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, 2017 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>
40 SYSCTL_DECL(_vfs_zfs);
43 * Note on space map block size:
45 * The data for a given space map can be kept on blocks of any size.
46 * Larger blocks entail fewer I/O operations, but they also cause the
47 * DMU to keep more data in-core, and also to waste more I/O bandwidth
48 * when only a few blocks have changed since the last transaction group.
52 * Enabled whenever we want to stress test the use of double-word
55 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
58 sm_entry_is_debug(uint64_t e)
60 return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX);
64 sm_entry_is_single_word(uint64_t e)
66 uint8_t prefix = SM_PREFIX_DECODE(e);
67 return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX);
71 sm_entry_is_double_word(uint64_t e)
73 return (SM_PREFIX_DECODE(e) == SM2_PREFIX);
77 * Iterate through the space map, invoking the callback on each (non-debug)
81 space_map_iterate(space_map_t *sm, sm_cb_t callback, void *arg)
83 uint64_t sm_len = space_map_length(sm);
84 ASSERT3U(sm->sm_blksz, !=, 0);
86 dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, sm_len,
87 ZIO_PRIORITY_SYNC_READ);
89 uint64_t blksz = sm->sm_blksz;
91 for (uint64_t block_base = 0; block_base < sm_len && error == 0;
92 block_base += blksz) {
94 error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
95 block_base, FTAG, &db, DMU_READ_PREFETCH);
99 uint64_t *block_start = db->db_data;
100 uint64_t block_length = MIN(sm_len - block_base, blksz);
101 uint64_t *block_end = block_start +
102 (block_length / sizeof (uint64_t));
104 VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
105 VERIFY3U(block_length, !=, 0);
106 ASSERT3U(blksz, ==, db->db_size);
108 for (uint64_t *block_cursor = block_start;
109 block_cursor < block_end && error == 0; block_cursor++) {
110 uint64_t e = *block_cursor;
112 if (sm_entry_is_debug(e)) /* Skip debug entries */
115 uint64_t raw_offset, raw_run, vdev_id;
117 if (sm_entry_is_single_word(e)) {
118 type = SM_TYPE_DECODE(e);
119 vdev_id = SM_NO_VDEVID;
120 raw_offset = SM_OFFSET_DECODE(e);
121 raw_run = SM_RUN_DECODE(e);
123 /* it is a two-word entry */
124 ASSERT(sm_entry_is_double_word(e));
125 raw_run = SM2_RUN_DECODE(e);
126 vdev_id = SM2_VDEV_DECODE(e);
128 /* move on to the second word */
131 VERIFY3P(block_cursor, <=, block_end);
133 type = SM2_TYPE_DECODE(e);
134 raw_offset = SM2_OFFSET_DECODE(e);
137 uint64_t entry_offset = (raw_offset << sm->sm_shift) +
139 uint64_t entry_run = raw_run << sm->sm_shift;
141 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
142 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
143 ASSERT3U(entry_offset, >=, sm->sm_start);
144 ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
145 ASSERT3U(entry_run, <=, sm->sm_size);
146 ASSERT3U(entry_offset + entry_run, <=,
147 sm->sm_start + sm->sm_size);
149 space_map_entry_t sme = {
152 .sme_offset = entry_offset,
155 error = callback(&sme, arg);
157 dmu_buf_rele(db, FTAG);
163 * Reads the entries from the last block of the space map into
164 * buf in reverse order. Populates nwords with number of words
167 * Refer to block comment within space_map_incremental_destroy()
168 * to understand why this function is needed.
171 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
172 uint64_t bufsz, uint64_t *nwords)
178 * Find the offset of the last word in the space map and use
179 * that to read the last block of the space map with
182 uint64_t last_word_offset =
183 sm->sm_phys->smp_objsize - sizeof (uint64_t);
184 error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
185 FTAG, &db, DMU_READ_NO_PREFETCH);
189 ASSERT3U(sm->sm_object, ==, db->db_object);
190 ASSERT3U(sm->sm_blksz, ==, db->db_size);
191 ASSERT3U(bufsz, >=, db->db_size);
192 ASSERT(nwords != NULL);
194 uint64_t *words = db->db_data;
196 (sm->sm_phys->smp_objsize - db->db_offset) / sizeof (uint64_t);
198 ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
200 uint64_t n = *nwords;
202 for (uint64_t i = 0; i < n; i++) {
203 uint64_t entry = words[i];
204 if (sm_entry_is_double_word(entry)) {
206 * Since we are populating the buffer backwards
207 * we have to be extra careful and add the two
208 * words of the double-word entry in the right
220 ASSERT(sm_entry_is_debug(entry) ||
221 sm_entry_is_single_word(entry));
228 * Assert that we wrote backwards all the
229 * way to the beginning of the buffer.
233 dmu_buf_rele(db, FTAG);
238 * Note: This function performs destructive actions - specifically
239 * it deletes entries from the end of the space map. Thus, callers
240 * should ensure that they are holding the appropriate locks for
241 * the space map that they provide.
244 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
247 uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
248 uint64_t *buf = zio_buf_alloc(bufsz);
250 dmu_buf_will_dirty(sm->sm_dbuf, tx);
253 * Ideally we would want to iterate from the beginning of the
254 * space map to the end in incremental steps. The issue with this
255 * approach is that we don't have any field on-disk that points
256 * us where to start between each step. We could try zeroing out
257 * entries that we've destroyed, but this doesn't work either as
258 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
260 * As a result, we destroy its entries incrementally starting from
261 * the end after applying the callback to each of them.
263 * The problem with this approach is that we cannot literally
264 * iterate through the words in the space map backwards as we
265 * can't distinguish two-word space map entries from their second
266 * word. Thus we do the following:
268 * 1] We get all the entries from the last block of the space map
269 * and put them into a buffer in reverse order. This way the
270 * last entry comes first in the buffer, the second to last is
272 * 2] We iterate through the entries in the buffer and we apply
273 * the callback to each one. As we move from entry to entry we
274 * we decrease the size of the space map, deleting effectively
276 * 3] If there are no more entries in the space map or the callback
277 * returns a value other than 0, we stop iterating over the
278 * space map. If there are entries remaining and the callback
279 * returned 0, we go back to step [1].
282 while (space_map_length(sm) > 0 && error == 0) {
284 error = space_map_reversed_last_block_entries(sm, buf, bufsz,
289 ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
291 for (uint64_t i = 0; i < nwords; i++) {
294 if (sm_entry_is_debug(e)) {
295 sm->sm_phys->smp_objsize -= sizeof (uint64_t);
296 space_map_update(sm);
301 uint64_t raw_offset, raw_run, vdev_id;
303 if (sm_entry_is_single_word(e)) {
304 type = SM_TYPE_DECODE(e);
305 vdev_id = SM_NO_VDEVID;
306 raw_offset = SM_OFFSET_DECODE(e);
307 raw_run = SM_RUN_DECODE(e);
309 ASSERT(sm_entry_is_double_word(e));
312 raw_run = SM2_RUN_DECODE(e);
313 vdev_id = SM2_VDEV_DECODE(e);
315 /* move to the second word */
319 ASSERT3P(i, <=, nwords);
321 type = SM2_TYPE_DECODE(e);
322 raw_offset = SM2_OFFSET_DECODE(e);
325 uint64_t entry_offset =
326 (raw_offset << sm->sm_shift) + sm->sm_start;
327 uint64_t entry_run = raw_run << sm->sm_shift;
329 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
330 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
331 VERIFY3U(entry_offset, >=, sm->sm_start);
332 VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
333 VERIFY3U(entry_run, <=, sm->sm_size);
334 VERIFY3U(entry_offset + entry_run, <=,
335 sm->sm_start + sm->sm_size);
337 space_map_entry_t sme = {
340 .sme_offset = entry_offset,
343 error = callback(&sme, arg);
347 if (type == SM_ALLOC)
348 sm->sm_phys->smp_alloc -= entry_run;
350 sm->sm_phys->smp_alloc += entry_run;
351 sm->sm_phys->smp_objsize -= words * sizeof (uint64_t);
352 space_map_update(sm);
356 if (space_map_length(sm) == 0) {
358 ASSERT0(sm->sm_phys->smp_objsize);
359 ASSERT0(sm->sm_alloc);
362 zio_buf_free(buf, bufsz);
366 typedef struct space_map_load_arg {
367 space_map_t *smla_sm;
368 range_tree_t *smla_rt;
370 } space_map_load_arg_t;
373 space_map_load_callback(space_map_entry_t *sme, void *arg)
375 space_map_load_arg_t *smla = arg;
376 if (sme->sme_type == smla->smla_type) {
377 VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
378 smla->smla_sm->sm_size);
379 range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
381 range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
388 * Load the space map disk into the specified range tree. Segments of maptype
389 * are added to the range tree, other segment types are removed.
392 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
396 space_map_load_arg_t smla;
398 VERIFY0(range_tree_space(rt));
399 space = space_map_allocated(sm);
401 if (maptype == SM_FREE) {
402 range_tree_add(rt, sm->sm_start, sm->sm_size);
403 space = sm->sm_size - space;
408 smla.smla_type = maptype;
409 err = space_map_iterate(sm, space_map_load_callback, &smla);
412 VERIFY3U(range_tree_space(rt), ==, space);
414 range_tree_vacate(rt, NULL, NULL);
421 space_map_histogram_clear(space_map_t *sm)
423 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
426 bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
430 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
433 * Verify that the in-core range tree does not have any
434 * ranges smaller than our sm_shift size.
436 for (int i = 0; i < sm->sm_shift; i++) {
437 if (rt->rt_histogram[i] != 0)
444 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
448 ASSERT(dmu_tx_is_syncing(tx));
449 VERIFY3U(space_map_object(sm), !=, 0);
451 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
454 dmu_buf_will_dirty(sm->sm_dbuf, tx);
456 ASSERT(space_map_histogram_verify(sm, rt));
458 * Transfer the content of the range tree histogram to the space
459 * map histogram. The space map histogram contains 32 buckets ranging
460 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
461 * however, can represent ranges from 2^0 to 2^63. Since the space
462 * map only cares about allocatable blocks (minimum of sm_shift) we
463 * can safely ignore all ranges in the range tree smaller than sm_shift.
465 for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
468 * Since the largest histogram bucket in the space map is
469 * 2^(32+sm_shift-1), we need to normalize the values in
470 * the range tree for any bucket larger than that size. For
471 * example given an sm_shift of 9, ranges larger than 2^40
472 * would get normalized as if they were 1TB ranges. Assume
473 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
474 * the calculation below would normalize this to 5 * 2^4 (16).
476 ASSERT3U(i, >=, idx + sm->sm_shift);
477 sm->sm_phys->smp_histogram[idx] +=
478 rt->rt_histogram[i] << (i - idx - sm->sm_shift);
481 * Increment the space map's index as long as we haven't
482 * reached the maximum bucket size. Accumulate all ranges
483 * larger than the max bucket size into the last bucket.
485 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
486 ASSERT3U(idx + sm->sm_shift, ==, i);
488 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
494 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
496 dmu_buf_will_dirty(sm->sm_dbuf, tx);
498 uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
499 SM_DEBUG_ACTION_ENCODE(maptype) |
500 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
501 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
503 dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_objsize,
504 sizeof (dentry), &dentry, tx);
506 sm->sm_phys->smp_objsize += sizeof (dentry);
510 * Writes one or more entries given a segment.
512 * Note: The function may release the dbuf from the pointer initially
513 * passed to it, and return a different dbuf. Also, the space map's
514 * dbuf must be dirty for the changes in sm_phys to take effect.
517 space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype,
518 uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx)
520 ASSERT3U(words, !=, 0);
521 ASSERT3U(words, <=, 2);
523 /* ensure the vdev_id can be represented by the space map */
524 ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
527 * if this is a single word entry, ensure that no vdev was
530 IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
532 dmu_buf_t *db = *dbp;
533 ASSERT3U(db->db_size, ==, sm->sm_blksz);
535 uint64_t *block_base = db->db_data;
536 uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
537 uint64_t *block_cursor = block_base +
538 (sm->sm_phys->smp_objsize - db->db_offset) / sizeof (uint64_t);
540 ASSERT3P(block_cursor, <=, block_end);
542 uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
543 uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
544 uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
546 ASSERT3U(rs->rs_start, >=, sm->sm_start);
547 ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size);
548 ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size);
549 ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size);
552 ASSERT3P(block_cursor, <=, block_end);
555 * If we are at the end of this block, flush it and start
556 * writing again from the beginning.
558 if (block_cursor == block_end) {
559 dmu_buf_rele(db, tag);
561 uint64_t next_word_offset = sm->sm_phys->smp_objsize;
562 VERIFY0(dmu_buf_hold(sm->sm_os,
563 space_map_object(sm), next_word_offset,
564 tag, &db, DMU_READ_PREFETCH));
565 dmu_buf_will_dirty(db, tx);
567 /* update caller's dbuf */
570 ASSERT3U(db->db_size, ==, sm->sm_blksz);
572 block_base = db->db_data;
573 block_cursor = block_base;
574 block_end = block_base +
575 (db->db_size / sizeof (uint64_t));
579 * If we are writing a two-word entry and we only have one
580 * word left on this block, just pad it with an empty debug
581 * entry and write the two-word entry in the next block.
583 uint64_t *next_entry = block_cursor + 1;
584 if (next_entry == block_end && words > 1) {
585 ASSERT3U(words, ==, 2);
586 *block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
587 SM_DEBUG_ACTION_ENCODE(0) |
588 SM_DEBUG_SYNCPASS_ENCODE(0) |
589 SM_DEBUG_TXG_ENCODE(0);
591 sm->sm_phys->smp_objsize += sizeof (uint64_t);
592 ASSERT3P(block_cursor, ==, block_end);
596 uint64_t run_len = MIN(size, run_max);
599 *block_cursor = SM_OFFSET_ENCODE(start) |
600 SM_TYPE_ENCODE(maptype) |
601 SM_RUN_ENCODE(run_len);
605 /* write the first word of the entry */
606 *block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
607 SM2_RUN_ENCODE(run_len) |
608 SM2_VDEV_ENCODE(vdev_id);
611 /* move on to the second word of the entry */
612 ASSERT3P(block_cursor, <, block_end);
613 *block_cursor = SM2_TYPE_ENCODE(maptype) |
614 SM2_OFFSET_ENCODE(start);
618 panic("%d-word space map entries are not supported",
622 sm->sm_phys->smp_objsize += words * sizeof (uint64_t);
632 * Note: The space map's dbuf must be dirty for the changes in sm_phys to
636 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
637 uint64_t vdev_id, dmu_tx_t *tx)
639 spa_t *spa = tx->tx_pool->dp_spa;
642 space_map_write_intro_debug(sm, maptype, tx);
646 * We do this right after we write the intro debug entry
647 * because the estimate does not take it into account.
649 uint64_t initial_objsize = sm->sm_phys->smp_objsize;
650 uint64_t estimated_growth =
651 space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
652 uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
656 * Find the offset right after the last word in the space map
657 * and use that to get a hold of the last block, so we can
658 * start appending to it.
660 uint64_t next_word_offset = sm->sm_phys->smp_objsize;
661 VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
662 next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
663 ASSERT3U(db->db_size, ==, sm->sm_blksz);
665 dmu_buf_will_dirty(db, tx);
667 avl_tree_t *t = &rt->rt_root;
668 for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
669 uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
670 uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
674 * We only write two-word entries when both of the following
677 * [1] The feature is enabled.
678 * [2] The offset or run is too big for a single-word entry,
679 * or the vdev_id is set (meaning not equal to
682 * Note that for purposes of testing we've added the case that
683 * we write two-word entries occasionally when the feature is
684 * enabled and zfs_force_some_double_word_sm_entries has been
687 if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
688 (offset >= (1ULL << SM_OFFSET_BITS) ||
689 length > SM_RUN_MAX ||
690 vdev_id != SM_NO_VDEVID ||
691 (zfs_force_some_double_word_sm_entries &&
692 spa_get_random(100) == 0)))
695 space_map_write_seg(sm, rs, maptype, vdev_id, words,
699 dmu_buf_rele(db, FTAG);
703 * We expect our estimation to be based on the worst case
704 * scenario [see comment in space_map_estimate_optimal_size()].
705 * Therefore we expect the actual objsize to be equal or less
706 * than whatever we estimated it to be.
708 ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_objsize);
713 * Note: This function manipulates the state of the given space map but
714 * does not hold any locks implicitly. Thus the caller is responsible
715 * for synchronizing writes to the space map.
718 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
719 uint64_t vdev_id, dmu_tx_t *tx)
721 objset_t *os = sm->sm_os;
723 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
724 VERIFY3U(space_map_object(sm), !=, 0);
726 dmu_buf_will_dirty(sm->sm_dbuf, tx);
729 * This field is no longer necessary since the in-core space map
730 * now contains the object number but is maintained for backwards
733 sm->sm_phys->smp_object = sm->sm_object;
735 if (range_tree_is_empty(rt)) {
736 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
740 if (maptype == SM_ALLOC)
741 sm->sm_phys->smp_alloc += range_tree_space(rt);
743 sm->sm_phys->smp_alloc -= range_tree_space(rt);
745 uint64_t nodes = avl_numnodes(&rt->rt_root);
746 uint64_t rt_space = range_tree_space(rt);
748 space_map_write_impl(sm, rt, maptype, vdev_id, tx);
751 * Ensure that the space_map's accounting wasn't changed
752 * while we were in the middle of writing it out.
754 VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
755 VERIFY3U(range_tree_space(rt), ==, rt_space);
759 space_map_open_impl(space_map_t *sm)
764 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
768 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
769 sm->sm_phys = sm->sm_dbuf->db_data;
774 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
775 uint64_t start, uint64_t size, uint8_t shift)
780 ASSERT(*smp == NULL);
784 sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
786 sm->sm_start = start;
788 sm->sm_shift = shift;
790 sm->sm_object = object;
792 error = space_map_open_impl(sm);
803 space_map_close(space_map_t *sm)
808 if (sm->sm_dbuf != NULL)
809 dmu_buf_rele(sm->sm_dbuf, sm);
813 kmem_free(sm, sizeof (*sm));
817 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
819 objset_t *os = sm->sm_os;
820 spa_t *spa = dmu_objset_spa(os);
821 dmu_object_info_t doi;
823 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
824 ASSERT(dmu_tx_is_syncing(tx));
825 VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
827 dmu_object_info_from_db(sm->sm_dbuf, &doi);
830 * If the space map has the wrong bonus size (because
831 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
832 * the wrong block size (because space_map_blksz has changed),
833 * free and re-allocate its object with the updated sizes.
835 * Otherwise, just truncate the current object.
837 if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
838 doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
839 doi.doi_data_block_size != blocksize) {
840 zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
841 "object[%llu]: old bonus %u, old blocksz %u",
842 dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
843 doi.doi_bonus_size, doi.doi_data_block_size);
845 space_map_free(sm, tx);
846 dmu_buf_rele(sm->sm_dbuf, sm);
848 sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
849 VERIFY0(space_map_open_impl(sm));
851 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
854 * If the spacemap is reallocated, its histogram
855 * will be reset. Do the same in the common case so that
856 * bugs related to the uncommon case do not go unnoticed.
858 bzero(sm->sm_phys->smp_histogram,
859 sizeof (sm->sm_phys->smp_histogram));
862 dmu_buf_will_dirty(sm->sm_dbuf, tx);
863 sm->sm_phys->smp_objsize = 0;
864 sm->sm_phys->smp_alloc = 0;
868 * Update the in-core space_map allocation and length values.
871 space_map_update(space_map_t *sm)
876 sm->sm_alloc = sm->sm_phys->smp_alloc;
877 sm->sm_length = sm->sm_phys->smp_objsize;
881 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
883 spa_t *spa = dmu_objset_spa(os);
887 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
888 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
889 bonuslen = sizeof (space_map_phys_t);
890 ASSERT3U(bonuslen, <=, dmu_bonus_max());
892 bonuslen = SPACE_MAP_SIZE_V0;
895 object = dmu_object_alloc(os, DMU_OT_SPACE_MAP, blocksize,
896 DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
902 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx)
904 spa_t *spa = dmu_objset_spa(os);
905 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
906 dmu_object_info_t doi;
908 VERIFY0(dmu_object_info(os, smobj, &doi));
909 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
910 spa_feature_decr(spa,
911 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
915 VERIFY0(dmu_object_free(os, smobj, tx));
919 space_map_free(space_map_t *sm, dmu_tx_t *tx)
924 space_map_free_obj(sm->sm_os, space_map_object(sm), tx);
929 * Given a range tree, it makes a worst-case estimate of how much
930 * space would the tree's segments take if they were written to
931 * the given space map.
934 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
937 spa_t *spa = dmu_objset_spa(sm->sm_os);
938 uint64_t shift = sm->sm_shift;
939 uint64_t *histogram = rt->rt_histogram;
940 uint64_t entries_for_seg = 0;
943 * In order to get a quick estimate of the optimal size that this
944 * range tree would have on-disk as a space map, we iterate through
945 * its histogram buckets instead of iterating through its nodes.
947 * Note that this is a highest-bound/worst-case estimate for the
950 * 1] We assume that we always add a debug padding for each block
951 * we write and we also assume that we start at the last word
952 * of a block attempting to write a two-word entry.
953 * 2] Rounding up errors due to the way segments are distributed
954 * in the buckets of the range tree's histogram.
955 * 3] The activation of zfs_force_some_double_word_sm_entries
956 * (tunable) when testing.
958 * = Math and Rounding Errors =
960 * rt_histogram[i] bucket of a range tree represents the number
961 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
962 * that, we want to divide the buckets into groups: Buckets that
963 * can be represented using a single-word entry, ones that can
964 * be represented with a double-word entry, and ones that can
965 * only be represented with multiple two-word entries.
967 * [Note that if the new encoding feature is not enabled there
968 * are only two groups: single-word entry buckets and multiple
969 * single-word entry buckets. The information below assumes
970 * two-word entries enabled, but it can easily applied when
971 * the feature is not enabled]
973 * To find the highest bucket that can be represented with a
974 * single-word entry we look at the maximum run that such entry
975 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
976 * the run of a space map entry is shifted by sm_shift, thus we
977 * add it to the exponent]. This way, excluding the value of the
978 * maximum run that can be represented by a single-word entry,
979 * all runs that are smaller exist in buckets 0 to
980 * SM_RUN_BITS + shift - 1.
982 * To find the highest bucket that can be represented with a
983 * double-word entry, we follow the same approach. Finally, any
984 * bucket higher than that are represented with multiple two-word
985 * entries. To be more specific, if the highest bucket whose
986 * segments can be represented with a single two-word entry is X,
987 * then bucket X+1 will need 2 two-word entries for each of its
988 * segments, X+2 will need 4, X+3 will need 8, ...etc.
990 * With all of the above we make our estimation based on bucket
991 * groups. There is a rounding error though. As we mentioned in
992 * the example with the one-word entry, the maximum run that can
993 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
994 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
995 * that length fall into the next bucket (and bucket group) where
996 * we start counting two-word entries and this is one more reason
997 * why the estimated size may end up being bigger than the actual
1003 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
1004 (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
1007 * If we are trying to force some double word entries just
1008 * assume the worst-case of every single word entry being
1009 * written as a double word entry.
1011 uint64_t entry_size =
1012 (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
1013 zfs_force_some_double_word_sm_entries) ?
1014 (2 * sizeof (uint64_t)) : sizeof (uint64_t);
1016 uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
1017 for (; idx <= single_entry_max_bucket; idx++)
1018 size += histogram[idx] * entry_size;
1020 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
1021 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1022 ASSERT3U(idx, >=, single_entry_max_bucket);
1024 1ULL << (idx - single_entry_max_bucket);
1025 size += histogram[idx] *
1026 entries_for_seg * entry_size;
1032 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
1034 uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
1035 for (; idx <= double_entry_max_bucket; idx++)
1036 size += histogram[idx] * 2 * sizeof (uint64_t);
1038 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1039 ASSERT3U(idx, >=, double_entry_max_bucket);
1040 entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
1041 size += histogram[idx] *
1042 entries_for_seg * 2 * sizeof (uint64_t);
1046 * Assume the worst case where we start with the padding at the end
1047 * of the current block and we add an extra padding entry at the end
1048 * of all subsequent blocks.
1050 size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
1056 space_map_object(space_map_t *sm)
1058 return (sm != NULL ? sm->sm_object : 0);
1062 * Returns the already synced, on-disk allocated space.
1065 space_map_allocated(space_map_t *sm)
1067 return (sm != NULL ? sm->sm_alloc : 0);
1071 * Returns the already synced, on-disk length;
1074 space_map_length(space_map_t *sm)
1076 return (sm != NULL ? sm->sm_length : 0);
1080 * Returns the allocated space that is currently syncing.
1083 space_map_alloc_delta(space_map_t *sm)
1087 ASSERT(sm->sm_dbuf != NULL);
1088 return (sm->sm_phys->smp_alloc - space_map_allocated(sm));