space_map.c revision 7bfdf011e081684f853a3242d0296695110d9d84
1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23 * Use is subject to license terms.
24 */
25
26#include <sys/zfs_context.h>
27#include <sys/spa.h>
28#include <sys/dmu.h>
29#include <sys/zio.h>
30#include <sys/space_map.h>
31
32/*
33 * Space map routines.
34 * NOTE: caller is responsible for all locking.
35 */
36static int
37space_map_seg_compare(const void *x1, const void *x2)
38{
39	const space_seg_t *s1 = x1;
40	const space_seg_t *s2 = x2;
41
42	if (s1->ss_start < s2->ss_start) {
43		if (s1->ss_end > s2->ss_start)
44			return (0);
45		return (-1);
46	}
47	if (s1->ss_start > s2->ss_start) {
48		if (s1->ss_start < s2->ss_end)
49			return (0);
50		return (1);
51	}
52	return (0);
53}
54
55void
56space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
57	kmutex_t *lp)
58{
59	bzero(sm, sizeof (*sm));
60
61	cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
62
63	avl_create(&sm->sm_root, space_map_seg_compare,
64	    sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
65
66	sm->sm_start = start;
67	sm->sm_size = size;
68	sm->sm_shift = shift;
69	sm->sm_lock = lp;
70}
71
72void
73space_map_destroy(space_map_t *sm)
74{
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);
79}
80
81void
82space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
83{
84	avl_index_t where;
85	space_seg_t ssearch, *ss_before, *ss_after, *ss;
86	uint64_t end = start + size;
87	int merge_before, merge_after;
88
89	ASSERT(MUTEX_HELD(sm->sm_lock));
90	VERIFY(size != 0);
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);
96
97	ssearch.ss_start = start;
98	ssearch.ss_end = end;
99	ss = avl_find(&sm->sm_root, &ssearch, &where);
100
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);
105		return;
106	}
107
108	/* Make sure we don't overlap with either of our neighbors */
109	VERIFY(ss == NULL);
110
111	ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
112	ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
113
114	merge_before = (ss_before != NULL && ss_before->ss_end == start);
115	merge_after = (ss_after != NULL && ss_after->ss_start == end);
116
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);
122		}
123		ss_after->ss_start = ss_before->ss_start;
124		kmem_free(ss_before, sizeof (*ss_before));
125		ss = ss_after;
126	} else if (merge_before) {
127		ss_before->ss_end = end;
128		if (sm->sm_pp_root)
129			avl_remove(sm->sm_pp_root, ss_before);
130		ss = ss_before;
131	} else if (merge_after) {
132		ss_after->ss_start = start;
133		if (sm->sm_pp_root)
134			avl_remove(sm->sm_pp_root, ss_after);
135		ss = ss_after;
136	} else {
137		ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
138		ss->ss_start = start;
139		ss->ss_end = end;
140		avl_insert(&sm->sm_root, ss, where);
141	}
142
143	if (sm->sm_pp_root)
144		avl_add(sm->sm_pp_root, ss);
145
146	sm->sm_space += size;
147}
148
149void
150space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
151{
152	avl_index_t where;
153	space_seg_t ssearch, *ss, *newseg;
154	uint64_t end = start + size;
155	int left_over, right_over;
156
157	ASSERT(MUTEX_HELD(sm->sm_lock));
158	VERIFY(size != 0);
159	VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
160	VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
161
162	ssearch.ss_start = start;
163	ssearch.ss_end = end;
164	ss = avl_find(&sm->sm_root, &ssearch, &where);
165
166	/* Make sure we completely overlap with someone */
167	if (ss == NULL) {
168		zfs_panic_recover("zfs: freeing free segment "
169		    "(offset=%llu size=%llu)",
170		    (longlong_t)start, (longlong_t)size);
171		return;
172	}
173	VERIFY3U(ss->ss_start, <=, start);
174	VERIFY3U(ss->ss_end, >=, end);
175	VERIFY(sm->sm_space - size <= sm->sm_size);
176
177	left_over = (ss->ss_start != start);
178	right_over = (ss->ss_end != end);
179
180	if (sm->sm_pp_root)
181		avl_remove(sm->sm_pp_root, ss);
182
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;
187		ss->ss_end = start;
188		avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
189		if (sm->sm_pp_root)
190			avl_add(sm->sm_pp_root, newseg);
191	} else if (left_over) {
192		ss->ss_end = start;
193	} else if (right_over) {
194		ss->ss_start = end;
195	} else {
196		avl_remove(&sm->sm_root, ss);
197		kmem_free(ss, sizeof (*ss));
198		ss = NULL;
199	}
200
201	if (sm->sm_pp_root && ss != NULL)
202		avl_add(sm->sm_pp_root, ss);
203
204	sm->sm_space -= size;
205}
206
207boolean_t
208space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
209{
210	avl_index_t where;
211	space_seg_t ssearch, *ss;
212	uint64_t end = start + size;
213
214	ASSERT(MUTEX_HELD(sm->sm_lock));
215	VERIFY(size != 0);
216	VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
217	VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
218
219	ssearch.ss_start = start;
220	ssearch.ss_end = end;
221	ss = avl_find(&sm->sm_root, &ssearch, &where);
222
223	return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
224}
225
226void
227space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
228{
229	space_seg_t *ss;
230	void *cookie = NULL;
231
232	ASSERT(MUTEX_HELD(sm->sm_lock));
233
234	while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
235		if (func != NULL)
236			func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
237		kmem_free(ss, sizeof (*ss));
238	}
239	sm->sm_space = 0;
240}
241
242void
243space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
244{
245	space_seg_t *ss;
246
247	ASSERT(MUTEX_HELD(sm->sm_lock));
248
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);
251}
252
253/*
254 * Wait for any in-progress space_map_load() to complete.
255 */
256void
257space_map_load_wait(space_map_t *sm)
258{
259	ASSERT(MUTEX_HELD(sm->sm_lock));
260
261	while (sm->sm_loading)
262		cv_wait(&sm->sm_load_cv, sm->sm_lock);
263}
264
265/*
266 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
267 * The caller must be OK with this.
268 */
269int
270space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
271	space_map_obj_t *smo, objset_t *os)
272{
273	uint64_t *entry, *entry_map, *entry_map_end;
274	uint64_t bufsize, size, offset, end, space;
275	uint64_t mapstart = sm->sm_start;
276	int error = 0;
277
278	ASSERT(MUTEX_HELD(sm->sm_lock));
279
280	space_map_load_wait(sm);
281
282	if (sm->sm_loaded)
283		return (0);
284
285	sm->sm_loading = B_TRUE;
286	end = smo->smo_objsize;
287	space = smo->smo_alloc;
288
289	ASSERT(sm->sm_ops == NULL);
290	VERIFY3U(sm->sm_space, ==, 0);
291
292	if (maptype == SM_FREE) {
293		space_map_add(sm, sm->sm_start, sm->sm_size);
294		space = sm->sm_size - space;
295	}
296
297	bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
298	entry_map = zio_buf_alloc(bufsize);
299
300	mutex_exit(sm->sm_lock);
301	if (end > bufsize)
302		dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
303	mutex_enter(sm->sm_lock);
304
305	for (offset = 0; offset < end; offset += bufsize) {
306		size = MIN(end - offset, bufsize);
307		VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
308		VERIFY(size != 0);
309
310		dprintf("object=%llu  offset=%llx  size=%llx\n",
311		    smo->smo_object, offset, size);
312
313		mutex_exit(sm->sm_lock);
314		error = dmu_read(os, smo->smo_object, offset, size, entry_map,
315		    DMU_READ_PREFETCH);
316		mutex_enter(sm->sm_lock);
317		if (error != 0)
318			break;
319
320		entry_map_end = entry_map + (size / sizeof (uint64_t));
321		for (entry = entry_map; entry < entry_map_end; entry++) {
322			uint64_t e = *entry;
323
324			if (SM_DEBUG_DECODE(e))		/* Skip debug entries */
325				continue;
326
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);
331		}
332	}
333
334	if (error == 0) {
335		VERIFY3U(sm->sm_space, ==, space);
336
337		sm->sm_loaded = B_TRUE;
338		sm->sm_ops = ops;
339		if (ops != NULL)
340			ops->smop_load(sm);
341	} else {
342		if (ops != NULL)
343			ops->smop_unload(sm);
344		sm->sm_ops = NULL;
345		space_map_vacate(sm, NULL, NULL);
346	}
347
348	zio_buf_free(entry_map, bufsize);
349
350	sm->sm_loading = B_FALSE;
351
352	cv_broadcast(&sm->sm_load_cv);
353
354	return (error);
355}
356
357void
358space_map_unload(space_map_t *sm)
359{
360	ASSERT(MUTEX_HELD(sm->sm_lock));
361
362	if (sm->sm_loaded && sm->sm_ops != NULL)
363		sm->sm_ops->smop_unload(sm);
364
365	sm->sm_loaded = B_FALSE;
366	sm->sm_ops = NULL;
367
368	space_map_vacate(sm, NULL, NULL);
369}
370
371uint64_t
372space_map_maxsize(space_map_t *sm)
373{
374	if (sm->sm_loaded && sm->sm_ops != NULL)
375		return (sm->sm_ops->smop_max(sm));
376	else
377		return (-1ULL);
378}
379
380uint64_t
381space_map_alloc(space_map_t *sm, uint64_t size)
382{
383	uint64_t start;
384
385	start = sm->sm_ops->smop_alloc(sm, size);
386	if (start != -1ULL)
387		space_map_remove(sm, start, size);
388	return (start);
389}
390
391void
392space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
393{
394	sm->sm_ops->smop_claim(sm, start, size);
395	space_map_remove(sm, start, size);
396}
397
398void
399space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
400{
401	space_map_add(sm, start, size);
402	sm->sm_ops->smop_free(sm, start, size);
403}
404
405/*
406 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
407 */
408void
409space_map_sync(space_map_t *sm, uint8_t maptype,
410	space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
411{
412	spa_t *spa = dmu_objset_spa(os);
413	void *cookie = NULL;
414	space_seg_t *ss;
415	uint64_t bufsize, start, size, run_len;
416	uint64_t *entry, *entry_map, *entry_map_end;
417
418	ASSERT(MUTEX_HELD(sm->sm_lock));
419
420	if (sm->sm_space == 0)
421		return;
422
423	dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
424	    smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
425	    maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
426	    sm->sm_space);
427
428	if (maptype == SM_ALLOC)
429		smo->smo_alloc += sm->sm_space;
430	else
431		smo->smo_alloc -= sm->sm_space;
432
433	bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
434	bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
435	entry_map = zio_buf_alloc(bufsize);
436	entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
437	entry = entry_map;
438
439	*entry++ = SM_DEBUG_ENCODE(1) |
440	    SM_DEBUG_ACTION_ENCODE(maptype) |
441	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
442	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
443
444	while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
445		size = ss->ss_end - ss->ss_start;
446		start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
447
448		sm->sm_space -= size;
449		size >>= sm->sm_shift;
450
451		while (size) {
452			run_len = MIN(size, SM_RUN_MAX);
453
454			if (entry == entry_map_end) {
455				mutex_exit(sm->sm_lock);
456				dmu_write(os, smo->smo_object, smo->smo_objsize,
457				    bufsize, entry_map, tx);
458				mutex_enter(sm->sm_lock);
459				smo->smo_objsize += bufsize;
460				entry = entry_map;
461			}
462
463			*entry++ = SM_OFFSET_ENCODE(start) |
464			    SM_TYPE_ENCODE(maptype) |
465			    SM_RUN_ENCODE(run_len);
466
467			start += run_len;
468			size -= run_len;
469		}
470		kmem_free(ss, sizeof (*ss));
471	}
472
473	if (entry != entry_map) {
474		size = (entry - entry_map) * sizeof (uint64_t);
475		mutex_exit(sm->sm_lock);
476		dmu_write(os, smo->smo_object, smo->smo_objsize,
477		    size, entry_map, tx);
478		mutex_enter(sm->sm_lock);
479		smo->smo_objsize += size;
480	}
481
482	zio_buf_free(entry_map, bufsize);
483
484	VERIFY3U(sm->sm_space, ==, 0);
485}
486
487void
488space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
489{
490	VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
491
492	smo->smo_objsize = 0;
493	smo->smo_alloc = 0;
494}
495
496/*
497 * Space map reference trees.
498 *
499 * A space map is a collection of integers.  Every integer is either
500 * in the map, or it's not.  A space map reference tree generalizes
501 * the idea: it allows its members to have arbitrary reference counts,
502 * as opposed to the implicit reference count of 0 or 1 in a space map.
503 * This representation comes in handy when computing the union or
504 * intersection of multiple space maps.  For example, the union of
505 * N space maps is the subset of the reference tree with refcnt >= 1.
506 * The intersection of N space maps is the subset with refcnt >= N.
507 *
508 * [It's very much like a Fourier transform.  Unions and intersections
509 * are hard to perform in the 'space map domain', so we convert the maps
510 * into the 'reference count domain', where it's trivial, then invert.]
511 *
512 * vdev_dtl_reassess() uses computations of this form to determine
513 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
514 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
515 * has an outage wherever refcnt >= vdev_children.
516 */
517static int
518space_map_ref_compare(const void *x1, const void *x2)
519{
520	const space_ref_t *sr1 = x1;
521	const space_ref_t *sr2 = x2;
522
523	if (sr1->sr_offset < sr2->sr_offset)
524		return (-1);
525	if (sr1->sr_offset > sr2->sr_offset)
526		return (1);
527
528	if (sr1 < sr2)
529		return (-1);
530	if (sr1 > sr2)
531		return (1);
532
533	return (0);
534}
535
536void
537space_map_ref_create(avl_tree_t *t)
538{
539	avl_create(t, space_map_ref_compare,
540	    sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
541}
542
543void
544space_map_ref_destroy(avl_tree_t *t)
545{
546	space_ref_t *sr;
547	void *cookie = NULL;
548
549	while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
550		kmem_free(sr, sizeof (*sr));
551
552	avl_destroy(t);
553}
554
555static void
556space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
557{
558	space_ref_t *sr;
559
560	sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
561	sr->sr_offset = offset;
562	sr->sr_refcnt = refcnt;
563
564	avl_add(t, sr);
565}
566
567void
568space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
569	int64_t refcnt)
570{
571	space_map_ref_add_node(t, start, refcnt);
572	space_map_ref_add_node(t, end, -refcnt);
573}
574
575/*
576 * Convert (or add) a space map into a reference tree.
577 */
578void
579space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
580{
581	space_seg_t *ss;
582
583	ASSERT(MUTEX_HELD(sm->sm_lock));
584
585	for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
586		space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
587}
588
589/*
590 * Convert a reference tree into a space map.  The space map will contain
591 * all members of the reference tree for which refcnt >= minref.
592 */
593void
594space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
595{
596	uint64_t start = -1ULL;
597	int64_t refcnt = 0;
598	space_ref_t *sr;
599
600	ASSERT(MUTEX_HELD(sm->sm_lock));
601
602	space_map_vacate(sm, NULL, NULL);
603
604	for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
605		refcnt += sr->sr_refcnt;
606		if (refcnt >= minref) {
607			if (start == -1ULL) {
608				start = sr->sr_offset;
609			}
610		} else {
611			if (start != -1ULL) {
612				uint64_t end = sr->sr_offset;
613				ASSERT(start <= end);
614				if (end > start)
615					space_map_add(sm, start, end - start);
616				start = -1ULL;
617			}
618		}
619	}
620	ASSERT(refcnt == 0);
621	ASSERT(start == -1ULL);
622}
623