HAMMER VFS - Correct seriuos bug in hammer rebalancing code
[dragonfly.git] / sys / vfs / hammer / hammer_cursor.c
CommitLineData
8cd0a023 1/*
b84de5af 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
8cd0a023
MD
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 *
5c8d05e2 34 * $DragonFly: src/sys/vfs/hammer/hammer_cursor.c,v 1.42 2008/08/06 15:38:58 dillon Exp $
8cd0a023
MD
35 */
36
37/*
38 * HAMMER B-Tree index - cursor support routines
39 */
40#include "hammer.h"
41
f36a9737 42static int hammer_load_cursor_parent(hammer_cursor_t cursor, int try_exclusive);
8cd0a023
MD
43
44/*
61aeeb33 45 * Initialize a fresh cursor using the B-Tree node cache. If the cache
47197d71 46 * is not available initialize a fresh cursor at the root of the filesystem.
8cd0a023
MD
47 */
48int
36f82b23 49hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
bcac4bbb 50 hammer_node_cache_t cache, hammer_inode_t ip)
8cd0a023 51{
47197d71 52 hammer_volume_t volume;
61aeeb33 53 hammer_node_t node;
8cd0a023
MD
54 int error;
55
56 bzero(cursor, sizeof(*cursor));
61aeeb33 57
36f82b23
MD
58 cursor->trans = trans;
59
61aeeb33 60 /*
4e17f465
MD
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 /*
61aeeb33
MD
73 * Step 1 - acquire a locked node from the cache if possible
74 */
bcac4bbb 75 if (cache && cache->node) {
4c286c36 76 node = hammer_ref_node_safe(trans, cache, &error);
61aeeb33 77 if (error == 0) {
98da6d8c 78 hammer_lock_sh(&node->lock);
61aeeb33
MD
79 if (node->flags & HAMMER_NODE_DELETED) {
80 hammer_unlock(&node->lock);
81 hammer_rel_node(node);
82 node = NULL;
83 }
61aeeb33 84 }
39d8fd63
MD
85 if (node == NULL)
86 ++hammer_stats_btree_root_iterations;
61aeeb33
MD
87 } else {
88 node = NULL;
39d8fd63 89 ++hammer_stats_btree_root_iterations;
61aeeb33
MD
90 }
91
92 /*
93 * Step 2 - If we couldn't get a node from the cache, get
47197d71 94 * the one from the root of the filesystem.
61aeeb33
MD
95 */
96 while (node == NULL) {
36f82b23 97 volume = hammer_get_root_volume(trans->hmp, &error);
61aeeb33
MD
98 if (error)
99 break;
82010f9f 100 node = hammer_get_node(trans, volume->ondisk->vol0_btree_root,
19619882 101 0, &error);
47197d71 102 hammer_rel_volume(volume, 0);
61aeeb33
MD
103 if (error)
104 break;
98da6d8c 105 hammer_lock_sh(&node->lock);
11ad5ade
MD
106
107 /*
108 * If someone got in before we could lock the node, retry.
109 */
61aeeb33
MD
110 if (node->flags & HAMMER_NODE_DELETED) {
111 hammer_unlock(&node->lock);
112 hammer_rel_node(node);
113 node = NULL;
11ad5ade
MD
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;
61aeeb33 121 }
8cd0a023 122 }
61aeeb33
MD
123
124 /*
125 * Step 3 - finish initializing the cursor by acquiring the parent
126 */
127 cursor->node = node;
8cd0a023 128 if (error == 0)
f36a9737 129 error = hammer_load_cursor_parent(cursor, 0);
0b075555 130 KKASSERT(error == 0);
a84a197d 131 /* if (error) hammer_done_cursor(cursor); */
8cd0a023
MD
132 return(error);
133}
134
4e17f465
MD
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 */
139void
140hammer_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
8cd0a023 149/*
8cd0a023
MD
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 */
153void
154hammer_done_cursor(hammer_cursor_t cursor)
155{
4e17f465
MD
156 hammer_inode_t ip;
157
adf01747 158 KKASSERT((cursor->flags & HAMMER_CURSOR_TRACKED) == 0);
8cd0a023
MD
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 }
4e17f465 173 if ((ip = cursor->ip) != NULL) {
4e17f465
MD
174 KKASSERT(ip->cursor_ip_refs > 0);
175 --ip->cursor_ip_refs;
176 hammer_unlock(&ip->lock);
177 cursor->ip = NULL;
178 }
3f43fb33
MD
179 if (cursor->iprec) {
180 hammer_rel_mem_record(cursor->iprec);
181 cursor->iprec = NULL;
182 }
4e17f465 183
6a37e7e4
MD
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) {
af209b0f 189 hammer_lock_ex_ident(&cursor->deadlk_node->lock, "hmrdlk");
6a37e7e4
MD
190 hammer_unlock(&cursor->deadlk_node->lock);
191 hammer_rel_node(cursor->deadlk_node);
192 cursor->deadlk_node = NULL;
193 }
b84de5af 194 if (cursor->deadlk_rec) {
a99b9ea2 195 hammer_wait_mem_record_ident(cursor->deadlk_rec, "hmmdlr");
b84de5af
MD
196 hammer_rel_mem_record(cursor->deadlk_rec);
197 cursor->deadlk_rec = NULL;
198 }
6a37e7e4 199
40043e7f 200 cursor->data = NULL;
11ad5ade 201 cursor->leaf = NULL;
8cd0a023
MD
202 cursor->left_bound = NULL;
203 cursor->right_bound = NULL;
36f82b23 204 cursor->trans = NULL;
8cd0a023
MD
205}
206
207/*
6a37e7e4
MD
208 * Upgrade cursor->node and cursor->parent to exclusive locks. This
209 * function can return EDEADLK.
210 *
7aa3b8a6
MD
211 * The lock must already be either held shared or already held exclusively
212 * by us.
213 *
6a37e7e4
MD
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 */
218int
219hammer_cursor_upgrade(hammer_cursor_t cursor)
220{
221 int error;
222
7aa3b8a6
MD
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) {
6a37e7e4
MD
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 }
6a37e7e4
MD
233 }
234 return(error);
235}
236
7bc5b8c2
MD
237int
238hammer_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
6a37e7e4
MD
250/*
251 * Downgrade cursor->node and cursor->parent to shared locks. This
252 * function can return EDEADLK.
253 */
254void
255hammer_cursor_downgrade(hammer_cursor_t cursor)
256{
7aa3b8a6 257 if (hammer_lock_excl_owned(&cursor->node->lock, curthread))
6a37e7e4 258 hammer_lock_downgrade(&cursor->node->lock);
7aa3b8a6
MD
259 if (cursor->parent &&
260 hammer_lock_excl_owned(&cursor->parent->lock, curthread)) {
6a37e7e4 261 hammer_lock_downgrade(&cursor->parent->lock);
7aa3b8a6 262 }
6a37e7e4
MD
263}
264
32c90105
MD
265/*
266 * Seek the cursor to the specified node and index.
cb51be26
MD
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.
32c90105
MD
271 */
272int
273hammer_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;
32c90105
MD
284 hammer_ref_node(node);
285 hammer_lock_sh(&node->lock);
f36a9737 286 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0);
32c90105
MD
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 }
f36a9737 294 error = hammer_load_cursor_parent(cursor, 0);
32c90105 295 }
a84a197d 296 cursor->index = index;
32c90105
MD
297 return (error);
298}
6a37e7e4 299
6a37e7e4 300/*
47197d71 301 * Load the parent of cursor->node into cursor->parent.
8cd0a023
MD
302 */
303static
304int
f36a9737 305hammer_load_cursor_parent(hammer_cursor_t cursor, int try_exclusive)
8cd0a023 306{
47197d71
MD
307 hammer_mount_t hmp;
308 hammer_node_t parent;
309 hammer_node_t node;
310 hammer_btree_elm_t elm;
8cd0a023 311 int error;
c82af904 312 int parent_index;
8cd0a023 313
36f82b23 314 hmp = cursor->trans->hmp;
8cd0a023
MD
315
316 if (cursor->node->ondisk->parent) {
47197d71 317 node = cursor->node;
82010f9f
MD
318 parent = hammer_btree_get_parent(cursor->trans, node,
319 &parent_index,
c82af904
MD
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;
a84a197d 327 }
8cd0a023
MD
328 } else {
329 cursor->parent = NULL;
330 cursor->parent_index = 0;
47197d71
MD
331 cursor->left_bound = &hmp->root_btree_beg;
332 cursor->right_bound = &hmp->root_btree_end;
8cd0a023
MD
333 error = 0;
334 }
335 return(error);
336}
337
8cd0a023
MD
338/*
339 * Cursor up to our parent node. Return ENOENT if we are at the root of
47197d71 340 * the filesystem.
8cd0a023
MD
341 */
342int
6a37e7e4 343hammer_cursor_up(hammer_cursor_t cursor)
8cd0a023
MD
344{
345 int error;
346
6a37e7e4 347 hammer_cursor_downgrade(cursor);
195c19a1
MD
348
349 /*
4e17f465
MD
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.
8cd0a023
MD
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
f36a9737 366 error = hammer_load_cursor_parent(cursor, 0);
8cd0a023
MD
367 return(error);
368}
369
370/*
f36a9737 371 * Special cursor up given a locked cursor. The orignal node is not
b3bad96f
MD
372 * unlocked or released and the cursor is not downgraded.
373 *
f3a4893b
MD
374 * This function can fail with EDEADLK.
375 *
376 * This function is only run when recursively deleting parent nodes
377 * to get rid of an empty leaf.
f36a9737
MD
378 */
379int
380hammer_cursor_up_locked(hammer_cursor_t cursor)
381{
382 hammer_node_t save;
383 int error;
f3a4893b 384 int save_index;
f36a9737
MD
385
386 /*
387 * If the parent is NULL we are at the root of the B-Tree and
388 * return ENOENT.
389 */
390 if (cursor->parent == NULL)
391 return (ENOENT);
392
393 save = cursor->node;
f3a4893b 394 save_index = cursor->index;
f36a9737
MD
395
396 /*
397 * Set the node to its parent.
398 */
399 cursor->node = cursor->parent;
400 cursor->index = cursor->parent_index;
401 cursor->parent = NULL;
402 cursor->parent_index = 0;
403
404 /*
405 * load the new parent, attempt to exclusively lock it. Note that
406 * we are still holding the old parent (now cursor->node) exclusively
f3a4893b
MD
407 * locked.
408 *
409 * This can return EDEADLK. Undo the operation on any error. These
410 * up sequences can occur during iterations so be sure to restore
411 * the index.
f36a9737
MD
412 */
413 error = hammer_load_cursor_parent(cursor, 1);
414 if (error) {
415 cursor->parent = cursor->node;
416 cursor->parent_index = cursor->index;
417 cursor->node = save;
f3a4893b 418 cursor->index = save_index;
f36a9737
MD
419 }
420 return(error);
421}
422
423
424/*
8cd0a023
MD
425 * Cursor down through the current node, which must be an internal node.
426 *
427 * This routine adjusts the cursor and sets index to 0.
428 */
429int
430hammer_cursor_down(hammer_cursor_t cursor)
431{
432 hammer_node_t node;
433 hammer_btree_elm_t elm;
8cd0a023
MD
434 int error;
435
436 /*
437 * The current node becomes the current parent
438 */
6a37e7e4 439 hammer_cursor_downgrade(cursor);
8cd0a023 440 node = cursor->node;
8cd0a023
MD
441 KKASSERT(cursor->index >= 0 && cursor->index < node->ondisk->count);
442 if (cursor->parent) {
443 hammer_unlock(&cursor->parent->lock);
444 hammer_rel_node(cursor->parent);
445 }
446 cursor->parent = node;
447 cursor->parent_index = cursor->index;
448 cursor->node = NULL;
449 cursor->index = 0;
450
451 /*
76376933 452 * Extract element to push into at (node,index), set bounds.
8cd0a023
MD
453 */
454 elm = &node->ondisk->elms[cursor->parent_index];
455
456 /*
fe7678ee
MD
457 * Ok, push down into elm. If elm specifies an internal or leaf
458 * node the current node must be an internal node. If elm specifies
459 * a spike then the current node must be a leaf node.
8cd0a023 460 */
fe7678ee
MD
461 switch(elm->base.btype) {
462 case HAMMER_BTREE_TYPE_INTERNAL:
463 case HAMMER_BTREE_TYPE_LEAF:
464 KKASSERT(node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
465 KKASSERT(elm->internal.subtree_offset != 0);
466 cursor->left_bound = &elm[0].internal.base;
467 cursor->right_bound = &elm[1].internal.base;
82010f9f 468 node = hammer_get_node(cursor->trans,
19619882 469 elm->internal.subtree_offset, 0, &error);
fe7678ee 470 if (error == 0) {
40043e7f 471 KASSERT(elm->base.btype == node->ondisk->type, ("BTYPE MISMATCH %c %c NODE %p\n", elm->base.btype, node->ondisk->type, node));
b33e2cc0 472 if (node->ondisk->parent != cursor->parent->node_offset)
973c11b9 473 panic("node %p %016llx vs %016llx\n", node, (long long)node->ondisk->parent, (long long)cursor->parent->node_offset);
fe7678ee
MD
474 KKASSERT(node->ondisk->parent == cursor->parent->node_offset);
475 }
476 break;
fe7678ee
MD
477 default:
478 panic("hammer_cursor_down: illegal btype %02x (%c)\n",
479 elm->base.btype,
480 (elm->base.btype ? elm->base.btype : '?'));
481 break;
8cd0a023
MD
482 }
483 if (error == 0) {
6a37e7e4 484 hammer_lock_sh(&node->lock);
f36a9737 485 KKASSERT ((node->flags & HAMMER_NODE_DELETED) == 0);
8cd0a023
MD
486 cursor->node = node;
487 cursor->index = 0;
488 }
489 return(error);
490}
491
b3bad96f
MD
492/************************************************************************
493 * DEADLOCK RECOVERY *
494 ************************************************************************
495 *
496 * These are the new deadlock recovery functions. Currently they are only
497 * used for the mirror propagation and physical node removal cases but
498 * ultimately the intention is to use them for all deadlock recovery
499 * operations.
c9ce54d6
MD
500 *
501 * WARNING! The contents of the cursor may be modified while unlocked.
502 * passive modifications including adjusting the node, parent,
503 * indexes, and leaf pointer.
504 *
505 * An outright removal of the element the cursor was pointing at
506 * will cause the HAMMER_CURSOR_TRACKED_RIPOUT flag to be set,
507 * which chains to causing the HAMMER_CURSOR_RETEST to be set
508 * when the cursor is locked again.
b3bad96f 509 */
adf01747 510void
982be4bf 511hammer_unlock_cursor(hammer_cursor_t cursor)
b3bad96f 512{
b3bad96f 513 hammer_node_t node;
b3bad96f 514
adf01747 515 KKASSERT((cursor->flags & HAMMER_CURSOR_TRACKED) == 0);
b3bad96f 516 KKASSERT(cursor->node);
19168d41
MD
517
518 /*
b3bad96f
MD
519 * Release the cursor's locks and track B-Tree operations on node.
520 * While being tracked our cursor can be modified by other threads
5c8d05e2 521 * and the node may be replaced.
b3bad96f
MD
522 */
523 if (cursor->parent) {
524 hammer_unlock(&cursor->parent->lock);
525 hammer_rel_node(cursor->parent);
526 cursor->parent = NULL;
527 }
528 node = cursor->node;
adf01747 529 cursor->flags |= HAMMER_CURSOR_TRACKED;
b3bad96f 530 TAILQ_INSERT_TAIL(&node->cursor_list, cursor, deadlk_entry);
b3bad96f 531 hammer_unlock(&node->lock);
adf01747
MD
532}
533
534/*
535 * Get the cursor heated up again. The cursor's node may have
536 * changed and we might have to locate the new parent.
537 *
538 * If the exact element we were on got deleted RIPOUT will be
539 * set and we must clear ATEDISK so an iteration does not skip
540 * the element after it.
541 */
542int
982be4bf 543hammer_lock_cursor(hammer_cursor_t cursor)
adf01747 544{
adf01747
MD
545 hammer_node_t node;
546 int error;
547
548 KKASSERT(cursor->flags & HAMMER_CURSOR_TRACKED);
549
550 /*
adf01747
MD
551 * Relock the node
552 */
553 for (;;) {
554 node = cursor->node;
555 hammer_ref_node(node);
556 hammer_lock_sh(&node->lock);
557 if (cursor->node == node) {
558 hammer_rel_node(node);
559 break;
560 }
561 hammer_unlock(&node->lock);
562 hammer_rel_node(node);
563 }
564
565 /*
566 * Untrack the cursor, clean up, and re-establish the parent node.
567 */
568 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry);
569 cursor->flags &= ~HAMMER_CURSOR_TRACKED;
570
5c8d05e2
MD
571 /*
572 * If a ripout has occured iterations must re-test the (new)
573 * current element. Clearing ATEDISK prevents the element from
574 * being skipped and RETEST causes it to be re-tested.
575 */
adf01747
MD
576 if (cursor->flags & HAMMER_CURSOR_TRACKED_RIPOUT) {
577 cursor->flags &= ~HAMMER_CURSOR_TRACKED_RIPOUT;
578 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
5c8d05e2 579 cursor->flags |= HAMMER_CURSOR_RETEST;
adf01747
MD
580 }
581 error = hammer_load_cursor_parent(cursor, 0);
582 return(error);
583}
584
585/*
586 * Recover from a deadlocked cursor, tracking any node removals or
587 * replacements. If the cursor's current node is removed by another
588 * thread (via btree_remove()) the cursor will be seeked upwards.
98da6d8c
MD
589 *
590 * The caller is working a modifying operation and must be holding the
591 * sync lock (shared). We do not release the sync lock because this
592 * would break atomicy.
adf01747
MD
593 */
594int
595hammer_recover_cursor(hammer_cursor_t cursor)
596{
597 int error;
598
982be4bf 599 hammer_unlock_cursor(cursor);
98da6d8c 600 KKASSERT(cursor->trans->sync_lock_refs > 0);
b3bad96f
MD
601
602 /*
603 * Wait for the deadlock to clear
604 */
605 if (cursor->deadlk_node) {
606 hammer_lock_ex_ident(&cursor->deadlk_node->lock, "hmrdlk");
607 hammer_unlock(&cursor->deadlk_node->lock);
608 hammer_rel_node(cursor->deadlk_node);
609 cursor->deadlk_node = NULL;
610 }
611 if (cursor->deadlk_rec) {
612 hammer_wait_mem_record_ident(cursor->deadlk_rec, "hmmdlr");
613 hammer_rel_mem_record(cursor->deadlk_rec);
614 cursor->deadlk_rec = NULL;
615 }
982be4bf 616 error = hammer_lock_cursor(cursor);
adf01747
MD
617 return(error);
618}
619
620/*
621 * Dup ocursor to ncursor. ncursor inherits ocursor's locks and ocursor
622 * is effectively unlocked and becomes tracked. If ocursor was not locked
623 * then ncursor also inherits the tracking.
624 *
625 * After the caller finishes working with ncursor it must be cleaned up
626 * with hammer_done_cursor(), and the caller must re-lock ocursor.
627 */
3f43fb33
MD
628hammer_cursor_t
629hammer_push_cursor(hammer_cursor_t ocursor)
adf01747 630{
3f43fb33 631 hammer_cursor_t ncursor;
adf01747
MD
632 hammer_inode_t ip;
633 hammer_node_t node;
bac808fe 634 hammer_mount_t hmp;
adf01747 635
bac808fe
MD
636 hmp = ocursor->trans->hmp;
637 ncursor = kmalloc(sizeof(*ncursor), hmp->m_misc, M_WAITOK | M_ZERO);
adf01747
MD
638 bcopy(ocursor, ncursor, sizeof(*ocursor));
639
640 node = ocursor->node;
641 hammer_ref_node(node);
642 if ((ocursor->flags & HAMMER_CURSOR_TRACKED) == 0) {
643 ocursor->flags |= HAMMER_CURSOR_TRACKED;
644 TAILQ_INSERT_TAIL(&node->cursor_list, ocursor, deadlk_entry);
b3bad96f 645 }
adf01747
MD
646 if (ncursor->parent)
647 ocursor->parent = NULL;
648 ocursor->data_buffer = NULL;
649 ocursor->leaf = NULL;
650 ocursor->data = NULL;
651 if (ncursor->flags & HAMMER_CURSOR_TRACKED)
652 TAILQ_INSERT_TAIL(&node->cursor_list, ncursor, deadlk_entry);
653 if ((ip = ncursor->ip) != NULL) {
654 ++ip->cursor_ip_refs;
b3bad96f 655 }
adf01747
MD
656 if (ncursor->iprec)
657 hammer_ref(&ncursor->iprec->lock);
3f43fb33
MD
658 return(ncursor);
659}
660
661/*
662 * Destroy ncursor and restore ocursor
663 *
664 * This is a temporary hack for the release. We can't afford to lose
665 * the IP lock until the IP object scan code is able to deal with it,
666 * so have ocursor inherit it back.
667 */
668void
669hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor)
670{
bac808fe 671 hammer_mount_t hmp;
3f43fb33
MD
672 hammer_inode_t ip;
673
bac808fe 674 hmp = ncursor->trans->hmp;
3f43fb33
MD
675 ip = ncursor->ip;
676 ncursor->ip = NULL;
677 if (ip)
678 --ip->cursor_ip_refs;
679 hammer_done_cursor(ncursor);
bac808fe 680 kfree(ncursor, hmp->m_misc);
3f43fb33 681 KKASSERT(ocursor->ip == ip);
982be4bf 682 hammer_lock_cursor(ocursor);
b3bad96f
MD
683}
684
685/*
686 * onode is being replaced by nnode by the reblocking code.
687 */
688void
689hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode)
690{
691 hammer_cursor_t cursor;
c9ce54d6
MD
692 hammer_node_ondisk_t ondisk;
693 hammer_node_ondisk_t nndisk;
694
695 ondisk = onode->ondisk;
696 nndisk = nnode->ondisk;
b3bad96f
MD
697
698 while ((cursor = TAILQ_FIRST(&onode->cursor_list)) != NULL) {
699 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry);
700 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry);
701 KKASSERT(cursor->node == onode);
c9ce54d6
MD
702 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
703 cursor->leaf = &nndisk->elms[cursor->index].leaf;
b3bad96f
MD
704 cursor->node = nnode;
705 hammer_ref_node(nnode);
706 hammer_rel_node(onode);
707 }
708}
709
710/*
f3a4893b
MD
711 * We have removed <node> from the parent and collapsed the parent.
712 *
713 * Cursors in deadlock recovery are seeked upward to the parent so the
714 * btree_remove() recursion works properly.
b3bad96f
MD
715 */
716void
717hammer_cursor_removed_node(hammer_node_t node, hammer_node_t parent, int index)
718{
719 hammer_cursor_t cursor;
c9ce54d6 720 hammer_node_ondisk_t ondisk;
b3bad96f
MD
721
722 KKASSERT(parent != NULL);
c9ce54d6
MD
723 ondisk = node->ondisk;
724
b3bad96f
MD
725 while ((cursor = TAILQ_FIRST(&node->cursor_list)) != NULL) {
726 KKASSERT(cursor->node == node);
727 KKASSERT(cursor->index == 0);
728 TAILQ_REMOVE(&node->cursor_list, cursor, deadlk_entry);
729 TAILQ_INSERT_TAIL(&parent->cursor_list, cursor, deadlk_entry);
c9ce54d6
MD
730 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
731 cursor->leaf = NULL;
adf01747 732 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT;
b3bad96f
MD
733 cursor->node = parent;
734 cursor->index = index;
735 hammer_ref_node(parent);
736 hammer_rel_node(node);
737 }
738}
739
740/*
741 * node was split at (onode, index) with elements >= index moved to nnode.
742 */
743void
744hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode, int index)
745{
746 hammer_cursor_t cursor;
c9ce54d6
MD
747 hammer_node_ondisk_t ondisk;
748 hammer_node_ondisk_t nndisk;
749
750 ondisk = onode->ondisk;
751 nndisk = nnode->ondisk;
b3bad96f
MD
752
753again:
754 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) {
755 KKASSERT(cursor->node == onode);
756 if (cursor->index < index)
757 continue;
758 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry);
759 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry);
c9ce54d6
MD
760 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
761 cursor->leaf = &nndisk->elms[cursor->index - index].leaf;
b3bad96f
MD
762 cursor->node = nnode;
763 cursor->index -= index;
764 hammer_ref_node(nnode);
765 hammer_rel_node(onode);
766 goto again;
767 }
768}
769
770/*
1775b6a0
MD
771 * An element was moved from one node to another or within a node. The
772 * index may also represent the end of the node (index == numelements).
773 *
bbb01e14
MD
774 * {oparent,pindex} is the parent node's pointer to onode/oindex.
775 *
1775b6a0
MD
776 * This is used by the rebalancing code. This is not an insertion or
777 * deletion and any additional elements, including the degenerate case at
778 * the end of the node, will be dealt with by additional distinct calls.
779 */
780void
bbb01e14
MD
781hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
782 hammer_node_t onode, int oindex,
783 hammer_node_t nnode, int nindex)
1775b6a0
MD
784{
785 hammer_cursor_t cursor;
c9ce54d6
MD
786 hammer_node_ondisk_t ondisk;
787 hammer_node_ondisk_t nndisk;
788
bbb01e14
MD
789 /*
790 * Adjust any cursors pointing at the element
791 */
c9ce54d6
MD
792 ondisk = onode->ondisk;
793 nndisk = nnode->ondisk;
bbb01e14 794again1:
1775b6a0
MD
795 TAILQ_FOREACH(cursor, &onode->cursor_list, deadlk_entry) {
796 KKASSERT(cursor->node == onode);
797 if (cursor->index != oindex)
798 continue;
799 TAILQ_REMOVE(&onode->cursor_list, cursor, deadlk_entry);
800 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry);
c9ce54d6
MD
801 if (cursor->leaf == &ondisk->elms[oindex].leaf)
802 cursor->leaf = &nndisk->elms[nindex].leaf;
1775b6a0
MD
803 cursor->node = nnode;
804 cursor->index = nindex;
805 hammer_ref_node(nnode);
806 hammer_rel_node(onode);
bbb01e14
MD
807 goto again1;
808 }
809
810 /*
811 * When moving the first element of onode to a different node any
812 * cursor which is pointing at (oparent,pindex) must be repointed
813 * to nnode and ATEDISK must be cleared.
814 *
815 * This prevents cursors from losing track due to insertions.
816 * Insertions temporarily release the cursor in order to update
817 * the mirror_tids. It primarily effects the mirror_write code.
818 * The other code paths generally only do a single insertion and
819 * then relookup or drop the cursor.
820 */
821 if (onode == nnode || oindex)
822 return;
823 ondisk = oparent->ondisk;
824again2:
825 TAILQ_FOREACH(cursor, &oparent->cursor_list, deadlk_entry) {
826 KKASSERT(cursor->node == oparent);
827 if (cursor->index != pindex)
828 continue;
829 kprintf("HAMMER debug: shifted cursor pointing at parent\n"
830 "parent %016jx:%d onode %016jx:%d nnode %016jx:%d\n",
831 (intmax_t)oparent->node_offset, pindex,
832 (intmax_t)onode->node_offset, oindex,
833 (intmax_t)nnode->node_offset, nindex);
834 print_backtrace();
835 TAILQ_REMOVE(&oparent->cursor_list, cursor, deadlk_entry);
836 TAILQ_INSERT_TAIL(&nnode->cursor_list, cursor, deadlk_entry);
837 if (cursor->leaf == &ondisk->elms[oindex].leaf)
838 cursor->leaf = &nndisk->elms[nindex].leaf;
839 cursor->node = nnode;
840 cursor->index = nindex;
841 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
842 hammer_ref_node(nnode);
843 hammer_rel_node(oparent);
844 goto again2;
1775b6a0
MD
845 }
846}
847
848/*
849 * The B-Tree element pointing to the specified node was moved from (oparent)
850 * to (nparent, nindex). We must locate any tracked cursors pointing at
851 * node and adjust their parent accordingly.
852 *
853 * This is used by the rebalancing code when packing elements causes an
854 * element to shift from one node to another.
855 */
856void
857hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
858 hammer_node_t nparent, int nindex)
859{
860 hammer_cursor_t cursor;
861
862again:
863 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) {
864 KKASSERT(cursor->node == node);
865 if (cursor->parent == oparent) {
866 cursor->parent = nparent;
867 cursor->parent_index = nindex;
868 hammer_ref_node(nparent);
869 hammer_rel_node(oparent);
870 goto again;
871 }
872 }
873}
874
875/*
e4a5ff06 876 * Deleted element at (node, index)
b3bad96f
MD
877 *
878 * Shift indexes >= index
879 */
880void
881hammer_cursor_deleted_element(hammer_node_t node, int index)
882{
883 hammer_cursor_t cursor;
c9ce54d6
MD
884 hammer_node_ondisk_t ondisk;
885
886 ondisk = node->ondisk;
b3bad96f
MD
887
888 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) {
889 KKASSERT(cursor->node == node);
890 if (cursor->index == index) {
adf01747 891 cursor->flags |= HAMMER_CURSOR_TRACKED_RIPOUT;
c9ce54d6
MD
892 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
893 cursor->leaf = NULL;
b3bad96f 894 } else if (cursor->index > index) {
c9ce54d6
MD
895 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
896 cursor->leaf = &ondisk->elms[cursor->index - 1].leaf;
b3bad96f
MD
897 --cursor->index;
898 }
899 }
900}
901
902/*
e4a5ff06 903 * Inserted element at (node, index)
b3bad96f
MD
904 *
905 * Shift indexes >= index
906 */
907void
908hammer_cursor_inserted_element(hammer_node_t node, int index)
909{
910 hammer_cursor_t cursor;
c9ce54d6
MD
911 hammer_node_ondisk_t ondisk;
912
913 ondisk = node->ondisk;
b3bad96f
MD
914
915 TAILQ_FOREACH(cursor, &node->cursor_list, deadlk_entry) {
916 KKASSERT(cursor->node == node);
c9ce54d6
MD
917 if (cursor->index >= index) {
918 if (cursor->leaf == &ondisk->elms[cursor->index].leaf)
919 cursor->leaf = &ondisk->elms[cursor->index + 1].leaf;
b3bad96f 920 ++cursor->index;
c9ce54d6 921 }
b3bad96f
MD
922 }
923}
924