2 * Copyright (c) 2009 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
37 static int rebalance_node(struct hammer_ioc_rebalance *rebal,
38 hammer_cursor_t cursor, hammer_node_lock_t lcache);
39 static void rebalance_closeout(hammer_node_lock_t base_item, int base_count,
40 hammer_btree_elm_t elm);
41 static void rebalance_parent_ptrs(hammer_node_lock_t base_item, int index,
42 hammer_node_lock_t item, hammer_node_lock_t chld_item);
45 * Iterate through the specified range of object ids and rebalance B-Tree
46 * leaf and internal nodes we encounter. A forwards iteration is used.
48 * All leafs are at the same depth. We use the b-tree scan code loosely
49 * to position ourselves and create degenerate cases to skip indices
50 * that we have rebalanced in bulk.
54 hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
55 struct hammer_ioc_rebalance *rebal)
57 struct hammer_cursor cursor;
58 struct hammer_node_lock lcache;
59 hammer_btree_leaf_elm_t elm;
63 if ((rebal->key_beg.localization | rebal->key_end.localization) &
64 HAMMER_LOCALIZE_PSEUDOFS_MASK) {
67 if (rebal->key_beg.localization > rebal->key_end.localization)
69 if (rebal->key_beg.localization == rebal->key_end.localization) {
70 if (rebal->key_beg.obj_id > rebal->key_end.obj_id)
72 /* key-space limitations - no check needed */
74 if (rebal->saturation < HAMMER_BTREE_INT_ELMS / 2)
75 rebal->saturation = HAMMER_BTREE_INT_ELMS / 2;
76 if (rebal->saturation > HAMMER_BTREE_INT_ELMS)
77 rebal->saturation = HAMMER_BTREE_INT_ELMS;
79 rebal->key_cur = rebal->key_beg;
80 rebal->key_cur.localization &= HAMMER_LOCALIZE_MASK;
81 rebal->key_cur.localization += ip->obj_localization;
83 hammer_btree_lcache_init(trans->hmp, &lcache, 2);
85 seq = trans->hmp->flusher.act;
88 * Scan forwards. Retries typically occur if a deadlock is detected.
91 error = hammer_init_cursor(trans, &cursor, NULL, NULL);
93 hammer_done_cursor(&cursor);
96 cursor.key_beg = rebal->key_cur;
97 cursor.key_end = rebal->key_end;
98 cursor.key_end.localization &= HAMMER_LOCALIZE_MASK;
99 cursor.key_end.localization += ip->obj_localization;
100 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
101 cursor.flags |= HAMMER_CURSOR_BACKEND;
104 * Cause internal nodes to be returned on the way up. Internal nodes
105 * are not returned on the way down so we can create a degenerate
106 * case to handle internal nodes as a trailing function of their
109 * Note that by not setting INSERTING or PRUNING no boundary
110 * corrections will be made and a sync lock is not needed for the
111 * B-Tree scan itself.
113 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
115 error = hammer_btree_first(&cursor);
119 * Rebalancing can be hard on the memory allocator, make
120 * sure there is enough free memory before doing it.
125 * We only care about internal nodes visited for the last
126 * time on the way up... that is, a trailing scan of the
127 * internal node after all of its children have been recursed
130 if (cursor.node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
132 * Leave cursor.index alone, we want to recurse
133 * through all children of the internal node before
136 * Process the internal node on the way up after
137 * the last child's sub-tree has been balanced.
139 if (cursor.index == cursor.node->ondisk->count - 1) {
140 hammer_sync_lock_sh(trans);
141 error = rebalance_node(rebal, &cursor, &lcache);
142 hammer_sync_unlock(trans);
146 * We don't need to iterate through all the leaf
147 * elements, we only care about the parent (internal)
150 cursor.index = cursor.node->ondisk->count - 1;
156 * Update returned scan position and do a flush if
159 * WARNING: We extract the base using the leaf element
160 * type but this could be an internal node. The
161 * base is the same either way.
163 * However, due to the rebalancing operation the
164 * cursor position may have exceeded the right-hand
167 * WARNING: See warnings in hammer_unlock_cursor()
170 elm = &cursor.node->ondisk->elms[cursor.index].leaf;
171 rebal->key_cur = elm->base;
172 ++rebal->stat_ncount;
174 while (hammer_flusher_meta_halflimit(trans->hmp) ||
175 hammer_flusher_undo_exhausted(trans, 2)) {
176 hammer_unlock_cursor(&cursor);
177 hammer_flusher_wait(trans->hmp, seq);
178 hammer_lock_cursor(&cursor);
179 seq = hammer_flusher_async_one(trans->hmp);
183 * Before iterating check if the rebalance operation caused
184 * the cursor to index past the right-hand boundary and make
185 * sure to stop if it does. Otherwise the iteration may
186 * panic e.g. due to the key maxing out its fields and no
187 * longer being within the strict bounds of the root node.
189 if (hammer_btree_cmp(&rebal->key_cur, &cursor.key_end) > 0) {
190 rebal->key_cur = cursor.key_end;
195 * Iterate, stop if a signal was received.
197 if ((error = hammer_signal_check(trans->hmp)) != 0)
199 error = hammer_btree_iterate(&cursor);
203 hammer_done_cursor(&cursor);
204 if (error == EDEADLK) {
205 ++rebal->stat_collisions;
208 if (error == EINTR) {
209 rebal->head.flags |= HAMMER_IOC_HEAD_INTR;
213 rebal->key_cur.localization &= HAMMER_LOCALIZE_MASK;
214 hammer_btree_lcache_free(trans->hmp, &lcache);
219 * Rebalance an internal node, called via a trailing upward recursion.
220 * All the children have already been individually rebalanced.
222 * To rebalance we scan the elements in the children and pack them,
223 * so we actually need to lock the children and the children's children.
227 * C C C C C C C children (first level) (internal or leaf nodes)
228 * children's elements (second level)
230 * <<<---------- pack children's elements, possibly remove excess
231 * children after packing.
233 * NOTE: The mirror_tids, parent pointers, and child pointers must be updated.
234 * Any live tracked B-Tree nodes must be updated (we worm out of that
235 * by not allowing any). And boundary elements must be preserved.
237 * NOTE: If the children are leaf nodes we may have a degenerate case
238 * case where there are no elements in the leafs.
243 rebalance_node(struct hammer_ioc_rebalance *rebal, hammer_cursor_t cursor,
244 struct hammer_node_lock *lcache)
246 struct hammer_node_lock lockroot;
247 hammer_node_lock_t base_item;
248 hammer_node_lock_t chld_item;
249 hammer_node_lock_t item;
250 hammer_btree_elm_t elm;
263 * Lock the parent node via the cursor, collect and lock our
264 * children and children's children.
266 * By the way, this is a LOT of locks.
268 hammer_node_lock_init(&lockroot, cursor->node);
269 error = hammer_cursor_upgrade(cursor);
272 error = hammer_btree_lock_children(cursor, 2, &lockroot, lcache);
277 * Make a copy of all the locked on-disk data to simplify the element
278 * shifting we are going to have to do. We will modify the copy
281 hammer_btree_lock_copy(cursor, &lockroot);
284 * Look at the first child node.
286 if (TAILQ_FIRST(&lockroot.list) == NULL)
288 type1 = TAILQ_FIRST(&lockroot.list)->node->ondisk->type;
291 * Figure out the total number of children's children and
292 * calculate the average number of elements per child.
294 * The minimum avg_elms is 1 when count > 0. avg_elms * root_elms
295 * is always greater or equal to count.
297 * If count == 0 we hit a degenerate case which will cause
298 * avg_elms to also calculate as 0.
300 if (hammer_debug_general & 0x1000)
301 kprintf("lockroot %p count %d\n", &lockroot, lockroot.count);
303 TAILQ_FOREACH(item, &lockroot.list, entry) {
304 if (hammer_debug_general & 0x1000)
305 kprintf("add count %d\n", item->count);
306 count += item->count;
307 KKASSERT(item->node->ondisk->type == type1);
309 avg_elms = (count + (lockroot.count - 1)) / lockroot.count;
310 KKASSERT(avg_elms >= 0);
313 * If the average number of elements per child is too low then
314 * calculate the desired number of children (n) such that the
315 * average number of elements is reasonable.
317 * If the desired number of children is 1 then avg_elms will
318 * wind up being count, which may still be smaller then saturation
321 if (count && avg_elms < rebal->saturation) {
322 n = (count + (rebal->saturation - 1)) / rebal->saturation;
323 avg_elms = (count + (n - 1)) / n;
327 * Pack the elements in the children. Elements for each item is
328 * packed into base_item until avg_elms is reached, then base_item
331 * hammer_cursor_moved_element() is called for each element moved
332 * to update tracked cursors, including the index beyond the last
333 * element (at count).
335 * Any cursors tracking the internal node itself must also be
336 * updated, potentially repointing them at a leaf and clearing
339 base_item = TAILQ_FIRST(&lockroot.list);
343 TAILQ_FOREACH(item, &lockroot.list, entry) {
345 KKASSERT(item->count == node->ondisk->count);
346 chld_item = TAILQ_FIRST(&item->list);
347 for (i = 0; i < item->count; ++i) {
349 * Closeout. If the next element is at index 0
350 * just use the existing separator in the parent.
352 if (base_count == avg_elms) {
354 elm = &lockroot.node->ondisk->elms[
357 elm = &node->ondisk->elms[i];
359 rebalance_closeout(base_item, base_count, elm);
360 base_item = TAILQ_NEXT(base_item, entry);
367 * Check degenerate no-work case. Otherwise pack
370 * All changes are made to the copy.
372 if (item == base_item && i == base_count) {
375 chld_item = TAILQ_NEXT(chld_item, entry);
382 elm = &base_item->copy->elms[base_count];
383 *elm = node->ondisk->elms[i];
384 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
387 * Adjust the mirror_tid of the target and the
388 * internal element linkage.
390 * The parent node (lockroot.node) should already
391 * have an aggregate mirror_tid so we do not have
392 * to update that. However, it is possible for us
393 * to catch a hammer_btree_mirror_propagate() with
394 * its pants down. Update the parent if necessary.
396 tid = node->ondisk->mirror_tid;
398 if (base_item->copy->mirror_tid < tid) {
399 base_item->copy->mirror_tid = tid;
400 if (lockroot.copy->mirror_tid < tid) {
401 lockroot.copy->mirror_tid = tid;
403 HAMMER_NODE_LOCK_UPDATED;
405 if (lockroot.copy->elms[root_count].
406 internal.mirror_tid < tid) {
407 lockroot.copy->elms[root_count].
408 internal.mirror_tid = tid;
410 HAMMER_NODE_LOCK_UPDATED;
412 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
416 * We moved elm. The parent pointers for any
417 * children of elm must be repointed.
419 if (item != base_item &&
420 node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
422 rebalance_parent_ptrs(base_item, base_count,
425 hammer_cursor_moved_element(item->parent->node,
432 chld_item = TAILQ_NEXT(chld_item, entry);
436 * Always call at the end (i == number of elements) in
437 * case a cursor is sitting indexed there.
439 hammer_cursor_moved_element(item->parent->node, item->index,
441 base_item->node, base_count);
445 * Packing complete, close-out base_item using the right-hand
446 * boundary of the original parent.
448 * If we will be deleting nodes from the root shift the old
449 * right-hand-boundary to the new ending index.
451 elm = &lockroot.node->ondisk->elms[lockroot.node->ondisk->count];
452 rebalance_closeout(base_item, base_count, elm);
454 if (lockroot.copy->count != root_count) {
455 lockroot.copy->count = root_count;
456 lockroot.copy->elms[root_count] = *elm;
457 lockroot.flags |= HAMMER_NODE_LOCK_UPDATED;
461 * Any extra items beyond base_item are now completely empty and
462 * can be destroyed. Queue the destruction up in the copy. Note
463 * that none of the destroyed nodes are part of our cursor.
465 * The cursor is locked so it isn't on the tracking list. It
466 * should have been pointing at the boundary element (at root_count).
467 * When deleting elements from the root (which is cursor.node), we
468 * have to update the cursor.index manually to keep it in bounds.
470 while ((base_item = TAILQ_NEXT(base_item, entry)) != NULL) {
471 hammer_cursor_removed_node(base_item->node, lockroot.node,
473 hammer_cursor_deleted_element(lockroot.node, base_count);
474 base_item->copy->type = HAMMER_BTREE_TYPE_DELETED;
475 base_item->copy->count = 0;
476 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
477 if (cursor->index > lockroot.copy->count)
479 ++rebal->stat_deletions;
483 * All done, sync the locked child tree to disk. This will also
484 * flush and delete deleted nodes.
486 rebal->stat_nrebal += hammer_btree_sync_copy(cursor, &lockroot);
488 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, lcache);
489 hammer_cursor_downgrade(cursor);
494 * Close-out the child base_item. This node contains base_count
497 * If the node is an internal node the right-hand boundary must be
502 rebalance_closeout(hammer_node_lock_t base_item, int base_count,
503 hammer_btree_elm_t elm)
505 hammer_node_lock_t parent;
506 hammer_btree_elm_t base_elm;
507 hammer_btree_elm_t rbound_elm;
511 * Update the count. NOTE: base_count can be 0 for the
512 * degenerate leaf case.
514 if (hammer_debug_general & 0x1000) {
515 kprintf("rebalance_closeout %016llx:",
516 (long long)base_item->node->node_offset);
518 if (base_item->copy->count != base_count) {
519 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
520 base_item->copy->count = base_count;
521 if (hammer_debug_general & 0x1000)
522 kprintf(" (count update)");
526 * If we are closing out an internal node we must assign
527 * a right-hand boundary. Use the element contents as the
528 * right-hand boundary.
530 * Internal nodes are required to have at least one child,
531 * otherwise the left and right boundary would end up being
532 * the same element. Only leaf nodes can be empty.
534 * Rebalancing may cut-off an internal node such that the
535 * new right hand boundary is the next element anyway, but
536 * we still have to make sure that subtree_offset, btype,
537 * and mirror_tid are all 0.
539 if (base_item->copy->type == HAMMER_BTREE_TYPE_INTERNAL) {
540 KKASSERT(base_count != 0);
541 base_elm = &base_item->copy->elms[base_count];
543 if (bcmp(base_elm, elm, sizeof(*elm)) != 0 ||
544 elm->internal.subtree_offset ||
545 elm->internal.mirror_tid ||
548 base_elm->internal.subtree_offset = 0;
549 base_elm->internal.mirror_tid = 0;
550 base_elm->base.btype = 0;
551 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
552 if (hammer_debug_general & 0x1000)
553 kprintf(" (rhs update)");
555 if (hammer_debug_general & 0x1000)
556 kprintf(" (rhs same)");
561 * The parent's boundary must be updated. Be careful to retain
562 * the btype and non-base internal fields as that information is
565 parent = base_item->parent;
566 rbound_elm = &parent->copy->elms[base_item->index + 1];
567 if (bcmp(&rbound_elm->base, &elm->base, sizeof(elm->base)) != 0) {
568 save = rbound_elm->base.btype;
569 rbound_elm->base = elm->base;
570 rbound_elm->base.btype = save;
571 parent->flags |= HAMMER_NODE_LOCK_UPDATED;
572 if (hammer_debug_general & 0x1000) {
573 kprintf(" (parent bound update %d)",
574 base_item->index + 1);
577 if (hammer_debug_general & 0x1000)
582 * An element in item has moved to base_item. We must update the parent
583 * pointer of the node the element points to (which is chld_item).
587 rebalance_parent_ptrs(hammer_node_lock_t base_item, int index,
588 hammer_node_lock_t item, hammer_node_lock_t chld_item)
590 KKASSERT(chld_item->node->ondisk->parent == item->node->node_offset);
591 chld_item->copy->parent = base_item->node->node_offset;
592 chld_item->flags |= HAMMER_NODE_LOCK_UPDATED;
593 hammer_cursor_parent_changed(chld_item->node,
594 item->node, base_item->node, index);