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 #include <sys/zfs_context.h>
30 #include <sys/space_map.h>
34 * NOTE: caller is responsible for all locking.
37 space_map_seg_compare(const void *x1, const void *x2)
39 const space_seg_t *s1 = x1;
40 const space_seg_t *s2 = x2;
42 if (s1->ss_start < s2->ss_start) {
43 if (s1->ss_end > s2->ss_start)
47 if (s1->ss_start > s2->ss_start) {
48 if (s1->ss_start < s2->ss_end)
56 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
59 bzero(sm, sizeof (*sm));
61 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
63 avl_create(&sm->sm_root, space_map_seg_compare,
64 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
73 space_map_destroy(space_map_t *sm)
75 ASSERT(!sm->sm_loaded && !sm->sm_loading);
76 VERIFY3U(sm->sm_space, ==, 0);
77 avl_destroy(&sm->sm_root);
78 cv_destroy(&sm->sm_load_cv);
82 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
85 space_seg_t ssearch, *ss_before, *ss_after, *ss;
86 uint64_t end = start + size;
87 int merge_before, merge_after;
89 ASSERT(MUTEX_HELD(sm->sm_lock));
91 VERIFY3U(start, >=, sm->sm_start);
92 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
93 VERIFY(sm->sm_space + size <= sm->sm_size);
94 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
95 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
97 ssearch.ss_start = start;
99 ss = avl_find(&sm->sm_root, &ssearch, &where);
101 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
102 zfs_panic_recover("zfs: allocating allocated segment"
103 "(offset=%llu size=%llu)\n",
104 (longlong_t)start, (longlong_t)size);
108 /* Make sure we don't overlap with either of our neighbors */
111 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
112 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
114 merge_before = (ss_before != NULL && ss_before->ss_end == start);
115 merge_after = (ss_after != NULL && ss_after->ss_start == end);
117 if (merge_before && merge_after) {
118 avl_remove(&sm->sm_root, ss_before);
119 if (sm->sm_pp_root) {
120 avl_remove(sm->sm_pp_root, ss_before);
121 avl_remove(sm->sm_pp_root, ss_after);
123 ss_after->ss_start = ss_before->ss_start;
124 kmem_free(ss_before, sizeof (*ss_before));
126 } else if (merge_before) {
127 ss_before->ss_end = end;
129 avl_remove(sm->sm_pp_root, ss_before);
131 } else if (merge_after) {
132 ss_after->ss_start = start;
134 avl_remove(sm->sm_pp_root, ss_after);
137 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
138 ss->ss_start = start;
140 avl_insert(&sm->sm_root, ss, where);
144 avl_add(sm->sm_pp_root, ss);
146 sm->sm_space += size;
150 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
153 space_seg_t ssearch, *ss, *newseg;
154 uint64_t end = start + size;
155 int left_over, right_over;
157 ASSERT(MUTEX_HELD(sm->sm_lock));
159 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
160 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
162 ssearch.ss_start = start;
163 ssearch.ss_end = end;
164 ss = avl_find(&sm->sm_root, &ssearch, &where);
166 /* Make sure we completely overlap with someone */
168 zfs_panic_recover("zfs: freeing free segment "
169 "(offset=%llu size=%llu)",
170 (longlong_t)start, (longlong_t)size);
173 VERIFY3U(ss->ss_start, <=, start);
174 VERIFY3U(ss->ss_end, >=, end);
175 VERIFY(sm->sm_space - size <= sm->sm_size);
177 left_over = (ss->ss_start != start);
178 right_over = (ss->ss_end != end);
181 avl_remove(sm->sm_pp_root, ss);
183 if (left_over && right_over) {
184 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
185 newseg->ss_start = end;
186 newseg->ss_end = ss->ss_end;
188 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
190 avl_add(sm->sm_pp_root, newseg);
191 } else if (left_over) {
193 } else if (right_over) {
196 avl_remove(&sm->sm_root, ss);
197 kmem_free(ss, sizeof (*ss));
201 if (sm->sm_pp_root && ss != NULL)
202 avl_add(sm->sm_pp_root, ss);
204 sm->sm_space -= size;
208 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
211 space_seg_t ssearch, *ss;
212 uint64_t end = start + size;
214 ASSERT(MUTEX_HELD(sm->sm_lock));
216 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
217 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
219 ssearch.ss_start = start;
220 ssearch.ss_end = end;
221 ss = avl_find(&sm->sm_root, &ssearch, &where);
223 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
227 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
232 ASSERT(MUTEX_HELD(sm->sm_lock));
234 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
236 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
237 kmem_free(ss, sizeof (*ss));
243 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
247 ASSERT(MUTEX_HELD(sm->sm_lock));
249 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
250 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
254 * Wait for any in-progress space_map_load() to complete.
257 space_map_load_wait(space_map_t *sm)
259 ASSERT(MUTEX_HELD(sm->sm_lock));
261 while (sm->sm_loading)
262 cv_wait(&sm->sm_load_cv, sm->sm_lock);
266 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
267 * The caller must be OK with this.
270 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
271 space_map_obj_t *smo, objset_t *os)
273 uint64_t *entry, *entry_map, *entry_map_end;
274 uint64_t bufsize, size, offset, end, space;
275 uint64_t mapstart = sm->sm_start;
278 ASSERT(MUTEX_HELD(sm->sm_lock));
280 space_map_load_wait(sm);
285 sm->sm_loading = B_TRUE;
286 end = smo->smo_objsize;
287 space = smo->smo_alloc;
289 ASSERT(sm->sm_ops == NULL);
290 VERIFY3U(sm->sm_space, ==, 0);
292 if (maptype == SM_FREE) {
293 space_map_add(sm, sm->sm_start, sm->sm_size);
294 space = sm->sm_size - space;
297 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
298 entry_map = zio_buf_alloc(bufsize);
300 mutex_exit(sm->sm_lock);
302 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
303 mutex_enter(sm->sm_lock);
305 for (offset = 0; offset < end; offset += bufsize) {
306 size = MIN(end - offset, bufsize);
307 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
310 dprintf("object=%llu offset=%llx size=%llx\n",
311 smo->smo_object, offset, size);
313 mutex_exit(sm->sm_lock);
314 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
316 mutex_enter(sm->sm_lock);
320 entry_map_end = entry_map + (size / sizeof (uint64_t));
321 for (entry = entry_map; entry < entry_map_end; entry++) {
324 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
327 (SM_TYPE_DECODE(e) == maptype ?
328 space_map_add : space_map_remove)(sm,
329 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
330 SM_RUN_DECODE(e) << sm->sm_shift);
335 VERIFY3U(sm->sm_space, ==, space);
337 sm->sm_loaded = B_TRUE;
342 space_map_vacate(sm, NULL, NULL);
345 zio_buf_free(entry_map, bufsize);
347 sm->sm_loading = B_FALSE;
349 cv_broadcast(&sm->sm_load_cv);
355 space_map_unload(space_map_t *sm)
357 ASSERT(MUTEX_HELD(sm->sm_lock));
359 if (sm->sm_loaded && sm->sm_ops != NULL)
360 sm->sm_ops->smop_unload(sm);
362 sm->sm_loaded = B_FALSE;
365 space_map_vacate(sm, NULL, NULL);
369 space_map_maxsize(space_map_t *sm)
371 ASSERT(sm->sm_ops != NULL);
372 return (sm->sm_ops->smop_max(sm));
376 space_map_alloc(space_map_t *sm, uint64_t size)
380 start = sm->sm_ops->smop_alloc(sm, size);
382 space_map_remove(sm, start, size);
387 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
389 sm->sm_ops->smop_claim(sm, start, size);
390 space_map_remove(sm, start, size);
394 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
396 space_map_add(sm, start, size);
397 sm->sm_ops->smop_free(sm, start, size);
401 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
404 space_map_sync(space_map_t *sm, uint8_t maptype,
405 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
407 spa_t *spa = dmu_objset_spa(os);
410 uint64_t bufsize, start, size, run_len;
411 uint64_t *entry, *entry_map, *entry_map_end;
413 ASSERT(MUTEX_HELD(sm->sm_lock));
415 if (sm->sm_space == 0)
418 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
419 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
420 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
423 if (maptype == SM_ALLOC)
424 smo->smo_alloc += sm->sm_space;
426 smo->smo_alloc -= sm->sm_space;
428 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
429 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
430 entry_map = zio_buf_alloc(bufsize);
431 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
434 *entry++ = SM_DEBUG_ENCODE(1) |
435 SM_DEBUG_ACTION_ENCODE(maptype) |
436 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
437 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
439 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
440 size = ss->ss_end - ss->ss_start;
441 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
443 sm->sm_space -= size;
444 size >>= sm->sm_shift;
447 run_len = MIN(size, SM_RUN_MAX);
449 if (entry == entry_map_end) {
450 mutex_exit(sm->sm_lock);
451 dmu_write(os, smo->smo_object, smo->smo_objsize,
452 bufsize, entry_map, tx);
453 mutex_enter(sm->sm_lock);
454 smo->smo_objsize += bufsize;
458 *entry++ = SM_OFFSET_ENCODE(start) |
459 SM_TYPE_ENCODE(maptype) |
460 SM_RUN_ENCODE(run_len);
465 kmem_free(ss, sizeof (*ss));
468 if (entry != entry_map) {
469 size = (entry - entry_map) * sizeof (uint64_t);
470 mutex_exit(sm->sm_lock);
471 dmu_write(os, smo->smo_object, smo->smo_objsize,
472 size, entry_map, tx);
473 mutex_enter(sm->sm_lock);
474 smo->smo_objsize += size;
477 zio_buf_free(entry_map, bufsize);
479 VERIFY3U(sm->sm_space, ==, 0);
483 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
485 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
487 smo->smo_objsize = 0;
492 * Space map reference trees.
494 * A space map is a collection of integers. Every integer is either
495 * in the map, or it's not. A space map reference tree generalizes
496 * the idea: it allows its members to have arbitrary reference counts,
497 * as opposed to the implicit reference count of 0 or 1 in a space map.
498 * This representation comes in handy when computing the union or
499 * intersection of multiple space maps. For example, the union of
500 * N space maps is the subset of the reference tree with refcnt >= 1.
501 * The intersection of N space maps is the subset with refcnt >= N.
503 * [It's very much like a Fourier transform. Unions and intersections
504 * are hard to perform in the 'space map domain', so we convert the maps
505 * into the 'reference count domain', where it's trivial, then invert.]
507 * vdev_dtl_reassess() uses computations of this form to determine
508 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
509 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
510 * has an outage wherever refcnt >= vdev_children.
513 space_map_ref_compare(const void *x1, const void *x2)
515 const space_ref_t *sr1 = x1;
516 const space_ref_t *sr2 = x2;
518 if (sr1->sr_offset < sr2->sr_offset)
520 if (sr1->sr_offset > sr2->sr_offset)
532 space_map_ref_create(avl_tree_t *t)
534 avl_create(t, space_map_ref_compare,
535 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
539 space_map_ref_destroy(avl_tree_t *t)
544 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
545 kmem_free(sr, sizeof (*sr));
551 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
555 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
556 sr->sr_offset = offset;
557 sr->sr_refcnt = refcnt;
563 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
566 space_map_ref_add_node(t, start, refcnt);
567 space_map_ref_add_node(t, end, -refcnt);
571 * Convert (or add) a space map into a reference tree.
574 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
578 ASSERT(MUTEX_HELD(sm->sm_lock));
580 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
581 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
585 * Convert a reference tree into a space map. The space map will contain
586 * all members of the reference tree for which refcnt >= minref.
589 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
591 uint64_t start = -1ULL;
595 ASSERT(MUTEX_HELD(sm->sm_lock));
597 space_map_vacate(sm, NULL, NULL);
599 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
600 refcnt += sr->sr_refcnt;
601 if (refcnt >= minref) {
602 if (start == -1ULL) {
603 start = sr->sr_offset;
606 if (start != -1ULL) {
607 uint64_t end = sr->sr_offset;
608 ASSERT(start <= end);
610 space_map_add(sm, start, end - start);
616 ASSERT(start == -1ULL);