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>
32 SYSCTL_DECL(_vfs_zfs);
33 static int space_map_last_hope;
34 TUNABLE_INT("vfs.zfs.space_map_last_hope", &space_map_last_hope);
35 SYSCTL_INT(_vfs_zfs, OID_AUTO, space_map_last_hope, CTLFLAG_RDTUN,
36 &space_map_last_hope, 0,
37 "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.");
41 * NOTE: caller is responsible for all locking.
44 space_map_seg_compare(const void *x1, const void *x2)
46 const space_seg_t *s1 = x1;
47 const space_seg_t *s2 = x2;
49 if (s1->ss_start < s2->ss_start) {
50 if (s1->ss_end > s2->ss_start)
54 if (s1->ss_start > s2->ss_start) {
55 if (s1->ss_start < s2->ss_end)
63 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
66 bzero(sm, sizeof (*sm));
68 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
70 avl_create(&sm->sm_root, space_map_seg_compare,
71 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
80 space_map_destroy(space_map_t *sm)
82 ASSERT(!sm->sm_loaded && !sm->sm_loading);
83 VERIFY3U(sm->sm_space, ==, 0);
84 avl_destroy(&sm->sm_root);
85 cv_destroy(&sm->sm_load_cv);
89 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
92 space_seg_t ssearch, *ss_before, *ss_after, *ss;
93 uint64_t end = start + size;
94 int merge_before, merge_after;
96 ASSERT(MUTEX_HELD(sm->sm_lock));
98 VERIFY3U(start, >=, sm->sm_start);
99 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
100 VERIFY(sm->sm_space + size <= sm->sm_size);
101 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
102 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
104 ssearch.ss_start = start;
105 ssearch.ss_end = end;
106 ss = avl_find(&sm->sm_root, &ssearch, &where);
108 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
109 zfs_panic_recover("zfs: allocating allocated segment"
110 "(offset=%llu size=%llu)\n",
111 (longlong_t)start, (longlong_t)size);
114 if (ss != NULL && space_map_last_hope) {
115 uint64_t sstart, ssize;
117 if (ss->ss_start > start)
118 sstart = ss->ss_start;
121 if (ss->ss_end > end)
122 ssize = end - sstart;
124 ssize = ss->ss_end - sstart;
126 "Removing colliding space_map range (start=%ju end=%ju). Good luck!",
127 (uintmax_t)sstart, (uintmax_t)(sstart + ssize));
128 space_map_remove(sm, sstart, ssize);
132 /* Make sure we don't overlap with either of our neighbors */
135 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
136 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
138 merge_before = (ss_before != NULL && ss_before->ss_end == start);
139 merge_after = (ss_after != NULL && ss_after->ss_start == end);
141 if (merge_before && merge_after) {
142 avl_remove(&sm->sm_root, ss_before);
143 if (sm->sm_pp_root) {
144 avl_remove(sm->sm_pp_root, ss_before);
145 avl_remove(sm->sm_pp_root, ss_after);
147 ss_after->ss_start = ss_before->ss_start;
148 kmem_free(ss_before, sizeof (*ss_before));
150 } else if (merge_before) {
151 ss_before->ss_end = end;
153 avl_remove(sm->sm_pp_root, ss_before);
155 } else if (merge_after) {
156 ss_after->ss_start = start;
158 avl_remove(sm->sm_pp_root, ss_after);
161 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
162 ss->ss_start = start;
164 avl_insert(&sm->sm_root, ss, where);
168 avl_add(sm->sm_pp_root, ss);
170 sm->sm_space += size;
174 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
176 space_seg_t ssearch, *ss, *newseg;
177 uint64_t end = start + size;
178 int left_over, right_over;
180 ASSERT(MUTEX_HELD(sm->sm_lock));
182 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
183 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
185 ssearch.ss_start = start;
186 ssearch.ss_end = end;
187 ss = avl_find(&sm->sm_root, &ssearch, NULL);
189 /* Make sure we completely overlap with someone */
191 zfs_panic_recover("zfs: freeing free segment "
192 "(offset=%llu size=%llu)",
193 (longlong_t)start, (longlong_t)size);
196 VERIFY3U(ss->ss_start, <=, start);
197 VERIFY3U(ss->ss_end, >=, end);
198 VERIFY(sm->sm_space - size < sm->sm_size);
200 left_over = (ss->ss_start != start);
201 right_over = (ss->ss_end != end);
204 avl_remove(sm->sm_pp_root, ss);
206 if (left_over && right_over) {
207 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
208 newseg->ss_start = end;
209 newseg->ss_end = ss->ss_end;
211 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
213 avl_add(sm->sm_pp_root, newseg);
214 } else if (left_over) {
216 } else if (right_over) {
219 avl_remove(&sm->sm_root, ss);
220 kmem_free(ss, sizeof (*ss));
224 if (sm->sm_pp_root && ss != NULL)
225 avl_add(sm->sm_pp_root, ss);
227 sm->sm_space -= size;
231 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
234 space_seg_t ssearch, *ss;
235 uint64_t end = start + size;
237 ASSERT(MUTEX_HELD(sm->sm_lock));
239 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
240 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
242 ssearch.ss_start = start;
243 ssearch.ss_end = end;
244 ss = avl_find(&sm->sm_root, &ssearch, &where);
246 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
250 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
255 ASSERT(MUTEX_HELD(sm->sm_lock));
257 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
259 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
260 kmem_free(ss, sizeof (*ss));
266 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
270 ASSERT(MUTEX_HELD(sm->sm_lock));
272 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
273 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
277 * Wait for any in-progress space_map_load() to complete.
280 space_map_load_wait(space_map_t *sm)
282 ASSERT(MUTEX_HELD(sm->sm_lock));
284 while (sm->sm_loading) {
285 ASSERT(!sm->sm_loaded);
286 cv_wait(&sm->sm_load_cv, sm->sm_lock);
291 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
292 * The caller must be OK with this.
295 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
296 space_map_obj_t *smo, objset_t *os)
298 uint64_t *entry, *entry_map, *entry_map_end;
299 uint64_t bufsize, size, offset, end, space;
300 uint64_t mapstart = sm->sm_start;
303 ASSERT(MUTEX_HELD(sm->sm_lock));
304 ASSERT(!sm->sm_loaded);
305 ASSERT(!sm->sm_loading);
307 sm->sm_loading = B_TRUE;
308 end = smo->smo_objsize;
309 space = smo->smo_alloc;
311 ASSERT(sm->sm_ops == NULL);
312 VERIFY3U(sm->sm_space, ==, 0);
314 if (maptype == SM_FREE) {
315 space_map_add(sm, sm->sm_start, sm->sm_size);
316 space = sm->sm_size - space;
319 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
320 entry_map = zio_buf_alloc(bufsize);
322 mutex_exit(sm->sm_lock);
324 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
325 mutex_enter(sm->sm_lock);
327 for (offset = 0; offset < end; offset += bufsize) {
328 size = MIN(end - offset, bufsize);
329 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
332 dprintf("object=%llu offset=%llx size=%llx\n",
333 smo->smo_object, offset, size);
335 mutex_exit(sm->sm_lock);
336 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
338 mutex_enter(sm->sm_lock);
342 entry_map_end = entry_map + (size / sizeof (uint64_t));
343 for (entry = entry_map; entry < entry_map_end; entry++) {
346 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
349 (SM_TYPE_DECODE(e) == maptype ?
350 space_map_add : space_map_remove)(sm,
351 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
352 SM_RUN_DECODE(e) << sm->sm_shift);
357 VERIFY3U(sm->sm_space, ==, space);
359 sm->sm_loaded = B_TRUE;
364 space_map_vacate(sm, NULL, NULL);
367 zio_buf_free(entry_map, bufsize);
369 sm->sm_loading = B_FALSE;
371 cv_broadcast(&sm->sm_load_cv);
377 space_map_unload(space_map_t *sm)
379 ASSERT(MUTEX_HELD(sm->sm_lock));
381 if (sm->sm_loaded && sm->sm_ops != NULL)
382 sm->sm_ops->smop_unload(sm);
384 sm->sm_loaded = B_FALSE;
387 space_map_vacate(sm, NULL, NULL);
391 space_map_maxsize(space_map_t *sm)
393 ASSERT(sm->sm_ops != NULL);
394 return (sm->sm_ops->smop_max(sm));
398 space_map_alloc(space_map_t *sm, uint64_t size)
402 start = sm->sm_ops->smop_alloc(sm, size);
404 space_map_remove(sm, start, size);
409 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
411 sm->sm_ops->smop_claim(sm, start, size);
412 space_map_remove(sm, start, size);
416 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
418 space_map_add(sm, start, size);
419 sm->sm_ops->smop_free(sm, start, size);
423 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
426 space_map_sync(space_map_t *sm, uint8_t maptype,
427 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
429 spa_t *spa = dmu_objset_spa(os);
432 uint64_t bufsize, start, size, run_len;
433 uint64_t *entry, *entry_map, *entry_map_end;
435 ASSERT(MUTEX_HELD(sm->sm_lock));
437 if (sm->sm_space == 0)
440 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
441 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
442 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
445 if (maptype == SM_ALLOC)
446 smo->smo_alloc += sm->sm_space;
448 smo->smo_alloc -= sm->sm_space;
450 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
451 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
452 entry_map = zio_buf_alloc(bufsize);
453 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
456 *entry++ = SM_DEBUG_ENCODE(1) |
457 SM_DEBUG_ACTION_ENCODE(maptype) |
458 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
459 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
461 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
462 size = ss->ss_end - ss->ss_start;
463 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
465 sm->sm_space -= size;
466 size >>= sm->sm_shift;
469 run_len = MIN(size, SM_RUN_MAX);
471 if (entry == entry_map_end) {
472 mutex_exit(sm->sm_lock);
473 dmu_write(os, smo->smo_object, smo->smo_objsize,
474 bufsize, entry_map, tx);
475 mutex_enter(sm->sm_lock);
476 smo->smo_objsize += bufsize;
480 *entry++ = SM_OFFSET_ENCODE(start) |
481 SM_TYPE_ENCODE(maptype) |
482 SM_RUN_ENCODE(run_len);
487 kmem_free(ss, sizeof (*ss));
490 if (entry != entry_map) {
491 size = (entry - entry_map) * sizeof (uint64_t);
492 mutex_exit(sm->sm_lock);
493 dmu_write(os, smo->smo_object, smo->smo_objsize,
494 size, entry_map, tx);
495 mutex_enter(sm->sm_lock);
496 smo->smo_objsize += size;
499 zio_buf_free(entry_map, bufsize);
501 VERIFY3U(sm->sm_space, ==, 0);
505 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
507 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
509 smo->smo_objsize = 0;
514 * Space map reference trees.
516 * A space map is a collection of integers. Every integer is either
517 * in the map, or it's not. A space map reference tree generalizes
518 * the idea: it allows its members to have arbitrary reference counts,
519 * as opposed to the implicit reference count of 0 or 1 in a space map.
520 * This representation comes in handy when computing the union or
521 * intersection of multiple space maps. For example, the union of
522 * N space maps is the subset of the reference tree with refcnt >= 1.
523 * The intersection of N space maps is the subset with refcnt >= N.
525 * [It's very much like a Fourier transform. Unions and intersections
526 * are hard to perform in the 'space map domain', so we convert the maps
527 * into the 'reference count domain', where it's trivial, then invert.]
529 * vdev_dtl_reassess() uses computations of this form to determine
530 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
531 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
532 * has an outage wherever refcnt >= vdev_children.
535 space_map_ref_compare(const void *x1, const void *x2)
537 const space_ref_t *sr1 = x1;
538 const space_ref_t *sr2 = x2;
540 if (sr1->sr_offset < sr2->sr_offset)
542 if (sr1->sr_offset > sr2->sr_offset)
554 space_map_ref_create(avl_tree_t *t)
556 avl_create(t, space_map_ref_compare,
557 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
561 space_map_ref_destroy(avl_tree_t *t)
566 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
567 kmem_free(sr, sizeof (*sr));
573 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
577 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
578 sr->sr_offset = offset;
579 sr->sr_refcnt = refcnt;
585 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
588 space_map_ref_add_node(t, start, refcnt);
589 space_map_ref_add_node(t, end, -refcnt);
593 * Convert (or add) a space map into a reference tree.
596 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
600 ASSERT(MUTEX_HELD(sm->sm_lock));
602 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
603 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
607 * Convert a reference tree into a space map. The space map will contain
608 * all members of the reference tree for which refcnt >= minref.
611 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
613 uint64_t start = -1ULL;
617 ASSERT(MUTEX_HELD(sm->sm_lock));
619 space_map_vacate(sm, NULL, NULL);
621 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
622 refcnt += sr->sr_refcnt;
623 if (refcnt >= minref) {
624 if (start == -1ULL) {
625 start = sr->sr_offset;
628 if (start != -1ULL) {
629 uint64_t end = sr->sr_offset;
630 ASSERT(start <= end);
632 space_map_add(sm, start, end - start);
638 ASSERT(start == -1ULL);