2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
34 * $DragonFly: src/sys/vfs/hammer/hammer_cursor.c,v 1.37 2008/07/05 18:59:27 dillon Exp $
38 * HAMMER B-Tree index - cursor support routines
42 static int hammer_load_cursor_parent(hammer_cursor_t cursor
, int try_exclusive
);
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.
49 hammer_init_cursor(hammer_transaction_t trans
, hammer_cursor_t cursor
,
50 hammer_node_cache_t cache
, hammer_inode_t ip
)
52 hammer_volume_t volume
;
56 bzero(cursor
, sizeof(*cursor
));
58 cursor
->trans
= trans
;
61 * If the cursor operation is on behalf of an inode, lock
64 if ((cursor
->ip
= ip
) != NULL
) {
66 if (trans
->type
== HAMMER_TRANS_FLS
)
67 hammer_lock_ex(&ip
->lock
);
69 hammer_lock_sh(&ip
->lock
);
73 * Step 1 - acquire a locked node from the cache if possible
75 if (cache
&& cache
->node
) {
76 node
= hammer_ref_node_safe(trans
->hmp
, cache
, &error
);
78 hammer_lock_sh_lowpri(&node
->lock
);
79 if (node
->flags
& HAMMER_NODE_DELETED
) {
80 hammer_unlock(&node
->lock
);
81 hammer_rel_node(node
);
90 * Step 2 - If we couldn't get a node from the cache, get
91 * the one from the root of the filesystem.
93 while (node
== NULL
) {
94 volume
= hammer_get_root_volume(trans
->hmp
, &error
);
97 node
= hammer_get_node(trans
->hmp
,
98 volume
->ondisk
->vol0_btree_root
,
100 hammer_rel_volume(volume
, 0);
103 hammer_lock_sh_lowpri(&node
->lock
);
106 * If someone got in before we could lock the node, retry.
108 if (node
->flags
& HAMMER_NODE_DELETED
) {
109 hammer_unlock(&node
->lock
);
110 hammer_rel_node(node
);
114 if (volume
->ondisk
->vol0_btree_root
!= node
->node_offset
) {
115 hammer_unlock(&node
->lock
);
116 hammer_rel_node(node
);
123 * Step 3 - finish initializing the cursor by acquiring the parent
127 error
= hammer_load_cursor_parent(cursor
, 0);
128 KKASSERT(error
== 0);
129 /* if (error) hammer_done_cursor(cursor); */
134 * Normalize a cursor. Sometimes cursors can be left in a state
135 * where node is NULL. If the cursor is in this state, cursor up.
138 hammer_normalize_cursor(hammer_cursor_t cursor
)
140 if (cursor
->node
== NULL
) {
141 KKASSERT(cursor
->parent
!= NULL
);
142 hammer_cursor_up(cursor
);
148 * We are finished with a cursor. We NULL out various fields as sanity
149 * check, in case the structure is inappropriately used afterwords.
152 hammer_done_cursor(hammer_cursor_t cursor
)
156 if (cursor
->parent
) {
157 hammer_unlock(&cursor
->parent
->lock
);
158 hammer_rel_node(cursor
->parent
);
159 cursor
->parent
= NULL
;
162 hammer_unlock(&cursor
->node
->lock
);
163 hammer_rel_node(cursor
->node
);
166 if (cursor
->data_buffer
) {
167 hammer_rel_buffer(cursor
->data_buffer
, 0);
168 cursor
->data_buffer
= NULL
;
170 if (cursor
->record_buffer
) {
171 hammer_rel_buffer(cursor
->record_buffer
, 0);
172 cursor
->record_buffer
= NULL
;
174 if ((ip
= cursor
->ip
) != NULL
) {
176 hammer_rel_mem_record(cursor
->iprec
);
177 cursor
->iprec
= NULL
;
179 KKASSERT(ip
->cursor_ip_refs
> 0);
180 --ip
->cursor_ip_refs
;
181 hammer_unlock(&ip
->lock
);
186 * If we deadlocked this node will be referenced. Do a quick
187 * lock/unlock to wait for the deadlock condition to clear.
189 if (cursor
->deadlk_node
) {
190 hammer_lock_ex_ident(&cursor
->deadlk_node
->lock
, "hmrdlk");
191 hammer_unlock(&cursor
->deadlk_node
->lock
);
192 hammer_rel_node(cursor
->deadlk_node
);
193 cursor
->deadlk_node
= NULL
;
195 if (cursor
->deadlk_rec
) {
196 hammer_wait_mem_record_ident(cursor
->deadlk_rec
, "hmmdlr");
197 hammer_rel_mem_record(cursor
->deadlk_rec
);
198 cursor
->deadlk_rec
= NULL
;
203 cursor
->left_bound
= NULL
;
204 cursor
->right_bound
= NULL
;
205 cursor
->trans
= NULL
;
209 * Upgrade cursor->node and cursor->parent to exclusive locks. This
210 * function can return EDEADLK.
212 * The lock must already be either held shared or already held exclusively
215 * If we fail to upgrade the lock and cursor->deadlk_node is NULL,
216 * we add another reference to the node that failed and set
217 * cursor->deadlk_node so hammer_done_cursor() can block on it.
220 hammer_cursor_upgrade(hammer_cursor_t cursor
)
224 error
= hammer_lock_upgrade(&cursor
->node
->lock
);
225 if (error
&& cursor
->deadlk_node
== NULL
) {
226 cursor
->deadlk_node
= cursor
->node
;
227 hammer_ref_node(cursor
->deadlk_node
);
228 } else if (error
== 0 && cursor
->parent
) {
229 error
= hammer_lock_upgrade(&cursor
->parent
->lock
);
230 if (error
&& cursor
->deadlk_node
== NULL
) {
231 cursor
->deadlk_node
= cursor
->parent
;
232 hammer_ref_node(cursor
->deadlk_node
);
239 hammer_cursor_upgrade_node(hammer_cursor_t cursor
)
243 error
= hammer_lock_upgrade(&cursor
->node
->lock
);
244 if (error
&& cursor
->deadlk_node
== NULL
) {
245 cursor
->deadlk_node
= cursor
->node
;
246 hammer_ref_node(cursor
->deadlk_node
);
252 * Downgrade cursor->node and cursor->parent to shared locks. This
253 * function can return EDEADLK.
256 hammer_cursor_downgrade(hammer_cursor_t cursor
)
258 if (hammer_lock_excl_owned(&cursor
->node
->lock
, curthread
))
259 hammer_lock_downgrade(&cursor
->node
->lock
);
260 if (cursor
->parent
&&
261 hammer_lock_excl_owned(&cursor
->parent
->lock
, curthread
)) {
262 hammer_lock_downgrade(&cursor
->parent
->lock
);
267 * Seek the cursor to the specified node and index.
269 * The caller must ref the node prior to calling this routine and release
270 * it after it returns. If the seek succeeds the cursor will gain its own
274 hammer_cursor_seek(hammer_cursor_t cursor
, hammer_node_t node
, int index
)
278 hammer_cursor_downgrade(cursor
);
281 if (cursor
->node
!= node
) {
282 hammer_unlock(&cursor
->node
->lock
);
283 hammer_rel_node(cursor
->node
);
285 hammer_ref_node(node
);
286 hammer_lock_sh(&node
->lock
);
287 KKASSERT ((node
->flags
& HAMMER_NODE_DELETED
) == 0);
289 if (cursor
->parent
) {
290 hammer_unlock(&cursor
->parent
->lock
);
291 hammer_rel_node(cursor
->parent
);
292 cursor
->parent
= NULL
;
293 cursor
->parent_index
= 0;
295 error
= hammer_load_cursor_parent(cursor
, 0);
297 cursor
->index
= index
;
302 * Load the parent of cursor->node into cursor->parent.
306 hammer_load_cursor_parent(hammer_cursor_t cursor
, int try_exclusive
)
309 hammer_node_t parent
;
311 hammer_btree_elm_t elm
;
315 hmp
= cursor
->trans
->hmp
;
317 if (cursor
->node
->ondisk
->parent
) {
319 parent
= hammer_btree_get_parent(node
, &parent_index
,
320 &error
, try_exclusive
);
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
;
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
;
339 * Cursor up to our parent node. Return ENOENT if we are at the root of
343 hammer_cursor_up(hammer_cursor_t cursor
)
347 hammer_cursor_downgrade(cursor
);
350 * If the parent is NULL we are at the root of the B-Tree and
353 if (cursor
->parent
== NULL
)
357 * Set the node to its parent.
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;
366 error
= hammer_load_cursor_parent(cursor
, 0);
371 * Special cursor up given a locked cursor. The orignal node is not
372 * unlocked or released and the cursor is not downgraded.
374 * This function will recover from deadlocks. EDEADLK cannot be returned.
377 hammer_cursor_up_locked(hammer_cursor_t cursor
)
383 * If the parent is NULL we are at the root of the B-Tree and
386 if (cursor
->parent
== NULL
)
392 * Set the node to its parent.
394 cursor
->node
= cursor
->parent
;
395 cursor
->index
= cursor
->parent_index
;
396 cursor
->parent
= NULL
;
397 cursor
->parent_index
= 0;
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.
404 error
= hammer_load_cursor_parent(cursor
, 1);
406 cursor
->parent
= cursor
->node
;
407 cursor
->parent_index
= cursor
->index
;
416 * Cursor down through the current node, which must be an internal node.
418 * This routine adjusts the cursor and sets index to 0.
421 hammer_cursor_down(hammer_cursor_t cursor
)
424 hammer_btree_elm_t elm
;
428 * The current node becomes the current parent
430 hammer_cursor_downgrade(cursor
);
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
);
437 cursor
->parent
= node
;
438 cursor
->parent_index
= cursor
->index
;
443 * Extract element to push into at (node,index), set bounds.
445 elm
= &node
->ondisk
->elms
[cursor
->parent_index
];
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.
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
->hmp
,
460 elm
->internal
.subtree_offset
, 0, &error
);
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
, node
->ondisk
->parent
, cursor
->parent
->node_offset
);
465 KKASSERT(node
->ondisk
->parent
== cursor
->parent
->node_offset
);
469 panic("hammer_cursor_down: illegal btype %02x (%c)\n",
471 (elm
->base
.btype
? elm
->base
.btype
: '?'));
475 hammer_lock_sh(&node
->lock
);
476 KKASSERT ((node
->flags
& HAMMER_NODE_DELETED
) == 0);
483 /************************************************************************
484 * DEADLOCK RECOVERY *
485 ************************************************************************
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
494 * Recover from a deadlocked cursor, tracking any node removals or
495 * replacements. If the cursor's current node is removed by another
496 * thread (via btree_remove()) the cursor will be seeked upwards.
499 hammer_recover_cursor(hammer_cursor_t cursor
)
507 KKASSERT((cursor
->flags
& HAMMER_CURSOR_DEADLK_RECOVER
) == 0);
508 KKASSERT(cursor
->node
);
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 node may be replaced.
514 if (cursor
->parent
) {
515 hammer_unlock(&cursor
->parent
->lock
);
516 hammer_rel_node(cursor
->parent
);
517 cursor
->parent
= NULL
;
520 cursor
->flags
|= HAMMER_CURSOR_DEADLK_RECOVER
;
521 TAILQ_INSERT_TAIL(&node
->cursor_list
, cursor
, deadlk_entry
);
522 status
= hammer_lock_status(&node
->lock
);
523 hammer_unlock(&node
->lock
);
525 if ((ip
= cursor
->ip
) != NULL
)
526 hammer_unlock(&ip
->lock
);
529 * Wait for the deadlock to clear
531 if (cursor
->deadlk_node
) {
532 hammer_lock_ex_ident(&cursor
->deadlk_node
->lock
, "hmrdlk");
533 hammer_unlock(&cursor
->deadlk_node
->lock
);
534 hammer_rel_node(cursor
->deadlk_node
);
535 cursor
->deadlk_node
= NULL
;
537 if (cursor
->deadlk_rec
) {
538 hammer_wait_mem_record_ident(cursor
->deadlk_rec
, "hmmdlr");
539 hammer_rel_mem_record(cursor
->deadlk_rec
);
540 cursor
->deadlk_rec
= NULL
;
544 * Get the cursor heated up again. The cursor's node may have
545 * changed and we might have to locate the new parent.
547 * If the exact element we were on got deleted RIPOUT will be
548 * set and we must clear ATEDISK so an iteration does not skip
549 * the element after it.
551 KKASSERT(cursor
->flags
& HAMMER_CURSOR_DEADLK_RECOVER
);
553 TAILQ_REMOVE(&node
->cursor_list
, cursor
, deadlk_entry
);
554 cursor
->flags
&= ~HAMMER_CURSOR_DEADLK_RECOVER
;
555 if (cursor
->flags
& HAMMER_CURSOR_DEADLK_RIPOUT
) {
556 cursor
->flags
&= ~HAMMER_CURSOR_DEADLK_RIPOUT
;
557 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
559 hammer_lock_sh(&node
->lock
);
560 error
= hammer_load_cursor_parent(cursor
, 0);
563 error
= hammer_cursor_upgrade(cursor
);
564 if (error
== EDEADLK
) {
574 * onode is being replaced by nnode by the reblocking code.
577 hammer_cursor_replaced_node(hammer_node_t onode
, hammer_node_t nnode
)
579 hammer_cursor_t cursor
;
581 while ((cursor
= TAILQ_FIRST(&onode
->cursor_list
)) != NULL
) {
582 TAILQ_REMOVE(&onode
->cursor_list
, cursor
, deadlk_entry
);
583 TAILQ_INSERT_TAIL(&nnode
->cursor_list
, cursor
, deadlk_entry
);
584 KKASSERT(cursor
->node
== onode
);
585 cursor
->node
= nnode
;
586 hammer_ref_node(nnode
);
587 hammer_rel_node(onode
);
592 * node is being removed, cursors in deadlock recovery are seeked upward
596 hammer_cursor_removed_node(hammer_node_t node
, hammer_node_t parent
, int index
)
598 hammer_cursor_t cursor
;
600 KKASSERT(parent
!= NULL
);
601 while ((cursor
= TAILQ_FIRST(&node
->cursor_list
)) != NULL
) {
602 KKASSERT(cursor
->node
== node
);
603 KKASSERT(cursor
->index
== 0);
604 TAILQ_REMOVE(&node
->cursor_list
, cursor
, deadlk_entry
);
605 TAILQ_INSERT_TAIL(&parent
->cursor_list
, cursor
, deadlk_entry
);
606 cursor
->flags
|= HAMMER_CURSOR_DEADLK_RIPOUT
;
607 cursor
->node
= parent
;
608 cursor
->index
= index
;
609 hammer_ref_node(parent
);
610 hammer_rel_node(node
);
615 * node was split at (onode, index) with elements >= index moved to nnode.
618 hammer_cursor_split_node(hammer_node_t onode
, hammer_node_t nnode
, int index
)
620 hammer_cursor_t cursor
;
623 TAILQ_FOREACH(cursor
, &onode
->cursor_list
, deadlk_entry
) {
624 KKASSERT(cursor
->node
== onode
);
625 if (cursor
->index
< index
)
627 TAILQ_REMOVE(&onode
->cursor_list
, cursor
, deadlk_entry
);
628 TAILQ_INSERT_TAIL(&nnode
->cursor_list
, cursor
, deadlk_entry
);
629 cursor
->node
= nnode
;
630 cursor
->index
-= index
;
631 hammer_ref_node(nnode
);
632 hammer_rel_node(onode
);
638 * Inserted element at (node, index)
640 * Shift indexes >= index
643 hammer_cursor_deleted_element(hammer_node_t node
, int index
)
645 hammer_cursor_t cursor
;
647 TAILQ_FOREACH(cursor
, &node
->cursor_list
, deadlk_entry
) {
648 KKASSERT(cursor
->node
== node
);
649 if (cursor
->index
== index
) {
650 cursor
->flags
|= HAMMER_CURSOR_DEADLK_RIPOUT
;
651 } else if (cursor
->index
> index
) {
658 * Deleted element at (node, index)
660 * Shift indexes >= index
663 hammer_cursor_inserted_element(hammer_node_t node
, int index
)
665 hammer_cursor_t cursor
;
667 TAILQ_FOREACH(cursor
, &node
->cursor_list
, deadlk_entry
) {
668 KKASSERT(cursor
->node
== node
);
669 if (cursor
->index
>= index
)