1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * HAMMER B-Tree index 37 * 38 * HAMMER implements a modified B+Tree. In documentation this will 39 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree 40 * looks like a B+Tree (A B-Tree which stores its records only at the leafs 41 * of the tree), but adds two additional boundary elements which describe 42 * the left-most and right-most element a node is able to represent. In 43 * otherwords, we have boundary elements at the two ends of a B-Tree node 44 * with no valid sub-tree pointer for the right-most element. 45 * 46 * A B-Tree internal node looks like this: 47 * 48 * B N N N N N N B <-- boundary and internal elements 49 * S S S S S S S <-- subtree pointers 50 * 51 * A B-Tree leaf node basically looks like this: 52 * 53 * L L L L L L L L <-- leaf elemenets 54 * 55 * The radix for an internal node is 1 less then a leaf but we get a 56 * number of significant benefits for our troubles. 57 * The left-hand boundary (B in the left) is integrated into the first 58 * element so it doesn't require 2 elements to accomodate boundaries. 59 * 60 * The big benefit to using a B-Tree containing boundary information 61 * is that it is possible to cache pointers into the middle of the tree 62 * and not have to start searches, insertions, OR deletions at the root 63 * node. In particular, searches are able to progress in a definitive 64 * direction from any point in the tree without revisting nodes. This 65 * greatly improves the efficiency of many operations, most especially 66 * record appends. 67 * 68 * B-Trees also make the stacking of trees fairly straightforward. 69 * 70 * INSERTIONS: A search performed with the intention of doing 71 * an insert will guarantee that the terminal leaf node is not full by 72 * splitting full nodes. Splits occur top-down during the dive down the 73 * B-Tree. 74 * 75 * DELETIONS: A deletion makes no attempt to proactively balance the 76 * tree and will recursively remove nodes that become empty. If a 77 * deadlock occurs a deletion may not be able to remove an empty leaf. 78 * Deletions never allow internal nodes to become empty (that would blow 79 * up the boundaries). 80 */ 81 #include "hammer.h" 82 83 static int btree_search(hammer_cursor_t cursor, int flags); 84 static int btree_split_internal(hammer_cursor_t cursor); 85 static int btree_split_leaf(hammer_cursor_t cursor); 86 static int btree_remove(hammer_cursor_t cursor, int *ndelete); 87 static __inline int btree_node_is_full(hammer_node_ondisk_t node); 88 static __inline int btree_max_elements(u_int8_t type); 89 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor, 90 hammer_tid_t mirror_tid); 91 static void hammer_make_separator(hammer_base_elm_t key1, 92 hammer_base_elm_t key2, hammer_base_elm_t dest); 93 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor); 94 static __inline void hammer_debug_btree_elm(hammer_cursor_t cursor, 95 hammer_btree_elm_t elm, const char *s, int res); 96 static __inline void hammer_debug_btree_parent(hammer_cursor_t cursor, 97 const char *s); 98 99 /* 100 * Iterate records after a search. The cursor is iterated forwards past 101 * the current record until a record matching the key-range requirements 102 * is found. ENOENT is returned if the iteration goes past the ending 103 * key. 104 * 105 * The iteration is inclusive of key_beg and can be inclusive or exclusive 106 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set. 107 * 108 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid 109 * may be modified by B-Tree functions. 110 * 111 * cursor->key_beg may or may not be modified by this function during 112 * the iteration. XXX future - in case of an inverted lock we may have 113 * to reinitiate the lookup and set key_beg to properly pick up where we 114 * left off. 115 * 116 * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor 117 * was reverse indexed due to being moved to a parent while unlocked, 118 * and something else might have inserted an element outside the iteration 119 * range. When this case occurs the iterator just keeps iterating until 120 * it gets back into the iteration range (instead of asserting). 121 * 122 * NOTE! EDEADLK *CANNOT* be returned by this procedure. 123 */ 124 int 125 hammer_btree_iterate(hammer_cursor_t cursor) 126 { 127 hammer_node_ondisk_t node; 128 hammer_btree_elm_t elm; 129 hammer_mount_t hmp; 130 int error = 0; 131 int r; 132 int s; 133 134 /* 135 * Skip past the current record 136 */ 137 hmp = cursor->trans->hmp; 138 node = cursor->node->ondisk; 139 if (node == NULL) 140 return(ENOENT); 141 if (cursor->index < node->count && 142 (cursor->flags & HAMMER_CURSOR_ATEDISK)) { 143 ++cursor->index; 144 } 145 146 /* 147 * HAMMER can wind up being cpu-bound. 148 */ 149 if (++hmp->check_yield > hammer_yield_check) { 150 hmp->check_yield = 0; 151 lwkt_user_yield(); 152 } 153 154 155 /* 156 * Loop until an element is found or we are done. 157 */ 158 for (;;) { 159 /* 160 * We iterate up the tree and then index over one element 161 * while we are at the last element in the current node. 162 * 163 * If we are at the root of the filesystem, cursor_up 164 * returns ENOENT. 165 * 166 * XXX this could be optimized by storing the information in 167 * the parent reference. 168 * 169 * XXX we can lose the node lock temporarily, this could mess 170 * up our scan. 171 */ 172 ++hammer_stats_btree_iterations; 173 hammer_flusher_clean_loose_ios(hmp); 174 175 if (cursor->index == node->count) { 176 if (hammer_debug_btree) { 177 hkprintf("BRACKETU %016llx[%d] -> %016llx[%d] td=%p\n", 178 (long long)cursor->node->node_offset, 179 cursor->index, 180 (long long)(cursor->parent ? cursor->parent->node_offset : -1), 181 cursor->parent_index, 182 curthread); 183 } 184 KKASSERT(cursor->parent == NULL || 185 cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset); 186 error = hammer_cursor_up(cursor); 187 if (error) 188 break; 189 /* reload stale pointer */ 190 node = cursor->node->ondisk; 191 KKASSERT(cursor->index != node->count); 192 193 /* 194 * If we are reblocking we want to return internal 195 * nodes. Note that the internal node will be 196 * returned multiple times, on each upward recursion 197 * from its children. The caller selects which 198 * revisit it cares about (usually first or last only). 199 */ 200 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) { 201 cursor->flags |= HAMMER_CURSOR_ATEDISK; 202 return(0); 203 } 204 ++cursor->index; 205 continue; 206 } 207 208 /* 209 * Check internal or leaf element. Determine if the record 210 * at the cursor has gone beyond the end of our range. 211 * 212 * We recurse down through internal nodes. 213 */ 214 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 215 elm = &node->elms[cursor->index]; 216 217 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base); 218 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base); 219 if (hammer_debug_btree) { 220 hammer_debug_btree_elm(cursor, elm, "BRACKETL", r); 221 hammer_debug_btree_elm(cursor, elm + 1, "BRACKETR", s); 222 } 223 224 if (r < 0) { 225 error = ENOENT; 226 break; 227 } 228 if (r == 0 && (cursor->flags & 229 HAMMER_CURSOR_END_INCLUSIVE) == 0) { 230 error = ENOENT; 231 break; 232 } 233 234 /* 235 * Better not be zero 236 */ 237 KKASSERT(elm->internal.subtree_offset != 0); 238 239 if (s <= 0) { 240 /* 241 * If running the mirror filter see if we 242 * can skip one or more entire sub-trees. 243 * If we can we return the internal node 244 * and the caller processes the skipped 245 * range (see mirror_read). 246 */ 247 if (cursor->flags & 248 HAMMER_CURSOR_MIRROR_FILTERED) { 249 if (elm->internal.mirror_tid < 250 cursor->cmirror->mirror_tid) { 251 hammer_cursor_mirror_filter(cursor); 252 return(0); 253 } 254 } 255 } else { 256 /* 257 * Normally it would be impossible for the 258 * cursor to have gotten back-indexed, 259 * but it can happen if a node is deleted 260 * and the cursor is moved to its parent 261 * internal node. ITERATE_CHECK will be set. 262 */ 263 KKASSERT(cursor->flags & 264 HAMMER_CURSOR_ITERATE_CHECK); 265 hdkprintf("DEBUG: Caught parent seek " 266 "in internal iteration\n"); 267 } 268 269 error = hammer_cursor_down(cursor); 270 if (error) 271 break; 272 KKASSERT(cursor->index == 0); 273 /* reload stale pointer */ 274 node = cursor->node->ondisk; 275 continue; 276 } else { 277 elm = &node->elms[cursor->index]; 278 r = hammer_btree_cmp(&cursor->key_end, &elm->base); 279 if (hammer_debug_btree) { 280 hammer_debug_btree_elm(cursor, elm, "ELEMENT", r); 281 } 282 if (r < 0) { 283 error = ENOENT; 284 break; 285 } 286 287 /* 288 * We support both end-inclusive and 289 * end-exclusive searches. 290 */ 291 if (r == 0 && 292 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) { 293 error = ENOENT; 294 break; 295 } 296 297 /* 298 * If ITERATE_CHECK is set an unlocked cursor may 299 * have been moved to a parent and the iterate can 300 * happen upon elements that are not in the requested 301 * range. 302 */ 303 if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) { 304 s = hammer_btree_cmp(&cursor->key_beg, 305 &elm->base); 306 if (s > 0) { 307 hdkprintf("DEBUG: Caught parent seek " 308 "in leaf iteration\n"); 309 ++cursor->index; 310 continue; 311 } 312 } 313 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 314 315 /* 316 * Return the element 317 */ 318 switch(elm->leaf.base.btype) { 319 case HAMMER_BTREE_TYPE_RECORD: 320 if ((cursor->flags & HAMMER_CURSOR_ASOF) && 321 hammer_btree_chkts(cursor->asof, &elm->base)) { 322 ++cursor->index; 323 continue; 324 } 325 error = 0; 326 break; 327 default: 328 error = EINVAL; 329 break; 330 } 331 if (error) 332 break; 333 } 334 335 /* 336 * Return entry 337 */ 338 if (hammer_debug_btree) { 339 elm = &cursor->node->ondisk->elms[cursor->index]; 340 hammer_debug_btree_elm(cursor, elm, "ITERATE", 0xffff); 341 } 342 return(0); 343 } 344 return(error); 345 } 346 347 /* 348 * We hit an internal element that we could skip as part of a mirroring 349 * scan. Calculate the entire range being skipped. 350 * 351 * It is important to include any gaps between the parent's left_bound 352 * and the node's left_bound, and same goes for the right side. 353 */ 354 static void 355 hammer_cursor_mirror_filter(hammer_cursor_t cursor) 356 { 357 struct hammer_cmirror *cmirror; 358 hammer_node_ondisk_t ondisk; 359 hammer_btree_elm_t elm; 360 361 ondisk = cursor->node->ondisk; 362 cmirror = cursor->cmirror; 363 364 /* 365 * Calculate the skipped range 366 */ 367 elm = &ondisk->elms[cursor->index]; 368 if (cursor->index == 0) 369 cmirror->skip_beg = *cursor->left_bound; 370 else 371 cmirror->skip_beg = elm->internal.base; 372 while (cursor->index < ondisk->count) { 373 if (elm->internal.mirror_tid >= cmirror->mirror_tid) 374 break; 375 ++cursor->index; 376 ++elm; 377 } 378 if (cursor->index == ondisk->count) 379 cmirror->skip_end = *cursor->right_bound; 380 else 381 cmirror->skip_end = elm->internal.base; 382 383 /* 384 * clip the returned result. 385 */ 386 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0) 387 cmirror->skip_beg = cursor->key_beg; 388 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0) 389 cmirror->skip_end = cursor->key_end; 390 } 391 392 /* 393 * Iterate in the reverse direction. This is used by the pruning code to 394 * avoid overlapping records. 395 */ 396 int 397 hammer_btree_iterate_reverse(hammer_cursor_t cursor) 398 { 399 hammer_node_ondisk_t node; 400 hammer_btree_elm_t elm; 401 hammer_mount_t hmp; 402 int error = 0; 403 int r; 404 int s; 405 406 /* mirror filtering not supported for reverse iteration */ 407 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0); 408 409 /* 410 * Skip past the current record. For various reasons the cursor 411 * may end up set to -1 or set to point at the end of the current 412 * node. These cases must be addressed. 413 */ 414 node = cursor->node->ondisk; 415 if (node == NULL) 416 return(ENOENT); 417 if (cursor->index != -1 && 418 (cursor->flags & HAMMER_CURSOR_ATEDISK)) { 419 --cursor->index; 420 } 421 if (cursor->index == cursor->node->ondisk->count) 422 --cursor->index; 423 424 /* 425 * HAMMER can wind up being cpu-bound. 426 */ 427 hmp = cursor->trans->hmp; 428 if (++hmp->check_yield > hammer_yield_check) { 429 hmp->check_yield = 0; 430 lwkt_user_yield(); 431 } 432 433 /* 434 * Loop until an element is found or we are done. 435 */ 436 for (;;) { 437 ++hammer_stats_btree_iterations; 438 hammer_flusher_clean_loose_ios(hmp); 439 440 /* 441 * We iterate up the tree and then index over one element 442 * while we are at the last element in the current node. 443 */ 444 if (cursor->index == -1) { 445 error = hammer_cursor_up(cursor); 446 if (error) { 447 cursor->index = 0; /* sanity */ 448 break; 449 } 450 /* reload stale pointer */ 451 node = cursor->node->ondisk; 452 KKASSERT(cursor->index != node->count); 453 --cursor->index; 454 continue; 455 } 456 457 /* 458 * Check internal or leaf element. Determine if the record 459 * at the cursor has gone beyond the end of our range. 460 * 461 * We recurse down through internal nodes. 462 */ 463 KKASSERT(cursor->index != node->count); 464 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 465 elm = &node->elms[cursor->index]; 466 467 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base); 468 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base); 469 if (hammer_debug_btree) { 470 hammer_debug_btree_elm(cursor, elm, "BRACKETL", r); 471 hammer_debug_btree_elm(cursor, elm + 1, "BRACKETR", s); 472 } 473 474 if (s >= 0) { 475 error = ENOENT; 476 break; 477 } 478 479 /* 480 * It shouldn't be possible to be seeked past key_end, 481 * even if the cursor got moved to a parent. 482 */ 483 KKASSERT(r >= 0); 484 485 /* 486 * Better not be zero 487 */ 488 KKASSERT(elm->internal.subtree_offset != 0); 489 490 error = hammer_cursor_down(cursor); 491 if (error) 492 break; 493 KKASSERT(cursor->index == 0); 494 /* reload stale pointer */ 495 node = cursor->node->ondisk; 496 497 /* this can assign -1 if the leaf was empty */ 498 cursor->index = node->count - 1; 499 continue; 500 } else { 501 elm = &node->elms[cursor->index]; 502 s = hammer_btree_cmp(&cursor->key_beg, &elm->base); 503 if (hammer_debug_btree) { 504 hammer_debug_btree_elm(cursor, elm, "ELEMENTR", s); 505 } 506 if (s > 0) { 507 error = ENOENT; 508 break; 509 } 510 511 /* 512 * It shouldn't be possible to be seeked past key_end, 513 * even if the cursor got moved to a parent. 514 */ 515 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 516 517 /* 518 * Return the element 519 */ 520 switch(elm->leaf.base.btype) { 521 case HAMMER_BTREE_TYPE_RECORD: 522 if ((cursor->flags & HAMMER_CURSOR_ASOF) && 523 hammer_btree_chkts(cursor->asof, &elm->base)) { 524 --cursor->index; 525 continue; 526 } 527 error = 0; 528 break; 529 default: 530 error = EINVAL; 531 break; 532 } 533 if (error) 534 break; 535 } 536 537 /* 538 * Return entry 539 */ 540 if (hammer_debug_btree) { 541 elm = &cursor->node->ondisk->elms[cursor->index]; 542 hammer_debug_btree_elm(cursor, elm, "ITERATER", 0xffff); 543 } 544 return(0); 545 } 546 return(error); 547 } 548 549 /* 550 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry 551 * could not be found, EDEADLK if inserting and a retry is needed, and a 552 * fatal error otherwise. When retrying, the caller must terminate the 553 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting. 554 * 555 * The cursor is suitably positioned for a deletion on success, and suitably 556 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was 557 * specified. 558 * 559 * The cursor may begin anywhere, the search will traverse the tree in 560 * either direction to locate the requested element. 561 * 562 * Most of the logic implementing historical searches is handled here. We 563 * do an initial lookup with create_tid set to the asof TID. Due to the 564 * way records are laid out, a backwards iteration may be required if 565 * ENOENT is returned to locate the historical record. Here's the 566 * problem: 567 * 568 * create_tid: 10 15 20 569 * LEAF1 LEAF2 570 * records: (11) (18) 571 * 572 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse 573 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is 574 * not visible and thus causes ENOENT to be returned. We really need 575 * to check record 11 in LEAF1. If it also fails then the search fails 576 * (e.g. it might represent the range 11-16 and thus still not match our 577 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring 578 * further iterations. 579 * 580 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK 581 * and the cursor->create_check TID if an iteration might be needed. 582 * In the above example create_check would be set to 14. 583 */ 584 int 585 hammer_btree_lookup(hammer_cursor_t cursor) 586 { 587 int error; 588 589 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 590 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 || 591 cursor->trans->sync_lock_refs > 0); 592 ++hammer_stats_btree_lookups; 593 if (cursor->flags & HAMMER_CURSOR_ASOF) { 594 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0); 595 cursor->key_beg.create_tid = cursor->asof; 596 for (;;) { 597 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK; 598 error = btree_search(cursor, 0); 599 if (error != ENOENT || 600 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) { 601 /* 602 * Stop if no error. 603 * Stop if error other then ENOENT. 604 * Stop if ENOENT and not special case. 605 */ 606 break; 607 } 608 if (hammer_debug_btree) { 609 hkprintf("CREATE_CHECK %016llx\n", 610 (long long)cursor->create_check); 611 } 612 cursor->key_beg.create_tid = cursor->create_check; 613 /* loop */ 614 } 615 } else { 616 error = btree_search(cursor, 0); 617 } 618 if (error == 0) 619 error = hammer_btree_extract(cursor, cursor->flags); 620 return(error); 621 } 622 623 /* 624 * Execute the logic required to start an iteration. The first record 625 * located within the specified range is returned and iteration control 626 * flags are adjusted for successive hammer_btree_iterate() calls. 627 * 628 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate 629 * in a loop without worrying about it. Higher-level merged searches will 630 * adjust the flag appropriately. 631 */ 632 int 633 hammer_btree_first(hammer_cursor_t cursor) 634 { 635 int error; 636 637 error = hammer_btree_lookup(cursor); 638 if (error == ENOENT) { 639 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 640 error = hammer_btree_iterate(cursor); 641 } 642 cursor->flags |= HAMMER_CURSOR_ATEDISK; 643 return(error); 644 } 645 646 /* 647 * Similarly but for an iteration in the reverse direction. 648 * 649 * Set ATEDISK when iterating backwards to skip the current entry, 650 * which after an ENOENT lookup will be pointing beyond our end point. 651 * 652 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse 653 * in a loop without worrying about it. Higher-level merged searches will 654 * adjust the flag appropriately. 655 */ 656 int 657 hammer_btree_last(hammer_cursor_t cursor) 658 { 659 struct hammer_base_elm save; 660 int error; 661 662 save = cursor->key_beg; 663 cursor->key_beg = cursor->key_end; 664 error = hammer_btree_lookup(cursor); 665 cursor->key_beg = save; 666 if (error == ENOENT || 667 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) { 668 cursor->flags |= HAMMER_CURSOR_ATEDISK; 669 error = hammer_btree_iterate_reverse(cursor); 670 } 671 cursor->flags |= HAMMER_CURSOR_ATEDISK; 672 return(error); 673 } 674 675 /* 676 * Extract the record and/or data associated with the cursor's current 677 * position. Any prior record or data stored in the cursor is replaced. 678 * 679 * NOTE: All extractions occur at the leaf of the B-Tree. 680 */ 681 int 682 hammer_btree_extract(hammer_cursor_t cursor, int flags) 683 { 684 hammer_node_ondisk_t node; 685 hammer_btree_elm_t elm; 686 hammer_off_t data_off; 687 hammer_mount_t hmp; 688 int32_t data_len; 689 int error; 690 691 /* 692 * Certain types of corruption can result in a NULL node pointer. 693 */ 694 if (cursor->node == NULL) { 695 hkprintf("NULL cursor->node, filesystem might " 696 "have gotten corrupted\n"); 697 return (EINVAL); 698 } 699 700 /* 701 * The case where the data reference resolves to the same buffer 702 * as the record reference must be handled. 703 */ 704 node = cursor->node->ondisk; 705 elm = &node->elms[cursor->index]; 706 cursor->data = NULL; 707 hmp = cursor->node->hmp; 708 709 /* 710 * There is nothing to extract for an internal element. 711 */ 712 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) 713 return(EINVAL); 714 715 /* 716 * Only record types have data. 717 */ 718 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF); 719 cursor->leaf = &elm->leaf; 720 721 if ((flags & HAMMER_CURSOR_GET_DATA) == 0) 722 return(0); 723 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) 724 return(EINVAL); 725 data_off = elm->leaf.data_offset; 726 data_len = elm->leaf.data_len; 727 if (data_off == 0) 728 return(0); 729 730 /* 731 * Load the data 732 */ 733 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE); 734 cursor->data = hammer_bread_ext(hmp, data_off, data_len, 735 &error, &cursor->data_buffer); 736 737 /* 738 * Mark the data buffer as not being meta-data if it isn't 739 * meta-data (sometimes bulk data is accessed via a volume 740 * block device). 741 */ 742 if (error == 0) { 743 switch(elm->leaf.base.rec_type) { 744 case HAMMER_RECTYPE_DATA: 745 case HAMMER_RECTYPE_DB: 746 if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0) 747 break; 748 if (hammer_double_buffer == 0 || 749 (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) { 750 hammer_io_notmeta(cursor->data_buffer); 751 } 752 break; 753 default: 754 break; 755 } 756 } 757 758 /* 759 * Deal with CRC errors on the extracted data. 760 */ 761 if (error == 0 && 762 hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) { 763 hdkprintf("CRC DATA @ %016llx/%d FAILED\n", 764 (long long)elm->leaf.data_offset, elm->leaf.data_len); 765 if (hammer_debug_critical) 766 Debugger("CRC FAILED: DATA"); 767 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM) 768 error = EDOM; /* less critical (mirroring) */ 769 else 770 error = EIO; /* critical */ 771 } 772 return(error); 773 } 774 775 776 /* 777 * Insert a leaf element into the B-Tree at the current cursor position. 778 * The cursor is positioned such that the element at and beyond the cursor 779 * are shifted to make room for the new record. 780 * 781 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT 782 * flag set and that call must return ENOENT before this function can be 783 * called. ENOSPC is returned if there is no room to insert a new record. 784 * 785 * The caller may depend on the cursor's exclusive lock after return to 786 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE). 787 */ 788 int 789 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm, 790 int *doprop) 791 { 792 hammer_node_ondisk_t node; 793 int i; 794 int error; 795 796 *doprop = 0; 797 if ((error = hammer_cursor_upgrade_node(cursor)) != 0) 798 return(error); 799 ++hammer_stats_btree_inserts; 800 801 /* 802 * Insert the element at the leaf node and update the count in the 803 * parent. It is possible for parent to be NULL, indicating that 804 * the filesystem's ROOT B-Tree node is a leaf itself, which is 805 * possible. The root inode can never be deleted so the leaf should 806 * never be empty. 807 * 808 * Remember that leaf nodes do not have boundaries. 809 */ 810 hammer_modify_node_all(cursor->trans, cursor->node); 811 node = cursor->node->ondisk; 812 i = cursor->index; 813 KKASSERT(elm->base.btype != 0); 814 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF); 815 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS); 816 if (i != node->count) { 817 bcopy(&node->elms[i], &node->elms[i+1], 818 (node->count - i) * sizeof(*elm)); 819 } 820 node->elms[i].leaf = *elm; 821 ++node->count; 822 hammer_cursor_inserted_element(cursor->node, i); 823 824 /* 825 * Update the leaf node's aggregate mirror_tid for mirroring 826 * support. 827 */ 828 if (node->mirror_tid < elm->base.delete_tid) { 829 node->mirror_tid = elm->base.delete_tid; 830 *doprop = 1; 831 } 832 if (node->mirror_tid < elm->base.create_tid) { 833 node->mirror_tid = elm->base.create_tid; 834 *doprop = 1; 835 } 836 hammer_modify_node_done(cursor->node); 837 838 /* 839 * Debugging sanity checks. 840 */ 841 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0); 842 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0); 843 if (i) { 844 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0); 845 } 846 if (i != node->count - 1) 847 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0); 848 849 return(0); 850 } 851 852 /* 853 * Delete a record from the B-Tree at the current cursor position. 854 * The cursor is positioned such that the current element is the one 855 * to be deleted. 856 * 857 * On return the cursor will be positioned after the deleted element and 858 * MAY point to an internal node. It will be suitable for the continuation 859 * of an iteration but not for an insertion or deletion. 860 * 861 * Deletions will attempt to partially rebalance the B-Tree in an upward 862 * direction, but will terminate rather then deadlock. Empty internal nodes 863 * are never allowed by a deletion which deadlocks may end up giving us an 864 * empty leaf. The pruner will clean up and rebalance the tree. 865 * 866 * This function can return EDEADLK, requiring the caller to retry the 867 * operation after clearing the deadlock. 868 * 869 * This function will store the number of deleted btree nodes in *ndelete 870 * if ndelete is not NULL. 871 */ 872 int 873 hammer_btree_delete(hammer_cursor_t cursor, int *ndelete) 874 { 875 hammer_node_ondisk_t ondisk; 876 hammer_node_t node; 877 hammer_node_t parent __debugvar; 878 int error; 879 int i; 880 881 KKASSERT (cursor->trans->sync_lock_refs > 0); 882 if (ndelete) 883 *ndelete = 0; 884 if ((error = hammer_cursor_upgrade(cursor)) != 0) 885 return(error); 886 ++hammer_stats_btree_deletes; 887 888 /* 889 * Delete the element from the leaf node. 890 * 891 * Remember that leaf nodes do not have boundaries. 892 */ 893 node = cursor->node; 894 ondisk = node->ondisk; 895 i = cursor->index; 896 897 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF); 898 KKASSERT(i >= 0 && i < ondisk->count); 899 hammer_modify_node_all(cursor->trans, node); 900 if (i + 1 != ondisk->count) { 901 bcopy(&ondisk->elms[i+1], &ondisk->elms[i], 902 (ondisk->count - i - 1) * sizeof(ondisk->elms[0])); 903 } 904 --ondisk->count; 905 hammer_modify_node_done(node); 906 hammer_cursor_deleted_element(node, i); 907 908 /* 909 * Validate local parent 910 */ 911 if (ondisk->parent) { 912 parent = cursor->parent; 913 914 KKASSERT(parent != NULL); 915 KKASSERT(parent->node_offset == ondisk->parent); 916 } 917 918 /* 919 * If the leaf becomes empty it must be detached from the parent, 920 * potentially recursing through to the filesystem root. 921 * 922 * This may reposition the cursor at one of the parent's of the 923 * current node. 924 * 925 * Ignore deadlock errors, that simply means that btree_remove 926 * was unable to recurse and had to leave us with an empty leaf. 927 */ 928 KKASSERT(cursor->index <= ondisk->count); 929 if (ondisk->count == 0) { 930 error = btree_remove(cursor, ndelete); 931 if (error == EDEADLK) 932 error = 0; 933 } else { 934 error = 0; 935 } 936 KKASSERT(cursor->parent == NULL || 937 cursor->parent_index < cursor->parent->ondisk->count); 938 return(error); 939 } 940 941 /* 942 * PRIMARY B-TREE SEARCH SUPPORT PROCEDURE 943 * 944 * Search the filesystem B-Tree for cursor->key_beg, return the matching node. 945 * 946 * The search can begin ANYWHERE in the B-Tree. As a first step the search 947 * iterates up the tree as necessary to properly position itself prior to 948 * actually doing the sarch. 949 * 950 * INSERTIONS: The search will split full nodes and leaves on its way down 951 * and guarentee that the leaf it ends up on is not full. If we run out 952 * of space the search continues to the leaf, but ENOSPC is returned. 953 * 954 * The search is only guarenteed to end up on a leaf if an error code of 0 955 * is returned, or if inserting and an error code of ENOENT is returned. 956 * Otherwise it can stop at an internal node. On success a search returns 957 * a leaf node. 958 * 959 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire 960 * filesystem, and it is not simple code. Please note the following facts: 961 * 962 * - Internal node recursions have a boundary on the left AND right. The 963 * right boundary is non-inclusive. The create_tid is a generic part 964 * of the key for internal nodes. 965 * 966 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a 967 * historical search. ASOF and INSERT are mutually exclusive. When 968 * doing an as-of lookup btree_search() checks for a right-edge boundary 969 * case. If while recursing down the left-edge differs from the key 970 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along 971 * with cursor->create_check. This is used by btree_lookup() to iterate. 972 * The iteration backwards because as-of searches can wind up going 973 * down the wrong branch of the B-Tree. 974 */ 975 static 976 int 977 btree_search(hammer_cursor_t cursor, int flags) 978 { 979 hammer_node_ondisk_t node; 980 hammer_btree_elm_t elm; 981 int error; 982 int enospc = 0; 983 int i; 984 int r; 985 int s; 986 987 flags |= cursor->flags; 988 ++hammer_stats_btree_searches; 989 990 if (hammer_debug_btree) { 991 hammer_debug_btree_elm(cursor, 992 (hammer_btree_elm_t)&cursor->key_beg, 993 "SEARCH", 0xffff); 994 if (cursor->parent) 995 hammer_debug_btree_parent(cursor, "SEARCHP"); 996 } 997 998 /* 999 * Move our cursor up the tree until we find a node whos range covers 1000 * the key we are trying to locate. 1001 * 1002 * The left bound is inclusive, the right bound is non-inclusive. 1003 * It is ok to cursor up too far. 1004 */ 1005 for (;;) { 1006 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound); 1007 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound); 1008 if (r >= 0 && s < 0) 1009 break; 1010 KKASSERT(cursor->parent); 1011 ++hammer_stats_btree_iterations; 1012 error = hammer_cursor_up(cursor); 1013 if (error) 1014 goto done; 1015 } 1016 1017 /* 1018 * The delete-checks below are based on node, not parent. Set the 1019 * initial delete-check based on the parent. 1020 */ 1021 if (r == 1) { 1022 KKASSERT(cursor->left_bound->create_tid != 1); 1023 cursor->create_check = cursor->left_bound->create_tid - 1; 1024 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK; 1025 } 1026 1027 /* 1028 * We better have ended up with a node somewhere. 1029 */ 1030 KKASSERT(cursor->node != NULL); 1031 1032 /* 1033 * If we are inserting we can't start at a full node if the parent 1034 * is also full (because there is no way to split the node), 1035 * continue running up the tree until the requirement is satisfied 1036 * or we hit the root of the filesystem. 1037 * 1038 * (If inserting we aren't doing an as-of search so we don't have 1039 * to worry about create_check). 1040 */ 1041 while (flags & HAMMER_CURSOR_INSERT) { 1042 if (btree_node_is_full(cursor->node->ondisk) == 0) 1043 break; 1044 if (cursor->node->ondisk->parent == 0 || 1045 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) { 1046 break; 1047 } 1048 ++hammer_stats_btree_iterations; 1049 error = hammer_cursor_up(cursor); 1050 /* node may have become stale */ 1051 if (error) 1052 goto done; 1053 } 1054 1055 /* 1056 * Push down through internal nodes to locate the requested key. 1057 */ 1058 node = cursor->node->ondisk; 1059 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 1060 /* 1061 * Scan the node to find the subtree index to push down into. 1062 * We go one-past, then back-up. 1063 * 1064 * We must proactively remove deleted elements which may 1065 * have been left over from a deadlocked btree_remove(). 1066 * 1067 * The left and right boundaries are included in the loop 1068 * in order to detect edge cases. 1069 * 1070 * If the separator only differs by create_tid (r == 1) 1071 * and we are doing an as-of search, we may end up going 1072 * down a branch to the left of the one containing the 1073 * desired key. This requires numerous special cases. 1074 */ 1075 ++hammer_stats_btree_iterations; 1076 if (hammer_debug_btree) { 1077 hkprintf("SEARCH-I %016llx count=%d\n", 1078 (long long)cursor->node->node_offset, 1079 node->count); 1080 } 1081 1082 /* 1083 * Try to shortcut the search before dropping into the 1084 * linear loop. Locate the first node where r <= 1. 1085 */ 1086 i = hammer_btree_search_node(&cursor->key_beg, node); 1087 while (i <= node->count) { 1088 ++hammer_stats_btree_elements; 1089 elm = &node->elms[i]; 1090 r = hammer_btree_cmp(&cursor->key_beg, &elm->base); 1091 if (hammer_debug_btree > 2) { 1092 hkprintf(" IELM %p [%d] r=%d\n", 1093 &node->elms[i], i, r); 1094 } 1095 if (r < 0) 1096 break; 1097 if (r == 1) { 1098 KKASSERT(elm->base.create_tid != 1); 1099 cursor->create_check = elm->base.create_tid - 1; 1100 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK; 1101 } 1102 ++i; 1103 } 1104 if (hammer_debug_btree) { 1105 hkprintf("SEARCH-I preI=%d/%d r=%d\n", 1106 i, node->count, r); 1107 } 1108 1109 /* 1110 * The first two cases (i == 0 or i == node->count + 1) 1111 * occur when the parent's idea of the boundary 1112 * is wider then the child's idea of the boundary, and 1113 * require special handling. If not inserting we can 1114 * terminate the search early for these cases but the 1115 * child's boundaries cannot be unconditionally modified. 1116 * 1117 * The last case (neither of the above) fits in child's 1118 * idea of the boundary, so we can simply push down the 1119 * cursor. 1120 */ 1121 if (i == 0) { 1122 /* 1123 * If i == 0 the search terminated to the LEFT of the 1124 * left_boundary but to the RIGHT of the parent's left 1125 * boundary. 1126 */ 1127 u_int8_t save; 1128 1129 elm = &node->elms[0]; 1130 1131 /* 1132 * If we aren't inserting we can stop here. 1133 */ 1134 if ((flags & (HAMMER_CURSOR_INSERT | 1135 HAMMER_CURSOR_PRUNING)) == 0) { 1136 cursor->index = 0; 1137 return(ENOENT); 1138 } 1139 1140 /* 1141 * Correct a left-hand boundary mismatch. 1142 * 1143 * We can only do this if we can upgrade the lock, 1144 * and synchronized as a background cursor (i.e. 1145 * inserting or pruning). 1146 * 1147 * WARNING: We can only do this if inserting, i.e. 1148 * we are running on the backend. 1149 */ 1150 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1151 return(error); 1152 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 1153 hammer_modify_node_field(cursor->trans, cursor->node, 1154 elms[0]); 1155 save = node->elms[0].base.btype; 1156 node->elms[0].base = *cursor->left_bound; 1157 node->elms[0].base.btype = save; 1158 hammer_modify_node_done(cursor->node); 1159 } else if (i == node->count + 1) { 1160 /* 1161 * If i == node->count + 1 the search terminated to 1162 * the RIGHT of the right boundary but to the LEFT 1163 * of the parent's right boundary. If we aren't 1164 * inserting we can stop here. 1165 * 1166 * Note that the last element in this case is 1167 * elms[i-2] prior to adjustments to 'i'. 1168 */ 1169 --i; 1170 if ((flags & (HAMMER_CURSOR_INSERT | 1171 HAMMER_CURSOR_PRUNING)) == 0) { 1172 cursor->index = i; 1173 return (ENOENT); 1174 } 1175 1176 /* 1177 * Correct a right-hand boundary mismatch. 1178 * (actual push-down record is i-2 prior to 1179 * adjustments to i). 1180 * 1181 * We can only do this if we can upgrade the lock, 1182 * and synchronized as a background cursor (i.e. 1183 * inserting or pruning). 1184 * 1185 * WARNING: We can only do this if inserting, i.e. 1186 * we are running on the backend. 1187 */ 1188 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1189 return(error); 1190 elm = &node->elms[i]; 1191 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 1192 hammer_modify_node(cursor->trans, cursor->node, 1193 &elm->base, sizeof(elm->base)); 1194 elm->base = *cursor->right_bound; 1195 hammer_modify_node_done(cursor->node); 1196 --i; 1197 } else { 1198 /* 1199 * The push-down index is now i - 1. If we had 1200 * terminated on the right boundary this will point 1201 * us at the last element. 1202 */ 1203 --i; 1204 } 1205 cursor->index = i; 1206 elm = &node->elms[i]; 1207 1208 if (hammer_debug_btree) { 1209 hammer_debug_btree_elm(cursor, elm, "RESULT-I", 0xffff); 1210 } 1211 1212 /* 1213 * We better have a valid subtree offset. 1214 */ 1215 KKASSERT(elm->internal.subtree_offset != 0); 1216 1217 /* 1218 * Handle insertion and deletion requirements. 1219 * 1220 * If inserting split full nodes. The split code will 1221 * adjust cursor->node and cursor->index if the current 1222 * index winds up in the new node. 1223 * 1224 * If inserting and a left or right edge case was detected, 1225 * we cannot correct the left or right boundary and must 1226 * prepend and append an empty leaf node in order to make 1227 * the boundary correction. 1228 * 1229 * If we run out of space we set enospc but continue on 1230 * to a leaf. 1231 */ 1232 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) { 1233 if (btree_node_is_full(node)) { 1234 error = btree_split_internal(cursor); 1235 if (error) { 1236 if (error != ENOSPC) 1237 goto done; 1238 enospc = 1; 1239 } 1240 /* 1241 * reload stale pointers 1242 */ 1243 i = cursor->index; 1244 node = cursor->node->ondisk; 1245 } 1246 } 1247 1248 /* 1249 * Push down (push into new node, existing node becomes 1250 * the parent) and continue the search. 1251 */ 1252 error = hammer_cursor_down(cursor); 1253 /* node may have become stale */ 1254 if (error) 1255 goto done; 1256 node = cursor->node->ondisk; 1257 } 1258 1259 /* 1260 * We are at a leaf, do a linear search of the key array. 1261 * 1262 * On success the index is set to the matching element and 0 1263 * is returned. 1264 * 1265 * On failure the index is set to the insertion point and ENOENT 1266 * is returned. 1267 * 1268 * Boundaries are not stored in leaf nodes, so the index can wind 1269 * up to the left of element 0 (index == 0) or past the end of 1270 * the array (index == node->count). It is also possible that the 1271 * leaf might be empty. 1272 */ 1273 ++hammer_stats_btree_iterations; 1274 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF); 1275 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS); 1276 if (hammer_debug_btree) { 1277 hkprintf("SEARCH-L %016llx count=%d\n", 1278 (long long)cursor->node->node_offset, 1279 node->count); 1280 } 1281 1282 /* 1283 * Try to shortcut the search before dropping into the 1284 * linear loop. Locate the first node where r <= 1. 1285 */ 1286 i = hammer_btree_search_node(&cursor->key_beg, node); 1287 while (i < node->count) { 1288 ++hammer_stats_btree_elements; 1289 elm = &node->elms[i]; 1290 1291 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base); 1292 1293 if (hammer_debug_btree > 1) 1294 hkprintf(" LELM %p [%d] r=%d\n", &node->elms[i], i, r); 1295 1296 /* 1297 * We are at a record element. Stop if we've flipped past 1298 * key_beg, not counting the create_tid test. Allow the 1299 * r == 1 case (key_beg > element but differs only by its 1300 * create_tid) to fall through to the AS-OF check. 1301 */ 1302 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD); 1303 1304 if (r < 0) 1305 goto failed; 1306 if (r > 1) { 1307 ++i; 1308 continue; 1309 } 1310 1311 /* 1312 * Check our as-of timestamp against the element. 1313 */ 1314 if (flags & HAMMER_CURSOR_ASOF) { 1315 if (hammer_btree_chkts(cursor->asof, 1316 &node->elms[i].base) != 0) { 1317 ++i; 1318 continue; 1319 } 1320 /* success */ 1321 } else { 1322 if (r > 0) { /* can only be +1 */ 1323 ++i; 1324 continue; 1325 } 1326 /* success */ 1327 } 1328 cursor->index = i; 1329 error = 0; 1330 if (hammer_debug_btree) { 1331 hkprintf("RESULT-L %016llx[%d] (SUCCESS)\n", 1332 (long long)cursor->node->node_offset, i); 1333 } 1334 goto done; 1335 } 1336 1337 /* 1338 * The search of the leaf node failed. i is the insertion point. 1339 */ 1340 failed: 1341 if (hammer_debug_btree) { 1342 hkprintf("RESULT-L %016llx[%d] (FAILED)\n", 1343 (long long)cursor->node->node_offset, i); 1344 } 1345 1346 /* 1347 * No exact match was found, i is now at the insertion point. 1348 * 1349 * If inserting split a full leaf before returning. This 1350 * may have the side effect of adjusting cursor->node and 1351 * cursor->index. 1352 */ 1353 cursor->index = i; 1354 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 && 1355 btree_node_is_full(node)) { 1356 error = btree_split_leaf(cursor); 1357 if (error) { 1358 if (error != ENOSPC) 1359 goto done; 1360 enospc = 1; 1361 } 1362 /* 1363 * reload stale pointers 1364 */ 1365 /* NOT USED 1366 i = cursor->index; 1367 node = &cursor->node->internal; 1368 */ 1369 } 1370 1371 /* 1372 * We reached a leaf but did not find the key we were looking for. 1373 * If this is an insert we will be properly positioned for an insert 1374 * (ENOENT) or unable to insert (ENOSPC). 1375 */ 1376 error = enospc ? ENOSPC : ENOENT; 1377 done: 1378 return(error); 1379 } 1380 1381 /* 1382 * Heuristical search for the first element whos comparison is <= 1. May 1383 * return an index whos compare result is > 1 but may only return an index 1384 * whos compare result is <= 1 if it is the first element with that result. 1385 */ 1386 int 1387 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node) 1388 { 1389 int b; 1390 int s; 1391 int i; 1392 int r; 1393 1394 /* 1395 * Don't bother if the node does not have very many elements 1396 */ 1397 b = 0; 1398 s = node->count; 1399 while (s - b > 4) { 1400 i = b + (s - b) / 2; 1401 ++hammer_stats_btree_elements; 1402 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base); 1403 if (r <= 1) { 1404 s = i; 1405 } else { 1406 b = i; 1407 } 1408 } 1409 return(b); 1410 } 1411 1412 1413 /************************************************************************ 1414 * SPLITTING AND MERGING * 1415 ************************************************************************ 1416 * 1417 * These routines do all the dirty work required to split and merge nodes. 1418 */ 1419 1420 /* 1421 * Split an internal node into two nodes and move the separator at the split 1422 * point to the parent. 1423 * 1424 * (cursor->node, cursor->index) indicates the element the caller intends 1425 * to push into. We will adjust node and index if that element winds 1426 * up in the split node. 1427 * 1428 * If we are at the root of the filesystem a new root must be created with 1429 * two elements, one pointing to the original root and one pointing to the 1430 * newly allocated split node. 1431 */ 1432 static 1433 int 1434 btree_split_internal(hammer_cursor_t cursor) 1435 { 1436 hammer_node_ondisk_t ondisk; 1437 hammer_node_t node; 1438 hammer_node_t parent; 1439 hammer_node_t new_node; 1440 hammer_btree_elm_t elm; 1441 hammer_btree_elm_t parent_elm; 1442 struct hammer_node_lock lockroot; 1443 hammer_mount_t hmp = cursor->trans->hmp; 1444 int parent_index; 1445 int made_root; 1446 int split; 1447 int error; 1448 int i; 1449 const int esize = sizeof(*elm); 1450 1451 hammer_node_lock_init(&lockroot, cursor->node); 1452 error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL); 1453 if (error) 1454 goto done; 1455 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1456 goto done; 1457 ++hammer_stats_btree_splits; 1458 1459 /* 1460 * Calculate the split point. If the insertion point is at the 1461 * end of the leaf we adjust the split point significantly to the 1462 * right to try to optimize node fill and flag it. If we hit 1463 * that same leaf again our heuristic failed and we don't try 1464 * to optimize node fill (it could lead to a degenerate case). 1465 */ 1466 node = cursor->node; 1467 ondisk = node->ondisk; 1468 KKASSERT(ondisk->count > 4); 1469 if (cursor->index == ondisk->count && 1470 (node->flags & HAMMER_NODE_NONLINEAR) == 0) { 1471 split = (ondisk->count + 1) * 3 / 4; 1472 node->flags |= HAMMER_NODE_NONLINEAR; 1473 } else { 1474 /* 1475 * We are splitting but elms[split] will be promoted to 1476 * the parent, leaving the right hand node with one less 1477 * element. If the insertion point will be on the 1478 * left-hand side adjust the split point to give the 1479 * right hand side one additional node. 1480 */ 1481 split = (ondisk->count + 1) / 2; 1482 if (cursor->index <= split) 1483 --split; 1484 } 1485 1486 /* 1487 * If we are at the root of the filesystem, create a new root node 1488 * with 1 element and split normally. Avoid making major 1489 * modifications until we know the whole operation will work. 1490 */ 1491 if (ondisk->parent == 0) { 1492 parent = hammer_alloc_btree(cursor->trans, 0, &error); 1493 if (parent == NULL) 1494 goto done; 1495 hammer_lock_ex(&parent->lock); 1496 hammer_modify_node_noundo(cursor->trans, parent); 1497 ondisk = parent->ondisk; 1498 ondisk->count = 1; 1499 ondisk->parent = 0; 1500 ondisk->mirror_tid = node->ondisk->mirror_tid; 1501 ondisk->signature = node->ondisk->signature; 1502 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1503 ondisk->elms[0].base = hmp->root_btree_beg; 1504 ondisk->elms[0].base.btype = node->ondisk->type; 1505 ondisk->elms[0].internal.subtree_offset = node->node_offset; 1506 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid; 1507 ondisk->elms[1].base = hmp->root_btree_end; 1508 hammer_modify_node_done(parent); 1509 made_root = 1; 1510 parent_index = 0; /* index of current node in parent */ 1511 } else { 1512 made_root = 0; 1513 parent = cursor->parent; 1514 parent_index = cursor->parent_index; 1515 } 1516 1517 /* 1518 * Split node into new_node at the split point. 1519 * 1520 * B O O O P N N B <-- P = node->elms[split] (index 4) 1521 * 0 1 2 3 4 5 6 <-- subtree indices 1522 * 1523 * x x P x x 1524 * s S S s 1525 * / \ 1526 * B O O O B B N N B <--- inner boundary points are 'P' 1527 * 0 1 2 3 4 5 6 1528 */ 1529 new_node = hammer_alloc_btree(cursor->trans, 0, &error); 1530 if (new_node == NULL) { 1531 if (made_root) { 1532 hammer_unlock(&parent->lock); 1533 hammer_delete_node(cursor->trans, parent); 1534 hammer_rel_node(parent); 1535 } 1536 goto done; 1537 } 1538 hammer_lock_ex(&new_node->lock); 1539 1540 /* 1541 * Create the new node. P becomes the left-hand boundary in the 1542 * new node. Copy the right-hand boundary as well. 1543 * 1544 * elm is the new separator. 1545 */ 1546 hammer_modify_node_noundo(cursor->trans, new_node); 1547 hammer_modify_node_all(cursor->trans, node); 1548 ondisk = node->ondisk; 1549 elm = &ondisk->elms[split]; 1550 bcopy(elm, &new_node->ondisk->elms[0], 1551 (ondisk->count - split + 1) * esize); /* +1 for boundary */ 1552 new_node->ondisk->count = ondisk->count - split; 1553 new_node->ondisk->parent = parent->node_offset; 1554 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1555 new_node->ondisk->mirror_tid = ondisk->mirror_tid; 1556 KKASSERT(ondisk->type == new_node->ondisk->type); 1557 hammer_cursor_split_node(node, new_node, split); 1558 1559 /* 1560 * Cleanup the original node. Elm (P) becomes the new boundary, 1561 * its subtree_offset was moved to the new node. If we had created 1562 * a new root its parent pointer may have changed. 1563 */ 1564 elm->base.btype = HAMMER_BTREE_TYPE_NONE; 1565 elm->internal.subtree_offset = 0; 1566 ondisk->count = split; 1567 1568 /* 1569 * Insert the separator into the parent, fixup the parent's 1570 * reference to the original node, and reference the new node. 1571 * The separator is P. 1572 * 1573 * Remember that ondisk->count does not include the right-hand boundary. 1574 */ 1575 hammer_modify_node_all(cursor->trans, parent); 1576 ondisk = parent->ondisk; 1577 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS); 1578 parent_elm = &ondisk->elms[parent_index+1]; 1579 bcopy(parent_elm, parent_elm + 1, 1580 (ondisk->count - parent_index) * esize); 1581 1582 /* 1583 * Why not use hammer_make_separator() here ? 1584 */ 1585 parent_elm->internal.base = elm->base; /* separator P */ 1586 parent_elm->internal.base.btype = new_node->ondisk->type; 1587 parent_elm->internal.subtree_offset = new_node->node_offset; 1588 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid; 1589 ++ondisk->count; 1590 hammer_modify_node_done(parent); 1591 hammer_cursor_inserted_element(parent, parent_index + 1); 1592 1593 /* 1594 * The children of new_node need their parent pointer set to new_node. 1595 * The children have already been locked by 1596 * hammer_btree_lock_children(). 1597 */ 1598 for (i = 0; i < new_node->ondisk->count; ++i) { 1599 elm = &new_node->ondisk->elms[i]; 1600 error = btree_set_parent_of_child(cursor->trans, new_node, elm); 1601 if (error) { 1602 hpanic("btree-fixup problem"); 1603 } 1604 } 1605 hammer_modify_node_done(new_node); 1606 1607 /* 1608 * The filesystem's root B-Tree pointer may have to be updated. 1609 */ 1610 if (made_root) { 1611 hammer_volume_t volume; 1612 1613 volume = hammer_get_root_volume(hmp, &error); 1614 KKASSERT(error == 0); 1615 1616 hammer_modify_volume_field(cursor->trans, volume, 1617 vol0_btree_root); 1618 volume->ondisk->vol0_btree_root = parent->node_offset; 1619 hammer_modify_volume_done(volume); 1620 node->ondisk->parent = parent->node_offset; 1621 /* node->ondisk->signature = 0; */ 1622 if (cursor->parent) { 1623 hammer_unlock(&cursor->parent->lock); 1624 hammer_rel_node(cursor->parent); 1625 } 1626 cursor->parent = parent; /* lock'd and ref'd */ 1627 hammer_rel_volume(volume, 0); 1628 } 1629 hammer_modify_node_done(node); 1630 1631 /* 1632 * Ok, now adjust the cursor depending on which element the original 1633 * index was pointing at. If we are >= the split point the push node 1634 * is now in the new node. 1635 * 1636 * NOTE: If we are at the split point itself we cannot stay with the 1637 * original node because the push index will point at the right-hand 1638 * boundary, which is illegal. 1639 * 1640 * NOTE: The cursor's parent or parent_index must be adjusted for 1641 * the case where a new parent (new root) was created, and the case 1642 * where the cursor is now pointing at the split node. 1643 */ 1644 if (cursor->index >= split) { 1645 cursor->parent_index = parent_index + 1; 1646 cursor->index -= split; 1647 hammer_unlock(&cursor->node->lock); 1648 hammer_rel_node(cursor->node); 1649 cursor->node = new_node; /* locked and ref'd */ 1650 } else { 1651 cursor->parent_index = parent_index; 1652 hammer_unlock(&new_node->lock); 1653 hammer_rel_node(new_node); 1654 } 1655 1656 /* 1657 * Fixup left and right bounds 1658 */ 1659 parent_elm = &parent->ondisk->elms[cursor->parent_index]; 1660 cursor->left_bound = &parent_elm[0].internal.base; 1661 cursor->right_bound = &parent_elm[1].internal.base; 1662 KKASSERT(hammer_btree_cmp(cursor->left_bound, 1663 &cursor->node->ondisk->elms[0].internal.base) <= 0); 1664 KKASSERT(hammer_btree_cmp(cursor->right_bound, 1665 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0); 1666 1667 done: 1668 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL); 1669 hammer_cursor_downgrade(cursor); 1670 return (error); 1671 } 1672 1673 /* 1674 * Same as the above, but splits a full leaf node. 1675 */ 1676 static 1677 int 1678 btree_split_leaf(hammer_cursor_t cursor) 1679 { 1680 hammer_node_ondisk_t ondisk; 1681 hammer_node_t parent; 1682 hammer_node_t leaf; 1683 hammer_mount_t hmp; 1684 hammer_node_t new_leaf; 1685 hammer_btree_elm_t elm; 1686 hammer_btree_elm_t parent_elm; 1687 hammer_base_elm_t mid_boundary; 1688 int parent_index; 1689 int made_root; 1690 int split; 1691 int error; 1692 const size_t esize = sizeof(*elm); 1693 1694 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1695 return(error); 1696 ++hammer_stats_btree_splits; 1697 1698 KKASSERT(hammer_btree_cmp(cursor->left_bound, 1699 &cursor->node->ondisk->elms[0].leaf.base) <= 0); 1700 KKASSERT(hammer_btree_cmp(cursor->right_bound, 1701 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0); 1702 1703 /* 1704 * Calculate the split point. If the insertion point is at the 1705 * end of the leaf we adjust the split point significantly to the 1706 * right to try to optimize node fill and flag it. If we hit 1707 * that same leaf again our heuristic failed and we don't try 1708 * to optimize node fill (it could lead to a degenerate case). 1709 */ 1710 leaf = cursor->node; 1711 ondisk = leaf->ondisk; 1712 KKASSERT(ondisk->count > 4); 1713 if (cursor->index == ondisk->count && 1714 (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) { 1715 split = (ondisk->count + 1) * 3 / 4; 1716 leaf->flags |= HAMMER_NODE_NONLINEAR; 1717 } else { 1718 split = (ondisk->count + 1) / 2; 1719 } 1720 1721 #if 0 1722 /* 1723 * If the insertion point is at the split point shift the 1724 * split point left so we don't have to worry about 1725 */ 1726 if (cursor->index == split) 1727 --split; 1728 #endif 1729 KKASSERT(split > 0 && split < ondisk->count); 1730 1731 error = 0; 1732 hmp = leaf->hmp; 1733 1734 elm = &ondisk->elms[split]; 1735 1736 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0); 1737 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0); 1738 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0); 1739 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0); 1740 1741 /* 1742 * If we are at the root of the tree, create a new root node with 1743 * 1 element and split normally. Avoid making major modifications 1744 * until we know the whole operation will work. 1745 */ 1746 if (ondisk->parent == 0) { 1747 parent = hammer_alloc_btree(cursor->trans, 0, &error); 1748 if (parent == NULL) 1749 goto done; 1750 hammer_lock_ex(&parent->lock); 1751 hammer_modify_node_noundo(cursor->trans, parent); 1752 ondisk = parent->ondisk; 1753 ondisk->count = 1; 1754 ondisk->parent = 0; 1755 ondisk->mirror_tid = leaf->ondisk->mirror_tid; 1756 ondisk->signature = leaf->ondisk->signature; 1757 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1758 ondisk->elms[0].base = hmp->root_btree_beg; 1759 ondisk->elms[0].base.btype = leaf->ondisk->type; 1760 ondisk->elms[0].internal.subtree_offset = leaf->node_offset; 1761 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid; 1762 ondisk->elms[1].base = hmp->root_btree_end; 1763 hammer_modify_node_done(parent); 1764 made_root = 1; 1765 parent_index = 0; /* insertion point in parent */ 1766 } else { 1767 made_root = 0; 1768 parent = cursor->parent; 1769 parent_index = cursor->parent_index; 1770 } 1771 1772 /* 1773 * Split leaf into new_leaf at the split point. Select a separator 1774 * value in-between the two leafs but with a bent towards the right 1775 * leaf since comparisons use an 'elm >= separator' inequality. 1776 * 1777 * L L L L L L L L 1778 * 1779 * x x P x x 1780 * s S S s 1781 * / \ 1782 * L L L L L L L L 1783 */ 1784 new_leaf = hammer_alloc_btree(cursor->trans, 0, &error); 1785 if (new_leaf == NULL) { 1786 if (made_root) { 1787 hammer_unlock(&parent->lock); 1788 hammer_delete_node(cursor->trans, parent); 1789 hammer_rel_node(parent); 1790 } 1791 goto done; 1792 } 1793 hammer_lock_ex(&new_leaf->lock); 1794 1795 /* 1796 * Create the new node and copy the leaf elements from the split 1797 * point on to the new node. 1798 */ 1799 hammer_modify_node_all(cursor->trans, leaf); 1800 hammer_modify_node_noundo(cursor->trans, new_leaf); 1801 ondisk = leaf->ondisk; 1802 elm = &ondisk->elms[split]; 1803 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize); 1804 new_leaf->ondisk->count = ondisk->count - split; 1805 new_leaf->ondisk->parent = parent->node_offset; 1806 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF; 1807 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid; 1808 KKASSERT(ondisk->type == new_leaf->ondisk->type); 1809 hammer_modify_node_done(new_leaf); 1810 hammer_cursor_split_node(leaf, new_leaf, split); 1811 1812 /* 1813 * Cleanup the original node. Because this is a leaf node and 1814 * leaf nodes do not have a right-hand boundary, there 1815 * aren't any special edge cases to clean up. We just fixup the 1816 * count. 1817 */ 1818 ondisk->count = split; 1819 1820 /* 1821 * Insert the separator into the parent, fixup the parent's 1822 * reference to the original node, and reference the new node. 1823 * The separator is P. 1824 * 1825 * Remember that ondisk->count does not include the right-hand boundary. 1826 * We are copying parent_index+1 to parent_index+2, not +0 to +1. 1827 */ 1828 hammer_modify_node_all(cursor->trans, parent); 1829 ondisk = parent->ondisk; 1830 KKASSERT(split != 0); 1831 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS); 1832 parent_elm = &ondisk->elms[parent_index+1]; 1833 bcopy(parent_elm, parent_elm + 1, 1834 (ondisk->count - parent_index) * esize); 1835 1836 /* 1837 * elm[-1] is the right-most elm in the original node. 1838 * elm[0] equals the left-most elm at index=0 in the new node. 1839 * parent_elm[-1] and parent_elm point to original and new node. 1840 * Update the parent_elm base to meet >elm[-1] and <=elm[0]. 1841 */ 1842 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base); 1843 parent_elm->internal.base.btype = new_leaf->ondisk->type; 1844 parent_elm->internal.subtree_offset = new_leaf->node_offset; 1845 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid; 1846 mid_boundary = &parent_elm->base; 1847 ++ondisk->count; 1848 hammer_modify_node_done(parent); 1849 hammer_cursor_inserted_element(parent, parent_index + 1); 1850 1851 /* 1852 * The filesystem's root B-Tree pointer may have to be updated. 1853 */ 1854 if (made_root) { 1855 hammer_volume_t volume; 1856 1857 volume = hammer_get_root_volume(hmp, &error); 1858 KKASSERT(error == 0); 1859 1860 hammer_modify_volume_field(cursor->trans, volume, 1861 vol0_btree_root); 1862 volume->ondisk->vol0_btree_root = parent->node_offset; 1863 hammer_modify_volume_done(volume); 1864 leaf->ondisk->parent = parent->node_offset; 1865 /* leaf->ondisk->signature = 0; */ 1866 if (cursor->parent) { 1867 hammer_unlock(&cursor->parent->lock); 1868 hammer_rel_node(cursor->parent); 1869 } 1870 cursor->parent = parent; /* lock'd and ref'd */ 1871 hammer_rel_volume(volume, 0); 1872 } 1873 hammer_modify_node_done(leaf); 1874 1875 /* 1876 * Ok, now adjust the cursor depending on which element the original 1877 * index was pointing at. If we are >= the split point the push node 1878 * is now in the new node. 1879 * 1880 * NOTE: If we are at the split point itself we need to select the 1881 * old or new node based on where key_beg's insertion point will be. 1882 * If we pick the wrong side the inserted element will wind up in 1883 * the wrong leaf node and outside that node's bounds. 1884 */ 1885 if (cursor->index > split || 1886 (cursor->index == split && 1887 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) { 1888 cursor->parent_index = parent_index + 1; 1889 cursor->index -= split; 1890 hammer_unlock(&cursor->node->lock); 1891 hammer_rel_node(cursor->node); 1892 cursor->node = new_leaf; 1893 } else { 1894 cursor->parent_index = parent_index; 1895 hammer_unlock(&new_leaf->lock); 1896 hammer_rel_node(new_leaf); 1897 } 1898 1899 /* 1900 * Fixup left and right bounds 1901 */ 1902 parent_elm = &parent->ondisk->elms[cursor->parent_index]; 1903 cursor->left_bound = &parent_elm[0].internal.base; 1904 cursor->right_bound = &parent_elm[1].internal.base; 1905 1906 /* 1907 * Assert that the bounds are correct. 1908 */ 1909 KKASSERT(hammer_btree_cmp(cursor->left_bound, 1910 &cursor->node->ondisk->elms[0].leaf.base) <= 0); 1911 KKASSERT(hammer_btree_cmp(cursor->right_bound, 1912 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0); 1913 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0); 1914 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0); 1915 1916 done: 1917 hammer_cursor_downgrade(cursor); 1918 return (error); 1919 } 1920 1921 #if 0 1922 1923 /* 1924 * Recursively correct the right-hand boundary's create_tid to (tid) as 1925 * long as the rest of the key matches. We have to recurse upward in 1926 * the tree as well as down the left side of each parent's right node. 1927 * 1928 * Return EDEADLK if we were only partially successful, forcing the caller 1929 * to try again. The original cursor is not modified. This routine can 1930 * also fail with EDEADLK if it is forced to throw away a portion of its 1931 * record history. 1932 * 1933 * The caller must pass a downgraded cursor to us (otherwise we can't dup it). 1934 */ 1935 struct hammer_rhb { 1936 TAILQ_ENTRY(hammer_rhb) entry; 1937 hammer_node_t node; 1938 int index; 1939 }; 1940 1941 TAILQ_HEAD(hammer_rhb_list, hammer_rhb); 1942 1943 int 1944 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid) 1945 { 1946 struct hammer_mount *hmp; 1947 struct hammer_rhb_list rhb_list; 1948 hammer_base_elm_t elm; 1949 hammer_node_t orig_node; 1950 struct hammer_rhb *rhb; 1951 int orig_index; 1952 int error; 1953 1954 TAILQ_INIT(&rhb_list); 1955 hmp = cursor->trans->hmp; 1956 1957 /* 1958 * Save our position so we can restore it on return. This also 1959 * gives us a stable 'elm'. 1960 */ 1961 orig_node = cursor->node; 1962 hammer_ref_node(orig_node); 1963 hammer_lock_sh(&orig_node->lock); 1964 orig_index = cursor->index; 1965 elm = &orig_node->ondisk->elms[orig_index].base; 1966 1967 /* 1968 * Now build a list of parents going up, allocating a rhb 1969 * structure for each one. 1970 */ 1971 while (cursor->parent) { 1972 /* 1973 * Stop if we no longer have any right-bounds to fix up 1974 */ 1975 if (elm->obj_id != cursor->right_bound->obj_id || 1976 elm->rec_type != cursor->right_bound->rec_type || 1977 elm->key != cursor->right_bound->key) { 1978 break; 1979 } 1980 1981 /* 1982 * Stop if the right-hand bound's create_tid does not 1983 * need to be corrected. 1984 */ 1985 if (cursor->right_bound->create_tid >= tid) 1986 break; 1987 1988 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO); 1989 rhb->node = cursor->parent; 1990 rhb->index = cursor->parent_index; 1991 hammer_ref_node(rhb->node); 1992 hammer_lock_sh(&rhb->node->lock); 1993 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry); 1994 1995 hammer_cursor_up(cursor); 1996 } 1997 1998 /* 1999 * now safely adjust the right hand bound for each rhb. This may 2000 * also require taking the right side of the tree and iterating down 2001 * ITS left side. 2002 */ 2003 error = 0; 2004 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2005 error = hammer_cursor_seek(cursor, rhb->node, rhb->index); 2006 if (error) 2007 break; 2008 TAILQ_REMOVE(&rhb_list, rhb, entry); 2009 hammer_unlock(&rhb->node->lock); 2010 hammer_rel_node(rhb->node); 2011 kfree(rhb, hmp->m_misc); 2012 2013 switch (cursor->node->ondisk->type) { 2014 case HAMMER_BTREE_TYPE_INTERNAL: 2015 /* 2016 * Right-boundary for parent at internal node 2017 * is one element to the right of the element whos 2018 * right boundary needs adjusting. We must then 2019 * traverse down the left side correcting any left 2020 * bounds (which may now be too far to the left). 2021 */ 2022 ++cursor->index; 2023 error = hammer_btree_correct_lhb(cursor, tid); 2024 break; 2025 default: 2026 hpanic("Bad node type"); 2027 error = EINVAL; 2028 break; 2029 } 2030 } 2031 2032 /* 2033 * Cleanup 2034 */ 2035 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2036 TAILQ_REMOVE(&rhb_list, rhb, entry); 2037 hammer_unlock(&rhb->node->lock); 2038 hammer_rel_node(rhb->node); 2039 kfree(rhb, hmp->m_misc); 2040 } 2041 error = hammer_cursor_seek(cursor, orig_node, orig_index); 2042 hammer_unlock(&orig_node->lock); 2043 hammer_rel_node(orig_node); 2044 return (error); 2045 } 2046 2047 /* 2048 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand 2049 * bound going downward starting at the current cursor position. 2050 * 2051 * This function does not restore the cursor after use. 2052 */ 2053 int 2054 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid) 2055 { 2056 struct hammer_rhb_list rhb_list; 2057 hammer_base_elm_t elm; 2058 hammer_base_elm_t cmp; 2059 struct hammer_rhb *rhb; 2060 struct hammer_mount *hmp; 2061 int error; 2062 2063 TAILQ_INIT(&rhb_list); 2064 hmp = cursor->trans->hmp; 2065 2066 cmp = &cursor->node->ondisk->elms[cursor->index].base; 2067 2068 /* 2069 * Record the node and traverse down the left-hand side for all 2070 * matching records needing a boundary correction. 2071 */ 2072 error = 0; 2073 for (;;) { 2074 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO); 2075 rhb->node = cursor->node; 2076 rhb->index = cursor->index; 2077 hammer_ref_node(rhb->node); 2078 hammer_lock_sh(&rhb->node->lock); 2079 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry); 2080 2081 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 2082 /* 2083 * Nothing to traverse down if we are at the right 2084 * boundary of an internal node. 2085 */ 2086 if (cursor->index == cursor->node->ondisk->count) 2087 break; 2088 } else { 2089 elm = &cursor->node->ondisk->elms[cursor->index].base; 2090 if (elm->btype == HAMMER_BTREE_TYPE_RECORD) 2091 break; 2092 hpanic("Illegal leaf record type %02x", elm->btype); 2093 } 2094 error = hammer_cursor_down(cursor); 2095 if (error) 2096 break; 2097 2098 elm = &cursor->node->ondisk->elms[cursor->index].base; 2099 if (elm->obj_id != cmp->obj_id || 2100 elm->rec_type != cmp->rec_type || 2101 elm->key != cmp->key) { 2102 break; 2103 } 2104 if (elm->create_tid >= tid) 2105 break; 2106 2107 } 2108 2109 /* 2110 * Now we can safely adjust the left-hand boundary from the bottom-up. 2111 * The last element we remove from the list is the caller's right hand 2112 * boundary, which must also be adjusted. 2113 */ 2114 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2115 error = hammer_cursor_seek(cursor, rhb->node, rhb->index); 2116 if (error) 2117 break; 2118 TAILQ_REMOVE(&rhb_list, rhb, entry); 2119 hammer_unlock(&rhb->node->lock); 2120 hammer_rel_node(rhb->node); 2121 kfree(rhb, hmp->m_misc); 2122 2123 elm = &cursor->node->ondisk->elms[cursor->index].base; 2124 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 2125 hammer_modify_node(cursor->trans, cursor->node, 2126 &elm->create_tid, 2127 sizeof(elm->create_tid)); 2128 elm->create_tid = tid; 2129 hammer_modify_node_done(cursor->node); 2130 } else { 2131 hpanic("Bad element type"); 2132 } 2133 } 2134 2135 /* 2136 * Cleanup 2137 */ 2138 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2139 TAILQ_REMOVE(&rhb_list, rhb, entry); 2140 hammer_unlock(&rhb->node->lock); 2141 hammer_rel_node(rhb->node); 2142 kfree(rhb, hmp->m_misc); 2143 } 2144 return (error); 2145 } 2146 2147 #endif 2148 2149 /* 2150 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at 2151 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete 2152 * the operation due to a deadlock, or some other error. 2153 * 2154 * This routine is initially called with an empty leaf and may be 2155 * recursively called with single-element internal nodes. 2156 * 2157 * It should also be noted that when removing empty leaves we must be sure 2158 * to test and update mirror_tid because another thread may have deadlocked 2159 * against us (or someone) trying to propagate it up and cannot retry once 2160 * the node has been deleted. 2161 * 2162 * On return the cursor may end up pointing to an internal node, suitable 2163 * for further iteration but not for an immediate insertion or deletion. 2164 */ 2165 static int 2166 btree_remove(hammer_cursor_t cursor, int *ndelete) 2167 { 2168 hammer_node_ondisk_t ondisk; 2169 hammer_btree_elm_t elm; 2170 hammer_node_t node; 2171 hammer_node_t parent; 2172 const int esize = sizeof(*elm); 2173 int error; 2174 2175 node = cursor->node; 2176 2177 /* 2178 * When deleting the root of the filesystem convert it to 2179 * an empty leaf node. Internal nodes cannot be empty. 2180 */ 2181 ondisk = node->ondisk; 2182 if (ondisk->parent == 0) { 2183 KKASSERT(cursor->parent == NULL); 2184 hammer_modify_node_all(cursor->trans, node); 2185 KKASSERT(ondisk == node->ondisk); 2186 ondisk->type = HAMMER_BTREE_TYPE_LEAF; 2187 ondisk->count = 0; 2188 hammer_modify_node_done(node); 2189 cursor->index = 0; 2190 return(0); 2191 } 2192 2193 parent = cursor->parent; 2194 2195 /* 2196 * Attempt to remove the parent's reference to the child. If the 2197 * parent would become empty we have to recurse. If we fail we 2198 * leave the parent pointing to an empty leaf node. 2199 * 2200 * We have to recurse successfully before we can delete the internal 2201 * node as it is illegal to have empty internal nodes. Even though 2202 * the operation may be aborted we must still fixup any unlocked 2203 * cursors as if we had deleted the element prior to recursing 2204 * (by calling hammer_cursor_deleted_element()) so those cursors 2205 * are properly forced up the chain by the recursion. 2206 */ 2207 if (parent->ondisk->count == 1) { 2208 /* 2209 * This special cursor_up_locked() call leaves the original 2210 * node exclusively locked and referenced, leaves the 2211 * original parent locked (as the new node), and locks the 2212 * new parent. It can return EDEADLK. 2213 * 2214 * We cannot call hammer_cursor_removed_node() until we are 2215 * actually able to remove the node. If we did then tracked 2216 * cursors in the middle of iterations could be repointed 2217 * to a parent node. If this occurs they could end up 2218 * scanning newly inserted records into the node (that could 2219 * not be deleted) when they push down again. 2220 * 2221 * Due to the way the recursion works the final parent is left 2222 * in cursor->parent after the recursion returns. Each 2223 * layer on the way back up is thus able to call 2224 * hammer_cursor_removed_node() and 'jump' the node up to 2225 * the (same) final parent. 2226 * 2227 * NOTE! The local variable 'parent' is invalid after we 2228 * call hammer_cursor_up_locked(). 2229 */ 2230 error = hammer_cursor_up_locked(cursor); 2231 parent = NULL; 2232 2233 if (error == 0) { 2234 hammer_cursor_deleted_element(cursor->node, 0); 2235 error = btree_remove(cursor, ndelete); 2236 if (error == 0) { 2237 KKASSERT(node != cursor->node); 2238 hammer_cursor_removed_node( 2239 node, cursor->node, cursor->index); 2240 hammer_modify_node_all(cursor->trans, node); 2241 ondisk = node->ondisk; 2242 ondisk->type = HAMMER_BTREE_TYPE_DELETED; 2243 ondisk->count = 0; 2244 hammer_modify_node_done(node); 2245 hammer_flush_node(node, 0); 2246 hammer_delete_node(cursor->trans, node); 2247 if (ndelete) 2248 (*ndelete)++; 2249 } else { 2250 /* 2251 * Defer parent removal because we could not 2252 * get the lock, just let the leaf remain 2253 * empty. 2254 */ 2255 /**/ 2256 } 2257 hammer_unlock(&node->lock); 2258 hammer_rel_node(node); 2259 } else { 2260 /* 2261 * Defer parent removal because we could not 2262 * get the lock, just let the leaf remain 2263 * empty. 2264 */ 2265 /**/ 2266 } 2267 } else { 2268 KKASSERT(parent->ondisk->count > 1); 2269 2270 hammer_modify_node_all(cursor->trans, parent); 2271 ondisk = parent->ondisk; 2272 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL); 2273 2274 elm = &ondisk->elms[cursor->parent_index]; 2275 KKASSERT(elm->internal.subtree_offset == node->node_offset); 2276 KKASSERT(ondisk->count > 0); 2277 2278 /* 2279 * We must retain the highest mirror_tid. The deleted 2280 * range is now encompassed by the element to the left. 2281 * If we are already at the left edge the new left edge 2282 * inherits mirror_tid. 2283 * 2284 * Note that bounds of the parent to our parent may create 2285 * a gap to the left of our left-most node or to the right 2286 * of our right-most node. The gap is silently included 2287 * in the mirror_tid's area of effect from the point of view 2288 * of the scan. 2289 */ 2290 if (cursor->parent_index) { 2291 if (elm[-1].internal.mirror_tid < 2292 elm[0].internal.mirror_tid) { 2293 elm[-1].internal.mirror_tid = 2294 elm[0].internal.mirror_tid; 2295 } 2296 } else { 2297 if (elm[1].internal.mirror_tid < 2298 elm[0].internal.mirror_tid) { 2299 elm[1].internal.mirror_tid = 2300 elm[0].internal.mirror_tid; 2301 } 2302 } 2303 2304 /* 2305 * Delete the subtree reference in the parent. Include 2306 * boundary element at end. 2307 */ 2308 bcopy(&elm[1], &elm[0], 2309 (ondisk->count - cursor->parent_index) * esize); 2310 --ondisk->count; 2311 hammer_modify_node_done(parent); 2312 hammer_cursor_removed_node(node, parent, cursor->parent_index); 2313 hammer_cursor_deleted_element(parent, cursor->parent_index); 2314 hammer_flush_node(node, 0); 2315 hammer_delete_node(cursor->trans, node); 2316 2317 /* 2318 * cursor->node is invalid, cursor up to make the cursor 2319 * valid again. We have to flag the condition in case 2320 * another thread wiggles an insertion in during an 2321 * iteration. 2322 */ 2323 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK; 2324 error = hammer_cursor_up(cursor); 2325 if (ndelete) 2326 (*ndelete)++; 2327 } 2328 return (error); 2329 } 2330 2331 /* 2332 * Propagate cursor->trans->tid up the B-Tree starting at the current 2333 * cursor position using pseudofs info gleaned from the passed inode. 2334 * 2335 * The passed inode has no relationship to the cursor position other 2336 * then being in the same pseudofs as the insertion or deletion we 2337 * are propagating the mirror_tid for. 2338 * 2339 * WARNING! Because we push and pop the passed cursor, it may be 2340 * modified by other B-Tree operations while it is unlocked 2341 * and things like the node & leaf pointers, and indexes might 2342 * change. 2343 */ 2344 void 2345 hammer_btree_do_propagation(hammer_cursor_t cursor, 2346 hammer_pseudofs_inmem_t pfsm, 2347 hammer_btree_leaf_elm_t leaf) 2348 { 2349 hammer_cursor_t ncursor; 2350 hammer_tid_t mirror_tid; 2351 int error __debugvar; 2352 2353 /* 2354 * We do not propagate a mirror_tid if the filesystem was mounted 2355 * in no-mirror mode. 2356 */ 2357 if (cursor->trans->hmp->master_id < 0) 2358 return; 2359 2360 /* 2361 * This is a bit of a hack because we cannot deadlock or return 2362 * EDEADLK here. The related operation has already completed and 2363 * we must propagate the mirror_tid now regardless. 2364 * 2365 * Generate a new cursor which inherits the original's locks and 2366 * unlock the original. Use the new cursor to propagate the 2367 * mirror_tid. Then clean up the new cursor and reacquire locks 2368 * on the original. 2369 * 2370 * hammer_dup_cursor() cannot dup locks. The dup inherits the 2371 * original's locks and the original is tracked and must be 2372 * re-locked. 2373 */ 2374 mirror_tid = cursor->node->ondisk->mirror_tid; 2375 KKASSERT(mirror_tid != 0); 2376 ncursor = hammer_push_cursor(cursor); 2377 error = hammer_btree_mirror_propagate(ncursor, mirror_tid); 2378 KKASSERT(error == 0); 2379 hammer_pop_cursor(cursor, ncursor); 2380 /* WARNING: cursor's leaf pointer may change after pop */ 2381 } 2382 2383 2384 /* 2385 * Propagate a mirror TID update upwards through the B-Tree to the root. 2386 * 2387 * A locked internal node must be passed in. The node will remain locked 2388 * on return. 2389 * 2390 * This function syncs mirror_tid at the specified internal node's element, 2391 * adjusts the node's aggregation mirror_tid, and then recurses upwards. 2392 */ 2393 static int 2394 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid) 2395 { 2396 hammer_btree_internal_elm_t elm; 2397 hammer_node_t node; 2398 int error; 2399 2400 for (;;) { 2401 error = hammer_cursor_up(cursor); 2402 if (error == 0) 2403 error = hammer_cursor_upgrade(cursor); 2404 2405 /* 2406 * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the 2407 * cursor will still be properly positioned for 2408 * mirror propagation, just not for iterations. 2409 */ 2410 while (error == EDEADLK) { 2411 hammer_recover_cursor(cursor); 2412 error = hammer_cursor_upgrade(cursor); 2413 } 2414 if (error) 2415 break; 2416 2417 /* 2418 * If the cursor deadlocked it could end up at a leaf 2419 * after we lost the lock. 2420 */ 2421 node = cursor->node; 2422 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL) 2423 continue; 2424 2425 /* 2426 * Adjust the node's element 2427 */ 2428 elm = &node->ondisk->elms[cursor->index].internal; 2429 if (elm->mirror_tid >= mirror_tid) 2430 break; 2431 hammer_modify_node(cursor->trans, node, &elm->mirror_tid, 2432 sizeof(elm->mirror_tid)); 2433 elm->mirror_tid = mirror_tid; 2434 hammer_modify_node_done(node); 2435 if (hammer_debug_general & 0x0002) { 2436 hdkprintf("propagate %016llx @%016llx:%d\n", 2437 (long long)mirror_tid, 2438 (long long)node->node_offset, 2439 cursor->index); 2440 } 2441 2442 2443 /* 2444 * Adjust the node's mirror_tid aggregator 2445 */ 2446 if (node->ondisk->mirror_tid >= mirror_tid) 2447 return(0); 2448 hammer_modify_node_field(cursor->trans, node, mirror_tid); 2449 node->ondisk->mirror_tid = mirror_tid; 2450 hammer_modify_node_done(node); 2451 if (hammer_debug_general & 0x0002) { 2452 hdkprintf("propagate %016llx @%016llx\n", 2453 (long long)mirror_tid, 2454 (long long)node->node_offset); 2455 } 2456 } 2457 if (error == ENOENT) 2458 error = 0; 2459 return(error); 2460 } 2461 2462 /* 2463 * Return a pointer to node's parent. If there is no error, 2464 * *parent_index is set to an index of parent's elm that points 2465 * to this node. 2466 */ 2467 hammer_node_t 2468 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node, 2469 int *parent_indexp, int *errorp, int try_exclusive) 2470 { 2471 hammer_node_t parent; 2472 hammer_btree_elm_t elm; 2473 int i; 2474 2475 /* 2476 * Get the node 2477 */ 2478 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp); 2479 if (*errorp) { 2480 KKASSERT(parent == NULL); 2481 return(NULL); 2482 } 2483 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0); 2484 2485 /* 2486 * Lock the node 2487 */ 2488 if (try_exclusive) { 2489 if (hammer_lock_ex_try(&parent->lock)) { 2490 hammer_rel_node(parent); 2491 *errorp = EDEADLK; 2492 return(NULL); 2493 } 2494 } else { 2495 hammer_lock_sh(&parent->lock); 2496 } 2497 2498 /* 2499 * Figure out which element in the parent is pointing to the 2500 * child. 2501 */ 2502 if (node->ondisk->count) { 2503 i = hammer_btree_search_node(&node->ondisk->elms[0].base, 2504 parent->ondisk); 2505 } else { 2506 i = 0; 2507 } 2508 while (i < parent->ondisk->count) { 2509 elm = &parent->ondisk->elms[i]; 2510 if (elm->internal.subtree_offset == node->node_offset) 2511 break; 2512 ++i; 2513 } 2514 if (i == parent->ondisk->count) { 2515 hammer_unlock(&parent->lock); 2516 hpanic("Bad B-Tree link: parent %p node %p", parent, node); 2517 } 2518 *parent_indexp = i; 2519 KKASSERT(*errorp == 0); 2520 return(parent); 2521 } 2522 2523 /* 2524 * The element (elm) has been moved to a new internal node (node). 2525 * 2526 * If the element represents a pointer to an internal node that node's 2527 * parent must be adjusted to the element's new location. 2528 * 2529 * XXX deadlock potential here with our exclusive locks 2530 */ 2531 int 2532 btree_set_parent_of_child(hammer_transaction_t trans, hammer_node_t node, 2533 hammer_btree_elm_t elm) 2534 { 2535 hammer_node_t child; 2536 int error; 2537 2538 error = 0; 2539 2540 if (hammer_is_internal_node_elm(elm)) { 2541 child = hammer_get_node(trans, elm->internal.subtree_offset, 2542 0, &error); 2543 if (error == 0) { 2544 hammer_modify_node_field(trans, child, parent); 2545 child->ondisk->parent = node->node_offset; 2546 hammer_modify_node_done(child); 2547 hammer_rel_node(child); 2548 } 2549 } 2550 return(error); 2551 } 2552 2553 /* 2554 * Initialize the root of a recursive B-Tree node lock list structure. 2555 */ 2556 void 2557 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node) 2558 { 2559 TAILQ_INIT(&parent->list); 2560 parent->parent = NULL; 2561 parent->node = node; 2562 parent->index = -1; 2563 parent->count = node->ondisk->count; 2564 parent->copy = NULL; 2565 parent->flags = 0; 2566 } 2567 2568 /* 2569 * Initialize a cache of hammer_node_lock's including space allocated 2570 * for node copies. 2571 * 2572 * This is used by the rebalancing code to preallocate the copy space 2573 * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER 2574 * locks, otherwise we can blow out the pageout daemon's emergency 2575 * reserve and deadlock it. 2576 * 2577 * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache. 2578 * The flag is set when the item is pulled off the cache for use. 2579 */ 2580 void 2581 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache, 2582 int depth) 2583 { 2584 hammer_node_lock_t item; 2585 int count; 2586 2587 for (count = 1; depth; --depth) 2588 count *= HAMMER_BTREE_LEAF_ELMS; 2589 bzero(lcache, sizeof(*lcache)); 2590 TAILQ_INIT(&lcache->list); 2591 while (count) { 2592 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO); 2593 item->copy = kmalloc(sizeof(*item->copy), 2594 hmp->m_misc, M_WAITOK); 2595 TAILQ_INIT(&item->list); 2596 TAILQ_INSERT_TAIL(&lcache->list, item, entry); 2597 --count; 2598 } 2599 } 2600 2601 void 2602 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache) 2603 { 2604 hammer_node_lock_t item; 2605 2606 while ((item = TAILQ_FIRST(&lcache->list)) != NULL) { 2607 TAILQ_REMOVE(&lcache->list, item, entry); 2608 KKASSERT(item->copy); 2609 KKASSERT(TAILQ_EMPTY(&item->list)); 2610 kfree(item->copy, hmp->m_misc); 2611 kfree(item, hmp->m_misc); 2612 } 2613 KKASSERT(lcache->copy == NULL); 2614 } 2615 2616 /* 2617 * Exclusively lock all the children of node. This is used by the split 2618 * code to prevent anyone from accessing the children of a cursor node 2619 * while we fix-up its parent offset. 2620 * 2621 * If we don't lock the children we can really mess up cursors which block 2622 * trying to cursor-up into our node. 2623 * 2624 * On failure EDEADLK (or some other error) is returned. If a deadlock 2625 * error is returned the cursor is adjusted to block on termination. 2626 * 2627 * The caller is responsible for managing parent->node, the root's node 2628 * is usually aliased from a cursor. 2629 */ 2630 int 2631 hammer_btree_lock_children(hammer_cursor_t cursor, int depth, 2632 hammer_node_lock_t parent, 2633 hammer_node_lock_t lcache) 2634 { 2635 hammer_node_t node; 2636 hammer_node_lock_t item; 2637 hammer_node_ondisk_t ondisk; 2638 hammer_btree_elm_t elm; 2639 hammer_node_t child; 2640 struct hammer_mount *hmp; 2641 int error; 2642 int i; 2643 2644 node = parent->node; 2645 ondisk = node->ondisk; 2646 error = 0; 2647 hmp = cursor->trans->hmp; 2648 2649 if (ondisk->type != HAMMER_BTREE_TYPE_INTERNAL) 2650 return(0); /* This could return non-zero */ 2651 2652 /* 2653 * We really do not want to block on I/O with exclusive locks held, 2654 * pre-get the children before trying to lock the mess. This is 2655 * only done one-level deep for now. 2656 */ 2657 for (i = 0; i < ondisk->count; ++i) { 2658 ++hammer_stats_btree_elements; 2659 elm = &ondisk->elms[i]; 2660 child = hammer_get_node(cursor->trans, 2661 elm->internal.subtree_offset, 2662 0, &error); 2663 if (child) 2664 hammer_rel_node(child); 2665 } 2666 2667 /* 2668 * Do it for real 2669 */ 2670 for (i = 0; error == 0 && i < ondisk->count; ++i) { 2671 ++hammer_stats_btree_elements; 2672 elm = &ondisk->elms[i]; 2673 2674 KKASSERT(elm->internal.subtree_offset != 0); 2675 child = hammer_get_node(cursor->trans, 2676 elm->internal.subtree_offset, 2677 0, &error); 2678 if (child) { 2679 if (hammer_lock_ex_try(&child->lock) != 0) { 2680 if (cursor->deadlk_node == NULL) { 2681 cursor->deadlk_node = child; 2682 hammer_ref_node(cursor->deadlk_node); 2683 } 2684 error = EDEADLK; 2685 hammer_rel_node(child); 2686 } else { 2687 if (lcache) { 2688 item = TAILQ_FIRST(&lcache->list); 2689 KKASSERT(item != NULL); 2690 item->flags |= HAMMER_NODE_LOCK_LCACHE; 2691 TAILQ_REMOVE(&lcache->list, item, entry); 2692 } else { 2693 item = kmalloc(sizeof(*item), 2694 hmp->m_misc, 2695 M_WAITOK|M_ZERO); 2696 TAILQ_INIT(&item->list); 2697 } 2698 2699 TAILQ_INSERT_TAIL(&parent->list, item, entry); 2700 item->parent = parent; 2701 item->node = child; 2702 item->index = i; 2703 item->count = child->ondisk->count; 2704 2705 /* 2706 * Recurse (used by the rebalancing code) 2707 */ 2708 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) { 2709 error = hammer_btree_lock_children( 2710 cursor, 2711 depth - 1, 2712 item, 2713 lcache); 2714 } 2715 } 2716 } 2717 } 2718 if (error) 2719 hammer_btree_unlock_children(hmp, parent, lcache); 2720 return(error); 2721 } 2722 2723 /* 2724 * Create an in-memory copy of all B-Tree nodes listed, recursively, 2725 * including the parent. 2726 */ 2727 void 2728 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent) 2729 { 2730 hammer_mount_t hmp = cursor->trans->hmp; 2731 hammer_node_lock_t item; 2732 2733 if (parent->copy == NULL) { 2734 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0); 2735 parent->copy = kmalloc(sizeof(*parent->copy), 2736 hmp->m_misc, M_WAITOK); 2737 } 2738 KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0); 2739 *parent->copy = *parent->node->ondisk; 2740 TAILQ_FOREACH(item, &parent->list, entry) { 2741 hammer_btree_lock_copy(cursor, item); 2742 } 2743 } 2744 2745 /* 2746 * Recursively sync modified copies to the media. 2747 */ 2748 int 2749 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent) 2750 { 2751 hammer_node_lock_t item; 2752 int count = 0; 2753 2754 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) { 2755 ++count; 2756 hammer_modify_node_all(cursor->trans, parent->node); 2757 *parent->node->ondisk = *parent->copy; 2758 hammer_modify_node_done(parent->node); 2759 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) { 2760 hammer_flush_node(parent->node, 0); 2761 hammer_delete_node(cursor->trans, parent->node); 2762 } 2763 } 2764 TAILQ_FOREACH(item, &parent->list, entry) { 2765 count += hammer_btree_sync_copy(cursor, item); 2766 } 2767 return(count); 2768 } 2769 2770 /* 2771 * Release previously obtained node locks. The caller is responsible for 2772 * cleaning up parent->node itself (its usually just aliased from a cursor), 2773 * but this function will take care of the copies. 2774 * 2775 * NOTE: The root node is not placed in the lcache and node->copy is not 2776 * deallocated when lcache != NULL. 2777 */ 2778 void 2779 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent, 2780 hammer_node_lock_t lcache) 2781 { 2782 hammer_node_lock_t item; 2783 hammer_node_ondisk_t copy; 2784 2785 while ((item = TAILQ_FIRST(&parent->list)) != NULL) { 2786 TAILQ_REMOVE(&parent->list, item, entry); 2787 hammer_btree_unlock_children(hmp, item, lcache); 2788 hammer_unlock(&item->node->lock); 2789 hammer_rel_node(item->node); 2790 if (lcache) { 2791 /* 2792 * NOTE: When placing the item back in the lcache 2793 * the flag is cleared by the bzero(). 2794 * Remaining fields are cleared as a safety 2795 * measure. 2796 */ 2797 KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE); 2798 KKASSERT(TAILQ_EMPTY(&item->list)); 2799 copy = item->copy; 2800 bzero(item, sizeof(*item)); 2801 TAILQ_INIT(&item->list); 2802 item->copy = copy; 2803 if (copy) 2804 bzero(copy, sizeof(*copy)); 2805 TAILQ_INSERT_TAIL(&lcache->list, item, entry); 2806 } else { 2807 kfree(item, hmp->m_misc); 2808 } 2809 } 2810 if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) { 2811 kfree(parent->copy, hmp->m_misc); 2812 parent->copy = NULL; /* safety */ 2813 } 2814 } 2815 2816 /************************************************************************ 2817 * MISCELLANIOUS SUPPORT * 2818 ************************************************************************/ 2819 2820 /* 2821 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp). 2822 * 2823 * Note that for this particular function a return value of -1, 0, or +1 2824 * can denote a match if create_tid is otherwise discounted. A create_tid 2825 * of zero is considered to be 'infinity' in comparisons. 2826 * 2827 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c. 2828 */ 2829 int 2830 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2) 2831 { 2832 if (key1->localization < key2->localization) 2833 return(-5); 2834 if (key1->localization > key2->localization) 2835 return(5); 2836 2837 if (key1->obj_id < key2->obj_id) 2838 return(-4); 2839 if (key1->obj_id > key2->obj_id) 2840 return(4); 2841 2842 if (key1->rec_type < key2->rec_type) 2843 return(-3); 2844 if (key1->rec_type > key2->rec_type) 2845 return(3); 2846 2847 if (key1->key < key2->key) 2848 return(-2); 2849 if (key1->key > key2->key) 2850 return(2); 2851 2852 /* 2853 * A create_tid of zero indicates a record which is undeletable 2854 * and must be considered to have a value of positive infinity. 2855 */ 2856 if (key1->create_tid == 0) { 2857 if (key2->create_tid == 0) 2858 return(0); 2859 return(1); 2860 } 2861 if (key2->create_tid == 0) 2862 return(-1); 2863 if (key1->create_tid < key2->create_tid) 2864 return(-1); 2865 if (key1->create_tid > key2->create_tid) 2866 return(1); 2867 return(0); 2868 } 2869 2870 /* 2871 * Test a timestamp against an element to determine whether the 2872 * element is visible. A timestamp of 0 means 'infinity'. 2873 */ 2874 int 2875 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base) 2876 { 2877 if (asof == 0) { 2878 if (base->delete_tid) 2879 return(1); 2880 return(0); 2881 } 2882 if (asof < base->create_tid) 2883 return(-1); 2884 if (base->delete_tid && asof >= base->delete_tid) 2885 return(1); 2886 return(0); 2887 } 2888 2889 /* 2890 * Create a separator half way inbetween key1 and key2. For fields just 2891 * one unit apart, the separator will match key2. key1 is on the left-hand 2892 * side and key2 is on the right-hand side. 2893 * 2894 * key2 must be >= the separator. It is ok for the separator to match key2. 2895 * 2896 * NOTE: Even if key1 does not match key2, the separator may wind up matching 2897 * key2. 2898 * 2899 * NOTE: It might be beneficial to just scrap this whole mess and just 2900 * set the separator to key2. 2901 */ 2902 #define MAKE_SEPARATOR(key1, key2, dest, field) \ 2903 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1); 2904 2905 static void 2906 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2, 2907 hammer_base_elm_t dest) 2908 { 2909 bzero(dest, sizeof(*dest)); 2910 2911 dest->rec_type = key2->rec_type; 2912 dest->key = key2->key; 2913 dest->obj_id = key2->obj_id; 2914 dest->create_tid = key2->create_tid; 2915 2916 MAKE_SEPARATOR(key1, key2, dest, localization); 2917 if (key1->localization == key2->localization) { 2918 MAKE_SEPARATOR(key1, key2, dest, obj_id); 2919 if (key1->obj_id == key2->obj_id) { 2920 MAKE_SEPARATOR(key1, key2, dest, rec_type); 2921 if (key1->rec_type == key2->rec_type) { 2922 MAKE_SEPARATOR(key1, key2, dest, key); 2923 /* 2924 * Don't bother creating a separator for 2925 * create_tid, which also conveniently avoids 2926 * having to handle the create_tid == 0 2927 * (infinity) case. Just leave create_tid 2928 * set to key2. 2929 * 2930 * Worst case, dest matches key2 exactly, 2931 * which is acceptable. 2932 */ 2933 } 2934 } 2935 } 2936 } 2937 2938 #undef MAKE_SEPARATOR 2939 2940 /* 2941 * Return whether a generic internal or leaf node is full 2942 */ 2943 static __inline 2944 int 2945 btree_node_is_full(hammer_node_ondisk_t node) 2946 { 2947 return(btree_max_elements(node->type) == node->count); 2948 } 2949 2950 static __inline 2951 int 2952 btree_max_elements(u_int8_t type) 2953 { 2954 int n; 2955 2956 n = hammer_node_max_elements(type); 2957 if (n == -1) 2958 hpanic("bad type %d", type); 2959 return(n); 2960 } 2961 2962 void 2963 hammer_print_btree_node(hammer_node_ondisk_t ondisk) 2964 { 2965 int i, n; 2966 2967 kprintf("node %p count=%d parent=%016llx type=%c\n", 2968 ondisk, ondisk->count, 2969 (long long)ondisk->parent, ondisk->type); 2970 2971 switch (ondisk->type) { 2972 case HAMMER_BTREE_TYPE_INTERNAL: 2973 n = ondisk->count + 1; /* count is NOT boundary inclusive */ 2974 break; 2975 case HAMMER_BTREE_TYPE_LEAF: 2976 n = ondisk->count; /* there is no boundary */ 2977 break; 2978 default: 2979 return; /* nothing to do */ 2980 } 2981 2982 /* 2983 * Dump elements including boundary. 2984 */ 2985 for (i = 0; i < n; ++i) { 2986 kprintf(" %2d", i); 2987 hammer_print_btree_elm(&ondisk->elms[i]); 2988 } 2989 } 2990 2991 void 2992 hammer_print_btree_elm(hammer_btree_elm_t elm) 2993 { 2994 kprintf("\tobj_id = %016llx\n", (long long)elm->base.obj_id); 2995 kprintf("\tkey = %016llx\n", (long long)elm->base.key); 2996 kprintf("\tcreate_tid = %016llx\n", (long long)elm->base.create_tid); 2997 kprintf("\tdelete_tid = %016llx\n", (long long)elm->base.delete_tid); 2998 kprintf("\trec_type = %04x\n", elm->base.rec_type); 2999 kprintf("\tobj_type = %02x\n", elm->base.obj_type); 3000 kprintf("\tbtype = %02x (%c)\n", elm->base.btype, 3001 hammer_elm_btype(elm)); 3002 kprintf("\tlocalization = %08x\n", elm->base.localization); 3003 3004 if (hammer_is_internal_node_elm(elm)) { 3005 kprintf("\tsubtree_off = %016llx\n", 3006 (long long)elm->internal.subtree_offset); 3007 } else if (hammer_is_leaf_node_elm(elm)) { 3008 kprintf("\tdata_offset = %016llx\n", 3009 (long long)elm->leaf.data_offset); 3010 kprintf("\tdata_len = %08x\n", elm->leaf.data_len); 3011 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc); 3012 } 3013 } 3014 3015 static __inline 3016 void 3017 hammer_debug_btree_elm(hammer_cursor_t cursor, hammer_btree_elm_t elm, 3018 const char *s, int res) 3019 { 3020 hkprintf("%-8s %016llx[%02d] %c " 3021 "lo=%08x obj=%016llx rec=%02x key=%016llx tid=%016llx td=%p " 3022 "r=%d\n", 3023 s, 3024 (long long)cursor->node->node_offset, 3025 cursor->index, 3026 hammer_elm_btype(elm), 3027 elm->base.localization, 3028 (long long)elm->base.obj_id, 3029 elm->base.rec_type, 3030 (long long)elm->base.key, 3031 (long long)elm->base.create_tid, 3032 curthread, 3033 res); 3034 } 3035 3036 static __inline 3037 void 3038 hammer_debug_btree_parent(hammer_cursor_t cursor, const char *s) 3039 { 3040 hammer_btree_elm_t elm = 3041 &cursor->parent->ondisk->elms[cursor->parent_index]; 3042 3043 hkprintf("%-8s %016llx[%d] %c " 3044 "(%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n", 3045 s, 3046 (long long)cursor->parent->node_offset, 3047 cursor->parent_index, 3048 hammer_elm_btype(elm), 3049 (long long)cursor->left_bound->obj_id, 3050 (long long)elm->internal.base.obj_id, 3051 (long long)cursor->right_bound->obj_id, 3052 (long long)(elm + 1)->internal.base.obj_id, 3053 cursor->left_bound, 3054 elm, 3055 cursor->right_bound, 3056 elm + 1); 3057 } 3058