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_cursor.c,v 1.42 2008/08/06 15:38:58 dillon Exp $ 35 */ 36 37 /* 38 * HAMMER B-Tree index - cursor support routines 39 */ 40 #include "hammer.h" 41 42 static int hammer_load_cursor_parent(hammer_cursor_t cursor, int try_exclusive); 43 44 /* 45 * Initialize a fresh cursor using the B-Tree node cache. If the cache 46 * is not available initialize a fresh cursor at the root of the filesystem. 47 */ 48 int 49 hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor, 50 hammer_node_cache_t cache, hammer_inode_t ip) 51 { 52 hammer_volume_t volume; 53 hammer_node_t node; 54 int error; 55 56 bzero(cursor, sizeof(*cursor)); 57 58 cursor->trans = trans; 59 60 /* 61 * If the cursor operation is on behalf of an inode, lock 62 * the inode. 63 */ 64 if ((cursor->ip = ip) != NULL) { 65 ++ip->cursor_ip_refs; 66 if (trans->type == HAMMER_TRANS_FLS) 67 hammer_lock_ex(&ip->lock); 68 else 69 hammer_lock_sh(&ip->lock); 70 } 71 72 /* 73 * Step 1 - acquire a locked node from the cache if possible 74 */ 75 if (cache && cache->node) { 76 node = hammer_ref_node_safe(trans, cache, &error); 77 if (error == 0) { 78 hammer_lock_sh(&node->lock); 79 if (node->flags & HAMMER_NODE_DELETED) { 80 hammer_unlock(&node->lock); 81 hammer_rel_node(node); 82 node = NULL; 83 } 84 } 85 if (node == NULL) 86 ++hammer_stats_btree_root_iterations; 87 } else { 88 node = NULL; 89 ++hammer_stats_btree_root_iterations; 90 } 91 92 /* 93 * Step 2 - If we couldn't get a node from the cache, get 94 * the one from the root of the filesystem. 95 */ 96 while (node == NULL) { 97 volume = hammer_get_root_volume(trans->hmp, &error); 98 if (error) 99 break; 100 node = hammer_get_node(trans, volume->ondisk->vol0_btree_root, 101 0, &error); 102 hammer_rel_volume(volume, 0); 103 if (error) 104 break; 105 hammer_lock_sh(&node->lock); 106 107 /* 108 * If someone got in before we could lock the node, retry. 109 */ 110 if (node->flags & HAMMER_NODE_DELETED) { 111 hammer_unlock(&node->lock); 112 hammer_rel_node(node); 113 node = NULL; 114 continue; 115 } 116 if (volume->ondisk->vol0_btree_root != node->node_offset) { 117 hammer_unlock(&node->lock); 118 hammer_rel_node(node); 119 node = NULL; 120 continue; 121 } 122 } 123 124 /* 125 * Step 3 - finish initializing the cursor by acquiring the parent 126 */ 127 cursor->node = node; 128 if (error == 0) 129 error = hammer_load_cursor_parent(cursor, 0); 130 KKASSERT(error == 0); 131 /* if (error) hammer_done_cursor(cursor); */ 132 return(error); 133 } 134 135 /* 136 * Normalize a cursor. Sometimes cursors can be left in a state 137 * where node is NULL. If the cursor is in this state, cursor up. 138 */ 139 void 140 hammer_normalize_cursor(hammer_cursor_t cursor) 141 { 142 if (cursor->node == NULL) { 143 KKASSERT(cursor->parent != NULL); 144 hammer_cursor_up(cursor); 145 } 146 } 147 148 149 /* 150 * We are finished with a cursor. We NULL out various fields as sanity 151 * check, in case the structure is inappropriately used afterwords. 152 */ 153 void 154 hammer_done_cursor(hammer_cursor_t cursor) 155 { 156 hammer_inode_t ip; 157 158 KKASSERT((cursor->flags & HAMMER_CURSOR_TRACKED) == 0); 159 if (cursor->parent) { 160 hammer_unlock(&cursor->parent->lock); 161 hammer_rel_node(cursor->parent); 162 cursor->parent = NULL; 163 } 164 if (cursor->node) { 165 hammer_unlock(&cursor->node->lock); 166 hammer_rel_node(cursor->node); 167 cursor->node = NULL; 168 } 169 if (cursor->data_buffer) { 170 hammer_rel_buffer(cursor->data_buffer, 0); 171 cursor->data_buffer = NULL; 172 } 173 if ((ip = cursor->ip) != NULL) { 174 KKASSERT(ip->cursor_ip_refs > 0); 175 --ip->cursor_ip_refs; 176 hammer_unlock(&ip->lock); 177 cursor->ip = NULL; 178 } 179 if (cursor->iprec) { 180 hammer_rel_mem_record(cursor->iprec); 181 cursor->iprec = NULL; 182 } 183 184 /* 185 * If we deadlocked this node will be referenced. Do a quick 186 * lock/unlock to wait for the deadlock condition to clear. 187 */ 188 if (cursor->deadlk_node) { 189 hammer_lock_ex_ident(&cursor->deadlk_node->lock, "hmrdlk"); 190 hammer_unlock(&cursor->deadlk_node->lock); 191 hammer_rel_node(cursor->deadlk_node); 192 cursor->deadlk_node = NULL; 193 } 194 if (cursor->deadlk_rec) { 195 hammer_wait_mem_record_ident(cursor->deadlk_rec, "hmmdlr"); 196 hammer_rel_mem_record(cursor->deadlk_rec); 197 cursor->deadlk_rec = NULL; 198 } 199 200 cursor->data = NULL; 201 cursor->leaf = NULL; 202 cursor->left_bound = NULL; 203 cursor->right_bound = NULL; 204 cursor->trans = NULL; 205 } 206 207 /* 208 * Upgrade cursor->node and cursor->parent to exclusive locks. This 209 * function can return EDEADLK. 210 * 211 * The lock must already be either held shared or already held exclusively 212 * by us. 213 * 214 * If we fail to upgrade the lock and cursor->deadlk_node is NULL, 215 * we add another reference to the node that failed and set 216 * cursor->deadlk_node so hammer_done_cursor() can block on it. 217 */ 218 int 219 hammer_cursor_upgrade(hammer_cursor_t cursor) 220 { 221 int error; 222 223 error = hammer_lock_upgrade(&cursor->node->lock); 224 if (error && cursor->deadlk_node == NULL) { 225 cursor->deadlk_node = cursor->node; 226 hammer_ref_node(cursor->deadlk_node); 227 } else if (error == 0 && cursor->parent) { 228 error = hammer_lock_upgrade(&cursor->parent->lock); 229 if (error && cursor->deadlk_node == NULL) { 230 cursor->deadlk_node = cursor->parent; 231 hammer_ref_node(cursor->deadlk_node); 232 } 233 } 234 return(error); 235 } 236 237 int 238 hammer_cursor_upgrade_node(hammer_cursor_t cursor) 239 { 240 int error; 241 242 error = hammer_lock_upgrade(&cursor->node->lock); 243 if (error && cursor->deadlk_node == NULL) { 244 cursor->deadlk_node = cursor->node; 245 hammer_ref_node(cursor->deadlk_node); 246 } 247 return(error); 248 } 249 250 /* 251 * Downgrade cursor->node and cursor->parent to shared locks. This 252 * function can return EDEADLK. 253 */ 254 void 255 hammer_cursor_downgrade(hammer_cursor_t cursor) 256 { 257 if (hammer_lock_excl_owned(&cursor->node->lock, curthread)) 258 hammer_lock_downgrade(&cursor->node->lock); 259 if (cursor->parent && 260 hammer_lock_excl_owned(&cursor->parent->lock, curthread)) { 261 hammer_lock_downgrade(&cursor->parent->lock); 262 } 263 } 264 265 /* 266 * Seek the cursor to the specified node and index. 267 * 268 * The caller must ref the node prior to calling this routine and release 269 * it after it returns. If the seek succeeds the cursor will gain its own 270 * ref on the node. 271 */ 272 int 273 hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node, int index) 274 { 275 int error; 276 277 hammer_cursor_downgrade(cursor); 278 error = 0; 279 280 if (cursor->node != node) { 281 hammer_unlock(&cursor->node->lock); 282 hammer_rel_node(cursor->node); 283 cursor->node = node; 284 hammer_ref_node(node); 285 hammer_lock_sh(&node->lock); 286 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0); 287 288 if (cursor->parent) { 289 hammer_unlock(&cursor->parent->lock); 290 hammer_rel_node(cursor->parent); 291 cursor->parent = NULL; 292 cursor->parent_index = 0; 293 } 294 error = hammer_load_cursor_parent(cursor, 0); 295 } 296 cursor->index = index; 297 return (error); 298 } 299 300 /* 301 * Load the parent of cursor->node into cursor->parent. 302 */ 303 static 304 int 305 hammer_load_cursor_parent(hammer_cursor_t cursor, int try_exclusive) 306 { 307 hammer_mount_t hmp; 308 hammer_node_t parent; 309 hammer_node_t node; 310 hammer_btree_elm_t elm; 311 int error; 312 int parent_index; 313 314 hmp = cursor->trans->hmp; 315 316 if (cursor->node->ondisk->parent) { 317 node = cursor->node; 318 parent = hammer_btree_get_parent(cursor->trans, node, 319 &parent_index, 320 &error, try_exclusive); 321 if (error == 0) { 322 elm = &parent->ondisk->elms[parent_index]; 323 cursor->parent = parent; 324 cursor->parent_index = parent_index; 325 cursor->left_bound = &elm[0].internal.base; 326 cursor->right_bound = &elm[1].internal.base; 327 } 328 } else { 329 cursor->parent = NULL; 330 cursor->parent_index = 0; 331 cursor->left_bound = &hmp->root_btree_beg; 332 cursor->right_bound = &hmp->root_btree_end; 333 error = 0; 334 } 335 return(error); 336 } 337 338 /* 339 * Cursor up to our parent node. Return ENOENT if we are at the root of 340 * the filesystem. 341 */ 342 int 343 hammer_cursor_up(hammer_cursor_t cursor) 344 { 345 int error; 346 347 hammer_cursor_downgrade(cursor); 348 349 /* 350 * If the parent is NULL we are at the root of the B-Tree and 351 * return ENOENT. 352 */ 353 if (cursor->parent == NULL) 354 return (ENOENT); 355 356 /* 357 * Set the node to its parent. 358 */ 359 hammer_unlock(&cursor->node->lock); 360 hammer_rel_node(cursor->node); 361 cursor->node = cursor->parent; 362 cursor->index = cursor->parent_index; 363 cursor->parent = NULL; 364 cursor->parent_index = 0; 365 366 error = hammer_load_cursor_parent(cursor, 0); 367 return(error); 368 } 369 370 /* 371 * Special cursor up given a locked cursor. The orignal node is not 372 * unlocked or released and the cursor is not downgraded. 373 * 374 * This function will recover from deadlocks. EDEADLK cannot be returned. 375 */ 376 int 377 hammer_cursor_up_locked(hammer_cursor_t cursor) 378 { 379 hammer_node_t save; 380 int error; 381 382 /* 383 * If the parent is NULL we are at the root of the B-Tree and 384 * return ENOENT. 385 */ 386 if (cursor->parent == NULL) 387 return (ENOENT); 388 389 save = cursor->node; 390 391 /* 392 * Set the node to its parent. 393 */ 394 cursor->node = cursor->parent; 395 cursor->index = cursor->parent_index; 396 cursor->parent = NULL; 397 cursor->parent_index = 0; 398 399 /* 400 * load the new parent, attempt to exclusively lock it. Note that 401 * we are still holding the old parent (now cursor->node) exclusively 402 * locked. This can return EDEADLK. 403 */ 404 error = hammer_load_cursor_parent(cursor, 1); 405 if (error) { 406 cursor->parent = cursor->node; 407 cursor->parent_index = cursor->index; 408 cursor->node = save; 409 cursor->index = 0; 410 } 411 return(error); 412 } 413 414 415 /* 416 * Cursor down through the current node, which must be an internal node. 417 * 418 * This routine adjusts the cursor and sets index to 0. 419 */ 420 int 421 hammer_cursor_down(hammer_cursor_t cursor) 422 { 423 hammer_node_t node; 424 hammer_btree_elm_t elm; 425 int error; 426 427 /* 428 * The current node becomes the current parent 429 */ 430 hammer_cursor_downgrade(cursor); 431 node = cursor->node; 432 KKASSERT(cursor->index >= 0 && cursor->index < node->ondisk->count); 433 if (cursor->parent) { 434 hammer_unlock(&cursor->parent->lock); 435 hammer_rel_node(cursor->parent); 436 } 437 cursor->parent = node; 438 cursor->parent_index = cursor->index; 439 cursor->node = NULL; 440 cursor->index = 0; 441 442 /* 443 * Extract element to push into at (node,index), set bounds. 444 */ 445 elm = &node->ondisk->elms[cursor->parent_index]; 446 447 /* 448 * Ok, push down into elm. If elm specifies an internal or leaf 449 * node the current node must be an internal node. If elm specifies 450 * a spike then the current node must be a leaf node. 451 */ 452 switch(elm->base.btype) { 453 case HAMMER_BTREE_TYPE_INTERNAL: 454 case HAMMER_BTREE_TYPE_LEAF: 455 KKASSERT(node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL); 456 KKASSERT(elm->internal.subtree_offset != 0); 457 cursor->left_bound = &elm[0].internal.base; 458 cursor->right_bound = &elm[1].internal.base; 459 node = hammer_get_node(cursor->trans, 460 elm->internal.subtree_offset, 0, &error); 461 if (error == 0) { 462 KASSERT(elm->base.btype == node->ondisk->type, ("BTYPE MISMATCH %c %c NODE %p\n", elm->base.btype, node->ondisk->type, node)); 463 if (node->ondisk->parent != cursor->parent->node_offset) 464 panic("node %p %016llx vs %016llx\n", node, (long long)node->ondisk->parent, (long long)cursor->parent->node_offset); 465 KKASSERT(node->ondisk->parent == cursor->parent->node_offset); 466 } 467 break; 468 default: 469 panic("hammer_cursor_down: illegal btype %02x (%c)\n", 470 elm->base.btype, 471 (elm->base.btype ? elm->base.btype : '?')); 472 break; 473 } 474 if (error == 0) { 475 hammer_lock_sh(&node->lock); 476 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0); 477 cursor->node = node; 478 cursor->index = 0; 479 } 480 return(error); 481 } 482 483 /************************************************************************ 484 * DEADLOCK RECOVERY * 485 ************************************************************************ 486 * 487 * These are the new deadlock recovery functions. Currently they are only 488 * used for the mirror propagation and physical node removal cases but 489 * ultimately the intention is to use them for all deadlock recovery 490 * operations. 491 * 492 * WARNING! The contents of the cursor may be modified while unlocked. 493 * passive modifications including adjusting the node, parent, 494 * indexes, and leaf pointer. 495 * 496 * An outright removal of the element the cursor was pointing at 497 * will cause the HAMMER_CURSOR_TRACKED_RIPOUT flag to be set, 498 * which chains to causing the HAMMER_CURSOR_RETEST to be set 499 * when the cursor is locked again. 500 */ 501 void 502 hammer_unlock_cursor(hammer_cursor_t cursor) 503 { 504 hammer_node_t node; 505 506 KKASSERT((cursor->flags & HAMMER_CURSOR_TRACKED) == 0); 507 KKASSERT(cursor->node); 508 509 /* 510 * Release the cursor's locks and track B-Tree operations on node. 511 * While being tracked our cursor can be modified by other threads 512 * and the node may be replaced. 513 */ 514 if (cursor->parent) { 515 hammer_unlock(&cursor->parent->lock); 516 hammer_rel_node(cursor->parent); 517 cursor->parent = NULL; 518 } 519 node = cursor->node; 520 cursor->flags |= HAMMER_CURSOR_TRACKED; 521 TAILQ_INSERT_TAIL(&node->cursor_list, cursor, deadlk_entry); 522 hammer_unlock(&node->lock); 523 } 524 525 /* 526 * Get the cursor heated up again. The cursor's node may have 527 * changed and we might have to locate the new parent. 528 * 529 * If the exact element we were on got deleted RIPOUT will be 530 * set and we must clear ATEDISK so an iteration does not skip 531 * the element after it. 532 */ 533 int 534 hammer_lock_cursor(hammer_cursor_t cursor) 535 { 536 hammer_node_t node; 537 int error; 538 539 KKASSERT(cursor->flags & HAMMER_CURSOR_TRACKED); 540 541 /* 542 * Relock the node 543 */ 544 for (;;) { 545 node = cursor->node; 546 hammer_ref_node(node); 547 hammer_lock_sh(&node->lock); 548 if (cursor->node == node) { 549 hammer_rel_node(node); 550 break; 551 } 552 hammer_unlock(&node->lock); 553 hammer_rel_node(node); 554 } 555 556 /* 557 * Untrack the cursor, clean up, and re-establish the parent node. 558 */ 559 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry); 560 cursor->flags &= ~HAMMER_CURSOR_TRACKED; 561 562 /* 563 * If a ripout has occured iterations must re-test the (new) 564 * current element. Clearing ATEDISK prevents the element from 565 * being skipped and RETEST causes it to be re-tested. 566 */ 567 if (cursor->flags & HAMMER_CURSOR_TRACKED_RIPOUT) { 568 cursor->flags &= ~HAMMER_CURSOR_TRACKED_RIPOUT; 569 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 570 cursor->flags |= HAMMER_CURSOR_RETEST; 571 } 572 error = hammer_load_cursor_parent(cursor, 0); 573 return(error); 574 } 575 576 /* 577 * Recover from a deadlocked cursor, tracking any node removals or 578 * replacements. If the cursor's current node is removed by another 579 * thread (via btree_remove()) the cursor will be seeked upwards. 580 * 581 * The caller is working a modifying operation and must be holding the 582 * sync lock (shared). We do not release the sync lock because this 583 * would break atomicy. 584 */ 585 int 586 hammer_recover_cursor(hammer_cursor_t cursor) 587 { 588 int error; 589 590 hammer_unlock_cursor(cursor); 591 KKASSERT(cursor->trans->sync_lock_refs > 0); 592 593 /* 594 * Wait for the deadlock to clear 595 */ 596 if (cursor->deadlk_node) { 597 hammer_lock_ex_ident(&cursor->deadlk_node->lock, "hmrdlk"); 598 hammer_unlock(&cursor->deadlk_node->lock); 599 hammer_rel_node(cursor->deadlk_node); 600 cursor->deadlk_node = NULL; 601 } 602 if (cursor->deadlk_rec) { 603 hammer_wait_mem_record_ident(cursor->deadlk_rec, "hmmdlr"); 604 hammer_rel_mem_record(cursor->deadlk_rec); 605 cursor->deadlk_rec = NULL; 606 } 607 error = hammer_lock_cursor(cursor); 608 return(error); 609 } 610 611 /* 612 * Dup ocursor to ncursor. ncursor inherits ocursor's locks and ocursor 613 * is effectively unlocked and becomes tracked. If ocursor was not locked 614 * then ncursor also inherits the tracking. 615 * 616 * After the caller finishes working with ncursor it must be cleaned up 617 * with hammer_done_cursor(), and the caller must re-lock ocursor. 618 */ 619 hammer_cursor_t 620 hammer_push_cursor(hammer_cursor_t ocursor) 621 { 622 hammer_cursor_t ncursor; 623 hammer_inode_t ip; 624 hammer_node_t node; 625 hammer_mount_t hmp; 626 627 hmp = ocursor->trans->hmp; 628 ncursor = kmalloc(sizeof(*ncursor), hmp->m_misc, M_WAITOK | M_ZERO); 629 bcopy(ocursor, ncursor, sizeof(*ocursor)); 630 631 node = ocursor->node; 632 hammer_ref_node(node); 633 if ((ocursor->flags & HAMMER_CURSOR_TRACKED) == 0) { 634 ocursor->flags |= HAMMER_CURSOR_TRACKED; 635 TAILQ_INSERT_TAIL(&node->cursor_list, ocursor, deadlk_entry); 636 } 637 if (ncursor->parent) 638 ocursor->parent = NULL; 639 ocursor->data_buffer = NULL; 640 ocursor->leaf = NULL; 641 ocursor->data = NULL; 642 if (ncursor->flags & HAMMER_CURSOR_TRACKED) 643 TAILQ_INSERT_TAIL(&node->cursor_list, ncursor, deadlk_entry); 644 if ((ip = ncursor->ip) != NULL) { 645 ++ip->cursor_ip_refs; 646 } 647 if (ncursor->iprec) 648 hammer_ref(&ncursor->iprec->lock); 649 return(ncursor); 650 } 651 652 /* 653 * Destroy ncursor and restore ocursor 654 * 655 * This is a temporary hack for the release. We can't afford to lose 656 * the IP lock until the IP object scan code is able to deal with it, 657 * so have ocursor inherit it back. 658 */ 659 void 660 hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor) 661 { 662 hammer_mount_t hmp; 663 hammer_inode_t ip; 664 665 hmp = ncursor->trans->hmp; 666 ip = ncursor->ip; 667 ncursor->ip = NULL; 668 if (ip) 669 --ip->cursor_ip_refs; 670 hammer_done_cursor(ncursor); 671 kfree(ncursor, hmp->m_misc); 672 KKASSERT(ocursor->ip == ip); 673 hammer_lock_cursor(ocursor); 674 } 675 676 /* 677 * onode is being replaced by nnode by the reblocking code. 678 */ 679 void 680 hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode) 681 { 682 hammer_cursor_t cursor; 683 hammer_node_ondisk_t ondisk; 684 hammer_node_ondisk_t nndisk; 685 686 ondisk = onode->ondisk; 687 nndisk = nnode->ondisk; 688 689 while ((cursor = TAILQ_FIRST(&onode->cursor_list)) != NULL) { 690 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 691 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 692 KKASSERT(cursor->node == onode); 693 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 694 cursor->leaf = &nndisk->elms[cursor->index].leaf; 695 cursor->node = nnode; 696 hammer_ref_node(nnode); 697 hammer_rel_node(onode); 698 } 699 } 700 701 /* 702 * node is being removed, cursors in deadlock recovery are seeked upward 703 * to the parent. 704 */ 705 void 706 hammer_cursor_removed_node(hammer_node_t node, hammer_node_t parent, int index) 707 { 708 hammer_cursor_t cursor; 709 hammer_node_ondisk_t ondisk; 710 711 KKASSERT(parent != NULL); 712 ondisk = node->ondisk; 713 714 while ((cursor = TAILQ_FIRST(&node->cursor_list)) != NULL) { 715 KKASSERT(cursor->node == node); 716 KKASSERT(cursor->index == 0); 717 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry); 718 TAILQ_INSERT_TAIL(&parent->cursor_list, cursor, deadlk_entry); 719 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 720 cursor->leaf = NULL; 721 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT; 722 cursor->node = parent; 723 cursor->index = index; 724 hammer_ref_node(parent); 725 hammer_rel_node(node); 726 } 727 } 728 729 /* 730 * node was split at (onode, index) with elements >= index moved to nnode. 731 */ 732 void 733 hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode, int index) 734 { 735 hammer_cursor_t cursor; 736 hammer_node_ondisk_t ondisk; 737 hammer_node_ondisk_t nndisk; 738 739 ondisk = onode->ondisk; 740 nndisk = nnode->ondisk; 741 742 again: 743 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) { 744 KKASSERT(cursor->node == onode); 745 if (cursor->index < index) 746 continue; 747 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 748 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 749 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 750 cursor->leaf = &nndisk->elms[cursor->index - index].leaf; 751 cursor->node = nnode; 752 cursor->index -= index; 753 hammer_ref_node(nnode); 754 hammer_rel_node(onode); 755 goto again; 756 } 757 } 758 759 /* 760 * An element was moved from one node to another or within a node. The 761 * index may also represent the end of the node (index == numelements). 762 * 763 * This is used by the rebalancing code. This is not an insertion or 764 * deletion and any additional elements, including the degenerate case at 765 * the end of the node, will be dealt with by additional distinct calls. 766 */ 767 void 768 hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode, 769 int oindex, int nindex) 770 { 771 hammer_cursor_t cursor; 772 hammer_node_ondisk_t ondisk; 773 hammer_node_ondisk_t nndisk; 774 775 ondisk = onode->ondisk; 776 nndisk = nnode->ondisk; 777 778 again: 779 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) { 780 KKASSERT(cursor->node == onode); 781 if (cursor->index != oindex) 782 continue; 783 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 784 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 785 if (cursor->leaf == &ondisk->elms[oindex].leaf) 786 cursor->leaf = &nndisk->elms[nindex].leaf; 787 cursor->node = nnode; 788 cursor->index = nindex; 789 hammer_ref_node(nnode); 790 hammer_rel_node(onode); 791 goto again; 792 } 793 } 794 795 /* 796 * The B-Tree element pointing to the specified node was moved from (oparent) 797 * to (nparent, nindex). We must locate any tracked cursors pointing at 798 * node and adjust their parent accordingly. 799 * 800 * This is used by the rebalancing code when packing elements causes an 801 * element to shift from one node to another. 802 */ 803 void 804 hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent, 805 hammer_node_t nparent, int nindex) 806 { 807 hammer_cursor_t cursor; 808 809 again: 810 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 811 KKASSERT(cursor->node == node); 812 if (cursor->parent == oparent) { 813 cursor->parent = nparent; 814 cursor->parent_index = nindex; 815 hammer_ref_node(nparent); 816 hammer_rel_node(oparent); 817 goto again; 818 } 819 } 820 } 821 822 /* 823 * Deleted element at (node, index) 824 * 825 * Shift indexes >= index 826 */ 827 void 828 hammer_cursor_deleted_element(hammer_node_t node, int index) 829 { 830 hammer_cursor_t cursor; 831 hammer_node_ondisk_t ondisk; 832 833 ondisk = node->ondisk; 834 835 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 836 KKASSERT(cursor->node == node); 837 if (cursor->index == index) { 838 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT; 839 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 840 cursor->leaf = NULL; 841 } else if (cursor->index > index) { 842 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 843 cursor->leaf = &ondisk->elms[cursor->index - 1].leaf; 844 --cursor->index; 845 } 846 } 847 } 848 849 /* 850 * Inserted element at (node, index) 851 * 852 * Shift indexes >= index 853 */ 854 void 855 hammer_cursor_inserted_element(hammer_node_t node, int index) 856 { 857 hammer_cursor_t cursor; 858 hammer_node_ondisk_t ondisk; 859 860 ondisk = node->ondisk; 861 862 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 863 KKASSERT(cursor->node == node); 864 if (cursor->index >= index) { 865 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 866 cursor->leaf = &ondisk->elms[cursor->index + 1].leaf; 867 ++cursor->index; 868 } 869 } 870 } 871 872