1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.76 2008/08/06 15:38:58 dillon Exp $ 35 */ 36 37 /* 38 * HAMMER B-Tree index 39 * 40 * HAMMER implements a modified B+Tree. In documentation this will 41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree 42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs 43 * of the tree), but adds two additional boundary elements which describe 44 * the left-most and right-most element a node is able to represent. In 45 * otherwords, we have boundary elements at the two ends of a B-Tree node 46 * instead of sub-tree pointers. 47 * 48 * A B-Tree internal node looks like this: 49 * 50 * B N N N N N N B <-- boundary and internal elements 51 * S S S S S S S <-- subtree pointers 52 * 53 * A B-Tree leaf node basically looks like this: 54 * 55 * L L L L L L L L <-- leaf elemenets 56 * 57 * The radix for an internal node is 1 less then a leaf but we get a 58 * number of significant benefits for our troubles. 59 * 60 * The big benefit to using a B-Tree containing boundary information 61 * is that it is possible to cache pointers into the middle of the tree 62 * and not have to start searches, insertions, OR deletions at the root 63 * node. In particular, searches are able to progress in a definitive 64 * direction from any point in the tree without revisting nodes. This 65 * greatly improves the efficiency of many operations, most especially 66 * record appends. 67 * 68 * B-Trees also make the stacking of trees fairly straightforward. 69 * 70 * INSERTIONS: A search performed with the intention of doing 71 * an insert will guarantee that the terminal leaf node is not full by 72 * splitting full nodes. Splits occur top-down during the dive down the 73 * B-Tree. 74 * 75 * DELETIONS: A deletion makes no attempt to proactively balance the 76 * tree and will recursively remove nodes that become empty. If a 77 * deadlock occurs a deletion may not be able to remove an empty leaf. 78 * Deletions never allow internal nodes to become empty (that would blow 79 * up the boundaries). 80 */ 81 #include "hammer.h" 82 #include <sys/buf.h> 83 #include <sys/buf2.h> 84 85 static int btree_search(hammer_cursor_t cursor, int flags); 86 static int btree_split_internal(hammer_cursor_t cursor); 87 static int btree_split_leaf(hammer_cursor_t cursor); 88 static int btree_remove(hammer_cursor_t cursor); 89 static int btree_node_is_full(hammer_node_ondisk_t node); 90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor, 91 hammer_tid_t mirror_tid); 92 static void hammer_make_separator(hammer_base_elm_t key1, 93 hammer_base_elm_t key2, hammer_base_elm_t dest); 94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor); 95 96 /* 97 * Iterate records after a search. The cursor is iterated forwards past 98 * the current record until a record matching the key-range requirements 99 * is found. ENOENT is returned if the iteration goes past the ending 100 * key. 101 * 102 * The iteration is inclusive of key_beg and can be inclusive or exclusive 103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set. 104 * 105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid 106 * may be modified by B-Tree functions. 107 * 108 * cursor->key_beg may or may not be modified by this function during 109 * the iteration. XXX future - in case of an inverted lock we may have 110 * to reinitiate the lookup and set key_beg to properly pick up where we 111 * left off. 112 * 113 * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor 114 * was reverse indexed due to being moved to a parent while unlocked, 115 * and something else might have inserted an element outside the iteration 116 * range. When this case occurs the iterator just keeps iterating until 117 * it gets back into the iteration range (instead of asserting). 118 * 119 * NOTE! EDEADLK *CANNOT* be returned by this procedure. 120 */ 121 int 122 hammer_btree_iterate(hammer_cursor_t cursor) 123 { 124 hammer_node_ondisk_t node; 125 hammer_btree_elm_t elm; 126 hammer_mount_t hmp; 127 int error = 0; 128 int r; 129 int s; 130 131 /* 132 * Skip past the current record 133 */ 134 hmp = cursor->trans->hmp; 135 node = cursor->node->ondisk; 136 if (node == NULL) 137 return(ENOENT); 138 if (cursor->index < node->count && 139 (cursor->flags & HAMMER_CURSOR_ATEDISK)) { 140 ++cursor->index; 141 } 142 143 /* 144 * HAMMER can wind up being cpu-bound. 145 */ 146 if (++hmp->check_yield > hammer_yield_check) { 147 hmp->check_yield = 0; 148 lwkt_user_yield(); 149 } 150 151 152 /* 153 * Loop until an element is found or we are done. 154 */ 155 for (;;) { 156 /* 157 * We iterate up the tree and then index over one element 158 * while we are at the last element in the current node. 159 * 160 * If we are at the root of the filesystem, cursor_up 161 * returns ENOENT. 162 * 163 * XXX this could be optimized by storing the information in 164 * the parent reference. 165 * 166 * XXX we can lose the node lock temporarily, this could mess 167 * up our scan. 168 */ 169 ++hammer_stats_btree_iterations; 170 hammer_flusher_clean_loose_ios(hmp); 171 172 if (cursor->index == node->count) { 173 if (hammer_debug_btree) { 174 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n", 175 (long long)cursor->node->node_offset, 176 cursor->index, 177 (long long)(cursor->parent ? cursor->parent->node_offset : -1), 178 cursor->parent_index, 179 curthread); 180 } 181 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset); 182 error = hammer_cursor_up(cursor); 183 if (error) 184 break; 185 /* reload stale pointer */ 186 node = cursor->node->ondisk; 187 KKASSERT(cursor->index != node->count); 188 189 /* 190 * If we are reblocking we want to return internal 191 * nodes. Note that the internal node will be 192 * returned multiple times, on each upward recursion 193 * from its children. The caller selects which 194 * revisit it cares about (usually first or last only). 195 */ 196 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) { 197 cursor->flags |= HAMMER_CURSOR_ATEDISK; 198 return(0); 199 } 200 ++cursor->index; 201 continue; 202 } 203 204 /* 205 * Check internal or leaf element. Determine if the record 206 * at the cursor has gone beyond the end of our range. 207 * 208 * We recurse down through internal nodes. 209 */ 210 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 211 elm = &node->elms[cursor->index]; 212 213 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base); 214 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base); 215 if (hammer_debug_btree) { 216 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n", 217 (long long)cursor->node->node_offset, 218 cursor->index, 219 (long long)elm[0].internal.base.obj_id, 220 elm[0].internal.base.rec_type, 221 (long long)elm[0].internal.base.key, 222 elm[0].internal.base.localization, 223 r, 224 curthread 225 ); 226 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n", 227 (long long)cursor->node->node_offset, 228 cursor->index + 1, 229 (long long)elm[1].internal.base.obj_id, 230 elm[1].internal.base.rec_type, 231 (long long)elm[1].internal.base.key, 232 elm[1].internal.base.localization, 233 s 234 ); 235 } 236 237 if (r < 0) { 238 error = ENOENT; 239 break; 240 } 241 if (r == 0 && (cursor->flags & 242 HAMMER_CURSOR_END_INCLUSIVE) == 0) { 243 error = ENOENT; 244 break; 245 } 246 247 /* 248 * Better not be zero 249 */ 250 KKASSERT(elm->internal.subtree_offset != 0); 251 252 if (s <= 0) { 253 /* 254 * If running the mirror filter see if we 255 * can skip one or more entire sub-trees. 256 * If we can we return the internal node 257 * and the caller processes the skipped 258 * range (see mirror_read). 259 */ 260 if (cursor->flags & 261 HAMMER_CURSOR_MIRROR_FILTERED) { 262 if (elm->internal.mirror_tid < 263 cursor->cmirror->mirror_tid) { 264 hammer_cursor_mirror_filter(cursor); 265 return(0); 266 } 267 } 268 } else { 269 /* 270 * Normally it would be impossible for the 271 * cursor to have gotten back-indexed, 272 * but it can happen if a node is deleted 273 * and the cursor is moved to its parent 274 * internal node. ITERATE_CHECK will be set. 275 */ 276 KKASSERT(cursor->flags & 277 HAMMER_CURSOR_ITERATE_CHECK); 278 kprintf("hammer_btree_iterate: " 279 "DEBUG: Caught parent seek " 280 "in internal iteration\n"); 281 } 282 283 error = hammer_cursor_down(cursor); 284 if (error) 285 break; 286 KKASSERT(cursor->index == 0); 287 /* reload stale pointer */ 288 node = cursor->node->ondisk; 289 continue; 290 } else { 291 elm = &node->elms[cursor->index]; 292 r = hammer_btree_cmp(&cursor->key_end, &elm->base); 293 if (hammer_debug_btree) { 294 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n", 295 (long long)cursor->node->node_offset, 296 cursor->index, 297 (elm[0].leaf.base.btype ? 298 elm[0].leaf.base.btype : '?'), 299 (long long)elm[0].leaf.base.obj_id, 300 elm[0].leaf.base.rec_type, 301 (long long)elm[0].leaf.base.key, 302 elm[0].leaf.base.localization, 303 r 304 ); 305 } 306 if (r < 0) { 307 error = ENOENT; 308 break; 309 } 310 311 /* 312 * We support both end-inclusive and 313 * end-exclusive searches. 314 */ 315 if (r == 0 && 316 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) { 317 error = ENOENT; 318 break; 319 } 320 321 /* 322 * If ITERATE_CHECK is set an unlocked cursor may 323 * have been moved to a parent and the iterate can 324 * happen upon elements that are not in the requested 325 * range. 326 */ 327 if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) { 328 s = hammer_btree_cmp(&cursor->key_beg, 329 &elm->base); 330 if (s > 0) { 331 kprintf("hammer_btree_iterate: " 332 "DEBUG: Caught parent seek " 333 "in leaf iteration\n"); 334 ++cursor->index; 335 continue; 336 } 337 } 338 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 339 340 /* 341 * Return the element 342 */ 343 switch(elm->leaf.base.btype) { 344 case HAMMER_BTREE_TYPE_RECORD: 345 if ((cursor->flags & HAMMER_CURSOR_ASOF) && 346 hammer_btree_chkts(cursor->asof, &elm->base)) { 347 ++cursor->index; 348 continue; 349 } 350 error = 0; 351 break; 352 default: 353 error = EINVAL; 354 break; 355 } 356 if (error) 357 break; 358 } 359 /* 360 * node pointer invalid after loop 361 */ 362 363 /* 364 * Return entry 365 */ 366 if (hammer_debug_btree) { 367 int i = cursor->index; 368 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i]; 369 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n", 370 cursor->node, i, 371 (long long)elm->internal.base.obj_id, 372 elm->internal.base.rec_type, 373 (long long)elm->internal.base.key, 374 elm->internal.base.localization 375 ); 376 } 377 return(0); 378 } 379 return(error); 380 } 381 382 /* 383 * We hit an internal element that we could skip as part of a mirroring 384 * scan. Calculate the entire range being skipped. 385 * 386 * It is important to include any gaps between the parent's left_bound 387 * and the node's left_bound, and same goes for the right side. 388 */ 389 static void 390 hammer_cursor_mirror_filter(hammer_cursor_t cursor) 391 { 392 struct hammer_cmirror *cmirror; 393 hammer_node_ondisk_t ondisk; 394 hammer_btree_elm_t elm; 395 396 ondisk = cursor->node->ondisk; 397 cmirror = cursor->cmirror; 398 399 /* 400 * Calculate the skipped range 401 */ 402 elm = &ondisk->elms[cursor->index]; 403 if (cursor->index == 0) 404 cmirror->skip_beg = *cursor->left_bound; 405 else 406 cmirror->skip_beg = elm->internal.base; 407 while (cursor->index < ondisk->count) { 408 if (elm->internal.mirror_tid >= cmirror->mirror_tid) 409 break; 410 ++cursor->index; 411 ++elm; 412 } 413 if (cursor->index == ondisk->count) 414 cmirror->skip_end = *cursor->right_bound; 415 else 416 cmirror->skip_end = elm->internal.base; 417 418 /* 419 * clip the returned result. 420 */ 421 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0) 422 cmirror->skip_beg = cursor->key_beg; 423 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0) 424 cmirror->skip_end = cursor->key_end; 425 } 426 427 /* 428 * Iterate in the reverse direction. This is used by the pruning code to 429 * avoid overlapping records. 430 */ 431 int 432 hammer_btree_iterate_reverse(hammer_cursor_t cursor) 433 { 434 hammer_node_ondisk_t node; 435 hammer_btree_elm_t elm; 436 hammer_mount_t hmp; 437 int error = 0; 438 int r; 439 int s; 440 441 /* mirror filtering not supported for reverse iteration */ 442 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0); 443 444 /* 445 * Skip past the current record. For various reasons the cursor 446 * may end up set to -1 or set to point at the end of the current 447 * node. These cases must be addressed. 448 */ 449 node = cursor->node->ondisk; 450 if (node == NULL) 451 return(ENOENT); 452 if (cursor->index != -1 && 453 (cursor->flags & HAMMER_CURSOR_ATEDISK)) { 454 --cursor->index; 455 } 456 if (cursor->index == cursor->node->ondisk->count) 457 --cursor->index; 458 459 /* 460 * HAMMER can wind up being cpu-bound. 461 */ 462 hmp = cursor->trans->hmp; 463 if (++hmp->check_yield > hammer_yield_check) { 464 hmp->check_yield = 0; 465 lwkt_user_yield(); 466 } 467 468 /* 469 * Loop until an element is found or we are done. 470 */ 471 for (;;) { 472 ++hammer_stats_btree_iterations; 473 hammer_flusher_clean_loose_ios(hmp); 474 475 /* 476 * We iterate up the tree and then index over one element 477 * while we are at the last element in the current node. 478 */ 479 if (cursor->index == -1) { 480 error = hammer_cursor_up(cursor); 481 if (error) { 482 cursor->index = 0; /* sanity */ 483 break; 484 } 485 /* reload stale pointer */ 486 node = cursor->node->ondisk; 487 KKASSERT(cursor->index != node->count); 488 --cursor->index; 489 continue; 490 } 491 492 /* 493 * Check internal or leaf element. Determine if the record 494 * at the cursor has gone beyond the end of our range. 495 * 496 * We recurse down through internal nodes. 497 */ 498 KKASSERT(cursor->index != node->count); 499 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 500 elm = &node->elms[cursor->index]; 501 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base); 502 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base); 503 if (hammer_debug_btree) { 504 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n", 505 (long long)cursor->node->node_offset, 506 cursor->index, 507 (long long)elm[0].internal.base.obj_id, 508 elm[0].internal.base.rec_type, 509 (long long)elm[0].internal.base.key, 510 elm[0].internal.base.localization, 511 r 512 ); 513 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n", 514 (long long)cursor->node->node_offset, 515 cursor->index + 1, 516 (long long)elm[1].internal.base.obj_id, 517 elm[1].internal.base.rec_type, 518 (long long)elm[1].internal.base.key, 519 elm[1].internal.base.localization, 520 s 521 ); 522 } 523 524 if (s >= 0) { 525 error = ENOENT; 526 break; 527 } 528 529 /* 530 * It shouldn't be possible to be seeked past key_end, 531 * even if the cursor got moved to a parent. 532 */ 533 KKASSERT(r >= 0); 534 535 /* 536 * Better not be zero 537 */ 538 KKASSERT(elm->internal.subtree_offset != 0); 539 540 error = hammer_cursor_down(cursor); 541 if (error) 542 break; 543 KKASSERT(cursor->index == 0); 544 /* reload stale pointer */ 545 node = cursor->node->ondisk; 546 547 /* this can assign -1 if the leaf was empty */ 548 cursor->index = node->count - 1; 549 continue; 550 } else { 551 elm = &node->elms[cursor->index]; 552 s = hammer_btree_cmp(&cursor->key_beg, &elm->base); 553 if (hammer_debug_btree) { 554 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n", 555 (long long)cursor->node->node_offset, 556 cursor->index, 557 (elm[0].leaf.base.btype ? 558 elm[0].leaf.base.btype : '?'), 559 (long long)elm[0].leaf.base.obj_id, 560 elm[0].leaf.base.rec_type, 561 (long long)elm[0].leaf.base.key, 562 elm[0].leaf.base.localization, 563 s 564 ); 565 } 566 if (s > 0) { 567 error = ENOENT; 568 break; 569 } 570 571 /* 572 * It shouldn't be possible to be seeked past key_end, 573 * even if the cursor got moved to a parent. 574 */ 575 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 576 577 /* 578 * Return the element 579 */ 580 switch(elm->leaf.base.btype) { 581 case HAMMER_BTREE_TYPE_RECORD: 582 if ((cursor->flags & HAMMER_CURSOR_ASOF) && 583 hammer_btree_chkts(cursor->asof, &elm->base)) { 584 --cursor->index; 585 continue; 586 } 587 error = 0; 588 break; 589 default: 590 error = EINVAL; 591 break; 592 } 593 if (error) 594 break; 595 } 596 /* 597 * node pointer invalid after loop 598 */ 599 600 /* 601 * Return entry 602 */ 603 if (hammer_debug_btree) { 604 int i = cursor->index; 605 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i]; 606 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n", 607 cursor->node, i, 608 (long long)elm->internal.base.obj_id, 609 elm->internal.base.rec_type, 610 (long long)elm->internal.base.key, 611 elm->internal.base.localization 612 ); 613 } 614 return(0); 615 } 616 return(error); 617 } 618 619 /* 620 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry 621 * could not be found, EDEADLK if inserting and a retry is needed, and a 622 * fatal error otherwise. When retrying, the caller must terminate the 623 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting. 624 * 625 * The cursor is suitably positioned for a deletion on success, and suitably 626 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was 627 * specified. 628 * 629 * The cursor may begin anywhere, the search will traverse the tree in 630 * either direction to locate the requested element. 631 * 632 * Most of the logic implementing historical searches is handled here. We 633 * do an initial lookup with create_tid set to the asof TID. Due to the 634 * way records are laid out, a backwards iteration may be required if 635 * ENOENT is returned to locate the historical record. Here's the 636 * problem: 637 * 638 * create_tid: 10 15 20 639 * LEAF1 LEAF2 640 * records: (11) (18) 641 * 642 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse 643 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is 644 * not visible and thus causes ENOENT to be returned. We really need 645 * to check record 11 in LEAF1. If it also fails then the search fails 646 * (e.g. it might represent the range 11-16 and thus still not match our 647 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring 648 * further iterations. 649 * 650 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK 651 * and the cursor->create_check TID if an iteration might be needed. 652 * In the above example create_check would be set to 14. 653 */ 654 int 655 hammer_btree_lookup(hammer_cursor_t cursor) 656 { 657 int error; 658 659 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK; 660 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 || 661 cursor->trans->sync_lock_refs > 0); 662 ++hammer_stats_btree_lookups; 663 if (cursor->flags & HAMMER_CURSOR_ASOF) { 664 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0); 665 cursor->key_beg.create_tid = cursor->asof; 666 for (;;) { 667 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK; 668 error = btree_search(cursor, 0); 669 if (error != ENOENT || 670 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) { 671 /* 672 * Stop if no error. 673 * Stop if error other then ENOENT. 674 * Stop if ENOENT and not special case. 675 */ 676 break; 677 } 678 if (hammer_debug_btree) { 679 kprintf("CREATE_CHECK %016llx\n", 680 (long long)cursor->create_check); 681 } 682 cursor->key_beg.create_tid = cursor->create_check; 683 /* loop */ 684 } 685 } else { 686 error = btree_search(cursor, 0); 687 } 688 if (error == 0) 689 error = hammer_btree_extract(cursor, cursor->flags); 690 return(error); 691 } 692 693 /* 694 * Execute the logic required to start an iteration. The first record 695 * located within the specified range is returned and iteration control 696 * flags are adjusted for successive hammer_btree_iterate() calls. 697 * 698 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate 699 * in a loop without worrying about it. Higher-level merged searches will 700 * adjust the flag appropriately. 701 */ 702 int 703 hammer_btree_first(hammer_cursor_t cursor) 704 { 705 int error; 706 707 error = hammer_btree_lookup(cursor); 708 if (error == ENOENT) { 709 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 710 error = hammer_btree_iterate(cursor); 711 } 712 cursor->flags |= HAMMER_CURSOR_ATEDISK; 713 return(error); 714 } 715 716 /* 717 * Similarly but for an iteration in the reverse direction. 718 * 719 * Set ATEDISK when iterating backwards to skip the current entry, 720 * which after an ENOENT lookup will be pointing beyond our end point. 721 * 722 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse 723 * in a loop without worrying about it. Higher-level merged searches will 724 * adjust the flag appropriately. 725 */ 726 int 727 hammer_btree_last(hammer_cursor_t cursor) 728 { 729 struct hammer_base_elm save; 730 int error; 731 732 save = cursor->key_beg; 733 cursor->key_beg = cursor->key_end; 734 error = hammer_btree_lookup(cursor); 735 cursor->key_beg = save; 736 if (error == ENOENT || 737 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) { 738 cursor->flags |= HAMMER_CURSOR_ATEDISK; 739 error = hammer_btree_iterate_reverse(cursor); 740 } 741 cursor->flags |= HAMMER_CURSOR_ATEDISK; 742 return(error); 743 } 744 745 /* 746 * Extract the record and/or data associated with the cursor's current 747 * position. Any prior record or data stored in the cursor is replaced. 748 * The cursor must be positioned at a leaf node. 749 * 750 * NOTE: All extractions occur at the leaf of the B-Tree. 751 */ 752 int 753 hammer_btree_extract(hammer_cursor_t cursor, int flags) 754 { 755 hammer_node_ondisk_t node; 756 hammer_btree_elm_t elm; 757 hammer_off_t data_off; 758 hammer_mount_t hmp; 759 int32_t data_len; 760 int error; 761 762 /* 763 * The case where the data reference resolves to the same buffer 764 * as the record reference must be handled. 765 */ 766 node = cursor->node->ondisk; 767 elm = &node->elms[cursor->index]; 768 cursor->data = NULL; 769 hmp = cursor->node->hmp; 770 771 /* 772 * There is nothing to extract for an internal element. 773 */ 774 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) 775 return(EINVAL); 776 777 /* 778 * Only record types have data. 779 */ 780 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF); 781 cursor->leaf = &elm->leaf; 782 783 if ((flags & HAMMER_CURSOR_GET_DATA) == 0) 784 return(0); 785 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) 786 return(0); 787 data_off = elm->leaf.data_offset; 788 data_len = elm->leaf.data_len; 789 if (data_off == 0) 790 return(0); 791 792 /* 793 * Load the data 794 */ 795 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE); 796 cursor->data = hammer_bread_ext(hmp, data_off, data_len, 797 &error, &cursor->data_buffer); 798 799 /* 800 * Mark the data buffer as not being meta-data if it isn't 801 * meta-data (sometimes bulk data is accessed via a volume 802 * block device). 803 */ 804 if (error == 0) { 805 switch(elm->leaf.base.rec_type) { 806 case HAMMER_RECTYPE_DATA: 807 case HAMMER_RECTYPE_DB: 808 hammer_io_notmeta(cursor->data_buffer); 809 break; 810 default: 811 break; 812 } 813 } 814 815 /* 816 * Deal with CRC errors on the extracted data. 817 */ 818 if (error == 0 && 819 hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) { 820 kprintf("CRC DATA @ %016llx/%d FAILED\n", 821 (long long)elm->leaf.data_offset, elm->leaf.data_len); 822 if (hammer_debug_critical) 823 Debugger("CRC FAILED: DATA"); 824 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM) 825 error = EDOM; /* less critical (mirroring) */ 826 else 827 error = EIO; /* critical */ 828 } 829 return(error); 830 } 831 832 833 /* 834 * Insert a leaf element into the B-Tree at the current cursor position. 835 * The cursor is positioned such that the element at and beyond the cursor 836 * are shifted to make room for the new record. 837 * 838 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT 839 * flag set and that call must return ENOENT before this function can be 840 * called. 841 * 842 * The caller may depend on the cursor's exclusive lock after return to 843 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE). 844 * 845 * ENOSPC is returned if there is no room to insert a new record. 846 */ 847 int 848 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm, 849 int *doprop) 850 { 851 hammer_node_ondisk_t node; 852 int i; 853 int error; 854 855 *doprop = 0; 856 if ((error = hammer_cursor_upgrade_node(cursor)) != 0) 857 return(error); 858 ++hammer_stats_btree_inserts; 859 860 /* 861 * Insert the element at the leaf node and update the count in the 862 * parent. It is possible for parent to be NULL, indicating that 863 * the filesystem's ROOT B-Tree node is a leaf itself, which is 864 * possible. The root inode can never be deleted so the leaf should 865 * never be empty. 866 * 867 * Remember that the right-hand boundary is not included in the 868 * count. 869 */ 870 hammer_modify_node_all(cursor->trans, cursor->node); 871 node = cursor->node->ondisk; 872 i = cursor->index; 873 KKASSERT(elm->base.btype != 0); 874 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF); 875 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS); 876 if (i != node->count) { 877 bcopy(&node->elms[i], &node->elms[i+1], 878 (node->count - i) * sizeof(*elm)); 879 } 880 node->elms[i].leaf = *elm; 881 ++node->count; 882 hammer_cursor_inserted_element(cursor->node, i); 883 884 /* 885 * Update the leaf node's aggregate mirror_tid for mirroring 886 * support. 887 */ 888 if (node->mirror_tid < elm->base.delete_tid) { 889 node->mirror_tid = elm->base.delete_tid; 890 *doprop = 1; 891 } 892 if (node->mirror_tid < elm->base.create_tid) { 893 node->mirror_tid = elm->base.create_tid; 894 *doprop = 1; 895 } 896 hammer_modify_node_done(cursor->node); 897 898 /* 899 * Debugging sanity checks. 900 */ 901 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0); 902 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0); 903 if (i) { 904 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0); 905 } 906 if (i != node->count - 1) 907 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0); 908 909 return(0); 910 } 911 912 /* 913 * Delete a record from the B-Tree at the current cursor position. 914 * The cursor is positioned such that the current element is the one 915 * to be deleted. 916 * 917 * On return the cursor will be positioned after the deleted element and 918 * MAY point to an internal node. It will be suitable for the continuation 919 * of an iteration but not for an insertion or deletion. 920 * 921 * Deletions will attempt to partially rebalance the B-Tree in an upward 922 * direction, but will terminate rather then deadlock. Empty internal nodes 923 * are never allowed by a deletion which deadlocks may end up giving us an 924 * empty leaf. The pruner will clean up and rebalance the tree. 925 * 926 * This function can return EDEADLK, requiring the caller to retry the 927 * operation after clearing the deadlock. 928 */ 929 int 930 hammer_btree_delete(hammer_cursor_t cursor) 931 { 932 hammer_node_ondisk_t ondisk; 933 hammer_node_t node; 934 hammer_node_t parent; 935 int error; 936 int i; 937 938 KKASSERT (cursor->trans->sync_lock_refs > 0); 939 if ((error = hammer_cursor_upgrade(cursor)) != 0) 940 return(error); 941 ++hammer_stats_btree_deletes; 942 943 /* 944 * Delete the element from the leaf node. 945 * 946 * Remember that leaf nodes do not have boundaries. 947 */ 948 node = cursor->node; 949 ondisk = node->ondisk; 950 i = cursor->index; 951 952 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF); 953 KKASSERT(i >= 0 && i < ondisk->count); 954 hammer_modify_node_all(cursor->trans, node); 955 if (i + 1 != ondisk->count) { 956 bcopy(&ondisk->elms[i+1], &ondisk->elms[i], 957 (ondisk->count - i - 1) * sizeof(ondisk->elms[0])); 958 } 959 --ondisk->count; 960 hammer_modify_node_done(node); 961 hammer_cursor_deleted_element(node, i); 962 963 /* 964 * Validate local parent 965 */ 966 if (ondisk->parent) { 967 parent = cursor->parent; 968 969 KKASSERT(parent != NULL); 970 KKASSERT(parent->node_offset == ondisk->parent); 971 } 972 973 /* 974 * If the leaf becomes empty it must be detached from the parent, 975 * potentially recursing through to the filesystem root. 976 * 977 * This may reposition the cursor at one of the parent's of the 978 * current node. 979 * 980 * Ignore deadlock errors, that simply means that btree_remove 981 * was unable to recurse and had to leave us with an empty leaf. 982 */ 983 KKASSERT(cursor->index <= ondisk->count); 984 if (ondisk->count == 0) { 985 error = btree_remove(cursor); 986 if (error == EDEADLK) 987 error = 0; 988 } else { 989 error = 0; 990 } 991 KKASSERT(cursor->parent == NULL || 992 cursor->parent_index < cursor->parent->ondisk->count); 993 return(error); 994 } 995 996 /* 997 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE 998 * 999 * Search the filesystem B-Tree for cursor->key_beg, return the matching node. 1000 * 1001 * The search can begin ANYWHERE in the B-Tree. As a first step the search 1002 * iterates up the tree as necessary to properly position itself prior to 1003 * actually doing the sarch. 1004 * 1005 * INSERTIONS: The search will split full nodes and leaves on its way down 1006 * and guarentee that the leaf it ends up on is not full. If we run out 1007 * of space the search continues to the leaf (to position the cursor for 1008 * the spike), but ENOSPC is returned. 1009 * 1010 * The search is only guarenteed to end up on a leaf if an error code of 0 1011 * is returned, or if inserting and an error code of ENOENT is returned. 1012 * Otherwise it can stop at an internal node. On success a search returns 1013 * a leaf node. 1014 * 1015 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire 1016 * filesystem, and it is not simple code. Please note the following facts: 1017 * 1018 * - Internal node recursions have a boundary on the left AND right. The 1019 * right boundary is non-inclusive. The create_tid is a generic part 1020 * of the key for internal nodes. 1021 * 1022 * - Leaf nodes contain terminal elements only now. 1023 * 1024 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a 1025 * historical search. ASOF and INSERT are mutually exclusive. When 1026 * doing an as-of lookup btree_search() checks for a right-edge boundary 1027 * case. If while recursing down the left-edge differs from the key 1028 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along 1029 * with cursor->create_check. This is used by btree_lookup() to iterate. 1030 * The iteration backwards because as-of searches can wind up going 1031 * down the wrong branch of the B-Tree. 1032 */ 1033 static 1034 int 1035 btree_search(hammer_cursor_t cursor, int flags) 1036 { 1037 hammer_node_ondisk_t node; 1038 hammer_btree_elm_t elm; 1039 int error; 1040 int enospc = 0; 1041 int i; 1042 int r; 1043 int s; 1044 1045 flags |= cursor->flags; 1046 ++hammer_stats_btree_searches; 1047 1048 if (hammer_debug_btree) { 1049 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n", 1050 (long long)cursor->node->node_offset, 1051 cursor->index, 1052 (long long)cursor->key_beg.obj_id, 1053 cursor->key_beg.rec_type, 1054 (long long)cursor->key_beg.key, 1055 (long long)cursor->key_beg.create_tid, 1056 cursor->key_beg.localization, 1057 curthread 1058 ); 1059 if (cursor->parent) 1060 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n", 1061 (long long)cursor->parent->node_offset, 1062 cursor->parent_index, 1063 (long long)cursor->left_bound->obj_id, 1064 (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id, 1065 (long long)cursor->right_bound->obj_id, 1066 (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id, 1067 cursor->left_bound, 1068 &cursor->parent->ondisk->elms[cursor->parent_index], 1069 cursor->right_bound, 1070 &cursor->parent->ondisk->elms[cursor->parent_index+1] 1071 ); 1072 } 1073 1074 /* 1075 * Move our cursor up the tree until we find a node whos range covers 1076 * the key we are trying to locate. 1077 * 1078 * The left bound is inclusive, the right bound is non-inclusive. 1079 * It is ok to cursor up too far. 1080 */ 1081 for (;;) { 1082 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound); 1083 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound); 1084 if (r >= 0 && s < 0) 1085 break; 1086 KKASSERT(cursor->parent); 1087 ++hammer_stats_btree_iterations; 1088 error = hammer_cursor_up(cursor); 1089 if (error) 1090 goto done; 1091 } 1092 1093 /* 1094 * The delete-checks below are based on node, not parent. Set the 1095 * initial delete-check based on the parent. 1096 */ 1097 if (r == 1) { 1098 KKASSERT(cursor->left_bound->create_tid != 1); 1099 cursor->create_check = cursor->left_bound->create_tid - 1; 1100 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK; 1101 } 1102 1103 /* 1104 * We better have ended up with a node somewhere. 1105 */ 1106 KKASSERT(cursor->node != NULL); 1107 1108 /* 1109 * If we are inserting we can't start at a full node if the parent 1110 * is also full (because there is no way to split the node), 1111 * continue running up the tree until the requirement is satisfied 1112 * or we hit the root of the filesystem. 1113 * 1114 * (If inserting we aren't doing an as-of search so we don't have 1115 * to worry about create_check). 1116 */ 1117 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) { 1118 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 1119 if (btree_node_is_full(cursor->node->ondisk) == 0) 1120 break; 1121 } else { 1122 if (btree_node_is_full(cursor->node->ondisk) ==0) 1123 break; 1124 } 1125 if (cursor->node->ondisk->parent == 0 || 1126 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) { 1127 break; 1128 } 1129 ++hammer_stats_btree_iterations; 1130 error = hammer_cursor_up(cursor); 1131 /* node may have become stale */ 1132 if (error) 1133 goto done; 1134 } 1135 1136 /* 1137 * Push down through internal nodes to locate the requested key. 1138 */ 1139 node = cursor->node->ondisk; 1140 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) { 1141 /* 1142 * Scan the node to find the subtree index to push down into. 1143 * We go one-past, then back-up. 1144 * 1145 * We must proactively remove deleted elements which may 1146 * have been left over from a deadlocked btree_remove(). 1147 * 1148 * The left and right boundaries are included in the loop 1149 * in order to detect edge cases. 1150 * 1151 * If the separator only differs by create_tid (r == 1) 1152 * and we are doing an as-of search, we may end up going 1153 * down a branch to the left of the one containing the 1154 * desired key. This requires numerous special cases. 1155 */ 1156 ++hammer_stats_btree_iterations; 1157 if (hammer_debug_btree) { 1158 kprintf("SEARCH-I %016llx count=%d\n", 1159 (long long)cursor->node->node_offset, 1160 node->count); 1161 } 1162 1163 /* 1164 * Try to shortcut the search before dropping into the 1165 * linear loop. Locate the first node where r <= 1. 1166 */ 1167 i = hammer_btree_search_node(&cursor->key_beg, node); 1168 while (i <= node->count) { 1169 ++hammer_stats_btree_elements; 1170 elm = &node->elms[i]; 1171 r = hammer_btree_cmp(&cursor->key_beg, &elm->base); 1172 if (hammer_debug_btree > 2) { 1173 kprintf(" IELM %p %d r=%d\n", 1174 &node->elms[i], i, r); 1175 } 1176 if (r < 0) 1177 break; 1178 if (r == 1) { 1179 KKASSERT(elm->base.create_tid != 1); 1180 cursor->create_check = elm->base.create_tid - 1; 1181 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK; 1182 } 1183 ++i; 1184 } 1185 if (hammer_debug_btree) { 1186 kprintf("SEARCH-I preI=%d/%d r=%d\n", 1187 i, node->count, r); 1188 } 1189 1190 /* 1191 * These cases occur when the parent's idea of the boundary 1192 * is wider then the child's idea of the boundary, and 1193 * require special handling. If not inserting we can 1194 * terminate the search early for these cases but the 1195 * child's boundaries cannot be unconditionally modified. 1196 */ 1197 if (i == 0) { 1198 /* 1199 * If i == 0 the search terminated to the LEFT of the 1200 * left_boundary but to the RIGHT of the parent's left 1201 * boundary. 1202 */ 1203 u_int8_t save; 1204 1205 elm = &node->elms[0]; 1206 1207 /* 1208 * If we aren't inserting we can stop here. 1209 */ 1210 if ((flags & (HAMMER_CURSOR_INSERT | 1211 HAMMER_CURSOR_PRUNING)) == 0) { 1212 cursor->index = 0; 1213 return(ENOENT); 1214 } 1215 1216 /* 1217 * Correct a left-hand boundary mismatch. 1218 * 1219 * We can only do this if we can upgrade the lock, 1220 * and synchronized as a background cursor (i.e. 1221 * inserting or pruning). 1222 * 1223 * WARNING: We can only do this if inserting, i.e. 1224 * we are running on the backend. 1225 */ 1226 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1227 return(error); 1228 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 1229 hammer_modify_node_field(cursor->trans, cursor->node, 1230 elms[0]); 1231 save = node->elms[0].base.btype; 1232 node->elms[0].base = *cursor->left_bound; 1233 node->elms[0].base.btype = save; 1234 hammer_modify_node_done(cursor->node); 1235 } else if (i == node->count + 1) { 1236 /* 1237 * If i == node->count + 1 the search terminated to 1238 * the RIGHT of the right boundary but to the LEFT 1239 * of the parent's right boundary. If we aren't 1240 * inserting we can stop here. 1241 * 1242 * Note that the last element in this case is 1243 * elms[i-2] prior to adjustments to 'i'. 1244 */ 1245 --i; 1246 if ((flags & (HAMMER_CURSOR_INSERT | 1247 HAMMER_CURSOR_PRUNING)) == 0) { 1248 cursor->index = i; 1249 return (ENOENT); 1250 } 1251 1252 /* 1253 * Correct a right-hand boundary mismatch. 1254 * (actual push-down record is i-2 prior to 1255 * adjustments to i). 1256 * 1257 * We can only do this if we can upgrade the lock, 1258 * and synchronized as a background cursor (i.e. 1259 * inserting or pruning). 1260 * 1261 * WARNING: We can only do this if inserting, i.e. 1262 * we are running on the backend. 1263 */ 1264 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1265 return(error); 1266 elm = &node->elms[i]; 1267 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 1268 hammer_modify_node(cursor->trans, cursor->node, 1269 &elm->base, sizeof(elm->base)); 1270 elm->base = *cursor->right_bound; 1271 hammer_modify_node_done(cursor->node); 1272 --i; 1273 } else { 1274 /* 1275 * The push-down index is now i - 1. If we had 1276 * terminated on the right boundary this will point 1277 * us at the last element. 1278 */ 1279 --i; 1280 } 1281 cursor->index = i; 1282 elm = &node->elms[i]; 1283 1284 if (hammer_debug_btree) { 1285 kprintf("RESULT-I %016llx[%d] %016llx %02x " 1286 "key=%016llx cre=%016llx lo=%02x\n", 1287 (long long)cursor->node->node_offset, 1288 i, 1289 (long long)elm->internal.base.obj_id, 1290 elm->internal.base.rec_type, 1291 (long long)elm->internal.base.key, 1292 (long long)elm->internal.base.create_tid, 1293 elm->internal.base.localization 1294 ); 1295 } 1296 1297 /* 1298 * We better have a valid subtree offset. 1299 */ 1300 KKASSERT(elm->internal.subtree_offset != 0); 1301 1302 /* 1303 * Handle insertion and deletion requirements. 1304 * 1305 * If inserting split full nodes. The split code will 1306 * adjust cursor->node and cursor->index if the current 1307 * index winds up in the new node. 1308 * 1309 * If inserting and a left or right edge case was detected, 1310 * we cannot correct the left or right boundary and must 1311 * prepend and append an empty leaf node in order to make 1312 * the boundary correction. 1313 * 1314 * If we run out of space we set enospc and continue on 1315 * to a leaf to provide the spike code with a good point 1316 * of entry. 1317 */ 1318 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) { 1319 if (btree_node_is_full(node)) { 1320 error = btree_split_internal(cursor); 1321 if (error) { 1322 if (error != ENOSPC) 1323 goto done; 1324 enospc = 1; 1325 } 1326 /* 1327 * reload stale pointers 1328 */ 1329 i = cursor->index; 1330 node = cursor->node->ondisk; 1331 } 1332 } 1333 1334 /* 1335 * Push down (push into new node, existing node becomes 1336 * the parent) and continue the search. 1337 */ 1338 error = hammer_cursor_down(cursor); 1339 /* node may have become stale */ 1340 if (error) 1341 goto done; 1342 node = cursor->node->ondisk; 1343 } 1344 1345 /* 1346 * We are at a leaf, do a linear search of the key array. 1347 * 1348 * On success the index is set to the matching element and 0 1349 * is returned. 1350 * 1351 * On failure the index is set to the insertion point and ENOENT 1352 * is returned. 1353 * 1354 * Boundaries are not stored in leaf nodes, so the index can wind 1355 * up to the left of element 0 (index == 0) or past the end of 1356 * the array (index == node->count). It is also possible that the 1357 * leaf might be empty. 1358 */ 1359 ++hammer_stats_btree_iterations; 1360 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF); 1361 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS); 1362 if (hammer_debug_btree) { 1363 kprintf("SEARCH-L %016llx count=%d\n", 1364 (long long)cursor->node->node_offset, 1365 node->count); 1366 } 1367 1368 /* 1369 * Try to shortcut the search before dropping into the 1370 * linear loop. Locate the first node where r <= 1. 1371 */ 1372 i = hammer_btree_search_node(&cursor->key_beg, node); 1373 while (i < node->count) { 1374 ++hammer_stats_btree_elements; 1375 elm = &node->elms[i]; 1376 1377 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base); 1378 1379 if (hammer_debug_btree > 1) 1380 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r); 1381 1382 /* 1383 * We are at a record element. Stop if we've flipped past 1384 * key_beg, not counting the create_tid test. Allow the 1385 * r == 1 case (key_beg > element but differs only by its 1386 * create_tid) to fall through to the AS-OF check. 1387 */ 1388 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD); 1389 1390 if (r < 0) 1391 goto failed; 1392 if (r > 1) { 1393 ++i; 1394 continue; 1395 } 1396 1397 /* 1398 * Check our as-of timestamp against the element. 1399 */ 1400 if (flags & HAMMER_CURSOR_ASOF) { 1401 if (hammer_btree_chkts(cursor->asof, 1402 &node->elms[i].base) != 0) { 1403 ++i; 1404 continue; 1405 } 1406 /* success */ 1407 } else { 1408 if (r > 0) { /* can only be +1 */ 1409 ++i; 1410 continue; 1411 } 1412 /* success */ 1413 } 1414 cursor->index = i; 1415 error = 0; 1416 if (hammer_debug_btree) { 1417 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n", 1418 (long long)cursor->node->node_offset, i); 1419 } 1420 goto done; 1421 } 1422 1423 /* 1424 * The search of the leaf node failed. i is the insertion point. 1425 */ 1426 failed: 1427 if (hammer_debug_btree) { 1428 kprintf("RESULT-L %016llx[%d] (FAILED)\n", 1429 (long long)cursor->node->node_offset, i); 1430 } 1431 1432 /* 1433 * No exact match was found, i is now at the insertion point. 1434 * 1435 * If inserting split a full leaf before returning. This 1436 * may have the side effect of adjusting cursor->node and 1437 * cursor->index. 1438 */ 1439 cursor->index = i; 1440 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 && 1441 btree_node_is_full(node)) { 1442 error = btree_split_leaf(cursor); 1443 if (error) { 1444 if (error != ENOSPC) 1445 goto done; 1446 enospc = 1; 1447 } 1448 /* 1449 * reload stale pointers 1450 */ 1451 /* NOT USED 1452 i = cursor->index; 1453 node = &cursor->node->internal; 1454 */ 1455 } 1456 1457 /* 1458 * We reached a leaf but did not find the key we were looking for. 1459 * If this is an insert we will be properly positioned for an insert 1460 * (ENOENT) or spike (ENOSPC) operation. 1461 */ 1462 error = enospc ? ENOSPC : ENOENT; 1463 done: 1464 return(error); 1465 } 1466 1467 /* 1468 * Heuristical search for the first element whos comparison is <= 1. May 1469 * return an index whos compare result is > 1 but may only return an index 1470 * whos compare result is <= 1 if it is the first element with that result. 1471 */ 1472 int 1473 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node) 1474 { 1475 int b; 1476 int s; 1477 int i; 1478 int r; 1479 1480 /* 1481 * Don't bother if the node does not have very many elements 1482 */ 1483 b = 0; 1484 s = node->count; 1485 while (s - b > 4) { 1486 i = b + (s - b) / 2; 1487 ++hammer_stats_btree_elements; 1488 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base); 1489 if (r <= 1) { 1490 s = i; 1491 } else { 1492 b = i; 1493 } 1494 } 1495 return(b); 1496 } 1497 1498 1499 /************************************************************************ 1500 * SPLITTING AND MERGING * 1501 ************************************************************************ 1502 * 1503 * These routines do all the dirty work required to split and merge nodes. 1504 */ 1505 1506 /* 1507 * Split an internal node into two nodes and move the separator at the split 1508 * point to the parent. 1509 * 1510 * (cursor->node, cursor->index) indicates the element the caller intends 1511 * to push into. We will adjust node and index if that element winds 1512 * up in the split node. 1513 * 1514 * If we are at the root of the filesystem a new root must be created with 1515 * two elements, one pointing to the original root and one pointing to the 1516 * newly allocated split node. 1517 */ 1518 static 1519 int 1520 btree_split_internal(hammer_cursor_t cursor) 1521 { 1522 hammer_node_ondisk_t ondisk; 1523 hammer_node_t node; 1524 hammer_node_t parent; 1525 hammer_node_t new_node; 1526 hammer_btree_elm_t elm; 1527 hammer_btree_elm_t parent_elm; 1528 struct hammer_node_lock lockroot; 1529 hammer_mount_t hmp = cursor->trans->hmp; 1530 hammer_off_t hint; 1531 int parent_index; 1532 int made_root; 1533 int split; 1534 int error; 1535 int i; 1536 const int esize = sizeof(*elm); 1537 1538 hammer_node_lock_init(&lockroot, cursor->node); 1539 error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL); 1540 if (error) 1541 goto done; 1542 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1543 goto done; 1544 ++hammer_stats_btree_splits; 1545 1546 /* 1547 * Calculate the split point. If the insertion point is at the 1548 * end of the leaf we adjust the split point significantly to the 1549 * right to try to optimize node fill and flag it. If we hit 1550 * that same leaf again our heuristic failed and we don't try 1551 * to optimize node fill (it could lead to a degenerate case). 1552 */ 1553 node = cursor->node; 1554 ondisk = node->ondisk; 1555 KKASSERT(ondisk->count > 4); 1556 if (cursor->index == ondisk->count && 1557 (node->flags & HAMMER_NODE_NONLINEAR) == 0) { 1558 split = (ondisk->count + 1) * 3 / 4; 1559 node->flags |= HAMMER_NODE_NONLINEAR; 1560 } else { 1561 /* 1562 * We are splitting but elms[split] will be promoted to 1563 * the parent, leaving the right hand node with one less 1564 * element. If the insertion point will be on the 1565 * left-hand side adjust the split point to give the 1566 * right hand side one additional node. 1567 */ 1568 split = (ondisk->count + 1) / 2; 1569 if (cursor->index <= split) 1570 --split; 1571 } 1572 1573 /* 1574 * If we are at the root of the filesystem, create a new root node 1575 * with 1 element and split normally. Avoid making major 1576 * modifications until we know the whole operation will work. 1577 */ 1578 if (ondisk->parent == 0) { 1579 parent = hammer_alloc_btree(cursor->trans, 0, &error); 1580 if (parent == NULL) 1581 goto done; 1582 hammer_lock_ex(&parent->lock); 1583 hammer_modify_node_noundo(cursor->trans, parent); 1584 ondisk = parent->ondisk; 1585 ondisk->count = 1; 1586 ondisk->parent = 0; 1587 ondisk->mirror_tid = node->ondisk->mirror_tid; 1588 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1589 ondisk->elms[0].base = hmp->root_btree_beg; 1590 ondisk->elms[0].base.btype = node->ondisk->type; 1591 ondisk->elms[0].internal.subtree_offset = node->node_offset; 1592 ondisk->elms[1].base = hmp->root_btree_end; 1593 hammer_modify_node_done(parent); 1594 /* ondisk->elms[1].base.btype - not used */ 1595 made_root = 1; 1596 parent_index = 0; /* index of current node in parent */ 1597 } else { 1598 made_root = 0; 1599 parent = cursor->parent; 1600 parent_index = cursor->parent_index; 1601 } 1602 1603 /* 1604 * Calculate a hint for the allocation of the new B-Tree node. 1605 * The most likely expansion is coming from the insertion point 1606 * at cursor->index, so try to localize the allocation of our 1607 * new node to accomodate that sub-tree. 1608 * 1609 * Use the right-most sub-tree when expandinging on the right edge. 1610 * This is a very common case when copying a directory tree. 1611 */ 1612 if (cursor->index == ondisk->count) 1613 hint = ondisk->elms[cursor->index - 1].internal.subtree_offset; 1614 else 1615 hint = ondisk->elms[cursor->index].internal.subtree_offset; 1616 1617 /* 1618 * Split node into new_node at the split point. 1619 * 1620 * B O O O P N N B <-- P = node->elms[split] (index 4) 1621 * 0 1 2 3 4 5 6 <-- subtree indices 1622 * 1623 * x x P x x 1624 * s S S s 1625 * / \ 1626 * B O O O B B N N B <--- inner boundary points are 'P' 1627 * 0 1 2 3 4 5 6 1628 */ 1629 new_node = hammer_alloc_btree(cursor->trans, 0, &error); 1630 if (new_node == NULL) { 1631 if (made_root) { 1632 hammer_unlock(&parent->lock); 1633 hammer_delete_node(cursor->trans, parent); 1634 hammer_rel_node(parent); 1635 } 1636 goto done; 1637 } 1638 hammer_lock_ex(&new_node->lock); 1639 1640 /* 1641 * Create the new node. P becomes the left-hand boundary in the 1642 * new node. Copy the right-hand boundary as well. 1643 * 1644 * elm is the new separator. 1645 */ 1646 hammer_modify_node_noundo(cursor->trans, new_node); 1647 hammer_modify_node_all(cursor->trans, node); 1648 ondisk = node->ondisk; 1649 elm = &ondisk->elms[split]; 1650 bcopy(elm, &new_node->ondisk->elms[0], 1651 (ondisk->count - split + 1) * esize); 1652 new_node->ondisk->count = ondisk->count - split; 1653 new_node->ondisk->parent = parent->node_offset; 1654 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1655 new_node->ondisk->mirror_tid = ondisk->mirror_tid; 1656 KKASSERT(ondisk->type == new_node->ondisk->type); 1657 hammer_cursor_split_node(node, new_node, split); 1658 1659 /* 1660 * Cleanup the original node. Elm (P) becomes the new boundary, 1661 * its subtree_offset was moved to the new node. If we had created 1662 * a new root its parent pointer may have changed. 1663 */ 1664 elm->internal.subtree_offset = 0; 1665 ondisk->count = split; 1666 1667 /* 1668 * Insert the separator into the parent, fixup the parent's 1669 * reference to the original node, and reference the new node. 1670 * The separator is P. 1671 * 1672 * Remember that base.count does not include the right-hand boundary. 1673 */ 1674 hammer_modify_node_all(cursor->trans, parent); 1675 ondisk = parent->ondisk; 1676 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS); 1677 parent_elm = &ondisk->elms[parent_index+1]; 1678 bcopy(parent_elm, parent_elm + 1, 1679 (ondisk->count - parent_index) * esize); 1680 parent_elm->internal.base = elm->base; /* separator P */ 1681 parent_elm->internal.base.btype = new_node->ondisk->type; 1682 parent_elm->internal.subtree_offset = new_node->node_offset; 1683 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid; 1684 ++ondisk->count; 1685 hammer_modify_node_done(parent); 1686 hammer_cursor_inserted_element(parent, parent_index + 1); 1687 1688 /* 1689 * The children of new_node need their parent pointer set to new_node. 1690 * The children have already been locked by 1691 * hammer_btree_lock_children(). 1692 */ 1693 for (i = 0; i < new_node->ondisk->count; ++i) { 1694 elm = &new_node->ondisk->elms[i]; 1695 error = btree_set_parent(cursor->trans, new_node, elm); 1696 if (error) { 1697 panic("btree_split_internal: btree-fixup problem"); 1698 } 1699 } 1700 hammer_modify_node_done(new_node); 1701 1702 /* 1703 * The filesystem's root B-Tree pointer may have to be updated. 1704 */ 1705 if (made_root) { 1706 hammer_volume_t volume; 1707 1708 volume = hammer_get_root_volume(hmp, &error); 1709 KKASSERT(error == 0); 1710 1711 hammer_modify_volume_field(cursor->trans, volume, 1712 vol0_btree_root); 1713 volume->ondisk->vol0_btree_root = parent->node_offset; 1714 hammer_modify_volume_done(volume); 1715 node->ondisk->parent = parent->node_offset; 1716 if (cursor->parent) { 1717 hammer_unlock(&cursor->parent->lock); 1718 hammer_rel_node(cursor->parent); 1719 } 1720 cursor->parent = parent; /* lock'd and ref'd */ 1721 hammer_rel_volume(volume, 0); 1722 } 1723 hammer_modify_node_done(node); 1724 1725 /* 1726 * Ok, now adjust the cursor depending on which element the original 1727 * index was pointing at. If we are >= the split point the push node 1728 * is now in the new node. 1729 * 1730 * NOTE: If we are at the split point itself we cannot stay with the 1731 * original node because the push index will point at the right-hand 1732 * boundary, which is illegal. 1733 * 1734 * NOTE: The cursor's parent or parent_index must be adjusted for 1735 * the case where a new parent (new root) was created, and the case 1736 * where the cursor is now pointing at the split node. 1737 */ 1738 if (cursor->index >= split) { 1739 cursor->parent_index = parent_index + 1; 1740 cursor->index -= split; 1741 hammer_unlock(&cursor->node->lock); 1742 hammer_rel_node(cursor->node); 1743 cursor->node = new_node; /* locked and ref'd */ 1744 } else { 1745 cursor->parent_index = parent_index; 1746 hammer_unlock(&new_node->lock); 1747 hammer_rel_node(new_node); 1748 } 1749 1750 /* 1751 * Fixup left and right bounds 1752 */ 1753 parent_elm = &parent->ondisk->elms[cursor->parent_index]; 1754 cursor->left_bound = &parent_elm[0].internal.base; 1755 cursor->right_bound = &parent_elm[1].internal.base; 1756 KKASSERT(hammer_btree_cmp(cursor->left_bound, 1757 &cursor->node->ondisk->elms[0].internal.base) <= 0); 1758 KKASSERT(hammer_btree_cmp(cursor->right_bound, 1759 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0); 1760 1761 done: 1762 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL); 1763 hammer_cursor_downgrade(cursor); 1764 return (error); 1765 } 1766 1767 /* 1768 * Same as the above, but splits a full leaf node. 1769 * 1770 * This function 1771 */ 1772 static 1773 int 1774 btree_split_leaf(hammer_cursor_t cursor) 1775 { 1776 hammer_node_ondisk_t ondisk; 1777 hammer_node_t parent; 1778 hammer_node_t leaf; 1779 hammer_mount_t hmp; 1780 hammer_node_t new_leaf; 1781 hammer_btree_elm_t elm; 1782 hammer_btree_elm_t parent_elm; 1783 hammer_base_elm_t mid_boundary; 1784 hammer_off_t hint; 1785 int parent_index; 1786 int made_root; 1787 int split; 1788 int error; 1789 const size_t esize = sizeof(*elm); 1790 1791 if ((error = hammer_cursor_upgrade(cursor)) != 0) 1792 return(error); 1793 ++hammer_stats_btree_splits; 1794 1795 KKASSERT(hammer_btree_cmp(cursor->left_bound, 1796 &cursor->node->ondisk->elms[0].leaf.base) <= 0); 1797 KKASSERT(hammer_btree_cmp(cursor->right_bound, 1798 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0); 1799 1800 /* 1801 * Calculate the split point. If the insertion point is at the 1802 * end of the leaf we adjust the split point significantly to the 1803 * right to try to optimize node fill and flag it. If we hit 1804 * that same leaf again our heuristic failed and we don't try 1805 * to optimize node fill (it could lead to a degenerate case). 1806 * 1807 * Spikes are made up of two leaf elements which cannot be 1808 * safely split. 1809 */ 1810 leaf = cursor->node; 1811 ondisk = leaf->ondisk; 1812 KKASSERT(ondisk->count > 4); 1813 if (cursor->index == ondisk->count && 1814 (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) { 1815 split = (ondisk->count + 1) * 3 / 4; 1816 leaf->flags |= HAMMER_NODE_NONLINEAR; 1817 } else { 1818 split = (ondisk->count + 1) / 2; 1819 } 1820 1821 #if 0 1822 /* 1823 * If the insertion point is at the split point shift the 1824 * split point left so we don't have to worry about 1825 */ 1826 if (cursor->index == split) 1827 --split; 1828 #endif 1829 KKASSERT(split > 0 && split < ondisk->count); 1830 1831 error = 0; 1832 hmp = leaf->hmp; 1833 1834 elm = &ondisk->elms[split]; 1835 1836 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0); 1837 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0); 1838 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0); 1839 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0); 1840 1841 /* 1842 * If we are at the root of the tree, create a new root node with 1843 * 1 element and split normally. Avoid making major modifications 1844 * until we know the whole operation will work. 1845 */ 1846 if (ondisk->parent == 0) { 1847 parent = hammer_alloc_btree(cursor->trans, 0, &error); 1848 if (parent == NULL) 1849 goto done; 1850 hammer_lock_ex(&parent->lock); 1851 hammer_modify_node_noundo(cursor->trans, parent); 1852 ondisk = parent->ondisk; 1853 ondisk->count = 1; 1854 ondisk->parent = 0; 1855 ondisk->mirror_tid = leaf->ondisk->mirror_tid; 1856 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL; 1857 ondisk->elms[0].base = hmp->root_btree_beg; 1858 ondisk->elms[0].base.btype = leaf->ondisk->type; 1859 ondisk->elms[0].internal.subtree_offset = leaf->node_offset; 1860 ondisk->elms[1].base = hmp->root_btree_end; 1861 /* ondisk->elms[1].base.btype = not used */ 1862 hammer_modify_node_done(parent); 1863 made_root = 1; 1864 parent_index = 0; /* insertion point in parent */ 1865 } else { 1866 made_root = 0; 1867 parent = cursor->parent; 1868 parent_index = cursor->parent_index; 1869 } 1870 1871 /* 1872 * Calculate a hint for the allocation of the new B-Tree leaf node. 1873 * For now just try to localize it within the same bigblock as 1874 * the current leaf. 1875 * 1876 * If the insertion point is at the end of the leaf we recognize a 1877 * likely append sequence of some sort (data, meta-data, inodes, 1878 * whatever). Set the hint to zero to allocate out of linear space 1879 * instead of trying to completely fill previously hinted space. 1880 * 1881 * This also sets the stage for recursive splits to localize using 1882 * the new space. 1883 */ 1884 ondisk = leaf->ondisk; 1885 if (cursor->index == ondisk->count) 1886 hint = 0; 1887 else 1888 hint = leaf->node_offset; 1889 1890 /* 1891 * Split leaf into new_leaf at the split point. Select a separator 1892 * value in-between the two leafs but with a bent towards the right 1893 * leaf since comparisons use an 'elm >= separator' inequality. 1894 * 1895 * L L L L L L L L 1896 * 1897 * x x P x x 1898 * s S S s 1899 * / \ 1900 * L L L L L L L L 1901 */ 1902 new_leaf = hammer_alloc_btree(cursor->trans, 0, &error); 1903 if (new_leaf == NULL) { 1904 if (made_root) { 1905 hammer_unlock(&parent->lock); 1906 hammer_delete_node(cursor->trans, parent); 1907 hammer_rel_node(parent); 1908 } 1909 goto done; 1910 } 1911 hammer_lock_ex(&new_leaf->lock); 1912 1913 /* 1914 * Create the new node and copy the leaf elements from the split 1915 * point on to the new node. 1916 */ 1917 hammer_modify_node_all(cursor->trans, leaf); 1918 hammer_modify_node_noundo(cursor->trans, new_leaf); 1919 ondisk = leaf->ondisk; 1920 elm = &ondisk->elms[split]; 1921 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize); 1922 new_leaf->ondisk->count = ondisk->count - split; 1923 new_leaf->ondisk->parent = parent->node_offset; 1924 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF; 1925 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid; 1926 KKASSERT(ondisk->type == new_leaf->ondisk->type); 1927 hammer_modify_node_done(new_leaf); 1928 hammer_cursor_split_node(leaf, new_leaf, split); 1929 1930 /* 1931 * Cleanup the original node. Because this is a leaf node and 1932 * leaf nodes do not have a right-hand boundary, there 1933 * aren't any special edge cases to clean up. We just fixup the 1934 * count. 1935 */ 1936 ondisk->count = split; 1937 1938 /* 1939 * Insert the separator into the parent, fixup the parent's 1940 * reference to the original node, and reference the new node. 1941 * The separator is P. 1942 * 1943 * Remember that base.count does not include the right-hand boundary. 1944 * We are copying parent_index+1 to parent_index+2, not +0 to +1. 1945 */ 1946 hammer_modify_node_all(cursor->trans, parent); 1947 ondisk = parent->ondisk; 1948 KKASSERT(split != 0); 1949 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS); 1950 parent_elm = &ondisk->elms[parent_index+1]; 1951 bcopy(parent_elm, parent_elm + 1, 1952 (ondisk->count - parent_index) * esize); 1953 1954 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base); 1955 parent_elm->internal.base.btype = new_leaf->ondisk->type; 1956 parent_elm->internal.subtree_offset = new_leaf->node_offset; 1957 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid; 1958 mid_boundary = &parent_elm->base; 1959 ++ondisk->count; 1960 hammer_modify_node_done(parent); 1961 hammer_cursor_inserted_element(parent, parent_index + 1); 1962 1963 /* 1964 * The filesystem's root B-Tree pointer may have to be updated. 1965 */ 1966 if (made_root) { 1967 hammer_volume_t volume; 1968 1969 volume = hammer_get_root_volume(hmp, &error); 1970 KKASSERT(error == 0); 1971 1972 hammer_modify_volume_field(cursor->trans, volume, 1973 vol0_btree_root); 1974 volume->ondisk->vol0_btree_root = parent->node_offset; 1975 hammer_modify_volume_done(volume); 1976 leaf->ondisk->parent = parent->node_offset; 1977 if (cursor->parent) { 1978 hammer_unlock(&cursor->parent->lock); 1979 hammer_rel_node(cursor->parent); 1980 } 1981 cursor->parent = parent; /* lock'd and ref'd */ 1982 hammer_rel_volume(volume, 0); 1983 } 1984 hammer_modify_node_done(leaf); 1985 1986 /* 1987 * Ok, now adjust the cursor depending on which element the original 1988 * index was pointing at. If we are >= the split point the push node 1989 * is now in the new node. 1990 * 1991 * NOTE: If we are at the split point itself we need to select the 1992 * old or new node based on where key_beg's insertion point will be. 1993 * If we pick the wrong side the inserted element will wind up in 1994 * the wrong leaf node and outside that node's bounds. 1995 */ 1996 if (cursor->index > split || 1997 (cursor->index == split && 1998 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) { 1999 cursor->parent_index = parent_index + 1; 2000 cursor->index -= split; 2001 hammer_unlock(&cursor->node->lock); 2002 hammer_rel_node(cursor->node); 2003 cursor->node = new_leaf; 2004 } else { 2005 cursor->parent_index = parent_index; 2006 hammer_unlock(&new_leaf->lock); 2007 hammer_rel_node(new_leaf); 2008 } 2009 2010 /* 2011 * Fixup left and right bounds 2012 */ 2013 parent_elm = &parent->ondisk->elms[cursor->parent_index]; 2014 cursor->left_bound = &parent_elm[0].internal.base; 2015 cursor->right_bound = &parent_elm[1].internal.base; 2016 2017 /* 2018 * Assert that the bounds are correct. 2019 */ 2020 KKASSERT(hammer_btree_cmp(cursor->left_bound, 2021 &cursor->node->ondisk->elms[0].leaf.base) <= 0); 2022 KKASSERT(hammer_btree_cmp(cursor->right_bound, 2023 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0); 2024 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0); 2025 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0); 2026 2027 done: 2028 hammer_cursor_downgrade(cursor); 2029 return (error); 2030 } 2031 2032 #if 0 2033 2034 /* 2035 * Recursively correct the right-hand boundary's create_tid to (tid) as 2036 * long as the rest of the key matches. We have to recurse upward in 2037 * the tree as well as down the left side of each parent's right node. 2038 * 2039 * Return EDEADLK if we were only partially successful, forcing the caller 2040 * to try again. The original cursor is not modified. This routine can 2041 * also fail with EDEADLK if it is forced to throw away a portion of its 2042 * record history. 2043 * 2044 * The caller must pass a downgraded cursor to us (otherwise we can't dup it). 2045 */ 2046 struct hammer_rhb { 2047 TAILQ_ENTRY(hammer_rhb) entry; 2048 hammer_node_t node; 2049 int index; 2050 }; 2051 2052 TAILQ_HEAD(hammer_rhb_list, hammer_rhb); 2053 2054 int 2055 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid) 2056 { 2057 struct hammer_mount *hmp; 2058 struct hammer_rhb_list rhb_list; 2059 hammer_base_elm_t elm; 2060 hammer_node_t orig_node; 2061 struct hammer_rhb *rhb; 2062 int orig_index; 2063 int error; 2064 2065 TAILQ_INIT(&rhb_list); 2066 hmp = cursor->trans->hmp; 2067 2068 /* 2069 * Save our position so we can restore it on return. This also 2070 * gives us a stable 'elm'. 2071 */ 2072 orig_node = cursor->node; 2073 hammer_ref_node(orig_node); 2074 hammer_lock_sh(&orig_node->lock); 2075 orig_index = cursor->index; 2076 elm = &orig_node->ondisk->elms[orig_index].base; 2077 2078 /* 2079 * Now build a list of parents going up, allocating a rhb 2080 * structure for each one. 2081 */ 2082 while (cursor->parent) { 2083 /* 2084 * Stop if we no longer have any right-bounds to fix up 2085 */ 2086 if (elm->obj_id != cursor->right_bound->obj_id || 2087 elm->rec_type != cursor->right_bound->rec_type || 2088 elm->key != cursor->right_bound->key) { 2089 break; 2090 } 2091 2092 /* 2093 * Stop if the right-hand bound's create_tid does not 2094 * need to be corrected. 2095 */ 2096 if (cursor->right_bound->create_tid >= tid) 2097 break; 2098 2099 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO); 2100 rhb->node = cursor->parent; 2101 rhb->index = cursor->parent_index; 2102 hammer_ref_node(rhb->node); 2103 hammer_lock_sh(&rhb->node->lock); 2104 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry); 2105 2106 hammer_cursor_up(cursor); 2107 } 2108 2109 /* 2110 * now safely adjust the right hand bound for each rhb. This may 2111 * also require taking the right side of the tree and iterating down 2112 * ITS left side. 2113 */ 2114 error = 0; 2115 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2116 error = hammer_cursor_seek(cursor, rhb->node, rhb->index); 2117 if (error) 2118 break; 2119 TAILQ_REMOVE(&rhb_list, rhb, entry); 2120 hammer_unlock(&rhb->node->lock); 2121 hammer_rel_node(rhb->node); 2122 kfree(rhb, hmp->m_misc); 2123 2124 switch (cursor->node->ondisk->type) { 2125 case HAMMER_BTREE_TYPE_INTERNAL: 2126 /* 2127 * Right-boundary for parent at internal node 2128 * is one element to the right of the element whos 2129 * right boundary needs adjusting. We must then 2130 * traverse down the left side correcting any left 2131 * bounds (which may now be too far to the left). 2132 */ 2133 ++cursor->index; 2134 error = hammer_btree_correct_lhb(cursor, tid); 2135 break; 2136 default: 2137 panic("hammer_btree_correct_rhb(): Bad node type"); 2138 error = EINVAL; 2139 break; 2140 } 2141 } 2142 2143 /* 2144 * Cleanup 2145 */ 2146 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2147 TAILQ_REMOVE(&rhb_list, rhb, entry); 2148 hammer_unlock(&rhb->node->lock); 2149 hammer_rel_node(rhb->node); 2150 kfree(rhb, hmp->m_misc); 2151 } 2152 error = hammer_cursor_seek(cursor, orig_node, orig_index); 2153 hammer_unlock(&orig_node->lock); 2154 hammer_rel_node(orig_node); 2155 return (error); 2156 } 2157 2158 /* 2159 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand 2160 * bound going downward starting at the current cursor position. 2161 * 2162 * This function does not restore the cursor after use. 2163 */ 2164 int 2165 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid) 2166 { 2167 struct hammer_rhb_list rhb_list; 2168 hammer_base_elm_t elm; 2169 hammer_base_elm_t cmp; 2170 struct hammer_rhb *rhb; 2171 struct hammer_mount *hmp; 2172 int error; 2173 2174 TAILQ_INIT(&rhb_list); 2175 hmp = cursor->trans->hmp; 2176 2177 cmp = &cursor->node->ondisk->elms[cursor->index].base; 2178 2179 /* 2180 * Record the node and traverse down the left-hand side for all 2181 * matching records needing a boundary correction. 2182 */ 2183 error = 0; 2184 for (;;) { 2185 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO); 2186 rhb->node = cursor->node; 2187 rhb->index = cursor->index; 2188 hammer_ref_node(rhb->node); 2189 hammer_lock_sh(&rhb->node->lock); 2190 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry); 2191 2192 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 2193 /* 2194 * Nothing to traverse down if we are at the right 2195 * boundary of an internal node. 2196 */ 2197 if (cursor->index == cursor->node->ondisk->count) 2198 break; 2199 } else { 2200 elm = &cursor->node->ondisk->elms[cursor->index].base; 2201 if (elm->btype == HAMMER_BTREE_TYPE_RECORD) 2202 break; 2203 panic("Illegal leaf record type %02x", elm->btype); 2204 } 2205 error = hammer_cursor_down(cursor); 2206 if (error) 2207 break; 2208 2209 elm = &cursor->node->ondisk->elms[cursor->index].base; 2210 if (elm->obj_id != cmp->obj_id || 2211 elm->rec_type != cmp->rec_type || 2212 elm->key != cmp->key) { 2213 break; 2214 } 2215 if (elm->create_tid >= tid) 2216 break; 2217 2218 } 2219 2220 /* 2221 * Now we can safely adjust the left-hand boundary from the bottom-up. 2222 * The last element we remove from the list is the caller's right hand 2223 * boundary, which must also be adjusted. 2224 */ 2225 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2226 error = hammer_cursor_seek(cursor, rhb->node, rhb->index); 2227 if (error) 2228 break; 2229 TAILQ_REMOVE(&rhb_list, rhb, entry); 2230 hammer_unlock(&rhb->node->lock); 2231 hammer_rel_node(rhb->node); 2232 kfree(rhb, hmp->m_misc); 2233 2234 elm = &cursor->node->ondisk->elms[cursor->index].base; 2235 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 2236 hammer_modify_node(cursor->trans, cursor->node, 2237 &elm->create_tid, 2238 sizeof(elm->create_tid)); 2239 elm->create_tid = tid; 2240 hammer_modify_node_done(cursor->node); 2241 } else { 2242 panic("hammer_btree_correct_lhb(): Bad element type"); 2243 } 2244 } 2245 2246 /* 2247 * Cleanup 2248 */ 2249 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) { 2250 TAILQ_REMOVE(&rhb_list, rhb, entry); 2251 hammer_unlock(&rhb->node->lock); 2252 hammer_rel_node(rhb->node); 2253 kfree(rhb, hmp->m_misc); 2254 } 2255 return (error); 2256 } 2257 2258 #endif 2259 2260 /* 2261 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at 2262 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete 2263 * the operation due to a deadlock, or some other error. 2264 * 2265 * This routine is initially called with an empty leaf and may be 2266 * recursively called with single-element internal nodes. 2267 * 2268 * It should also be noted that when removing empty leaves we must be sure 2269 * to test and update mirror_tid because another thread may have deadlocked 2270 * against us (or someone) trying to propagate it up and cannot retry once 2271 * the node has been deleted. 2272 * 2273 * On return the cursor may end up pointing to an internal node, suitable 2274 * for further iteration but not for an immediate insertion or deletion. 2275 */ 2276 static int 2277 btree_remove(hammer_cursor_t cursor) 2278 { 2279 hammer_node_ondisk_t ondisk; 2280 hammer_btree_elm_t elm; 2281 hammer_node_t node; 2282 hammer_node_t parent; 2283 const int esize = sizeof(*elm); 2284 int error; 2285 2286 node = cursor->node; 2287 2288 /* 2289 * When deleting the root of the filesystem convert it to 2290 * an empty leaf node. Internal nodes cannot be empty. 2291 */ 2292 ondisk = node->ondisk; 2293 if (ondisk->parent == 0) { 2294 KKASSERT(cursor->parent == NULL); 2295 hammer_modify_node_all(cursor->trans, node); 2296 KKASSERT(ondisk == node->ondisk); 2297 ondisk->type = HAMMER_BTREE_TYPE_LEAF; 2298 ondisk->count = 0; 2299 hammer_modify_node_done(node); 2300 cursor->index = 0; 2301 return(0); 2302 } 2303 2304 parent = cursor->parent; 2305 2306 /* 2307 * Attempt to remove the parent's reference to the child. If the 2308 * parent would become empty we have to recurse. If we fail we 2309 * leave the parent pointing to an empty leaf node. 2310 * 2311 * We have to recurse successfully before we can delete the internal 2312 * node as it is illegal to have empty internal nodes. Even though 2313 * the operation may be aborted we must still fixup any unlocked 2314 * cursors as if we had deleted the element prior to recursing 2315 * (by calling hammer_cursor_deleted_element()) so those cursors 2316 * are properly forced up the chain by the recursion. 2317 */ 2318 if (parent->ondisk->count == 1) { 2319 /* 2320 * This special cursor_up_locked() call leaves the original 2321 * node exclusively locked and referenced, leaves the 2322 * original parent locked (as the new node), and locks the 2323 * new parent. It can return EDEADLK. 2324 * 2325 * We cannot call hammer_cursor_removed_node() until we are 2326 * actually able to remove the node. If we did then tracked 2327 * cursors in the middle of iterations could be repointed 2328 * to a parent node. If this occurs they could end up 2329 * scanning newly inserted records into the node (that could 2330 * not be deleted) when they push down again. 2331 * 2332 * Due to the way the recursion works the final parent is left 2333 * in cursor->parent after the recursion returns. Each 2334 * layer on the way back up is thus able to call 2335 * hammer_cursor_removed_node() and 'jump' the node up to 2336 * the (same) final parent. 2337 * 2338 * NOTE! The local variable 'parent' is invalid after we 2339 * call hammer_cursor_up_locked(). 2340 */ 2341 error = hammer_cursor_up_locked(cursor); 2342 parent = NULL; 2343 2344 if (error == 0) { 2345 hammer_cursor_deleted_element(cursor->node, 0); 2346 error = btree_remove(cursor); 2347 if (error == 0) { 2348 KKASSERT(node != cursor->node); 2349 hammer_cursor_removed_node( 2350 node, cursor->node, 2351 cursor->index); 2352 hammer_modify_node_all(cursor->trans, node); 2353 ondisk = node->ondisk; 2354 ondisk->type = HAMMER_BTREE_TYPE_DELETED; 2355 ondisk->count = 0; 2356 hammer_modify_node_done(node); 2357 hammer_flush_node(node, 0); 2358 hammer_delete_node(cursor->trans, node); 2359 } else { 2360 /* 2361 * Defer parent removal because we could not 2362 * get the lock, just let the leaf remain 2363 * empty. 2364 */ 2365 /**/ 2366 } 2367 hammer_unlock(&node->lock); 2368 hammer_rel_node(node); 2369 } else { 2370 /* 2371 * Defer parent removal because we could not 2372 * get the lock, just let the leaf remain 2373 * empty. 2374 */ 2375 /**/ 2376 } 2377 } else { 2378 KKASSERT(parent->ondisk->count > 1); 2379 2380 hammer_modify_node_all(cursor->trans, parent); 2381 ondisk = parent->ondisk; 2382 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL); 2383 2384 elm = &ondisk->elms[cursor->parent_index]; 2385 KKASSERT(elm->internal.subtree_offset == node->node_offset); 2386 KKASSERT(ondisk->count > 0); 2387 2388 /* 2389 * We must retain the highest mirror_tid. The deleted 2390 * range is now encompassed by the element to the left. 2391 * If we are already at the left edge the new left edge 2392 * inherits mirror_tid. 2393 * 2394 * Note that bounds of the parent to our parent may create 2395 * a gap to the left of our left-most node or to the right 2396 * of our right-most node. The gap is silently included 2397 * in the mirror_tid's area of effect from the point of view 2398 * of the scan. 2399 */ 2400 if (cursor->parent_index) { 2401 if (elm[-1].internal.mirror_tid < 2402 elm[0].internal.mirror_tid) { 2403 elm[-1].internal.mirror_tid = 2404 elm[0].internal.mirror_tid; 2405 } 2406 } else { 2407 if (elm[1].internal.mirror_tid < 2408 elm[0].internal.mirror_tid) { 2409 elm[1].internal.mirror_tid = 2410 elm[0].internal.mirror_tid; 2411 } 2412 } 2413 2414 /* 2415 * Delete the subtree reference in the parent. Include 2416 * boundary element at end. 2417 */ 2418 bcopy(&elm[1], &elm[0], 2419 (ondisk->count - cursor->parent_index) * esize); 2420 --ondisk->count; 2421 hammer_modify_node_done(parent); 2422 hammer_cursor_removed_node(node, parent, cursor->parent_index); 2423 hammer_cursor_deleted_element(parent, cursor->parent_index); 2424 hammer_flush_node(node, 0); 2425 hammer_delete_node(cursor->trans, node); 2426 2427 /* 2428 * cursor->node is invalid, cursor up to make the cursor 2429 * valid again. We have to flag the condition in case 2430 * another thread wiggles an insertion in during an 2431 * iteration. 2432 */ 2433 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK; 2434 error = hammer_cursor_up(cursor); 2435 } 2436 return (error); 2437 } 2438 2439 /* 2440 * Propagate cursor->trans->tid up the B-Tree starting at the current 2441 * cursor position using pseudofs info gleaned from the passed inode. 2442 * 2443 * The passed inode has no relationship to the cursor position other 2444 * then being in the same pseudofs as the insertion or deletion we 2445 * are propagating the mirror_tid for. 2446 * 2447 * WARNING! Because we push and pop the passed cursor, it may be 2448 * modified by other B-Tree operations while it is unlocked 2449 * and things like the node & leaf pointers, and indexes might 2450 * change. 2451 */ 2452 void 2453 hammer_btree_do_propagation(hammer_cursor_t cursor, 2454 hammer_pseudofs_inmem_t pfsm, 2455 hammer_btree_leaf_elm_t leaf) 2456 { 2457 hammer_cursor_t ncursor; 2458 hammer_tid_t mirror_tid; 2459 int error; 2460 2461 /* 2462 * We do not propagate a mirror_tid if the filesystem was mounted 2463 * in no-mirror mode. 2464 */ 2465 if (cursor->trans->hmp->master_id < 0) 2466 return; 2467 2468 /* 2469 * This is a bit of a hack because we cannot deadlock or return 2470 * EDEADLK here. The related operation has already completed and 2471 * we must propagate the mirror_tid now regardless. 2472 * 2473 * Generate a new cursor which inherits the original's locks and 2474 * unlock the original. Use the new cursor to propagate the 2475 * mirror_tid. Then clean up the new cursor and reacquire locks 2476 * on the original. 2477 * 2478 * hammer_dup_cursor() cannot dup locks. The dup inherits the 2479 * original's locks and the original is tracked and must be 2480 * re-locked. 2481 */ 2482 mirror_tid = cursor->node->ondisk->mirror_tid; 2483 KKASSERT(mirror_tid != 0); 2484 ncursor = hammer_push_cursor(cursor); 2485 error = hammer_btree_mirror_propagate(ncursor, mirror_tid); 2486 KKASSERT(error == 0); 2487 hammer_pop_cursor(cursor, ncursor); 2488 /* WARNING: cursor's leaf pointer may change after pop */ 2489 } 2490 2491 2492 /* 2493 * Propagate a mirror TID update upwards through the B-Tree to the root. 2494 * 2495 * A locked internal node must be passed in. The node will remain locked 2496 * on return. 2497 * 2498 * This function syncs mirror_tid at the specified internal node's element, 2499 * adjusts the node's aggregation mirror_tid, and then recurses upwards. 2500 */ 2501 static int 2502 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid) 2503 { 2504 hammer_btree_internal_elm_t elm; 2505 hammer_node_t node; 2506 int error; 2507 2508 for (;;) { 2509 error = hammer_cursor_up(cursor); 2510 if (error == 0) 2511 error = hammer_cursor_upgrade(cursor); 2512 2513 /* 2514 * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the 2515 * cursor will still be properly positioned for 2516 * mirror propagation, just not for iterations. 2517 */ 2518 while (error == EDEADLK) { 2519 hammer_recover_cursor(cursor); 2520 error = hammer_cursor_upgrade(cursor); 2521 } 2522 if (error) 2523 break; 2524 2525 /* 2526 * If the cursor deadlocked it could end up at a leaf 2527 * after we lost the lock. 2528 */ 2529 node = cursor->node; 2530 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL) 2531 continue; 2532 2533 /* 2534 * Adjust the node's element 2535 */ 2536 elm = &node->ondisk->elms[cursor->index].internal; 2537 if (elm->mirror_tid >= mirror_tid) 2538 break; 2539 hammer_modify_node(cursor->trans, node, &elm->mirror_tid, 2540 sizeof(elm->mirror_tid)); 2541 elm->mirror_tid = mirror_tid; 2542 hammer_modify_node_done(node); 2543 if (hammer_debug_general & 0x0002) { 2544 kprintf("mirror_propagate: propagate " 2545 "%016llx @%016llx:%d\n", 2546 (long long)mirror_tid, 2547 (long long)node->node_offset, 2548 cursor->index); 2549 } 2550 2551 2552 /* 2553 * Adjust the node's mirror_tid aggregator 2554 */ 2555 if (node->ondisk->mirror_tid >= mirror_tid) 2556 return(0); 2557 hammer_modify_node_field(cursor->trans, node, mirror_tid); 2558 node->ondisk->mirror_tid = mirror_tid; 2559 hammer_modify_node_done(node); 2560 if (hammer_debug_general & 0x0002) { 2561 kprintf("mirror_propagate: propagate " 2562 "%016llx @%016llx\n", 2563 (long long)mirror_tid, 2564 (long long)node->node_offset); 2565 } 2566 } 2567 if (error == ENOENT) 2568 error = 0; 2569 return(error); 2570 } 2571 2572 hammer_node_t 2573 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node, 2574 int *parent_indexp, int *errorp, int try_exclusive) 2575 { 2576 hammer_node_t parent; 2577 hammer_btree_elm_t elm; 2578 int i; 2579 2580 /* 2581 * Get the node 2582 */ 2583 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp); 2584 if (*errorp) { 2585 KKASSERT(parent == NULL); 2586 return(NULL); 2587 } 2588 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0); 2589 2590 /* 2591 * Lock the node 2592 */ 2593 if (try_exclusive) { 2594 if (hammer_lock_ex_try(&parent->lock)) { 2595 hammer_rel_node(parent); 2596 *errorp = EDEADLK; 2597 return(NULL); 2598 } 2599 } else { 2600 hammer_lock_sh(&parent->lock); 2601 } 2602 2603 /* 2604 * Figure out which element in the parent is pointing to the 2605 * child. 2606 */ 2607 if (node->ondisk->count) { 2608 i = hammer_btree_search_node(&node->ondisk->elms[0].base, 2609 parent->ondisk); 2610 } else { 2611 i = 0; 2612 } 2613 while (i < parent->ondisk->count) { 2614 elm = &parent->ondisk->elms[i]; 2615 if (elm->internal.subtree_offset == node->node_offset) 2616 break; 2617 ++i; 2618 } 2619 if (i == parent->ondisk->count) { 2620 hammer_unlock(&parent->lock); 2621 panic("Bad B-Tree link: parent %p node %p\n", parent, node); 2622 } 2623 *parent_indexp = i; 2624 KKASSERT(*errorp == 0); 2625 return(parent); 2626 } 2627 2628 /* 2629 * The element (elm) has been moved to a new internal node (node). 2630 * 2631 * If the element represents a pointer to an internal node that node's 2632 * parent must be adjusted to the element's new location. 2633 * 2634 * XXX deadlock potential here with our exclusive locks 2635 */ 2636 int 2637 btree_set_parent(hammer_transaction_t trans, hammer_node_t node, 2638 hammer_btree_elm_t elm) 2639 { 2640 hammer_node_t child; 2641 int error; 2642 2643 error = 0; 2644 2645 switch(elm->base.btype) { 2646 case HAMMER_BTREE_TYPE_INTERNAL: 2647 case HAMMER_BTREE_TYPE_LEAF: 2648 child = hammer_get_node(trans, elm->internal.subtree_offset, 2649 0, &error); 2650 if (error == 0) { 2651 hammer_modify_node_field(trans, child, parent); 2652 child->ondisk->parent = node->node_offset; 2653 hammer_modify_node_done(child); 2654 hammer_rel_node(child); 2655 } 2656 break; 2657 default: 2658 break; 2659 } 2660 return(error); 2661 } 2662 2663 /* 2664 * Initialize the root of a recursive B-Tree node lock list structure. 2665 */ 2666 void 2667 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node) 2668 { 2669 TAILQ_INIT(&parent->list); 2670 parent->parent = NULL; 2671 parent->node = node; 2672 parent->index = -1; 2673 parent->count = node->ondisk->count; 2674 parent->copy = NULL; 2675 parent->flags = 0; 2676 } 2677 2678 /* 2679 * Initialize a cache of hammer_node_lock's including space allocated 2680 * for node copies. 2681 * 2682 * This is used by the rebalancing code to preallocate the copy space 2683 * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER 2684 * locks, otherwise we can blow out the pageout daemon's emergency 2685 * reserve and deadlock it. 2686 * 2687 * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache. 2688 * The flag is set when the item is pulled off the cache for use. 2689 */ 2690 void 2691 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache, 2692 int depth) 2693 { 2694 hammer_node_lock_t item; 2695 int count; 2696 2697 for (count = 1; depth; --depth) 2698 count *= HAMMER_BTREE_LEAF_ELMS; 2699 bzero(lcache, sizeof(*lcache)); 2700 TAILQ_INIT(&lcache->list); 2701 while (count) { 2702 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO); 2703 item->copy = kmalloc(sizeof(*item->copy), 2704 hmp->m_misc, M_WAITOK); 2705 TAILQ_INIT(&item->list); 2706 TAILQ_INSERT_TAIL(&lcache->list, item, entry); 2707 --count; 2708 } 2709 } 2710 2711 void 2712 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache) 2713 { 2714 hammer_node_lock_t item; 2715 2716 while ((item = TAILQ_FIRST(&lcache->list)) != NULL) { 2717 TAILQ_REMOVE(&lcache->list, item, entry); 2718 KKASSERT(item->copy); 2719 KKASSERT(TAILQ_EMPTY(&item->list)); 2720 kfree(item->copy, hmp->m_misc); 2721 kfree(item, hmp->m_misc); 2722 } 2723 KKASSERT(lcache->copy == NULL); 2724 } 2725 2726 /* 2727 * Exclusively lock all the children of node. This is used by the split 2728 * code to prevent anyone from accessing the children of a cursor node 2729 * while we fix-up its parent offset. 2730 * 2731 * If we don't lock the children we can really mess up cursors which block 2732 * trying to cursor-up into our node. 2733 * 2734 * On failure EDEADLK (or some other error) is returned. If a deadlock 2735 * error is returned the cursor is adjusted to block on termination. 2736 * 2737 * The caller is responsible for managing parent->node, the root's node 2738 * is usually aliased from a cursor. 2739 */ 2740 int 2741 hammer_btree_lock_children(hammer_cursor_t cursor, int depth, 2742 hammer_node_lock_t parent, 2743 hammer_node_lock_t lcache) 2744 { 2745 hammer_node_t node; 2746 hammer_node_lock_t item; 2747 hammer_node_ondisk_t ondisk; 2748 hammer_btree_elm_t elm; 2749 hammer_node_t child; 2750 struct hammer_mount *hmp; 2751 int error; 2752 int i; 2753 2754 node = parent->node; 2755 ondisk = node->ondisk; 2756 error = 0; 2757 hmp = cursor->trans->hmp; 2758 2759 /* 2760 * We really do not want to block on I/O with exclusive locks held, 2761 * pre-get the children before trying to lock the mess. This is 2762 * only done one-level deep for now. 2763 */ 2764 for (i = 0; i < ondisk->count; ++i) { 2765 ++hammer_stats_btree_elements; 2766 elm = &ondisk->elms[i]; 2767 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF && 2768 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) { 2769 continue; 2770 } 2771 child = hammer_get_node(cursor->trans, 2772 elm->internal.subtree_offset, 2773 0, &error); 2774 if (child) 2775 hammer_rel_node(child); 2776 } 2777 2778 /* 2779 * Do it for real 2780 */ 2781 for (i = 0; error == 0 && i < ondisk->count; ++i) { 2782 ++hammer_stats_btree_elements; 2783 elm = &ondisk->elms[i]; 2784 2785 switch(elm->base.btype) { 2786 case HAMMER_BTREE_TYPE_INTERNAL: 2787 case HAMMER_BTREE_TYPE_LEAF: 2788 KKASSERT(elm->internal.subtree_offset != 0); 2789 child = hammer_get_node(cursor->trans, 2790 elm->internal.subtree_offset, 2791 0, &error); 2792 break; 2793 default: 2794 child = NULL; 2795 break; 2796 } 2797 if (child) { 2798 if (hammer_lock_ex_try(&child->lock) != 0) { 2799 if (cursor->deadlk_node == NULL) { 2800 cursor->deadlk_node = child; 2801 hammer_ref_node(cursor->deadlk_node); 2802 } 2803 error = EDEADLK; 2804 hammer_rel_node(child); 2805 } else { 2806 if (lcache) { 2807 item = TAILQ_FIRST(&lcache->list); 2808 KKASSERT(item != NULL); 2809 item->flags |= HAMMER_NODE_LOCK_LCACHE; 2810 TAILQ_REMOVE(&lcache->list, 2811 item, entry); 2812 } else { 2813 item = kmalloc(sizeof(*item), 2814 hmp->m_misc, 2815 M_WAITOK|M_ZERO); 2816 TAILQ_INIT(&item->list); 2817 } 2818 2819 TAILQ_INSERT_TAIL(&parent->list, item, entry); 2820 item->parent = parent; 2821 item->node = child; 2822 item->index = i; 2823 item->count = child->ondisk->count; 2824 2825 /* 2826 * Recurse (used by the rebalancing code) 2827 */ 2828 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) { 2829 error = hammer_btree_lock_children( 2830 cursor, 2831 depth - 1, 2832 item, 2833 lcache); 2834 } 2835 } 2836 } 2837 } 2838 if (error) 2839 hammer_btree_unlock_children(hmp, parent, lcache); 2840 return(error); 2841 } 2842 2843 /* 2844 * Create an in-memory copy of all B-Tree nodes listed, recursively, 2845 * including the parent. 2846 */ 2847 void 2848 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent) 2849 { 2850 hammer_mount_t hmp = cursor->trans->hmp; 2851 hammer_node_lock_t item; 2852 2853 if (parent->copy == NULL) { 2854 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0); 2855 parent->copy = kmalloc(sizeof(*parent->copy), 2856 hmp->m_misc, M_WAITOK); 2857 } 2858 KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0); 2859 *parent->copy = *parent->node->ondisk; 2860 TAILQ_FOREACH(item, &parent->list, entry) { 2861 hammer_btree_lock_copy(cursor, item); 2862 } 2863 } 2864 2865 /* 2866 * Recursively sync modified copies to the media. 2867 */ 2868 int 2869 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent) 2870 { 2871 hammer_node_lock_t item; 2872 int count = 0; 2873 2874 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) { 2875 ++count; 2876 hammer_modify_node_all(cursor->trans, parent->node); 2877 *parent->node->ondisk = *parent->copy; 2878 hammer_modify_node_done(parent->node); 2879 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) { 2880 hammer_flush_node(parent->node, 0); 2881 hammer_delete_node(cursor->trans, parent->node); 2882 } 2883 } 2884 TAILQ_FOREACH(item, &parent->list, entry) { 2885 count += hammer_btree_sync_copy(cursor, item); 2886 } 2887 return(count); 2888 } 2889 2890 /* 2891 * Release previously obtained node locks. The caller is responsible for 2892 * cleaning up parent->node itself (its usually just aliased from a cursor), 2893 * but this function will take care of the copies. 2894 * 2895 * NOTE: The root node is not placed in the lcache and node->copy is not 2896 * deallocated when lcache != NULL. 2897 */ 2898 void 2899 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent, 2900 hammer_node_lock_t lcache) 2901 { 2902 hammer_node_lock_t item; 2903 hammer_node_ondisk_t copy; 2904 2905 while ((item = TAILQ_FIRST(&parent->list)) != NULL) { 2906 TAILQ_REMOVE(&parent->list, item, entry); 2907 hammer_btree_unlock_children(hmp, item, lcache); 2908 hammer_unlock(&item->node->lock); 2909 hammer_rel_node(item->node); 2910 if (lcache) { 2911 /* 2912 * NOTE: When placing the item back in the lcache 2913 * the flag is cleared by the bzero(). 2914 * Remaining fields are cleared as a safety 2915 * measure. 2916 */ 2917 KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE); 2918 KKASSERT(TAILQ_EMPTY(&item->list)); 2919 copy = item->copy; 2920 bzero(item, sizeof(*item)); 2921 TAILQ_INIT(&item->list); 2922 item->copy = copy; 2923 if (copy) 2924 bzero(copy, sizeof(*copy)); 2925 TAILQ_INSERT_TAIL(&lcache->list, item, entry); 2926 } else { 2927 kfree(item, hmp->m_misc); 2928 } 2929 } 2930 if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) { 2931 kfree(parent->copy, hmp->m_misc); 2932 parent->copy = NULL; /* safety */ 2933 } 2934 } 2935 2936 /************************************************************************ 2937 * MISCELLANIOUS SUPPORT * 2938 ************************************************************************/ 2939 2940 /* 2941 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp). 2942 * 2943 * Note that for this particular function a return value of -1, 0, or +1 2944 * can denote a match if create_tid is otherwise discounted. A create_tid 2945 * of zero is considered to be 'infinity' in comparisons. 2946 * 2947 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c. 2948 */ 2949 int 2950 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2) 2951 { 2952 if (key1->localization < key2->localization) 2953 return(-5); 2954 if (key1->localization > key2->localization) 2955 return(5); 2956 2957 if (key1->obj_id < key2->obj_id) 2958 return(-4); 2959 if (key1->obj_id > key2->obj_id) 2960 return(4); 2961 2962 if (key1->rec_type < key2->rec_type) 2963 return(-3); 2964 if (key1->rec_type > key2->rec_type) 2965 return(3); 2966 2967 if (key1->key < key2->key) 2968 return(-2); 2969 if (key1->key > key2->key) 2970 return(2); 2971 2972 /* 2973 * A create_tid of zero indicates a record which is undeletable 2974 * and must be considered to have a value of positive infinity. 2975 */ 2976 if (key1->create_tid == 0) { 2977 if (key2->create_tid == 0) 2978 return(0); 2979 return(1); 2980 } 2981 if (key2->create_tid == 0) 2982 return(-1); 2983 if (key1->create_tid < key2->create_tid) 2984 return(-1); 2985 if (key1->create_tid > key2->create_tid) 2986 return(1); 2987 return(0); 2988 } 2989 2990 /* 2991 * Test a timestamp against an element to determine whether the 2992 * element is visible. A timestamp of 0 means 'infinity'. 2993 */ 2994 int 2995 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base) 2996 { 2997 if (asof == 0) { 2998 if (base->delete_tid) 2999 return(1); 3000 return(0); 3001 } 3002 if (asof < base->create_tid) 3003 return(-1); 3004 if (base->delete_tid && asof >= base->delete_tid) 3005 return(1); 3006 return(0); 3007 } 3008 3009 /* 3010 * Create a separator half way inbetween key1 and key2. For fields just 3011 * one unit apart, the separator will match key2. key1 is on the left-hand 3012 * side and key2 is on the right-hand side. 3013 * 3014 * key2 must be >= the separator. It is ok for the separator to match key2. 3015 * 3016 * NOTE: Even if key1 does not match key2, the separator may wind up matching 3017 * key2. 3018 * 3019 * NOTE: It might be beneficial to just scrap this whole mess and just 3020 * set the separator to key2. 3021 */ 3022 #define MAKE_SEPARATOR(key1, key2, dest, field) \ 3023 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1); 3024 3025 static void 3026 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2, 3027 hammer_base_elm_t dest) 3028 { 3029 bzero(dest, sizeof(*dest)); 3030 3031 dest->rec_type = key2->rec_type; 3032 dest->key = key2->key; 3033 dest->obj_id = key2->obj_id; 3034 dest->create_tid = key2->create_tid; 3035 3036 MAKE_SEPARATOR(key1, key2, dest, localization); 3037 if (key1->localization == key2->localization) { 3038 MAKE_SEPARATOR(key1, key2, dest, obj_id); 3039 if (key1->obj_id == key2->obj_id) { 3040 MAKE_SEPARATOR(key1, key2, dest, rec_type); 3041 if (key1->rec_type == key2->rec_type) { 3042 MAKE_SEPARATOR(key1, key2, dest, key); 3043 /* 3044 * Don't bother creating a separator for 3045 * create_tid, which also conveniently avoids 3046 * having to handle the create_tid == 0 3047 * (infinity) case. Just leave create_tid 3048 * set to key2. 3049 * 3050 * Worst case, dest matches key2 exactly, 3051 * which is acceptable. 3052 */ 3053 } 3054 } 3055 } 3056 } 3057 3058 #undef MAKE_SEPARATOR 3059 3060 /* 3061 * Return whether a generic internal or leaf node is full 3062 */ 3063 static int 3064 btree_node_is_full(hammer_node_ondisk_t node) 3065 { 3066 switch(node->type) { 3067 case HAMMER_BTREE_TYPE_INTERNAL: 3068 if (node->count == HAMMER_BTREE_INT_ELMS) 3069 return(1); 3070 break; 3071 case HAMMER_BTREE_TYPE_LEAF: 3072 if (node->count == HAMMER_BTREE_LEAF_ELMS) 3073 return(1); 3074 break; 3075 default: 3076 panic("illegal btree subtype"); 3077 } 3078 return(0); 3079 } 3080 3081 #if 0 3082 static int 3083 btree_max_elements(u_int8_t type) 3084 { 3085 if (type == HAMMER_BTREE_TYPE_LEAF) 3086 return(HAMMER_BTREE_LEAF_ELMS); 3087 if (type == HAMMER_BTREE_TYPE_INTERNAL) 3088 return(HAMMER_BTREE_INT_ELMS); 3089 panic("btree_max_elements: bad type %d\n", type); 3090 } 3091 #endif 3092 3093 void 3094 hammer_print_btree_node(hammer_node_ondisk_t ondisk) 3095 { 3096 hammer_btree_elm_t elm; 3097 int i; 3098 3099 kprintf("node %p count=%d parent=%016llx type=%c\n", 3100 ondisk, ondisk->count, 3101 (long long)ondisk->parent, ondisk->type); 3102 3103 /* 3104 * Dump both boundary elements if an internal node 3105 */ 3106 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) { 3107 for (i = 0; i <= ondisk->count; ++i) { 3108 elm = &ondisk->elms[i]; 3109 hammer_print_btree_elm(elm, ondisk->type, i); 3110 } 3111 } else { 3112 for (i = 0; i < ondisk->count; ++i) { 3113 elm = &ondisk->elms[i]; 3114 hammer_print_btree_elm(elm, ondisk->type, i); 3115 } 3116 } 3117 } 3118 3119 void 3120 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i) 3121 { 3122 kprintf(" %2d", i); 3123 kprintf("\tobj_id = %016llx\n", (long long)elm->base.obj_id); 3124 kprintf("\tkey = %016llx\n", (long long)elm->base.key); 3125 kprintf("\tcreate_tid = %016llx\n", (long long)elm->base.create_tid); 3126 kprintf("\tdelete_tid = %016llx\n", (long long)elm->base.delete_tid); 3127 kprintf("\trec_type = %04x\n", elm->base.rec_type); 3128 kprintf("\tobj_type = %02x\n", elm->base.obj_type); 3129 kprintf("\tbtype = %02x (%c)\n", 3130 elm->base.btype, 3131 (elm->base.btype ? elm->base.btype : '?')); 3132 kprintf("\tlocalization = %02x\n", elm->base.localization); 3133 3134 switch(type) { 3135 case HAMMER_BTREE_TYPE_INTERNAL: 3136 kprintf("\tsubtree_off = %016llx\n", 3137 (long long)elm->internal.subtree_offset); 3138 break; 3139 case HAMMER_BTREE_TYPE_RECORD: 3140 kprintf("\tdata_offset = %016llx\n", 3141 (long long)elm->leaf.data_offset); 3142 kprintf("\tdata_len = %08x\n", elm->leaf.data_len); 3143 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc); 3144 break; 3145 } 3146 } 3147