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 SYSCTL_DECL(_vfs_zfs);
36 static int space_map_last_hope;
37 TUNABLE_INT("vfs.zfs.space_map_last_hope", &space_map_last_hope);
38 SYSCTL_INT(_vfs_zfs, OID_AUTO, space_map_last_hope, CTLFLAG_RDTUN,
39 &space_map_last_hope, 0,
40 "If kernel panic in space_map code on pool import, import the pool in readonly mode and backup all your data before trying this option.");
44 * NOTE: caller is responsible for all locking.
47 space_map_seg_compare(const void *x1, const void *x2)
49 const space_seg_t *s1 = x1;
50 const space_seg_t *s2 = x2;
52 if (s1->ss_start < s2->ss_start) {
53 if (s1->ss_end > s2->ss_start)
57 if (s1->ss_start > s2->ss_start) {
58 if (s1->ss_start < s2->ss_end)
66 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
69 bzero(sm, sizeof (*sm));
71 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
73 avl_create(&sm->sm_root, space_map_seg_compare,
74 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
83 space_map_destroy(space_map_t *sm)
85 ASSERT(!sm->sm_loaded && !sm->sm_loading);
86 VERIFY0(sm->sm_space);
87 avl_destroy(&sm->sm_root);
88 cv_destroy(&sm->sm_load_cv);
92 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
95 space_seg_t ssearch, *ss_before, *ss_after, *ss;
96 uint64_t end = start + size;
97 int merge_before, merge_after;
99 ASSERT(MUTEX_HELD(sm->sm_lock));
101 VERIFY3U(start, >=, sm->sm_start);
102 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
103 VERIFY(sm->sm_space + size <= sm->sm_size);
104 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
105 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
107 ssearch.ss_start = start;
108 ssearch.ss_end = end;
109 ss = avl_find(&sm->sm_root, &ssearch, &where);
111 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
112 zfs_panic_recover("zfs: allocating allocated segment"
113 "(offset=%llu size=%llu)\n",
114 (longlong_t)start, (longlong_t)size);
117 if (ss != NULL && space_map_last_hope) {
118 uint64_t sstart, ssize;
120 if (ss->ss_start > start)
121 sstart = ss->ss_start;
124 if (ss->ss_end > end)
125 ssize = end - sstart;
127 ssize = ss->ss_end - sstart;
129 "Removing colliding space_map range (start=%ju end=%ju). Good luck!",
130 (uintmax_t)sstart, (uintmax_t)(sstart + ssize));
131 space_map_remove(sm, sstart, ssize);
135 /* Make sure we don't overlap with either of our neighbors */
138 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
139 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
141 merge_before = (ss_before != NULL && ss_before->ss_end == start);
142 merge_after = (ss_after != NULL && ss_after->ss_start == end);
144 if (merge_before && merge_after) {
145 avl_remove(&sm->sm_root, ss_before);
146 if (sm->sm_pp_root) {
147 avl_remove(sm->sm_pp_root, ss_before);
148 avl_remove(sm->sm_pp_root, ss_after);
150 ss_after->ss_start = ss_before->ss_start;
151 kmem_free(ss_before, sizeof (*ss_before));
153 } else if (merge_before) {
154 ss_before->ss_end = end;
156 avl_remove(sm->sm_pp_root, ss_before);
158 } else if (merge_after) {
159 ss_after->ss_start = start;
161 avl_remove(sm->sm_pp_root, ss_after);
164 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
165 ss->ss_start = start;
167 avl_insert(&sm->sm_root, ss, where);
171 avl_add(sm->sm_pp_root, ss);
173 sm->sm_space += size;
177 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
179 space_seg_t ssearch, *ss, *newseg;
180 uint64_t end = start + size;
181 int left_over, right_over;
183 ASSERT(MUTEX_HELD(sm->sm_lock));
185 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
186 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
188 ssearch.ss_start = start;
189 ssearch.ss_end = end;
190 ss = avl_find(&sm->sm_root, &ssearch, NULL);
192 /* Make sure we completely overlap with someone */
194 zfs_panic_recover("zfs: freeing free segment "
195 "(offset=%llu size=%llu)",
196 (longlong_t)start, (longlong_t)size);
199 VERIFY3U(ss->ss_start, <=, start);
200 VERIFY3U(ss->ss_end, >=, end);
201 VERIFY(sm->sm_space - size < sm->sm_size);
203 left_over = (ss->ss_start != start);
204 right_over = (ss->ss_end != end);
207 avl_remove(sm->sm_pp_root, ss);
209 if (left_over && right_over) {
210 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
211 newseg->ss_start = end;
212 newseg->ss_end = ss->ss_end;
214 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
216 avl_add(sm->sm_pp_root, newseg);
217 } else if (left_over) {
219 } else if (right_over) {
222 avl_remove(&sm->sm_root, ss);
223 kmem_free(ss, sizeof (*ss));
227 if (sm->sm_pp_root && ss != NULL)
228 avl_add(sm->sm_pp_root, ss);
230 sm->sm_space -= size;
234 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
237 space_seg_t ssearch, *ss;
238 uint64_t end = start + size;
240 ASSERT(MUTEX_HELD(sm->sm_lock));
242 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
243 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
245 ssearch.ss_start = start;
246 ssearch.ss_end = end;
247 ss = avl_find(&sm->sm_root, &ssearch, &where);
249 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
253 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
258 ASSERT(MUTEX_HELD(sm->sm_lock));
260 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
262 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
263 kmem_free(ss, sizeof (*ss));
269 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
273 ASSERT(MUTEX_HELD(sm->sm_lock));
275 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
276 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
280 * Wait for any in-progress space_map_load() to complete.
283 space_map_load_wait(space_map_t *sm)
285 ASSERT(MUTEX_HELD(sm->sm_lock));
287 while (sm->sm_loading) {
288 ASSERT(!sm->sm_loaded);
289 cv_wait(&sm->sm_load_cv, sm->sm_lock);
294 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
295 * The caller must be OK with this.
298 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
299 space_map_obj_t *smo, objset_t *os)
301 uint64_t *entry, *entry_map, *entry_map_end;
302 uint64_t bufsize, size, offset, end, space;
303 uint64_t mapstart = sm->sm_start;
306 ASSERT(MUTEX_HELD(sm->sm_lock));
307 ASSERT(!sm->sm_loaded);
308 ASSERT(!sm->sm_loading);
310 sm->sm_loading = B_TRUE;
311 end = smo->smo_objsize;
312 space = smo->smo_alloc;
314 ASSERT(sm->sm_ops == NULL);
315 VERIFY0(sm->sm_space);
317 if (maptype == SM_FREE) {
318 space_map_add(sm, sm->sm_start, sm->sm_size);
319 space = sm->sm_size - space;
322 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
323 entry_map = zio_buf_alloc(bufsize);
325 mutex_exit(sm->sm_lock);
327 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
328 mutex_enter(sm->sm_lock);
330 for (offset = 0; offset < end; offset += bufsize) {
331 size = MIN(end - offset, bufsize);
332 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
335 dprintf("object=%llu offset=%llx size=%llx\n",
336 smo->smo_object, offset, size);
338 mutex_exit(sm->sm_lock);
339 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
341 mutex_enter(sm->sm_lock);
345 entry_map_end = entry_map + (size / sizeof (uint64_t));
346 for (entry = entry_map; entry < entry_map_end; entry++) {
349 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
352 (SM_TYPE_DECODE(e) == maptype ?
353 space_map_add : space_map_remove)(sm,
354 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
355 SM_RUN_DECODE(e) << sm->sm_shift);
360 VERIFY3U(sm->sm_space, ==, space);
362 sm->sm_loaded = B_TRUE;
367 space_map_vacate(sm, NULL, NULL);
370 zio_buf_free(entry_map, bufsize);
372 sm->sm_loading = B_FALSE;
374 cv_broadcast(&sm->sm_load_cv);
380 space_map_unload(space_map_t *sm)
382 ASSERT(MUTEX_HELD(sm->sm_lock));
384 if (sm->sm_loaded && sm->sm_ops != NULL)
385 sm->sm_ops->smop_unload(sm);
387 sm->sm_loaded = B_FALSE;
390 space_map_vacate(sm, NULL, NULL);
394 space_map_maxsize(space_map_t *sm)
396 ASSERT(sm->sm_ops != NULL);
397 return (sm->sm_ops->smop_max(sm));
401 space_map_alloc(space_map_t *sm, uint64_t size)
405 start = sm->sm_ops->smop_alloc(sm, size);
407 space_map_remove(sm, start, size);
412 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
414 sm->sm_ops->smop_claim(sm, start, size);
415 space_map_remove(sm, start, size);
419 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
421 space_map_add(sm, start, size);
422 sm->sm_ops->smop_free(sm, start, size);
426 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
429 space_map_sync(space_map_t *sm, uint8_t maptype,
430 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
432 spa_t *spa = dmu_objset_spa(os);
435 uint64_t bufsize, start, size, run_len;
436 uint64_t *entry, *entry_map, *entry_map_end;
438 ASSERT(MUTEX_HELD(sm->sm_lock));
440 if (sm->sm_space == 0)
443 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
444 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
445 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
448 if (maptype == SM_ALLOC)
449 smo->smo_alloc += sm->sm_space;
451 smo->smo_alloc -= sm->sm_space;
453 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
454 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
455 entry_map = zio_buf_alloc(bufsize);
456 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
459 *entry++ = SM_DEBUG_ENCODE(1) |
460 SM_DEBUG_ACTION_ENCODE(maptype) |
461 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
462 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
464 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
465 size = ss->ss_end - ss->ss_start;
466 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
468 sm->sm_space -= size;
469 size >>= sm->sm_shift;
472 run_len = MIN(size, SM_RUN_MAX);
474 if (entry == entry_map_end) {
475 mutex_exit(sm->sm_lock);
476 dmu_write(os, smo->smo_object, smo->smo_objsize,
477 bufsize, entry_map, tx);
478 mutex_enter(sm->sm_lock);
479 smo->smo_objsize += bufsize;
483 *entry++ = SM_OFFSET_ENCODE(start) |
484 SM_TYPE_ENCODE(maptype) |
485 SM_RUN_ENCODE(run_len);
490 kmem_free(ss, sizeof (*ss));
493 if (entry != entry_map) {
494 size = (entry - entry_map) * sizeof (uint64_t);
495 mutex_exit(sm->sm_lock);
496 dmu_write(os, smo->smo_object, smo->smo_objsize,
497 size, entry_map, tx);
498 mutex_enter(sm->sm_lock);
499 smo->smo_objsize += size;
502 zio_buf_free(entry_map, bufsize);
504 VERIFY0(sm->sm_space);
508 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
510 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
512 smo->smo_objsize = 0;
517 * Space map reference trees.
519 * A space map is a collection of integers. Every integer is either
520 * in the map, or it's not. A space map reference tree generalizes
521 * the idea: it allows its members to have arbitrary reference counts,
522 * as opposed to the implicit reference count of 0 or 1 in a space map.
523 * This representation comes in handy when computing the union or
524 * intersection of multiple space maps. For example, the union of
525 * N space maps is the subset of the reference tree with refcnt >= 1.
526 * The intersection of N space maps is the subset with refcnt >= N.
528 * [It's very much like a Fourier transform. Unions and intersections
529 * are hard to perform in the 'space map domain', so we convert the maps
530 * into the 'reference count domain', where it's trivial, then invert.]
532 * vdev_dtl_reassess() uses computations of this form to determine
533 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
534 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
535 * has an outage wherever refcnt >= vdev_children.
538 space_map_ref_compare(const void *x1, const void *x2)
540 const space_ref_t *sr1 = x1;
541 const space_ref_t *sr2 = x2;
543 if (sr1->sr_offset < sr2->sr_offset)
545 if (sr1->sr_offset > sr2->sr_offset)
557 space_map_ref_create(avl_tree_t *t)
559 avl_create(t, space_map_ref_compare,
560 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
564 space_map_ref_destroy(avl_tree_t *t)
569 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
570 kmem_free(sr, sizeof (*sr));
576 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
580 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
581 sr->sr_offset = offset;
582 sr->sr_refcnt = refcnt;
588 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
591 space_map_ref_add_node(t, start, refcnt);
592 space_map_ref_add_node(t, end, -refcnt);
596 * Convert (or add) a space map into a reference tree.
599 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
603 ASSERT(MUTEX_HELD(sm->sm_lock));
605 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
606 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
610 * Convert a reference tree into a space map. The space map will contain
611 * all members of the reference tree for which refcnt >= minref.
614 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
616 uint64_t start = -1ULL;
620 ASSERT(MUTEX_HELD(sm->sm_lock));
622 space_map_vacate(sm, NULL, NULL);
624 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
625 refcnt += sr->sr_refcnt;
626 if (refcnt >= minref) {
627 if (start == -1ULL) {
628 start = sr->sr_offset;
631 if (start != -1ULL) {
632 uint64_t end = sr->sr_offset;
633 ASSERT(start <= end);
635 space_map_add(sm, start, end - start);
641 ASSERT(start == -1ULL);