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);
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 hammer_btree_leaf_elm_t elm;
62 if ((rebal->key_beg.localization | rebal->key_end.localization) &
63 HAMMER_LOCALIZE_PSEUDOFS_MASK) {
66 if (rebal->key_beg.localization > rebal->key_end.localization)
68 if (rebal->key_beg.localization == rebal->key_end.localization) {
69 if (rebal->key_beg.obj_id > rebal->key_end.obj_id)
71 /* key-space limitations - no check needed */
73 if (rebal->saturation < HAMMER_BTREE_INT_ELMS / 2)
74 rebal->saturation = HAMMER_BTREE_INT_ELMS / 2;
75 if (rebal->saturation > HAMMER_BTREE_INT_ELMS)
76 rebal->saturation = HAMMER_BTREE_INT_ELMS;
78 rebal->key_cur = rebal->key_beg;
79 rebal->key_cur.localization &= HAMMER_LOCALIZE_MASK;
80 rebal->key_cur.localization += ip->obj_localization;
82 seq = trans->hmp->flusher.act;
85 * Scan forwards. Retries typically occur if a deadlock is detected.
88 error = hammer_init_cursor(trans, &cursor, NULL, NULL);
90 hammer_done_cursor(&cursor);
93 cursor.key_beg = rebal->key_cur;
94 cursor.key_end = rebal->key_end;
95 cursor.key_end.localization &= HAMMER_LOCALIZE_MASK;
96 cursor.key_end.localization += ip->obj_localization;
97 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
98 cursor.flags |= HAMMER_CURSOR_BACKEND;
101 * Cause internal nodes to be returned on the way up. Internal nodes
102 * are not returned on the way down so we can create a degenerate
103 * case to handle internal nodes as a trailing function of their
106 * Note that by not setting INSERTING or PRUNING no boundary
107 * corrections will be made and a sync lock is not needed for the
108 * B-Tree scan itself.
110 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
112 error = hammer_btree_first(&cursor);
116 * Rebalancing can be hard on the memory allocator, make
117 * sure there is enough free memory before doing it.
122 * We only care about internal nodes visited for the last
123 * time on the way up... that is, a trailing scan of the
124 * internal node after all of its children have been recursed
127 if (cursor.node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
129 * Leave cursor.index alone, we want to recurse
130 * through all children of the internal node before
133 * Process the internal node on the way up after
134 * the last child's sub-tree has been balanced.
136 if (cursor.index == cursor.node->ondisk->count - 1) {
137 hammer_sync_lock_sh(trans);
138 error = rebalance_node(rebal, &cursor);
139 hammer_sync_unlock(trans);
143 * We don't need to iterate through all the leaf
144 * elements, we only care about the parent (internal)
147 cursor.index = cursor.node->ondisk->count - 1;
153 * Update returned scan position and do a flush if
156 * WARNING: We extract the base using the leaf element
157 * type but this could be an internal node. The
158 * base is the same either way.
160 * However, due to the rebalancing operation the
161 * cursor position may have exceeded the right-hand
164 * WARNING: See warnings in hammer_unlock_cursor()
167 elm = &cursor.node->ondisk->elms[cursor.index].leaf;
168 rebal->key_cur = elm->base;
169 ++rebal->stat_ncount;
171 while (hammer_flusher_meta_halflimit(trans->hmp) ||
172 hammer_flusher_undo_exhausted(trans, 2)) {
173 hammer_unlock_cursor(&cursor);
174 hammer_flusher_wait(trans->hmp, seq);
175 hammer_lock_cursor(&cursor);
176 seq = hammer_flusher_async_one(trans->hmp);
180 * Before iterating check if the rebalance operation caused
181 * the cursor to index past the right-hand boundary and make
182 * sure to stop if it does. Otherwise the iteration may
183 * panic e.g. due to the key maxing out its fields and no
184 * longer being within the strict bounds of the root node.
186 if (hammer_btree_cmp(&rebal->key_cur, &cursor.key_end) > 0) {
187 rebal->key_cur = cursor.key_end;
192 * Iterate, stop if a signal was received.
194 if ((error = hammer_signal_check(trans->hmp)) != 0)
196 error = hammer_btree_iterate(&cursor);
200 hammer_done_cursor(&cursor);
201 if (error == EDEADLK) {
202 ++rebal->stat_collisions;
205 if (error == EINTR) {
206 rebal->head.flags |= HAMMER_IOC_HEAD_INTR;
210 rebal->key_cur.localization &= HAMMER_LOCALIZE_MASK;
215 * Rebalance an internal node, called via a trailing upward recursion.
216 * All the children have already been individually rebalanced.
218 * To rebalance we scan the elements in the children and pack them,
219 * so we actually need to lock the children and the children's children.
223 * C C C C C C C children (first level) (internal or leaf nodes)
224 * children's elements (second level)
226 * <<<---------- pack children's elements, possibly remove excess
227 * children after packing.
229 * NOTE: The mirror_tids, parent pointers, and child pointers must be updated.
230 * Any live tracked B-Tree nodes must be updated (we worm out of that
231 * by not allowing any). And boundary elements must be preserved.
233 * NOTE: If the children are leaf nodes we may have a degenerate case
234 * case where there are no elements in the leafs.
239 rebalance_node(struct hammer_ioc_rebalance *rebal, hammer_cursor_t cursor)
241 struct hammer_node_lock lockroot;
242 hammer_node_lock_t base_item;
243 hammer_node_lock_t chld_item;
244 hammer_node_lock_t item;
245 hammer_btree_elm_t elm;
258 * Lock the parent node via the cursor, collect and lock our
259 * children and children's children.
261 * By the way, this is a LOT of locks.
263 hammer_node_lock_init(&lockroot, cursor->node);
264 error = hammer_cursor_upgrade(cursor);
267 error = hammer_btree_lock_children(cursor, 2, &lockroot);
272 * Make a copy of all the locked on-disk data to simplify the element
273 * shifting we are going to have to do. We will modify the copy
276 hammer_btree_lock_copy(cursor, &lockroot);
279 * Look at the first child node.
281 if (TAILQ_FIRST(&lockroot.list) == NULL)
283 type1 = TAILQ_FIRST(&lockroot.list)->node->ondisk->type;
286 * Figure out the total number of children's children and
287 * calculate the average number of elements per child.
289 * The minimum avg_elms is 1 when count > 0. avg_elms * root_elms
290 * is always greater or equal to count.
292 * If count == 0 we hit a degenerate case which will cause
293 * avg_elms to also calculate as 0.
295 if (hammer_debug_general & 0x1000)
296 kprintf("lockroot %p count %d\n", &lockroot, lockroot.count);
298 TAILQ_FOREACH(item, &lockroot.list, entry) {
299 if (hammer_debug_general & 0x1000)
300 kprintf("add count %d\n", item->count);
301 count += item->count;
302 KKASSERT(item->node->ondisk->type == type1);
304 avg_elms = (count + (lockroot.count - 1)) / lockroot.count;
305 KKASSERT(avg_elms >= 0);
308 * If the average number of elements per child is too low then
309 * calculate the desired number of children (n) such that the
310 * average number of elements is reasonable.
312 * If the desired number of children is 1 then avg_elms will
313 * wind up being count, which may still be smaller then saturation
316 if (count && avg_elms < rebal->saturation) {
317 n = (count + (rebal->saturation - 1)) / rebal->saturation;
318 avg_elms = (count + (n - 1)) / n;
322 * Pack the elements in the children. Elements for each item is
323 * packed into base_item until avg_elms is reached, then base_item
326 * hammer_cursor_moved_element() is called for each element moved
327 * to update tracked cursors, including the index beyond the last
328 * element (at count).
330 * Any cursors tracking the internal node itself must also be
331 * updated, potentially repointing them at a leaf and clearing
334 base_item = TAILQ_FIRST(&lockroot.list);
338 TAILQ_FOREACH(item, &lockroot.list, entry) {
340 KKASSERT(item->count == node->ondisk->count);
341 chld_item = TAILQ_FIRST(&item->list);
342 for (i = 0; i < item->count; ++i) {
344 * Closeout. If the next element is at index 0
345 * just use the existing separator in the parent.
347 if (base_count == avg_elms) {
349 elm = &lockroot.node->ondisk->elms[
352 elm = &node->ondisk->elms[i];
354 rebalance_closeout(base_item, base_count, elm);
355 base_item = TAILQ_NEXT(base_item, entry);
362 * Check degenerate no-work case. Otherwise pack
365 * All changes are made to the copy.
367 if (item == base_item && i == base_count) {
370 chld_item = TAILQ_NEXT(chld_item, entry);
377 elm = &base_item->copy->elms[base_count];
378 *elm = node->ondisk->elms[i];
379 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
382 * Adjust the mirror_tid of the target and the
383 * internal element linkage.
385 * The parent node (lockroot.node) should already
386 * have an aggregate mirror_tid so we do not have
387 * to update that. However, it is possible for us
388 * to catch a hammer_btree_mirror_propagate() with
389 * its pants down. Update the parent if necessary.
391 tid = node->ondisk->mirror_tid;
393 if (base_item->copy->mirror_tid < tid) {
394 base_item->copy->mirror_tid = tid;
395 if (lockroot.copy->mirror_tid < tid) {
396 lockroot.copy->mirror_tid = tid;
398 HAMMER_NODE_LOCK_UPDATED;
400 if (lockroot.copy->elms[root_count].
401 internal.mirror_tid < tid) {
402 lockroot.copy->elms[root_count].
403 internal.mirror_tid = tid;
405 HAMMER_NODE_LOCK_UPDATED;
407 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
411 * We moved elm. The parent pointers for any
412 * children of elm must be repointed.
414 if (item != base_item &&
415 node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
417 rebalance_parent_ptrs(base_item, base_count,
420 hammer_cursor_moved_element(item->parent->node,
427 chld_item = TAILQ_NEXT(chld_item, entry);
431 * Always call at the end (i == number of elements) in
432 * case a cursor is sitting indexed there.
434 hammer_cursor_moved_element(item->parent->node, item->index,
436 base_item->node, base_count);
440 * Packing complete, close-out base_item using the right-hand
441 * boundary of the original parent.
443 * If we will be deleting nodes from the root shift the old
444 * right-hand-boundary to the new ending index.
446 elm = &lockroot.node->ondisk->elms[lockroot.node->ondisk->count];
447 rebalance_closeout(base_item, base_count, elm);
449 if (lockroot.copy->count != root_count) {
450 lockroot.copy->count = root_count;
451 lockroot.copy->elms[root_count] = *elm;
452 lockroot.flags |= HAMMER_NODE_LOCK_UPDATED;
456 * Any extra items beyond base_item are now completely empty and
457 * can be destroyed. Queue the destruction up in the copy. Note
458 * that none of the destroyed nodes are part of our cursor.
460 * The cursor is locked so it isn't on the tracking list. It
461 * should have been pointing at the boundary element (at root_count).
462 * When deleting elements from the root (which is cursor.node), we
463 * have to update the cursor.index manually to keep it in bounds.
465 while ((base_item = TAILQ_NEXT(base_item, entry)) != NULL) {
466 hammer_cursor_removed_node(base_item->node, lockroot.node,
468 hammer_cursor_deleted_element(lockroot.node, base_count);
469 base_item->copy->type = HAMMER_BTREE_TYPE_DELETED;
470 base_item->copy->count = 0;
471 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
472 if (cursor->index > lockroot.copy->count)
474 ++rebal->stat_deletions;
478 * All done, sync the locked child tree to disk. This will also
479 * flush and delete deleted nodes.
481 rebal->stat_nrebal += hammer_btree_sync_copy(cursor, &lockroot);
483 hammer_btree_unlock_children(cursor, &lockroot);
484 hammer_cursor_downgrade(cursor);
489 * Close-out the child base_item. This node contains base_count
492 * If the node is an internal node the right-hand boundary must be
497 rebalance_closeout(hammer_node_lock_t base_item, int base_count,
498 hammer_btree_elm_t elm)
500 hammer_node_lock_t parent;
501 hammer_btree_elm_t base_elm;
502 hammer_btree_elm_t rbound_elm;
506 * Update the count. NOTE: base_count can be 0 for the
507 * degenerate leaf case.
509 if (hammer_debug_general & 0x1000) {
510 kprintf("rebalance_closeout %016llx:",
511 (long long)base_item->node->node_offset);
513 if (base_item->copy->count != base_count) {
514 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
515 base_item->copy->count = base_count;
516 if (hammer_debug_general & 0x1000)
517 kprintf(" (count update)");
521 * If we are closing out an internal node we must assign
522 * a right-hand boundary. Use the element contents as the
523 * right-hand boundary.
525 * Internal nodes are required to have at least one child,
526 * otherwise the left and right boundary would end up being
527 * the same element. Only leaf nodes can be empty.
529 * Rebalancing may cut-off an internal node such that the
530 * new right hand boundary is the next element anyway, but
531 * we still have to make sure that subtree_offset, btype,
532 * and mirror_tid are all 0.
534 if (base_item->copy->type == HAMMER_BTREE_TYPE_INTERNAL) {
535 KKASSERT(base_count != 0);
536 base_elm = &base_item->copy->elms[base_count];
538 if (bcmp(base_elm, elm, sizeof(*elm)) != 0 ||
539 elm->internal.subtree_offset ||
540 elm->internal.mirror_tid ||
543 base_elm->internal.subtree_offset = 0;
544 base_elm->internal.mirror_tid = 0;
545 base_elm->base.btype = 0;
546 base_item->flags |= HAMMER_NODE_LOCK_UPDATED;
547 if (hammer_debug_general & 0x1000)
548 kprintf(" (rhs update)");
550 if (hammer_debug_general & 0x1000)
551 kprintf(" (rhs same)");
556 * The parent's boundary must be updated. Be careful to retain
557 * the btype and non-base internal fields as that information is
560 parent = base_item->parent;
561 rbound_elm = &parent->copy->elms[base_item->index + 1];
562 if (bcmp(&rbound_elm->base, &elm->base, sizeof(elm->base)) != 0) {
563 save = rbound_elm->base.btype;
564 rbound_elm->base = elm->base;
565 rbound_elm->base.btype = save;
566 parent->flags |= HAMMER_NODE_LOCK_UPDATED;
567 if (hammer_debug_general & 0x1000) {
568 kprintf(" (parent bound update %d)",
569 base_item->index + 1);
572 if (hammer_debug_general & 0x1000)
577 * An element in item has moved to base_item. We must update the parent
578 * pointer of the node the element points to (which is chld_item).
582 rebalance_parent_ptrs(hammer_node_lock_t base_item, int index,
583 hammer_node_lock_t item, hammer_node_lock_t chld_item)
585 KKASSERT(chld_item->node->ondisk->parent == item->node->node_offset);
586 chld_item->copy->parent = base_item->node->node_offset;
587 chld_item->flags |= HAMMER_NODE_LOCK_UPDATED;
588 hammer_cursor_parent_changed(chld_item->node,
589 item->node, base_item->node, index);