xref: /illumos-gate/usr/src/uts/common/fs/zfs/space_map.c (revision 67a4bb8f9ad4c49e9aa9e21e2114a7c093c3a73a)
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  * 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), bufsize,
80 		    end - bufsize);
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 	/*
175 	 * Transfer the content of the range tree histogram to the space
176 	 * map histogram. The space map histogram contains 32 buckets ranging
177 	 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
178 	 * however, can represent ranges from 2^0 to 2^63. Since the space
179 	 * map only cares about allocatable blocks (minimum of sm_shift) we
180 	 * can safely ignore all ranges in the range tree smaller than sm_shift.
181 	 */
182 	for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
183 
184 		/*
185 		 * Since the largest histogram bucket in the space map is
186 		 * 2^(32+sm_shift-1), we need to normalize the values in
187 		 * the range tree for any bucket larger than that size. For
188 		 * example given an sm_shift of 9, ranges larger than 2^40
189 		 * would get normalized as if they were 1TB ranges. Assume
190 		 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
191 		 * the calculation below would normalize this to 5 * 2^4 (16).
192 		 */
193 		ASSERT3U(i, >=, idx + sm->sm_shift);
194 		sm->sm_phys->smp_histogram[idx] +=
195 		    rt->rt_histogram[i] << (i - idx - sm->sm_shift);
196 
197 		/*
198 		 * Increment the space map's index as long as we haven't
199 		 * reached the maximum bucket size. Accumulate all ranges
200 		 * larger than the max bucket size into the last bucket.
201 		 */
202 		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
203 			ASSERT3U(idx + sm->sm_shift, ==, i);
204 			idx++;
205 			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
206 		}
207 	}
208 }
209 
210 uint64_t
211 space_map_entries(space_map_t *sm, range_tree_t *rt)
212 {
213 	avl_tree_t *t = &rt->rt_root;
214 	range_seg_t *rs;
215 	uint64_t size, entries;
216 
217 	/*
218 	 * All space_maps always have a debug entry so account for it here.
219 	 */
220 	entries = 1;
221 
222 	/*
223 	 * Traverse the range tree and calculate the number of space map
224 	 * entries that would be required to write out the range tree.
225 	 */
226 	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
227 		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
228 		entries += howmany(size, SM_RUN_MAX);
229 	}
230 	return (entries);
231 }
232 
233 /*
234  * Note: space_map_write() will drop sm_lock across dmu_write() calls.
235  */
236 void
237 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
238     dmu_tx_t *tx)
239 {
240 	objset_t *os = sm->sm_os;
241 	spa_t *spa = dmu_objset_spa(os);
242 	avl_tree_t *t = &rt->rt_root;
243 	range_seg_t *rs;
244 	uint64_t size, total, rt_space, nodes;
245 	uint64_t *entry, *entry_map, *entry_map_end;
246 	uint64_t expected_entries, actual_entries = 1;
247 
248 	ASSERT(MUTEX_HELD(rt->rt_lock));
249 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
250 	VERIFY3U(space_map_object(sm), !=, 0);
251 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
252 
253 	/*
254 	 * This field is no longer necessary since the in-core space map
255 	 * now contains the object number but is maintained for backwards
256 	 * compatibility.
257 	 */
258 	sm->sm_phys->smp_object = sm->sm_object;
259 
260 	if (range_tree_space(rt) == 0) {
261 		VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
262 		return;
263 	}
264 
265 	if (maptype == SM_ALLOC)
266 		sm->sm_phys->smp_alloc += range_tree_space(rt);
267 	else
268 		sm->sm_phys->smp_alloc -= range_tree_space(rt);
269 
270 	expected_entries = space_map_entries(sm, rt);
271 
272 	entry_map = zio_buf_alloc(sm->sm_blksz);
273 	entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
274 	entry = entry_map;
275 
276 	*entry++ = SM_DEBUG_ENCODE(1) |
277 	    SM_DEBUG_ACTION_ENCODE(maptype) |
278 	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
279 	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
280 
281 	total = 0;
282 	nodes = avl_numnodes(&rt->rt_root);
283 	rt_space = range_tree_space(rt);
284 	for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
285 		uint64_t start;
286 
287 		size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
288 		start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
289 
290 		total += size << sm->sm_shift;
291 
292 		while (size != 0) {
293 			uint64_t run_len;
294 
295 			run_len = MIN(size, SM_RUN_MAX);
296 
297 			if (entry == entry_map_end) {
298 				mutex_exit(rt->rt_lock);
299 				dmu_write(os, space_map_object(sm),
300 				    sm->sm_phys->smp_objsize, sm->sm_blksz,
301 				    entry_map, tx);
302 				mutex_enter(rt->rt_lock);
303 				sm->sm_phys->smp_objsize += sm->sm_blksz;
304 				entry = entry_map;
305 			}
306 
307 			*entry++ = SM_OFFSET_ENCODE(start) |
308 			    SM_TYPE_ENCODE(maptype) |
309 			    SM_RUN_ENCODE(run_len);
310 
311 			start += run_len;
312 			size -= run_len;
313 			actual_entries++;
314 		}
315 	}
316 
317 	if (entry != entry_map) {
318 		size = (entry - entry_map) * sizeof (uint64_t);
319 		mutex_exit(rt->rt_lock);
320 		dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
321 		    size, entry_map, tx);
322 		mutex_enter(rt->rt_lock);
323 		sm->sm_phys->smp_objsize += size;
324 	}
325 	ASSERT3U(expected_entries, ==, actual_entries);
326 
327 	/*
328 	 * Ensure that the space_map's accounting wasn't changed
329 	 * while we were in the middle of writing it out.
330 	 */
331 	VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
332 	VERIFY3U(range_tree_space(rt), ==, rt_space);
333 	VERIFY3U(range_tree_space(rt), ==, total);
334 
335 	zio_buf_free(entry_map, sm->sm_blksz);
336 }
337 
338 static int
339 space_map_open_impl(space_map_t *sm)
340 {
341 	int error;
342 	u_longlong_t blocks;
343 
344 	error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
345 	if (error)
346 		return (error);
347 
348 	dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
349 	sm->sm_phys = sm->sm_dbuf->db_data;
350 	return (0);
351 }
352 
353 int
354 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
355     uint64_t start, uint64_t size, uint8_t shift, kmutex_t *lp)
356 {
357 	space_map_t *sm;
358 	int error;
359 
360 	ASSERT(*smp == NULL);
361 	ASSERT(os != NULL);
362 	ASSERT(object != 0);
363 
364 	sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
365 
366 	sm->sm_start = start;
367 	sm->sm_size = size;
368 	sm->sm_shift = shift;
369 	sm->sm_lock = lp;
370 	sm->sm_os = os;
371 	sm->sm_object = object;
372 
373 	error = space_map_open_impl(sm);
374 	if (error != 0) {
375 		space_map_close(sm);
376 		return (error);
377 	}
378 
379 	*smp = sm;
380 
381 	return (0);
382 }
383 
384 void
385 space_map_close(space_map_t *sm)
386 {
387 	if (sm == NULL)
388 		return;
389 
390 	if (sm->sm_dbuf != NULL)
391 		dmu_buf_rele(sm->sm_dbuf, sm);
392 	sm->sm_dbuf = NULL;
393 	sm->sm_phys = NULL;
394 
395 	kmem_free(sm, sizeof (*sm));
396 }
397 
398 void
399 space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
400 {
401 	objset_t *os = sm->sm_os;
402 	spa_t *spa = dmu_objset_spa(os);
403 	dmu_object_info_t doi;
404 
405 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
406 	ASSERT(dmu_tx_is_syncing(tx));
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, reallocating: "
422 		    "old bonus %u, old blocksz %u", dmu_tx_get_txg(tx),
423 		    spa_name(spa), doi.doi_bonus_size, doi.doi_data_block_size);
424 
425 		space_map_free(sm, tx);
426 		dmu_buf_rele(sm->sm_dbuf, sm);
427 
428 		sm->sm_object = space_map_alloc(sm->sm_os, tx);
429 		VERIFY0(space_map_open_impl(sm));
430 	} else {
431 		VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
432 
433 		/*
434 		 * If the spacemap is reallocated, its histogram
435 		 * will be reset.  Do the same in the common case so that
436 		 * bugs related to the uncommon case do not go unnoticed.
437 		 */
438 		bzero(sm->sm_phys->smp_histogram,
439 		    sizeof (sm->sm_phys->smp_histogram));
440 	}
441 
442 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
443 	sm->sm_phys->smp_objsize = 0;
444 	sm->sm_phys->smp_alloc = 0;
445 }
446 
447 /*
448  * Update the in-core space_map allocation and length values.
449  */
450 void
451 space_map_update(space_map_t *sm)
452 {
453 	if (sm == NULL)
454 		return;
455 
456 	ASSERT(MUTEX_HELD(sm->sm_lock));
457 
458 	sm->sm_alloc = sm->sm_phys->smp_alloc;
459 	sm->sm_length = sm->sm_phys->smp_objsize;
460 }
461 
462 uint64_t
463 space_map_alloc(objset_t *os, dmu_tx_t *tx)
464 {
465 	spa_t *spa = dmu_objset_spa(os);
466 	uint64_t object;
467 	int bonuslen;
468 
469 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
470 		spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
471 		bonuslen = sizeof (space_map_phys_t);
472 		ASSERT3U(bonuslen, <=, dmu_bonus_max());
473 	} else {
474 		bonuslen = SPACE_MAP_SIZE_V0;
475 	}
476 
477 	object = dmu_object_alloc(os,
478 	    DMU_OT_SPACE_MAP, space_map_blksz,
479 	    DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
480 
481 	return (object);
482 }
483 
484 void
485 space_map_free(space_map_t *sm, dmu_tx_t *tx)
486 {
487 	spa_t *spa;
488 
489 	if (sm == NULL)
490 		return;
491 
492 	spa = dmu_objset_spa(sm->sm_os);
493 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
494 		dmu_object_info_t doi;
495 
496 		dmu_object_info_from_db(sm->sm_dbuf, &doi);
497 		if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
498 			VERIFY(spa_feature_is_active(spa,
499 			    SPA_FEATURE_SPACEMAP_HISTOGRAM));
500 			spa_feature_decr(spa,
501 			    SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
502 		}
503 	}
504 
505 	VERIFY3U(dmu_object_free(sm->sm_os, space_map_object(sm), tx), ==, 0);
506 	sm->sm_object = 0;
507 }
508 
509 uint64_t
510 space_map_object(space_map_t *sm)
511 {
512 	return (sm != NULL ? sm->sm_object : 0);
513 }
514 
515 /*
516  * Returns the already synced, on-disk allocated space.
517  */
518 uint64_t
519 space_map_allocated(space_map_t *sm)
520 {
521 	return (sm != NULL ? sm->sm_alloc : 0);
522 }
523 
524 /*
525  * Returns the already synced, on-disk length;
526  */
527 uint64_t
528 space_map_length(space_map_t *sm)
529 {
530 	return (sm != NULL ? sm->sm_length : 0);
531 }
532 
533 /*
534  * Returns the allocated space that is currently syncing.
535  */
536 int64_t
537 space_map_alloc_delta(space_map_t *sm)
538 {
539 	if (sm == NULL)
540 		return (0);
541 	ASSERT(sm->sm_dbuf != NULL);
542 	return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
543 }
544