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, 2014 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 * 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.
48 int space_map_blksz = (1 << 12);
49 SYSCTL_INT(_vfs_zfs, OID_AUTO, space_map_blksz, CTLFLAG_RDTUN, &space_map_blksz, 0,
50 "Maximum block size for space map. Must be power of 2 and greater than 4096.");
53 * Load the space map disk into the specified range tree. Segments of maptype
54 * are added to the range tree, other segment types are removed.
56 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
57 * The caller must be OK with this.
60 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
62 uint64_t *entry, *entry_map, *entry_map_end;
63 uint64_t bufsize, size, offset, end, space;
66 ASSERT(MUTEX_HELD(sm->sm_lock));
68 end = space_map_length(sm);
69 space = space_map_allocated(sm);
71 VERIFY0(range_tree_space(rt));
73 if (maptype == SM_FREE) {
74 range_tree_add(rt, sm->sm_start, sm->sm_size);
75 space = sm->sm_size - space;
78 bufsize = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
79 entry_map = zio_buf_alloc(bufsize);
81 mutex_exit(sm->sm_lock);
83 dmu_prefetch(sm->sm_os, space_map_object(sm), bufsize,
86 mutex_enter(sm->sm_lock);
88 for (offset = 0; offset < end; offset += bufsize) {
89 size = MIN(end - offset, bufsize);
90 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
92 ASSERT3U(sm->sm_blksz, !=, 0);
94 dprintf("object=%llu offset=%llx size=%llx\n",
95 space_map_object(sm), offset, size);
97 mutex_exit(sm->sm_lock);
98 error = dmu_read(sm->sm_os, space_map_object(sm), offset, size,
99 entry_map, DMU_READ_PREFETCH);
100 mutex_enter(sm->sm_lock);
104 entry_map_end = entry_map + (size / sizeof (uint64_t));
105 for (entry = entry_map; entry < entry_map_end; entry++) {
107 uint64_t offset, size;
109 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
112 offset = (SM_OFFSET_DECODE(e) << sm->sm_shift) +
114 size = SM_RUN_DECODE(e) << sm->sm_shift;
116 VERIFY0(P2PHASE(offset, 1ULL << sm->sm_shift));
117 VERIFY0(P2PHASE(size, 1ULL << sm->sm_shift));
118 VERIFY3U(offset, >=, sm->sm_start);
119 VERIFY3U(offset + size, <=, sm->sm_start + sm->sm_size);
120 if (SM_TYPE_DECODE(e) == maptype) {
121 VERIFY3U(range_tree_space(rt) + size, <=,
123 range_tree_add(rt, offset, size);
125 range_tree_remove(rt, offset, size);
131 VERIFY3U(range_tree_space(rt), ==, space);
133 range_tree_vacate(rt, NULL, NULL);
135 zio_buf_free(entry_map, bufsize);
140 space_map_histogram_clear(space_map_t *sm)
142 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
145 bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
149 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
152 * Verify that the in-core range tree does not have any
153 * ranges smaller than our sm_shift size.
155 for (int i = 0; i < sm->sm_shift; i++) {
156 if (rt->rt_histogram[i] != 0)
163 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
167 ASSERT(MUTEX_HELD(rt->rt_lock));
168 ASSERT(dmu_tx_is_syncing(tx));
169 VERIFY3U(space_map_object(sm), !=, 0);
171 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
174 dmu_buf_will_dirty(sm->sm_dbuf, tx);
176 ASSERT(space_map_histogram_verify(sm, rt));
179 * Transfer the content of the range tree histogram to the space
180 * map histogram. The space map histogram contains 32 buckets ranging
181 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
182 * however, can represent ranges from 2^0 to 2^63. Since the space
183 * map only cares about allocatable blocks (minimum of sm_shift) we
184 * can safely ignore all ranges in the range tree smaller than sm_shift.
186 for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
189 * Since the largest histogram bucket in the space map is
190 * 2^(32+sm_shift-1), we need to normalize the values in
191 * the range tree for any bucket larger than that size. For
192 * example given an sm_shift of 9, ranges larger than 2^40
193 * would get normalized as if they were 1TB ranges. Assume
194 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
195 * the calculation below would normalize this to 5 * 2^4 (16).
197 ASSERT3U(i, >=, idx + sm->sm_shift);
198 sm->sm_phys->smp_histogram[idx] +=
199 rt->rt_histogram[i] << (i - idx - sm->sm_shift);
202 * Increment the space map's index as long as we haven't
203 * reached the maximum bucket size. Accumulate all ranges
204 * larger than the max bucket size into the last bucket.
206 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
207 ASSERT3U(idx + sm->sm_shift, ==, i);
209 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
215 space_map_entries(space_map_t *sm, range_tree_t *rt)
217 avl_tree_t *t = &rt->rt_root;
219 uint64_t size, entries;
222 * All space_maps always have a debug entry so account for it here.
227 * Traverse the range tree and calculate the number of space map
228 * entries that would be required to write out the range tree.
230 for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
231 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
232 entries += howmany(size, SM_RUN_MAX);
238 * Note: space_map_write() will drop sm_lock across dmu_write() calls.
241 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
244 objset_t *os = sm->sm_os;
245 spa_t *spa = dmu_objset_spa(os);
246 avl_tree_t *t = &rt->rt_root;
248 uint64_t size, total, rt_space, nodes;
249 uint64_t *entry, *entry_map, *entry_map_end;
250 uint64_t expected_entries, actual_entries = 1;
252 ASSERT(MUTEX_HELD(rt->rt_lock));
253 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
254 VERIFY3U(space_map_object(sm), !=, 0);
255 dmu_buf_will_dirty(sm->sm_dbuf, tx);
258 * This field is no longer necessary since the in-core space map
259 * now contains the object number but is maintained for backwards
262 sm->sm_phys->smp_object = sm->sm_object;
264 if (range_tree_space(rt) == 0) {
265 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
269 if (maptype == SM_ALLOC)
270 sm->sm_phys->smp_alloc += range_tree_space(rt);
272 sm->sm_phys->smp_alloc -= range_tree_space(rt);
274 expected_entries = space_map_entries(sm, rt);
276 entry_map = zio_buf_alloc(sm->sm_blksz);
277 entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
280 *entry++ = SM_DEBUG_ENCODE(1) |
281 SM_DEBUG_ACTION_ENCODE(maptype) |
282 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
283 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
286 nodes = avl_numnodes(&rt->rt_root);
287 rt_space = range_tree_space(rt);
288 for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
291 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
292 start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
294 total += size << sm->sm_shift;
299 run_len = MIN(size, SM_RUN_MAX);
301 if (entry == entry_map_end) {
302 mutex_exit(rt->rt_lock);
303 dmu_write(os, space_map_object(sm),
304 sm->sm_phys->smp_objsize, sm->sm_blksz,
306 mutex_enter(rt->rt_lock);
307 sm->sm_phys->smp_objsize += sm->sm_blksz;
311 *entry++ = SM_OFFSET_ENCODE(start) |
312 SM_TYPE_ENCODE(maptype) |
313 SM_RUN_ENCODE(run_len);
321 if (entry != entry_map) {
322 size = (entry - entry_map) * sizeof (uint64_t);
323 mutex_exit(rt->rt_lock);
324 dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
325 size, entry_map, tx);
326 mutex_enter(rt->rt_lock);
327 sm->sm_phys->smp_objsize += size;
329 ASSERT3U(expected_entries, ==, actual_entries);
332 * Ensure that the space_map's accounting wasn't changed
333 * while we were in the middle of writing it out.
335 VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
336 VERIFY3U(range_tree_space(rt), ==, rt_space);
337 VERIFY3U(range_tree_space(rt), ==, total);
339 zio_buf_free(entry_map, sm->sm_blksz);
343 space_map_open_impl(space_map_t *sm)
348 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
352 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
353 sm->sm_phys = sm->sm_dbuf->db_data;
358 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
359 uint64_t start, uint64_t size, uint8_t shift, kmutex_t *lp)
364 ASSERT(*smp == NULL);
368 sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
370 sm->sm_start = start;
372 sm->sm_shift = shift;
375 sm->sm_object = object;
377 error = space_map_open_impl(sm);
389 space_map_close(space_map_t *sm)
394 if (sm->sm_dbuf != NULL)
395 dmu_buf_rele(sm->sm_dbuf, sm);
399 kmem_free(sm, sizeof (*sm));
403 space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
405 objset_t *os = sm->sm_os;
406 spa_t *spa = dmu_objset_spa(os);
407 dmu_object_info_t doi;
409 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
410 ASSERT(dmu_tx_is_syncing(tx));
412 dmu_object_info_from_db(sm->sm_dbuf, &doi);
415 * If the space map has the wrong bonus size (because
416 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
417 * the wrong block size (because space_map_blksz has changed),
418 * free and re-allocate its object with the updated sizes.
420 * Otherwise, just truncate the current object.
422 if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
423 doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
424 doi.doi_data_block_size != space_map_blksz) {
425 zfs_dbgmsg("txg %llu, spa %s, reallocating: "
426 "old bonus %u, old blocksz %u", dmu_tx_get_txg(tx),
427 spa_name(spa), doi.doi_bonus_size, doi.doi_data_block_size);
429 space_map_free(sm, tx);
430 dmu_buf_rele(sm->sm_dbuf, sm);
432 sm->sm_object = space_map_alloc(sm->sm_os, tx);
433 VERIFY0(space_map_open_impl(sm));
435 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
438 * If the spacemap is reallocated, its histogram
439 * will be reset. Do the same in the common case so that
440 * bugs related to the uncommon case do not go unnoticed.
442 bzero(sm->sm_phys->smp_histogram,
443 sizeof (sm->sm_phys->smp_histogram));
446 dmu_buf_will_dirty(sm->sm_dbuf, tx);
447 sm->sm_phys->smp_objsize = 0;
448 sm->sm_phys->smp_alloc = 0;
452 * Update the in-core space_map allocation and length values.
455 space_map_update(space_map_t *sm)
460 ASSERT(MUTEX_HELD(sm->sm_lock));
462 sm->sm_alloc = sm->sm_phys->smp_alloc;
463 sm->sm_length = sm->sm_phys->smp_objsize;
467 space_map_alloc(objset_t *os, dmu_tx_t *tx)
469 spa_t *spa = dmu_objset_spa(os);
473 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
474 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
475 bonuslen = sizeof (space_map_phys_t);
476 ASSERT3U(bonuslen, <=, dmu_bonus_max());
478 bonuslen = SPACE_MAP_SIZE_V0;
481 object = dmu_object_alloc(os,
482 DMU_OT_SPACE_MAP, space_map_blksz,
483 DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
489 space_map_free(space_map_t *sm, dmu_tx_t *tx)
496 spa = dmu_objset_spa(sm->sm_os);
497 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
498 dmu_object_info_t doi;
500 dmu_object_info_from_db(sm->sm_dbuf, &doi);
501 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
502 VERIFY(spa_feature_is_active(spa,
503 SPA_FEATURE_SPACEMAP_HISTOGRAM));
504 spa_feature_decr(spa,
505 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
509 VERIFY3U(dmu_object_free(sm->sm_os, space_map_object(sm), tx), ==, 0);
514 space_map_object(space_map_t *sm)
516 return (sm != NULL ? sm->sm_object : 0);
520 * Returns the already synced, on-disk allocated space.
523 space_map_allocated(space_map_t *sm)
525 return (sm != NULL ? sm->sm_alloc : 0);
529 * Returns the already synced, on-disk length;
532 space_map_length(space_map_t *sm)
534 return (sm != NULL ? sm->sm_length : 0);
538 * Returns the allocated space that is currently syncing.
541 space_map_alloc_delta(space_map_t *sm)
545 ASSERT(sm->sm_dbuf != NULL);
546 return (sm->sm_phys->smp_alloc - space_map_allocated(sm));