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>
37 * NOTE: caller is responsible for all locking.
40 space_map_seg_compare(const void *x1, const void *x2)
42 const space_seg_t *s1 = x1;
43 const space_seg_t *s2 = x2;
45 if (s1->ss_start < s2->ss_start) {
46 if (s1->ss_end > s2->ss_start)
50 if (s1->ss_start > s2->ss_start) {
51 if (s1->ss_start < s2->ss_end)
59 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
62 bzero(sm, sizeof (*sm));
64 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
66 avl_create(&sm->sm_root, space_map_seg_compare,
67 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
76 space_map_destroy(space_map_t *sm)
78 ASSERT(!sm->sm_loaded && !sm->sm_loading);
79 VERIFY0(sm->sm_space);
80 avl_destroy(&sm->sm_root);
81 cv_destroy(&sm->sm_load_cv);
85 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
88 space_seg_t ssearch, *ss_before, *ss_after, *ss;
89 uint64_t end = start + size;
90 int merge_before, merge_after;
92 ASSERT(MUTEX_HELD(sm->sm_lock));
94 VERIFY3U(start, >=, sm->sm_start);
95 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
96 VERIFY(sm->sm_space + size <= sm->sm_size);
97 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
98 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
100 ssearch.ss_start = start;
101 ssearch.ss_end = end;
102 ss = avl_find(&sm->sm_root, &ssearch, &where);
104 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
105 zfs_panic_recover("zfs: allocating allocated segment"
106 "(offset=%llu size=%llu)\n",
107 (longlong_t)start, (longlong_t)size);
111 /* Make sure we don't overlap with either of our neighbors */
114 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
115 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
117 merge_before = (ss_before != NULL && ss_before->ss_end == start);
118 merge_after = (ss_after != NULL && ss_after->ss_start == end);
120 if (merge_before && merge_after) {
121 avl_remove(&sm->sm_root, ss_before);
122 if (sm->sm_pp_root) {
123 avl_remove(sm->sm_pp_root, ss_before);
124 avl_remove(sm->sm_pp_root, ss_after);
126 ss_after->ss_start = ss_before->ss_start;
127 kmem_free(ss_before, sizeof (*ss_before));
129 } else if (merge_before) {
130 ss_before->ss_end = end;
132 avl_remove(sm->sm_pp_root, ss_before);
134 } else if (merge_after) {
135 ss_after->ss_start = start;
137 avl_remove(sm->sm_pp_root, ss_after);
140 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
141 ss->ss_start = start;
143 avl_insert(&sm->sm_root, ss, where);
147 avl_add(sm->sm_pp_root, ss);
149 sm->sm_space += size;
153 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
155 space_seg_t ssearch, *ss, *newseg;
156 uint64_t end = start + size;
157 int left_over, right_over;
159 ASSERT(MUTEX_HELD(sm->sm_lock));
161 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
162 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
164 ssearch.ss_start = start;
165 ssearch.ss_end = end;
166 ss = avl_find(&sm->sm_root, &ssearch, NULL);
168 /* Make sure we completely overlap with someone */
170 zfs_panic_recover("zfs: freeing free segment "
171 "(offset=%llu size=%llu)",
172 (longlong_t)start, (longlong_t)size);
175 VERIFY3U(ss->ss_start, <=, start);
176 VERIFY3U(ss->ss_end, >=, end);
177 VERIFY(sm->sm_space - size < sm->sm_size);
179 left_over = (ss->ss_start != start);
180 right_over = (ss->ss_end != end);
183 avl_remove(sm->sm_pp_root, ss);
185 if (left_over && right_over) {
186 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
187 newseg->ss_start = end;
188 newseg->ss_end = ss->ss_end;
190 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
192 avl_add(sm->sm_pp_root, newseg);
193 } else if (left_over) {
195 } else if (right_over) {
198 avl_remove(&sm->sm_root, ss);
199 kmem_free(ss, sizeof (*ss));
203 if (sm->sm_pp_root && ss != NULL)
204 avl_add(sm->sm_pp_root, ss);
206 sm->sm_space -= size;
210 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
213 space_seg_t ssearch, *ss;
214 uint64_t end = start + size;
216 ASSERT(MUTEX_HELD(sm->sm_lock));
218 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
219 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
221 ssearch.ss_start = start;
222 ssearch.ss_end = end;
223 ss = avl_find(&sm->sm_root, &ssearch, &where);
225 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
229 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
234 ASSERT(MUTEX_HELD(sm->sm_lock));
236 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
238 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
239 kmem_free(ss, sizeof (*ss));
245 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
249 ASSERT(MUTEX_HELD(sm->sm_lock));
251 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
252 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
256 * Wait for any in-progress space_map_load() to complete.
259 space_map_load_wait(space_map_t *sm)
261 ASSERT(MUTEX_HELD(sm->sm_lock));
263 while (sm->sm_loading) {
264 ASSERT(!sm->sm_loaded);
265 cv_wait(&sm->sm_load_cv, sm->sm_lock);
270 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
271 * The caller must be OK with this.
274 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
275 space_map_obj_t *smo, objset_t *os)
277 uint64_t *entry, *entry_map, *entry_map_end;
278 uint64_t bufsize, size, offset, end, space;
279 uint64_t mapstart = sm->sm_start;
282 ASSERT(MUTEX_HELD(sm->sm_lock));
283 ASSERT(!sm->sm_loaded);
284 ASSERT(!sm->sm_loading);
286 sm->sm_loading = B_TRUE;
287 end = smo->smo_objsize;
288 space = smo->smo_alloc;
290 ASSERT(sm->sm_ops == NULL);
291 VERIFY0(sm->sm_space);
293 if (maptype == SM_FREE) {
294 space_map_add(sm, sm->sm_start, sm->sm_size);
295 space = sm->sm_size - space;
298 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
299 entry_map = zio_buf_alloc(bufsize);
301 mutex_exit(sm->sm_lock);
303 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
304 mutex_enter(sm->sm_lock);
306 for (offset = 0; offset < end; offset += bufsize) {
307 size = MIN(end - offset, bufsize);
308 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
311 dprintf("object=%llu offset=%llx size=%llx\n",
312 smo->smo_object, offset, size);
314 mutex_exit(sm->sm_lock);
315 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
317 mutex_enter(sm->sm_lock);
321 entry_map_end = entry_map + (size / sizeof (uint64_t));
322 for (entry = entry_map; entry < entry_map_end; entry++) {
325 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
328 (SM_TYPE_DECODE(e) == maptype ?
329 space_map_add : space_map_remove)(sm,
330 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
331 SM_RUN_DECODE(e) << sm->sm_shift);
336 VERIFY3U(sm->sm_space, ==, space);
338 sm->sm_loaded = B_TRUE;
343 space_map_vacate(sm, NULL, NULL);
346 zio_buf_free(entry_map, bufsize);
348 sm->sm_loading = B_FALSE;
350 cv_broadcast(&sm->sm_load_cv);
356 space_map_unload(space_map_t *sm)
358 ASSERT(MUTEX_HELD(sm->sm_lock));
360 if (sm->sm_loaded && sm->sm_ops != NULL)
361 sm->sm_ops->smop_unload(sm);
363 sm->sm_loaded = B_FALSE;
366 space_map_vacate(sm, NULL, NULL);
370 space_map_maxsize(space_map_t *sm)
372 ASSERT(sm->sm_ops != NULL);
373 return (sm->sm_ops->smop_max(sm));
377 space_map_alloc(space_map_t *sm, uint64_t size)
381 start = sm->sm_ops->smop_alloc(sm, size);
383 space_map_remove(sm, start, size);
388 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
390 sm->sm_ops->smop_claim(sm, start, size);
391 space_map_remove(sm, start, size);
395 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
397 space_map_add(sm, start, size);
398 sm->sm_ops->smop_free(sm, start, size);
402 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
405 space_map_sync(space_map_t *sm, uint8_t maptype,
406 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
408 spa_t *spa = dmu_objset_spa(os);
411 uint64_t bufsize, start, size, run_len;
412 uint64_t *entry, *entry_map, *entry_map_end;
414 ASSERT(MUTEX_HELD(sm->sm_lock));
416 if (sm->sm_space == 0)
419 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
420 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
421 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
424 if (maptype == SM_ALLOC)
425 smo->smo_alloc += sm->sm_space;
427 smo->smo_alloc -= sm->sm_space;
429 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
430 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
431 entry_map = zio_buf_alloc(bufsize);
432 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
435 *entry++ = SM_DEBUG_ENCODE(1) |
436 SM_DEBUG_ACTION_ENCODE(maptype) |
437 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
438 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
440 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
441 size = ss->ss_end - ss->ss_start;
442 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
444 sm->sm_space -= size;
445 size >>= sm->sm_shift;
448 run_len = MIN(size, SM_RUN_MAX);
450 if (entry == entry_map_end) {
451 mutex_exit(sm->sm_lock);
452 dmu_write(os, smo->smo_object, smo->smo_objsize,
453 bufsize, entry_map, tx);
454 mutex_enter(sm->sm_lock);
455 smo->smo_objsize += bufsize;
459 *entry++ = SM_OFFSET_ENCODE(start) |
460 SM_TYPE_ENCODE(maptype) |
461 SM_RUN_ENCODE(run_len);
466 kmem_free(ss, sizeof (*ss));
469 if (entry != entry_map) {
470 size = (entry - entry_map) * sizeof (uint64_t);
471 mutex_exit(sm->sm_lock);
472 dmu_write(os, smo->smo_object, smo->smo_objsize,
473 size, entry_map, tx);
474 mutex_enter(sm->sm_lock);
475 smo->smo_objsize += size;
478 zio_buf_free(entry_map, bufsize);
480 VERIFY0(sm->sm_space);
484 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
486 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
488 smo->smo_objsize = 0;
493 * Space map reference trees.
495 * A space map is a collection of integers. Every integer is either
496 * in the map, or it's not. A space map reference tree generalizes
497 * the idea: it allows its members to have arbitrary reference counts,
498 * as opposed to the implicit reference count of 0 or 1 in a space map.
499 * This representation comes in handy when computing the union or
500 * intersection of multiple space maps. For example, the union of
501 * N space maps is the subset of the reference tree with refcnt >= 1.
502 * The intersection of N space maps is the subset with refcnt >= N.
504 * [It's very much like a Fourier transform. Unions and intersections
505 * are hard to perform in the 'space map domain', so we convert the maps
506 * into the 'reference count domain', where it's trivial, then invert.]
508 * vdev_dtl_reassess() uses computations of this form to determine
509 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
510 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
511 * has an outage wherever refcnt >= vdev_children.
514 space_map_ref_compare(const void *x1, const void *x2)
516 const space_ref_t *sr1 = x1;
517 const space_ref_t *sr2 = x2;
519 if (sr1->sr_offset < sr2->sr_offset)
521 if (sr1->sr_offset > sr2->sr_offset)
533 space_map_ref_create(avl_tree_t *t)
535 avl_create(t, space_map_ref_compare,
536 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
540 space_map_ref_destroy(avl_tree_t *t)
545 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
546 kmem_free(sr, sizeof (*sr));
552 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
556 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
557 sr->sr_offset = offset;
558 sr->sr_refcnt = refcnt;
564 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
567 space_map_ref_add_node(t, start, refcnt);
568 space_map_ref_add_node(t, end, -refcnt);
572 * Convert (or add) a space map into a reference tree.
575 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
579 ASSERT(MUTEX_HELD(sm->sm_lock));
581 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
582 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
586 * Convert a reference tree into a space map. The space map will contain
587 * all members of the reference tree for which refcnt >= minref.
590 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
592 uint64_t start = -1ULL;
596 ASSERT(MUTEX_HELD(sm->sm_lock));
598 space_map_vacate(sm, NULL, NULL);
600 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
601 refcnt += sr->sr_refcnt;
602 if (refcnt >= minref) {
603 if (start == -1ULL) {
604 start = sr->sr_offset;
607 if (start != -1ULL) {
608 uint64_t end = sr->sr_offset;
609 ASSERT(start <= end);
611 space_map_add(sm, start, end - start);
617 ASSERT(start == -1ULL);