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 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
33 #include <sys/space_map.h>
35 static kmem_cache_t *space_seg_cache;
40 ASSERT(space_seg_cache == NULL);
41 space_seg_cache = kmem_cache_create("space_seg_cache",
42 sizeof (space_seg_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
48 kmem_cache_destroy(space_seg_cache);
49 space_seg_cache = NULL;
54 * NOTE: caller is responsible for all locking.
57 space_map_seg_compare(const void *x1, const void *x2)
59 const space_seg_t *s1 = x1;
60 const space_seg_t *s2 = x2;
62 if (s1->ss_start < s2->ss_start) {
63 if (s1->ss_end > s2->ss_start)
67 if (s1->ss_start > s2->ss_start) {
68 if (s1->ss_start < s2->ss_end)
76 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
79 bzero(sm, sizeof (*sm));
81 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
83 avl_create(&sm->sm_root, space_map_seg_compare,
84 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
93 space_map_destroy(space_map_t *sm)
95 ASSERT(!sm->sm_loaded && !sm->sm_loading);
96 VERIFY0(sm->sm_space);
97 avl_destroy(&sm->sm_root);
98 cv_destroy(&sm->sm_load_cv);
102 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
105 space_seg_t ssearch, *ss_before, *ss_after, *ss;
106 uint64_t end = start + size;
107 int merge_before, merge_after;
109 ASSERT(MUTEX_HELD(sm->sm_lock));
111 VERIFY3U(start, >=, sm->sm_start);
112 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
113 VERIFY(sm->sm_space + size <= sm->sm_size);
114 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
115 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
117 ssearch.ss_start = start;
118 ssearch.ss_end = end;
119 ss = avl_find(&sm->sm_root, &ssearch, &where);
121 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
122 zfs_panic_recover("zfs: allocating allocated segment"
123 "(offset=%llu size=%llu)\n",
124 (longlong_t)start, (longlong_t)size);
128 /* Make sure we don't overlap with either of our neighbors */
131 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
132 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
134 merge_before = (ss_before != NULL && ss_before->ss_end == start);
135 merge_after = (ss_after != NULL && ss_after->ss_start == end);
137 if (merge_before && merge_after) {
138 avl_remove(&sm->sm_root, ss_before);
139 if (sm->sm_pp_root) {
140 avl_remove(sm->sm_pp_root, ss_before);
141 avl_remove(sm->sm_pp_root, ss_after);
143 ss_after->ss_start = ss_before->ss_start;
144 kmem_cache_free(space_seg_cache, ss_before);
146 } else if (merge_before) {
147 ss_before->ss_end = end;
149 avl_remove(sm->sm_pp_root, ss_before);
151 } else if (merge_after) {
152 ss_after->ss_start = start;
154 avl_remove(sm->sm_pp_root, ss_after);
157 ss = kmem_cache_alloc(space_seg_cache, KM_PUSHPAGE);
158 ss->ss_start = start;
160 avl_insert(&sm->sm_root, ss, where);
164 avl_add(sm->sm_pp_root, ss);
166 sm->sm_space += size;
170 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
173 space_seg_t ssearch, *ss, *newseg;
174 uint64_t end = start + size;
175 int left_over, right_over;
177 ASSERT(MUTEX_HELD(sm->sm_lock));
179 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
180 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
182 ssearch.ss_start = start;
183 ssearch.ss_end = end;
184 ss = avl_find(&sm->sm_root, &ssearch, &where);
186 /* Make sure we completely overlap with someone */
188 zfs_panic_recover("zfs: freeing free segment "
189 "(offset=%llu size=%llu)",
190 (longlong_t)start, (longlong_t)size);
193 VERIFY3U(ss->ss_start, <=, start);
194 VERIFY3U(ss->ss_end, >=, end);
195 VERIFY(sm->sm_space - size <= sm->sm_size);
197 left_over = (ss->ss_start != start);
198 right_over = (ss->ss_end != end);
201 avl_remove(sm->sm_pp_root, ss);
203 if (left_over && right_over) {
204 newseg = kmem_cache_alloc(space_seg_cache, KM_PUSHPAGE);
205 newseg->ss_start = end;
206 newseg->ss_end = ss->ss_end;
208 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
210 avl_add(sm->sm_pp_root, newseg);
211 } else if (left_over) {
213 } else if (right_over) {
216 avl_remove(&sm->sm_root, ss);
217 kmem_cache_free(space_seg_cache, ss);
221 if (sm->sm_pp_root && ss != NULL)
222 avl_add(sm->sm_pp_root, ss);
224 sm->sm_space -= size;
228 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
231 space_seg_t ssearch, *ss;
232 uint64_t end = start + size;
234 ASSERT(MUTEX_HELD(sm->sm_lock));
236 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
237 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
239 ssearch.ss_start = start;
240 ssearch.ss_end = end;
241 ss = avl_find(&sm->sm_root, &ssearch, &where);
243 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
247 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
252 ASSERT(MUTEX_HELD(sm->sm_lock));
254 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
256 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
257 kmem_cache_free(space_seg_cache, ss);
263 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
267 ASSERT(MUTEX_HELD(sm->sm_lock));
269 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
270 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
274 * Wait for any in-progress space_map_load() to complete.
277 space_map_load_wait(space_map_t *sm)
279 ASSERT(MUTEX_HELD(sm->sm_lock));
281 while (sm->sm_loading) {
282 ASSERT(!sm->sm_loaded);
283 cv_wait(&sm->sm_load_cv, sm->sm_lock);
288 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
289 * The caller must be OK with this.
292 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
293 space_map_obj_t *smo, objset_t *os)
295 uint64_t *entry, *entry_map, *entry_map_end;
296 uint64_t bufsize, size, offset, end, space;
297 uint64_t mapstart = sm->sm_start;
300 ASSERT(MUTEX_HELD(sm->sm_lock));
301 ASSERT(!sm->sm_loaded);
302 ASSERT(!sm->sm_loading);
304 sm->sm_loading = B_TRUE;
305 end = smo->smo_objsize;
306 space = smo->smo_alloc;
308 ASSERT(sm->sm_ops == NULL);
309 VERIFY0(sm->sm_space);
311 if (maptype == SM_FREE) {
312 space_map_add(sm, sm->sm_start, sm->sm_size);
313 space = sm->sm_size - space;
316 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
317 entry_map = zio_buf_alloc(bufsize);
319 mutex_exit(sm->sm_lock);
321 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
322 mutex_enter(sm->sm_lock);
324 for (offset = 0; offset < end; offset += bufsize) {
325 size = MIN(end - offset, bufsize);
326 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
329 dprintf("object=%llu offset=%llx size=%llx\n",
330 smo->smo_object, offset, size);
332 mutex_exit(sm->sm_lock);
333 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
335 mutex_enter(sm->sm_lock);
339 entry_map_end = entry_map + (size / sizeof (uint64_t));
340 for (entry = entry_map; entry < entry_map_end; entry++) {
343 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
346 (SM_TYPE_DECODE(e) == maptype ?
347 space_map_add : space_map_remove)(sm,
348 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
349 SM_RUN_DECODE(e) << sm->sm_shift);
354 VERIFY3U(sm->sm_space, ==, space);
356 sm->sm_loaded = B_TRUE;
361 space_map_vacate(sm, NULL, NULL);
364 zio_buf_free(entry_map, bufsize);
366 sm->sm_loading = B_FALSE;
368 cv_broadcast(&sm->sm_load_cv);
374 space_map_unload(space_map_t *sm)
376 ASSERT(MUTEX_HELD(sm->sm_lock));
378 if (sm->sm_loaded && sm->sm_ops != NULL)
379 sm->sm_ops->smop_unload(sm);
381 sm->sm_loaded = B_FALSE;
384 space_map_vacate(sm, NULL, NULL);
388 space_map_maxsize(space_map_t *sm)
390 ASSERT(sm->sm_ops != NULL);
391 return (sm->sm_ops->smop_max(sm));
395 space_map_alloc(space_map_t *sm, uint64_t size)
399 start = sm->sm_ops->smop_alloc(sm, size);
401 space_map_remove(sm, start, size);
406 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
408 sm->sm_ops->smop_claim(sm, start, size);
409 space_map_remove(sm, start, size);
413 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
415 space_map_add(sm, start, size);
416 sm->sm_ops->smop_free(sm, start, size);
420 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
423 space_map_sync(space_map_t *sm, uint8_t maptype,
424 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
426 spa_t *spa = dmu_objset_spa(os);
429 uint64_t bufsize, start, size, run_len, delta, sm_space;
430 uint64_t *entry, *entry_map, *entry_map_end;
432 ASSERT(MUTEX_HELD(sm->sm_lock));
434 if (sm->sm_space == 0)
437 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
438 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
439 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
442 if (maptype == SM_ALLOC)
443 smo->smo_alloc += sm->sm_space;
445 smo->smo_alloc -= sm->sm_space;
447 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
448 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
449 entry_map = zio_buf_alloc(bufsize);
450 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
453 *entry++ = SM_DEBUG_ENCODE(1) |
454 SM_DEBUG_ACTION_ENCODE(maptype) |
455 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
456 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
459 sm_space = sm->sm_space;
460 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
461 size = ss->ss_end - ss->ss_start;
462 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
465 size >>= sm->sm_shift;
468 run_len = MIN(size, SM_RUN_MAX);
470 if (entry == entry_map_end) {
471 mutex_exit(sm->sm_lock);
472 dmu_write(os, smo->smo_object, smo->smo_objsize,
473 bufsize, entry_map, tx);
474 mutex_enter(sm->sm_lock);
475 smo->smo_objsize += bufsize;
479 *entry++ = SM_OFFSET_ENCODE(start) |
480 SM_TYPE_ENCODE(maptype) |
481 SM_RUN_ENCODE(run_len);
486 kmem_cache_free(space_seg_cache, ss);
489 if (entry != entry_map) {
490 size = (entry - entry_map) * sizeof (uint64_t);
491 mutex_exit(sm->sm_lock);
492 dmu_write(os, smo->smo_object, smo->smo_objsize,
493 size, entry_map, tx);
494 mutex_enter(sm->sm_lock);
495 smo->smo_objsize += size;
499 * Ensure that the space_map's accounting wasn't changed
500 * while we were in the middle of writing it out.
502 VERIFY3U(sm->sm_space, ==, sm_space);
504 zio_buf_free(entry_map, bufsize);
506 sm->sm_space -= delta;
507 VERIFY0(sm->sm_space);
511 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
513 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
515 smo->smo_objsize = 0;
520 * Space map reference trees.
522 * A space map is a collection of integers. Every integer is either
523 * in the map, or it's not. A space map reference tree generalizes
524 * the idea: it allows its members to have arbitrary reference counts,
525 * as opposed to the implicit reference count of 0 or 1 in a space map.
526 * This representation comes in handy when computing the union or
527 * intersection of multiple space maps. For example, the union of
528 * N space maps is the subset of the reference tree with refcnt >= 1.
529 * The intersection of N space maps is the subset with refcnt >= N.
531 * [It's very much like a Fourier transform. Unions and intersections
532 * are hard to perform in the 'space map domain', so we convert the maps
533 * into the 'reference count domain', where it's trivial, then invert.]
535 * vdev_dtl_reassess() uses computations of this form to determine
536 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
537 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
538 * has an outage wherever refcnt >= vdev_children.
541 space_map_ref_compare(const void *x1, const void *x2)
543 const space_ref_t *sr1 = x1;
544 const space_ref_t *sr2 = x2;
546 if (sr1->sr_offset < sr2->sr_offset)
548 if (sr1->sr_offset > sr2->sr_offset)
560 space_map_ref_create(avl_tree_t *t)
562 avl_create(t, space_map_ref_compare,
563 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
567 space_map_ref_destroy(avl_tree_t *t)
572 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
573 kmem_free(sr, sizeof (*sr));
579 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
583 sr = kmem_alloc(sizeof (*sr), KM_PUSHPAGE);
584 sr->sr_offset = offset;
585 sr->sr_refcnt = refcnt;
591 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
594 space_map_ref_add_node(t, start, refcnt);
595 space_map_ref_add_node(t, end, -refcnt);
599 * Convert (or add) a space map into a reference tree.
602 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
606 ASSERT(MUTEX_HELD(sm->sm_lock));
608 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
609 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
613 * Convert a reference tree into a space map. The space map will contain
614 * all members of the reference tree for which refcnt >= minref.
617 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
619 uint64_t start = -1ULL;
623 ASSERT(MUTEX_HELD(sm->sm_lock));
625 space_map_vacate(sm, NULL, NULL);
627 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
628 refcnt += sr->sr_refcnt;
629 if (refcnt >= minref) {
630 if (start == -1ULL) {
631 start = sr->sr_offset;
634 if (start != -1ULL) {
635 uint64_t end = sr->sr_offset;
636 ASSERT(start <= end);
638 space_map_add(sm, start, end - start);
644 ASSERT(start == -1ULL);