4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 * Copyright (c) 2012, 2019 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/dsl_pool.h>
36 #include <sys/space_map.h>
37 #include <sys/refcount.h>
38 #include <sys/zfeature.h>
41 * Note on space map block size:
43 * The data for a given space map can be kept on blocks of any size.
44 * Larger blocks entail fewer I/O operations, but they also cause the
45 * DMU to keep more data in-core, and also to waste more I/O bandwidth
46 * when only a few blocks have changed since the last transaction group.
50 * Enabled whenever we want to stress test the use of double-word
53 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
56 * Override the default indirect block size of 128K, instead use 16K for
57 * spacemaps (2^14 bytes). This dramatically reduces write inflation since
58 * appending to a spacemap typically has to write one data block (4KB) and one
59 * or two indirect blocks (16K-32K, rather than 128K).
61 int space_map_ibs = 14;
64 sm_entry_is_debug(uint64_t e)
66 return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX);
70 sm_entry_is_single_word(uint64_t e)
72 uint8_t prefix = SM_PREFIX_DECODE(e);
73 return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX);
77 sm_entry_is_double_word(uint64_t e)
79 return (SM_PREFIX_DECODE(e) == SM2_PREFIX);
83 * Iterate through the space map, invoking the callback on each (non-debug)
84 * space map entry. Stop after reading 'end' bytes of the space map.
87 space_map_iterate(space_map_t *sm, uint64_t end, sm_cb_t callback, void *arg)
89 uint64_t blksz = sm->sm_blksz;
91 ASSERT3U(blksz, !=, 0);
92 ASSERT3U(end, <=, space_map_length(sm));
93 ASSERT0(P2PHASE(end, sizeof (uint64_t)));
95 dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, end,
96 ZIO_PRIORITY_SYNC_READ);
99 for (uint64_t block_base = 0; block_base < end && error == 0;
100 block_base += blksz) {
102 error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
103 block_base, FTAG, &db, DMU_READ_PREFETCH);
107 uint64_t *block_start = db->db_data;
108 uint64_t block_length = MIN(end - block_base, blksz);
109 uint64_t *block_end = block_start +
110 (block_length / sizeof (uint64_t));
112 VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
113 VERIFY3U(block_length, !=, 0);
114 ASSERT3U(blksz, ==, db->db_size);
116 for (uint64_t *block_cursor = block_start;
117 block_cursor < block_end && error == 0; block_cursor++) {
118 uint64_t e = *block_cursor;
120 if (sm_entry_is_debug(e)) /* Skip debug entries */
123 uint64_t raw_offset, raw_run, vdev_id;
125 if (sm_entry_is_single_word(e)) {
126 type = SM_TYPE_DECODE(e);
127 vdev_id = SM_NO_VDEVID;
128 raw_offset = SM_OFFSET_DECODE(e);
129 raw_run = SM_RUN_DECODE(e);
131 /* it is a two-word entry */
132 ASSERT(sm_entry_is_double_word(e));
133 raw_run = SM2_RUN_DECODE(e);
134 vdev_id = SM2_VDEV_DECODE(e);
136 /* move on to the second word */
139 VERIFY3P(block_cursor, <=, block_end);
141 type = SM2_TYPE_DECODE(e);
142 raw_offset = SM2_OFFSET_DECODE(e);
145 uint64_t entry_offset = (raw_offset << sm->sm_shift) +
147 uint64_t entry_run = raw_run << sm->sm_shift;
149 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
150 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
151 ASSERT3U(entry_offset, >=, sm->sm_start);
152 ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
153 ASSERT3U(entry_run, <=, sm->sm_size);
154 ASSERT3U(entry_offset + entry_run, <=,
155 sm->sm_start + sm->sm_size);
157 space_map_entry_t sme = {
160 .sme_offset = entry_offset,
163 error = callback(&sme, arg);
165 dmu_buf_rele(db, FTAG);
171 * Reads the entries from the last block of the space map into
172 * buf in reverse order. Populates nwords with number of words
175 * Refer to block comment within space_map_incremental_destroy()
176 * to understand why this function is needed.
179 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
180 uint64_t bufsz, uint64_t *nwords)
186 * Find the offset of the last word in the space map and use
187 * that to read the last block of the space map with
190 uint64_t last_word_offset =
191 sm->sm_phys->smp_length - sizeof (uint64_t);
192 error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
193 FTAG, &db, DMU_READ_NO_PREFETCH);
197 ASSERT3U(sm->sm_object, ==, db->db_object);
198 ASSERT3U(sm->sm_blksz, ==, db->db_size);
199 ASSERT3U(bufsz, >=, db->db_size);
200 ASSERT(nwords != NULL);
202 uint64_t *words = db->db_data;
204 (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
206 ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
208 uint64_t n = *nwords;
210 for (uint64_t i = 0; i < n; i++) {
211 uint64_t entry = words[i];
212 if (sm_entry_is_double_word(entry)) {
214 * Since we are populating the buffer backwards
215 * we have to be extra careful and add the two
216 * words of the double-word entry in the right
228 ASSERT(sm_entry_is_debug(entry) ||
229 sm_entry_is_single_word(entry));
236 * Assert that we wrote backwards all the
237 * way to the beginning of the buffer.
241 dmu_buf_rele(db, FTAG);
246 * Note: This function performs destructive actions - specifically
247 * it deletes entries from the end of the space map. Thus, callers
248 * should ensure that they are holding the appropriate locks for
249 * the space map that they provide.
252 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
255 uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
256 uint64_t *buf = zio_buf_alloc(bufsz);
258 dmu_buf_will_dirty(sm->sm_dbuf, tx);
261 * Ideally we would want to iterate from the beginning of the
262 * space map to the end in incremental steps. The issue with this
263 * approach is that we don't have any field on-disk that points
264 * us where to start between each step. We could try zeroing out
265 * entries that we've destroyed, but this doesn't work either as
266 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
268 * As a result, we destroy its entries incrementally starting from
269 * the end after applying the callback to each of them.
271 * The problem with this approach is that we cannot literally
272 * iterate through the words in the space map backwards as we
273 * can't distinguish two-word space map entries from their second
274 * word. Thus we do the following:
276 * 1] We get all the entries from the last block of the space map
277 * and put them into a buffer in reverse order. This way the
278 * last entry comes first in the buffer, the second to last is
280 * 2] We iterate through the entries in the buffer and we apply
281 * the callback to each one. As we move from entry to entry we
282 * we decrease the size of the space map, deleting effectively
284 * 3] If there are no more entries in the space map or the callback
285 * returns a value other than 0, we stop iterating over the
286 * space map. If there are entries remaining and the callback
287 * returned 0, we go back to step [1].
290 while (space_map_length(sm) > 0 && error == 0) {
292 error = space_map_reversed_last_block_entries(sm, buf, bufsz,
297 ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
299 for (uint64_t i = 0; i < nwords; i++) {
302 if (sm_entry_is_debug(e)) {
303 sm->sm_phys->smp_length -= sizeof (uint64_t);
308 uint64_t raw_offset, raw_run, vdev_id;
310 if (sm_entry_is_single_word(e)) {
311 type = SM_TYPE_DECODE(e);
312 vdev_id = SM_NO_VDEVID;
313 raw_offset = SM_OFFSET_DECODE(e);
314 raw_run = SM_RUN_DECODE(e);
316 ASSERT(sm_entry_is_double_word(e));
319 raw_run = SM2_RUN_DECODE(e);
320 vdev_id = SM2_VDEV_DECODE(e);
322 /* move to the second word */
326 ASSERT3P(i, <=, nwords);
328 type = SM2_TYPE_DECODE(e);
329 raw_offset = SM2_OFFSET_DECODE(e);
332 uint64_t entry_offset =
333 (raw_offset << sm->sm_shift) + sm->sm_start;
334 uint64_t entry_run = raw_run << sm->sm_shift;
336 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
337 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
338 VERIFY3U(entry_offset, >=, sm->sm_start);
339 VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
340 VERIFY3U(entry_run, <=, sm->sm_size);
341 VERIFY3U(entry_offset + entry_run, <=,
342 sm->sm_start + sm->sm_size);
344 space_map_entry_t sme = {
347 .sme_offset = entry_offset,
350 error = callback(&sme, arg);
354 if (type == SM_ALLOC)
355 sm->sm_phys->smp_alloc -= entry_run;
357 sm->sm_phys->smp_alloc += entry_run;
358 sm->sm_phys->smp_length -= words * sizeof (uint64_t);
362 if (space_map_length(sm) == 0) {
364 ASSERT0(space_map_allocated(sm));
367 zio_buf_free(buf, bufsz);
371 typedef struct space_map_load_arg {
372 space_map_t *smla_sm;
373 range_tree_t *smla_rt;
375 } space_map_load_arg_t;
378 space_map_load_callback(space_map_entry_t *sme, void *arg)
380 space_map_load_arg_t *smla = arg;
381 if (sme->sme_type == smla->smla_type) {
382 VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
383 smla->smla_sm->sm_size);
384 range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
386 range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
393 * Load the spacemap into the rangetree, like space_map_load. But only
394 * read the first 'length' bytes of the spacemap.
397 space_map_load_length(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
400 space_map_load_arg_t smla;
402 VERIFY0(range_tree_space(rt));
404 if (maptype == SM_FREE)
405 range_tree_add(rt, sm->sm_start, sm->sm_size);
409 smla.smla_type = maptype;
410 int err = space_map_iterate(sm, length,
411 space_map_load_callback, &smla);
414 range_tree_vacate(rt, NULL, NULL);
420 * Load the space map disk into the specified range tree. Segments of maptype
421 * are added to the range tree, other segment types are removed.
424 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
426 return (space_map_load_length(sm, rt, maptype, space_map_length(sm)));
430 space_map_histogram_clear(space_map_t *sm)
432 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
435 bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
439 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
442 * Verify that the in-core range tree does not have any
443 * ranges smaller than our sm_shift size.
445 for (int i = 0; i < sm->sm_shift; i++) {
446 if (rt->rt_histogram[i] != 0)
453 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
457 ASSERT(dmu_tx_is_syncing(tx));
458 VERIFY3U(space_map_object(sm), !=, 0);
460 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
463 dmu_buf_will_dirty(sm->sm_dbuf, tx);
465 ASSERT(space_map_histogram_verify(sm, rt));
467 * Transfer the content of the range tree histogram to the space
468 * map histogram. The space map histogram contains 32 buckets ranging
469 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
470 * however, can represent ranges from 2^0 to 2^63. Since the space
471 * map only cares about allocatable blocks (minimum of sm_shift) we
472 * can safely ignore all ranges in the range tree smaller than sm_shift.
474 for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
477 * Since the largest histogram bucket in the space map is
478 * 2^(32+sm_shift-1), we need to normalize the values in
479 * the range tree for any bucket larger than that size. For
480 * example given an sm_shift of 9, ranges larger than 2^40
481 * would get normalized as if they were 1TB ranges. Assume
482 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
483 * the calculation below would normalize this to 5 * 2^4 (16).
485 ASSERT3U(i, >=, idx + sm->sm_shift);
486 sm->sm_phys->smp_histogram[idx] +=
487 rt->rt_histogram[i] << (i - idx - sm->sm_shift);
490 * Increment the space map's index as long as we haven't
491 * reached the maximum bucket size. Accumulate all ranges
492 * larger than the max bucket size into the last bucket.
494 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
495 ASSERT3U(idx + sm->sm_shift, ==, i);
497 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
503 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
505 dmu_buf_will_dirty(sm->sm_dbuf, tx);
507 uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
508 SM_DEBUG_ACTION_ENCODE(maptype) |
509 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
510 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
512 dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_length,
513 sizeof (dentry), &dentry, tx);
515 sm->sm_phys->smp_length += sizeof (dentry);
519 * Writes one or more entries given a segment.
521 * Note: The function may release the dbuf from the pointer initially
522 * passed to it, and return a different dbuf. Also, the space map's
523 * dbuf must be dirty for the changes in sm_phys to take effect.
526 space_map_write_seg(space_map_t *sm, uint64_t rstart, uint64_t rend,
527 maptype_t maptype, uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp,
528 void *tag, dmu_tx_t *tx)
530 ASSERT3U(words, !=, 0);
531 ASSERT3U(words, <=, 2);
533 /* ensure the vdev_id can be represented by the space map */
534 ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
537 * if this is a single word entry, ensure that no vdev was
540 IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
542 dmu_buf_t *db = *dbp;
543 ASSERT3U(db->db_size, ==, sm->sm_blksz);
545 uint64_t *block_base = db->db_data;
546 uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
547 uint64_t *block_cursor = block_base +
548 (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
550 ASSERT3P(block_cursor, <=, block_end);
552 uint64_t size = (rend - rstart) >> sm->sm_shift;
553 uint64_t start = (rstart - sm->sm_start) >> sm->sm_shift;
554 uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
556 ASSERT3U(rstart, >=, sm->sm_start);
557 ASSERT3U(rstart, <, sm->sm_start + sm->sm_size);
558 ASSERT3U(rend - rstart, <=, sm->sm_size);
559 ASSERT3U(rend, <=, sm->sm_start + sm->sm_size);
562 ASSERT3P(block_cursor, <=, block_end);
565 * If we are at the end of this block, flush it and start
566 * writing again from the beginning.
568 if (block_cursor == block_end) {
569 dmu_buf_rele(db, tag);
571 uint64_t next_word_offset = sm->sm_phys->smp_length;
572 VERIFY0(dmu_buf_hold(sm->sm_os,
573 space_map_object(sm), next_word_offset,
574 tag, &db, DMU_READ_PREFETCH));
575 dmu_buf_will_dirty(db, tx);
577 /* update caller's dbuf */
580 ASSERT3U(db->db_size, ==, sm->sm_blksz);
582 block_base = db->db_data;
583 block_cursor = block_base;
584 block_end = block_base +
585 (db->db_size / sizeof (uint64_t));
589 * If we are writing a two-word entry and we only have one
590 * word left on this block, just pad it with an empty debug
591 * entry and write the two-word entry in the next block.
593 uint64_t *next_entry = block_cursor + 1;
594 if (next_entry == block_end && words > 1) {
595 ASSERT3U(words, ==, 2);
596 *block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
597 SM_DEBUG_ACTION_ENCODE(0) |
598 SM_DEBUG_SYNCPASS_ENCODE(0) |
599 SM_DEBUG_TXG_ENCODE(0);
601 sm->sm_phys->smp_length += sizeof (uint64_t);
602 ASSERT3P(block_cursor, ==, block_end);
606 uint64_t run_len = MIN(size, run_max);
609 *block_cursor = SM_OFFSET_ENCODE(start) |
610 SM_TYPE_ENCODE(maptype) |
611 SM_RUN_ENCODE(run_len);
615 /* write the first word of the entry */
616 *block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
617 SM2_RUN_ENCODE(run_len) |
618 SM2_VDEV_ENCODE(vdev_id);
621 /* move on to the second word of the entry */
622 ASSERT3P(block_cursor, <, block_end);
623 *block_cursor = SM2_TYPE_ENCODE(maptype) |
624 SM2_OFFSET_ENCODE(start);
628 panic("%d-word space map entries are not supported",
632 sm->sm_phys->smp_length += words * sizeof (uint64_t);
642 * Note: The space map's dbuf must be dirty for the changes in sm_phys to
646 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
647 uint64_t vdev_id, dmu_tx_t *tx)
649 spa_t *spa = tx->tx_pool->dp_spa;
652 space_map_write_intro_debug(sm, maptype, tx);
656 * We do this right after we write the intro debug entry
657 * because the estimate does not take it into account.
659 uint64_t initial_objsize = sm->sm_phys->smp_length;
660 uint64_t estimated_growth =
661 space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
662 uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
666 * Find the offset right after the last word in the space map
667 * and use that to get a hold of the last block, so we can
668 * start appending to it.
670 uint64_t next_word_offset = sm->sm_phys->smp_length;
671 VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
672 next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
673 ASSERT3U(db->db_size, ==, sm->sm_blksz);
675 dmu_buf_will_dirty(db, tx);
677 zfs_btree_t *t = &rt->rt_root;
678 zfs_btree_index_t where;
679 for (range_seg_t *rs = zfs_btree_first(t, &where); rs != NULL;
680 rs = zfs_btree_next(t, &where, &where)) {
681 uint64_t offset = (rs_get_start(rs, rt) - sm->sm_start) >>
683 uint64_t length = (rs_get_end(rs, rt) - rs_get_start(rs, rt)) >>
688 * We only write two-word entries when both of the following
691 * [1] The feature is enabled.
692 * [2] The offset or run is too big for a single-word entry,
693 * or the vdev_id is set (meaning not equal to
696 * Note that for purposes of testing we've added the case that
697 * we write two-word entries occasionally when the feature is
698 * enabled and zfs_force_some_double_word_sm_entries has been
701 if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
702 (offset >= (1ULL << SM_OFFSET_BITS) ||
703 length > SM_RUN_MAX ||
704 vdev_id != SM_NO_VDEVID ||
705 (zfs_force_some_double_word_sm_entries &&
706 spa_get_random(100) == 0)))
709 space_map_write_seg(sm, rs_get_start(rs, rt), rs_get_end(rs,
710 rt), maptype, vdev_id, words, &db, FTAG, tx);
713 dmu_buf_rele(db, FTAG);
717 * We expect our estimation to be based on the worst case
718 * scenario [see comment in space_map_estimate_optimal_size()].
719 * Therefore we expect the actual objsize to be equal or less
720 * than whatever we estimated it to be.
722 ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_length);
727 * Note: This function manipulates the state of the given space map but
728 * does not hold any locks implicitly. Thus the caller is responsible
729 * for synchronizing writes to the space map.
732 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
733 uint64_t vdev_id, dmu_tx_t *tx)
735 ASSERT(dsl_pool_sync_context(dmu_objset_pool(sm->sm_os)));
736 VERIFY3U(space_map_object(sm), !=, 0);
738 dmu_buf_will_dirty(sm->sm_dbuf, tx);
741 * This field is no longer necessary since the in-core space map
742 * now contains the object number but is maintained for backwards
745 sm->sm_phys->smp_object = sm->sm_object;
747 if (range_tree_is_empty(rt)) {
748 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
752 if (maptype == SM_ALLOC)
753 sm->sm_phys->smp_alloc += range_tree_space(rt);
755 sm->sm_phys->smp_alloc -= range_tree_space(rt);
757 uint64_t nodes = zfs_btree_numnodes(&rt->rt_root);
758 uint64_t rt_space = range_tree_space(rt);
760 space_map_write_impl(sm, rt, maptype, vdev_id, tx);
763 * Ensure that the space_map's accounting wasn't changed
764 * while we were in the middle of writing it out.
766 VERIFY3U(nodes, ==, zfs_btree_numnodes(&rt->rt_root));
767 VERIFY3U(range_tree_space(rt), ==, rt_space);
771 space_map_open_impl(space_map_t *sm)
776 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
780 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
781 sm->sm_phys = sm->sm_dbuf->db_data;
786 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
787 uint64_t start, uint64_t size, uint8_t shift)
792 ASSERT(*smp == NULL);
796 sm = kmem_alloc(sizeof (space_map_t), KM_SLEEP);
798 sm->sm_start = start;
800 sm->sm_shift = shift;
802 sm->sm_object = object;
807 error = space_map_open_impl(sm);
818 space_map_close(space_map_t *sm)
823 if (sm->sm_dbuf != NULL)
824 dmu_buf_rele(sm->sm_dbuf, sm);
828 kmem_free(sm, sizeof (*sm));
832 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
834 objset_t *os = sm->sm_os;
835 spa_t *spa = dmu_objset_spa(os);
836 dmu_object_info_t doi;
838 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
839 ASSERT(dmu_tx_is_syncing(tx));
840 VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
842 dmu_object_info_from_db(sm->sm_dbuf, &doi);
845 * If the space map has the wrong bonus size (because
846 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
847 * the wrong block size (because space_map_blksz has changed),
848 * free and re-allocate its object with the updated sizes.
850 * Otherwise, just truncate the current object.
852 if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
853 doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
854 doi.doi_data_block_size != blocksize ||
855 doi.doi_metadata_block_size != 1 << space_map_ibs) {
856 zfs_dbgmsg("txg %llu, spa %s, sm %px, reallocating "
857 "object[%llu]: old bonus %u, old blocksz %u",
858 dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
859 doi.doi_bonus_size, doi.doi_data_block_size);
861 space_map_free(sm, tx);
862 dmu_buf_rele(sm->sm_dbuf, sm);
864 sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
865 VERIFY0(space_map_open_impl(sm));
867 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
870 * If the spacemap is reallocated, its histogram
871 * will be reset. Do the same in the common case so that
872 * bugs related to the uncommon case do not go unnoticed.
874 bzero(sm->sm_phys->smp_histogram,
875 sizeof (sm->sm_phys->smp_histogram));
878 dmu_buf_will_dirty(sm->sm_dbuf, tx);
879 sm->sm_phys->smp_length = 0;
880 sm->sm_phys->smp_alloc = 0;
884 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
886 spa_t *spa = dmu_objset_spa(os);
890 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
891 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
892 bonuslen = sizeof (space_map_phys_t);
893 ASSERT3U(bonuslen, <=, dmu_bonus_max());
895 bonuslen = SPACE_MAP_SIZE_V0;
898 object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
899 space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
905 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx)
907 spa_t *spa = dmu_objset_spa(os);
908 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
909 dmu_object_info_t doi;
911 VERIFY0(dmu_object_info(os, smobj, &doi));
912 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
913 spa_feature_decr(spa,
914 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
918 VERIFY0(dmu_object_free(os, smobj, tx));
922 space_map_free(space_map_t *sm, dmu_tx_t *tx)
927 space_map_free_obj(sm->sm_os, space_map_object(sm), tx);
932 * Given a range tree, it makes a worst-case estimate of how much
933 * space would the tree's segments take if they were written to
934 * the given space map.
937 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
940 spa_t *spa = dmu_objset_spa(sm->sm_os);
941 uint64_t shift = sm->sm_shift;
942 uint64_t *histogram = rt->rt_histogram;
943 uint64_t entries_for_seg = 0;
946 * In order to get a quick estimate of the optimal size that this
947 * range tree would have on-disk as a space map, we iterate through
948 * its histogram buckets instead of iterating through its nodes.
950 * Note that this is a highest-bound/worst-case estimate for the
953 * 1] We assume that we always add a debug padding for each block
954 * we write and we also assume that we start at the last word
955 * of a block attempting to write a two-word entry.
956 * 2] Rounding up errors due to the way segments are distributed
957 * in the buckets of the range tree's histogram.
958 * 3] The activation of zfs_force_some_double_word_sm_entries
959 * (tunable) when testing.
961 * = Math and Rounding Errors =
963 * rt_histogram[i] bucket of a range tree represents the number
964 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
965 * that, we want to divide the buckets into groups: Buckets that
966 * can be represented using a single-word entry, ones that can
967 * be represented with a double-word entry, and ones that can
968 * only be represented with multiple two-word entries.
970 * [Note that if the new encoding feature is not enabled there
971 * are only two groups: single-word entry buckets and multiple
972 * single-word entry buckets. The information below assumes
973 * two-word entries enabled, but it can easily applied when
974 * the feature is not enabled]
976 * To find the highest bucket that can be represented with a
977 * single-word entry we look at the maximum run that such entry
978 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
979 * the run of a space map entry is shifted by sm_shift, thus we
980 * add it to the exponent]. This way, excluding the value of the
981 * maximum run that can be represented by a single-word entry,
982 * all runs that are smaller exist in buckets 0 to
983 * SM_RUN_BITS + shift - 1.
985 * To find the highest bucket that can be represented with a
986 * double-word entry, we follow the same approach. Finally, any
987 * bucket higher than that are represented with multiple two-word
988 * entries. To be more specific, if the highest bucket whose
989 * segments can be represented with a single two-word entry is X,
990 * then bucket X+1 will need 2 two-word entries for each of its
991 * segments, X+2 will need 4, X+3 will need 8, ...etc.
993 * With all of the above we make our estimation based on bucket
994 * groups. There is a rounding error though. As we mentioned in
995 * the example with the one-word entry, the maximum run that can
996 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
997 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
998 * that length fall into the next bucket (and bucket group) where
999 * we start counting two-word entries and this is one more reason
1000 * why the estimated size may end up being bigger than the actual
1006 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
1007 (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
1010 * If we are trying to force some double word entries just
1011 * assume the worst-case of every single word entry being
1012 * written as a double word entry.
1014 uint64_t entry_size =
1015 (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
1016 zfs_force_some_double_word_sm_entries) ?
1017 (2 * sizeof (uint64_t)) : sizeof (uint64_t);
1019 uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
1020 for (; idx <= single_entry_max_bucket; idx++)
1021 size += histogram[idx] * entry_size;
1023 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
1024 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1025 ASSERT3U(idx, >=, single_entry_max_bucket);
1027 1ULL << (idx - single_entry_max_bucket);
1028 size += histogram[idx] *
1029 entries_for_seg * entry_size;
1035 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
1037 uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
1038 for (; idx <= double_entry_max_bucket; idx++)
1039 size += histogram[idx] * 2 * sizeof (uint64_t);
1041 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1042 ASSERT3U(idx, >=, double_entry_max_bucket);
1043 entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
1044 size += histogram[idx] *
1045 entries_for_seg * 2 * sizeof (uint64_t);
1049 * Assume the worst case where we start with the padding at the end
1050 * of the current block and we add an extra padding entry at the end
1051 * of all subsequent blocks.
1053 size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
1059 space_map_object(space_map_t *sm)
1061 return (sm != NULL ? sm->sm_object : 0);
1065 space_map_allocated(space_map_t *sm)
1067 return (sm != NULL ? sm->sm_phys->smp_alloc : 0);
1071 space_map_length(space_map_t *sm)
1073 return (sm != NULL ? sm->sm_phys->smp_length : 0);
1077 space_map_nblocks(space_map_t *sm)
1081 return (DIV_ROUND_UP(space_map_length(sm), sm->sm_blksz));