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_SLEEP);
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)
172 space_seg_t ssearch, *ss, *newseg;
173 uint64_t end = start + size;
174 int left_over, right_over;
176 ASSERT(MUTEX_HELD(sm->sm_lock));
178 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
179 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
181 ssearch.ss_start = start;
182 ssearch.ss_end = end;
183 ss = avl_find(&sm->sm_root, &ssearch, NULL);
185 /* Make sure we completely overlap with someone */
187 zfs_panic_recover("zfs: freeing free segment "
188 "(offset=%llu size=%llu)",
189 (longlong_t)start, (longlong_t)size);
192 VERIFY3U(ss->ss_start, <=, start);
193 VERIFY3U(ss->ss_end, >=, end);
194 VERIFY(sm->sm_space - size < sm->sm_size);
196 left_over = (ss->ss_start != start);
197 right_over = (ss->ss_end != end);
200 avl_remove(sm->sm_pp_root, ss);
202 if (left_over && right_over) {
203 newseg = kmem_cache_alloc(space_seg_cache, KM_SLEEP);
204 newseg->ss_start = end;
205 newseg->ss_end = ss->ss_end;
207 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
209 avl_add(sm->sm_pp_root, newseg);
210 } else if (left_over) {
212 } else if (right_over) {
215 avl_remove(&sm->sm_root, ss);
216 kmem_cache_free(space_seg_cache, ss);
220 if (sm->sm_pp_root && ss != NULL)
221 avl_add(sm->sm_pp_root, ss);
223 sm->sm_space -= size;
227 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
230 space_seg_t ssearch, *ss;
231 uint64_t end = start + size;
233 ASSERT(MUTEX_HELD(sm->sm_lock));
235 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
236 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
238 ssearch.ss_start = start;
239 ssearch.ss_end = end;
240 ss = avl_find(&sm->sm_root, &ssearch, &where);
242 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
246 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
251 ASSERT(MUTEX_HELD(sm->sm_lock));
253 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
255 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
256 kmem_cache_free(space_seg_cache, ss);
262 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
266 ASSERT(MUTEX_HELD(sm->sm_lock));
268 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
269 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
273 * Wait for any in-progress space_map_load() to complete.
276 space_map_load_wait(space_map_t *sm)
278 ASSERT(MUTEX_HELD(sm->sm_lock));
280 while (sm->sm_loading) {
281 ASSERT(!sm->sm_loaded);
282 cv_wait(&sm->sm_load_cv, sm->sm_lock);
287 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
288 * The caller must be OK with this.
291 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
292 space_map_obj_t *smo, objset_t *os)
294 uint64_t *entry, *entry_map, *entry_map_end;
295 uint64_t bufsize, size, offset, end, space;
296 uint64_t mapstart = sm->sm_start;
299 ASSERT(MUTEX_HELD(sm->sm_lock));
300 ASSERT(!sm->sm_loaded);
301 ASSERT(!sm->sm_loading);
303 sm->sm_loading = B_TRUE;
304 end = smo->smo_objsize;
305 space = smo->smo_alloc;
307 ASSERT(sm->sm_ops == NULL);
308 VERIFY0(sm->sm_space);
310 if (maptype == SM_FREE) {
311 space_map_add(sm, sm->sm_start, sm->sm_size);
312 space = sm->sm_size - space;
315 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
316 entry_map = zio_buf_alloc(bufsize);
318 mutex_exit(sm->sm_lock);
320 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
321 mutex_enter(sm->sm_lock);
323 for (offset = 0; offset < end; offset += bufsize) {
324 size = MIN(end - offset, bufsize);
325 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
328 dprintf("object=%llu offset=%llx size=%llx\n",
329 smo->smo_object, offset, size);
331 mutex_exit(sm->sm_lock);
332 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
334 mutex_enter(sm->sm_lock);
338 entry_map_end = entry_map + (size / sizeof (uint64_t));
339 for (entry = entry_map; entry < entry_map_end; entry++) {
342 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
345 (SM_TYPE_DECODE(e) == maptype ?
346 space_map_add : space_map_remove)(sm,
347 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
348 SM_RUN_DECODE(e) << sm->sm_shift);
353 VERIFY3U(sm->sm_space, ==, space);
355 sm->sm_loaded = B_TRUE;
360 space_map_vacate(sm, NULL, NULL);
363 zio_buf_free(entry_map, bufsize);
365 sm->sm_loading = B_FALSE;
367 cv_broadcast(&sm->sm_load_cv);
373 space_map_unload(space_map_t *sm)
375 ASSERT(MUTEX_HELD(sm->sm_lock));
377 if (sm->sm_loaded && sm->sm_ops != NULL)
378 sm->sm_ops->smop_unload(sm);
380 sm->sm_loaded = B_FALSE;
383 space_map_vacate(sm, NULL, NULL);
387 space_map_maxsize(space_map_t *sm)
389 ASSERT(sm->sm_ops != NULL);
390 return (sm->sm_ops->smop_max(sm));
394 space_map_alloc(space_map_t *sm, uint64_t size)
398 start = sm->sm_ops->smop_alloc(sm, size);
400 space_map_remove(sm, start, size);
405 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
407 sm->sm_ops->smop_claim(sm, start, size);
408 space_map_remove(sm, start, size);
412 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
414 space_map_add(sm, start, size);
415 sm->sm_ops->smop_free(sm, start, size);
419 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
422 space_map_sync(space_map_t *sm, uint8_t maptype,
423 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
425 spa_t *spa = dmu_objset_spa(os);
428 uint64_t bufsize, start, size, run_len, delta, sm_space;
429 uint64_t *entry, *entry_map, *entry_map_end;
431 ASSERT(MUTEX_HELD(sm->sm_lock));
433 if (sm->sm_space == 0)
436 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
437 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
438 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
441 if (maptype == SM_ALLOC)
442 smo->smo_alloc += sm->sm_space;
444 smo->smo_alloc -= sm->sm_space;
446 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
447 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
448 entry_map = zio_buf_alloc(bufsize);
449 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
452 *entry++ = SM_DEBUG_ENCODE(1) |
453 SM_DEBUG_ACTION_ENCODE(maptype) |
454 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
455 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
458 sm_space = sm->sm_space;
459 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
460 size = ss->ss_end - ss->ss_start;
461 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
464 size >>= sm->sm_shift;
467 run_len = MIN(size, SM_RUN_MAX);
469 if (entry == entry_map_end) {
470 mutex_exit(sm->sm_lock);
471 dmu_write(os, smo->smo_object, smo->smo_objsize,
472 bufsize, entry_map, tx);
473 mutex_enter(sm->sm_lock);
474 smo->smo_objsize += bufsize;
478 *entry++ = SM_OFFSET_ENCODE(start) |
479 SM_TYPE_ENCODE(maptype) |
480 SM_RUN_ENCODE(run_len);
485 kmem_cache_free(space_seg_cache, ss);
488 if (entry != entry_map) {
489 size = (entry - entry_map) * sizeof (uint64_t);
490 mutex_exit(sm->sm_lock);
491 dmu_write(os, smo->smo_object, smo->smo_objsize,
492 size, entry_map, tx);
493 mutex_enter(sm->sm_lock);
494 smo->smo_objsize += size;
498 * Ensure that the space_map's accounting wasn't changed
499 * while we were in the middle of writing it out.
501 VERIFY3U(sm->sm_space, ==, sm_space);
503 zio_buf_free(entry_map, bufsize);
505 sm->sm_space -= delta;
506 VERIFY0(sm->sm_space);
510 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
512 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
514 smo->smo_objsize = 0;
519 * Space map reference trees.
521 * A space map is a collection of integers. Every integer is either
522 * in the map, or it's not. A space map reference tree generalizes
523 * the idea: it allows its members to have arbitrary reference counts,
524 * as opposed to the implicit reference count of 0 or 1 in a space map.
525 * This representation comes in handy when computing the union or
526 * intersection of multiple space maps. For example, the union of
527 * N space maps is the subset of the reference tree with refcnt >= 1.
528 * The intersection of N space maps is the subset with refcnt >= N.
530 * [It's very much like a Fourier transform. Unions and intersections
531 * are hard to perform in the 'space map domain', so we convert the maps
532 * into the 'reference count domain', where it's trivial, then invert.]
534 * vdev_dtl_reassess() uses computations of this form to determine
535 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
536 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
537 * has an outage wherever refcnt >= vdev_children.
540 space_map_ref_compare(const void *x1, const void *x2)
542 const space_ref_t *sr1 = x1;
543 const space_ref_t *sr2 = x2;
545 if (sr1->sr_offset < sr2->sr_offset)
547 if (sr1->sr_offset > sr2->sr_offset)
559 space_map_ref_create(avl_tree_t *t)
561 avl_create(t, space_map_ref_compare,
562 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
566 space_map_ref_destroy(avl_tree_t *t)
571 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
572 kmem_free(sr, sizeof (*sr));
578 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
582 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
583 sr->sr_offset = offset;
584 sr->sr_refcnt = refcnt;
590 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
593 space_map_ref_add_node(t, start, refcnt);
594 space_map_ref_add_node(t, end, -refcnt);
598 * Convert (or add) a space map into a reference tree.
601 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
605 ASSERT(MUTEX_HELD(sm->sm_lock));
607 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
608 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
612 * Convert a reference tree into a space map. The space map will contain
613 * all members of the reference tree for which refcnt >= minref.
616 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
618 uint64_t start = -1ULL;
622 ASSERT(MUTEX_HELD(sm->sm_lock));
624 space_map_vacate(sm, NULL, NULL);
626 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
627 refcnt += sr->sr_refcnt;
628 if (refcnt >= minref) {
629 if (start == -1ULL) {
630 start = sr->sr_offset;
633 if (start != -1ULL) {
634 uint64_t end = sr->sr_offset;
635 ASSERT(start <= end);
637 space_map_add(sm, start, end - start);
643 ASSERT(start == -1ULL);