2 * Copyright (c) 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_reblock.c,v 1.34 2008/11/13 02:18:43 dillon Exp $
37 * HAMMER reblocker - This code frees up fragmented physical space
39 * HAMMER only keeps track of free space on a big-block basis. A big-block
40 * containing holes can only be freed by migrating the remaining data in
41 * that big-block into a new big-block, then freeing the big-block.
43 * This function is called from an ioctl or via the hammer support thread.
48 static int hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
49 hammer_cursor_t cursor,
50 hammer_btree_elm_t elm);
51 static int hammer_reblock_data(struct hammer_ioc_reblock *reblock,
52 hammer_cursor_t cursor, hammer_btree_elm_t elm);
53 static int hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
54 hammer_cursor_t cursor, hammer_btree_elm_t elm);
55 static int hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
56 hammer_cursor_t cursor, hammer_btree_elm_t elm);
59 hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
60 struct hammer_ioc_reblock *reblock)
62 struct hammer_cursor cursor;
63 hammer_btree_elm_t elm;
70 * A fill level <= 20% is considered an emergency. free_level is
71 * inverted from fill_level.
73 if (reblock->free_level >= HAMMER_LARGEBLOCK_SIZE * 8 / 10)
74 slop = HAMMER_CHKSPC_EMERGENCY;
76 slop = HAMMER_CHKSPC_REBLOCK;
78 if ((reblock->key_beg.localization | reblock->key_end.localization) &
79 HAMMER_LOCALIZE_PSEUDOFS_MASK) {
82 if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
84 if (reblock->free_level < 0)
87 reblock->key_cur = reblock->key_beg;
88 reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
89 reblock->key_cur.localization += ip->obj_localization;
92 seq = trans->hmp->flusher.act;
94 error = hammer_init_cursor(trans, &cursor, NULL, NULL);
96 hammer_done_cursor(&cursor);
99 cursor.key_beg.localization = reblock->key_cur.localization;
100 cursor.key_beg.obj_id = reblock->key_cur.obj_id;
101 cursor.key_beg.key = HAMMER_MIN_KEY;
102 cursor.key_beg.create_tid = 1;
103 cursor.key_beg.delete_tid = 0;
104 cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
105 cursor.key_beg.obj_type = 0;
107 cursor.key_end.localization = (reblock->key_end.localization &
108 HAMMER_LOCALIZE_MASK) +
109 ip->obj_localization;
110 cursor.key_end.obj_id = reblock->key_end.obj_id;
111 cursor.key_end.key = HAMMER_MAX_KEY;
112 cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
113 cursor.key_end.delete_tid = 0;
114 cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
115 cursor.key_end.obj_type = 0;
117 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
118 cursor.flags |= HAMMER_CURSOR_BACKEND;
119 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
122 * This flag allows the btree scan code to return internal nodes,
123 * so we can reblock them in addition to the leafs. Only specify it
124 * if we intend to reblock B-Tree nodes.
126 if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
127 cursor.flags |= HAMMER_CURSOR_REBLOCKING;
129 error = hammer_btree_first(&cursor);
132 * Internal or Leaf node
134 KKASSERT(cursor.index < cursor.node->ondisk->count);
135 elm = &cursor.node->ondisk->elms[cursor.index];
136 reblock->key_cur.obj_id = elm->base.obj_id;
137 reblock->key_cur.localization = elm->base.localization;
140 * Yield to more important tasks
142 if ((error = hammer_signal_check(trans->hmp)) != 0)
146 * If there is insufficient free space it may be due to
147 * reserved bigblocks, which flushing might fix.
149 * We must force a retest in case the unlocked cursor is
150 * moved to the end of the leaf, or moved to an internal
153 * WARNING: See warnings in hammer_unlock_cursor() function.
155 if (hammer_checkspace(trans->hmp, slop)) {
156 if (++checkspace_count == 10) {
160 hammer_unlock_cursor(&cursor);
161 cursor.flags |= HAMMER_CURSOR_RETEST;
162 hammer_flusher_wait(trans->hmp, seq);
163 hammer_lock_cursor(&cursor);
164 seq = hammer_flusher_async(trans->hmp, NULL);
169 * Acquiring the sync_lock prevents the operation from
170 * crossing a synchronization boundary.
172 * NOTE: cursor.node may have changed on return.
174 * WARNING: See warnings in hammer_unlock_cursor() function.
176 hammer_sync_lock_sh(trans);
177 error = hammer_reblock_helper(reblock, &cursor, elm);
178 hammer_sync_unlock(trans);
180 while (hammer_flusher_meta_halflimit(trans->hmp) ||
181 hammer_flusher_undo_exhausted(trans, 2)) {
182 hammer_unlock_cursor(&cursor);
183 hammer_flusher_wait(trans->hmp, seq);
184 hammer_lock_cursor(&cursor);
185 seq = hammer_flusher_async_one(trans->hmp);
189 * Setup for iteration, our cursor flags may be modified by
190 * other threads while we are unlocked.
192 cursor.flags |= HAMMER_CURSOR_ATEDISK;
195 * We allocate data buffers, which atm we don't track
196 * dirty levels for because we allow the kernel to write
197 * them. But if we allocate too many we can still deadlock
200 * WARNING: See warnings in hammer_unlock_cursor() function.
201 * (The cursor's node and element may change!)
204 hammer_unlock_cursor(&cursor);
205 bwillwrite(HAMMER_XBUFSIZE);
206 hammer_lock_cursor(&cursor);
210 error = hammer_btree_iterate(&cursor);
215 hammer_done_cursor(&cursor);
216 if (error == EWOULDBLOCK) {
217 hammer_flusher_sync(trans->hmp);
220 if (error == EDEADLK)
222 if (error == EINTR) {
223 reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
227 reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
232 * Reblock the B-Tree (leaf) node, record, and/or data if necessary.
234 * XXX We have no visibility into internal B-Tree nodes at the moment,
238 hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
239 hammer_cursor_t cursor, hammer_btree_elm_t elm)
242 hammer_off_t tmp_offset;
243 hammer_node_ondisk_t ondisk;
244 struct hammer_btree_leaf_elm leaf;
251 hmp = cursor->trans->hmp;
254 * Reblock data. Note that data embedded in a record is reblocked
255 * by the record reblock code. Data processing only occurs at leaf
256 * nodes and for RECORD element types.
258 if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
260 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
262 tmp_offset = elm->leaf.data_offset;
269 * NOTE: Localization restrictions may also have been set-up, we can't
270 * just set the match flags willy-nilly here.
272 switch(elm->leaf.base.rec_type) {
273 case HAMMER_RECTYPE_INODE:
274 case HAMMER_RECTYPE_SNAPSHOT:
275 case HAMMER_RECTYPE_CONFIG:
276 iocflags = HAMMER_IOC_DO_INODES;
278 case HAMMER_RECTYPE_EXT:
279 case HAMMER_RECTYPE_FIX:
280 case HAMMER_RECTYPE_PFS:
281 case HAMMER_RECTYPE_DIRENTRY:
282 iocflags = HAMMER_IOC_DO_DIRS;
284 case HAMMER_RECTYPE_DATA:
285 case HAMMER_RECTYPE_DB:
286 iocflags = HAMMER_IOC_DO_DATA;
292 if (reblock->head.flags & iocflags) {
293 ++reblock->data_count;
294 reblock->data_byte_count += elm->leaf.data_len;
295 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
296 if (hammer_debug_general & 0x4000)
297 kprintf("D %6d/%d\n", bytes, reblock->free_level);
298 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
299 bytes >= reblock->free_level) {
301 * This is nasty, the uncache code may have to get
302 * vnode locks and because of that we can't hold
305 * WARNING: See warnings in hammer_unlock_cursor()
309 hammer_unlock_cursor(cursor);
310 hammer_io_direct_uncache(hmp, &leaf);
311 hammer_lock_cursor(cursor);
314 * elm may have become stale or invalid, reload it.
315 * ondisk variable is temporary only. Note that
316 * cursor->node and thus cursor->node->ondisk may
319 ondisk = cursor->node->ondisk;
320 elm = &ondisk->elms[cursor->index];
321 if (cursor->flags & HAMMER_CURSOR_RETEST) {
322 kprintf("hammer: debug: retest on "
323 "reblocker uncache\n");
325 } else if (ondisk->type != HAMMER_BTREE_TYPE_LEAF ||
326 cursor->index >= ondisk->count) {
327 kprintf("hammer: debug: shifted on "
328 "reblocker uncache\n");
330 } else if (bcmp(&elm->leaf, &leaf, sizeof(leaf))) {
331 kprintf("hammer: debug: changed on "
332 "reblocker uncache\n");
336 error = hammer_cursor_upgrade(cursor);
338 KKASSERT(cursor->index < ondisk->count);
339 error = hammer_reblock_data(reblock,
343 ++reblock->data_moves;
344 reblock->data_byte_moves += elm->leaf.data_len;
351 * Reblock a B-Tree internal or leaf node. A leaf node is reblocked
352 * on initial entry only (element 0). An internal node is reblocked
353 * when entered upward from its first leaf node only (also element 0).
354 * Further revisits of the internal node (index > 0) are ignored.
356 tmp_offset = cursor->node->node_offset;
357 if (cursor->index == 0 &&
358 error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
359 ++reblock->btree_count;
360 bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
361 if (hammer_debug_general & 0x4000)
362 kprintf("B %6d/%d\n", bytes, reblock->free_level);
363 if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
364 bytes >= reblock->free_level) {
365 error = hammer_cursor_upgrade(cursor);
367 if (cursor->parent) {
368 KKASSERT(cursor->parent_index <
369 cursor->parent->ondisk->count);
370 elm = &cursor->parent->ondisk->elms[cursor->parent_index];
374 switch(cursor->node->ondisk->type) {
375 case HAMMER_BTREE_TYPE_LEAF:
376 error = hammer_reblock_leaf_node(
377 reblock, cursor, elm);
379 case HAMMER_BTREE_TYPE_INTERNAL:
380 error = hammer_reblock_int_node(
381 reblock, cursor, elm);
384 panic("Illegal B-Tree node type");
388 ++reblock->btree_moves;
393 hammer_cursor_downgrade(cursor);
398 * Reblock a record's data. Both the B-Tree element and record pointers
399 * to the data must be adjusted.
402 hammer_reblock_data(struct hammer_ioc_reblock *reblock,
403 hammer_cursor_t cursor, hammer_btree_elm_t elm)
405 struct hammer_buffer *data_buffer = NULL;
406 hammer_off_t ndata_offset;
410 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
411 HAMMER_CURSOR_GET_LEAF);
414 ndata = hammer_alloc_data(cursor->trans, elm->leaf.data_len,
415 elm->leaf.base.rec_type,
416 &ndata_offset, &data_buffer,
420 hammer_io_notmeta(data_buffer);
423 * Move the data. Note that we must invalidate any cached
424 * data buffer in the cursor before calling blockmap_free.
425 * The blockmap_free may free up the entire large-block and
426 * will not be able to invalidate it if the cursor is holding
427 * a data buffer cached in that large block.
429 hammer_modify_buffer(cursor->trans, data_buffer, NULL, 0);
430 bcopy(cursor->data, ndata, elm->leaf.data_len);
431 hammer_modify_buffer_done(data_buffer);
432 hammer_cursor_invalidate_cache(cursor);
434 hammer_blockmap_free(cursor->trans,
435 elm->leaf.data_offset, elm->leaf.data_len);
437 hammer_modify_node(cursor->trans, cursor->node,
438 &elm->leaf.data_offset, sizeof(hammer_off_t));
439 elm->leaf.data_offset = ndata_offset;
440 hammer_modify_node_done(cursor->node);
444 hammer_rel_buffer(data_buffer, 0);
449 * Reblock a B-Tree leaf node. The parent must be adjusted to point to
450 * the new copy of the leaf node.
452 * elm is a pointer to the parent element pointing at cursor.node.
455 hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
456 hammer_cursor_t cursor, hammer_btree_elm_t elm)
463 * Don't supply a hint when allocating the leaf. Fills are done
464 * from the leaf upwards.
466 onode = cursor->node;
467 nnode = hammer_alloc_btree(cursor->trans, 0, &error);
475 hammer_lock_ex(&nnode->lock);
476 hammer_modify_node_noundo(cursor->trans, nnode);
477 bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
481 * We are not the root of the B-Tree
483 hammer_modify_node(cursor->trans, cursor->parent,
484 &elm->internal.subtree_offset,
485 sizeof(elm->internal.subtree_offset));
486 elm->internal.subtree_offset = nnode->node_offset;
487 hammer_modify_node_done(cursor->parent);
490 * We are the root of the B-Tree
492 hammer_volume_t volume;
494 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
495 KKASSERT(error == 0);
497 hammer_modify_volume_field(cursor->trans, volume,
499 volume->ondisk->vol0_btree_root = nnode->node_offset;
500 hammer_modify_volume_done(volume);
501 hammer_rel_volume(volume, 0);
504 hammer_cursor_replaced_node(onode, nnode);
505 hammer_delete_node(cursor->trans, onode);
507 if (hammer_debug_general & 0x4000) {
508 kprintf("REBLOCK LNODE %016llx -> %016llx\n",
509 (long long)onode->node_offset,
510 (long long)nnode->node_offset);
512 hammer_modify_node_done(nnode);
513 cursor->node = nnode;
515 hammer_unlock(&onode->lock);
516 hammer_rel_node(onode);
522 * Reblock a B-Tree internal node. The parent must be adjusted to point to
523 * the new copy of the internal node, and the node's children's parent
524 * pointers must also be adjusted to point to the new copy.
526 * elm is a pointer to the parent element pointing at cursor.node.
529 hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
530 hammer_cursor_t cursor, hammer_btree_elm_t elm)
532 struct hammer_node_lock lockroot;
539 hammer_node_lock_init(&lockroot, cursor->node);
540 error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
545 * The internal node is visited after recursing through its
546 * first element. Use the subtree offset allocated for that
547 * element as a hint for allocating the internal node.
549 onode = cursor->node;
550 if (onode->ondisk->count)
551 hint = onode->ondisk->elms[0].internal.subtree_offset;
554 nnode = hammer_alloc_btree(cursor->trans, 0, &error);
560 * Move the node. Adjust the parent's pointer to us first.
562 hammer_lock_ex(&nnode->lock);
563 hammer_modify_node_noundo(cursor->trans, nnode);
564 bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
568 * We are not the root of the B-Tree
570 hammer_modify_node(cursor->trans, cursor->parent,
571 &elm->internal.subtree_offset,
572 sizeof(elm->internal.subtree_offset));
573 elm->internal.subtree_offset = nnode->node_offset;
574 hammer_modify_node_done(cursor->parent);
577 * We are the root of the B-Tree
579 hammer_volume_t volume;
581 volume = hammer_get_root_volume(cursor->trans->hmp, &error);
582 KKASSERT(error == 0);
584 hammer_modify_volume_field(cursor->trans, volume,
586 volume->ondisk->vol0_btree_root = nnode->node_offset;
587 hammer_modify_volume_done(volume);
588 hammer_rel_volume(volume, 0);
592 * Now adjust our children's pointers to us.
594 for (i = 0; i < nnode->ondisk->count; ++i) {
595 elm = &nnode->ondisk->elms[i];
596 error = btree_set_parent(cursor->trans, nnode, elm);
598 panic("reblock internal node: fixup problem");
604 * The new node replaces the current node in the cursor. The cursor
605 * expects it to be locked so leave it locked. Discard onode.
607 hammer_cursor_replaced_node(onode, nnode);
608 hammer_delete_node(cursor->trans, onode);
610 if (hammer_debug_general & 0x4000) {
611 kprintf("REBLOCK INODE %016llx -> %016llx\n",
612 (long long)onode->node_offset,
613 (long long)nnode->node_offset);
615 hammer_modify_node_done(nnode);
616 cursor->node = nnode;
618 hammer_unlock(&onode->lock);
619 hammer_rel_node(onode);
622 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);