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 *ss_before, *ss_after, *ss;
106 uint64_t end = start + size;
107 int merge_before, merge_after;
109 ASSERT(MUTEX_HELD(sm->sm_lock));
110 VERIFY(!sm->sm_condensing);
112 VERIFY3U(start, >=, sm->sm_start);
113 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
114 VERIFY(sm->sm_space + size <= sm->sm_size);
115 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
116 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
118 ss = space_map_find(sm, start, size, &where);
120 zfs_panic_recover("zfs: allocating allocated segment"
121 "(offset=%llu size=%llu)\n",
122 (longlong_t)start, (longlong_t)size);
126 /* Make sure we don't overlap with either of our neighbors */
129 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
130 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
132 merge_before = (ss_before != NULL && ss_before->ss_end == start);
133 merge_after = (ss_after != NULL && ss_after->ss_start == end);
135 if (merge_before && merge_after) {
136 avl_remove(&sm->sm_root, ss_before);
137 if (sm->sm_pp_root) {
138 avl_remove(sm->sm_pp_root, ss_before);
139 avl_remove(sm->sm_pp_root, ss_after);
141 ss_after->ss_start = ss_before->ss_start;
142 kmem_cache_free(space_seg_cache, ss_before);
144 } else if (merge_before) {
145 ss_before->ss_end = end;
147 avl_remove(sm->sm_pp_root, ss_before);
149 } else if (merge_after) {
150 ss_after->ss_start = start;
152 avl_remove(sm->sm_pp_root, ss_after);
155 ss = kmem_cache_alloc(space_seg_cache, KM_SLEEP);
156 ss->ss_start = start;
158 avl_insert(&sm->sm_root, ss, where);
162 avl_add(sm->sm_pp_root, ss);
164 sm->sm_space += size;
168 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
173 space_seg_t *ss, *newseg;
174 uint64_t end = start + size;
175 int left_over, right_over;
177 VERIFY(!sm->sm_condensing);
179 ss = space_map_find(sm, start, size, &where);
181 ss = space_map_find(sm, start, size, NULL);
184 /* Make sure we completely overlap with someone */
186 zfs_panic_recover("zfs: freeing free segment "
187 "(offset=%llu size=%llu)",
188 (longlong_t)start, (longlong_t)size);
191 VERIFY3U(ss->ss_start, <=, start);
192 VERIFY3U(ss->ss_end, >=, end);
193 VERIFY(sm->sm_space - size < sm->sm_size);
195 left_over = (ss->ss_start != start);
196 right_over = (ss->ss_end != end);
199 avl_remove(sm->sm_pp_root, ss);
201 if (left_over && right_over) {
202 newseg = kmem_cache_alloc(space_seg_cache, KM_SLEEP);
203 newseg->ss_start = end;
204 newseg->ss_end = ss->ss_end;
206 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
208 avl_add(sm->sm_pp_root, newseg);
209 } else if (left_over) {
211 } else if (right_over) {
214 avl_remove(&sm->sm_root, ss);
215 kmem_cache_free(space_seg_cache, ss);
219 if (sm->sm_pp_root && ss != NULL)
220 avl_add(sm->sm_pp_root, ss);
222 sm->sm_space -= size;
226 space_map_find(space_map_t *sm, uint64_t start, uint64_t size,
229 space_seg_t ssearch, *ss;
231 ASSERT(MUTEX_HELD(sm->sm_lock));
233 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
234 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
236 ssearch.ss_start = start;
237 ssearch.ss_end = start + size;
238 ss = avl_find(&sm->sm_root, &ssearch, wherep);
240 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= start + size)
246 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
250 return (space_map_find(sm, start, size, &where) != 0);
254 space_map_swap(space_map_t **msrc, space_map_t **mdst)
258 ASSERT(MUTEX_HELD((*msrc)->sm_lock));
259 ASSERT0((*mdst)->sm_space);
260 ASSERT0(avl_numnodes(&(*mdst)->sm_root));
268 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
273 ASSERT(MUTEX_HELD(sm->sm_lock));
275 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
277 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
278 kmem_cache_free(space_seg_cache, ss);
284 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
288 ASSERT(MUTEX_HELD(sm->sm_lock));
290 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
291 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
295 * Wait for any in-progress space_map_load() to complete.
298 space_map_load_wait(space_map_t *sm)
300 ASSERT(MUTEX_HELD(sm->sm_lock));
302 while (sm->sm_loading) {
303 ASSERT(!sm->sm_loaded);
304 cv_wait(&sm->sm_load_cv, sm->sm_lock);
309 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
310 * The caller must be OK with this.
313 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
314 space_map_obj_t *smo, objset_t *os)
316 uint64_t *entry, *entry_map, *entry_map_end;
317 uint64_t bufsize, size, offset, end, space;
318 uint64_t mapstart = sm->sm_start;
321 ASSERT(MUTEX_HELD(sm->sm_lock));
322 ASSERT(!sm->sm_loaded);
323 ASSERT(!sm->sm_loading);
325 sm->sm_loading = B_TRUE;
326 end = smo->smo_objsize;
327 space = smo->smo_alloc;
329 ASSERT(sm->sm_ops == NULL);
330 VERIFY0(sm->sm_space);
332 if (maptype == SM_FREE) {
333 space_map_add(sm, sm->sm_start, sm->sm_size);
334 space = sm->sm_size - space;
337 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
338 entry_map = zio_buf_alloc(bufsize);
340 mutex_exit(sm->sm_lock);
342 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
343 mutex_enter(sm->sm_lock);
345 for (offset = 0; offset < end; offset += bufsize) {
346 size = MIN(end - offset, bufsize);
347 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
350 dprintf("object=%llu offset=%llx size=%llx\n",
351 smo->smo_object, offset, size);
353 mutex_exit(sm->sm_lock);
354 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
356 mutex_enter(sm->sm_lock);
360 entry_map_end = entry_map + (size / sizeof (uint64_t));
361 for (entry = entry_map; entry < entry_map_end; entry++) {
364 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
367 (SM_TYPE_DECODE(e) == maptype ?
368 space_map_add : space_map_remove)(sm,
369 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
370 SM_RUN_DECODE(e) << sm->sm_shift);
375 VERIFY3U(sm->sm_space, ==, space);
377 sm->sm_loaded = B_TRUE;
382 space_map_vacate(sm, NULL, NULL);
385 zio_buf_free(entry_map, bufsize);
387 sm->sm_loading = B_FALSE;
389 cv_broadcast(&sm->sm_load_cv);
395 space_map_unload(space_map_t *sm)
397 ASSERT(MUTEX_HELD(sm->sm_lock));
399 if (sm->sm_loaded && sm->sm_ops != NULL)
400 sm->sm_ops->smop_unload(sm);
402 sm->sm_loaded = B_FALSE;
405 space_map_vacate(sm, NULL, NULL);
409 space_map_maxsize(space_map_t *sm)
411 ASSERT(sm->sm_ops != NULL);
412 return (sm->sm_ops->smop_max(sm));
416 space_map_alloc(space_map_t *sm, uint64_t size)
420 start = sm->sm_ops->smop_alloc(sm, size);
422 space_map_remove(sm, start, size);
427 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
429 sm->sm_ops->smop_claim(sm, start, size);
430 space_map_remove(sm, start, size);
434 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
436 space_map_add(sm, start, size);
437 sm->sm_ops->smop_free(sm, start, size);
441 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
444 space_map_sync(space_map_t *sm, uint8_t maptype,
445 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
447 spa_t *spa = dmu_objset_spa(os);
448 avl_tree_t *t = &sm->sm_root;
450 uint64_t bufsize, start, size, run_len, total, sm_space, nodes;
451 uint64_t *entry, *entry_map, *entry_map_end;
453 ASSERT(MUTEX_HELD(sm->sm_lock));
455 if (sm->sm_space == 0)
458 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
459 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
460 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
463 if (maptype == SM_ALLOC)
464 smo->smo_alloc += sm->sm_space;
466 smo->smo_alloc -= sm->sm_space;
468 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
469 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
470 entry_map = zio_buf_alloc(bufsize);
471 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
474 *entry++ = SM_DEBUG_ENCODE(1) |
475 SM_DEBUG_ACTION_ENCODE(maptype) |
476 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
477 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
480 nodes = avl_numnodes(&sm->sm_root);
481 sm_space = sm->sm_space;
482 for (ss = avl_first(t); ss != NULL; ss = AVL_NEXT(t, ss)) {
483 size = ss->ss_end - ss->ss_start;
484 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
487 size >>= sm->sm_shift;
490 run_len = MIN(size, SM_RUN_MAX);
492 if (entry == entry_map_end) {
493 mutex_exit(sm->sm_lock);
494 dmu_write(os, smo->smo_object, smo->smo_objsize,
495 bufsize, entry_map, tx);
496 mutex_enter(sm->sm_lock);
497 smo->smo_objsize += bufsize;
501 *entry++ = SM_OFFSET_ENCODE(start) |
502 SM_TYPE_ENCODE(maptype) |
503 SM_RUN_ENCODE(run_len);
510 if (entry != entry_map) {
511 size = (entry - entry_map) * sizeof (uint64_t);
512 mutex_exit(sm->sm_lock);
513 dmu_write(os, smo->smo_object, smo->smo_objsize,
514 size, entry_map, tx);
515 mutex_enter(sm->sm_lock);
516 smo->smo_objsize += size;
520 * Ensure that the space_map's accounting wasn't changed
521 * while we were in the middle of writing it out.
523 VERIFY3U(nodes, ==, avl_numnodes(&sm->sm_root));
524 VERIFY3U(sm->sm_space, ==, sm_space);
525 VERIFY3U(sm->sm_space, ==, total);
527 zio_buf_free(entry_map, bufsize);
531 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
533 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
535 smo->smo_objsize = 0;
540 * Space map reference trees.
542 * A space map is a collection of integers. Every integer is either
543 * in the map, or it's not. A space map reference tree generalizes
544 * the idea: it allows its members to have arbitrary reference counts,
545 * as opposed to the implicit reference count of 0 or 1 in a space map.
546 * This representation comes in handy when computing the union or
547 * intersection of multiple space maps. For example, the union of
548 * N space maps is the subset of the reference tree with refcnt >= 1.
549 * The intersection of N space maps is the subset with refcnt >= N.
551 * [It's very much like a Fourier transform. Unions and intersections
552 * are hard to perform in the 'space map domain', so we convert the maps
553 * into the 'reference count domain', where it's trivial, then invert.]
555 * vdev_dtl_reassess() uses computations of this form to determine
556 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
557 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
558 * has an outage wherever refcnt >= vdev_children.
561 space_map_ref_compare(const void *x1, const void *x2)
563 const space_ref_t *sr1 = x1;
564 const space_ref_t *sr2 = x2;
566 if (sr1->sr_offset < sr2->sr_offset)
568 if (sr1->sr_offset > sr2->sr_offset)
580 space_map_ref_create(avl_tree_t *t)
582 avl_create(t, space_map_ref_compare,
583 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
587 space_map_ref_destroy(avl_tree_t *t)
592 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
593 kmem_free(sr, sizeof (*sr));
599 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
603 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
604 sr->sr_offset = offset;
605 sr->sr_refcnt = refcnt;
611 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
614 space_map_ref_add_node(t, start, refcnt);
615 space_map_ref_add_node(t, end, -refcnt);
619 * Convert (or add) a space map into a reference tree.
622 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
626 ASSERT(MUTEX_HELD(sm->sm_lock));
628 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
629 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
633 * Convert a reference tree into a space map. The space map will contain
634 * all members of the reference tree for which refcnt >= minref.
637 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
639 uint64_t start = -1ULL;
643 ASSERT(MUTEX_HELD(sm->sm_lock));
645 space_map_vacate(sm, NULL, NULL);
647 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
648 refcnt += sr->sr_refcnt;
649 if (refcnt >= minref) {
650 if (start == -1ULL) {
651 start = sr->sr_offset;
654 if (start != -1ULL) {
655 uint64_t end = sr->sr_offset;
656 ASSERT(start <= end);
658 space_map_add(sm, start, end - start);
664 ASSERT(start == -1ULL);