ref'd */
1942                 hammer_rel_volume(volume, 0);
1943         }
1944         hammer_modify_node_done(leaf);
1945
1946         /*
1947          * Ok, now adjust the cursor depending on which element the original
1948          * index was pointing at.  If we are >= the split point the push node
1949          * is now in the new node.
1950          *
1951          * NOTE: If we are at the split point itself we need to select the
1952          * old or new node based on where key_beg's insertion point will be.
1953          * If we pick the wrong side the inserted element will wind up in
1954          * the wrong leaf node and outside that node's bounds.
1955          */
1956         if (cursor->index > split ||
1957             (cursor->index == split &&
1958              hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1959                 cursor->parent_index = parent_index + 1;
1960                 cursor->index -= split;
1961                 hammer_unlock(&cursor->node->lock);
1962                 hammer_rel_node(cursor->node);
1963                 cursor->node = new_leaf;
1964         } else {
1965                 cursor->parent_index = parent_index;
1966                 hammer_unlock(&new_leaf->lock);
1967                 hammer_rel_node(new_leaf);
1968         }
1969
1970         /*
1971          * Fixup left and right bounds
1972          */
1973         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1974         cursor->left_bound = &parent_elm[0].internal.base;
1975         cursor->right_bound = &parent_elm[1].internal.base;
1976
1977         /*
1978          * Assert that the bounds are correct.
1979          */
1980         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1981                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1982         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1983                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1984         KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1985         KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1986
1987 done:
1988         hammer_cursor_downgrade(cursor);
1989         return (error);
1990 }
1991
1992 #if 0
1993
1994 /*
1995  * Recursively correct the right-hand boundary's create_tid to (tid) as
1996  * long as the rest of the key matches.  We have to recurse upward in
1997  * the tree as well as down the left side of each parent's right node.
1998  *
1999  * Return EDEADLK if we were only partially successful, forcing the caller
2000  * to try again.  The original cursor is not modified.  This routine can
2001  * also fail with EDEADLK if it is forced to throw away a portion of its
2002  * record history.
2003  *
2004  * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2005  */
2006 struct hammer_rhb {
2007         TAILQ_ENTRY(hammer_rhb) entry;
2008         hammer_node_t   node;
2009         int             index;
2010 };
2011
2012 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2013
2014 int
2015 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2016 {
2017         struct hammer_mount *hmp;
2018         struct hammer_rhb_list rhb_list;
2019         hammer_base_elm_t elm;
2020         hammer_node_t orig_node;
2021         struct hammer_rhb *rhb;
2022         int orig_index;
2023         int error;
2024
2025         TAILQ_INIT(&rhb_list);
2026         hmp = cursor->trans->hmp;
2027
2028         /*
2029          * Save our position so we can restore it on return.  This also
2030          * gives us a stable 'elm'.
2031          */
2032         orig_node = cursor->node;
2033         hammer_ref_node(orig_node);
2034         hammer_lock_sh(&orig_node->lock);
2035         orig_index = cursor->index;
2036         elm = &orig_node->ondisk->elms[orig_index].base;
2037
2038         /*
2039          * Now build a list of parents going up, allocating a rhb
2040          * structure for each one.
2041          */
2042         while (cursor->parent) {
2043                 /*
2044                  * Stop if we no longer have any right-bounds to fix up
2045                  */
2046                 if (elm->obj_id != cursor->right_bound->obj_id ||
2047                     elm->rec_type != cursor->right_bound->rec_type ||
2048                     elm->key != cursor->right_bound->key) {
2049                         break;
2050                 }
2051
2052                 /*
2053                  * Stop if the right-hand bound's create_tid does not
2054                  * need to be corrected.
2055                  */
2056                 if (cursor->right_bound->create_tid >= tid)
2057                         break;
2058
2059                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2060                 rhb->node = cursor->parent;
2061                 rhb->index = cursor->parent_index;
2062                 hammer_ref_node(rhb->node);
2063                 hammer_lock_sh(&rhb->node->lock);
2064                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2065
2066                 hammer_cursor_up(cursor);
2067         }
2068
2069         /*
2070          * now safely adjust the right hand bound for each rhb.  This may
2071          * also require taking the right side of the tree and iterating down
2072          * ITS left side.
2073          */
2074         error = 0;
2075         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2076                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2077                 if (error)
2078                         break;
2079                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2080                 hammer_unlock(&rhb->node->lock);
2081                 hammer_rel_node(rhb->node);
2082                 kfree(rhb, hmp->m_misc);
2083
2084                 switch (cursor->node->ondisk->type) {
2085                 case HAMMER_BTREE_TYPE_INTERNAL:
2086                         /*
2087                          * Right-boundary for parent at internal node
2088                          * is one element to the right of the element whos
2089                          * right boundary needs adjusting.  We must then
2090                          * traverse down the left side correcting any left
2091                          * bounds (which may now be too far to the left).
2092                          */
2093                         ++cursor->index;
2094                         error = hammer_btree_correct_lhb(cursor, tid);
2095                         break;
2096                 default:
2097                         panic("hammer_btree_correct_rhb(): Bad node type");
2098                         error = EINVAL;
2099                         break;
2100                 }
2101         }
2102
2103         /*
2104          * Cleanup
2105          */
2106         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2107                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2108                 hammer_unlock(&rhb->node->lock);
2109                 hammer_rel_node(rhb->node);
2110                 kfree(rhb, hmp->m_misc);
2111         }
2112         error = hammer_cursor_seek(cursor, orig_node, orig_index);
2113         hammer_unlock(&orig_node->lock);
2114         hammer_rel_node(orig_node);
2115         return (error);
2116 }
2117
2118 /*
2119  * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2120  * bound going downward starting at the current cursor position.
2121  *
2122  * This function does not restore the cursor after use.
2123  */
2124 int
2125 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2126 {
2127         struct hammer_rhb_list rhb_list;
2128         hammer_base_elm_t elm;
2129         hammer_base_elm_t cmp;
2130         struct hammer_rhb *rhb;
2131         struct hammer_mount *hmp;
2132         int error;
2133
2134         TAILQ_INIT(&rhb_list);
2135         hmp = cursor->trans->hmp;
2136
2137         cmp = &cursor->node->ondisk->elms[cursor->index].base;
2138
2139         /*
2140          * Record the node and traverse down the left-hand side for all
2141          * matching records needing a boundary correction.
2142          */
2143         error = 0;
2144         for (;;) {
2145                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2146                 rhb->node = cursor->node;
2147                 rhb->index = cursor->index;
2148                 hammer_ref_node(rhb->node);
2149                 hammer_lock_sh(&rhb->node->lock);
2150                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2151
2152                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2153                         /*
2154                          * Nothing to traverse down if we are at the right
2155                          * boundary of an internal node.
2156                          */
2157                         if (cursor->index == cursor->node->ondisk->count)
2158                                 break;
2159                 } else {
2160                         elm = &cursor->node->ondisk->elms[cursor->index].base;
2161                         if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2162                                 break;
2163                         panic("Illegal leaf record type %02x", elm->btype);
2164                 }
2165                 error = hammer_cursor_down(cursor);
2166                 if (error)
2167                         break;
2168
2169                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2170                 if (elm->obj_id != cmp->obj_id ||
2171                     elm->rec_type != cmp->rec_type ||
2172                     elm->key != cmp->key) {
2173                         break;
2174                 }
2175                 if (elm->create_tid >= tid)
2176                         break;
2177
2178         }
2179
2180         /*
2181          * Now we can safely adjust the left-hand boundary from the bottom-up.
2182          * The last element we remove from the list is the caller's right hand
2183          * boundary, which must also be adjusted.
2184          */
2185         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2186                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2187                 if (error)
2188                         break;
2189                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2190                 hammer_unlock(&rhb->node->lock);
2191                 hammer_rel_node(rhb->node);
2192                 kfree(rhb, hmp->m_misc);
2193
2194                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2195                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2196                         hammer_modify_node(cursor->trans, cursor->node,
2197                                            &elm->create_tid,
2198                                            sizeof(elm->create_tid));
2199                         elm->create_tid = tid;
2200                         hammer_modify_node_done(cursor->node);
2201                 } else {
2202                         panic("hammer_btree_correct_lhb(): Bad element type");
2203                 }
2204         }
2205
2206         /*
2207          * Cleanup
2208          */
2209         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2210                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2211                 hammer_unlock(&rhb->node->lock);
2212                 hammer_rel_node(rhb->node);
2213                 kfree(rhb, hmp->m_misc);
2214         }
2215         return (error);
2216 }
2217
2218 #endif
2219
2220 /*
2221  * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2222  * (cursor->node).  Returns 0 on success, EDEADLK if we could not complete
2223  * the operation due to a deadlock, or some other error.
2224  *
2225  * This routine is initially called with an empty leaf and may be
2226  * recursively called with single-element internal nodes.
2227  *
2228  * It should also be noted that when removing empty leaves we must be sure
2229  * to test and update mirror_tid because another thread may have deadlocked
2230  * against us (or someone) trying to propagate it up and cannot retry once
2231  * the node has been deleted.
2232  *
2233  * On return the cursor may end up pointing to an internal node, suitable
2234  * for further iteration but not for an immediate insertion or deletion.
2235  */
2236 static int
2237 btree_remove(hammer_cursor_t cursor)
2238 {
2239         hammer_node_ondisk_t ondisk;
2240         hammer_btree_elm_t elm;
2241         hammer_node_t node;
2242         hammer_node_t parent;
2243         const int esize = sizeof(*elm);
2244         int error;
2245
2246         node = cursor->node;
2247
2248         /*
2249          * When deleting the root of the filesystem convert it to
2250          * an empty leaf node.  Internal nodes cannot be empty.
2251          */
2252         ondisk = node->ondisk;
2253         if (ondisk->parent == 0) {
2254                 KKASSERT(cursor->parent == NULL);
2255                 hammer_modify_node_all(cursor->trans, node);
2256                 KKASSERT(ondisk == node->ondisk);
2257                 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2258                 ondisk->count = 0;
2259                 hammer_modify_node_done(node);
2260                 cursor->index = 0;
2261                 return(0);
2262         }
2263
2264         parent = cursor->parent;
2265
2266         /*
2267          * Attempt to remove the parent's reference to the child.  If the
2268          * parent would become empty we have to recurse.  If we fail we
2269          * leave the parent pointing to an empty leaf node.
2270          *
2271          * We have to recurse successfully before we can delete the internal
2272          * node as it is illegal to have empty internal nodes.  Even though
2273          * the operation may be aborted we must still fixup any unlocked
2274          * cursors as if we had deleted the element prior to recursing
2275          * (by calling hammer_cursor_deleted_element()) so those cursors
2276          * are properly forced up the chain by the recursion.
2277          */
2278         if (parent->ondisk->count == 1) {
2279                 /*
2280                  * This special cursor_up_locked() call leaves the original
2281                  * node exclusively locked and referenced, leaves the
2282                  * original parent locked (as the new node), and locks the
2283                  * new parent.  It can return EDEADLK.
2284                  *
2285                  * We cannot call hammer_cursor_removed_node() until we are
2286                  * actually able to remove the node.  If we did then tracked
2287                  * cursors in the middle of iterations could be repointed
2288                  * to a parent node.  If this occurs they could end up
2289                  * scanning newly inserted records into the node (that could
2290                  * not be deleted) when they push down again.
2291                  *
2292                  * Due to the way the recursion works the final parent is left
2293                  * in cursor->parent after the recursion returns.  Each
2294                  * layer on the way back up is thus able to call
2295                  * hammer_cursor_removed_node() and 'jump' the node up to
2296                  * the (same) final parent.
2297                  *
2298                  * NOTE!  The local variable 'parent' is invalid after we
2299                  *        call hammer_cursor_up_locked().
2300                  */
2301                 error = hammer_cursor_up_locked(cursor);
2302                 parent = NULL;
2303
2304                 if (error == 0) {
2305                         hammer_cursor_deleted_element(cursor->node, 0);
2306                         error = btree_remove(cursor);
2307                         if (error == 0) {
2308                                 KKASSERT(node != cursor->node);
2309                                 hammer_cursor_removed_node(
2310                                         node, cursor->node,
2311                                         cursor->index);
2312                                 hammer_modify_node_all(cursor->trans, node);
2313                                 ondisk = node->ondisk;
2314                                 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2315                                 ondisk->count = 0;
2316                                 hammer_modify_node_done(node);
2317                                 hammer_flush_node(node, 0);
2318                                 hammer_delete_node(cursor->trans, node);
2319                         } else {
2320                                 /*
2321                                  * Defer parent removal because we could not
2322                                  * get the lock, just let the leaf remain
2323                                  * empty.
2324                                  */
2325                                 /**/
2326                         }
2327                         hammer_unlock(&node->lock);
2328                         hammer_rel_node(node);
2329                 } else {
2330                         /*
2331                          * Defer parent removal because we could not
2332                          * get the lock, just let the leaf remain
2333                          * empty.
2334                          */
2335                         /**/
2336                 }
2337         } else {
2338                 KKASSERT(parent->ondisk->count > 1);
2339
2340                 hammer_modify_node_all(cursor->trans, parent);
2341                 ondisk = parent->ondisk;
2342                 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2343
2344                 elm = &ondisk->elms[cursor->parent_index];
2345                 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2346                 KKASSERT(ondisk->count > 0);
2347
2348                 /*
2349                  * We must retain the highest mirror_tid.  The deleted
2350                  * range is now encompassed by the element to the left.
2351                  * If we are already at the left edge the new left edge
2352                  * inherits mirror_tid.
2353                  *
2354                  * Note that bounds of the parent to our parent may create
2355                  * a gap to the left of our left-most node or to the right
2356                  * of our right-most node.  The gap is silently included
2357                  * in the mirror_tid's area of effect from the point of view
2358                  * of the scan.
2359                  */
2360                 if (cursor->parent_index) {
2361                         if (elm[-1].internal.mirror_tid <
2362                             elm[0].internal.mirror_tid) {
2363                                 elm[-1].internal.mirror_tid =
2364                                     elm[0].internal.mirror_tid;
2365                         }
2366                 } else {
2367                         if (elm[1].internal.mirror_tid <
2368                             elm[0].internal.mirror_tid) {
2369                                 elm[1].internal.mirror_tid =
2370                                     elm[0].internal.mirror_tid;
2371                         }
2372                 }
2373
2374                 /*
2375                  * Delete the subtree reference in the parent.  Include
2376                  * boundary element at end.
2377                  */
2378                 bcopy(&elm[1], &elm[0],
2379                       (ondisk->count - cursor->parent_index) * esize);
2380                 --ondisk->count;
2381                 hammer_modify_node_done(parent);
2382                 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2383                 hammer_cursor_deleted_element(parent, cursor->parent_index);
2384                 hammer_flush_node(node, 0);
2385                 hammer_delete_node(cursor->trans, node);
2386
2387                 /*
2388                  * cursor->node is invalid, cursor up to make the cursor
2389                  * valid again.  We have to flag the condition in case
2390                  * another thread wiggles an insertion in during an
2391                  * iteration.
2392                  */
2393                 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK;
2394                 error = hammer_cursor_up(cursor);
2395         }
2396         return (error);
2397 }
2398
2399 /*
2400  * Propagate cursor->trans->tid up the B-Tree starting at the current
2401  * cursor position using pseudofs info gleaned from the passed inode.
2402  *
2403  * The passed inode has no relationship to the cursor position other
2404  * then being in the same pseudofs as the insertion or deletion we
2405  * are propagating the mirror_tid for.
2406  *
2407  * WARNING!  Because we push and pop the passed cursor, it may be
2408  *           modified by other B-Tree operations while it is unlocked
2409  *           and things like the node & leaf pointers, and indexes might
2410  *           change.
2411  */
2412 void
2413 hammer_btree_do_propagation(hammer_cursor_t cursor,
2414                             hammer_pseudofs_inmem_t pfsm,
2415                             hammer_btree_leaf_elm_t leaf)
2416 {
2417         hammer_cursor_t ncursor;
2418         hammer_tid_t mirror_tid;
2419         int error __debugvar;
2420
2421         /*
2422          * We do not propagate a mirror_tid if the filesystem was mounted
2423          * in no-mirror mode.
2424          */
2425         if (cursor->trans->hmp->master_id < 0)
2426                 return;
2427
2428         /*
2429          * This is a bit of a hack because we cannot deadlock or return
2430          * EDEADLK here.  The related operation has already completed and
2431          * we must propagate the mirror_tid now regardless.
2432          *
2433          * Generate a new cursor which inherits the original's locks and
2434          * unlock the original.  Use the new cursor to propagate the
2435          * mirror_tid.  Then clean up the new cursor and reacquire locks
2436          * on the original.
2437          *
2438          * hammer_dup_cursor() cannot dup locks.  The dup inherits the
2439          * original's locks and the original is tracked and must be
2440          * re-locked.
2441          */
2442         mirror_tid = cursor->node->ondisk->mirror_tid;
2443         KKASSERT(mirror_tid != 0);
2444         ncursor = hammer_push_cursor(cursor);
2445         error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2446         KKASSERT(error == 0);
2447         hammer_pop_cursor(cursor, ncursor);
2448         /* WARNING: cursor's leaf pointer may change after pop */
2449 }
2450
2451
2452 /*
2453  * Propagate a mirror TID update upwards through the B-Tree to the root.
2454  *
2455  * A locked internal node must be passed in.  The node will remain locked
2456  * on return.
2457  *
2458  * This function syncs mirror_tid at the specified internal node's element,
2459  * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2460  */
2461 static int
2462 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2463 {
2464         hammer_btree_internal_elm_t elm;
2465         hammer_node_t node;
2466         int error;
2467
2468         for (;;) {
2469                 error = hammer_cursor_up(cursor);
2470                 if (error == 0)
2471                         error = hammer_cursor_upgrade(cursor);
2472
2473                 /*
2474                  * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the
2475                  * cursor will still be properly positioned for
2476                  * mirror propagation, just not for iterations.
2477                  */
2478                 while (error == EDEADLK) {
2479                         hammer_recover_cursor(cursor);
2480                         error = hammer_cursor_upgrade(cursor);
2481                 }
2482                 if (error)
2483                         break;
2484
2485                 /*
2486                  * If the cursor deadlocked it could end up at a leaf
2487                  * after we lost the lock.
2488                  */
2489                 node = cursor->node;
2490                 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL)
2491                         continue;
2492
2493                 /*
2494                  * Adjust the node's element
2495                  */
2496                 elm = &node->ondisk->elms[cursor->index].internal;
2497                 if (elm->mirror_tid >= mirror_tid)
2498                         break;
2499                 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2500                                    sizeof(elm->mirror_tid));
2501                 elm->mirror_tid = mirror_tid;
2502                 hammer_modify_node_done(node);
2503                 if (hammer_debug_general & 0x0002) {
2504                         kprintf("mirror_propagate: propagate "
2505                                 "%016llx @%016llx:%d\n",
2506                                 (long long)mirror_tid,
2507                                 (long long)node->node_offset,
2508                                 cursor->index);
2509                 }
2510
2511
2512                 /*
2513                  * Adjust the node's mirror_tid aggregator
2514                  */
2515                 if (node->ondisk->mirror_tid >= mirror_tid)
2516                         return(0);
2517                 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2518                 node->ondisk->mirror_tid = mirror_tid;
2519                 hammer_modify_node_done(node);
2520                 if (hammer_debug_general & 0x0002) {
2521                         kprintf("mirror_propagate: propagate "
2522                                 "%016llx @%016llx\n",
2523                                 (long long)mirror_tid,
2524                                 (long long)node->node_offset);
2525                 }
2526         }
2527         if (error == ENOENT)
2528                 error = 0;
2529         return(error);
2530 }
2531
2532 hammer_node_t
2533 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2534                         int *parent_indexp, int *errorp, int try_exclusive)
2535 {
2536         hammer_node_t parent;
2537         hammer_btree_elm_t elm;
2538         int i;
2539
2540         /*
2541          * Get the node
2542          */
2543         parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2544         if (*errorp) {
2545                 KKASSERT(parent == NULL);
2546                 return(NULL);
2547         }
2548         KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2549
2550         /*
2551          * Lock the node
2552          */
2553         if (try_exclusive) {
2554                 if (hammer_lock_ex_try(&parent->lock)) {
2555                         hammer_rel_node(parent);
2556                         *errorp = EDEADLK;
2557                         return(NULL);
2558                 }
2559         } else {
2560                 hammer_lock_sh(&parent->lock);
2561         }
2562
2563         /*
2564          * Figure out which element in the parent is pointing to the
2565          * child.
2566          */
2567         if (node->ondisk->count) {
2568                 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2569                                              parent->ondisk);
2570         } else {
2571                 i = 0;
2572         }
2573         while (i < parent->ondisk->count) {
2574                 elm = &parent->ondisk->elms[i];
2575                 if (elm->internal.subtree_offset == node->node_offset)
2576                         break;
2577                 ++i;
2578         }
2579         if (i == parent->ondisk->count) {
2580                 hammer_unlock(&parent->lock);
2581                 panic("Bad B-Tree link: parent %p node %p", parent, node);
2582         }
2583         *parent_indexp = i;
2584         KKASSERT(*errorp == 0);
2585         return(parent);
2586 }
2587
2588 /*
2589  * The element (elm) has been moved to a new internal node (node).
2590  *
2591  * If the element represents a pointer to an internal node that node's
2592  * parent must be adjusted to the element's new location.
2593  *
2594  * XXX deadlock potential here with our exclusive locks
2595  */
2596 int
2597 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2598                  hammer_btree_elm_t elm)
2599 {
2600         hammer_node_t child;
2601         int error;
2602
2603         error = 0;
2604
2605         switch(elm->base.btype) {
2606         case HAMMER_BTREE_TYPE_INTERNAL:
2607         case HAMMER_BTREE_TYPE_LEAF:
2608                 child = hammer_get_node(trans, elm->internal.subtree_offset,
2609                                         0, &error);
2610                 if (error == 0) {
2611                         hammer_modify_node_field(trans, child, parent);
2612                         child->ondisk->parent = node->node_offset;
2613                         hammer_modify_node_done(child);
2614                         hammer_rel_node(child);
2615                 }
2616                 break;
2617         default:
2618                 break;
2619         }
2620         return(error);
2621 }
2622
2623 /*
2624  * Initialize the root of a recursive B-Tree node lock list structure.
2625  */
2626 void
2627 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2628 {
2629         TAILQ_INIT(&parent->list);
2630         parent->parent = NULL;
2631         parent->node = node;
2632         parent->index = -1;
2633         parent->count = node->ondisk->count;
2634         parent->copy = NULL;
2635         parent->flags = 0;
2636 }
2637
2638 /*
2639  * Initialize a cache of hammer_node_lock's including space allocated
2640  * for node copies.
2641  *
2642  * This is used by the rebalancing code to preallocate the copy space
2643  * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER
2644  * locks, otherwise we can blow out the pageout daemon's emergency
2645  * reserve and deadlock it.
2646  *
2647  * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache.
2648  *       The flag is set when the item is pulled off the cache for use.
2649  */
2650 void
2651 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
2652                          int depth)
2653 {
2654         hammer_node_lock_t item;
2655         int count;
2656
2657         for (count = 1; depth; --depth)
2658                 count *= HAMMER_BTREE_LEAF_ELMS;
2659         bzero(lcache, sizeof(*lcache));
2660         TAILQ_INIT(&lcache->list);
2661         while (count) {
2662                 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO);
2663                 item->copy = kmalloc(sizeof(*item->copy),
2664                                      hmp->m_misc, M_WAITOK);
2665                 TAILQ_INIT(&item->list);
2666                 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2667                 --count;
2668         }
2669 }
2670
2671 void
2672 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache)
2673 {
2674         hammer_node_lock_t item;
2675
2676         while ((item = TAILQ_FIRST(&lcache->list)) != NULL) {
2677                 TAILQ_REMOVE(&lcache->list, item, entry);
2678                 KKASSERT(item->copy);
2679                 KKASSERT(TAILQ_EMPTY(&item->list));
2680                 kfree(item->copy, hmp->m_misc);
2681                 kfree(item, hmp->m_misc);
2682         }
2683         KKASSERT(lcache->copy == NULL);
2684 }
2685
2686 /*
2687  * Exclusively lock all the children of node.  This is used by the split
2688  * code to prevent anyone from accessing the children of a cursor node
2689  * while we fix-up its parent offset.
2690  *
2691  * If we don't lock the children we can really mess up cursors which block
2692  * trying to cursor-up into our node.
2693  *
2694  * On failure EDEADLK (or some other error) is returned.  If a deadlock
2695  * error is returned the cursor is adjusted to block on termination.
2696  *
2697  * The caller is responsible for managing parent->node, the root's node
2698  * is usually aliased from a cursor.
2699  */
2700 int
2701 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2702                            hammer_node_lock_t parent,
2703                            hammer_node_lock_t lcache)
2704 {
2705         hammer_node_t node;
2706         hammer_node_lock_t item;
2707         hammer_node_ondisk_t ondisk;
2708         hammer_btree_elm_t elm;
2709         hammer_node_t child;
2710         struct hammer_mount *hmp;
2711         int error;
2712         int i;
2713
2714         node = parent->node;
2715         ondisk = node->ondisk;
2716         error = 0;
2717         hmp = cursor->trans->hmp;
2718
2719         /*
2720          * We really do not want to block on I/O with exclusive locks held,
2721          * pre-get the children before trying to lock the mess.  This is
2722          * only done one-level deep for now.
2723          */
2724         for (i = 0; i < ondisk->count; ++i) {
2725                 ++hammer_stats_btree_elements;
2726                 elm = &ondisk->elms[i];
2727                 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2728                     elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2729                         continue;
2730                 }
2731                 child = hammer_get_node(cursor->trans,
2732                                         elm->internal.subtree_offset,
2733                                         0, &error);
2734                 if (child)
2735                         hammer_rel_node(child);
2736         }
2737
2738         /*
2739          * Do it for real
2740          */
2741         for (i = 0; error == 0 && i < ondisk->count; ++i) {
2742                 ++hammer_stats_btree_elements;
2743                 elm = &ondisk->elms[i];
2744
2745                 switch(elm->base.btype) {
2746                 case HAMMER_BTREE_TYPE_INTERNAL:
2747                 case HAMMER_BTREE_TYPE_LEAF:
2748                         KKASSERT(elm->internal.subtree_offset != 0);
2749                         child = hammer_get_node(cursor->trans,
2750                                                 elm->internal.subtree_offset,
2751                                                 0, &error);
2752                         break;
2753                 default:
2754                         child = NULL;
2755                         break;
2756                 }
2757                 if (child) {
2758                         if (hammer_lock_ex_try(&child->lock) != 0) {
2759                                 if (cursor->deadlk_node == NULL) {
2760                                         cursor->deadlk_node = child;
2761                                         hammer_ref_node(cursor->deadlk_node);
2762                                 }
2763                                 error = EDEADLK;
2764                                 hammer_rel_node(child);
2765                         } else {
2766                                 if (lcache) {
2767                                         item = TAILQ_FIRST(&lcache->list);
2768                                         KKASSERT(item != NULL);
2769                                         item->flags |= HAMMER_NODE_LOCK_LCACHE;
2770                                         TAILQ_REMOVE(&lcache->list,
2771                                                      item, entry);
2772                                 } else {
2773                                         item = kmalloc(sizeof(*item),
2774                                                        hmp->m_misc,
2775                                                        M_WAITOK|M_ZERO);
2776                                         TAILQ_INIT(&item->list);
2777                                 }
2778
2779                                 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2780                                 item->parent = parent;
2781                                 item->node = child;
2782                                 item->index = i;
2783                                 item->count = child->ondisk->count;
2784
2785                                 /*
2786                                  * Recurse (used by the rebalancing code)
2787                                  */
2788                                 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2789                                         error = hammer_btree_lock_children(
2790                                                         cursor,
2791                                                         depth - 1,
2792                                                         item,
2793                                                         lcache);
2794                                 }
2795                         }
2796                 }
2797         }
2798         if (error)
2799                 hammer_btree_unlock_children(hmp, parent, lcache);
2800         return(error);
2801 }
2802
2803 /*
2804  * Create an in-memory copy of all B-Tree nodes listed, recursively,
2805  * including the parent.
2806  */
2807 void
2808 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2809 {
2810         hammer_mount_t hmp = cursor->trans->hmp;
2811         hammer_node_lock_t item;
2812
2813         if (parent->copy == NULL) {
2814                 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0);
2815                 parent->copy = kmalloc(sizeof(*parent->copy),
2816                                        hmp->m_misc, M_WAITOK);
2817         }
2818         KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0);
2819         *parent->copy = *parent->node->ondisk;
2820         TAILQ_FOREACH(item, &parent->list, entry) {
2821                 hammer_btree_lock_copy(cursor, item);
2822         }
2823 }
2824
2825 /*
2826  * Recursively sync modified copies to the media.
2827  */
2828 int
2829 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2830 {
2831         hammer_node_lock_t item;
2832         int count = 0;
2833
2834         if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2835                 ++count;
2836                 hammer_modify_node_all(cursor->trans, parent->node);
2837                 *parent->node->ondisk = *parent->copy;
2838                 hammer_modify_node_done(parent->node);
2839                 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2840                         hammer_flush_node(parent->node, 0);
2841                         hammer_delete_node(cursor->trans, parent->node);
2842                 }
2843         }
2844         TAILQ_FOREACH(item, &parent->list, entry) {
2845                 count += hammer_btree_sync_copy(cursor, item);
2846         }
2847         return(count);
2848 }
2849
2850 /*
2851  * Release previously obtained node locks.  The caller is responsible for
2852  * cleaning up parent->node itself (its usually just aliased from a cursor),
2853  * but this function will take care of the copies.
2854  *
2855  * NOTE: The root node is not placed in the lcache and node->copy is not
2856  *       deallocated when lcache != NULL.
2857  */
2858 void
2859 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent,
2860                              hammer_node_lock_t lcache)
2861 {
2862         hammer_node_lock_t item;
2863         hammer_node_ondisk_t copy;
2864
2865         while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2866                 TAILQ_REMOVE(&parent->list, item, entry);
2867                 hammer_btree_unlock_children(hmp, item, lcache);
2868                 hammer_unlock(&item->node->lock);
2869                 hammer_rel_node(item->node);
2870                 if (lcache) {
2871                         /*
2872                          * NOTE: When placing the item back in the lcache
2873                          *       the flag is cleared by the bzero().
2874                          *       Remaining fields are cleared as a safety
2875                          *       measure.
2876                          */
2877                         KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE);
2878                         KKASSERT(TAILQ_EMPTY(&item->list));
2879                         copy = item->copy;
2880                         bzero(item, sizeof(*item));
2881                         TAILQ_INIT(&item->list);
2882                         item->copy = copy;
2883                         if (copy)
2884                                 bzero(copy, sizeof(*copy));
2885                         TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2886                 } else {
2887                         kfree(item, hmp->m_misc);
2888                 }
2889         }
2890         if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) {
2891                 kfree(parent->copy, hmp->m_misc);
2892                 parent->copy = NULL;    /* safety */
2893         }
2894 }
2895
2896 /************************************************************************
2897  *                         MISCELLANIOUS SUPPORT                        *
2898  ************************************************************************/
2899
2900 /*
2901  * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2902  *
2903  * Note that for this particular function a return value of -1, 0, or +1
2904  * can denote a match if create_tid is otherwise discounted.  A create_tid
2905  * of zero is considered to be 'infinity' in comparisons.
2906  *
2907  * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2908  */
2909 int
2910 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2911 {
2912         if (key1->localization < key2->localization)
2913                 return(-5);
2914         if (key1->localization > key2->localization)
2915                 return(5);
2916
2917         if (key1->obj_id < key2->obj_id)
2918                 return(-4);
2919         if (key1->obj_id > key2->obj_id)
2920                 return(4);
2921
2922         if (key1->rec_type < key2->rec_type)
2923                 return(-3);
2924         if (key1->rec_type > key2->rec_type)
2925                 return(3);
2926
2927         if (key1->key < key2->key)
2928                 return(-2);
2929         if (key1->key > key2->key)
2930                 return(2);
2931
2932         /*
2933          * A create_tid of zero indicates a record which is undeletable
2934          * and must be considered to have a value of positive infinity.
2935          */
2936         if (key1->create_tid == 0) {
2937                 if (key2->create_tid == 0)
2938                         return(0);
2939                 return(1);
2940         }
2941         if (key2->create_tid == 0)
2942                 return(-1);
2943         if (key1->create_tid < key2->create_tid)
2944                 return(-1);
2945         if (key1->create_tid > key2->create_tid)
2946                 return(1);
2947         return(0);
2948 }
2949
2950 /*
2951  * Test a timestamp against an element to determine whether the
2952  * element is visible.  A timestamp of 0 means 'infinity'.
2953  */
2954 int
2955 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2956 {
2957         if (asof == 0) {
2958                 if (base->delete_tid)
2959                         return(1);
2960                 return(0);
2961         }
2962         if (asof < base->create_tid)
2963                 return(-1);
2964         if (base->delete_tid && asof >= base->delete_tid)
2965                 return(1);
2966         return(0);
2967 }
2968
2969 /*
2970  * Create a separator half way inbetween key1 and key2.  For fields just
2971  * one unit apart, the separator will match key2.  key1 is on the left-hand
2972  * side and key2 is on the right-hand side.
2973  *
2974  * key2 must be >= the separator.  It is ok for the separator to match key2.
2975  *
2976  * NOTE: Even if key1 does not match key2, the separator may wind up matching
2977  * key2.
2978  *
2979  * NOTE: It might be beneficial to just scrap this whole mess and just
2980  * set the separator to key2.
2981  */
2982 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2983         dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2984
2985 static void
2986 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2987                       hammer_base_elm_t dest)
2988 {
2989         bzero(dest, sizeof(*dest));
2990
2991         dest->rec_type = key2->rec_type;
2992         dest->key = key2->key;
2993         dest->obj_id = key2->obj_id;
2994         dest->create_tid = key2->create_tid;
2995
2996         MAKE_SEPARATOR(key1, key2, dest, localization);
2997         if (key1->localization == key2->localization) {
2998                 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2999                 if (key1->obj_id == key2->obj_id) {
3000                         MAKE_SEPARATOR(key1, key2, dest, rec_type);
3001                         if (key1->rec_type == key2->rec_type) {
3002                                 MAKE_SEPARATOR(key1, key2, dest, key);
3003                                 /*
3004                                  * Don't bother creating a separator for
3005                                  * create_tid, which also conveniently avoids
3006                                  * having to handle the create_tid == 0
3007                                  * (infinity) case.  Just leave create_tid
3008                                  * set to key2.
3009                                  *
3010                                  * Worst case, dest matches key2 exactly,
3011                                  * which is acceptable.
3012                                  */
3013                         }
3014                 }
3015         }
3016 }
3017
3018 #undef MAKE_SEPARATOR
3019
3020 /*
3021  * Return whether a generic internal or leaf node is full
3022  */
3023 static int
3024 btree_node_is_full(hammer_node_ondisk_t node)
3025 {
3026         switch(node->type) {
3027         case HAMMER_BTREE_TYPE_INTERNAL:
3028                 if (node->count == HAMMER_BTREE_INT_ELMS)
3029                         return(1);
3030                 break;
3031         case HAMMER_BTREE_TYPE_LEAF:
3032                 if (node->count == HAMMER_BTREE_LEAF_ELMS)
3033                         return(1);
3034                 break;
3035         default:
3036                 panic("illegal btree type");
3037         }
3038         return(0);
3039 }
3040
3041 #if 0
3042 static int
3043 btree_max_elements(u_int8_t type)
3044 {
3045         if (type == HAMMER_BTREE_TYPE_LEAF)
3046                 return(HAMMER_BTREE_LEAF_ELMS);
3047         if (type == HAMMER_BTREE_TYPE_INTERNAL)
3048                 return(HAMMER_BTREE_INT_ELMS);
3049         panic("btree_max_elements: bad type %d", type);
3050 }
3051 #endif
3052
3053 void
3054 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
3055 {
3056         hammer_btree_elm_t elm;
3057         int i;
3058
3059         kprintf("node %p count=%d parent=%016llx type=%c\n",
3060                 ondisk, ondisk->count,
3061                 (long long)ondisk->parent, ondisk->type);
3062
3063         /*
3064          * Dump both boundary elements if an internal node
3065          */
3066         if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
3067                 for (i = 0; i <= ondisk->count; ++i) {
3068                         elm = &ondisk->elms[i];
3069                         hammer_print_btree_elm(elm, ondisk->type, i);
3070                 }
3071         } else {
3072                 for (i = 0; i < ondisk->count; ++i) {
3073                         elm = &ondisk->elms[i];
3074                         hammer_print_btree_elm(elm, ondisk->type, i);
3075                 }
3076         }
3077 }
3078
3079 void
3080 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
3081 {
3082         kprintf("  %2d", i);
3083         kprintf("\tobj_id       = %016llx\n", (long long)elm->base.obj_id);
3084         kprintf("\tkey          = %016llx\n", (long long)elm->base.key);
3085         kprintf("\tcreate_tid   = %016llx\n", (long long)elm->base.create_tid);
3086         kprintf("\tdelete_tid   = %016llx\n", (long long)elm->base.delete_tid);
3087         kprintf("\trec_type     = %04x\n", elm->base.rec_type);
3088         kprintf("\tobj_type     = %02x\n", elm->base.obj_type);
3089         kprintf("\tbtype        = %02x (%c)\n",
3090                 elm->base.btype,
3091                 (elm->base.btype ? elm->base.btype : '?'));
3092         kprintf("\tlocalization = %02x\n", elm->base.localization);
3093
3094         switch(type) {
3095         case HAMMER_BTREE_TYPE_INTERNAL:
3096                 kprintf("\tsubtree_off  = %016llx\n",
3097                         (long long)elm->internal.subtree_offset);
3098                 break;
3099         case HAMMER_BTREE_TYPE_RECORD:
3100                 kprintf("\tdata_offset  = %016llx\n",
3101                         (long long)elm->leaf.data_offset);
3102                 kprintf("\tdata_len     = %08x\n", elm->leaf.data_len);
3103                 kprintf("\tdata_crc     = %08x\n", elm->leaf.data_crc);
3104                 break;
3105         }
3106 }