xref: /illumos-gate/usr/src/uts/common/fs/zfs/space_map.c (revision b4203d757c7c247e39c94c09a94021a3a8121062)
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, 2016 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  * The data for a given space map can be kept on blocks of any size.
42  * Larger blocks entail fewer i/o operations, but they also cause the
43  * DMU to keep more data in-core, and also to waste more i/o bandwidth
44  * when only a few blocks have changed since the last transaction group.
45  */
46 int space_map_blksz = (1 << 12);
47 
48 /*
49  * Load the space map disk into the specified range tree. Segments of maptype
50  * are added to the range tree, other segment types are removed.
51  *
52  * Note: space_map_load() will drop sm_lock across dmu_read() calls.
53  * The caller must be OK with this.
54  */
55 int
56 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
57 {
58 	uint64_t *entry, *entry_map, *entry_map_end;
59 	uint64_t bufsize, size, offset, end, space;
60 	int error = 0;
61 
62 	ASSERT(MUTEX_HELD(sm->sm_lock));
63 
64 	end = space_map_length(sm);
65 	space = space_map_allocated(sm);
66 
67 	VERIFY0(range_tree_space(rt));
68 
69 	if (maptype == SM_FREE) {
70 		range_tree_add(rt, sm->sm_start, sm->sm_size);
71 		space = sm->sm_size - space;
72 	}
73 
74 	bufsize = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
75 	entry_map = zio_buf_alloc(bufsize);
76 
77 	mutex_exit(sm->sm_lock);
78 	if (end > bufsize) {
79 		dmu_prefetch(sm->sm_os, space_map_object(sm), 0, bufsize,
80 		    end - bufsize, ZIO_PRIORITY_SYNC_READ);
81 	}
82 	mutex_enter(sm->sm_lock);
83 
84 	for (offset = 0; offset < end; offset += bufsize) {
85 		size = MIN(end - offset, bufsize);
86 		VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
87 		VERIFY(size != 0);
88 		ASSERT3U(sm->sm_blksz, !=, 0);
89 
90 		dprintf("object=%llu  offset=%llx  size=%llx\n",
91 		    space_map_object(sm), offset, size);
92 
93 		mutex_exit(sm->sm_lock);
94 		error = dmu_read(sm->sm_os, space_map_object(sm), offset, size,
95 		    entry_map, DMU_READ_PREFETCH);
96 		mutex_enter(sm->sm_lock);
97 		if (error != 0)
98 			break;
99 
100 		entry_map_end = entry_map + (size / sizeof (uint64_t));
101 		for (entry = entry_map; entry < entry_map_end; entry++) {
102 			uint64_t e = *entry;
103 			uint64_t offset, size;
104 
105 			if (SM_DEBUG_DECODE(e))		/* Skip debug entries */
106 				continue;
107 
108 			offset = (SM_OFFSET_DECODE(e) << sm->sm_shift) +
109 			    sm->sm_start;
110 			size = SM_RUN_DECODE(e) << sm->sm_shift;
111 
112 			VERIFY0(P2PHASE(offset, 1ULL << sm->sm_shift));
113 			VERIFY0(P2PHASE(size, 1ULL << sm->sm_shift));
114 			VERIFY3U(offset, >=, sm->sm_start);
115 			VERIFY3U(offset + size, <=, sm->sm_start + sm->sm_size);
116 			if (SM_TYPE_DECODE(e) == maptype) {
117 				VERIFY3U(range_tree_space(rt) + size, <=,
118 				    sm->sm_size);
119 				range_tree_add(rt, offset, size);
120 			} else {
121 				range_tree_remove(rt, offset, size);
122 			}
123 		}
124 	}
125 
126 	if (error == 0)
127 		VERIFY3U(range_tree_space(rt), ==, space);
128 	else
129 		range_tree_vacate(rt, NULL, NULL);
130 
131 	zio_buf_free(entry_map, bufsize);
132 	return (error);
133 }
134 
135 void
136 space_map_histogram_clear(space_map_t *sm)
137 {
138 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
139 		return;
140 
141 	bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
142 }
143 
144 boolean_t
145 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
146 {
147 	/*
148 	 * Verify that the in-core range tree does not have any
149 	 * ranges smaller than our sm_shift size.
150 	 */
151 	for (int i = 0; i < sm->sm_shift; i++) {
152 		if (rt->rt_histogram[i] != 0)
153 			return (B_FALSE);
154 	}
155 	return (B_TRUE);
156 }
157 
158 void
159 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
160 {
161 	int idx = 0;
162 
163 	ASSERT(MUTEX_HELD(rt->rt_lock));
164 	ASSERT(dmu_tx_is_syncing(tx));
165 	VERIFY3U(space_map_object(sm), !=, 0);
166 
167 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
168 		return;
169 
170 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
171 
172 	ASSERT(space_map_histogram_verify(sm, rt));
173 	/*
174 	 * Transfer the content of the range tree histogram to the space
175 	 * map histogram. The space map histogram contains 32 buckets ranging
176 	 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
177 	 * however, can represent ranges from 2^0 to 2^63. Since the space
178 	 * map only cares about allocatable blocks (minimum of sm_shift) we
179 	 * can safely ignore all ranges in the range tree smaller than sm_shift.
180 	 */
181 	for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
182 
183 		/*
184 		 * Since the largest histogram bucket in the space map is
185 		 * 2^(32+sm_shift-1), we need to normalize the values in
186 		 * the range tree for any bucket larger than that size. For
187 		 * example given an sm_shift of 9, ranges larger than 2^40
188 		 * would get normalized as if they were 1TB ranges. Assume
189 		 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
190 		 * the calculation below would normalize this to 5 * 2^4 (16).
191 		 */
192 		ASSERT3U(i, >=, idx + sm->sm_shift);
193 		sm->sm_phys->smp_histogram[idx] +=
194 		    rt->rt_histogram[i] << (i - idx - sm->sm_shift);
195 
196 		/*
197 		 * Increment the space map's index as long as we haven't
198 		 * reached the maximum bucket size. Accumulate all ranges
199 		 * larger than the max bucket size into the last bucket.
200 		 */
201 		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
202 			ASSERT3U(idx + sm->sm_shift, ==, i);
203 			idx++;
204 			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
205 		}
206 	}
207 }
208 
209 uint64_t
210 space_map_entries(space_map_t *sm, range_tree_t *rt)
211 {
212 	avl_tree_t *t = &rt->rt_root;
213 	range_seg_t *rs;
214 	uint64_t size, entries;
215 
216 	/*
217 	 * All space_maps always have a debug entry so account for it here.
218 	 */
219 	entries = 1;
220 
221 	/*
222 	 * Traverse the range tree and calculate the number of space map
223 	 * entries that would be required to write out the range tree.
224 	 */
225 	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
226 		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
227 		entries += howmany(size, SM_RUN_MAX);
228 	}
229 	return (entries);
230 }
231 
232 /*
233  * Note: space_map_write() will drop sm_lock across dmu_write() calls.
234  */
235 void
236 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
237     dmu_tx_t *tx)
238 {
239 	objset_t *os = sm->sm_os;
240 	spa_t *spa = dmu_objset_spa(os);
241 	avl_tree_t *t = &rt->rt_root;
242 	range_seg_t *rs;
243 	uint64_t size, total, rt_space, nodes;
244 	uint64_t *entry, *entry_map, *entry_map_end;
245 	uint64_t expected_entries, actual_entries = 1;
246 
247 	ASSERT(MUTEX_HELD(rt->rt_lock));
248 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
249 	VERIFY3U(space_map_object(sm), !=, 0);
250 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
251 
252 	/*
253 	 * This field is no longer necessary since the in-core space map
254 	 * now contains the object number but is maintained for backwards
255 	 * compatibility.
256 	 */
257 	sm->sm_phys->smp_object = sm->sm_object;
258 
259 	if (range_tree_space(rt) == 0) {
260 		VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
261 		return;
262 	}
263 
264 	if (maptype == SM_ALLOC)
265 		sm->sm_phys->smp_alloc += range_tree_space(rt);
266 	else
267 		sm->sm_phys->smp_alloc -= range_tree_space(rt);
268 
269 	expected_entries = space_map_entries(sm, rt);
270 
271 	entry_map = zio_buf_alloc(sm->sm_blksz);
272 	entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
273 	entry = entry_map;
274 
275 	*entry++ = SM_DEBUG_ENCODE(1) |
276 	    SM_DEBUG_ACTION_ENCODE(maptype) |
277 	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
278 	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
279 
280 	total = 0;
281 	nodes = avl_numnodes(&rt->rt_root);
282 	rt_space = range_tree_space(rt);
283 	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
284 		uint64_t start;
285 
286 		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
287 		start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
288 
289 		total += size << sm->sm_shift;
290 
291 		while (size != 0) {
292 			uint64_t run_len;
293 
294 			run_len = MIN(size, SM_RUN_MAX);
295 
296 			if (entry == entry_map_end) {
297 				mutex_exit(rt->rt_lock);
298 				dmu_write(os, space_map_object(sm),
299 				    sm->sm_phys->smp_objsize, sm->sm_blksz,
300 				    entry_map, tx);
301 				mutex_enter(rt->rt_lock);
302 				sm->sm_phys->smp_objsize += sm->sm_blksz;
303 				entry = entry_map;
304 			}
305 
306 			*entry++ = SM_OFFSET_ENCODE(start) |
307 			    SM_TYPE_ENCODE(maptype) |
308 			    SM_RUN_ENCODE(run_len);
309 
310 			start += run_len;
311 			size -= run_len;
312 			actual_entries++;
313 		}
314 	}
315 
316 	if (entry != entry_map) {
317 		size = (entry - entry_map) * sizeof (uint64_t);
318 		mutex_exit(rt->rt_lock);
319 		dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
320 		    size, entry_map, tx);
321 		mutex_enter(rt->rt_lock);
322 		sm->sm_phys->smp_objsize += size;
323 	}
324 	ASSERT3U(expected_entries, ==, actual_entries);
325 
326 	/*
327 	 * Ensure that the space_map's accounting wasn't changed
328 	 * while we were in the middle of writing it out.
329 	 */
330 	VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
331 	VERIFY3U(range_tree_space(rt), ==, rt_space);
332 	VERIFY3U(range_tree_space(rt), ==, total);
333 
334 	zio_buf_free(entry_map, sm->sm_blksz);
335 }
336 
337 static int
338 space_map_open_impl(space_map_t *sm)
339 {
340 	int error;
341 	u_longlong_t blocks;
342 
343 	error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
344 	if (error)
345 		return (error);
346 
347 	dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
348 	sm->sm_phys = sm->sm_dbuf->db_data;
349 	return (0);
350 }
351 
352 int
353 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
354     uint64_t start, uint64_t size, uint8_t shift, kmutex_t *lp)
355 {
356 	space_map_t *sm;
357 	int error;
358 
359 	ASSERT(*smp == NULL);
360 	ASSERT(os != NULL);
361 	ASSERT(object != 0);
362 
363 	sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
364 
365 	sm->sm_start = start;
366 	sm->sm_size = size;
367 	sm->sm_shift = shift;
368 	sm->sm_lock = lp;
369 	sm->sm_os = os;
370 	sm->sm_object = object;
371 
372 	error = space_map_open_impl(sm);
373 	if (error != 0) {
374 		space_map_close(sm);
375 		return (error);
376 	}
377 
378 	*smp = sm;
379 
380 	return (0);
381 }
382 
383 void
384 space_map_close(space_map_t *sm)
385 {
386 	if (sm == NULL)
387 		return;
388 
389 	if (sm->sm_dbuf != NULL)
390 		dmu_buf_rele(sm->sm_dbuf, sm);
391 	sm->sm_dbuf = NULL;
392 	sm->sm_phys = NULL;
393 
394 	kmem_free(sm, sizeof (*sm));
395 }
396 
397 void
398 space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
399 {
400 	objset_t *os = sm->sm_os;
401 	spa_t *spa = dmu_objset_spa(os);
402 	dmu_object_info_t doi;
403 
404 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
405 	ASSERT(dmu_tx_is_syncing(tx));
406 	VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
407 
408 	dmu_object_info_from_db(sm->sm_dbuf, &doi);
409 
410 	/*
411 	 * If the space map has the wrong bonus size (because
412 	 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
413 	 * the wrong block size (because space_map_blksz has changed),
414 	 * free and re-allocate its object with the updated sizes.
415 	 *
416 	 * Otherwise, just truncate the current object.
417 	 */
418 	if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
419 	    doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
420 	    doi.doi_data_block_size != space_map_blksz) {
421 		zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
422 		    "object[%llu]: old bonus %u, old blocksz %u",
423 		    dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
424 		    doi.doi_bonus_size, doi.doi_data_block_size);
425 
426 		space_map_free(sm, tx);
427 		dmu_buf_rele(sm->sm_dbuf, sm);
428 
429 		sm->sm_object = space_map_alloc(sm->sm_os, tx);
430 		VERIFY0(space_map_open_impl(sm));
431 	} else {
432 		VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
433 
434 		/*
435 		 * If the spacemap is reallocated, its histogram
436 		 * will be reset.  Do the same in the common case so that
437 		 * bugs related to the uncommon case do not go unnoticed.
438 		 */
439 		bzero(sm->sm_phys->smp_histogram,
440 		    sizeof (sm->sm_phys->smp_histogram));
441 	}
442 
443 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
444 	sm->sm_phys->smp_objsize = 0;
445 	sm->sm_phys->smp_alloc = 0;
446 }
447 
448 /*
449  * Update the in-core space_map allocation and length values.
450  */
451 void
452 space_map_update(space_map_t *sm)
453 {
454 	if (sm == NULL)
455 		return;
456 
457 	ASSERT(MUTEX_HELD(sm->sm_lock));
458 
459 	sm->sm_alloc = sm->sm_phys->smp_alloc;
460 	sm->sm_length = sm->sm_phys->smp_objsize;
461 }
462 
463 uint64_t
464 space_map_alloc(objset_t *os, dmu_tx_t *tx)
465 {
466 	spa_t *spa = dmu_objset_spa(os);
467 	uint64_t object;
468 	int bonuslen;
469 
470 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
471 		spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
472 		bonuslen = sizeof (space_map_phys_t);
473 		ASSERT3U(bonuslen, <=, dmu_bonus_max());
474 	} else {
475 		bonuslen = SPACE_MAP_SIZE_V0;
476 	}
477 
478 	object = dmu_object_alloc(os,
479 	    DMU_OT_SPACE_MAP, space_map_blksz,
480 	    DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
481 
482 	return (object);
483 }
484 
485 void
486 space_map_free(space_map_t *sm, dmu_tx_t *tx)
487 {
488 	spa_t *spa;
489 
490 	if (sm == NULL)
491 		return;
492 
493 	spa = dmu_objset_spa(sm->sm_os);
494 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
495 		dmu_object_info_t doi;
496 
497 		dmu_object_info_from_db(sm->sm_dbuf, &doi);
498 		if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
499 			VERIFY(spa_feature_is_active(spa,
500 			    SPA_FEATURE_SPACEMAP_HISTOGRAM));
501 			spa_feature_decr(spa,
502 			    SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
503 		}
504 	}
505 
506 	VERIFY3U(dmu_object_free(sm->sm_os, space_map_object(sm), tx), ==, 0);
507 	sm->sm_object = 0;
508 }
509 
510 uint64_t
511 space_map_object(space_map_t *sm)
512 {
513 	return (sm != NULL ? sm->sm_object : 0);
514 }
515 
516 /*
517  * Returns the already synced, on-disk allocated space.
518  */
519 uint64_t
520 space_map_allocated(space_map_t *sm)
521 {
522 	return (sm != NULL ? sm->sm_alloc : 0);
523 }
524 
525 /*
526  * Returns the already synced, on-disk length;
527  */
528 uint64_t
529 space_map_length(space_map_t *sm)
530 {
531 	return (sm != NULL ? sm->sm_length : 0);
532 }
533 
534 /*
535  * Returns the allocated space that is currently syncing.
536  */
537 int64_t
538 space_map_alloc_delta(space_map_t *sm)
539 {
540 	if (sm == NULL)
541 		return (0);
542 	ASSERT(sm->sm_dbuf != NULL);
543 	return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
544 }
545