sys/vfs/hammer/hammer_btree.c

   1 /*
   2  * Copyright (c) 2007 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.28 2008/02/05 07:58:43 dillon Exp $
  35  */
  36
  37 /*
  38  * HAMMER B-Tree index
  39  *
  40  * HAMMER implements a modified B+Tree.  In documentation this will
  41  * simply be refered to as the HAMMER B-Tree.  Basically a HAMMER B-Tree
  42  * looks like a B+Tree (A B-Tree which stores its records only at the leafs
  43  * of the tree), but adds two additional boundary elements which describe
  44  * the left-most and right-most element a node is able to represent.  In
  45  * otherwords, we have boundary elements at the two ends of a B-Tree node
  46  * instead of sub-tree pointers.
  47  *
  48  * A B-Tree internal node looks like this:
  49  *
  50  *      B N N N N N N B   <-- boundary and internal elements
  51  *       S S S S S S S    <-- subtree pointers
  52  *
  53  * A B-Tree leaf node basically looks like this:
  54  *
  55  *      L L L L L L L L   <-- leaf elemenets
  56  *
  57  * The radix for an internal node is 1 less then a leaf but we get a
  58  * number of significant benefits for our troubles.
  59  *
  60  * The big benefit to using a B-Tree containing boundary information
  61  * is that it is possible to cache pointers into the middle of the tree
  62  * and not have to start searches, insertions, OR deletions at the root
  63  * node.   In particular, searches are able to progress in a definitive
  64  * direction from any point in the tree without revisting nodes.  This
  65  * greatly improves the efficiency of many operations, most especially
  66  * record appends.
  67  *
  68  * B-Trees also make the stacking of trees fairly straightforward.
  69  *
  70  * SPIKES: Two leaf elements denoting a sub-range of keys may represent
  71  * a spike, or a recursion into another cluster.  Most standard B-Tree
  72  * searches traverse spikes.  The ending spike element is range-inclusive
  73  * and does not operate quite like a right-bound.
  74  *
  75  * INSERTIONS:  A search performed with the intention of doing
  76  * an insert will guarantee that the terminal leaf node is not full by
  77  * splitting full nodes.  Splits occur top-down during the dive down the
  78  * B-Tree.
  79  *
  80  * DELETIONS: A deletion makes no attempt to proactively balance the
  81  * tree and will recursively remove nodes that become empty.  Empty
  82  * nodes are not allowed and a deletion may recurse upwards from the leaf.
  83  * Rather then allow a deadlock a deletion may terminate early by setting
  84  * an internal node's element's subtree_offset to 0.  The deletion will
  85  * then be resumed the next time a search encounters the element.
  86  */
  87 #include "hammer.h"
  88 #include <sys/buf.h>
  89 #include <sys/buf2.h>
  90
  91 static int btree_search(hammer_cursor_t cursor, int flags);
  92 static int btree_split_internal(hammer_cursor_t cursor);
  93 static int btree_split_leaf(hammer_cursor_t cursor);
  94 static int btree_remove(hammer_cursor_t cursor);
  95 static int btree_remove_deleted_element(hammer_cursor_t cursor);
  96 static int btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm);
  97 static int btree_node_is_almost_full(hammer_node_ondisk_t node);
  98 static int btree_node_is_full(hammer_node_ondisk_t node);
  99 static void hammer_make_separator(hammer_base_elm_t key1,
 100                         hammer_base_elm_t key2, hammer_base_elm_t dest);
 101
 102 /*
 103  * Iterate records after a search.  The cursor is iterated forwards past
 104  * the current record until a record matching the key-range requirements
 105  * is found.  ENOENT is returned if the iteration goes past the ending
 106  * key.
 107  *
 108  * The iteration is inclusive of key_beg and can be inclusive or exclusive
 109  * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
 110  *
 111  * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
 112  * may be modified by B-Tree functions.
 113  *
 114  * cursor->key_beg may or may not be modified by this function during
 115  * the iteration.  XXX future - in case of an inverted lock we may have
 116  * to reinitiate the lookup and set key_beg to properly pick up where we
 117  * left off.
 118  *
 119  * NOTE!  EDEADLK *CANNOT* be returned by this procedure.
 120  */
 121 int
 122 hammer_btree_iterate(hammer_cursor_t cursor)
 123 {
 124         hammer_node_ondisk_t node;
 125         hammer_btree_elm_t elm;
 126         int error;
 127         int r;
 128         int s;
 129
 130         /*
 131          * Skip past the current record
 132          */
 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          * Loop until an element is found or we are done.
 143          */
 144         for (;;) {
 145                 /*
 146                  * We iterate up the tree and then index over one element
 147                  * while we are at the last element in the current node.
 148                  *
 149                  * NOTE: This can pop us up to another cluster.
 150                  *
 151                  * If we are at the root of the root cluster, cursor_up
 152                  * returns ENOENT.
 153                  *
 154                  * NOTE: hammer_cursor_up() will adjust cursor->key_beg
 155                  * when told to re-search for the cluster tag.
 156                  *
 157                  * XXX this could be optimized by storing the information in
 158                  * the parent reference.
 159                  *
 160                  * XXX we can lose the node lock temporarily, this could mess
 161                  * up our scan.
 162                  */
 163                 if (cursor->index == node->count) {
 164                         error = hammer_cursor_up(cursor);
 165                         if (error)
 166                                 break;
 167                         /* reload stale pointer */
 168                         node = cursor->node->ondisk;
 169                         KKASSERT(cursor->index != node->count);
 170                         ++cursor->index;
 171                         continue;
 172                 }
 173
 174                 /*
 175                  * Check internal or leaf element.  Determine if the record
 176                  * at the cursor has gone beyond the end of our range.
 177                  *
 178                  * Generally we recurse down through internal nodes.  An
 179                  * internal node can only be returned if INCLUSTER is set
 180                  * and the node represents a cluster-push record.
 181                  */
 182                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 183                         elm = &node->elms[cursor->index];
 184                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 185                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 186                         if (hammer_debug_btree) {
 187                                 kprintf("BRACKETL %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
 188                                         cursor->node->cluster->volume->vol_no,
 189                                         cursor->node->cluster->clu_no,
 190                                         cursor->node->node_offset,
 191                                         cursor->index,
 192                                         elm[0].internal.base.obj_id,
 193                                         elm[0].internal.base.rec_type,
 194                                         elm[0].internal.base.key,
 195                                         r
 196                                 );
 197                                 kprintf("BRACKETR %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
 198                                         cursor->node->cluster->volume->vol_no,
 199                                         cursor->node->cluster->clu_no,
 200                                         cursor->node->node_offset,
 201                                         cursor->index + 1,
 202                                         elm[1].internal.base.obj_id,
 203                                         elm[1].internal.base.rec_type,
 204                                         elm[1].internal.base.key,
 205                                         s
 206                                 );
 207                         }
 208
 209                         if (r < 0) {
 210                                 error = ENOENT;
 211                                 break;
 212                         }
 213                         if (r == 0 && (cursor->flags &
 214                                        HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 215                                 error = ENOENT;
 216                                 break;
 217                         }
 218                         KKASSERT(s <= 0);
 219
 220                         /*
 221                          * When iterating try to clean up any deleted
 222                          * internal elements left over from btree_remove()
 223                          * deadlocks, but it is ok if we can't.
 224                          */
 225                         if (elm->internal.subtree_offset == 0) {
 226                                 btree_remove_deleted_element(cursor);
 227                                 /* note: elm also invalid */
 228                         } else if (elm->internal.subtree_offset != 0) {
 229                                 error = hammer_cursor_down(cursor);
 230                                 if (error)
 231                                         break;
 232                                 KKASSERT(cursor->index == 0);
 233                         }
 234                         /* reload stale pointer */
 235                         node = cursor->node->ondisk;
 236                         continue;
 237                 } else {
 238                         elm = &node->elms[cursor->index];
 239                         r = hammer_btree_cmp(&cursor->key_end, &elm->base);
 240                         if (hammer_debug_btree) {
 241                                 kprintf("ELEMENT  %d:%d:%08x:%d %c %016llx %02x %016llx %d\n",
 242                                         cursor->node->cluster->volume->vol_no,
 243                                         cursor->node->cluster->clu_no,
 244                                         cursor->node->node_offset,
 245                                         cursor->index,
 246                                         (elm[0].leaf.base.btype ?
 247                                          elm[0].leaf.base.btype : '?'),
 248                                         elm[0].leaf.base.obj_id,
 249                                         elm[0].leaf.base.rec_type,
 250                                         elm[0].leaf.base.key,
 251                                         r
 252                                 );
 253                         }
 254                         if (r < 0) {
 255                                 error = ENOENT;
 256                                 break;
 257                         }
 258
 259                         /*
 260                          * We support both end-inclusive and
 261                          * end-exclusive searches.
 262                          */
 263                         if (r == 0 &&
 264                            (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 265                                 error = ENOENT;
 266                                 break;
 267                         }
 268
 269                         switch(elm->leaf.base.btype) {
 270                         case HAMMER_BTREE_TYPE_RECORD:
 271                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 272                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 273                                         ++cursor->index;
 274                                         continue;
 275                                 }
 276                                 break;
 277                         case HAMMER_BTREE_TYPE_SPIKE_BEG:
 278                                 /*
 279                                  * NOTE: This code assumes that the spike
 280                                  * ending element immediately follows the
 281                                  * spike beginning element.
 282                                  */
 283                                 /*
 284                                  * We must cursor-down via the SPIKE_END
 285                                  * element, otherwise cursor->parent will
 286                                  * not be set correctly for deletions.
 287                                  *
 288                                  * fall-through to avoid an improper
 289                                  * termination from the conditional above.
 290                                  */
 291                                 KKASSERT(cursor->index + 1 < node->count);
 292                                 ++elm;
 293                                 KKASSERT(elm->leaf.base.btype ==
 294                                          HAMMER_BTREE_TYPE_SPIKE_END);
 295                                 ++cursor->index;
 296                                 /* fall through */
 297                         case HAMMER_BTREE_TYPE_SPIKE_END:
 298                                 /*
 299                                  * The SPIKE_END element is inclusive, NOT
 300                                  * like a boundary, so be careful with the
 301                                  * match check.
 302                                  *
 303                                  * This code assumes that a preceding SPIKE_BEG
 304                                  * has already been checked.
 305                                  */
 306                                 if (cursor->flags & HAMMER_CURSOR_INCLUSTER)
 307                                         break;
 308                                 error = hammer_cursor_down(cursor);
 309                                 if (error)
 310                                         break;
 311                                 KKASSERT(cursor->index == 0);
 312                                 /* reload stale pointer */
 313                                 node = cursor->node->ondisk;
 314
 315                                 /*
 316                                  * If the cluster root is empty it and its
 317                                  * related spike can be deleted.  Ignore
 318                                  * errors.  Cursor
 319                                  */
 320                                 if (node->count == 0) {
 321                                         error = hammer_cursor_upgrade(cursor);
 322                                         if (error == 0)
 323                                                 error = btree_remove(cursor);
 324                                         hammer_cursor_downgrade(cursor);
 325                                         error = 0;
 326                                         /* reload stale pointer */
 327                                         node = cursor->node->ondisk;
 328                                 }
 329                                 continue;
 330                         default:
 331                                 error = EINVAL;
 332                                 break;
 333                         }
 334                         if (error)
 335                                 break;
 336                 }
 337                 /*
 338                  * node pointer invalid after loop
 339                  */
 340
 341                 /*
 342                  * Return entry
 343                  */
 344                 if (hammer_debug_btree) {
 345                         int i = cursor->index;
 346                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 347                         kprintf("ITERATE  %p:%d %016llx %02x %016llx\n",
 348                                 cursor->node, i,
 349                                 elm->internal.base.obj_id,
 350                                 elm->internal.base.rec_type,
 351                                 elm->internal.base.key
 352                         );
 353                 }
 354                 return(0);
 355         }
 356         return(error);
 357 }
 358
 359 /*
 360  * Iterate in the reverse direction.  This is used by the pruning code to
 361  * avoid overlapping records.
 362  */
 363 int
 364 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
 365 {
 366         hammer_node_ondisk_t node;
 367         hammer_btree_elm_t elm;
 368         int error;
 369         int r;
 370         int s;
 371
 372         /*
 373          * Skip past the current record.  For various reasons the cursor
 374          * may end up set to -1 or set to point at the end of the current
 375          * node.  These cases must be addressed.
 376          */
 377         node = cursor->node->ondisk;
 378         if (node == NULL)
 379                 return(ENOENT);
 380         if (cursor->index != -1 &&
 381             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
 382                 --cursor->index;
 383         }
 384         if (cursor->index == cursor->node->ondisk->count)
 385                 --cursor->index;
 386
 387         /*
 388          * Loop until an element is found or we are done.
 389          */
 390         for (;;) {
 391                 /*
 392                  * We iterate up the tree and then index over one element
 393                  * while we are at the last element in the current node.
 394                  *
 395                  * NOTE: This can pop us up to another cluster.
 396                  *
 397                  * If we are at the root of the root cluster, cursor_up
 398                  * returns ENOENT.
 399                  *
 400                  * NOTE: hammer_cursor_up() will adjust cursor->key_beg
 401                  * when told to re-search for the cluster tag.
 402                  *
 403                  * XXX this could be optimized by storing the information in
 404                  * the parent reference.
 405                  *
 406                  * XXX we can lose the node lock temporarily, this could mess
 407                  * up our scan.
 408                  */
 409                 if (cursor->index == -1) {
 410                         error = hammer_cursor_up(cursor);
 411                         if (error) {
 412                                 cursor->index = 0; /* sanity */
 413                                 break;
 414                         }
 415                         /* reload stale pointer */
 416                         node = cursor->node->ondisk;
 417                         KKASSERT(cursor->index != node->count);
 418                         --cursor->index;
 419                         continue;
 420                 }
 421
 422                 /*
 423                  * Check internal or leaf element.  Determine if the record
 424                  * at the cursor has gone beyond the end of our range.
 425                  *
 426                  * Generally we recurse down through internal nodes.  An
 427                  * internal node can only be returned if INCLUSTER is set
 428                  * and the node represents a cluster-push record.
 429                  */
 430                 KKASSERT(cursor->index != node->count);
 431                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
 432                         elm = &node->elms[cursor->index];
 433                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
 434                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
 435                         if (hammer_debug_btree) {
 436                                 kprintf("BRACKETL %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
 437                                         cursor->node->cluster->volume->vol_no,
 438                                         cursor->node->cluster->clu_no,
 439                                         cursor->node->node_offset,
 440                                         cursor->index,
 441                                         elm[0].internal.base.obj_id,
 442                                         elm[0].internal.base.rec_type,
 443                                         elm[0].internal.base.key,
 444                                         r
 445                                 );
 446                                 kprintf("BRACKETR %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
 447                                         cursor->node->cluster->volume->vol_no,
 448                                         cursor->node->cluster->clu_no,
 449                                         cursor->node->node_offset,
 450                                         cursor->index + 1,
 451                                         elm[1].internal.base.obj_id,
 452                                         elm[1].internal.base.rec_type,
 453                                         elm[1].internal.base.key,
 454                                         s
 455                                 );
 456                         }
 457
 458                         if (s >= 0) {
 459                                 error = ENOENT;
 460                                 break;
 461                         }
 462                         KKASSERT(r >= 0);
 463
 464                         /*
 465                          * When iterating try to clean up any deleted
 466                          * internal elements left over from btree_remove()
 467                          * deadlocks, but it is ok if we can't.
 468                          */
 469                         if (elm->internal.subtree_offset == 0) {
 470                                 btree_remove_deleted_element(cursor);
 471                                 /* note: elm also invalid */
 472                         } else if (elm->internal.subtree_offset != 0) {
 473                                 error = hammer_cursor_down(cursor);
 474                                 if (error)
 475                                         break;
 476                                 KKASSERT(cursor->index == 0);
 477                                 cursor->index = cursor->node->ondisk->count - 1;
 478                         }
 479                         /* reload stale pointer */
 480                         node = cursor->node->ondisk;
 481                         continue;
 482                 } else {
 483                         elm = &node->elms[cursor->index];
 484                         s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
 485                         if (hammer_debug_btree) {
 486                                 kprintf("ELEMENT  %d:%d:%08x:%d %c %016llx %02x %016llx %d\n",
 487                                         cursor->node->cluster->volume->vol_no,
 488                                         cursor->node->cluster->clu_no,
 489                                         cursor->node->node_offset,
 490                                         cursor->index,
 491                                         (elm[0].leaf.base.btype ?
 492                                          elm[0].leaf.base.btype : '?'),
 493                                         elm[0].leaf.base.obj_id,
 494                                         elm[0].leaf.base.rec_type,
 495                                         elm[0].leaf.base.key,
 496                                         s
 497                                 );
 498                         }
 499                         if (s > 0) {
 500                                 error = ENOENT;
 501                                 break;
 502                         }
 503
 504                         switch(elm->leaf.base.btype) {
 505                         case HAMMER_BTREE_TYPE_RECORD:
 506                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
 507                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
 508                                         --cursor->index;
 509                                         continue;
 510                                 }
 511                                 break;
 512                         case HAMMER_BTREE_TYPE_SPIKE_BEG:
 513                                 /*
 514                                  * Skip the spike BEG record.  We will hit
 515                                  * the END record first since we are
 516                                  * iterating backwards.
 517                                  */
 518                                 --cursor->index;
 519                                 continue;
 520                         case HAMMER_BTREE_TYPE_SPIKE_END:
 521                                 /*
 522                                  * The SPIKE_END element is inclusive, NOT
 523                                  * like a boundary, so be careful with the
 524                                  * match check.
 525                                  *
 526                                  * This code assumes that a preceding SPIKE_BEG
 527                                  * has already been checked.
 528                                  */
 529                                 if (cursor->flags & HAMMER_CURSOR_INCLUSTER)
 530                                         break;
 531                                 error = hammer_cursor_down(cursor);
 532                                 if (error)
 533                                         break;
 534                                 KKASSERT(cursor->index == 0);
 535                                 /* reload stale pointer */
 536                                 node = cursor->node->ondisk;
 537
 538                                 /*
 539                                  * If the cluster root is empty it and its
 540                                  * related spike can be deleted.  Ignore
 541                                  * errors.  Cursor
 542                                  */
 543                                 if (node->count == 0) {
 544                                         error = hammer_cursor_upgrade(cursor);
 545                                         if (error == 0)
 546                                                 error = btree_remove(cursor);
 547                                         hammer_cursor_downgrade(cursor);
 548                                         error = 0;
 549                                         /* reload stale pointer */
 550                                         node = cursor->node->ondisk;
 551                                 }
 552                                 cursor->index = node->count - 1;
 553                                 continue;
 554                         default:
 555                                 error = EINVAL;
 556                                 break;
 557                         }
 558                         if (error)
 559                                 break;
 560                 }
 561                 /*
 562                  * node pointer invalid after loop
 563                  */
 564
 565                 /*
 566                  * Return entry
 567                  */
 568                 if (hammer_debug_btree) {
 569                         int i = cursor->index;
 570                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
 571                         kprintf("ITERATE  %p:%d %016llx %02x %016llx\n",
 572                                 cursor->node, i,
 573                                 elm->internal.base.obj_id,
 574                                 elm->internal.base.rec_type,
 575                                 elm->internal.base.key
 576                         );
 577                 }
 578                 return(0);
 579         }
 580         return(error);
 581 }
 582
 583 /*
 584  * Lookup cursor->key_beg.  0 is returned on success, ENOENT if the entry
 585  * could not be found, EDEADLK if inserting and a retry is needed, and a
 586  * fatal error otherwise.  When retrying, the caller must terminate the
 587  * cursor and reinitialize it.  EDEADLK cannot be returned if not inserting.
 588  *
 589  * The cursor is suitably positioned for a deletion on success, and suitably
 590  * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
 591  * specified.
 592  *
 593  * The cursor may begin anywhere, the search will traverse clusters in
 594  * either direction to locate the requested element.
 595  *
 596  * Most of the logic implementing historical searches is handled here.  We
 597  * do an initial lookup with create_tid set to the asof TID.  Due to the
 598  * way records are laid out, a backwards iteration may be required if
 599  * ENOENT is returned to locate the historical record.  Here's the
 600  * problem:
 601  *
 602  * create_tid:    10      15       20
 603  *                   LEAF1   LEAF2
 604  * records:         (11)        (18)
 605  *
 606  * Lets say we want to do a lookup AS-OF timestamp 17.  We will traverse
 607  * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
 608  * not visible and thus causes ENOENT to be returned.  We really need
 609  * to check record 11 in LEAF1.  If it also fails then the search fails
 610  * (e.g. it might represent the range 11-16 and thus still not match our
 611  * AS-OF timestamp of 17).
 612  *
 613  * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
 614  * and the cursor->create_check TID if an iteration might be needed.
 615  * In the above example create_check would be set to 14.
 616  */
 617 int
 618 hammer_btree_lookup(hammer_cursor_t cursor)
 619 {
 620         int error;
 621
 622         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 623                 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
 624                 cursor->key_beg.create_tid = cursor->asof;
 625                 for (;;) {
 626                         cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
 627                         error = btree_search(cursor, 0);
 628                         if (error != ENOENT ||
 629                             (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
 630                                 /*
 631                                  * Stop if no error.
 632                                  * Stop if error other then ENOENT.
 633                                  * Stop if ENOENT and not special case.
 634                                  */
 635                                 break;
 636                         }
 637                         if (hammer_debug_btree) {
 638                                 kprintf("CREATE_CHECK %016llx\n",
 639                                         cursor->create_check);
 640                         }
 641                         cursor->key_beg.create_tid = cursor->create_check;
 642                         /* loop */
 643                 }
 644         } else {
 645                 error = btree_search(cursor, 0);
 646         }
 647         if (error == 0 && cursor->flags)
 648                 error = hammer_btree_extract(cursor, cursor->flags);
 649         return(error);
 650 }
 651
 652 /*
 653  * Execute the logic required to start an iteration.  The first record
 654  * located within the specified range is returned and iteration control
 655  * flags are adjusted for successive hammer_btree_iterate() calls.
 656  */
 657 int
 658 hammer_btree_first(hammer_cursor_t cursor)
 659 {
 660         int error;
 661
 662         error = hammer_btree_lookup(cursor);
 663         if (error == ENOENT) {
 664                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 665                 error = hammer_btree_iterate(cursor);
 666         }
 667         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 668         return(error);
 669 }
 670
 671 /*
 672  * Similarly but for an iteration in the reverse direction.
 673  */
 674 int
 675 hammer_btree_last(hammer_cursor_t cursor)
 676 {
 677         struct hammer_base_elm save;
 678         int error;
 679
 680         save = cursor->key_beg;
 681         cursor->key_beg = cursor->key_end;
 682         error = hammer_btree_lookup(cursor);
 683         cursor->key_beg = save;
 684         if (error == ENOENT ||
 685             (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
 686                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 687                 error = hammer_btree_iterate_reverse(cursor);
 688         }
 689         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 690         return(error);
 691 }
 692
 693 /*
 694  * Extract the record and/or data associated with the cursor's current
 695  * position.  Any prior record or data stored in the cursor is replaced.
 696  * The cursor must be positioned at a leaf node.
 697  *
 698  * NOTE: Most extractions occur at the leaf of the B-Tree.  The only
 699  *       extraction allowed at an internal element is at a cluster-push.
 700  *       Cluster-push elements have records but no data.
 701  */
 702 int
 703 hammer_btree_extract(hammer_cursor_t cursor, int flags)
 704 {
 705         hammer_node_ondisk_t node;
 706         hammer_btree_elm_t elm;
 707         hammer_cluster_t cluster;
 708         u_int64_t buf_type;
 709         int32_t cloff;
 710         int32_t roff;
 711         int error;
 712
 713         /*
 714          * A cluster record type has no data reference, the information
 715          * is stored directly in the record and B-Tree element.
 716          *
 717          * The case where the data reference resolves to the same buffer
 718          * as the record reference must be handled.
 719          */
 720         node = cursor->node->ondisk;
 721         elm = &node->elms[cursor->index];
 722         cluster = cursor->node->cluster;
 723         cursor->flags &= ~HAMMER_CURSOR_DATA_EMBEDDED;
 724         cursor->data = NULL;
 725
 726         /*
 727          * There is nothing to extract for an internal element.
 728          */
 729         if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
 730                 return(EINVAL);
 731
 732         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 733
 734         /*
 735          * Leaf element.
 736          */
 737         if ((flags & HAMMER_CURSOR_GET_RECORD)) {
 738                 cloff = elm->leaf.rec_offset;
 739                 cursor->record = hammer_bread(cluster, cloff,
 740                                               HAMMER_FSBUF_RECORDS, &error,
 741                                               &cursor->record_buffer);
 742         } else {
 743                 cloff = 0;
 744                 error = 0;
 745         }
 746         if ((flags & HAMMER_CURSOR_GET_DATA) && error == 0) {
 747                 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) {
 748                         /*
 749                          * Only records have data references.  Spike elements
 750                          * do not.
 751                          */
 752                         cursor->data = NULL;
 753                 } else if ((cloff ^ elm->leaf.data_offset) & ~HAMMER_BUFMASK) {
 754                         /*
 755                          * The data is not in the same buffer as the last
 756                          * record we cached, but it could still be embedded
 757                          * in a record.  Note that we may not have loaded the
 758                          * record's buffer above, depending on flags.
 759                          */
 760                         if ((elm->leaf.rec_offset ^ elm->leaf.data_offset) &
 761                             ~HAMMER_BUFMASK) {
 762                                 if (elm->leaf.data_len & HAMMER_BUFMASK)
 763                                         buf_type = HAMMER_FSBUF_DATA;
 764                                 else
 765                                         buf_type = 0;   /* pure data buffer */
 766                         } else {
 767                                 buf_type = HAMMER_FSBUF_RECORDS;
 768                         }
 769                         cursor->data = hammer_bread(cluster,
 770                                                   elm->leaf.data_offset,
 771                                                   buf_type, &error,
 772                                                   &cursor->data_buffer);
 773                 } else {
 774                         /*
 775                          * Data in same buffer as record.  Note that we
 776                          * leave any existing data_buffer intact, even
 777                          * though we don't use it in this case, in case
 778                          * other records extracted during an iteration
 779                          * go back to it.
 780                          *
 781                          * The data must be embedded in the record for this
 782                          * case to be hit.
 783                          *
 784                          * Just assume the buffer type is correct.
 785                          */
 786                         cursor->data = (void *)
 787                                 ((char *)cursor->record_buffer->ondisk +
 788                                  (elm->leaf.data_offset & HAMMER_BUFMASK));
 789                         roff = (char *)cursor->data - (char *)cursor->record;
 790                         KKASSERT (roff >= 0 && roff < HAMMER_RECORD_SIZE);
 791                         cursor->flags |= HAMMER_CURSOR_DATA_EMBEDDED;
 792                 }
 793         }
 794         return(error);
 795 }
 796
 797
 798 /*
 799  * Insert a leaf element into the B-Tree at the current cursor position.
 800  * The cursor is positioned such that the element at and beyond the cursor
 801  * are shifted to make room for the new record.
 802  *
 803  * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
 804  * flag set and that call must return ENOENT before this function can be
 805  * called.
 806  *
 807  * ENOSPC is returned if there is no room to insert a new record.
 808  */
 809 int
 810 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm)
 811 {
 812         hammer_node_ondisk_t node;
 813         int i;
 814         int error;
 815
 816         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 817                 return(error);
 818
 819         /*
 820          * Insert the element at the leaf node and update the count in the
 821          * parent.  It is possible for parent to be NULL, indicating that
 822          * the root of the B-Tree in the cluster is a leaf.  It is also
 823          * possible for the leaf to be empty.
 824          *
 825          * Remember that the right-hand boundary is not included in the
 826          * count.
 827          */
 828         hammer_modify_node(cursor->node);
 829         node = cursor->node->ondisk;
 830         i = cursor->index;
 831         KKASSERT(elm->base.btype != 0);
 832         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 833         KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
 834         if (i != node->count) {
 835                 bcopy(&node->elms[i], &node->elms[i+1],
 836                       (node->count - i) * sizeof(*elm));
 837         }
 838         node->elms[i] = *elm;
 839         ++node->count;
 840
 841         /*
 842          * Debugging sanity checks.  Note that the element to the left
 843          * can match the element we are inserting if it is a SPIKE_END,
 844          * because spike-end's represent a non-inclusive end to a range.
 845          */
 846         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
 847         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
 848         if (i) {
 849                 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->leaf.base) < 0);
 850         }
 851         if (i != node->count - 1)
 852                 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->leaf.base) > 0);
 853
 854         return(0);
 855 }
 856
 857 /*
 858  * Insert a cluster spike into the B-Tree at the current cursor position.
 859  * The caller pre-positions the insertion cursor at ncluster's
 860  * left bound in the originating cluster.  Both the originating cluster
 861  * and the target cluster must be serialized, EDEADLK is fatal.
 862  *
 863  * Basically we have to lay down the two spike elements and assert that
 864  * the leaf's right bound does not bisect the ending element.  The ending
 865  * spike element is non-inclusive, just like a boundary.  The target cluster's
 866  * clu_btree_parent_offset may have to adjusted.
 867  *
 868  * NOTE: Serialization is usually accoplished by virtue of being the
 869  * initial accessor of a cluster.
 870  */
 871 int
 872 hammer_btree_insert_cluster(hammer_cursor_t cursor, hammer_cluster_t ncluster,
 873                             int32_t rec_offset)
 874 {
 875         hammer_node_ondisk_t node;
 876         hammer_btree_elm_t   elm;
 877         hammer_cluster_t     ocluster;
 878         const int esize = sizeof(*elm);
 879         int error;
 880         int i;
 881         int32_t node_offset;
 882
 883         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 884                 return(error);
 885         hammer_modify_node(cursor->node);
 886         node = cursor->node->ondisk;
 887         node_offset = cursor->node->node_offset;
 888         i = cursor->index;
 889
 890         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
 891         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS - 2);
 892         KKASSERT(i >= 0 && i <= node->count);
 893
 894         /*
 895          * Make sure the spike is legal or the B-Tree code will get really
 896          * confused.
 897          *
 898          * XXX the right bound my bisect the two spike elements.  We
 899          * need code here to 'fix' the right bound going up the tree
 900          * instead of an assertion.
 901          */
 902         KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_beg,
 903                                   cursor->left_bound) >= 0);
 904         KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
 905                                   cursor->right_bound) <= 0);
 906         if (i != node->count) {
 907                 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
 908                                           &node->elms[i].leaf.base) <= 0);
 909         }
 910
 911         elm = &node->elms[i];
 912         bcopy(elm, elm + 2, (node->count - i) * esize);
 913         bzero(elm, 2 * esize);
 914         node->count += 2;
 915
 916         elm[0].leaf.base = ncluster->ondisk->clu_btree_beg;
 917         elm[0].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_BEG;
 918         elm[0].leaf.rec_offset = rec_offset;
 919         elm[0].leaf.spike_clu_no = ncluster->clu_no;
 920         elm[0].leaf.spike_vol_no = ncluster->volume->vol_no;
 921
 922         elm[1].leaf.base = ncluster->ondisk->clu_btree_end;
 923         elm[1].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_END;
 924         elm[1].leaf.rec_offset = rec_offset;
 925         elm[1].leaf.spike_clu_no = ncluster->clu_no;
 926         elm[1].leaf.spike_vol_no = ncluster->volume->vol_no;
 927
 928         /*
 929          * SPIKE_END must be inclusive, not exclusive.
 930          */
 931         KKASSERT(elm[1].leaf.base.create_tid != 1);
 932         --elm[1].leaf.base.create_tid;
 933
 934         /*
 935          * The target cluster's parent offset may have to be updated.
 936          *
 937          * NOTE: Modifying a cluster header does not mark it open, and
 938          * flushing it will only clear an existing open flag if the cluster
 939          * has been validated.
 940          */
 941         if (hammer_debug_general & 0x40) {
 942                 kprintf("INSERT CLUSTER %d:%d -> %d:%d ",
 943                         ncluster->ondisk->clu_btree_parent_vol_no,
 944                         ncluster->ondisk->clu_btree_parent_clu_no,
 945                         ncluster->volume->vol_no,
 946                         ncluster->clu_no);
 947         }
 948
 949         ocluster = cursor->node->cluster;
 950         if (ncluster->ondisk->clu_btree_parent_offset != node_offset ||
 951             ncluster->ondisk->clu_btree_parent_clu_no != ocluster->clu_no ||
 952             ncluster->ondisk->clu_btree_parent_vol_no != ocluster->volume->vol_no) {
 953                 hammer_modify_cluster(ncluster);
 954                 ncluster->ondisk->clu_btree_parent_offset = node_offset;
 955                 ncluster->ondisk->clu_btree_parent_clu_no = ocluster->clu_no;
 956                 ncluster->ondisk->clu_btree_parent_vol_no = ocluster->volume->vol_no;
 957                 if (hammer_debug_general & 0x40)
 958                         kprintf("(offset fixup)\n");
 959         } else {
 960                 if (hammer_debug_general & 0x40)
 961                         kprintf("(offset unchanged)\n");
 962         }
 963
 964         return(0);
 965 }
 966
 967 /*
 968  * Delete a record from the B-Tree at the current cursor position.
 969  * The cursor is positioned such that the current element is the one
 970  * to be deleted.
 971  *
 972  * On return the cursor will be positioned after the deleted element and
 973  * MAY point to an internal node.  It will be suitable for the continuation
 974  * of an iteration but not for an insertion or deletion.
 975  *
 976  * Deletions will attempt to partially rebalance the B-Tree in an upward
 977  * direction, but will terminate rather then deadlock.  Empty leaves are
 978  * not allowed except at the root node of a cluster.  An early termination
 979  * will leave an internal node with an element whos subtree_offset is 0,
 980  * a case detected and handled by btree_search().
 981  *
 982  * This function can return EDEADLK, requiring the caller to retry the
 983  * operation after clearing the deadlock.
 984  */
 985 int
 986 hammer_btree_delete(hammer_cursor_t cursor)
 987 {
 988         hammer_node_ondisk_t ondisk;
 989         hammer_node_t node;
 990         hammer_node_t parent;
 991         int error;
 992         int i;
 993
 994         if ((error = hammer_cursor_upgrade(cursor)) != 0)
 995                 return(error);
 996
 997         /*
 998          * Delete the element from the leaf node.
 999          *
1000          * Remember that leaf nodes do not have boundaries.
1001          */
1002         node = cursor->node;
1003         ondisk = node->ondisk;
1004         i = cursor->index;
1005
1006         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
1007         KKASSERT(i >= 0 && i < ondisk->count);
1008         hammer_modify_node(node);
1009         if (i + 1 != ondisk->count) {
1010                 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
1011                       (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
1012         }
1013         --ondisk->count;
1014
1015         /*
1016          * Validate local parent
1017          */
1018         if (ondisk->parent) {
1019                 parent = cursor->parent;
1020
1021                 KKASSERT(parent != NULL);
1022                 KKASSERT(parent->node_offset == ondisk->parent);
1023                 KKASSERT(parent->cluster == node->cluster);
1024         }
1025
1026         /*
1027          * If the leaf becomes empty it must be detached from the parent,
1028          * potentially recursing through to the cluster root.
1029          *
1030          * This may reposition the cursor at one of the parent's of the
1031          * current node.
1032          *
1033          * Ignore deadlock errors, that simply means that btree_remove
1034          * was unable to recurse and had to leave the subtree_offset
1035          * in the parent set to 0.
1036          */
1037         KKASSERT(cursor->index <= ondisk->count);
1038         if (ondisk->count == 0) {
1039                 do {
1040                         error = btree_remove(cursor);
1041                 } while (error == EAGAIN);
1042                 if (error == EDEADLK)
1043                         error = 0;
1044         } else {
1045                 error = 0;
1046         }
1047         KKASSERT(cursor->parent == NULL ||
1048                  cursor->parent_index < cursor->parent->ondisk->count);
1049         return(error);
1050 }
1051
1052 /*
1053  * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
1054  *
1055  * Search a cluster's B-Tree for cursor->key_beg, return the matching node.
1056  *
1057  * The search can begin ANYWHERE in the B-Tree.  As a first step the search
1058  * iterates up the tree as necessary to properly position itself prior to
1059  * actually doing the sarch.
1060  *
1061  * INSERTIONS: The search will split full nodes and leaves on its way down
1062  * and guarentee that the leaf it ends up on is not full.  If we run out
1063  * of space the search continues to the leaf (to position the cursor for
1064  * the spike), but ENOSPC is returned.
1065  *
1066  * The search is only guarenteed to end up on a leaf if an error code of 0
1067  * is returned, or if inserting and an error code of ENOENT is returned.
1068  * Otherwise it can stop at an internal node.  On success a search returns
1069  * a leaf node unless INCLUSTER is set and the search located a cluster push
1070  * node (which is an internal node).
1071  *
1072  * COMPLEXITY WARNING!  This is the core B-Tree search code for the entire
1073  * filesystem, and it is not simple code.  Please note the following facts:
1074  *
1075  * - Internal node recursions have a boundary on the left AND right.  The
1076  *   right boundary is non-inclusive.  The create_tid is a generic part
1077  *   of the key for internal nodes.
1078  *
1079  * - Leaf nodes contain terminal elements AND spikes.  A spike recurses into
1080  *   another cluster and contains two leaf elements.. a beginning and an
1081  *   ending element.  The SPIKE_END element is RANGE-EXCLUSIVE, just like a
1082  *   boundary.  This means that it is possible to have two elements
1083  *   (a spike ending element and a record) side by side with the same key.
1084  *
1085  * - Because the SPIKE_END element is range inclusive, it cannot match the
1086  *   right boundary of the parent node.  SPIKE_BEG and SPIKE_END elements
1087  *   always come in pairs, and always exist side by side in the same leaf.
1088  *
1089  * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1090  *   historical search.  ASOF and INSERT are mutually exclusive.  When
1091  *   doing an as-of lookup btree_search() checks for a right-edge boundary
1092  *   case.  If while recursing down the left-edge differs from the key
1093  *   by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1094  *   with cursor->create_check.  This is used by btree_lookup() to iterate.
1095  *   The iteration backwards because as-of searches can wind up going
1096  *   down the wrong branch of the B-Tree.
1097  */
1098 static
1099 int
1100 btree_search(hammer_cursor_t cursor, int flags)
1101 {
1102         hammer_node_ondisk_t node;
1103         hammer_cluster_t cluster;
1104         hammer_btree_elm_t elm;
1105         int error;
1106         int enospc = 0;
1107         int i;
1108         int r;
1109         int s;
1110
1111         flags |= cursor->flags;
1112
1113         if (hammer_debug_btree) {
1114                 kprintf("SEARCH   %d:%d:%08x[%d] %016llx %02x key=%016llx cre=%016llx\n",
1115                         cursor->node->cluster->volume->vol_no,
1116                         cursor->node->cluster->clu_no,
1117                         cursor->node->node_offset,
1118                         cursor->index,
1119                         cursor->key_beg.obj_id,
1120                         cursor->key_beg.rec_type,
1121                         cursor->key_beg.key,
1122                         cursor->key_beg.create_tid
1123                 );
1124         }
1125
1126         /*
1127          * Move our cursor up the tree until we find a node whos range covers
1128          * the key we are trying to locate.  This may move us between
1129          * clusters.
1130          *
1131          * The left bound is inclusive, the right bound is non-inclusive.
1132          * It is ok to cursor up too far so when cursoring across a cluster
1133          * boundary.
1134          *
1135          * First see if we can skip the whole cluster.  hammer_cursor_up()
1136          * handles both cases but this way we don't check the cluster
1137          * bounds when going up the tree within a cluster.
1138          *
1139          * NOTE: If INCLUSTER is set and we are at the root of the cluster,
1140          * hammer_cursor_up() will return ENOENT.
1141          */
1142         cluster = cursor->node->cluster;
1143         for (;;) {
1144                 r = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_beg);
1145                 s = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_end);
1146
1147                 if (r >= 0 && s < 0)
1148                         break;
1149                 error = hammer_cursor_toroot(cursor);
1150                 if (error)
1151                         goto done;
1152                 KKASSERT(cursor->parent);
1153                 error = hammer_cursor_up(cursor);
1154                 if (error)
1155                         goto done;
1156                 cluster = cursor->node->cluster;
1157         }
1158         for (;;) {
1159                 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1160                 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1161                 if (r >= 0 && s < 0)
1162                         break;
1163                 KKASSERT(cursor->parent);
1164                 error = hammer_cursor_up(cursor);
1165                 if (error)
1166                         goto done;
1167         }
1168
1169         /*
1170          * The delete-checks below are based on node, not parent.  Set the
1171          * initial delete-check based on the parent.
1172          */
1173         if (r == 1) {
1174                 KKASSERT(cursor->left_bound->create_tid != 1);
1175                 cursor->create_check = cursor->left_bound->create_tid - 1;
1176                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1177         }
1178
1179         /*
1180          * We better have ended up with a node somewhere, and our second
1181          * while loop had better not have traversed up a cluster.
1182          */
1183         KKASSERT(cursor->node != NULL && cursor->node->cluster == cluster);
1184
1185         /*
1186          * If we are inserting we can't start at a full node if the parent
1187          * is also full (because there is no way to split the node),
1188          * continue running up the tree until the requirement is satisfied
1189          * or we hit the root of the current cluster.
1190          *
1191          * (If inserting we aren't doing an as-of search so we don't have
1192          *  to worry about create_check).
1193          */
1194         while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1195                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1196                         if (btree_node_is_full(cursor->node->ondisk) == 0)
1197                                 break;
1198                 } else {
1199                         if (btree_node_is_almost_full(cursor->node->ondisk) ==0)
1200                                 break;
1201                 }
1202                 if (cursor->node->ondisk->parent == 0 ||
1203                     cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1204                         break;
1205                 }
1206                 error = hammer_cursor_up(cursor);
1207                 /* cluster and node are now may become stale */
1208                 if (error)
1209                         goto done;
1210         }
1211         /* cluster = cursor->node->cluster; not needed until next cluster = */
1212
1213 new_cluster:
1214         /*
1215          * Push down through internal nodes to locate the requested key.
1216          */
1217         cluster = cursor->node->cluster;
1218         node = cursor->node->ondisk;
1219         while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1220                 /*
1221                  * Scan the node to find the subtree index to push down into.
1222                  * We go one-past, then back-up.
1223                  *
1224                  * We must proactively remove deleted elements which may
1225                  * have been left over from a deadlocked btree_remove().
1226                  *
1227                  * The left and right boundaries are included in the loop
1228                  * in order to detect edge cases.
1229                  *
1230                  * If the separator only differs by create_tid (r == 1)
1231                  * and we are doing an as-of search, we may end up going
1232                  * down a branch to the left of the one containing the
1233                  * desired key.  This requires numerous special cases.
1234                  */
1235                 if (hammer_debug_btree) {
1236                         kprintf("SEARCH-I %d:%d:%08x count=%d\n",
1237                                 cursor->node->cluster->volume->vol_no,
1238                                 cursor->node->cluster->clu_no,
1239                                 cursor->node->node_offset,
1240                                 node->count);
1241                 }
1242                 for (i = 0; i <= node->count; ++i) {
1243                         elm = &node->elms[i];
1244                         r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1245                         if (hammer_debug_btree > 2) {
1246                                 kprintf(" IELM %p %d r=%d\n",
1247                                         &node->elms[i], i, r);
1248                         }
1249                         if (r < 0)
1250                                 break;
1251                         if (r == 1) {
1252                                 KKASSERT(elm->base.create_tid != 1);
1253                                 cursor->create_check = elm->base.create_tid - 1;
1254                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1255                         }
1256                 }
1257                 if (hammer_debug_btree) {
1258                         kprintf("SEARCH-I preI=%d/%d r=%d\n",
1259                                 i, node->count, r);
1260                 }
1261
1262                 /*
1263                  * These cases occur when the parent's idea of the boundary
1264                  * is wider then the child's idea of the boundary, and
1265                  * require special handling.  If not inserting we can
1266                  * terminate the search early for these cases but the
1267                  * child's boundaries cannot be unconditionally modified.
1268                  */
1269                 if (i == 0) {
1270                         /*
1271                          * If i == 0 the search terminated to the LEFT of the
1272                          * left_boundary but to the RIGHT of the parent's left
1273                          * boundary.
1274                          */
1275                         u_int8_t save;
1276
1277                         elm = &node->elms[0];
1278
1279                         /*
1280                          * If we aren't inserting we can stop here.
1281                          */
1282                         if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1283                                 cursor->index = 0;
1284                                 return(ENOENT);
1285                         }
1286
1287                         /*
1288                          * Correct a left-hand boundary mismatch.
1289                          *
1290                          * We can only do this if we can upgrade the lock.
1291                          */
1292                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1293                                 return(error);
1294                         hammer_modify_node(cursor->node);
1295                         save = node->elms[0].base.btype;
1296                         node->elms[0].base = *cursor->left_bound;
1297                         node->elms[0].base.btype = save;
1298                 } else if (i == node->count + 1) {
1299                         /*
1300                          * If i == node->count + 1 the search terminated to
1301                          * the RIGHT of the right boundary but to the LEFT
1302                          * of the parent's right boundary.  If we aren't
1303                          * inserting we can stop here.
1304                          *
1305                          * Note that the last element in this case is
1306                          * elms[i-2] prior to adjustments to 'i'.
1307                          */
1308                         --i;
1309                         if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1310                                 cursor->index = i;
1311                                 return (ENOENT);
1312                         }
1313
1314                         /*
1315                          * Correct a right-hand boundary mismatch.
1316                          * (actual push-down record is i-2 prior to
1317                          * adjustments to i).
1318                          *
1319                          * We can only do this if we can upgrade the lock.
1320                          */
1321                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1322                                 return(error);
1323                         elm = &node->elms[i];
1324                         hammer_modify_node(cursor->node);
1325                         elm->base = *cursor->right_bound;
1326                         --i;
1327                 } else {
1328                         /*
1329                          * The push-down index is now i - 1.  If we had
1330                          * terminated on the right boundary this will point
1331                          * us at the last element.
1332                          */
1333                         --i;
1334                 }
1335                 cursor->index = i;
1336                 elm = &node->elms[i];
1337
1338                 if (hammer_debug_btree) {
1339                         kprintf("RESULT-I %d:%d:%08x[%d] %016llx %02x "
1340                                 "key=%016llx cre=%016llx\n",
1341                                 cursor->node->cluster->volume->vol_no,
1342                                 cursor->node->cluster->clu_no,
1343                                 cursor->node->node_offset,
1344                                 i,
1345                                 elm->internal.base.obj_id,
1346                                 elm->internal.base.rec_type,
1347                                 elm->internal.base.key,
1348                                 elm->internal.base.create_tid
1349                         );
1350                 }
1351
1352                 /*
1353                  * When searching try to clean up any deleted
1354                  * internal elements left over from btree_remove()
1355                  * deadlocks.
1356                  *
1357                  * If we fail and we are doing an insertion lookup,
1358                  * we have to return EDEADLK, because an insertion lookup
1359                  * must terminate at a leaf.
1360                  */
1361                 if (elm->internal.subtree_offset == 0) {
1362                         error = btree_remove_deleted_element(cursor);
1363                         if (error == 0)
1364                                 goto new_cluster;
1365                         if (error == EDEADLK &&
1366                             (flags & HAMMER_CURSOR_INSERT) == 0) {
1367                                 error = ENOENT;
1368                         }
1369                         return(error);
1370                 }
1371
1372
1373                 /*
1374                  * Handle insertion and deletion requirements.
1375                  *
1376                  * If inserting split full nodes.  The split code will
1377                  * adjust cursor->node and cursor->index if the current
1378                  * index winds up in the new node.
1379                  *
1380                  * If inserting and a left or right edge case was detected,
1381                  * we cannot correct the left or right boundary and must
1382                  * prepend and append an empty leaf node in order to make
1383                  * the boundary correction.
1384                  *
1385                  * If we run out of space we set enospc and continue on
1386                  * to a leaf to provide the spike code with a good point
1387                  * of entry.  Enospc is reset if we cross a cluster boundary.
1388                  */
1389                 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1390                         if (btree_node_is_full(node)) {
1391                                 error = btree_split_internal(cursor);
1392                                 if (error) {
1393                                         if (error != ENOSPC)
1394                                                 goto done;
1395                                         enospc = 1;
1396                                 }
1397                                 /*
1398                                  * reload stale pointers
1399                                  */
1400                                 i = cursor->index;
1401                                 node = cursor->node->ondisk;
1402                         }
1403                 }
1404
1405                 /*
1406                  * Push down (push into new node, existing node becomes
1407                  * the parent) and continue the search.
1408                  */
1409                 error = hammer_cursor_down(cursor);
1410                 /* node and cluster become stale */
1411                 if (error)
1412                         goto done;
1413                 node = cursor->node->ondisk;
1414                 cluster = cursor->node->cluster;
1415         }
1416
1417         /*
1418          * We are at a leaf, do a linear search of the key array.
1419          *
1420          * If we encounter a spike element type within the necessary
1421          * range we push into it.  Note that SPIKE_END is non-inclusive
1422          * of the spike range.
1423          *
1424          * On success the index is set to the matching element and 0
1425          * is returned.
1426          *
1427          * On failure the index is set to the insertion point and ENOENT
1428          * is returned.
1429          *
1430          * Boundaries are not stored in leaf nodes, so the index can wind
1431          * up to the left of element 0 (index == 0) or past the end of
1432          * the array (index == node->count).
1433          */
1434         KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1435         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1436         if (hammer_debug_btree) {
1437                 kprintf("SEARCH-L %d:%d:%08x count=%d\n",
1438                         cursor->node->cluster->volume->vol_no,
1439                         cursor->node->cluster->clu_no,
1440                         cursor->node->node_offset,
1441                         node->count);
1442         }
1443
1444         for (i = 0; i < node->count; ++i) {
1445                 elm = &node->elms[i];
1446
1447                 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1448
1449                 if (hammer_debug_btree > 1)
1450                         kprintf("  ELM %p %d r=%d\n", &node->elms[i], i, r);
1451
1452                 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG) {
1453                         /*
1454                          * SPIKE_BEG.  Stop if we are to the left of the
1455                          * spike begin element.
1456                          *
1457                          * If we are not the last element in the leaf continue
1458                          * the loop looking for the SPIKE_END.  If we are
1459                          * the last element, however, then push into the
1460                          * spike.
1461                          *
1462                          * If doing an as-of search a Spike demark on a
1463                          * create_tid boundary must be pushed into and an
1464                          * iteration will be forced if it turned out to be
1465                          * the wrong choice.
1466                          *
1467                          * If not doing an as-of search exact comparisons
1468                          * must be used.
1469                          *
1470                          * enospc must be reset because we have crossed a
1471                          * cluster boundary.
1472                          */
1473                         if (r < 0)
1474                                 goto failed;
1475
1476                         /*
1477                          * Set the create_check if the spike element
1478                          * only differs by its create_tid.
1479                          */
1480                         if (r == 1) {
1481                                 cursor->create_check = elm->base.create_tid - 1;
1482                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1483                         }
1484                         if (i != node->count - 1)
1485                                 continue;
1486                         panic("btree_search: illegal spike, no SPIKE_END "
1487                               "in leaf node! %p\n", cursor->node);
1488                 }
1489                 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1490                         /*
1491                          * SPIKE_END.  We can only hit this case if we are
1492                          * greater or equal to SPIKE_BEG.
1493                          *
1494                          * If we are <= SPIKE_END we must push into
1495                          * it, otherwise continue the search.  The SPIKE_END
1496                          * element is range-inclusive.
1497                          *
1498                          * enospc must be reset because we have crossed a
1499                          * cluster boundary.
1500                          */
1501                         if (r > 0) {
1502                                 /*
1503                                  * Continue the search but check for a
1504                                  * create_tid boundary.  Because the
1505                                  * SPIKE_END is inclusive we do not have
1506                                  * to subtract 1 to force an iteration to
1507                                  * go down the spike.
1508                                  */
1509                                 if (r == 1) {
1510                                         cursor->create_check =
1511                                                 elm->base.create_tid - 1;
1512                                         cursor->flags |=
1513                                                 HAMMER_CURSOR_CREATE_CHECK;
1514                                 }
1515                                 continue;
1516                         }
1517                         if (flags & HAMMER_CURSOR_INCLUSTER)
1518                                 goto success;
1519                         cursor->index = i;
1520                         error = hammer_cursor_down(cursor);
1521                         enospc = 0;
1522                         if (error)
1523                                 goto done;
1524                         goto new_cluster;
1525                 }
1526
1527                 /*
1528                  * We are at a record element.  Stop if we've flipped past
1529                  * key_beg, not counting the create_tid test.  Allow the
1530                  * r == 1 case (key_beg > element but differs only by its
1531                  * create_tid) to fall through to the AS-OF check.
1532                  */
1533                 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1534
1535                 if (r < 0)
1536                         goto failed;
1537                 if (r > 1)
1538                         continue;
1539
1540                 /*
1541                  * Check our as-of timestamp against the element.
1542                  */
1543                 if (flags & HAMMER_CURSOR_ASOF) {
1544                         if (hammer_btree_chkts(cursor->asof,
1545                                                &node->elms[i].base) != 0) {
1546                                 continue;
1547                         }
1548                         /* success */
1549                 } else {
1550                         if (r > 0)      /* can only be +1 */
1551                                 continue;
1552                         /* success */
1553                 }
1554 success:
1555                 cursor->index = i;
1556                 error = 0;
1557                 if (hammer_debug_btree) {
1558                         kprintf("RESULT-L %d:%d:%08x[%d] (SUCCESS)\n",
1559                                 cursor->node->cluster->volume->vol_no,
1560                                 cursor->node->cluster->clu_no,
1561                                 cursor->node->node_offset,
1562                                 i);
1563                 }
1564                 goto done;
1565         }
1566
1567         /*
1568          * The search of the leaf node failed.  i is the insertion point.
1569          */
1570 failed:
1571         if (hammer_debug_btree) {
1572                 kprintf("RESULT-L %d:%d:%08x[%d] (FAILED)\n",
1573                         cursor->node->cluster->volume->vol_no,
1574                         cursor->node->cluster->clu_no,
1575                         cursor->node->node_offset,
1576                         i);
1577         }
1578
1579         /*
1580          * No exact match was found, i is now at the insertion point.
1581          *
1582          * If inserting split a full leaf before returning.  This
1583          * may have the side effect of adjusting cursor->node and
1584          * cursor->index.
1585          *
1586          * For now the leaf must have at least 2 free elements to accomodate
1587          * the insertion of a spike during recovery.  See the
1588          * hammer_btree_insert_cluster() function.
1589          */
1590         cursor->index = i;
1591         if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1592              btree_node_is_almost_full(node)) {
1593                 error = btree_split_leaf(cursor);
1594                 if (error) {
1595                         if (error != ENOSPC)
1596                                 goto done;
1597                         enospc = 1;
1598                 }
1599                 /*
1600                  * reload stale pointers
1601                  */
1602                 /* NOT USED
1603                 i = cursor->index;
1604                 node = &cursor->node->internal;
1605                 */
1606         }
1607
1608         /*
1609          * We reached a leaf but did not find the key we were looking for.
1610          * If this is an insert we will be properly positioned for an insert
1611          * (ENOENT) or spike (ENOSPC) operation.
1612          */
1613         error = enospc ? ENOSPC : ENOENT;
1614 done:
1615         return(error);
1616 }
1617
1618
1619 /************************************************************************
1620  *                         SPLITTING AND MERGING                        *
1621  ************************************************************************
1622  *
1623  * These routines do all the dirty work required to split and merge nodes.
1624  */
1625
1626 /*
1627  * Split an internal node into two nodes and move the separator at the split
1628  * point to the parent.
1629  *
1630  * (cursor->node, cursor->index) indicates the element the caller intends
1631  * to push into.  We will adjust node and index if that element winds
1632  * up in the split node.
1633  *
1634  * If we are at the root of a cluster a new root must be created with two
1635  * elements, one pointing to the original root and one pointing to the
1636  * newly allocated split node.
1637  *
1638  * NOTE! Being at the root of a cluster is different from being at the
1639  * root of the root cluster.  cursor->parent will not be NULL and
1640  * cursor->node->ondisk.parent must be tested against 0.  Theoretically
1641  * we could propogate the algorithm into the parent and deal with multiple
1642  * 'roots' in the cluster header, but it's easier not to.
1643  */
1644 static
1645 int
1646 btree_split_internal(hammer_cursor_t cursor)
1647 {
1648         hammer_node_ondisk_t ondisk;
1649         hammer_node_t node;
1650         hammer_node_t parent;
1651         hammer_node_t new_node;
1652         hammer_btree_elm_t elm;
1653         hammer_btree_elm_t parent_elm;
1654         hammer_node_locklist_t locklist = NULL;
1655         int parent_index;
1656         int made_root;
1657         int split;
1658         int error;
1659         int i;
1660         const int esize = sizeof(*elm);
1661
1662         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1663                 return(error);
1664         if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1665                 error = hammer_btree_lock_children(cursor, &locklist);
1666                 if (error)
1667                         goto done;
1668         }
1669
1670         /*
1671          * We are splitting but elms[split] will be promoted to the parent,
1672          * leaving the right hand node with one less element.  If the
1673          * insertion point will be on the left-hand side adjust the split
1674          * point to give the right hand side one additional node.
1675          */
1676         node = cursor->node;
1677         ondisk = node->ondisk;
1678         split = (ondisk->count + 1) / 2;
1679         if (cursor->index <= split)
1680                 --split;
1681
1682         /*
1683          * If we are at the root of the cluster, create a new root node with
1684          * 1 element and split normally.  Avoid making major modifications
1685          * until we know the whole operation will work.
1686          *
1687          * The root of the cluster is different from the root of the root
1688          * cluster.  Use the node's on-disk structure's parent offset to
1689          * detect the case.
1690          */
1691         if (ondisk->parent == 0) {
1692                 parent = hammer_alloc_btree(node->cluster, &error);
1693                 if (parent == NULL)
1694                         goto done;
1695                 hammer_lock_ex(&parent->lock);
1696                 hammer_modify_node(parent);
1697                 ondisk = parent->ondisk;
1698                 ondisk->count = 1;
1699                 ondisk->parent = 0;
1700                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1701                 ondisk->elms[0].base = node->cluster->clu_btree_beg;
1702                 ondisk->elms[0].base.btype = node->ondisk->type;
1703                 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1704                 ondisk->elms[1].base = node->cluster->clu_btree_end;
1705                 /* ondisk->elms[1].base.btype - not used */
1706                 made_root = 1;
1707                 parent_index = 0;       /* index of current node in parent */
1708         } else {
1709                 made_root = 0;
1710                 parent = cursor->parent;
1711                 parent_index = cursor->parent_index;
1712                 KKASSERT(parent->cluster == node->cluster);
1713         }
1714
1715         /*
1716          * Split node into new_node at the split point.
1717          *
1718          *  B O O O P N N B     <-- P = node->elms[split]
1719          *   0 1 2 3 4 5 6      <-- subtree indices
1720          *
1721          *       x x P x x
1722          *        s S S s
1723          *         /   \
1724          *  B O O O B    B N N B        <--- inner boundary points are 'P'
1725          *   0 1 2 3      4 5 6
1726          *
1727          */
1728         new_node = hammer_alloc_btree(node->cluster, &error);
1729         if (new_node == NULL) {
1730                 if (made_root) {
1731                         hammer_unlock(&parent->lock);
1732                         parent->flags |= HAMMER_NODE_DELETED;
1733                         hammer_rel_node(parent);
1734                 }
1735                 goto done;
1736         }
1737         hammer_lock_ex(&new_node->lock);
1738
1739         /*
1740          * Create the new node.  P becomes the left-hand boundary in the
1741          * new node.  Copy the right-hand boundary as well.
1742          *
1743          * elm is the new separator.
1744          */
1745         hammer_modify_node(new_node);
1746         hammer_modify_node(node);
1747         ondisk = node->ondisk;
1748         elm = &ondisk->elms[split];
1749         bcopy(elm, &new_node->ondisk->elms[0],
1750               (ondisk->count - split + 1) * esize);
1751         new_node->ondisk->count = ondisk->count - split;
1752         new_node->ondisk->parent = parent->node_offset;
1753         new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1754         KKASSERT(ondisk->type == new_node->ondisk->type);
1755
1756         /*
1757          * Cleanup the original node.  Elm (P) becomes the new boundary,
1758          * its subtree_offset was moved to the new node.  If we had created
1759          * a new root its parent pointer may have changed.
1760          */
1761         elm->internal.subtree_offset = 0;
1762         ondisk->count = split;
1763
1764         /*
1765          * Insert the separator into the parent, fixup the parent's
1766          * reference to the original node, and reference the new node.
1767          * The separator is P.
1768          *
1769          * Remember that base.count does not include the right-hand boundary.
1770          */
1771         hammer_modify_node(parent);
1772         ondisk = parent->ondisk;
1773         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1774         parent_elm = &ondisk->elms[parent_index+1];
1775         bcopy(parent_elm, parent_elm + 1,
1776               (ondisk->count - parent_index) * esize);
1777         parent_elm->internal.base = elm->base;  /* separator P */
1778         parent_elm->internal.base.btype = new_node->ondisk->type;
1779         parent_elm->internal.subtree_offset = new_node->node_offset;
1780         ++ondisk->count;
1781
1782         /*
1783          * The children of new_node need their parent pointer set to new_node.
1784          * The children have already been locked by
1785          * hammer_btree_lock_children().
1786          */
1787         for (i = 0; i < new_node->ondisk->count; ++i) {
1788                 elm = &new_node->ondisk->elms[i];
1789                 error = btree_set_parent(new_node, elm);
1790                 if (error) {
1791                         panic("btree_split_internal: btree-fixup problem");
1792                 }
1793         }
1794
1795         /*
1796          * The cluster's root pointer may have to be updated.
1797          */
1798         if (made_root) {
1799                 hammer_modify_cluster(node->cluster);
1800                 node->cluster->ondisk->clu_btree_root = parent->node_offset;
1801                 node->ondisk->parent = parent->node_offset;
1802                 if (cursor->parent) {
1803                         hammer_unlock(&cursor->parent->lock);
1804                         hammer_rel_node(cursor->parent);
1805                 }
1806                 cursor->parent = parent;        /* lock'd and ref'd */
1807         }
1808
1809
1810         /*
1811          * Ok, now adjust the cursor depending on which element the original
1812          * index was pointing at.  If we are >= the split point the push node
1813          * is now in the new node.
1814          *
1815          * NOTE: If we are at the split point itself we cannot stay with the
1816          * original node because the push index will point at the right-hand
1817          * boundary, which is illegal.
1818          *
1819          * NOTE: The cursor's parent or parent_index must be adjusted for
1820          * the case where a new parent (new root) was created, and the case
1821          * where the cursor is now pointing at the split node.
1822          */
1823         if (cursor->index >= split) {
1824                 cursor->parent_index = parent_index + 1;
1825                 cursor->index -= split;
1826                 hammer_unlock(&cursor->node->lock);
1827                 hammer_rel_node(cursor->node);
1828                 cursor->node = new_node;        /* locked and ref'd */
1829         } else {
1830                 cursor->parent_index = parent_index;
1831                 hammer_unlock(&new_node->lock);
1832                 hammer_rel_node(new_node);
1833         }
1834
1835         /*
1836          * Fixup left and right bounds
1837          */
1838         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1839         cursor->left_bound = &parent_elm[0].internal.base;
1840         cursor->right_bound = &parent_elm[1].internal.base;
1841         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1842                  &cursor->node->ondisk->elms[0].internal.base) <= 0);
1843         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1844                  &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1845
1846 done:
1847         hammer_btree_unlock_children(&locklist);
1848         hammer_cursor_downgrade(cursor);
1849         return (error);
1850 }
1851
1852 /*
1853  * Same as the above, but splits a full leaf node.
1854  *
1855  * This function
1856  */
1857 static
1858 int
1859 btree_split_leaf(hammer_cursor_t cursor)
1860 {
1861         hammer_node_ondisk_t ondisk;
1862         hammer_node_t parent;
1863         hammer_node_t leaf;
1864         hammer_node_t new_leaf;
1865         hammer_btree_elm_t elm;
1866         hammer_btree_elm_t parent_elm;
1867         hammer_base_elm_t mid_boundary;
1868         hammer_node_locklist_t locklist = NULL;
1869         int parent_index;
1870         int made_root;
1871         int split;
1872         int error;
1873         int i;
1874         const size_t esize = sizeof(*elm);
1875
1876         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1877                 return(error);
1878         if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1879                 error = hammer_btree_lock_children(cursor, &locklist);
1880                 if (error)
1881                         goto done;
1882         }
1883
1884         /*
1885          * Calculate the split point.  If the insertion point will be on
1886          * the left-hand side adjust the split point to give the right
1887          * hand side one additional node.
1888          *
1889          * Spikes are made up of two leaf elements which cannot be
1890          * safely split.
1891          */
1892         leaf = cursor->node;
1893         ondisk = leaf->ondisk;
1894         split = (ondisk->count + 1) / 2;
1895         if (cursor->index <= split)
1896                 --split;
1897         error = 0;
1898
1899         elm = &ondisk->elms[split];
1900         if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1901                 KKASSERT(split &&
1902                         elm[-1].leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG);
1903                 --split;
1904         }
1905
1906         /*
1907          * If we are at the root of the tree, create a new root node with
1908          * 1 element and split normally.  Avoid making major modifications
1909          * until we know the whole operation will work.
1910          */
1911         if (ondisk->parent == 0) {
1912                 parent = hammer_alloc_btree(leaf->cluster, &error);
1913                 if (parent == NULL)
1914                         goto done;
1915                 hammer_lock_ex(&parent->lock);
1916                 hammer_modify_node(parent);
1917                 ondisk = parent->ondisk;
1918                 ondisk->count = 1;
1919                 ondisk->parent = 0;
1920                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1921                 ondisk->elms[0].base = leaf->cluster->clu_btree_beg;
1922                 ondisk->elms[0].base.btype = leaf->ondisk->type;
1923                 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1924                 ondisk->elms[1].base = leaf->cluster->clu_btree_end;
1925                 /* ondisk->elms[1].base.btype = not used */
1926                 made_root = 1;
1927                 parent_index = 0;       /* insertion point in parent */
1928         } else {
1929                 made_root = 0;
1930                 parent = cursor->parent;
1931                 parent_index = cursor->parent_index;
1932                 KKASSERT(parent->cluster == leaf->cluster);
1933         }
1934
1935         /*
1936          * Split leaf into new_leaf at the split point.  Select a separator
1937          * value in-between the two leafs but with a bent towards the right
1938          * leaf since comparisons use an 'elm >= separator' inequality.
1939          *
1940          *  L L L L L L L L
1941          *
1942          *       x x P x x
1943          *        s S S s
1944          *         /   \
1945          *  L L L L     L L L L
1946          */
1947         new_leaf = hammer_alloc_btree(leaf->cluster, &error);
1948         if (new_leaf == NULL) {
1949                 if (made_root) {
1950                         hammer_unlock(&parent->lock);
1951                         parent->flags |= HAMMER_NODE_DELETED;
1952                         hammer_rel_node(parent);
1953                 }
1954                 goto done;
1955         }
1956         hammer_lock_ex(&new_leaf->lock);
1957
1958         /*
1959          * Create the new node.  P (elm) become the left-hand boundary in the
1960          * new node.  Copy the right-hand boundary as well.
1961          */
1962         hammer_modify_node(leaf);
1963         hammer_modify_node(new_leaf);
1964         ondisk = leaf->ondisk;
1965         elm = &ondisk->elms[split];
1966         bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1967         new_leaf->ondisk->count = ondisk->count - split;
1968         new_leaf->ondisk->parent = parent->node_offset;
1969         new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1970         KKASSERT(ondisk->type == new_leaf->ondisk->type);
1971
1972         /*
1973          * Cleanup the original node.  Because this is a leaf node and
1974          * leaf nodes do not have a right-hand boundary, there
1975          * aren't any special edge cases to clean up.  We just fixup the
1976          * count.
1977          */
1978         ondisk->count = split;
1979
1980         /*
1981          * Insert the separator into the parent, fixup the parent's
1982          * reference to the original node, and reference the new node.
1983          * The separator is P.
1984          *
1985          * Remember that base.count does not include the right-hand boundary.
1986          * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1987          */
1988         hammer_modify_node(parent);
1989         ondisk = parent->ondisk;
1990         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1991         parent_elm = &ondisk->elms[parent_index+1];
1992         bcopy(parent_elm, parent_elm + 1,
1993               (ondisk->count - parent_index) * esize);
1994
1995         /*
1996          * Create the separator.  XXX At the moment use exactly the
1997          * right-hand element if this is a recovery operation in order
1998          * to guarantee that it does not bisect the spike elements in a
1999          * later call to hammer_btree_insert_cluster().
2000          */
2001         if (cursor->flags & HAMMER_CURSOR_RECOVER) {
2002                 parent_elm->base = elm[0].base;
2003         } else {
2004                 hammer_make_separator(&elm[-1].base, &elm[0].base,
2005                                       &parent_elm->base);
2006         }
2007         parent_elm->internal.base.btype = new_leaf->ondisk->type;
2008         parent_elm->internal.subtree_offset = new_leaf->node_offset;
2009         mid_boundary = &parent_elm->base;
2010         ++ondisk->count;
2011
2012         /*
2013          * The children of new_leaf need their parent pointer set to new_leaf.
2014          * The children have already been locked by btree_lock_children().
2015          *
2016          * The leaf's elements are either TYPE_RECORD or TYPE_SPIKE_*.  Only
2017          * elements of BTREE_TYPE_SPIKE_END really requires any action.
2018          */
2019         for (i = 0; i < new_leaf->ondisk->count; ++i) {
2020                 elm = &new_leaf->ondisk->elms[i];
2021                 error = btree_set_parent(new_leaf, elm);
2022                 if (error) {
2023                         panic("btree_split_internal: btree-fixup problem");
2024                 }
2025         }
2026
2027         /*
2028          * The cluster's root pointer may have to be updated.
2029          */
2030         if (made_root) {
2031                 hammer_modify_cluster(leaf->cluster);
2032                 leaf->cluster->ondisk->clu_btree_root = parent->node_offset;
2033                 leaf->ondisk->parent = parent->node_offset;
2034                 if (cursor->parent) {
2035                         hammer_unlock(&cursor->parent->lock);
2036                         hammer_rel_node(cursor->parent);
2037                 }
2038                 cursor->parent = parent;        /* lock'd and ref'd */
2039         }
2040
2041         /*
2042          * Ok, now adjust the cursor depending on which element the original
2043          * index was pointing at.  If we are >= the split point the push node
2044          * is now in the new node.
2045          *
2046          * NOTE: If we are at the split point itself we need to select the
2047          * old or new node based on where key_beg's insertion point will be.
2048          * If we pick the wrong side the inserted element will wind up in
2049          * the wrong leaf node and outside that node's bounds.
2050          */
2051         if (cursor->index > split ||
2052             (cursor->index == split &&
2053              hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
2054                 cursor->parent_index = parent_index + 1;
2055                 cursor->index -= split;
2056                 hammer_unlock(&cursor->node->lock);
2057                 hammer_rel_node(cursor->node);
2058                 cursor->node = new_leaf;
2059         } else {
2060                 cursor->parent_index = parent_index;
2061                 hammer_unlock(&new_leaf->lock);
2062                 hammer_rel_node(new_leaf);
2063         }
2064
2065         /*
2066          * Fixup left and right bounds
2067          */
2068         parent_elm = &parent->ondisk->elms[cursor->parent_index];
2069         cursor->left_bound = &parent_elm[0].internal.base;
2070         cursor->right_bound = &parent_elm[1].internal.base;
2071
2072         /*
2073          * Note: The right assertion is typically > 0, but if the last element
2074          * is a SPIKE_END it can be == 0 because the spike-end is non-inclusive
2075          * of the range being spiked.
2076          *
2077          * This may seem a bit odd but it works.
2078          */
2079         KKASSERT(hammer_btree_cmp(cursor->left_bound,
2080                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
2081         KKASSERT(hammer_btree_cmp(cursor->right_bound,
2082                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) >= 0);
2083
2084 done:
2085         hammer_btree_unlock_children(&locklist);
2086         hammer_cursor_downgrade(cursor);
2087         return (error);
2088 }
2089
2090 /*
2091  * Recursively correct the right-hand boundary's create_tid to (tid) as
2092  * long as the rest of the key matches.  We have to recurse upward in
2093  * the tree as well as down the left side of each parent's right node.
2094  *
2095  * Return EDEADLK if we were only partially successful, forcing the caller
2096  * to try again.  The original cursor is not modified.  This routine can
2097  * also fail with EDEADLK if it is forced to throw away a portion of its
2098  * record history.
2099  *
2100  * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2101  */
2102 struct hammer_rhb {
2103         TAILQ_ENTRY(hammer_rhb) entry;
2104         hammer_node_t   node;
2105         int             index;
2106 };
2107
2108 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2109
2110 int
2111 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2112 {
2113         struct hammer_rhb_list rhb_list;
2114         hammer_base_elm_t elm;
2115         hammer_node_t orig_node;
2116         struct hammer_rhb *rhb;
2117         int orig_index;
2118         int error;
2119
2120         TAILQ_INIT(&rhb_list);
2121
2122         /*
2123          * Save our position so we can restore it on return.  This also
2124          * gives us a stable 'elm'.
2125          */
2126         orig_node = cursor->node;
2127         hammer_ref_node(orig_node);
2128         hammer_lock_sh(&orig_node->lock);
2129         orig_index = cursor->index;
2130         elm = &orig_node->ondisk->elms[orig_index].base;
2131
2132         /*
2133          * Now build a list of parents going up, allocating a rhb
2134          * structure for each one.
2135          */
2136         while (cursor->parent) {
2137                 /*
2138                  * Stop if we no longer have any right-bounds to fix up
2139                  */
2140                 if (elm->obj_id != cursor->right_bound->obj_id ||
2141                     elm->rec_type != cursor->right_bound->rec_type ||
2142                     elm->key != cursor->right_bound->key) {
2143                         break;
2144                 }
2145
2146                 /*
2147                  * Stop if the right-hand bound's create_tid does not
2148                  * need to be corrected.   Note that if the parent is
2149                  * a cluster the bound is pointing at the actual bound
2150                  * in the cluster header, not the SPIKE_END element in
2151                  * the parent cluster, so we don't have to worry about
2152                  * the fact that SPIKE_END is range-inclusive.
2153                  */
2154                 if (cursor->right_bound->create_tid >= tid)
2155                         break;
2156
2157                 KKASSERT(cursor->parent->ondisk->elms[cursor->parent_index].base.btype != HAMMER_BTREE_TYPE_SPIKE_BEG);
2158
2159                 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
2160                 rhb->node = cursor->parent;
2161                 rhb->index = cursor->parent_index;
2162                 hammer_ref_node(rhb->node);
2163                 hammer_lock_sh(&rhb->node->lock);
2164                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2165
2166                 hammer_cursor_up(cursor);
2167         }
2168
2169         /*
2170          * now safely adjust the right hand bound for each rhb.  This may
2171          * also require taking the right side of the tree and iterating down
2172          * ITS left side.
2173          */
2174         error = 0;
2175         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2176                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2177                 kprintf("CORRECT RHB %d:%d:%08x index %d type=%c\n",
2178                         rhb->node->cluster->volume->vol_no,
2179                         rhb->node->cluster->clu_no, rhb->node->node_offset,
2180                         rhb->index, cursor->node->ondisk->type);
2181                 if (error)
2182                         break;
2183                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2184                 hammer_unlock(&rhb->node->lock);
2185                 hammer_rel_node(rhb->node);
2186                 kfree(rhb, M_HAMMER);
2187
2188                 switch (cursor->node->ondisk->type) {
2189                 case HAMMER_BTREE_TYPE_INTERNAL:
2190                         /*
2191                          * Right-boundary for parent at internal node
2192                          * is one element to the right of the element whos
2193                          * right boundary needs adjusting.  We must then
2194                          * traverse down the left side correcting any left
2195                          * bounds (which may now be too far to the left).
2196                          */
2197                         ++cursor->index;
2198                         error = hammer_btree_correct_lhb(cursor, tid);
2199                         break;
2200                 case HAMMER_BTREE_TYPE_LEAF:
2201                         /*
2202                          * Right-boundary for parent at leaf node.  Both
2203                          * the SPIKE_END and the cluster header must be
2204                          * corrected, but we don't have to traverse down
2205                          * (there's nothing TO traverse down other then what
2206                          * we've already recorded).
2207                          *
2208                          * The SPIKE_END is range-inclusive.
2209                          */
2210                         error = hammer_cursor_down(cursor);
2211                         if (error == 0)
2212                                 error = hammer_lock_upgrade(&cursor->parent->lock);
2213                         if (error == 0) {
2214                                 kprintf("hammer_btree_correct_rhb-X @%d:%d:%08x\n",
2215                                         cursor->parent->cluster->volume->vol_no,
2216                                         cursor->parent->cluster->clu_no,
2217                                         cursor->parent->node_offset);
2218                                 hammer_modify_node(cursor->parent);
2219                                 elm = &cursor->parent->ondisk->elms[cursor->parent_index].base;
2220                                 KKASSERT(elm->btype == HAMMER_BTREE_TYPE_SPIKE_END);
2221                                 elm->create_tid = tid - 1;
2222                                 hammer_modify_cluster(cursor->node->cluster);
2223                                 cursor->node->cluster->ondisk->clu_btree_end.create_tid = tid;
2224                                 cursor->node->cluster->clu_btree_end.create_tid = tid;
2225                         }
2226                         break;
2227                 default:
2228                         panic("hammer_btree_correct_rhb(): Bad node type");
2229                         error = EINVAL;
2230                         break;
2231                 }
2232         }
2233
2234         /*
2235          * Cleanup
2236          */
2237         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2238                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2239                 hammer_unlock(&rhb->node->lock);
2240                 hammer_rel_node(rhb->node);
2241                 kfree(rhb, M_HAMMER);
2242         }
2243         error = hammer_cursor_seek(cursor, orig_node, orig_index);
2244         hammer_unlock(&orig_node->lock);
2245         hammer_rel_node(orig_node);
2246         return (error);
2247 }
2248
2249 /*
2250  * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2251  * bound going downward starting at the current cursor position.
2252  *
2253  * This function does not restore the cursor after use.
2254  */
2255 int
2256 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2257 {
2258         struct hammer_rhb_list rhb_list;
2259         hammer_base_elm_t elm;
2260         hammer_base_elm_t cmp;
2261         struct hammer_rhb *rhb;
2262         int error;
2263
2264         TAILQ_INIT(&rhb_list);
2265
2266         cmp = &cursor->node->ondisk->elms[cursor->index].base;
2267
2268         /*
2269          * Record the node and traverse down the left-hand side for all
2270          * matching records needing a boundary correction.
2271          */
2272         error = 0;
2273         for (;;) {
2274                 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
2275                 rhb->node = cursor->node;
2276                 rhb->index = cursor->index;
2277                 hammer_ref_node(rhb->node);
2278                 hammer_lock_sh(&rhb->node->lock);
2279                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2280
2281                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2282                         /*
2283                          * Nothing to traverse down if we are at the right
2284                          * boundary of an internal node.
2285                          */
2286                         if (cursor->index == cursor->node->ondisk->count)
2287                                 break;
2288                 } else {
2289                         elm = &cursor->node->ondisk->elms[cursor->index].base;
2290                         if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2291                                 break;
2292                         KKASSERT(elm->btype == HAMMER_BTREE_TYPE_SPIKE_BEG);
2293                 }
2294                 error = hammer_cursor_down(cursor);
2295                 if (error)
2296                         break;
2297
2298                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2299                 if (elm->obj_id != cmp->obj_id ||
2300                     elm->rec_type != cmp->rec_type ||
2301                     elm->key != cmp->key) {
2302                         break;
2303                 }
2304                 if (elm->create_tid >= tid)
2305                         break;
2306
2307         }
2308
2309         /*
2310          * Now we can safely adjust the left-hand boundary from the bottom-up.
2311          * The last element we remove from the list is the caller's right hand
2312          * boundary, which must also be adjusted.
2313          */
2314         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2315                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2316                 if (error)
2317                         break;
2318                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2319                 hammer_unlock(&rhb->node->lock);
2320                 hammer_rel_node(rhb->node);
2321                 kfree(rhb, M_HAMMER);
2322
2323                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2324                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2325                         kprintf("hammer_btree_correct_lhb-I @%d:%d:%08x @%d\n",
2326                                 cursor->node->cluster->volume->vol_no,
2327                                 cursor->node->cluster->clu_no,
2328                                 cursor->node->node_offset, cursor->index);
2329                         hammer_modify_node(cursor->node);
2330                         elm->create_tid = tid;
2331                 } else if (elm->btype == HAMMER_BTREE_TYPE_SPIKE_BEG) {
2332                         /*
2333                          * SPIKE_BEG, also correct cluster header.  Occurs
2334                          * only while we are traversing the left-hand
2335                          * boundary.
2336                          */
2337                         kprintf("hammer_btree_correct_lhb-B @%d:%d:%08x\n",
2338                                 cursor->node->cluster->volume->vol_no,
2339                                 cursor->node->cluster->clu_no,
2340                                 cursor->node->node_offset);
2341                         hammer_modify_node(cursor->node);
2342                         elm->create_tid = tid;
2343
2344                         /*
2345                          * We can only cursor down through SPIKE_END.
2346                          */
2347                         ++cursor->index;
2348                         error = hammer_cursor_down(cursor);
2349                         if (error == 0)
2350                                 error = hammer_lock_upgrade(&cursor->parent->lock);
2351                         if (error == 0) {
2352                                 hammer_modify_node(cursor->parent);
2353                                 elm = &cursor->parent->ondisk->elms[cursor->parent_index - 1].base;
2354                                 KKASSERT(elm->btype == HAMMER_BTREE_TYPE_SPIKE_BEG);
2355                                 elm->create_tid = tid;
2356                                 hammer_modify_cluster(cursor->node->cluster);
2357                                 cursor->node->cluster->ondisk->clu_btree_end.create_tid = tid;
2358                                 cursor->node->cluster->clu_btree_end.create_tid = tid;
2359                         }
2360                 } else {
2361                         panic("hammer_btree_correct_lhb(): Bad element type");
2362                 }
2363         }
2364
2365         /*
2366          * Cleanup
2367          */
2368         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2369                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2370                 hammer_unlock(&rhb->node->lock);
2371                 hammer_rel_node(rhb->node);
2372                 kfree(rhb, M_HAMMER);
2373         }
2374         return (error);
2375 }
2376
2377 /*
2378  * Attempt to remove the empty B-Tree node at (cursor->node).  Returns 0
2379  * on success, EAGAIN if we could not acquire the necessary locks, or some
2380  * other error.  This node can be a leaf node or an internal node.
2381  *
2382  * On return the cursor may end up pointing at an internal node, suitable
2383  * for further iteration but not for an immediate insertion or deletion.
2384  *
2385  * cursor->node may be an internal node or a leaf node.
2386  *
2387  * NOTE: If cursor->node has one element it is the parent trying to delete
2388  * that element, make sure cursor->index is properly adjusted on success.
2389  */
2390 int
2391 btree_remove(hammer_cursor_t cursor)
2392 {
2393         hammer_node_ondisk_t ondisk;
2394         hammer_btree_elm_t elm;
2395         hammer_node_t node;
2396         hammer_node_t save;
2397         hammer_node_t parent;
2398         const int esize = sizeof(*elm);
2399         int error;
2400
2401         /*
2402          * If we are at the root of the cluster we must be able to
2403          * successfully delete the HAMMER_BTREE_SPIKE_* leaf elements in
2404          * the parent in order to be able to destroy the cluster.
2405          */
2406         node = cursor->node;
2407
2408         if (node->ondisk->parent == 0) {
2409                 hammer_modify_node(node);
2410                 ondisk = node->ondisk;
2411                 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2412                 ondisk->count = 0;
2413                 cursor->index = 0;
2414                 error = 0;
2415
2416                 /*
2417                  * When trying to delete a cluster we need to exclusively
2418                  * lock the cluster root, its parent (leaf in parent cluster),
2419                  * AND the parent of that leaf if it's going to be empty,
2420                  * because we can't leave around an empty leaf.
2421                  *
2422                  * XXX this is messy due to potentially recursive locks.
2423                  * downgrade the cursor, get a second shared lock on the
2424                  * node that cannot deadlock because we only own shared locks
2425                  * then, cursor-up, and re-upgrade everything.  If the
2426                  * upgrades EDEADLK then don't try to remove the cluster
2427                  * at this time.
2428                  */
2429                 if ((parent = cursor->parent) != NULL) {
2430                         hammer_cursor_downgrade(cursor);
2431                         save = node;
2432                         hammer_ref_node(save);
2433                         hammer_lock_sh(&save->lock);
2434
2435                         /*
2436                          * After the cursor up save has the empty root node
2437                          * of the target cluster to be deleted, cursor->node
2438                          * is at the leaf containing the spikes, and
2439                          * cursor->parent is the parent of that leaf.
2440                          *
2441                          * cursor->node and cursor->parent are both in the
2442                          * parent cluster of the cluster being deleted.
2443                          */
2444                         error = hammer_cursor_up(cursor);
2445
2446                         if (error == 0)
2447                                 error = hammer_cursor_upgrade(cursor);
2448                         if (error == 0)
2449                                 error = hammer_lock_upgrade(&save->lock);
2450
2451                         if (error) {
2452                                 /* may be EDEADLK */
2453                                 kprintf("BTREE_REMOVE: Cannot delete cluster\n");
2454                                 Debugger("BTREE_REMOVE");
2455                         } else {
2456                                 /*
2457                                  * cursor->node is now the leaf in the parent
2458                                  * cluster containing the spike elements.
2459                                  *
2460                                  * The cursor should be pointing at the
2461                                  * SPIKE_END element.
2462                                  *
2463                                  * Remove the spike elements and recurse
2464                                  * if the leaf becomes empty.
2465                                  */
2466                                 node = cursor->node;
2467                                 hammer_modify_node(node);
2468                                 ondisk = node->ondisk;
2469                                 KKASSERT(cursor->index > 0);
2470                                 --cursor->index;
2471                                 elm = &ondisk->elms[cursor->index];
2472                                 KKASSERT(elm[0].leaf.base.btype ==
2473                                          HAMMER_BTREE_TYPE_SPIKE_BEG);
2474                                 KKASSERT(elm[1].leaf.base.btype ==
2475                                          HAMMER_BTREE_TYPE_SPIKE_END);
2476
2477                                 /*
2478                                  * Ok, remove it and the underlying record.
2479                                  */
2480                                 hammer_free_record(node->cluster,
2481                                                    elm->leaf.rec_offset,
2482                                                    HAMMER_RECTYPE_CLUSTER);
2483                                 bcopy(elm + 2, elm, (ondisk->count -
2484                                                     cursor->index - 2) * esize);
2485                                 ondisk->count -= 2;
2486                                 save->flags |= HAMMER_NODE_DELETED;
2487                                 save->cluster->flags |= HAMMER_CLUSTER_DELETED;
2488                                 hammer_flush_node(save);
2489                                 hammer_unlock(&save->lock);
2490                                 hammer_rel_node(save);
2491                                 if (ondisk->count == 0)
2492                                         error = EAGAIN;
2493                         }
2494                 }
2495                 return(error);
2496         }
2497
2498         /*
2499          * Zero-out the parent's reference to the child and flag the
2500          * child for destruction.  This ensures that the child is not
2501          * reused while other references to it exist.
2502          */
2503         parent = cursor->parent;
2504         hammer_modify_node(parent);
2505         ondisk = parent->ondisk;
2506         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2507         elm = &ondisk->elms[cursor->parent_index];
2508         KKASSERT(elm->internal.subtree_offset == node->node_offset);
2509         elm->internal.subtree_offset = 0;
2510
2511         hammer_flush_node(node);
2512         node->flags |= HAMMER_NODE_DELETED;
2513
2514         /*
2515          * If the parent would otherwise not become empty we can physically
2516          * remove the zero'd element.  Note however that in order to
2517          * guarentee a valid cursor we still need to be able to cursor up
2518          * because we no longer have a node.
2519          *
2520          * This collapse will change the parent's boundary elements, making
2521          * them wider.  The new boundaries are recursively corrected in
2522          * btree_search().
2523          *
2524          * XXX we can theoretically recalculate the midpoint but there isn't
2525          * much of a reason to do it.
2526          */
2527         error = hammer_cursor_up(cursor);
2528         if (error == 0)
2529                 error = hammer_cursor_upgrade(cursor);
2530
2531         if (error) {
2532                 kprintf("BTREE_REMOVE: Cannot lock parent, skipping\n");
2533                 Debugger("BTREE_REMOVE");
2534                 return (0);
2535         }
2536
2537         /*
2538          * Remove the internal element from the parent.  The bcopy must
2539          * include the right boundary element.
2540          */
2541         KKASSERT(parent == cursor->node && ondisk == parent->ondisk);
2542         node = parent;
2543         parent = NULL;
2544         /* ondisk is node's ondisk */
2545         /* elm is node's element */
2546
2547         /*
2548          * Remove the internal element that we zero'd out.  Tell the caller
2549          * to loop if it hits zero (to try to avoid eating up precious kernel
2550          * stack).
2551          */
2552         KKASSERT(ondisk->count > 0);
2553         bcopy(&elm[1], &elm[0], (ondisk->count - cursor->index) * esize);
2554         --ondisk->count;
2555         if (ondisk->count == 0)
2556                 error = EAGAIN;
2557         return(error);
2558 }
2559
2560 /*
2561  * Attempt to remove the deleted internal element at the current cursor
2562  * position.  If we are unable to remove the element we return EDEADLK.
2563  *
2564  * If the current internal node becomes empty we delete it in the parent
2565  * and cursor up, looping until we finish or we deadlock.
2566  *
2567  * On return, if successful, the cursor will be pointing at the next
2568  * iterative position in the B-Tree.  If unsuccessful the cursor will be
2569  * pointing at the last deleted internal element that could not be
2570  * removed.
2571  */
2572 static
2573 int
2574 btree_remove_deleted_element(hammer_cursor_t cursor)
2575 {
2576         hammer_node_t node;
2577         hammer_btree_elm_t elm;
2578         int error;
2579
2580         if ((error = hammer_cursor_upgrade(cursor)) != 0)
2581                 return(error);
2582         node = cursor->node;
2583         elm = &node->ondisk->elms[cursor->index];
2584         if (elm->internal.subtree_offset == 0) {
2585                 do {
2586                         error = btree_remove(cursor);
2587                         kprintf("BTREE REMOVE DELETED ELEMENT %d\n", error);
2588                 } while (error == EAGAIN);
2589         }
2590         return(error);
2591 }
2592
2593 /*
2594  * The element (elm) has been moved to a new internal node (node).
2595  *
2596  * If the element represents a pointer to an internal node that node's
2597  * parent must be adjusted to the element's new location.
2598  *
2599  * If the element represents a spike the target cluster's header must
2600  * be adjusted to point to the element's new location.  This only
2601  * applies to HAMMER_SPIKE_END.
2602  *
2603  * GET_CLUSTER_NORECOVER must be used to avoid a recovery recursion during
2604  * the rebuild of the recovery cluster's B-Tree, which can blow the kernel
2605  * stack.
2606  *
2607  * XXX deadlock potential here with our exclusive locks
2608  */
2609 static
2610 int
2611 btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm)
2612 {
2613         hammer_volume_t volume;
2614         hammer_cluster_t cluster;
2615         hammer_node_t child;
2616         int error;
2617
2618         error = 0;
2619
2620         switch(elm->base.btype) {
2621         case HAMMER_BTREE_TYPE_INTERNAL:
2622         case HAMMER_BTREE_TYPE_LEAF:
2623                 child = hammer_get_node(node->cluster,
2624                                         elm->internal.subtree_offset, &error);
2625                 if (error == 0) {
2626                         hammer_modify_node(child);
2627                         child->ondisk->parent = node->node_offset;
2628                         hammer_rel_node(child);
2629                 }
2630                 break;
2631         case HAMMER_BTREE_TYPE_SPIKE_END:
2632                 volume = hammer_get_volume(node->cluster->volume->hmp,
2633                                            elm->leaf.spike_vol_no, &error);
2634                 if (error)
2635                         break;
2636                 cluster = hammer_get_cluster(volume, elm->leaf.spike_clu_no,
2637                                              &error, GET_CLUSTER_NORECOVER);
2638                 hammer_rel_volume(volume, 0);
2639                 if (error)
2640                         break;
2641                 hammer_modify_cluster(cluster);
2642                 cluster->ondisk->clu_btree_parent_offset = node->node_offset;
2643                 KKASSERT(cluster->ondisk->clu_btree_parent_clu_no ==
2644                          node->cluster->clu_no);
2645                 KKASSERT(cluster->ondisk->clu_btree_parent_vol_no ==
2646                          node->cluster->volume->vol_no);
2647                 hammer_rel_cluster(cluster, 0);
2648                 break;
2649         default:
2650                 break;
2651         }
2652         return(error);
2653 }
2654
2655 /*
2656  * Exclusively lock all the children of node.  This is used by the split
2657  * code to prevent anyone from accessing the children of a cursor node
2658  * while we fix-up its parent offset.
2659  *
2660  * If we don't lock the children we can really mess up cursors which block
2661  * trying to cursor-up into our node.
2662  *
2663  * WARNING: Cannot be used when doing B-tree operations on a recovery
2664  * cluster because the target cluster may require recovery, resulting
2665  * in a deep recursion which blows the kernel stack.
2666  *
2667  * On failure EDEADLK (or some other error) is returned.  If a deadlock
2668  * error is returned the cursor is adjusted to block on termination.
2669  */
2670 int
2671 hammer_btree_lock_children(hammer_cursor_t cursor,
2672                            struct hammer_node_locklist **locklistp)
2673 {
2674         hammer_node_t node;
2675         hammer_node_locklist_t item;
2676         hammer_node_ondisk_t ondisk;
2677         hammer_btree_elm_t elm;
2678         hammer_volume_t volume;
2679         hammer_cluster_t cluster;
2680         hammer_node_t child;
2681         int error;
2682         int i;
2683
2684         node = cursor->node;
2685         ondisk = node->ondisk;
2686         error = 0;
2687         for (i = 0; error == 0 && i < ondisk->count; ++i) {
2688                 elm = &ondisk->elms[i];
2689
2690                 child = NULL;
2691                 switch(elm->base.btype) {
2692                 case HAMMER_BTREE_TYPE_INTERNAL:
2693                 case HAMMER_BTREE_TYPE_LEAF:
2694                         child = hammer_get_node(node->cluster,
2695                                                 elm->internal.subtree_offset,
2696                                                 &error);
2697                         break;
2698                 case HAMMER_BTREE_TYPE_SPIKE_END:
2699                         volume = hammer_get_volume(node->cluster->volume->hmp,
2700                                                    elm->leaf.spike_vol_no,
2701                                                    &error);
2702                         if (error)
2703                                 break;
2704                         cluster = hammer_get_cluster(volume,
2705                                                      elm->leaf.spike_clu_no,
2706                                                      &error,
2707                                                      0);
2708                         hammer_rel_volume(volume, 0);
2709                         if (error)
2710                                 break;
2711                         KKASSERT(cluster->ondisk->clu_btree_root != 0);
2712                         child = hammer_get_node(cluster,
2713                                                 cluster->ondisk->clu_btree_root,
2714                                                 &error);
2715                         hammer_rel_cluster(cluster, 0);
2716                         break;
2717                 default:
2718                         break;
2719                 }
2720                 if (child) {
2721                         if (hammer_lock_ex_try(&child->lock) != 0) {
2722                                 if (cursor->deadlk_node == NULL) {
2723                                         cursor->deadlk_node = node;
2724                                         hammer_ref_node(cursor->deadlk_node);
2725                                 }
2726                                 error = EDEADLK;
2727                         } else {
2728                                 item = kmalloc(sizeof(*item),
2729                                                 M_HAMMER, M_WAITOK);
2730                                 item->next = *locklistp;
2731                                 item->node = child;
2732                                 *locklistp = item;
2733                         }
2734                 }
2735         }
2736         if (error)
2737                 hammer_btree_unlock_children(locklistp);
2738         return(error);
2739 }
2740
2741
2742 /*
2743  * Release previously obtained node locks.
2744  */
2745 void
2746 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2747 {
2748         hammer_node_locklist_t item;
2749
2750         while ((item = *locklistp) != NULL) {
2751                 *locklistp = item->next;
2752                 hammer_unlock(&item->node->lock);
2753                 hammer_rel_node(item->node);
2754                 kfree(item, M_HAMMER);
2755         }
2756 }
2757
2758 /************************************************************************
2759  *                         MISCELLANIOUS SUPPORT                        *
2760  ************************************************************************/
2761
2762 /*
2763  * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2764  *
2765  * Note that for this particular function a return value of -1, 0, or +1
2766  * can denote a match if create_tid is otherwise discounted.  A create_tid
2767  * of zero is considered to be 'infinity' in comparisons.
2768  *
2769  * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2770  */
2771 int
2772 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2773 {
2774         if (key1->obj_id < key2->obj_id)
2775                 return(-4);
2776         if (key1->obj_id > key2->obj_id)
2777                 return(4);
2778
2779         if (key1->rec_type < key2->rec_type)
2780                 return(-3);
2781         if (key1->rec_type > key2->rec_type)
2782                 return(3);
2783
2784         if (key1->key < key2->key)
2785                 return(-2);
2786         if (key1->key > key2->key)
2787                 return(2);
2788
2789         /*
2790          * A create_tid of zero indicates a record which is undeletable
2791          * and must be considered to have a value of positive infinity.
2792          */
2793         if (key1->create_tid == 0) {
2794                 if (key2->create_tid == 0)
2795                         return(0);
2796                 return(1);
2797         }
2798         if (key2->create_tid == 0)
2799                 return(-1);
2800         if (key1->create_tid < key2->create_tid)
2801                 return(-1);
2802         if (key1->create_tid > key2->create_tid)
2803                 return(1);
2804         return(0);
2805 }
2806
2807 /*
2808  * Test a timestamp against an element to determine whether the
2809  * element is visible.  A timestamp of 0 means 'infinity'.
2810  */
2811 int
2812 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2813 {
2814         if (asof == 0) {
2815                 if (base->delete_tid)
2816                         return(1);
2817                 return(0);
2818         }
2819         if (asof < base->create_tid)
2820                 return(-1);
2821         if (base->delete_tid && asof >= base->delete_tid)
2822                 return(1);
2823         return(0);
2824 }
2825
2826 /*
2827  * Create a separator half way inbetween key1 and key2.  For fields just
2828  * one unit apart, the separator will match key2.  key1 is on the left-hand
2829  * side and key2 is on the right-hand side.
2830  *
2831  * create_tid has to be special cased because a value of 0 represents
2832  * infinity.
2833  */
2834 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2835         dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2836
2837 static void
2838 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2839                       hammer_base_elm_t dest)
2840 {
2841         bzero(dest, sizeof(*dest));
2842         MAKE_SEPARATOR(key1, key2, dest, obj_id);
2843         MAKE_SEPARATOR(key1, key2, dest, rec_type);
2844         MAKE_SEPARATOR(key1, key2, dest, key);
2845
2846         if (key1->obj_id == key2->obj_id &&
2847             key1->rec_type == key2->rec_type &&
2848             key1->key == key2->key) {
2849                 if (key1->create_tid == 0) {
2850                         /*
2851                          * Oops, a create_tid of 0 means 'infinity', so
2852                          * if everything matches this just isn't legal.
2853                          */
2854                         panic("key1->create_tid of 0 is impossible here");
2855                 } else if (key2->create_tid == 0) {
2856                         dest->create_tid = key1->create_tid + 1;
2857                 } else {
2858                         MAKE_SEPARATOR(key1, key2, dest, create_tid);
2859                 }
2860         } else {
2861                 dest->create_tid = 0;
2862         }
2863 }
2864
2865 #undef MAKE_SEPARATOR
2866
2867 /*
2868  * Return whether a generic internal or leaf node is full
2869  */
2870 static int
2871 btree_node_is_full(hammer_node_ondisk_t node)
2872 {
2873         switch(node->type) {
2874         case HAMMER_BTREE_TYPE_INTERNAL:
2875                 if (node->count == HAMMER_BTREE_INT_ELMS)
2876                         return(1);
2877                 break;
2878         case HAMMER_BTREE_TYPE_LEAF:
2879                 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2880                         return(1);
2881                 break;
2882         default:
2883                 panic("illegal btree subtype");
2884         }
2885         return(0);
2886 }
2887
2888 /*
2889  * Return whether a generic internal or leaf node is almost full.  This
2890  * routine is used as a helper for search insertions to guarentee at
2891  * least 2 available slots in the internal node(s) leading up to a leaf,
2892  * so hammer_btree_insert_cluster() will function properly.
2893  */
2894 static int
2895 btree_node_is_almost_full(hammer_node_ondisk_t node)
2896 {
2897         switch(node->type) {
2898         case HAMMER_BTREE_TYPE_INTERNAL:
2899                 if (node->count > HAMMER_BTREE_INT_ELMS - 2)
2900                         return(1);
2901                 break;
2902         case HAMMER_BTREE_TYPE_LEAF:
2903                 if (node->count > HAMMER_BTREE_LEAF_ELMS - 2)
2904                         return(1);
2905                 break;
2906         default:
2907                 panic("illegal btree subtype");
2908         }
2909         return(0);
2910 }
2911
2912 #if 0
2913 static int
2914 btree_max_elements(u_int8_t type)
2915 {
2916         if (type == HAMMER_BTREE_TYPE_LEAF)
2917                 return(HAMMER_BTREE_LEAF_ELMS);
2918         if (type == HAMMER_BTREE_TYPE_INTERNAL)
2919                 return(HAMMER_BTREE_INT_ELMS);
2920         panic("btree_max_elements: bad type %d\n", type);
2921 }
2922 #endif
2923
2924 void
2925 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2926 {
2927         hammer_btree_elm_t elm;
2928         int i;
2929
2930         kprintf("node %p count=%d parent=%d type=%c\n",
2931                 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2932
2933         /*
2934          * Dump both boundary elements if an internal node
2935          */
2936         if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2937                 for (i = 0; i <= ondisk->count; ++i) {
2938                         elm = &ondisk->elms[i];
2939                         hammer_print_btree_elm(elm, ondisk->type, i);
2940                 }
2941         } else {
2942                 for (i = 0; i < ondisk->count; ++i) {
2943                         elm = &ondisk->elms[i];
2944                         hammer_print_btree_elm(elm, ondisk->type, i);
2945                 }
2946         }
2947 }
2948
2949 void
2950 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2951 {
2952         kprintf("  %2d", i);
2953         kprintf("\tobj_id       = %016llx\n", elm->base.obj_id);
2954         kprintf("\tkey          = %016llx\n", elm->base.key);
2955         kprintf("\tcreate_tid   = %016llx\n", elm->base.create_tid);
2956         kprintf("\tdelete_tid   = %016llx\n", elm->base.delete_tid);
2957         kprintf("\trec_type     = %04x\n", elm->base.rec_type);
2958         kprintf("\tobj_type     = %02x\n", elm->base.obj_type);
2959         kprintf("\tbtype        = %02x (%c)\n",
2960                 elm->base.btype,
2961                 (elm->base.btype ? elm->base.btype : '?'));
2962
2963         switch(type) {
2964         case HAMMER_BTREE_TYPE_INTERNAL:
2965                 kprintf("\tsubtree_off  = %08x\n",
2966                         elm->internal.subtree_offset);
2967                 break;
2968         case HAMMER_BTREE_TYPE_SPIKE_BEG:
2969         case HAMMER_BTREE_TYPE_SPIKE_END:
2970                 kprintf("\tspike_clu_no = %d\n", elm->leaf.spike_clu_no);
2971                 kprintf("\tspike_vol_no = %d\n", elm->leaf.spike_vol_no);
2972                 break;
2973         case HAMMER_BTREE_TYPE_RECORD:
2974                 kprintf("\trec_offset   = %08x\n", elm->leaf.rec_offset);
2975                 kprintf("\tdata_offset  = %08x\n", elm->leaf.data_offset);
2976                 kprintf("\tdata_len     = %08x\n", elm->leaf.data_len);
2977                 kprintf("\tdata_crc     = %08x\n", elm->leaf.data_crc);
2978                 break;
2979         }
2980 }