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  * instead of sub-tree pointers.
  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  *
  58  * The big benefit to using a B-Tree containing boundary information
  59  * is that it is possible to cache pointers into the middle of the tree
  60  * and not have to start searches, insertions, OR deletions at the root
  61  * node.   In particular, searches are able to progress in a definitive
  62  * direction from any point in the tree without revisting nodes.  This
  63  * greatly improves the efficiency of many operations, most especially
  64  * record appends.
  65  *
  66  * B-Trees also make the stacking of trees fairly straightforward.
  67  *
  68  * INSERTIONS:  A search performed with the intention of doing
  69  * an insert will guarantee that the terminal leaf node is not full by
  70  * splitting full nodes.  Splits occur top-down during the dive down the
  71  * B-Tree.
  72  *
  73  * DELETIONS: A deletion makes no attempt to proactively balance the
  74  * tree and will recursively remove nodes that become empty.  If a
  75  * deadlock occurs a deletion may not be able to remove an empty leaf.
  76  * Deletions never allow internal nodes to become empty (that would blow
  77  * up the boundaries).
  78  */
  79 #include "hammer.h"
  80 #include <sys/buf.h>
  81 #include <sys/buf2.h>
  82
  83 static int btree_search(hammer_cursor_t cursor, int flags);
  84 static int btree_split_internal(hammer_cursor_t cursor);
  85 static int btree_split_leaf(hammer_cursor_t cursor);
  86 static int btree_remove(hammer_cursor_t cursor);
  87 static int btree_node_is_full(hammer_node_ondisk_t node);
  88 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
  89                         hammer_tid_t mirror_tid);
  90 static void hammer_make_separator(hammer_base_elm_t key1,
  91                         hammer_base_elm_t key2, hammer_base_elm_t dest);
  92 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
  93
  94 /*
  95  * Iterate records after a search.  The cursor is iterated forwards past
  96  * the current record until a record matching the key-range requirements
  97  * is found.  ENOENT is returned if the iteration goes past the ending
  98  * key.
  99  *
 100  * The iteration is inclusive of key_beg and can be inclusive or exclusive
 101  * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
 102  *
 103  * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
 104  * may be modified by B-Tree functions.
 105  *
 106  * cursor->key_beg may or may not be modified by this function during
 107  * the iteration.  XXX future - in case of an inverted lock we may have
 108  * to reinitiate the lookup and set key_beg to properly pick up where we
 109  * left off.
 110  *
 111  * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor
 112  * was reverse indexed due to being moved to a parent while unlocked,
 113  * and something else might have inserted an element outside the iteration
 114  * range.  When this case occurs the iterator just keeps iterating until
 115  * it gets back into the iteration range (instead of asserting).
 116  *
 117  * NOTE!  EDEADLK *CANNOT* be returned by this procedure.
 118  */
 119 int
 120 hammer_btree_iterate(hammer_cursor_t cursor)
 121 {
 122         hammer_node_ondisk_t node;
 123         hammer_btree_elm_t elm;
 124         hammer_mount_t hmp;
 125         int error = 0;
 126         int r;
 127         int s;
 128
 129         /*
 130          * Skip past the current record
 131          */
 132         hmp = cursor->trans->hmp;
 133         node = cursor->node->ondisk;
 134         if (node == NULL)
 135                 return(ENOENT);
 136         if (cursor->index < node->count &&
 137             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 138                 ++cursor->index;
 139         }
 140
 141         /*
 142          * HAMMER can wind up being cpu-bound.
 143          */
 144         if (++hmp->check_yield > hammer_yield_check) {
 145                 hmp->check_yield = 0;
 146                 lwkt_user_yield();
 147         }
 148
 149
 150         /*
 151          * Loop until an element is found or we are done.
 152          */
 153         for (;;) {
 154                 /*
 155                  * We iterate up the tree and then index over one element
 156                  * while we are at the last element in the current node.
 157                  *
 158                  * If we are at the root of the filesystem, cursor_up
 159                  * returns ENOENT.
 160                  *
 161                  * XXX this could be optimized by storing the information in
 162                  * the parent reference.
 163                  *
 164                  * XXX we can lose the node lock temporarily, this could mess
 165                  * up our scan.
 166                  */
 167                 ++hammer_stats_btree_iterations;
 168                 hammer_flusher_clean_loose_ios(hmp);
 169
 170                 if (cursor->index == node->count) {
 171                         if (hammer_debug_btree) {
 172                                 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
 173                                         (long long)cursor->node->node_offset,
 174                                         cursor->index,
 175                                         (long long)(cursor->parent ? cursor->parent->node_offset : -1),
 176                                         cursor->parent_index,
 177                                         curthread);
 178                         }
 179                         KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
 180                         error = hammer_cursor_up(cursor);
 181                         if (error)
 182                                 break;
 183                         /* reload stale pointer */
 184                         node = cursor->node->ondisk;
 185                         KKASSERT(cursor->index != node->count);
 186
 187                         /*
 188                          * If we are reblocking we want to return internal
 189                          * nodes.  Note that the internal node will be
 190                          * returned multiple times, on each upward recursion
 191                          * from its children.  The caller selects which
 192                          * revisit it cares about (usually first or last only).
 193                          */
 194                         if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
 195                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 196                                 return(0);
 197                         }
 198                         ++cursor->index;
 199                         continue;
 200                 }
 201
 202                 /*
 203                  * Check internal or leaf element.  Determine if the record
 204                  * at the cursor has gone beyond the end of our range.
 205                  *
 206                  * We recurse down through internal nodes.
 207                  */
 208                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 209                         elm = &node->elms[cursor->index];
 210
 211                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 212                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 213                         if (hammer_debug_btree) {
 214                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
 215                                         (long long)cursor->node->node_offset,
 216                                         cursor->index,
 217                                         (long long)elm[0].internal.base.obj_id,
 218                                         elm[0].internal.base.rec_type,
 219                                         (long long)elm[0].internal.base.key,
 220                                         elm[0].internal.base.localization,
 221                                         r,
 222                                         curthread
 223                                 );
 224                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 225                                         (long long)cursor->node->node_offset,
 226                                         cursor->index + 1,
 227                                         (long long)elm[1].internal.base.obj_id,
 228                                         elm[1].internal.base.rec_type,
 229                                         (long long)elm[1].internal.base.key,
 230                                         elm[1].internal.base.localization,
 231                                         s
 232                                 );
 233                         }
 234
 235                         if (r < 0) {
 236                                 error = ENOENT;
 237                                 break;
 238                         }
 239                         if (r == 0 && (cursor->flags &
 240                                        HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 241                                 error = ENOENT;
 242                                 break;
 243                         }
 244
 245                         /*
 246                          * Better not be zero
 247                          */
 248                         KKASSERT(elm->internal.subtree_offset != 0);
 249
 250                         if (s <= 0) {
 251                                 /*
 252                                  * If running the mirror filter see if we
 253                                  * can skip one or more entire sub-trees.
 254                                  * If we can we return the internal node
 255                                  * and the caller processes the skipped
 256                                  * range (see mirror_read).
 257                                  */
 258                                 if (cursor->flags &
 259                                     HAMMER_CURSOR_MIRROR_FILTERED) {
 260                                         if (elm->internal.mirror_tid <
 261                                             cursor->cmirror->mirror_tid) {
 262                                                 hammer_cursor_mirror_filter(cursor);
 263                                                 return(0);
 264                                         }
 265                                 }
 266                         } else {
 267                                 /*
 268                                  * Normally it would be impossible for the
 269                                  * cursor to have gotten back-indexed,
 270                                  * but it can happen if a node is deleted
 271                                  * and the cursor is moved to its parent
 272                                  * internal node.  ITERATE_CHECK will be set.
 273                                  */
 274                                 KKASSERT(cursor->flags &
 275                                          HAMMER_CURSOR_ITERATE_CHECK);
 276                                 kprintf("hammer_btree_iterate: "
 277                                         "DEBUG: Caught parent seek "
 278                                         "in internal iteration\n");
 279                         }
 280
 281                         error = hammer_cursor_down(cursor);
 282                         if (error)
 283                                 break;
 284                         KKASSERT(cursor->index == 0);
 285                         /* reload stale pointer */
 286                         node = cursor->node->ondisk;
 287                         continue;
 288                 } else {
 289                         elm = &node->elms[cursor->index];
 290                         r = hammer_btree_cmp(&cursor->key_end, &elm->base);
 291                         if (hammer_debug_btree) {
 292                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
 293                                         (long long)cursor->node->node_offset,
 294                                         cursor->index,
 295                                         (elm[0].leaf.base.btype ?
 296                                          elm[0].leaf.base.btype : '?'),
 297                                         (long long)elm[0].leaf.base.obj_id,
 298                                         elm[0].leaf.base.rec_type,
 299                                         (long long)elm[0].leaf.base.key,
 300                                         elm[0].leaf.base.localization,
 301                                         r
 302                                 );
 303                         }
 304                         if (r < 0) {
 305                                 error = ENOENT;
 306                                 break;
 307                         }
 308
 309                         /*
 310                          * We support both end-inclusive and
 311                          * end-exclusive searches.
 312                          */
 313                         if (r == 0 &&
 314                            (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 315                                 error = ENOENT;
 316                                 break;
 317                         }
 318
 319                         /*
 320                          * If ITERATE_CHECK is set an unlocked cursor may
 321                          * have been moved to a parent and the iterate can
 322                          * happen upon elements that are not in the requested
 323                          * range.
 324                          */
 325                         if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) {
 326                                 s = hammer_btree_cmp(&cursor->key_beg,
 327                                                      &elm->base);
 328                                 if (s > 0) {
 329                                         kprintf("hammer_btree_iterate: "
 330                                                 "DEBUG: Caught parent seek "
 331                                                 "in leaf iteration\n");
 332                                         ++cursor->index;
 333                                         continue;
 334                                 }
 335                         }
 336                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 337
 338                         /*
 339                          * Return the element
 340                          */
 341                         switch(elm->leaf.base.btype) {
 342                         case HAMMER_BTREE_TYPE_RECORD:
 343                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 344                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 345                                         ++cursor->index;
 346                                         continue;
 347                                 }
 348                                 error = 0;
 349                                 break;
 350                         default:
 351                                 error = EINVAL;
 352                                 break;
 353                         }
 354                         if (error)
 355                                 break;
 356                 }
 357                 /*
 358                  * node pointer invalid after loop
 359                  */
 360
 361                 /*
 362                  * Return entry
 363                  */
 364                 if (hammer_debug_btree) {
 365                         int i = cursor->index;
 366                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 367                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
 368                                 cursor->node, i,
 369                                 (long long)elm->internal.base.obj_id,
 370                                 elm->internal.base.rec_type,
 371                                 (long long)elm->internal.base.key,
 372                                 elm->internal.base.localization
 373                         );
 374                 }
 375                 return(0);
 376         }
 377         return(error);
 378 }
 379
 380 /*
 381  * We hit an internal element that we could skip as part of a mirroring
 382  * scan.  Calculate the entire range being skipped.
 383  *
 384  * It is important to include any gaps between the parent's left_bound
 385  * and the node's left_bound, and same goes for the right side.
 386  */
 387 static void
 388 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
 389 {
 390         struct hammer_cmirror *cmirror;
 391         hammer_node_ondisk_t ondisk;
 392         hammer_btree_elm_t elm;
 393
 394         ondisk = cursor->node->ondisk;
 395         cmirror = cursor->cmirror;
 396
 397         /*
 398          * Calculate the skipped range
 399          */
 400         elm = &ondisk->elms[cursor->index];
 401         if (cursor->index == 0)
 402                 cmirror->skip_beg = *cursor->left_bound;
 403         else
 404                 cmirror->skip_beg = elm->internal.base;
 405         while (cursor->index < ondisk->count) {
 406                 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
 407                         break;
 408                 ++cursor->index;
 409                 ++elm;
 410         }
 411         if (cursor->index == ondisk->count)
 412                 cmirror->skip_end = *cursor->right_bound;
 413         else
 414                 cmirror->skip_end = elm->internal.base;
 415
 416         /*
 417          * clip the returned result.
 418          */
 419         if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
 420                 cmirror->skip_beg = cursor->key_beg;
 421         if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
 422                 cmirror->skip_end = cursor->key_end;
 423 }
 424
 425 /*
 426  * Iterate in the reverse direction.  This is used by the pruning code to
 427  * avoid overlapping records.
 428  */
 429 int
 430 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
 431 {
 432         hammer_node_ondisk_t node;
 433         hammer_btree_elm_t elm;
 434         hammer_mount_t hmp;
 435         int error = 0;
 436         int r;
 437         int s;
 438
 439         /* mirror filtering not supported for reverse iteration */
 440         KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
 441
 442         /*
 443          * Skip past the current record.  For various reasons the cursor
 444          * may end up set to -1 or set to point at the end of the current
 445          * node.  These cases must be addressed.
 446          */
 447         node = cursor->node->ondisk;
 448         if (node == NULL)
 449                 return(ENOENT);
 450         if (cursor->index != -1 &&
 451             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 452                 --cursor->index;
 453         }
 454         if (cursor->index == cursor->node->ondisk->count)
 455                 --cursor->index;
 456
 457         /*
 458          * HAMMER can wind up being cpu-bound.
 459          */
 460         hmp = cursor->trans->hmp;
 461         if (++hmp->check_yield > hammer_yield_check) {
 462                 hmp->check_yield = 0;
 463                 lwkt_user_yield();
 464         }
 465
 466         /*
 467          * Loop until an element is found or we are done.
 468          */
 469         for (;;) {
 470                 ++hammer_stats_btree_iterations;
 471                 hammer_flusher_clean_loose_ios(hmp);
 472
 473                 /*
 474                  * We iterate up the tree and then index over one element
 475                  * while we are at the last element in the current node.
 476                  */
 477                 if (cursor->index == -1) {
 478                         error = hammer_cursor_up(cursor);
 479                         if (error) {
 480                                 cursor->index = 0; /* sanity */
 481                                 break;
 482                         }
 483                         /* reload stale pointer */
 484                         node = cursor->node->ondisk;
 485                         KKASSERT(cursor->index != node->count);
 486                         --cursor->index;
 487                         continue;
 488                 }
 489
 490                 /*
 491                  * Check internal or leaf element.  Determine if the record
 492                  * at the cursor has gone beyond the end of our range.
 493                  *
 494                  * We recurse down through internal nodes.
 495                  */
 496                 KKASSERT(cursor->index != node->count);
 497                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 498                         elm = &node->elms[cursor->index];
 499                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 500                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 501                         if (hammer_debug_btree) {
 502                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 503                                         (long long)cursor->node->node_offset,
 504                                         cursor->index,
 505                                         (long long)elm[0].internal.base.obj_id,
 506                                         elm[0].internal.base.rec_type,
 507                                         (long long)elm[0].internal.base.key,
 508                                         elm[0].internal.base.localization,
 509                                         r
 510                                 );
 511                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
 512                                         (long long)cursor->node->node_offset,
 513                                         cursor->index + 1,
 514                                         (long long)elm[1].internal.base.obj_id,
 515                                         elm[1].internal.base.rec_type,
 516                                         (long long)elm[1].internal.base.key,
 517                                         elm[1].internal.base.localization,
 518                                         s
 519                                 );
 520                         }
 521
 522                         if (s >= 0) {
 523                                 error = ENOENT;
 524                                 break;
 525                         }
 526
 527                         /*
 528                          * It shouldn't be possible to be seeked past key_end,
 529                          * even if the cursor got moved to a parent.
 530                          */
 531                         KKASSERT(r >= 0);
 532
 533                         /*
 534                          * Better not be zero
 535                          */
 536                         KKASSERT(elm->internal.subtree_offset != 0);
 537
 538                         error = hammer_cursor_down(cursor);
 539                         if (error)
 540                                 break;
 541                         KKASSERT(cursor->index == 0);
 542                         /* reload stale pointer */
 543                         node = cursor->node->ondisk;
 544
 545                         /* this can assign -1 if the leaf was empty */
 546                         cursor->index = node->count - 1;
 547                         continue;
 548                 } else {
 549                         elm = &node->elms[cursor->index];
 550                         s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
 551                         if (hammer_debug_btree) {
 552                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
 553                                         (long long)cursor->node->node_offset,
 554                                         cursor->index,
 555                                         (elm[0].leaf.base.btype ?
 556                                          elm[0].leaf.base.btype : '?'),
 557                                         (long long)elm[0].leaf.base.obj_id,
 558                                         elm[0].leaf.base.rec_type,
 559                                         (long long)elm[0].leaf.base.key,
 560                                         elm[0].leaf.base.localization,
 561                                         s
 562                                 );
 563                         }
 564                         if (s > 0) {
 565                                 error = ENOENT;
 566                                 break;
 567                         }
 568
 569                         /*
 570                          * It shouldn't be possible to be seeked past key_end,
 571                          * even if the cursor got moved to a parent.
 572                          */
 573                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 574
 575                         /*
 576                          * Return the element
 577                          */
 578                         switch(elm->leaf.base.btype) {
 579                         case HAMMER_BTREE_TYPE_RECORD:
 580                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 581                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 582                                         --cursor->index;
 583                                         continue;
 584                                 }
 585                                 error = 0;
 586                                 break;
 587                         default:
 588                                 error = EINVAL;
 589                                 break;
 590                         }
 591                         if (error)
 592                                 break;
 593                 }
 594                 /*
 595                  * node pointer invalid after loop
 596                  */
 597
 598                 /*
 599                  * Return entry
 600                  */
 601                 if (hammer_debug_btree) {
 602                         int i = cursor->index;
 603                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 604                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
 605                                 cursor->node, i,
 606                                 (long long)elm->internal.base.obj_id,
 607                                 elm->internal.base.rec_type,
 608                                 (long long)elm->internal.base.key,
 609                                 elm->internal.base.localization
 610                         );
 611                 }
 612                 return(0);
 613         }
 614         return(error);
 615 }
 616
 617 /*
 618  * Lookup cursor->key_beg.  0 is returned on success, ENOENT if the entry
 619  * could not be found, EDEADLK if inserting and a retry is needed, and a
 620  * fatal error otherwise.  When retrying, the caller must terminate the
 621  * cursor and reinitialize it.  EDEADLK cannot be returned if not inserting.
 622  *
 623  * The cursor is suitably positioned for a deletion on success, and suitably
 624  * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
 625  * specified.
 626  *
 627  * The cursor may begin anywhere, the search will traverse the tree in
 628  * either direction to locate the requested element.
 629  *
 630  * Most of the logic implementing historical searches is handled here.  We
 631  * do an initial lookup with create_tid set to the asof TID.  Due to the
 632  * way records are laid out, a backwards iteration may be required if
 633  * ENOENT is returned to locate the historical record.  Here's the
 634  * problem:
 635  *
 636  * create_tid:    10      15       20
 637  *                   LEAF1   LEAF2
 638  * records:         (11)        (18)
 639  *
 640  * Lets say we want to do a lookup AS-OF timestamp 17.  We will traverse
 641  * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
 642  * not visible and thus causes ENOENT to be returned.  We really need
 643  * to check record 11 in LEAF1.  If it also fails then the search fails
 644  * (e.g. it might represent the range 11-16 and thus still not match our
 645  * AS-OF timestamp of 17).  Note that LEAF1 could be empty, requiring
 646  * further iterations.
 647  *
 648  * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
 649  * and the cursor->create_check TID if an iteration might be needed.
 650  * In the above example create_check would be set to 14.
 651  */
 652 int
 653 hammer_btree_lookup(hammer_cursor_t cursor)
 654 {
 655         int error;
 656
 657         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
 658         KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
 659                   cursor->trans->sync_lock_refs > 0);
 660         ++hammer_stats_btree_lookups;
 661         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 662                 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
 663                 cursor->key_beg.create_tid = cursor->asof;
 664                 for (;;) {
 665                         cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
 666                         error = btree_search(cursor, 0);
 667                         if (error != ENOENT ||
 668                             (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
 669                                 /*
 670                                  * Stop if no error.
 671                                  * Stop if error other then ENOENT.
 672                                  * Stop if ENOENT and not special case.
 673                                  */
 674                                 break;
 675                         }
 676                         if (hammer_debug_btree) {
 677                                 kprintf("CREATE_CHECK %016llx\n",
 678                                         (long long)cursor->create_check);
 679                         }
 680                         cursor->key_beg.create_tid = cursor->create_check;
 681                         /* loop */
 682                 }
 683         } else {
 684                 error = btree_search(cursor, 0);
 685         }
 686         if (error == 0)
 687                 error = hammer_btree_extract(cursor, cursor->flags);
 688         return(error);
 689 }
 690
 691 /*
 692  * Execute the logic required to start an iteration.  The first record
 693  * located within the specified range is returned and iteration control
 694  * flags are adjusted for successive hammer_btree_iterate() calls.
 695  *
 696  * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
 697  * in a loop without worrying about it.  Higher-level merged searches will
 698  * adjust the flag appropriately.
 699  */
 700 int
 701 hammer_btree_first(hammer_cursor_t cursor)
 702 {
 703         int error;
 704
 705         error = hammer_btree_lookup(cursor);
 706         if (error == ENOENT) {
 707                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 708                 error = hammer_btree_iterate(cursor);
 709         }
 710         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 711         return(error);
 712 }
 713
 714 /*
 715  * Similarly but for an iteration in the reverse direction.
 716  *
 717  * Set ATEDISK when iterating backwards to skip the current entry,
 718  * which after an ENOENT lookup will be pointing beyond our end point.
 719  *
 720  * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
 721  * in a loop without worrying about it.  Higher-level merged searches will
 722  * adjust the flag appropriately.
 723  */
 724 int
 725 hammer_btree_last(hammer_cursor_t cursor)
 726 {
 727         struct hammer_base_elm save;
 728         int error;
 729
 730         save = cursor->key_beg;
 731         cursor->key_beg = cursor->key_end;
 732         error = hammer_btree_lookup(cursor);
 733         cursor->key_beg = save;
 734         if (error == ENOENT ||
 735             (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 736                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
 737                 error = hammer_btree_iterate_reverse(cursor);
 738         }
 739         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 740         return(error);
 741 }
 742
 743 /*
 744  * Extract the record and/or data associated with the cursor's current
 745  * position.  Any prior record or data stored in the cursor is replaced.
 746  * The cursor must be positioned at a leaf node.
 747  *
 748  * NOTE: All extractions occur at the leaf of the B-Tree.
 749  */
 750 int
 751 hammer_btree_extract(hammer_cursor_t cursor, int flags)
 752 {
 753         hammer_node_ondisk_t node;
 754         hammer_btree_elm_t elm;
 755         hammer_off_t data_off;
 756         hammer_mount_t hmp;
 757         int32_t data_len;
 758         int error;
 759
 760         /*
 761          * The case where the data reference resolves to the same buffer
 762          * as the record reference must be handled.
 763          */
 764         node = cursor->node->ondisk;
 765         elm = &node->elms[cursor->index];
 766         cursor->data = NULL;
 767         hmp = cursor->node->hmp;
 768
 769         /*
 770          * There is nothing to extract for an internal element.
 771          */
 772         if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
 773                 return(EINVAL);
 774
 775         /*
 776          * Only record types have data.
 777          */
 778         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 779         cursor->leaf = &elm->leaf;
 780
 781         if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
 782                 return(0);
 783         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
 784                 return(0);
 785         data_off = elm->leaf.data_offset;
 786         data_len = elm->leaf.data_len;
 787         if (data_off == 0)
 788                 return(0);
 789
 790         /*
 791          * Load the data
 792          */
 793         KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
 794         cursor->data = hammer_bread_ext(hmp, data_off, data_len,
 795                                         &error, &cursor->data_buffer);
 796
 797         /*
 798          * Mark the data buffer as not being meta-data if it isn't
 799          * meta-data (sometimes bulk data is accessed via a volume
 800          * block device).
 801          */
 802         if (error == 0) {
 803                 switch(elm->leaf.base.rec_type) {
 804                 case HAMMER_RECTYPE_DATA:
 805                 case HAMMER_RECTYPE_DB:
 806                         if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0)
 807                                 break;
 808                         if (hammer_double_buffer == 0 ||
 809                             (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) {
 810                                 hammer_io_notmeta(cursor->data_buffer);
 811                         }
 812                         break;
 813                 default:
 814                         break;
 815                 }
 816         }
 817
 818         /*
 819          * Deal with CRC errors on the extracted data.
 820          */
 821         if (error == 0 &&
 822             hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
 823                 kprintf("CRC DATA @ %016llx/%d FAILED\n",
 824                         (long long)elm->leaf.data_offset, elm->leaf.data_len);
 825                 if (hammer_debug_critical)
 826                         Debugger("CRC FAILED: DATA");
 827                 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
 828                         error = EDOM;   /* less critical (mirroring) */
 829                 else
 830                         error = EIO;    /* critical */
 831         }
 832         return(error);
 833 }
 834
 835
 836 /*
 837  * Insert a leaf element into the B-Tree at the current cursor position.
 838  * The cursor is positioned such that the element at and beyond the cursor
 839  * are shifted to make room for the new record.
 840  *
 841  * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
 842  * flag set and that call must return ENOENT before this function can be
 843  * called.
 844  *
 845  * The caller may depend on the cursor's exclusive lock after return to
 846  * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
 847  *
 848  * ENOSPC is returned if there is no room to insert a new record.
 849  */
 850 int
 851 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
 852                     int *doprop)
 853 {
 854         hammer_node_ondisk_t node;
 855         int i;
 856         int error;
 857
 858         *doprop = 0;
 859         if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
 860                 return(error);
 861         ++hammer_stats_btree_inserts;
 862
 863         /*
 864          * Insert the element at the leaf node and update the count in the
 865          * parent.  It is possible for parent to be NULL, indicating that
 866          * the filesystem's ROOT B-Tree node is a leaf itself, which is
 867          * possible.  The root inode can never be deleted so the leaf should
 868          * never be empty.
 869          *
 870          * Remember that the right-hand boundary is not included in the
 871          * count.
 872          */
 873         hammer_modify_node_all(cursor->trans, cursor->node);
 874         node = cursor->node->ondisk;
 875         i = cursor->index;
 876         KKASSERT(elm->base.btype != 0);
 877         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 878         KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
 879         if (i != node->count) {
 880                 bcopy(&node->elms[i], &node->elms[i+1],
 881                       (node->count - i) * sizeof(*elm));
 882         }
 883         node->elms[i].leaf = *elm;
 884         ++node->count;
 885         hammer_cursor_inserted_element(cursor->node, i);
 886
 887         /*
 888          * Update the leaf node's aggregate mirror_tid for mirroring
 889          * support.
 890          */
 891         if (node->mirror_tid < elm->base.delete_tid) {
 892                 node->mirror_tid = elm->base.delete_tid;
 893                 *doprop = 1;
 894         }
 895         if (node->mirror_tid < elm->base.create_tid) {
 896                 node->mirror_tid = elm->base.create_tid;
 897                 *doprop = 1;
 898         }
 899         hammer_modify_node_done(cursor->node);
 900
 901         /*
 902          * Debugging sanity checks.
 903          */
 904         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
 905         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
 906         if (i) {
 907                 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
 908         }
 909         if (i != node->count - 1)
 910                 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
 911
 912         return(0);
 913 }
 914
 915 /*
 916  * Delete a record from the B-Tree at the current cursor position.
 917  * The cursor is positioned such that the current element is the one
 918  * to be deleted.
 919  *
 920  * On return the cursor will be positioned after the deleted element and
 921  * MAY point to an internal node.  It will be suitable for the continuation
 922  * of an iteration but not for an insertion or deletion.
 923  *
 924  * Deletions will attempt to partially rebalance the B-Tree in an upward
 925  * direction, but will terminate rather then deadlock.  Empty internal nodes
 926  * are never allowed by a deletion which deadlocks may end up giving us an
 927  * empty leaf.  The pruner will clean up and rebalance the tree.
 928  *
 929  * This function can return EDEADLK, requiring the caller to retry the
 930  * operation after clearing the deadlock.
 931  */
 932 int
 933 hammer_btree_delete(hammer_cursor_t cursor)
 934 {
 935         hammer_node_ondisk_t ondisk;
 936         hammer_node_t node;
 937         hammer_node_t parent __debugvar;
 938         int error;
 939         int i;
 940
 941         KKASSERT (cursor->trans->sync_lock_refs > 0);
 942         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 943                 return(error);
 944         ++hammer_stats_btree_deletes;
 945
 946         /*
 947          * Delete the element from the leaf node.
 948          *
 949          * Remember that leaf nodes do not have boundaries.
 950          */
 951         node = cursor->node;
 952         ondisk = node->ondisk;
 953         i = cursor->index;
 954
 955         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
 956         KKASSERT(i >= 0 && i < ondisk->count);
 957         hammer_modify_node_all(cursor->trans, node);
 958         if (i + 1 != ondisk->count) {
 959                 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
 960                       (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
 961         }
 962         --ondisk->count;
 963         hammer_modify_node_done(node);
 964         hammer_cursor_deleted_element(node, i);
 965
 966         /*
 967          * Validate local parent
 968          */
 969         if (ondisk->parent) {
 970                 parent = cursor->parent;
 971
 972                 KKASSERT(parent != NULL);
 973                 KKASSERT(parent->node_offset == ondisk->parent);
 974         }
 975
 976         /*
 977          * If the leaf becomes empty it must be detached from the parent,
 978          * potentially recursing through to the filesystem root.
 979          *
 980          * This may reposition the cursor at one of the parent's of the
 981          * current node.
 982          *
 983          * Ignore deadlock errors, that simply means that btree_remove
 984          * was unable to recurse and had to leave us with an empty leaf.
 985          */
 986         KKASSERT(cursor->index <= ondisk->count);
 987         if (ondisk->count == 0) {
 988                 error = btree_remove(cursor);
 989                 if (error == EDEADLK)
 990                         error = 0;
 991         } else {
 992                 error = 0;
 993         }
 994         KKASSERT(cursor->parent == NULL ||
 995                  cursor->parent_index < cursor->parent->ondisk->count);
 996         return(error);
 997 }
 998
 999 /*
1000  * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
1001  *
1002  * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
1003  *
1004  * The search can begin ANYWHERE in the B-Tree.  As a first step the search
1005  * iterates up the tree as necessary to properly position itself prior to
1006  * actually doing the sarch.
1007  *
1008  * INSERTIONS: The search will split full nodes and leaves on its way down
1009  * and guarentee that the leaf it ends up on is not full.  If we run out
1010  * of space the search continues to the leaf, but ENOSPC is returned.
1011  *
1012  * The search is only guarenteed to end up on a leaf if an error code of 0
1013  * is returned, or if inserting and an error code of ENOENT is returned.
1014  * Otherwise it can stop at an internal node.  On success a search returns
1015  * a leaf node.
1016  *
1017  * COMPLEXITY WARNING!  This is the core B-Tree search code for the entire
1018  * filesystem, and it is not simple code.  Please note the following facts:
1019  *
1020  * - Internal node recursions have a boundary on the left AND right.  The
1021  *   right boundary is non-inclusive.  The create_tid is a generic part
1022  *   of the key for internal nodes.
1023  *
1024  * - Leaf nodes contain terminal elements only now.
1025  *
1026  * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1027  *   historical search.  ASOF and INSERT are mutually exclusive.  When
1028  *   doing an as-of lookup btree_search() checks for a right-edge boundary
1029  *   case.  If while recursing down the left-edge differs from the key
1030  *   by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1031  *   with cursor->create_check.  This is used by btree_lookup() to iterate.
1032  *   The iteration backwards because as-of searches can wind up going
1033  *   down the wrong branch of the B-Tree.
1034  */
1035 static
1036 int
1037 btree_search(hammer_cursor_t cursor, int flags)
1038 {
1039         hammer_node_ondisk_t node;
1040         hammer_btree_elm_t elm;
1041         int error;
1042         int enospc = 0;
1043         int i;
1044         int r;
1045         int s;
1046
1047         flags |= cursor->flags;
1048         ++hammer_stats_btree_searches;
1049
1050         if (hammer_debug_btree) {
1051                 kprintf("SEARCH   %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
1052                         (long long)cursor->node->node_offset,
1053                         cursor->index,
1054                         (long long)cursor->key_beg.obj_id,
1055                         cursor->key_beg.rec_type,
1056                         (long long)cursor->key_beg.key,
1057                         (long long)cursor->key_beg.create_tid,
1058                         cursor->key_beg.localization,
1059                         curthread
1060                 );
1061                 if (cursor->parent)
1062                     kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
1063                         (long long)cursor->parent->node_offset,
1064                         cursor->parent_index,
1065                         (long long)cursor->left_bound->obj_id,
1066                         (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
1067                         (long long)cursor->right_bound->obj_id,
1068                         (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
1069                         cursor->left_bound,
1070                         &cursor->parent->ondisk->elms[cursor->parent_index],
1071                         cursor->right_bound,
1072                         &cursor->parent->ondisk->elms[cursor->parent_index+1]
1073                     );
1074         }
1075
1076         /*
1077          * Move our cursor up the tree until we find a node whos range covers
1078          * the key we are trying to locate.
1079          *
1080          * The left bound is inclusive, the right bound is non-inclusive.
1081          * It is ok to cursor up too far.
1082          */
1083         for (;;) {
1084                 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1085                 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1086                 if (r >= 0 && s < 0)
1087                         break;
1088                 KKASSERT(cursor->parent);
1089                 ++hammer_stats_btree_iterations;
1090                 error = hammer_cursor_up(cursor);
1091                 if (error)
1092                         goto done;
1093         }
1094
1095         /*
1096          * The delete-checks below are based on node, not parent.  Set the
1097          * initial delete-check based on the parent.
1098          */
1099         if (r == 1) {
1100                 KKASSERT(cursor->left_bound->create_tid != 1);
1101                 cursor->create_check = cursor->left_bound->create_tid - 1;
1102                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1103         }
1104
1105         /*
1106          * We better have ended up with a node somewhere.
1107          */
1108         KKASSERT(cursor->node != NULL);
1109
1110         /*
1111          * If we are inserting we can't start at a full node if the parent
1112          * is also full (because there is no way to split the node),
1113          * continue running up the tree until the requirement is satisfied
1114          * or we hit the root of the filesystem.
1115          *
1116          * (If inserting we aren't doing an as-of search so we don't have
1117          *  to worry about create_check).
1118          */
1119         while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1120                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1121                         if (btree_node_is_full(cursor->node->ondisk) == 0)
1122                                 break;
1123                 } else {
1124                         if (btree_node_is_full(cursor->node->ondisk) ==0)
1125                                 break;
1126                 }
1127                 if (cursor->node->ondisk->parent == 0 ||
1128                     cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1129                         break;
1130                 }
1131                 ++hammer_stats_btree_iterations;
1132                 error = hammer_cursor_up(cursor);
1133                 /* node may have become stale */
1134                 if (error)
1135                         goto done;
1136         }
1137
1138         /*
1139          * Push down through internal nodes to locate the requested key.
1140          */
1141         node = cursor->node->ondisk;
1142         while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1143                 /*
1144                  * Scan the node to find the subtree index to push down into.
1145                  * We go one-past, then back-up.
1146                  *
1147                  * We must proactively remove deleted elements which may
1148                  * have been left over from a deadlocked btree_remove().
1149                  *
1150                  * The left and right boundaries are included in the loop
1151                  * in order to detect edge cases.
1152                  *
1153                  * If the separator only differs by create_tid (r == 1)
1154                  * and we are doing an as-of search, we may end up going
1155                  * down a branch to the left of the one containing the
1156                  * desired key.  This requires numerous special cases.
1157                  */
1158                 ++hammer_stats_btree_iterations;
1159                 if (hammer_debug_btree) {
1160                         kprintf("SEARCH-I %016llx count=%d\n",
1161                                 (long long)cursor->node->node_offset,
1162                                 node->count);
1163                 }
1164
1165                 /*
1166                  * Try to shortcut the search before dropping into the
1167                  * linear loop.  Locate the first node where r <= 1.
1168                  */
1169                 i = hammer_btree_search_node(&cursor->key_beg, node);
1170                 while (i <= node->count) {
1171                         ++hammer_stats_btree_elements;
1172                         elm = &node->elms[i];
1173                         r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1174                         if (hammer_debug_btree > 2) {
1175                                 kprintf(" IELM %p %d r=%d\n",
1176                                         &node->elms[i], i, r);
1177                         }
1178                         if (r < 0)
1179                                 break;
1180                         if (r == 1) {
1181                                 KKASSERT(elm->base.create_tid != 1);
1182                                 cursor->create_check = elm->base.create_tid - 1;
1183                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1184                         }
1185                         ++i;
1186                 }
1187                 if (hammer_debug_btree) {
1188                         kprintf("SEARCH-I preI=%d/%d r=%d\n",
1189                                 i, node->count, r);
1190                 }
1191
1192                 /*
1193                  * These cases occur when the parent's idea of the boundary
1194                  * is wider then the child's idea of the boundary, and
1195                  * require special handling.  If not inserting we can
1196                  * terminate the search early for these cases but the
1197                  * child's boundaries cannot be unconditionally modified.
1198                  */
1199                 if (i == 0) {
1200                         /*
1201                          * If i == 0 the search terminated to the LEFT of the
1202                          * left_boundary but to the RIGHT of the parent's left
1203                          * boundary.
1204                          */
1205                         u_int8_t save;
1206
1207                         elm = &node->elms[0];
1208
1209                         /*
1210                          * If we aren't inserting we can stop here.
1211                          */
1212                         if ((flags & (HAMMER_CURSOR_INSERT |
1213                                       HAMMER_CURSOR_PRUNING)) == 0) {
1214                                 cursor->index = 0;
1215                                 return(ENOENT);
1216                         }
1217
1218                         /*
1219                          * Correct a left-hand boundary mismatch.
1220                          *
1221                          * We can only do this if we can upgrade the lock,
1222                          * and synchronized as a background cursor (i.e.
1223                          * inserting or pruning).
1224                          *
1225                          * WARNING: We can only do this if inserting, i.e.
1226                          * we are running on the backend.
1227                          */
1228                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1229                                 return(error);
1230                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1231                         hammer_modify_node_field(cursor->trans, cursor->node,
1232                                                  elms[0]);
1233                         save = node->elms[0].base.btype;
1234                         node->elms[0].base = *cursor->left_bound;
1235                         node->elms[0].base.btype = save;
1236                         hammer_modify_node_done(cursor->node);
1237                 } else if (i == node->count + 1) {
1238                         /*
1239                          * If i == node->count + 1 the search terminated to
1240                          * the RIGHT of the right boundary but to the LEFT
1241                          * of the parent's right boundary.  If we aren't
1242                          * inserting we can stop here.
1243                          *
1244                          * Note that the last element in this case is
1245                          * elms[i-2] prior to adjustments to 'i'.
1246                          */
1247                         --i;
1248                         if ((flags & (HAMMER_CURSOR_INSERT |
1249                                       HAMMER_CURSOR_PRUNING)) == 0) {
1250                                 cursor->index = i;
1251                                 return (ENOENT);
1252                         }
1253
1254                         /*
1255                          * Correct a right-hand boundary mismatch.
1256                          * (actual push-down record is i-2 prior to
1257                          * adjustments to i).
1258                          *
1259                          * We can only do this if we can upgrade the lock,
1260                          * and synchronized as a background cursor (i.e.
1261                          * inserting or pruning).
1262                          *
1263                          * WARNING: We can only do this if inserting, i.e.
1264                          * we are running on the backend.
1265                          */
1266                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1267                                 return(error);
1268                         elm = &node->elms[i];
1269                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1270                         hammer_modify_node(cursor->trans, cursor->node,
1271                                            &elm->base, sizeof(elm->base));
1272                         elm->base = *cursor->right_bound;
1273                         hammer_modify_node_done(cursor->node);
1274                         --i;
1275                 } else {
1276                         /*
1277                          * The push-down index is now i - 1.  If we had
1278                          * terminated on the right boundary this will point
1279                          * us at the last element.
1280                          */
1281                         --i;
1282                 }
1283                 cursor->index = i;
1284                 elm = &node->elms[i];
1285
1286                 if (hammer_debug_btree) {
1287                         kprintf("RESULT-I %016llx[%d] %016llx %02x "
1288                                 "key=%016llx cre=%016llx lo=%02x\n",
1289                                 (long long)cursor->node->node_offset,
1290                                 i,
1291                                 (long long)elm->internal.base.obj_id,
1292                                 elm->internal.base.rec_type,
1293                                 (long long)elm->internal.base.key,
1294                                 (long long)elm->internal.base.create_tid,
1295                                 elm->internal.base.localization
1296                         );
1297                 }
1298
1299                 /*
1300                  * We better have a valid subtree offset.
1301                  */
1302                 KKASSERT(elm->internal.subtree_offset != 0);
1303
1304                 /*
1305                  * Handle insertion and deletion requirements.
1306                  *
1307                  * If inserting split full nodes.  The split code will
1308                  * adjust cursor->node and cursor->index if the current
1309                  * index winds up in the new node.
1310                  *
1311                  * If inserting and a left or right edge case was detected,
1312                  * we cannot correct the left or right boundary and must
1313                  * prepend and append an empty leaf node in order to make
1314                  * the boundary correction.
1315                  *
1316                  * If we run out of space we set enospc but continue on
1317                  * to a leaf.
1318                  */
1319                 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1320                         if (btree_node_is_full(node)) {
1321                                 error = btree_split_internal(cursor);
1322                                 if (error) {
1323                                         if (error != ENOSPC)
1324                                                 goto done;
1325                                         enospc = 1;
1326                                 }
1327                                 /*
1328                                  * reload stale pointers
1329                                  */
1330                                 i = cursor->index;
1331                                 node = cursor->node->ondisk;
1332                         }
1333                 }
1334
1335                 /*
1336                  * Push down (push into new node, existing node becomes
1337                  * the parent) and continue the search.
1338                  */
1339                 error = hammer_cursor_down(cursor);
1340                 /* node may have become stale */
1341                 if (error)
1342                         goto done;
1343                 node = cursor->node->ondisk;
1344         }
1345
1346         /*
1347          * We are at a leaf, do a linear search of the key array.
1348          *
1349          * On success the index is set to the matching element and 0
1350          * is returned.
1351          *
1352          * On failure the index is set to the insertion point and ENOENT
1353          * is returned.
1354          *
1355          * Boundaries are not stored in leaf nodes, so the index can wind
1356          * up to the left of element 0 (index == 0) or past the end of
1357          * the array (index == node->count).  It is also possible that the
1358          * leaf might be empty.
1359          */
1360         ++hammer_stats_btree_iterations;
1361         KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1362         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1363         if (hammer_debug_btree) {
1364                 kprintf("SEARCH-L %016llx count=%d\n",
1365                         (long long)cursor->node->node_offset,
1366                         node->count);
1367         }
1368
1369         /*
1370          * Try to shortcut the search before dropping into the
1371          * linear loop.  Locate the first node where r <= 1.
1372          */
1373         i = hammer_btree_search_node(&cursor->key_beg, node);
1374         while (i < node->count) {
1375                 ++hammer_stats_btree_elements;
1376                 elm = &node->elms[i];
1377
1378                 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1379
1380                 if (hammer_debug_btree > 1)
1381                         kprintf("  ELM %p %d r=%d\n", &node->elms[i], i, r);
1382
1383                 /*
1384                  * We are at a record element.  Stop if we've flipped past
1385                  * key_beg, not counting the create_tid test.  Allow the
1386                  * r == 1 case (key_beg > element but differs only by its
1387                  * create_tid) to fall through to the AS-OF check.
1388                  */
1389                 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1390
1391                 if (r < 0)
1392                         goto failed;
1393                 if (r > 1) {
1394                         ++i;
1395                         continue;
1396                 }
1397
1398                 /*
1399                  * Check our as-of timestamp against the element.
1400                  */
1401                 if (flags & HAMMER_CURSOR_ASOF) {
1402                         if (hammer_btree_chkts(cursor->asof,
1403                                                &node->elms[i].base) != 0) {
1404                                 ++i;
1405                                 continue;
1406                         }
1407                         /* success */
1408                 } else {
1409                         if (r > 0) {    /* can only be +1 */
1410                                 ++i;
1411                                 continue;
1412                         }
1413                         /* success */
1414                 }
1415                 cursor->index = i;
1416                 error = 0;
1417                 if (hammer_debug_btree) {
1418                         kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1419                                 (long long)cursor->node->node_offset, i);
1420                 }
1421                 goto done;
1422         }
1423
1424         /*
1425          * The search of the leaf node failed.  i is the insertion point.
1426          */
1427 failed:
1428         if (hammer_debug_btree) {
1429                 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1430                         (long long)cursor->node->node_offset, i);
1431         }
1432
1433         /*
1434          * No exact match was found, i is now at the insertion point.
1435          *
1436          * If inserting split a full leaf before returning.  This
1437          * may have the side effect of adjusting cursor->node and
1438          * cursor->index.
1439          */
1440         cursor->index = i;
1441         if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1442              btree_node_is_full(node)) {
1443                 error = btree_split_leaf(cursor);
1444                 if (error) {
1445                         if (error != ENOSPC)
1446                                 goto done;
1447                         enospc = 1;
1448                 }
1449                 /*
1450                  * reload stale pointers
1451                  */
1452                 /* NOT USED
1453                 i = cursor->index;
1454                 node = &cursor->node->internal;
1455                 */
1456         }
1457
1458         /*
1459          * We reached a leaf but did not find the key we were looking for.
1460          * If this is an insert we will be properly positioned for an insert
1461          * (ENOENT) or unable to insert (ENOSPC).
1462          */
1463         error = enospc ? ENOSPC : ENOENT;
1464 done:
1465         return(error);
1466 }
1467
1468 /*
1469  * Heuristical search for the first element whos comparison is <= 1.  May
1470  * return an index whos compare result is > 1 but may only return an index
1471  * whos compare result is <= 1 if it is the first element with that result.
1472  */
1473 int
1474 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1475 {
1476         int b;
1477         int s;
1478         int i;
1479         int r;
1480
1481         /*
1482          * Don't bother if the node does not have very many elements
1483          */
1484         b = 0;
1485         s = node->count;
1486         while (s - b > 4) {
1487                 i = b + (s - b) / 2;
1488                 ++hammer_stats_btree_elements;
1489                 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1490                 if (r <= 1) {
1491                         s = i;
1492                 } else {
1493                         b = i;
1494                 }
1495         }
1496         return(b);
1497 }
1498
1499
1500 /************************************************************************
1501  *                         SPLITTING AND MERGING                        *
1502  ************************************************************************
1503  *
1504  * These routines do all the dirty work required to split and merge nodes.
1505  */
1506
1507 /*
1508  * Split an internal node into two nodes and move the separator at the split
1509  * point to the parent.
1510  *
1511  * (cursor->node, cursor->index) indicates the element the caller intends
1512  * to push into.  We will adjust node and index if that element winds
1513  * up in the split node.
1514  *
1515  * If we are at the root of the filesystem a new root must be created with
1516  * two elements, one pointing to the original root and one pointing to the
1517  * newly allocated split node.
1518  */
1519 static
1520 int
1521 btree_split_internal(hammer_cursor_t cursor)
1522 {
1523         hammer_node_ondisk_t ondisk;
1524         hammer_node_t node;
1525         hammer_node_t parent;
1526         hammer_node_t new_node;
1527         hammer_btree_elm_t elm;
1528         hammer_btree_elm_t parent_elm;
1529         struct hammer_node_lock lockroot;
1530         hammer_mount_t hmp = cursor->trans->hmp;
1531         int parent_index;
1532         int made_root;
1533         int split;
1534         int error;
1535         int i;
1536         const int esize = sizeof(*elm);
1537
1538         hammer_node_lock_init(&lockroot, cursor->node);
1539         error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
1540         if (error)
1541                 goto done;
1542         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1543                 goto done;
1544         ++hammer_stats_btree_splits;
1545
1546         /*
1547          * Calculate the split point.  If the insertion point is at the
1548          * end of the leaf we adjust the split point significantly to the
1549          * right to try to optimize node fill and flag it.  If we hit
1550          * that same leaf again our heuristic failed and we don't try
1551          * to optimize node fill (it could lead to a degenerate case).
1552          */
1553         node = cursor->node;
1554         ondisk = node->ondisk;
1555         KKASSERT(ondisk->count > 4);
1556         if (cursor->index == ondisk->count &&
1557             (node->flags & HAMMER_NODE_NONLINEAR) == 0) {
1558                 split = (ondisk->count + 1) * 3 / 4;
1559                 node->flags |= HAMMER_NODE_NONLINEAR;
1560         } else {
1561                 /*
1562                  * We are splitting but elms[split] will be promoted to
1563                  * the parent, leaving the right hand node with one less
1564                  * element.  If the insertion point will be on the
1565                  * left-hand side adjust the split point to give the
1566                  * right hand side one additional node.
1567                  */
1568                 split = (ondisk->count + 1) / 2;
1569                 if (cursor->index <= split)
1570                         --split;
1571         }
1572
1573         /*
1574          * If we are at the root of the filesystem, create a new root node
1575          * with 1 element and split normally.  Avoid making major
1576          * modifications until we know the whole operation will work.
1577          */
1578         if (ondisk->parent == 0) {
1579                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1580                 if (parent == NULL)
1581                         goto done;
1582                 hammer_lock_ex(&parent->lock);
1583                 hammer_modify_node_noundo(cursor->trans, parent);
1584                 ondisk = parent->ondisk;
1585                 ondisk->count = 1;
1586                 ondisk->parent = 0;
1587                 ondisk->mirror_tid = node->ondisk->mirror_tid;
1588                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1589                 ondisk->elms[0].base = hmp->root_btree_beg;
1590                 ondisk->elms[0].base.btype = node->ondisk->type;
1591                 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1592                 ondisk->elms[1].base = hmp->root_btree_end;
1593                 hammer_modify_node_done(parent);
1594                 /* ondisk->elms[1].base.btype - not used */
1595                 made_root = 1;
1596                 parent_index = 0;       /* index of current node in parent */
1597         } else {
1598                 made_root = 0;
1599                 parent = cursor->parent;
1600                 parent_index = cursor->parent_index;
1601         }
1602
1603         /*
1604          * Split node into new_node at the split point.
1605          *
1606          *  B O O O P N N B     <-- P = node->elms[split] (index 4)
1607          *   0 1 2 3 4 5 6      <-- subtree indices
1608          *
1609          *       x x P x x
1610          *        s S S s
1611          *         /   \
1612          *  B O O O B    B N N B        <--- inner boundary points are 'P'
1613          *   0 1 2 3      4 5 6
1614          */
1615         new_node = hammer_alloc_btree(cursor->trans, 0, &error);
1616         if (new_node == NULL) {
1617                 if (made_root) {
1618                         hammer_unlock(&parent->lock);
1619                         hammer_delete_node(cursor->trans, parent);
1620                         hammer_rel_node(parent);
1621                 }
1622                 goto done;
1623         }
1624         hammer_lock_ex(&new_node->lock);
1625
1626         /*
1627          * Create the new node.  P becomes the left-hand boundary in the
1628          * new node.  Copy the right-hand boundary as well.
1629          *
1630          * elm is the new separator.
1631          */
1632         hammer_modify_node_noundo(cursor->trans, new_node);
1633         hammer_modify_node_all(cursor->trans, node);
1634         ondisk = node->ondisk;
1635         elm = &ondisk->elms[split];
1636         bcopy(elm, &new_node->ondisk->elms[0],
1637               (ondisk->count - split + 1) * esize);
1638         new_node->ondisk->count = ondisk->count - split;
1639         new_node->ondisk->parent = parent->node_offset;
1640         new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1641         new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1642         KKASSERT(ondisk->type == new_node->ondisk->type);
1643         hammer_cursor_split_node(node, new_node, split);
1644
1645         /*
1646          * Cleanup the original node.  Elm (P) becomes the new boundary,
1647          * its subtree_offset was moved to the new node.  If we had created
1648          * a new root its parent pointer may have changed.
1649          */
1650         elm->internal.subtree_offset = 0;
1651         ondisk->count = split;
1652
1653         /*
1654          * Insert the separator into the parent, fixup the parent's
1655          * reference to the original node, and reference the new node.
1656          * The separator is P.
1657          *
1658          * Remember that base.count does not include the right-hand boundary.
1659          */
1660         hammer_modify_node_all(cursor->trans, parent);
1661         ondisk = parent->ondisk;
1662         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1663         parent_elm = &ondisk->elms[parent_index+1];
1664         bcopy(parent_elm, parent_elm + 1,
1665               (ondisk->count - parent_index) * esize);
1666         parent_elm->internal.base = elm->base;  /* separator P */
1667         parent_elm->internal.base.btype = new_node->ondisk->type;
1668         parent_elm->internal.subtree_offset = new_node->node_offset;
1669         parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1670         ++ondisk->count;
1671         hammer_modify_node_done(parent);
1672         hammer_cursor_inserted_element(parent, parent_index + 1);
1673
1674         /*
1675          * The children of new_node need their parent pointer set to new_node.
1676          * The children have already been locked by
1677          * hammer_btree_lock_children().
1678          */
1679         for (i = 0; i < new_node->ondisk->count; ++i) {
1680                 elm = &new_node->ondisk->elms[i];
1681                 error = btree_set_parent(cursor->trans, new_node, elm);
1682                 if (error) {
1683                         panic("btree_split_internal: btree-fixup problem");
1684                 }
1685         }
1686         hammer_modify_node_done(new_node);
1687
1688         /*
1689          * The filesystem's root B-Tree pointer may have to be updated.
1690          */
1691         if (made_root) {
1692                 hammer_volume_t volume;
1693
1694                 volume = hammer_get_root_volume(hmp, &error);
1695                 KKASSERT(error == 0);
1696
1697                 hammer_modify_volume_field(cursor->trans, volume,
1698                                            vol0_btree_root);
1699                 volume->ondisk->vol0_btree_root = parent->node_offset;
1700                 hammer_modify_volume_done(volume);
1701                 node->ondisk->parent = parent->node_offset;
1702                 if (cursor->parent) {
1703                         hammer_unlock(&cursor->parent->lock);
1704                         hammer_rel_node(cursor->parent);
1705                 }
1706                 cursor->parent = parent;        /* lock'd and ref'd */
1707                 hammer_rel_volume(volume, 0);
1708         }
1709         hammer_modify_node_done(node);
1710
1711         /*
1712          * Ok, now adjust the cursor depending on which element the original
1713          * index was pointing at.  If we are >= the split point the push node
1714          * is now in the new node.
1715          *
1716          * NOTE: If we are at the split point itself we cannot stay with the
1717          * original node because the push index will point at the right-hand
1718          * boundary, which is illegal.
1719          *
1720          * NOTE: The cursor's parent or parent_index must be adjusted for
1721          * the case where a new parent (new root) was created, and the case
1722          * where the cursor is now pointing at the split node.
1723          */
1724         if (cursor->index >= split) {
1725                 cursor->parent_index = parent_index + 1;
1726                 cursor->index -= split;
1727                 hammer_unlock(&cursor->node->lock);
1728                 hammer_rel_node(cursor->node);
1729                 cursor->node = new_node;        /* locked and ref'd */
1730         } else {
1731                 cursor->parent_index = parent_index;
1732                 hammer_unlock(&new_node->lock);
1733                 hammer_rel_node(new_node);
1734         }
1735
1736         /*
1737          * Fixup left and right bounds
1738          */
1739         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1740         cursor->left_bound = &parent_elm[0].internal.base;
1741         cursor->right_bound = &parent_elm[1].internal.base;
1742         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1743                  &cursor->node->ondisk->elms[0].internal.base) <= 0);
1744         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1745                  &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1746
1747 done:
1748         hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
1749         hammer_cursor_downgrade(cursor);
1750         return (error);
1751 }
1752
1753 /*
1754  * Same as the above, but splits a full leaf node.
1755  *
1756  * This function
1757  */
1758 static
1759 int
1760 btree_split_leaf(hammer_cursor_t cursor)
1761 {
1762         hammer_node_ondisk_t ondisk;
1763         hammer_node_t parent;
1764         hammer_node_t leaf;
1765         hammer_mount_t hmp;
1766         hammer_node_t new_leaf;
1767         hammer_btree_elm_t elm;
1768         hammer_btree_elm_t parent_elm;
1769         hammer_base_elm_t mid_boundary;
1770         int parent_index;
1771         int made_root;
1772         int split;
1773         int error;
1774         const size_t esize = sizeof(*elm);
1775
1776         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1777                 return(error);
1778         ++hammer_stats_btree_splits;
1779
1780         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1781                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1782         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1783                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1784
1785         /*
1786          * Calculate the split point.  If the insertion point is at the
1787          * end of the leaf we adjust the split point significantly to the
1788          * right to try to optimize node fill and flag it.  If we hit
1789          * that same leaf again our heuristic failed and we don't try
1790          * to optimize node fill (it could lead to a degenerate case).
1791          */
1792         leaf = cursor->node;
1793         ondisk = leaf->ondisk;
1794         KKASSERT(ondisk->count > 4);
1795         if (cursor->index == ondisk->count &&
1796             (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) {
1797                 split = (ondisk->count + 1) * 3 / 4;
1798                 leaf->flags |= HAMMER_NODE_NONLINEAR;
1799         } else {
1800                 split = (ondisk->count + 1) / 2;
1801         }
1802
1803 #if 0
1804         /*
1805          * If the insertion point is at the split point shift the
1806          * split point left so we don't have to worry about
1807          */
1808         if (cursor->index == split)
1809                 --split;
1810 #endif
1811         KKASSERT(split > 0 && split < ondisk->count);
1812
1813         error = 0;
1814         hmp = leaf->hmp;
1815
1816         elm = &ondisk->elms[split];
1817
1818         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1819         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1820         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1821         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1822
1823         /*
1824          * If we are at the root of the tree, create a new root node with
1825          * 1 element and split normally.  Avoid making major modifications
1826          * until we know the whole operation will work.
1827          */
1828         if (ondisk->parent == 0) {
1829                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1830                 if (parent == NULL)
1831                         goto done;
1832                 hammer_lock_ex(&parent->lock);
1833                 hammer_modify_node_noundo(cursor->trans, parent);
1834                 ondisk = parent->ondisk;
1835                 ondisk->count = 1;
1836                 ondisk->parent = 0;
1837                 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1838                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1839                 ondisk->elms[0].base = hmp->root_btree_beg;
1840                 ondisk->elms[0].base.btype = leaf->ondisk->type;
1841                 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1842                 ondisk->elms[1].base = hmp->root_btree_end;
1843                 /* ondisk->elms[1].base.btype = not used */
1844                 hammer_modify_node_done(parent);
1845                 made_root = 1;
1846                 parent_index = 0;       /* insertion point in parent */
1847         } else {
1848                 made_root = 0;
1849                 parent = cursor->parent;
1850                 parent_index = cursor->parent_index;
1851         }
1852
1853         /*
1854          * Split leaf into new_leaf at the split point.  Select a separator
1855          * value in-between the two leafs but with a bent towards the right
1856          * leaf since comparisons use an 'elm >= separator' inequality.
1857          *
1858          *  L L L L L L L L
1859          *
1860          *       x x P x x
1861          *        s S S s
1862          *         /   \
1863          *  L L L L     L L L L
1864          */
1865         new_leaf = hammer_alloc_btree(cursor->trans, 0, &error);
1866         if (new_leaf == NULL) {
1867                 if (made_root) {
1868                         hammer_unlock(&parent->lock);
1869                         hammer_delete_node(cursor->trans, parent);
1870                         hammer_rel_node(parent);
1871                 }
1872                 goto done;
1873         }
1874         hammer_lock_ex(&new_leaf->lock);
1875
1876         /*
1877          * Create the new node and copy the leaf elements from the split
1878          * point on to the new node.
1879          */
1880         hammer_modify_node_all(cursor->trans, leaf);
1881         hammer_modify_node_noundo(cursor->trans, new_leaf);
1882         ondisk = leaf->ondisk;
1883         elm = &ondisk->elms[split];
1884         bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1885         new_leaf->ondisk->count = ondisk->count - split;
1886         new_leaf->ondisk->parent = parent->node_offset;
1887         new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1888         new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1889         KKASSERT(ondisk->type == new_leaf->ondisk->type);
1890         hammer_modify_node_done(new_leaf);
1891         hammer_cursor_split_node(leaf, new_leaf, split);
1892
1893         /*
1894          * Cleanup the original node.  Because this is a leaf node and
1895          * leaf nodes do not have a right-hand boundary, there
1896          * aren't any special edge cases to clean up.  We just fixup the
1897          * count.
1898          */
1899         ondisk->count = split;
1900
1901         /*
1902          * Insert the separator into the parent, fixup the parent's
1903          * reference to the original node, and reference the new node.
1904          * The separator is P.
1905          *
1906          * Remember that base.count does not include the right-hand boundary.
1907          * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1908          */
1909         hammer_modify_node_all(cursor->trans, parent);
1910         ondisk = parent->ondisk;
1911         KKASSERT(split != 0);
1912         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1913         parent_elm = &ondisk->elms[parent_index+1];
1914         bcopy(parent_elm, parent_elm + 1,
1915               (ondisk->count - parent_index) * esize);
1916
1917         hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1918         parent_elm->internal.base.btype = new_leaf->ondisk->type;
1919         parent_elm->internal.subtree_offset = new_leaf->node_offset;
1920         parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1921         mid_boundary = &parent_elm->base;
1922         ++ondisk->count;
1923         hammer_modify_node_done(parent);
1924         hammer_cursor_inserted_element(parent, parent_index + 1);
1925
1926         /*
1927          * The filesystem's root B-Tree pointer may have to be updated.
1928          */
1929         if (made_root) {
1930                 hammer_volume_t volume;
1931
1932                 volume = hammer_get_root_volume(hmp, &error);
1933                 KKASSERT(error == 0);
1934
1935                 hammer_modify_volume_field(cursor->trans, volume,
1936                                            vol0_btree_root);
1937                 volume->ondisk->vol0_btree_root = parent->node_offset;
1938                 hammer_modify_volume_done(volume);
1939                 leaf->ondisk->parent = parent->node_offset;
1940                 if (cursor->parent) {
1941                         hammer_unlock(&cursor->parent->lock);
1942                         hammer_rel_node(cursor->parent);
1943                 }
1944                 cursor->parent = parent;        /* lock'd and ref'd */
1945                 hammer_rel_volume(volume, 0);
1946         }
1947         hammer_modify_node_done(leaf);
1948
1949         /*
1950          * Ok, now adjust the cursor depending on which element the original
1951          * index was pointing at.  If we are >= the split point the push node
1952          * is now in the new node.
1953          *
1954          * NOTE: If we are at the split point itself we need to select the
1955          * old or new node based on where key_beg's insertion point will be.
1956          * If we pick the wrong side the inserted element will wind up in
1957          * the wrong leaf node and outside that node's bounds.
1958          */
1959         if (cursor->index > split ||
1960             (cursor->index == split &&
1961              hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1962                 cursor->parent_index = parent_index + 1;
1963                 cursor->index -= split;
1964                 hammer_unlock(&cursor->node->lock);
1965                 hammer_rel_node(cursor->node);
1966                 cursor->node = new_leaf;
1967         } else {
1968                 cursor->parent_index = parent_index;
1969                 hammer_unlock(&new_leaf->lock);
1970                 hammer_rel_node(new_leaf);
1971         }
1972
1973         /*
1974          * Fixup left and right bounds
1975          */
1976         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1977         cursor->left_bound = &parent_elm[0].internal.base;
1978         cursor->right_bound = &parent_elm[1].internal.base;
1979
1980         /*
1981          * Assert that the bounds are correct.
1982          */
1983         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1984                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1985         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1986                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1987         KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1988         KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1989
1990 done:
1991         hammer_cursor_downgrade(cursor);
1992         return (error);
1993 }
1994
1995 #if 0
1996
1997 /*
1998  * Recursively correct the right-hand boundary's create_tid to (tid) as
1999  * long as the rest of the key matches.  We have to recurse upward in
2000  * the tree as well as down the left side of each parent's right node.
2001  *
2002  * Return EDEADLK if we were only partially successful, forcing the caller
2003  * to try again.  The original cursor is not modified.  This routine can
2004  * also fail with EDEADLK if it is forced to throw away a portion of its
2005  * record history.
2006  *
2007  * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2008  */
2009 struct hammer_rhb {
2010         TAILQ_ENTRY(hammer_rhb) entry;
2011         hammer_node_t   node;
2012         int             index;
2013 };
2014
2015 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2016
2017 int
2018 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2019 {
2020         struct hammer_mount *hmp;
2021         struct hammer_rhb_list rhb_list;
2022         hammer_base_elm_t elm;
2023         hammer_node_t orig_node;
2024         struct hammer_rhb *rhb;
2025         int orig_index;
2026         int error;
2027
2028         TAILQ_INIT(&rhb_list);
2029         hmp = cursor->trans->hmp;
2030
2031         /*
2032          * Save our position so we can restore it on return.  This also
2033          * gives us a stable 'elm'.
2034          */
2035         orig_node = cursor->node;
2036         hammer_ref_node(orig_node);
2037         hammer_lock_sh(&orig_node->lock);
2038         orig_index = cursor->index;
2039         elm = &orig_node->ondisk->elms[orig_index].base;
2040
2041         /*
2042          * Now build a list of parents going up, allocating a rhb
2043          * structure for each one.
2044          */
2045         while (cursor->parent) {
2046                 /*
2047                  * Stop if we no longer have any right-bounds to fix up
2048                  */
2049                 if (elm->obj_id != cursor->right_bound->obj_id ||
2050                     elm->rec_type != cursor->right_bound->rec_type ||
2051                     elm->key != cursor->right_bound->key) {
2052                         break;
2053                 }
2054
2055                 /*
2056                  * Stop if the right-hand bound's create_tid does not
2057                  * need to be corrected.
2058                  */
2059                 if (cursor->right_bound->create_tid >= tid)
2060                         break;
2061
2062                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2063                 rhb->node = cursor->parent;
2064                 rhb->index = cursor->parent_index;
2065                 hammer_ref_node(rhb->node);
2066                 hammer_lock_sh(&rhb->node->lock);
2067                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2068
2069                 hammer_cursor_up(cursor);
2070         }
2071
2072         /*
2073          * now safely adjust the right hand bound for each rhb.  This may
2074          * also require taking the right side of the tree and iterating down
2075          * ITS left side.
2076          */
2077         error = 0;
2078         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2079                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2080                 if (error)
2081                         break;
2082                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2083                 hammer_unlock(&rhb->node->lock);
2084                 hammer_rel_node(rhb->node);
2085                 kfree(rhb, hmp->m_misc);
2086
2087                 switch (cursor->node->ondisk->type) {
2088                 case HAMMER_BTREE_TYPE_INTERNAL:
2089                         /*
2090                          * Right-boundary for parent at internal node
2091                          * is one element to the right of the element whos
2092                          * right boundary needs adjusting.  We must then
2093                          * traverse down the left side correcting any left
2094                          * bounds (which may now be too far to the left).
2095                          */
2096                         ++cursor->index;
2097                         error = hammer_btree_correct_lhb(cursor, tid);
2098                         break;
2099                 default:
2100                         panic("hammer_btree_correct_rhb(): Bad node type");
2101                         error = EINVAL;
2102                         break;
2103                 }
2104         }
2105
2106         /*
2107          * Cleanup
2108          */
2109         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2110                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2111                 hammer_unlock(&rhb->node->lock);
2112                 hammer_rel_node(rhb->node);
2113                 kfree(rhb, hmp->m_misc);
2114         }
2115         error = hammer_cursor_seek(cursor, orig_node, orig_index);
2116         hammer_unlock(&orig_node->lock);
2117         hammer_rel_node(orig_node);
2118         return (error);
2119 }
2120
2121 /*
2122  * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2123  * bound going downward starting at the current cursor position.
2124  *
2125  * This function does not restore the cursor after use.
2126  */
2127 int
2128 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2129 {
2130         struct hammer_rhb_list rhb_list;
2131         hammer_base_elm_t elm;
2132         hammer_base_elm_t cmp;
2133         struct hammer_rhb *rhb;
2134         struct hammer_mount *hmp;
2135         int error;
2136
2137         TAILQ_INIT(&rhb_list);
2138         hmp = cursor->trans->hmp;
2139
2140         cmp = &cursor->node->ondisk->elms[cursor->index].base;
2141
2142         /*
2143          * Record the node and traverse down the left-hand side for all
2144          * matching records needing a boundary correction.
2145          */
2146         error = 0;
2147         for (;;) {
2148                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2149                 rhb->node = cursor->node;
2150                 rhb->index = cursor->index;
2151                 hammer_ref_node(rhb->node);
2152                 hammer_lock_sh(&rhb->node->lock);
2153                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2154
2155                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2156                         /*
2157                          * Nothing to traverse down if we are at the right
2158                          * boundary of an internal node.
2159                          */
2160                         if (cursor->index == cursor->node->ondisk->count)
2161                                 break;
2162                 } else {
2163                         elm = &cursor->node->ondisk->elms[cursor->index].base;
2164                         if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2165                                 break;
2166                         panic("Illegal leaf record type %02x", elm->btype);
2167                 }
2168                 error = hammer_cursor_down(cursor);
2169                 if (error)
2170                         break;
2171
2172                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2173                 if (elm->obj_id != cmp->obj_id ||
2174                     elm->rec_type != cmp->rec_type ||
2175                     elm->key != cmp->key) {
2176                         break;
2177                 }
2178                 if (elm->create_tid >= tid)
2179                         break;
2180
2181         }
2182
2183         /*
2184          * Now we can safely adjust the left-hand boundary from the bottom-up.
2185          * The last element we remove from the list is the caller's right hand
2186          * boundary, which must also be adjusted.
2187          */
2188         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2189                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2190                 if (error)
2191                         break;
2192                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2193                 hammer_unlock(&rhb->node->lock);
2194                 hammer_rel_node(rhb->node);
2195                 kfree(rhb, hmp->m_misc);
2196
2197                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2198                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2199                         hammer_modify_node(cursor->trans, cursor->node,
2200                                            &elm->create_tid,
2201                                            sizeof(elm->create_tid));
2202                         elm->create_tid = tid;
2203                         hammer_modify_node_done(cursor->node);
2204                 } else {
2205                         panic("hammer_btree_correct_lhb(): Bad element type");
2206                 }
2207         }
2208
2209         /*
2210          * Cleanup
2211          */
2212         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2213                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2214                 hammer_unlock(&rhb->node->lock);
2215                 hammer_rel_node(rhb->node);
2216                 kfree(rhb, hmp->m_misc);
2217         }
2218         return (error);
2219 }
2220
2221 #endif
2222
2223 /*
2224  * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2225  * (cursor->node).  Returns 0 on success, EDEADLK if we could not complete
2226  * the operation due to a deadlock, or some other error.
2227  *
2228  * This routine is initially called with an empty leaf and may be
2229  * recursively called with single-element internal nodes.
2230  *
2231  * It should also be noted that when removing empty leaves we must be sure
2232  * to test and update mirror_tid because another thread may have deadlocked
2233  * against us (or someone) trying to propagate it up and cannot retry once
2234  * the node has been deleted.
2235  *
2236  * On return the cursor may end up pointing to an internal node, suitable
2237  * for further iteration but not for an immediate insertion or deletion.
2238  */
2239 static int
2240 btree_remove(hammer_cursor_t cursor)
2241 {
2242         hammer_node_ondisk_t ondisk;
2243         hammer_btree_elm_t elm;
2244         hammer_node_t node;
2245         hammer_node_t parent;
2246         const int esize = sizeof(*elm);
2247         int error;
2248
2249         node = cursor->node;
2250
2251         /*
2252          * When deleting the root of the filesystem convert it to
2253          * an empty leaf node.  Internal nodes cannot be empty.
2254          */
2255         ondisk = node->ondisk;
2256         if (ondisk->parent == 0) {
2257                 KKASSERT(cursor->parent == NULL);
2258                 hammer_modify_node_all(cursor->trans, node);
2259                 KKASSERT(ondisk == node->ondisk);
2260                 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2261                 ondisk->count = 0;
2262                 hammer_modify_node_done(node);
2263                 cursor->index = 0;
2264                 return(0);
2265         }
2266
2267         parent = cursor->parent;
2268
2269         /*
2270          * Attempt to remove the parent's reference to the child.  If the
2271          * parent would become empty we have to recurse.  If we fail we
2272          * leave the parent pointing to an empty leaf node.
2273          *
2274          * We have to recurse successfully before we can delete the internal
2275          * node as it is illegal to have empty internal nodes.  Even though
2276          * the operation may be aborted we must still fixup any unlocked
2277          * cursors as if we had deleted the element prior to recursing
2278          * (by calling hammer_cursor_deleted_element()) so those cursors
2279          * are properly forced up the chain by the recursion.
2280          */
2281         if (parent->ondisk->count == 1) {
2282                 /*
2283                  * This special cursor_up_locked() call leaves the original
2284                  * node exclusively locked and referenced, leaves the
2285                  * original parent locked (as the new node), and locks the
2286                  * new parent.  It can return EDEADLK.
2287                  *
2288                  * We cannot call hammer_cursor_removed_node() until we are
2289                  * actually able to remove the node.  If we did then tracked
2290                  * cursors in the middle of iterations could be repointed
2291                  * to a parent node.  If this occurs they could end up
2292                  * scanning newly inserted records into the node (that could
2293                  * not be deleted) when they push down again.
2294                  *
2295                  * Due to the way the recursion works the final parent is left
2296                  * in cursor->parent after the recursion returns.  Each
2297                  * layer on the way back up is thus able to call
2298                  * hammer_cursor_removed_node() and 'jump' the node up to
2299                  * the (same) final parent.
2300                  *
2301                  * NOTE!  The local variable 'parent' is invalid after we
2302                  *        call hammer_cursor_up_locked().
2303                  */
2304                 error = hammer_cursor_up_locked(cursor);
2305                 parent = NULL;
2306
2307                 if (error == 0) {
2308                         hammer_cursor_deleted_element(cursor->node, 0);
2309                         error = btree_remove(cursor);
2310                         if (error == 0) {
2311                                 KKASSERT(node != cursor->node);
2312                                 hammer_cursor_removed_node(
2313                                         node, cursor->node,
2314                                         cursor->index);
2315                                 hammer_modify_node_all(cursor->trans, node);
2316                                 ondisk = node->ondisk;
2317                                 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2318                                 ondisk->count = 0;
2319                                 hammer_modify_node_done(node);
2320                                 hammer_flush_node(node, 0);
2321                                 hammer_delete_node(cursor->trans, node);
2322                         } else {
2323                                 /*
2324                                  * Defer parent removal because we could not
2325                                  * get the lock, just let the leaf remain
2326                                  * empty.
2327                                  */
2328                                 /**/
2329                         }
2330                         hammer_unlock(&node->lock);
2331                         hammer_rel_node(node);
2332                 } else {
2333                         /*
2334                          * Defer parent removal because we could not
2335                          * get the lock, just let the leaf remain
2336                          * empty.
2337                          */
2338                         /**/
2339                 }
2340         } else {
2341                 KKASSERT(parent->ondisk->count > 1);
2342
2343                 hammer_modify_node_all(cursor->trans, parent);
2344                 ondisk = parent->ondisk;
2345                 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2346
2347                 elm = &ondisk->elms[cursor->parent_index];
2348                 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2349                 KKASSERT(ondisk->count > 0);
2350
2351                 /*
2352                  * We must retain the highest mirror_tid.  The deleted
2353                  * range is now encompassed by the element to the left.
2354                  * If we are already at the left edge the new left edge
2355                  * inherits mirror_tid.
2356                  *
2357                  * Note that bounds of the parent to our parent may create
2358                  * a gap to the left of our left-most node or to the right
2359                  * of our right-most node.  The gap is silently included
2360                  * in the mirror_tid's area of effect from the point of view
2361                  * of the scan.
2362                  */
2363                 if (cursor->parent_index) {
2364                         if (elm[-1].internal.mirror_tid <
2365                             elm[0].internal.mirror_tid) {
2366                                 elm[-1].internal.mirror_tid =
2367                                     elm[0].internal.mirror_tid;
2368                         }
2369                 } else {
2370                         if (elm[1].internal.mirror_tid <
2371                             elm[0].internal.mirror_tid) {
2372                                 elm[1].internal.mirror_tid =
2373                                     elm[0].internal.mirror_tid;
2374                         }
2375                 }
2376
2377                 /*
2378                  * Delete the subtree reference in the parent.  Include
2379                  * boundary element at end.
2380                  */
2381                 bcopy(&elm[1], &elm[0],
2382                       (ondisk->count - cursor->parent_index) * esize);
2383                 --ondisk->count;
2384                 hammer_modify_node_done(parent);
2385                 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2386                 hammer_cursor_deleted_element(parent, cursor->parent_index);
2387                 hammer_flush_node(node, 0);
2388                 hammer_delete_node(cursor->trans, node);
2389
2390                 /*
2391                  * cursor->node is invalid, cursor up to make the cursor
2392                  * valid again.  We have to flag the condition in case
2393                  * another thread wiggles an insertion in during an
2394                  * iteration.
2395                  */
2396                 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK;
2397                 error = hammer_cursor_up(cursor);
2398         }
2399         return (error);
2400 }
2401
2402 /*
2403  * Propagate cursor->trans->tid up the B-Tree starting at the current
2404  * cursor position using pseudofs info gleaned from the passed inode.
2405  *
2406  * The passed inode has no relationship to the cursor position other
2407  * then being in the same pseudofs as the insertion or deletion we
2408  * are propagating the mirror_tid for.
2409  *
2410  * WARNING!  Because we push and pop the passed cursor, it may be
2411  *           modified by other B-Tree operations while it is unlocked
2412  *           and things like the node & leaf pointers, and indexes might
2413  *           change.
2414  */
2415 void
2416 hammer_btree_do_propagation(hammer_cursor_t cursor,
2417                             hammer_pseudofs_inmem_t pfsm,
2418                             hammer_btree_leaf_elm_t leaf)
2419 {
2420         hammer_cursor_t ncursor;
2421         hammer_tid_t mirror_tid;
2422         int error __debugvar;
2423
2424         /*
2425          * We do not propagate a mirror_tid if the filesystem was mounted
2426          * in no-mirror mode.
2427          */
2428         if (cursor->trans->hmp->master_id < 0)
2429                 return;
2430
2431         /*
2432          * This is a bit of a hack because we cannot deadlock or return
2433          * EDEADLK here.  The related operation has already completed and
2434          * we must propagate the mirror_tid now regardless.
2435          *
2436          * Generate a new cursor which inherits the original's locks and
2437          * unlock the original.  Use the new cursor to propagate the
2438          * mirror_tid.  Then clean up the new cursor and reacquire locks
2439          * on the original.
2440          *
2441          * hammer_dup_cursor() cannot dup locks.  The dup inherits the
2442          * original's locks and the original is tracked and must be
2443          * re-locked.
2444          */
2445         mirror_tid = cursor->node->ondisk->mirror_tid;
2446         KKASSERT(mirror_tid != 0);
2447         ncursor = hammer_push_cursor(cursor);
2448         error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2449         KKASSERT(error == 0);
2450         hammer_pop_cursor(cursor, ncursor);
2451         /* WARNING: cursor's leaf pointer may change after pop */
2452 }
2453
2454
2455 /*
2456  * Propagate a mirror TID update upwards through the B-Tree to the root.
2457  *
2458  * A locked internal node must be passed in.  The node will remain locked
2459  * on return.
2460  *
2461  * This function syncs mirror_tid at the specified internal node's element,
2462  * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2463  */
2464 static int
2465 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2466 {
2467         hammer_btree_internal_elm_t elm;
2468         hammer_node_t node;
2469         int error;
2470
2471         for (;;) {
2472                 error = hammer_cursor_up(cursor);
2473                 if (error == 0)
2474                         error = hammer_cursor_upgrade(cursor);
2475
2476                 /*
2477                  * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the
2478                  * cursor will still be properly positioned for
2479                  * mirror propagation, just not for iterations.
2480                  */
2481                 while (error == EDEADLK) {
2482                         hammer_recover_cursor(cursor);
2483                         error = hammer_cursor_upgrade(cursor);
2484                 }
2485                 if (error)
2486                         break;
2487
2488                 /*
2489                  * If the cursor deadlocked it could end up at a leaf
2490                  * after we lost the lock.
2491                  */
2492                 node = cursor->node;
2493                 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL)
2494                         continue;
2495
2496                 /*
2497                  * Adjust the node's element
2498                  */
2499                 elm = &node->ondisk->elms[cursor->index].internal;
2500                 if (elm->mirror_tid >= mirror_tid)
2501                         break;
2502                 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2503                                    sizeof(elm->mirror_tid));
2504                 elm->mirror_tid = mirror_tid;
2505                 hammer_modify_node_done(node);
2506                 if (hammer_debug_general & 0x0002) {
2507                         kprintf("mirror_propagate: propagate "
2508                                 "%016llx @%016llx:%d\n",
2509                                 (long long)mirror_tid,
2510                                 (long long)node->node_offset,
2511                                 cursor->index);
2512                 }
2513
2514
2515                 /*
2516                  * Adjust the node's mirror_tid aggregator
2517                  */
2518                 if (node->ondisk->mirror_tid >= mirror_tid)
2519                         return(0);
2520                 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2521                 node->ondisk->mirror_tid = mirror_tid;
2522                 hammer_modify_node_done(node);
2523                 if (hammer_debug_general & 0x0002) {
2524                         kprintf("mirror_propagate: propagate "
2525                                 "%016llx @%016llx\n",
2526                                 (long long)mirror_tid,
2527                                 (long long)node->node_offset);
2528                 }
2529         }
2530         if (error == ENOENT)
2531                 error = 0;
2532         return(error);
2533 }
2534
2535 hammer_node_t
2536 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2537                         int *parent_indexp, int *errorp, int try_exclusive)
2538 {
2539         hammer_node_t parent;
2540         hammer_btree_elm_t elm;
2541         int i;
2542
2543         /*
2544          * Get the node
2545          */
2546         parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2547         if (*errorp) {
2548                 KKASSERT(parent == NULL);
2549                 return(NULL);
2550         }
2551         KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2552
2553         /*
2554          * Lock the node
2555          */
2556         if (try_exclusive) {
2557                 if (hammer_lock_ex_try(&parent->lock)) {
2558                         hammer_rel_node(parent);
2559                         *errorp = EDEADLK;
2560                         return(NULL);
2561                 }
2562         } else {
2563                 hammer_lock_sh(&parent->lock);
2564         }
2565
2566         /*
2567          * Figure out which element in the parent is pointing to the
2568          * child.
2569          */
2570         if (node->ondisk->count) {
2571                 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2572                                              parent->ondisk);
2573         } else {
2574                 i = 0;
2575         }
2576         while (i < parent->ondisk->count) {
2577                 elm = &parent->ondisk->elms[i];
2578                 if (elm->internal.subtree_offset == node->node_offset)
2579                         break;
2580                 ++i;
2581         }
2582         if (i == parent->ondisk->count) {
2583                 hammer_unlock(&parent->lock);
2584                 panic("Bad B-Tree link: parent %p node %p", parent, node);
2585         }
2586         *parent_indexp = i;
2587         KKASSERT(*errorp == 0);
2588         return(parent);
2589 }
2590
2591 /*
2592  * The element (elm) has been moved to a new internal node (node).
2593  *
2594  * If the element represents a pointer to an internal node that node's
2595  * parent must be adjusted to the element's new location.
2596  *
2597  * XXX deadlock potential here with our exclusive locks
2598  */
2599 int
2600 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2601                  hammer_btree_elm_t elm)
2602 {
2603         hammer_node_t child;
2604         int error;
2605
2606         error = 0;
2607
2608         switch(elm->base.btype) {
2609         case HAMMER_BTREE_TYPE_INTERNAL:
2610         case HAMMER_BTREE_TYPE_LEAF:
2611                 child = hammer_get_node(trans, elm->internal.subtree_offset,
2612                                         0, &error);
2613                 if (error == 0) {
2614                         hammer_modify_node_field(trans, child, parent);
2615                         child->ondisk->parent = node->node_offset;
2616                         hammer_modify_node_done(child);
2617                         hammer_rel_node(child);
2618                 }
2619                 break;
2620         default:
2621                 break;
2622         }
2623         return(error);
2624 }
2625
2626 /*
2627  * Initialize the root of a recursive B-Tree node lock list structure.
2628  */
2629 void
2630 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2631 {
2632         TAILQ_INIT(&parent->list);
2633         parent->parent = NULL;
2634         parent->node = node;
2635         parent->index = -1;
2636         parent->count = node->ondisk->count;
2637         parent->copy = NULL;
2638         parent->flags = 0;
2639 }
2640
2641 /*
2642  * Initialize a cache of hammer_node_lock's including space allocated
2643  * for node copies.
2644  *
2645  * This is used by the rebalancing code to preallocate the copy space
2646  * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER
2647  * locks, otherwise we can blow out the pageout daemon's emergency
2648  * reserve and deadlock it.
2649  *
2650  * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache.
2651  *       The flag is set when the item is pulled off the cache for use.
2652  */
2653 void
2654 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
2655                          int depth)
2656 {
2657         hammer_node_lock_t item;
2658         int count;
2659
2660         for (count = 1; depth; --depth)
2661                 count *= HAMMER_BTREE_LEAF_ELMS;
2662         bzero(lcache, sizeof(*lcache));
2663         TAILQ_INIT(&lcache->list);
2664         while (count) {
2665                 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO);
2666                 item->copy = kmalloc(sizeof(*item->copy),
2667                                      hmp->m_misc, M_WAITOK);
2668                 TAILQ_INIT(&item->list);
2669                 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2670                 --count;
2671         }
2672 }
2673
2674 void
2675 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache)
2676 {
2677         hammer_node_lock_t item;
2678
2679         while ((item = TAILQ_FIRST(&lcache->list)) != NULL) {
2680                 TAILQ_REMOVE(&lcache->list, item, entry);
2681                 KKASSERT(item->copy);
2682                 KKASSERT(TAILQ_EMPTY(&item->list));
2683                 kfree(item->copy, hmp->m_misc);
2684                 kfree(item, hmp->m_misc);
2685         }
2686         KKASSERT(lcache->copy == NULL);
2687 }
2688
2689 /*
2690  * Exclusively lock all the children of node.  This is used by the split
2691  * code to prevent anyone from accessing the children of a cursor node
2692  * while we fix-up its parent offset.
2693  *
2694  * If we don't lock the children we can really mess up cursors which block
2695  * trying to cursor-up into our node.
2696  *
2697  * On failure EDEADLK (or some other error) is returned.  If a deadlock
2698  * error is returned the cursor is adjusted to block on termination.
2699  *
2700  * The caller is responsible for managing parent->node, the root's node
2701  * is usually aliased from a cursor.
2702  */
2703 int
2704 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2705                            hammer_node_lock_t parent,
2706                            hammer_node_lock_t lcache)
2707 {
2708         hammer_node_t node;
2709         hammer_node_lock_t item;
2710         hammer_node_ondisk_t ondisk;
2711         hammer_btree_elm_t elm;
2712         hammer_node_t child;
2713         struct hammer_mount *hmp;
2714         int error;
2715         int i;
2716
2717         node = parent->node;
2718         ondisk = node->ondisk;
2719         error = 0;
2720         hmp = cursor->trans->hmp;
2721
2722         /*
2723          * We really do not want to block on I/O with exclusive locks held,
2724          * pre-get the children before trying to lock the mess.  This is
2725          * only done one-level deep for now.
2726          */
2727         for (i = 0; i < ondisk->count; ++i) {
2728                 ++hammer_stats_btree_elements;
2729                 elm = &ondisk->elms[i];
2730                 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2731                     elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2732                         continue;
2733                 }
2734                 child = hammer_get_node(cursor->trans,
2735                                         elm->internal.subtree_offset,
2736                                         0, &error);
2737                 if (child)
2738                         hammer_rel_node(child);
2739         }
2740
2741         /*
2742          * Do it for real
2743          */
2744         for (i = 0; error == 0 && i < ondisk->count; ++i) {
2745                 ++hammer_stats_btree_elements;
2746                 elm = &ondisk->elms[i];
2747
2748                 switch(elm->base.btype) {
2749                 case HAMMER_BTREE_TYPE_INTERNAL:
2750                 case HAMMER_BTREE_TYPE_LEAF:
2751                         KKASSERT(elm->internal.subtree_offset != 0);
2752                         child = hammer_get_node(cursor->trans,
2753                                                 elm->internal.subtree_offset,
2754                                                 0, &error);
2755                         break;
2756                 default:
2757                         child = NULL;
2758                         break;
2759                 }
2760                 if (child) {
2761                         if (hammer_lock_ex_try(&child->lock) != 0) {
2762                                 if (cursor->deadlk_node == NULL) {
2763                                         cursor->deadlk_node = child;
2764                                         hammer_ref_node(cursor->deadlk_node);
2765                                 }
2766                                 error = EDEADLK;
2767                                 hammer_rel_node(child);
2768                         } else {
2769                                 if (lcache) {
2770                                         item = TAILQ_FIRST(&lcache->list);
2771                                         KKASSERT(item != NULL);
2772                                         item->flags |= HAMMER_NODE_LOCK_LCACHE;
2773                                         TAILQ_REMOVE(&lcache->list,
2774                                                      item, entry);
2775                                 } else {
2776                                         item = kmalloc(sizeof(*item),
2777                                                        hmp->m_misc,
2778                                                        M_WAITOK|M_ZERO);
2779                                         TAILQ_INIT(&item->list);
2780                                 }
2781
2782                                 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2783                                 item->parent = parent;
2784                                 item->node = child;
2785                                 item->index = i;
2786                                 item->count = child->ondisk->count;
2787
2788                                 /*
2789                                  * Recurse (used by the rebalancing code)
2790                                  */
2791                                 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2792                                         error = hammer_btree_lock_children(
2793                                                         cursor,
2794                                                         depth - 1,
2795                                                         item,
2796                                                         lcache);
2797                                 }
2798                         }
2799                 }
2800         }
2801         if (error)
2802                 hammer_btree_unlock_children(hmp, parent, lcache);
2803         return(error);
2804 }
2805
2806 /*
2807  * Create an in-memory copy of all B-Tree nodes listed, recursively,
2808  * including the parent.
2809  */
2810 void
2811 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2812 {
2813         hammer_mount_t hmp = cursor->trans->hmp;
2814         hammer_node_lock_t item;
2815
2816         if (parent->copy == NULL) {
2817                 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0);
2818                 parent->copy = kmalloc(sizeof(*parent->copy),
2819                                        hmp->m_misc, M_WAITOK);
2820         }
2821         KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0);
2822         *parent->copy = *parent->node->ondisk;
2823         TAILQ_FOREACH(item, &parent->list, entry) {
2824                 hammer_btree_lock_copy(cursor, item);
2825         }
2826 }
2827
2828 /*
2829  * Recursively sync modified copies to the media.
2830  */
2831 int
2832 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2833 {
2834         hammer_node_lock_t item;
2835         int count = 0;
2836
2837         if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2838                 ++count;
2839                 hammer_modify_node_all(cursor->trans, parent->node);
2840                 *parent->node->ondisk = *parent->copy;
2841                 hammer_modify_node_done(parent->node);
2842                 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2843                         hammer_flush_node(parent->node, 0);
2844                         hammer_delete_node(cursor->trans, parent->node);
2845                 }
2846         }
2847         TAILQ_FOREACH(item, &parent->list, entry) {
2848                 count += hammer_btree_sync_copy(cursor, item);
2849         }
2850         return(count);
2851 }
2852
2853 /*
2854  * Release previously obtained node locks.  The caller is responsible for
2855  * cleaning up parent->node itself (its usually just aliased from a cursor),
2856  * but this function will take care of the copies.
2857  *
2858  * NOTE: The root node is not placed in the lcache and node->copy is not
2859  *       deallocated when lcache != NULL.
2860  */
2861 void
2862 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent,
2863                              hammer_node_lock_t lcache)
2864 {
2865         hammer_node_lock_t item;
2866         hammer_node_ondisk_t copy;
2867
2868         while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2869                 TAILQ_REMOVE(&parent->list, item, entry);
2870                 hammer_btree_unlock_children(hmp, item, lcache);
2871                 hammer_unlock(&item->node->lock);
2872                 hammer_rel_node(item->node);
2873                 if (lcache) {
2874                         /*
2875                          * NOTE: When placing the item back in the lcache
2876                          *       the flag is cleared by the bzero().
2877                          *       Remaining fields are cleared as a safety
2878                          *       measure.
2879                          */
2880                         KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE);
2881                         KKASSERT(TAILQ_EMPTY(&item->list));
2882                         copy = item->copy;
2883                         bzero(item, sizeof(*item));
2884                         TAILQ_INIT(&item->list);
2885                         item->copy = copy;
2886                         if (copy)
2887                                 bzero(copy, sizeof(*copy));
2888                         TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2889                 } else {
2890                         kfree(item, hmp->m_misc);
2891                 }
2892         }
2893         if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) {
2894                 kfree(parent->copy, hmp->m_misc);
2895                 parent->copy = NULL;    /* safety */
2896         }
2897 }
2898
2899 /************************************************************************
2900  *                         MISCELLANIOUS SUPPORT                        *
2901  ************************************************************************/
2902
2903 /*
2904  * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2905  *
2906  * Note that for this particular function a return value of -1, 0, or +1
2907  * can denote a match if create_tid is otherwise discounted.  A create_tid
2908  * of zero is considered to be 'infinity' in comparisons.
2909  *
2910  * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2911  */
2912 int
2913 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2914 {
2915         if (key1->localization < key2->localization)
2916                 return(-5);
2917         if (key1->localization > key2->localization)
2918                 return(5);
2919
2920         if (key1->obj_id < key2->obj_id)
2921                 return(-4);
2922         if (key1->obj_id > key2->obj_id)
2923                 return(4);
2924
2925         if (key1->rec_type < key2->rec_type)
2926                 return(-3);
2927         if (key1->rec_type > key2->rec_type)
2928                 return(3);
2929
2930         if (key1->key < key2->key)
2931                 return(-2);
2932         if (key1->key > key2->key)
2933                 return(2);
2934
2935         /*
2936          * A create_tid of zero indicates a record which is undeletable
2937          * and must be considered to have a value of positive infinity.
2938          */
2939         if (key1->create_tid == 0) {
2940                 if (key2->create_tid == 0)
2941                         return(0);
2942                 return(1);
2943         }
2944         if (key2->create_tid == 0)
2945                 return(-1);
2946         if (key1->create_tid < key2->create_tid)
2947                 return(-1);
2948         if (key1->create_tid > key2->create_tid)
2949                 return(1);
2950         return(0);
2951 }
2952
2953 /*
2954  * Test a timestamp against an element to determine whether the
2955  * element is visible.  A timestamp of 0 means 'infinity'.
2956  */
2957 int
2958 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2959 {
2960         if (asof == 0) {
2961                 if (base->delete_tid)
2962                         return(1);
2963                 return(0);
2964         }
2965         if (asof < base->create_tid)
2966                 return(-1);
2967         if (base->delete_tid && asof >= base->delete_tid)
2968                 return(1);
2969         return(0);
2970 }
2971
2972 /*
2973  * Create a separator half way inbetween key1 and key2.  For fields just
2974  * one unit apart, the separator will match key2.  key1 is on the left-hand
2975  * side and key2 is on the right-hand side.
2976  *
2977  * key2 must be >= the separator.  It is ok for the separator to match key2.
2978  *
2979  * NOTE: Even if key1 does not match key2, the separator may wind up matching
2980  * key2.
2981  *
2982  * NOTE: It might be beneficial to just scrap this whole mess and just
2983  * set the separator to key2.
2984  */
2985 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2986         dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2987
2988 static void
2989 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2990                       hammer_base_elm_t dest)
2991 {
2992         bzero(dest, sizeof(*dest));
2993
2994         dest->rec_type = key2->rec_type;
2995         dest->key = key2->key;
2996         dest->obj_id = key2->obj_id;
2997         dest->create_tid = key2->create_tid;
2998
2999         MAKE_SEPARATOR(key1, key2, dest, localization);
3000         if (key1->localization == key2->localization) {
3001                 MAKE_SEPARATOR(key1, key2, dest, obj_id);
3002                 if (key1->obj_id == key2->obj_id) {
3003                         MAKE_SEPARATOR(key1, key2, dest, rec_type);
3004                         if (key1->rec_type == key2->rec_type) {
3005                                 MAKE_SEPARATOR(key1, key2, dest, key);
3006                                 /*
3007                                  * Don't bother creating a separator for
3008                                  * create_tid, which also conveniently avoids
3009                                  * having to handle the create_tid == 0
3010                                  * (infinity) case.  Just leave create_tid
3011                                  * set to key2.
3012                                  *
3013                                  * Worst case, dest matches key2 exactly,
3014                                  * which is acceptable.
3015                                  */
3016                         }
3017                 }
3018         }
3019 }
3020
3021 #undef MAKE_SEPARATOR
3022
3023 /*
3024  * Return whether a generic internal or leaf node is full
3025  */
3026 static int
3027 btree_node_is_full(hammer_node_ondisk_t node)
3028 {
3029         switch(node->type) {
3030         case HAMMER_BTREE_TYPE_INTERNAL:
3031                 if (node->count == HAMMER_BTREE_INT_ELMS)
3032                         return(1);
3033                 break;
3034         case HAMMER_BTREE_TYPE_LEAF:
3035                 if (node->count == HAMMER_BTREE_LEAF_ELMS)
3036                         return(1);
3037                 break;
3038         default:
3039                 panic("illegal btree subtype");
3040         }
3041         return(0);
3042 }
3043
3044 #if 0
3045 static int
3046 btree_max_elements(u_int8_t type)
3047 {
3048         if (type == HAMMER_BTREE_TYPE_LEAF)
3049                 return(HAMMER_BTREE_LEAF_ELMS);
3050         if (type == HAMMER_BTREE_TYPE_INTERNAL)
3051                 return(HAMMER_BTREE_INT_ELMS);
3052         panic("btree_max_elements: bad type %d", type);
3053 }
3054 #endif
3055
3056 void
3057 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
3058 {
3059         hammer_btree_elm_t elm;
3060         int i;
3061
3062         kprintf("node %p count=%d parent=%016llx type=%c\n",
3063                 ondisk, ondisk->count,
3064                 (long long)ondisk->parent, ondisk->type);
3065
3066         /*
3067          * Dump both boundary elements if an internal node
3068          */
3069         if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
3070                 for (i = 0; i <= ondisk->count; ++i) {
3071                         elm = &ondisk->elms[i];
3072                         hammer_print_btree_elm(elm, ondisk->type, i);
3073                 }
3074         } else {
3075                 for (i = 0; i < ondisk->count; ++i) {
3076                         elm = &ondisk->elms[i];
3077                         hammer_print_btree_elm(elm, ondisk->type, i);
3078                 }
3079         }
3080 }
3081
3082 void
3083 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
3084 {
3085         kprintf("  %2d", i);
3086         kprintf("\tobj_id       = %016llx\n", (long long)elm->base.obj_id);
3087         kprintf("\tkey          = %016llx\n", (long long)elm->base.key);
3088         kprintf("\tcreate_tid   = %016llx\n", (long long)elm->base.create_tid);
3089         kprintf("\tdelete_tid   = %016llx\n", (long long)elm->base.delete_tid);
3090         kprintf("\trec_type     = %04x\n", elm->base.rec_type);
3091         kprintf("\tobj_type     = %02x\n", elm->base.obj_type);
3092         kprintf("\tbtype        = %02x (%c)\n",
3093                 elm->base.btype,
3094                 (elm->base.btype ? elm->base.btype : '?'));
3095         kprintf("\tlocalization = %02x\n", elm->base.localization);
3096
3097         switch(type) {
3098         case HAMMER_BTREE_TYPE_INTERNAL:
3099                 kprintf("\tsubtree_off  = %016llx\n",
3100                         (long long)elm->internal.subtree_offset);
3101                 break;
3102         case HAMMER_BTREE_TYPE_RECORD:
3103                 kprintf("\tdata_offset  = %016llx\n",
3104                         (long long)elm->leaf.data_offset);
3105                 kprintf("\tdata_len     = %08x\n", elm->leaf.data_len);
3106                 kprintf("\tdata_crc     = %08x\n", elm->leaf.data_crc);
3107                 break;
3108         }
3109 }