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, 1); 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, 1); 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, 1); 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, 1); 259 if (cursor->parent && 260 hammer_lock_excl_owned(&cursor->parent->lock, curthread)) { 261 hammer_lock_downgrade(&cursor->parent->lock, 1); 262 } 263 } 264 265 /* 266 * Upgrade and downgrade pairs of cursors. This is used by the dedup 267 * code which must deal with two cursors. A special function is needed 268 * because some of the nodes may be shared between the two cursors, 269 * resulting in share counts > 1 which will normally cause an upgrade 270 * to fail. 271 */ 272 static __noinline 273 int 274 collect_node(hammer_node_t *array, int *counts, int n, hammer_node_t node) 275 { 276 int i; 277 278 for (i = 0; i < n; ++i) { 279 if (array[i] == node) 280 break; 281 } 282 if (i == n) { 283 array[i] = node; 284 counts[i] = 1; 285 ++i; 286 } else { 287 ++counts[i]; 288 } 289 return(i); 290 } 291 292 int 293 hammer_cursor_upgrade2(hammer_cursor_t cursor1, hammer_cursor_t cursor2) 294 { 295 hammer_node_t nodes[4]; 296 int counts[4]; 297 int error; 298 int i; 299 int n; 300 301 n = collect_node(nodes, counts, 0, cursor1->node); 302 if (cursor1->parent) 303 n = collect_node(nodes, counts, n, cursor1->parent); 304 n = collect_node(nodes, counts, n, cursor2->node); 305 if (cursor2->parent) 306 n = collect_node(nodes, counts, n, cursor2->parent); 307 308 error = 0; 309 for (i = 0; i < n; ++i) { 310 error = hammer_lock_upgrade(&nodes[i]->lock, counts[i]); 311 if (error) 312 break; 313 } 314 if (error) { 315 while (--i >= 0) 316 hammer_lock_downgrade(&nodes[i]->lock, counts[i]); 317 } 318 return (error); 319 } 320 321 void 322 hammer_cursor_downgrade2(hammer_cursor_t cursor1, hammer_cursor_t cursor2) 323 { 324 hammer_node_t nodes[4]; 325 int counts[4]; 326 int i; 327 int n; 328 329 n = collect_node(nodes, counts, 0, cursor1->node); 330 if (cursor1->parent) 331 n = collect_node(nodes, counts, n, cursor1->parent); 332 n = collect_node(nodes, counts, n, cursor2->node); 333 if (cursor2->parent) 334 n = collect_node(nodes, counts, n, cursor2->parent); 335 336 for (i = 0; i < n; ++i) 337 hammer_lock_downgrade(&nodes[i]->lock, counts[i]); 338 } 339 340 /* 341 * Seek the cursor to the specified node and index. 342 * 343 * The caller must ref the node prior to calling this routine and release 344 * it after it returns. If the seek succeeds the cursor will gain its own 345 * ref on the node. 346 */ 347 int 348 hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node, int index) 349 { 350 int error; 351 352 hammer_cursor_downgrade(cursor); 353 error = 0; 354 355 if (cursor->node != node) { 356 hammer_unlock(&cursor->node->lock); 357 hammer_rel_node(cursor->node); 358 cursor->node = node; 359 hammer_ref_node(node); 360 hammer_lock_sh(&node->lock); 361 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0); 362 363 if (cursor->parent) { 364 hammer_unlock(&cursor->parent->lock); 365 hammer_rel_node(cursor->parent); 366 cursor->parent = NULL; 367 cursor->parent_index = 0; 368 } 369 error = hammer_load_cursor_parent(cursor, 0); 370 } 371 cursor->index = index; 372 return (error); 373 } 374 375 /* 376 * Load the parent of cursor->node into cursor->parent. 377 */ 378 static 379 int 380 hammer_load_cursor_parent(hammer_cursor_t cursor, int try_exclusive) 381 { 382 hammer_mount_t hmp; 383 hammer_node_t parent; 384 hammer_node_t node; 385 hammer_btree_elm_t elm; 386 int error; 387 int parent_index; 388 389 hmp = cursor->trans->hmp; 390 391 if (cursor->node->ondisk->parent) { 392 node = cursor->node; 393 parent = hammer_btree_get_parent(cursor->trans, node, 394 &parent_index, 395 &error, try_exclusive); 396 if (error == 0) { 397 elm = &parent->ondisk->elms[parent_index]; 398 cursor->parent = parent; 399 cursor->parent_index = parent_index; 400 cursor->left_bound = &elm[0].internal.base; 401 cursor->right_bound = &elm[1].internal.base; 402 } 403 } else { 404 cursor->parent = NULL; 405 cursor->parent_index = 0; 406 cursor->left_bound = &hmp->root_btree_beg; 407 cursor->right_bound = &hmp->root_btree_end; 408 error = 0; 409 } 410 return(error); 411 } 412 413 /* 414 * Cursor up to our parent node. Return ENOENT if we are at the root of 415 * the filesystem. 416 */ 417 int 418 hammer_cursor_up(hammer_cursor_t cursor) 419 { 420 int error; 421 422 hammer_cursor_downgrade(cursor); 423 424 /* 425 * If the parent is NULL we are at the root of the B-Tree and 426 * return ENOENT. 427 */ 428 if (cursor->parent == NULL) 429 return (ENOENT); 430 431 /* 432 * Set the node to its parent. 433 */ 434 hammer_unlock(&cursor->node->lock); 435 hammer_rel_node(cursor->node); 436 cursor->node = cursor->parent; 437 cursor->index = cursor->parent_index; 438 cursor->parent = NULL; 439 cursor->parent_index = 0; 440 441 error = hammer_load_cursor_parent(cursor, 0); 442 return(error); 443 } 444 445 /* 446 * Special cursor up given a locked cursor. The orignal node is not 447 * unlocked or released and the cursor is not downgraded. 448 * 449 * This function can fail with EDEADLK. 450 * 451 * This function is only run when recursively deleting parent nodes 452 * to get rid of an empty leaf. 453 */ 454 int 455 hammer_cursor_up_locked(hammer_cursor_t cursor) 456 { 457 hammer_node_t save; 458 int error; 459 int save_index; 460 461 /* 462 * If the parent is NULL we are at the root of the B-Tree and 463 * return ENOENT. 464 */ 465 if (cursor->parent == NULL) 466 return (ENOENT); 467 468 save = cursor->node; 469 save_index = cursor->index; 470 471 /* 472 * Set the node to its parent. 473 */ 474 cursor->node = cursor->parent; 475 cursor->index = cursor->parent_index; 476 cursor->parent = NULL; 477 cursor->parent_index = 0; 478 479 /* 480 * load the new parent, attempt to exclusively lock it. Note that 481 * we are still holding the old parent (now cursor->node) exclusively 482 * locked. 483 * 484 * This can return EDEADLK. Undo the operation on any error. These 485 * up sequences can occur during iterations so be sure to restore 486 * the index. 487 */ 488 error = hammer_load_cursor_parent(cursor, 1); 489 if (error) { 490 cursor->parent = cursor->node; 491 cursor->parent_index = cursor->index; 492 cursor->node = save; 493 cursor->index = save_index; 494 } 495 return(error); 496 } 497 498 499 /* 500 * Cursor down through the current node, which must be an internal node. 501 * 502 * This routine adjusts the cursor and sets index to 0. 503 */ 504 int 505 hammer_cursor_down(hammer_cursor_t cursor) 506 { 507 hammer_node_t node; 508 hammer_btree_elm_t elm; 509 int error; 510 511 /* 512 * The current node becomes the current parent 513 */ 514 hammer_cursor_downgrade(cursor); 515 node = cursor->node; 516 KKASSERT(cursor->index >= 0 && cursor->index < node->ondisk->count); 517 if (cursor->parent) { 518 hammer_unlock(&cursor->parent->lock); 519 hammer_rel_node(cursor->parent); 520 } 521 cursor->parent = node; 522 cursor->parent_index = cursor->index; 523 cursor->node = NULL; 524 cursor->index = 0; 525 526 /* 527 * Extract element to push into at (node,index), set bounds. 528 */ 529 elm = &node->ondisk->elms[cursor->parent_index]; 530 531 /* 532 * Ok, push down into elm. If elm specifies an internal or leaf 533 * node the current node must be an internal node. If elm specifies 534 * a spike then the current node must be a leaf node. 535 */ 536 switch(elm->base.btype) { 537 case HAMMER_BTREE_TYPE_INTERNAL: 538 case HAMMER_BTREE_TYPE_LEAF: 539 KKASSERT(node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL); 540 KKASSERT(elm->internal.subtree_offset != 0); 541 cursor->left_bound = &elm[0].internal.base; 542 cursor->right_bound = &elm[1].internal.base; 543 node = hammer_get_node(cursor->trans, 544 elm->internal.subtree_offset, 0, &error); 545 if (error == 0) { 546 KASSERT(elm->base.btype == node->ondisk->type, ("BTYPE MISMATCH %c %c NODE %p\n", elm->base.btype, node->ondisk->type, node)); 547 if (node->ondisk->parent != cursor->parent->node_offset) 548 panic("node %p %016llx vs %016llx\n", node, (long long)node->ondisk->parent, (long long)cursor->parent->node_offset); 549 KKASSERT(node->ondisk->parent == cursor->parent->node_offset); 550 } 551 break; 552 default: 553 panic("hammer_cursor_down: illegal btype %02x (%c)\n", 554 elm->base.btype, 555 (elm->base.btype ? elm->base.btype : '?')); 556 break; 557 } 558 if (error == 0) { 559 hammer_lock_sh(&node->lock); 560 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0); 561 cursor->node = node; 562 cursor->index = 0; 563 } 564 return(error); 565 } 566 567 /************************************************************************ 568 * DEADLOCK RECOVERY * 569 ************************************************************************ 570 * 571 * These are the new deadlock recovery functions. Currently they are only 572 * used for the mirror propagation and physical node removal cases but 573 * ultimately the intention is to use them for all deadlock recovery 574 * operations. 575 * 576 * WARNING! The contents of the cursor may be modified while unlocked. 577 * passive modifications including adjusting the node, parent, 578 * indexes, and leaf pointer. 579 * 580 * An outright removal of the element the cursor was pointing at 581 * will cause the HAMMER_CURSOR_TRACKED_RIPOUT flag to be set, 582 * which chains to causing the HAMMER_CURSOR_RETEST to be set 583 * when the cursor is locked again. 584 */ 585 void 586 hammer_unlock_cursor(hammer_cursor_t cursor) 587 { 588 hammer_node_t node; 589 590 KKASSERT((cursor->flags & HAMMER_CURSOR_TRACKED) == 0); 591 KKASSERT(cursor->node); 592 593 /* 594 * Release the cursor's locks and track B-Tree operations on node. 595 * While being tracked our cursor can be modified by other threads 596 * and the node may be replaced. 597 */ 598 if (cursor->parent) { 599 hammer_unlock(&cursor->parent->lock); 600 hammer_rel_node(cursor->parent); 601 cursor->parent = NULL; 602 } 603 node = cursor->node; 604 cursor->flags |= HAMMER_CURSOR_TRACKED; 605 TAILQ_INSERT_TAIL(&node->cursor_list, cursor, deadlk_entry); 606 hammer_unlock(&node->lock); 607 } 608 609 /* 610 * Get the cursor heated up again. The cursor's node may have 611 * changed and we might have to locate the new parent. 612 * 613 * If the exact element we were on got deleted RIPOUT will be 614 * set and we must clear ATEDISK so an iteration does not skip 615 * the element after it. 616 */ 617 int 618 hammer_lock_cursor(hammer_cursor_t cursor) 619 { 620 hammer_node_t node; 621 int error; 622 623 KKASSERT(cursor->flags & HAMMER_CURSOR_TRACKED); 624 625 /* 626 * Relock the node 627 */ 628 for (;;) { 629 node = cursor->node; 630 hammer_ref_node(node); 631 hammer_lock_sh(&node->lock); 632 if (cursor->node == node) { 633 hammer_rel_node(node); 634 break; 635 } 636 hammer_unlock(&node->lock); 637 hammer_rel_node(node); 638 } 639 640 /* 641 * Untrack the cursor, clean up, and re-establish the parent node. 642 */ 643 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry); 644 cursor->flags &= ~HAMMER_CURSOR_TRACKED; 645 646 /* 647 * If a ripout has occured iterations must re-test the (new) 648 * current element. Clearing ATEDISK prevents the element from 649 * being skipped and RETEST causes it to be re-tested. 650 */ 651 if (cursor->flags & HAMMER_CURSOR_TRACKED_RIPOUT) { 652 cursor->flags &= ~HAMMER_CURSOR_TRACKED_RIPOUT; 653 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 654 cursor->flags |= HAMMER_CURSOR_RETEST; 655 } 656 error = hammer_load_cursor_parent(cursor, 0); 657 return(error); 658 } 659 660 /* 661 * Recover from a deadlocked cursor, tracking any node removals or 662 * replacements. If the cursor's current node is removed by another 663 * thread (via btree_remove()) the cursor will be seeked upwards. 664 * 665 * The caller is working a modifying operation and must be holding the 666 * sync lock (shared). We do not release the sync lock because this 667 * would break atomicy. 668 */ 669 int 670 hammer_recover_cursor(hammer_cursor_t cursor) 671 { 672 int error; 673 674 hammer_unlock_cursor(cursor); 675 KKASSERT(cursor->trans->sync_lock_refs > 0); 676 677 /* 678 * Wait for the deadlock to clear 679 */ 680 if (cursor->deadlk_node) { 681 hammer_lock_ex_ident(&cursor->deadlk_node->lock, "hmrdlk"); 682 hammer_unlock(&cursor->deadlk_node->lock); 683 hammer_rel_node(cursor->deadlk_node); 684 cursor->deadlk_node = NULL; 685 } 686 if (cursor->deadlk_rec) { 687 hammer_wait_mem_record_ident(cursor->deadlk_rec, "hmmdlr"); 688 hammer_rel_mem_record(cursor->deadlk_rec); 689 cursor->deadlk_rec = NULL; 690 } 691 error = hammer_lock_cursor(cursor); 692 return(error); 693 } 694 695 /* 696 * Dup ocursor to ncursor. ncursor inherits ocursor's locks and ocursor 697 * is effectively unlocked and becomes tracked. If ocursor was not locked 698 * then ncursor also inherits the tracking. 699 * 700 * After the caller finishes working with ncursor it must be cleaned up 701 * with hammer_done_cursor(), and the caller must re-lock ocursor. 702 */ 703 hammer_cursor_t 704 hammer_push_cursor(hammer_cursor_t ocursor) 705 { 706 hammer_cursor_t ncursor; 707 hammer_inode_t ip; 708 hammer_node_t node; 709 hammer_mount_t hmp; 710 711 hmp = ocursor->trans->hmp; 712 ncursor = kmalloc(sizeof(*ncursor), hmp->m_misc, M_WAITOK | M_ZERO); 713 bcopy(ocursor, ncursor, sizeof(*ocursor)); 714 715 node = ocursor->node; 716 hammer_ref_node(node); 717 if ((ocursor->flags & HAMMER_CURSOR_TRACKED) == 0) { 718 ocursor->flags |= HAMMER_CURSOR_TRACKED; 719 TAILQ_INSERT_TAIL(&node->cursor_list, ocursor, deadlk_entry); 720 } 721 if (ncursor->parent) 722 ocursor->parent = NULL; 723 ocursor->data_buffer = NULL; 724 ocursor->leaf = NULL; 725 ocursor->data = NULL; 726 if (ncursor->flags & HAMMER_CURSOR_TRACKED) 727 TAILQ_INSERT_TAIL(&node->cursor_list, ncursor, deadlk_entry); 728 if ((ip = ncursor->ip) != NULL) { 729 ++ip->cursor_ip_refs; 730 } 731 if (ncursor->iprec) 732 hammer_ref(&ncursor->iprec->lock); 733 return(ncursor); 734 } 735 736 /* 737 * Destroy ncursor and restore ocursor 738 * 739 * This is a temporary hack for the release. We can't afford to lose 740 * the IP lock until the IP object scan code is able to deal with it, 741 * so have ocursor inherit it back. 742 */ 743 void 744 hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor) 745 { 746 hammer_mount_t hmp; 747 hammer_inode_t ip; 748 749 hmp = ncursor->trans->hmp; 750 ip = ncursor->ip; 751 ncursor->ip = NULL; 752 if (ip) 753 --ip->cursor_ip_refs; 754 hammer_done_cursor(ncursor); 755 kfree(ncursor, hmp->m_misc); 756 KKASSERT(ocursor->ip == ip); 757 hammer_lock_cursor(ocursor); 758 } 759 760 /* 761 * onode is being replaced by nnode by the reblocking code. 762 */ 763 void 764 hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode) 765 { 766 hammer_cursor_t cursor; 767 hammer_node_ondisk_t ondisk; 768 hammer_node_ondisk_t nndisk; 769 770 ondisk = onode->ondisk; 771 nndisk = nnode->ondisk; 772 773 while ((cursor = TAILQ_FIRST(&onode->cursor_list)) != NULL) { 774 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 775 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 776 KKASSERT(cursor->node == onode); 777 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 778 cursor->leaf = &nndisk->elms[cursor->index].leaf; 779 cursor->node = nnode; 780 hammer_ref_node(nnode); 781 hammer_rel_node(onode); 782 } 783 } 784 785 /* 786 * We have removed <node> from the parent and collapsed the parent. 787 * 788 * Cursors in deadlock recovery are seeked upward to the parent so the 789 * btree_remove() recursion works properly even though we have marked 790 * the cursor as requiring a reseek. 791 * 792 * This is the only cursor function which sets HAMMER_CURSOR_ITERATE_CHECK, 793 * meaning the cursor is no longer definitively pointing at an element 794 * within its iteration (if the cursor is being used to iterate). The 795 * iteration code will take this into account instead of asserting if the 796 * cursor is outside the iteration range. 797 */ 798 void 799 hammer_cursor_removed_node(hammer_node_t node, hammer_node_t parent, int index) 800 { 801 hammer_cursor_t cursor; 802 hammer_node_ondisk_t ondisk; 803 804 KKASSERT(parent != NULL); 805 ondisk = node->ondisk; 806 807 while ((cursor = TAILQ_FIRST(&node->cursor_list)) != NULL) { 808 KKASSERT(cursor->node == node); 809 KKASSERT(cursor->index == 0); 810 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry); 811 TAILQ_INSERT_TAIL(&parent->cursor_list, cursor, deadlk_entry); 812 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 813 cursor->leaf = NULL; 814 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT; 815 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK; 816 cursor->node = parent; 817 cursor->index = index; 818 hammer_ref_node(parent); 819 hammer_rel_node(node); 820 } 821 } 822 823 /* 824 * node was split at (onode, index) with elements >= index moved to nnode. 825 */ 826 void 827 hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode, int index) 828 { 829 hammer_cursor_t cursor; 830 hammer_node_ondisk_t ondisk; 831 hammer_node_ondisk_t nndisk; 832 833 ondisk = onode->ondisk; 834 nndisk = nnode->ondisk; 835 836 again: 837 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) { 838 KKASSERT(cursor->node == onode); 839 if (cursor->index < index) 840 continue; 841 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 842 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 843 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 844 cursor->leaf = &nndisk->elms[cursor->index - index].leaf; 845 cursor->node = nnode; 846 cursor->index -= index; 847 hammer_ref_node(nnode); 848 hammer_rel_node(onode); 849 goto again; 850 } 851 } 852 853 /* 854 * An element was moved from one node to another or within a node. The 855 * index may also represent the end of the node (index == numelements). 856 * 857 * {oparent,pindex} is the parent node's pointer to onode/oindex. 858 * 859 * This is used by the rebalancing code. This is not an insertion or 860 * deletion and any additional elements, including the degenerate case at 861 * the end of the node, will be dealt with by additional distinct calls. 862 */ 863 void 864 hammer_cursor_moved_element(hammer_node_t oparent, int pindex, 865 hammer_node_t onode, int oindex, 866 hammer_node_t nnode, int nindex) 867 { 868 hammer_cursor_t cursor; 869 hammer_node_ondisk_t ondisk; 870 hammer_node_ondisk_t nndisk; 871 872 /* 873 * Adjust any cursors pointing at the element 874 */ 875 ondisk = onode->ondisk; 876 nndisk = nnode->ondisk; 877 again1: 878 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) { 879 KKASSERT(cursor->node == onode); 880 if (cursor->index != oindex) 881 continue; 882 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry); 883 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 884 if (cursor->leaf == &ondisk->elms[oindex].leaf) 885 cursor->leaf = &nndisk->elms[nindex].leaf; 886 cursor->node = nnode; 887 cursor->index = nindex; 888 hammer_ref_node(nnode); 889 hammer_rel_node(onode); 890 goto again1; 891 } 892 893 /* 894 * When moving the first element of onode to a different node any 895 * cursor which is pointing at (oparent,pindex) must be repointed 896 * to nnode and ATEDISK must be cleared. 897 * 898 * This prevents cursors from losing track due to insertions. 899 * Insertions temporarily release the cursor in order to update 900 * the mirror_tids. It primarily effects the mirror_write code. 901 * The other code paths generally only do a single insertion and 902 * then relookup or drop the cursor. 903 */ 904 if (onode == nnode || oindex) 905 return; 906 ondisk = oparent->ondisk; 907 again2: 908 TAILQ_FOREACH(cursor, &oparent->cursor_list, deadlk_entry) { 909 KKASSERT(cursor->node == oparent); 910 if (cursor->index != pindex) 911 continue; 912 kprintf("HAMMER debug: shifted cursor pointing at parent\n" 913 "parent %016jx:%d onode %016jx:%d nnode %016jx:%d\n", 914 (intmax_t)oparent->node_offset, pindex, 915 (intmax_t)onode->node_offset, oindex, 916 (intmax_t)nnode->node_offset, nindex); 917 TAILQ_REMOVE(&oparent->cursor_list, cursor, deadlk_entry); 918 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry); 919 if (cursor->leaf == &ondisk->elms[oindex].leaf) 920 cursor->leaf = &nndisk->elms[nindex].leaf; 921 cursor->node = nnode; 922 cursor->index = nindex; 923 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 924 hammer_ref_node(nnode); 925 hammer_rel_node(oparent); 926 goto again2; 927 } 928 } 929 930 /* 931 * The B-Tree element pointing to the specified node was moved from (oparent) 932 * to (nparent, nindex). We must locate any tracked cursors pointing at 933 * node and adjust their parent accordingly. 934 * 935 * This is used by the rebalancing code when packing elements causes an 936 * element to shift from one node to another. 937 */ 938 void 939 hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent, 940 hammer_node_t nparent, int nindex) 941 { 942 hammer_cursor_t cursor; 943 944 again: 945 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 946 KKASSERT(cursor->node == node); 947 if (cursor->parent == oparent) { 948 cursor->parent = nparent; 949 cursor->parent_index = nindex; 950 hammer_ref_node(nparent); 951 hammer_rel_node(oparent); 952 goto again; 953 } 954 } 955 } 956 957 /* 958 * Deleted element at (node, index) 959 * 960 * Shift indexes >= index 961 */ 962 void 963 hammer_cursor_deleted_element(hammer_node_t node, int index) 964 { 965 hammer_cursor_t cursor; 966 hammer_node_ondisk_t ondisk; 967 968 ondisk = node->ondisk; 969 970 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 971 KKASSERT(cursor->node == node); 972 if (cursor->index == index) { 973 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT; 974 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 975 cursor->leaf = NULL; 976 } else if (cursor->index > index) { 977 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 978 cursor->leaf = &ondisk->elms[cursor->index - 1].leaf; 979 --cursor->index; 980 } 981 } 982 } 983 984 /* 985 * Inserted element at (node, index) 986 * 987 * Shift indexes >= index 988 */ 989 void 990 hammer_cursor_inserted_element(hammer_node_t node, int index) 991 { 992 hammer_cursor_t cursor; 993 hammer_node_ondisk_t ondisk; 994 995 ondisk = node->ondisk; 996 997 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) { 998 KKASSERT(cursor->node == node); 999 if (cursor->index >= index) { 1000 if (cursor->leaf == &ondisk->elms[cursor->index].leaf) 1001 cursor->leaf = &ondisk->elms[cursor->index + 1].leaf; 1002 ++cursor->index; 1003 } 1004 } 1005 } 1006 1007 /* 1008 * Invalidate the cached data buffer associated with a cursor. 1009 * 1010 * This needs to be done when the underlying block is being freed or 1011 * the referenced buffer can prevent the related buffer cache buffer 1012 * from being properly invalidated. 1013 */ 1014 void 1015 hammer_cursor_invalidate_cache(hammer_cursor_t cursor) 1016 { 1017 if (cursor->data_buffer) { 1018 hammer_rel_buffer(cursor->data_buffer, 0); 1019 cursor->data_buffer = NULL; 1020 cursor->data = NULL; 1021 } 1022 } 1023 1024