space_map.c revision 2e4c998613148111f2fc5371085331ffb39122ff
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 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
27 */
28
29#include <sys/zfs_context.h>
30#include <sys/spa.h>
31#include <sys/dmu.h>
32#include <sys/dmu_tx.h>
33#include <sys/dnode.h>
34#include <sys/dsl_pool.h>
35#include <sys/zio.h>
36#include <sys/space_map.h>
37#include <sys/refcount.h>
38#include <sys/zfeature.h>
39
40/*
41 * This value controls how the space map's block size is allowed to grow.
42 * If the value is set to the same size as SPACE_MAP_INITIAL_BLOCKSIZE then
43 * the space map block size will remain fixed. Setting this value to something
44 * greater than SPACE_MAP_INITIAL_BLOCKSIZE will allow the space map to
45 * increase its block size as needed. To maintain backwards compatibilty the
46 * space map's block size must be a power of 2 and SPACE_MAP_INITIAL_BLOCKSIZE
47 * or larger.
48 */
49int space_map_max_blksz = (1 << 12);
50
51/*
52 * Load the space map disk into the specified range tree. Segments of maptype
53 * are added to the range tree, other segment types are removed.
54 *
55 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
56 * The caller must be OK with this.
57 */
58int
59space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
60{
61	uint64_t *entry, *entry_map, *entry_map_end;
62	uint64_t bufsize, size, offset, end, space;
63	int error = 0;
64
65	ASSERT(MUTEX_HELD(sm->sm_lock));
66
67	end = space_map_length(sm);
68	space = space_map_allocated(sm);
69
70	VERIFY0(range_tree_space(rt));
71
72	if (maptype == SM_FREE) {
73		range_tree_add(rt, sm->sm_start, sm->sm_size);
74		space = sm->sm_size - space;
75	}
76
77	bufsize = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
78	entry_map = zio_buf_alloc(bufsize);
79
80	mutex_exit(sm->sm_lock);
81	if (end > bufsize) {
82		dmu_prefetch(sm->sm_os, space_map_object(sm), bufsize,
83		    end - bufsize);
84	}
85	mutex_enter(sm->sm_lock);
86
87	for (offset = 0; offset < end; offset += bufsize) {
88		size = MIN(end - offset, bufsize);
89		VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
90		VERIFY(size != 0);
91		ASSERT3U(sm->sm_blksz, !=, 0);
92
93		dprintf("object=%llu  offset=%llx  size=%llx\n",
94		    space_map_object(sm), offset, size);
95
96		mutex_exit(sm->sm_lock);
97		error = dmu_read(sm->sm_os, space_map_object(sm), offset, size,
98		    entry_map, DMU_READ_PREFETCH);
99		mutex_enter(sm->sm_lock);
100		if (error != 0)
101			break;
102
103		entry_map_end = entry_map + (size / sizeof (uint64_t));
104		for (entry = entry_map; entry < entry_map_end; entry++) {
105			uint64_t e = *entry;
106			uint64_t offset, size;
107
108			if (SM_DEBUG_DECODE(e))		/* Skip debug entries */
109				continue;
110
111			offset = (SM_OFFSET_DECODE(e) << sm->sm_shift) +
112			    sm->sm_start;
113			size = SM_RUN_DECODE(e) << sm->sm_shift;
114
115			VERIFY0(P2PHASE(offset, 1ULL << sm->sm_shift));
116			VERIFY0(P2PHASE(size, 1ULL << sm->sm_shift));
117			VERIFY3U(offset, >=, sm->sm_start);
118			VERIFY3U(offset + size, <=, sm->sm_start + sm->sm_size);
119			if (SM_TYPE_DECODE(e) == maptype) {
120				VERIFY3U(range_tree_space(rt) + size, <=,
121				    sm->sm_size);
122				range_tree_add(rt, offset, size);
123			} else {
124				range_tree_remove(rt, offset, size);
125			}
126		}
127	}
128
129	if (error == 0)
130		VERIFY3U(range_tree_space(rt), ==, space);
131	else
132		range_tree_vacate(rt, NULL, NULL);
133
134	zio_buf_free(entry_map, bufsize);
135	return (error);
136}
137
138void
139space_map_histogram_clear(space_map_t *sm)
140{
141	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
142		return;
143
144	bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
145}
146
147boolean_t
148space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
149{
150	/*
151	 * Verify that the in-core range tree does not have any
152	 * ranges smaller than our sm_shift size.
153	 */
154	for (int i = 0; i < sm->sm_shift; i++) {
155		if (rt->rt_histogram[i] != 0)
156			return (B_FALSE);
157	}
158	return (B_TRUE);
159}
160
161void
162space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
163{
164	int idx = 0;
165
166	ASSERT(MUTEX_HELD(rt->rt_lock));
167	ASSERT(dmu_tx_is_syncing(tx));
168	VERIFY3U(space_map_object(sm), !=, 0);
169
170	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
171		return;
172
173	dmu_buf_will_dirty(sm->sm_dbuf, tx);
174
175	ASSERT(space_map_histogram_verify(sm, rt));
176
177	/*
178	 * Transfer the content of the range tree histogram to the space
179	 * map histogram. The space map histogram contains 32 buckets ranging
180	 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
181	 * however, can represent ranges from 2^0 to 2^63. Since the space
182	 * map only cares about allocatable blocks (minimum of sm_shift) we
183	 * can safely ignore all ranges in the range tree smaller than sm_shift.
184	 */
185	for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
186
187		/*
188		 * Since the largest histogram bucket in the space map is
189		 * 2^(32+sm_shift-1), we need to normalize the values in
190		 * the range tree for any bucket larger than that size. For
191		 * example given an sm_shift of 9, ranges larger than 2^40
192		 * would get normalized as if they were 1TB ranges. Assume
193		 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
194		 * the calculation below would normalize this to 5 * 2^4 (16).
195		 */
196		ASSERT3U(i, >=, idx + sm->sm_shift);
197		sm->sm_phys->smp_histogram[idx] +=
198		    rt->rt_histogram[i] << (i - idx - sm->sm_shift);
199
200		/*
201		 * Increment the space map's index as long as we haven't
202		 * reached the maximum bucket size. Accumulate all ranges
203		 * larger than the max bucket size into the last bucket.
204		 */
205		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
206			ASSERT3U(idx + sm->sm_shift, ==, i);
207			idx++;
208			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
209		}
210	}
211}
212
213uint64_t
214space_map_entries(space_map_t *sm, range_tree_t *rt)
215{
216	avl_tree_t *t = &rt->rt_root;
217	range_seg_t *rs;
218	uint64_t size, entries;
219
220	/*
221	 * All space_maps always have a debug entry so account for it here.
222	 */
223	entries = 1;
224
225	/*
226	 * Traverse the range tree and calculate the number of space map
227	 * entries that would be required to write out the range tree.
228	 */
229	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
230		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
231		entries += howmany(size, SM_RUN_MAX);
232	}
233	return (entries);
234}
235
236void
237space_map_set_blocksize(space_map_t *sm, uint64_t size, dmu_tx_t *tx)
238{
239	uint32_t blksz;
240	u_longlong_t blocks;
241
242	ASSERT3U(sm->sm_blksz, !=, 0);
243	ASSERT3U(space_map_object(sm), !=, 0);
244	ASSERT(sm->sm_dbuf != NULL);
245	VERIFY(ISP2(space_map_max_blksz));
246
247	if (sm->sm_blksz >= space_map_max_blksz)
248		return;
249
250	/*
251	 * The object contains more than one block so we can't adjust
252	 * its size.
253	 */
254	if (sm->sm_phys->smp_objsize > sm->sm_blksz)
255		return;
256
257	if (size > sm->sm_blksz) {
258		uint64_t newsz;
259
260		/*
261		 * Older software versions treat space map blocks as fixed
262		 * entities. The DMU is capable of handling different block
263		 * sizes making it possible for us to increase the
264		 * block size and maintain backwards compatibility. The
265		 * caveat is that the new block sizes must be a
266		 * power of 2 so that old software can append to the file,
267		 * adding more blocks. The block size can grow until it
268		 * reaches space_map_max_blksz.
269		 */
270		newsz = ISP2(size) ? size : 1ULL << highbit64(size);
271		if (newsz > space_map_max_blksz)
272			newsz = space_map_max_blksz;
273
274		VERIFY0(dmu_object_set_blocksize(sm->sm_os,
275		    space_map_object(sm), newsz, 0, tx));
276		dmu_object_size_from_db(sm->sm_dbuf, &blksz, &blocks);
277
278		zfs_dbgmsg("txg %llu, spa %s, increasing blksz from %d to %d",
279		    dmu_tx_get_txg(tx), spa_name(dmu_objset_spa(sm->sm_os)),
280		    sm->sm_blksz, blksz);
281
282		VERIFY3U(newsz, ==, blksz);
283		VERIFY3U(sm->sm_blksz, <, blksz);
284		sm->sm_blksz = blksz;
285	}
286}
287
288/*
289 * Note: space_map_write() will drop sm_lock across dmu_write() calls.
290 */
291void
292space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
293    dmu_tx_t *tx)
294{
295	objset_t *os = sm->sm_os;
296	spa_t *spa = dmu_objset_spa(os);
297	avl_tree_t *t = &rt->rt_root;
298	range_seg_t *rs;
299	uint64_t size, total, rt_space, nodes;
300	uint64_t *entry, *entry_map, *entry_map_end;
301	uint64_t newsz, expected_entries, actual_entries = 1;
302
303	ASSERT(MUTEX_HELD(rt->rt_lock));
304	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
305	VERIFY3U(space_map_object(sm), !=, 0);
306	dmu_buf_will_dirty(sm->sm_dbuf, tx);
307
308	/*
309	 * This field is no longer necessary since the in-core space map
310	 * now contains the object number but is maintained for backwards
311	 * compatibility.
312	 */
313	sm->sm_phys->smp_object = sm->sm_object;
314
315	if (range_tree_space(rt) == 0) {
316		VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
317		return;
318	}
319
320	if (maptype == SM_ALLOC)
321		sm->sm_phys->smp_alloc += range_tree_space(rt);
322	else
323		sm->sm_phys->smp_alloc -= range_tree_space(rt);
324
325	expected_entries = space_map_entries(sm, rt);
326
327	/*
328	 * Calculate the new size for the space map on-disk and see if
329	 * we can grow the block size to accommodate the new size.
330	 */
331	newsz = sm->sm_phys->smp_objsize + expected_entries * sizeof (uint64_t);
332	space_map_set_blocksize(sm, newsz, tx);
333
334	entry_map = zio_buf_alloc(sm->sm_blksz);
335	entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
336	entry = entry_map;
337
338	*entry++ = SM_DEBUG_ENCODE(1) |
339	    SM_DEBUG_ACTION_ENCODE(maptype) |
340	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
341	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
342
343	total = 0;
344	nodes = avl_numnodes(&rt->rt_root);
345	rt_space = range_tree_space(rt);
346	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
347		uint64_t start;
348
349		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
350		start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
351
352		total += size << sm->sm_shift;
353
354		while (size != 0) {
355			uint64_t run_len;
356
357			run_len = MIN(size, SM_RUN_MAX);
358
359			if (entry == entry_map_end) {
360				mutex_exit(rt->rt_lock);
361				dmu_write(os, space_map_object(sm),
362				    sm->sm_phys->smp_objsize, sm->sm_blksz,
363				    entry_map, tx);
364				mutex_enter(rt->rt_lock);
365				sm->sm_phys->smp_objsize += sm->sm_blksz;
366				entry = entry_map;
367			}
368
369			*entry++ = SM_OFFSET_ENCODE(start) |
370			    SM_TYPE_ENCODE(maptype) |
371			    SM_RUN_ENCODE(run_len);
372
373			start += run_len;
374			size -= run_len;
375			actual_entries++;
376		}
377	}
378
379	if (entry != entry_map) {
380		size = (entry - entry_map) * sizeof (uint64_t);
381		mutex_exit(rt->rt_lock);
382		dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
383		    size, entry_map, tx);
384		mutex_enter(rt->rt_lock);
385		sm->sm_phys->smp_objsize += size;
386	}
387	ASSERT3U(expected_entries, ==, actual_entries);
388
389	/*
390	 * Ensure that the space_map's accounting wasn't changed
391	 * while we were in the middle of writing it out.
392	 */
393	VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
394	VERIFY3U(range_tree_space(rt), ==, rt_space);
395	VERIFY3U(range_tree_space(rt), ==, total);
396
397	zio_buf_free(entry_map, sm->sm_blksz);
398}
399
400static int
401space_map_open_impl(space_map_t *sm)
402{
403	int error;
404	u_longlong_t blocks;
405
406	error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
407	if (error)
408		return (error);
409
410	dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
411	sm->sm_phys = sm->sm_dbuf->db_data;
412	return (0);
413}
414
415int
416space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
417    uint64_t start, uint64_t size, uint8_t shift, kmutex_t *lp)
418{
419	space_map_t *sm;
420	int error;
421
422	ASSERT(*smp == NULL);
423	ASSERT(os != NULL);
424	ASSERT(object != 0);
425
426	sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
427
428	sm->sm_start = start;
429	sm->sm_size = size;
430	sm->sm_shift = shift;
431	sm->sm_lock = lp;
432	sm->sm_os = os;
433	sm->sm_object = object;
434
435	error = space_map_open_impl(sm);
436	if (error != 0) {
437		space_map_close(sm);
438		return (error);
439	}
440
441	*smp = sm;
442
443	return (0);
444}
445
446void
447space_map_close(space_map_t *sm)
448{
449	if (sm == NULL)
450		return;
451
452	if (sm->sm_dbuf != NULL)
453		dmu_buf_rele(sm->sm_dbuf, sm);
454	sm->sm_dbuf = NULL;
455	sm->sm_phys = NULL;
456
457	kmem_free(sm, sizeof (*sm));
458}
459
460static void
461space_map_reallocate(space_map_t *sm, dmu_tx_t *tx)
462{
463	ASSERT(dmu_tx_is_syncing(tx));
464
465	space_map_free(sm, tx);
466	dmu_buf_rele(sm->sm_dbuf, sm);
467
468	sm->sm_object = space_map_alloc(sm->sm_os, tx);
469	VERIFY0(space_map_open_impl(sm));
470}
471
472void
473space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
474{
475	objset_t *os = sm->sm_os;
476	spa_t *spa = dmu_objset_spa(os);
477	dmu_object_info_t doi;
478	int bonuslen;
479
480	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
481	ASSERT(dmu_tx_is_syncing(tx));
482
483	VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
484	dmu_object_info_from_db(sm->sm_dbuf, &doi);
485
486	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
487		bonuslen = sizeof (space_map_phys_t);
488		ASSERT3U(bonuslen, <=, dmu_bonus_max());
489	} else {
490		bonuslen = SPACE_MAP_SIZE_V0;
491	}
492
493	if (bonuslen != doi.doi_bonus_size ||
494	    doi.doi_data_block_size != SPACE_MAP_INITIAL_BLOCKSIZE) {
495		zfs_dbgmsg("txg %llu, spa %s, reallocating: "
496		    "old bonus %u, old blocksz %u", dmu_tx_get_txg(tx),
497		    spa_name(spa), doi.doi_bonus_size, doi.doi_data_block_size);
498		space_map_reallocate(sm, tx);
499		VERIFY3U(sm->sm_blksz, ==, SPACE_MAP_INITIAL_BLOCKSIZE);
500	}
501
502	dmu_buf_will_dirty(sm->sm_dbuf, tx);
503	sm->sm_phys->smp_objsize = 0;
504	sm->sm_phys->smp_alloc = 0;
505}
506
507/*
508 * Update the in-core space_map allocation and length values.
509 */
510void
511space_map_update(space_map_t *sm)
512{
513	if (sm == NULL)
514		return;
515
516	ASSERT(MUTEX_HELD(sm->sm_lock));
517
518	sm->sm_alloc = sm->sm_phys->smp_alloc;
519	sm->sm_length = sm->sm_phys->smp_objsize;
520}
521
522uint64_t
523space_map_alloc(objset_t *os, dmu_tx_t *tx)
524{
525	spa_t *spa = dmu_objset_spa(os);
526	uint64_t object;
527	int bonuslen;
528
529	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
530		spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
531		bonuslen = sizeof (space_map_phys_t);
532		ASSERT3U(bonuslen, <=, dmu_bonus_max());
533	} else {
534		bonuslen = SPACE_MAP_SIZE_V0;
535	}
536
537	object = dmu_object_alloc(os,
538	    DMU_OT_SPACE_MAP, SPACE_MAP_INITIAL_BLOCKSIZE,
539	    DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
540
541	return (object);
542}
543
544void
545space_map_free(space_map_t *sm, dmu_tx_t *tx)
546{
547	spa_t *spa;
548
549	if (sm == NULL)
550		return;
551
552	spa = dmu_objset_spa(sm->sm_os);
553	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
554		dmu_object_info_t doi;
555
556		dmu_object_info_from_db(sm->sm_dbuf, &doi);
557		if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
558			VERIFY(spa_feature_is_active(spa,
559			    SPA_FEATURE_SPACEMAP_HISTOGRAM));
560			spa_feature_decr(spa,
561			    SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
562		}
563	}
564
565	VERIFY3U(dmu_object_free(sm->sm_os, space_map_object(sm), tx), ==, 0);
566	sm->sm_object = 0;
567}
568
569uint64_t
570space_map_object(space_map_t *sm)
571{
572	return (sm != NULL ? sm->sm_object : 0);
573}
574
575/*
576 * Returns the already synced, on-disk allocated space.
577 */
578uint64_t
579space_map_allocated(space_map_t *sm)
580{
581	return (sm != NULL ? sm->sm_alloc : 0);
582}
583
584/*
585 * Returns the already synced, on-disk length;
586 */
587uint64_t
588space_map_length(space_map_t *sm)
589{
590	return (sm != NULL ? sm->sm_length : 0);
591}
592
593/*
594 * Returns the allocated space that is currently syncing.
595 */
596int64_t
597space_map_alloc_delta(space_map_t *sm)
598{
599	if (sm == NULL)
600		return (0);
601	ASSERT(sm->sm_dbuf != NULL);
602	return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
603}
604