]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/space_map.c
Illumos 4171, 4172
[mirror_zfs.git] / module / zfs / space_map.c
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) 2013 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 */
49 int 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 */
58 int
59 space_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
138 void
139 space_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
147 boolean_t
148 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
149 {
150 int i;
151
152 /*
153 * Verify that the in-core range tree does not have any
154 * ranges smaller than our sm_shift size.
155 */
156 for (i = 0; i < sm->sm_shift; i++) {
157 if (rt->rt_histogram[i] != 0)
158 return (B_FALSE);
159 }
160 return (B_TRUE);
161 }
162
163 void
164 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
165 {
166 int idx = 0;
167 int i;
168
169 ASSERT(MUTEX_HELD(rt->rt_lock));
170 ASSERT(dmu_tx_is_syncing(tx));
171 VERIFY3U(space_map_object(sm), !=, 0);
172
173 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
174 return;
175
176 dmu_buf_will_dirty(sm->sm_dbuf, tx);
177
178 ASSERT(space_map_histogram_verify(sm, rt));
179
180 /*
181 * Transfer the content of the range tree histogram to the space
182 * map histogram. The space map histogram contains 32 buckets ranging
183 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
184 * however, can represent ranges from 2^0 to 2^63. Since the space
185 * map only cares about allocatable blocks (minimum of sm_shift) we
186 * can safely ignore all ranges in the range tree smaller than sm_shift.
187 */
188 for (i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
189
190 /*
191 * Since the largest histogram bucket in the space map is
192 * 2^(32+sm_shift-1), we need to normalize the values in
193 * the range tree for any bucket larger than that size. For
194 * example given an sm_shift of 9, ranges larger than 2^40
195 * would get normalized as if they were 1TB ranges. Assume
196 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
197 * the calculation below would normalize this to 5 * 2^4 (16).
198 */
199 ASSERT3U(i, >=, idx + sm->sm_shift);
200 sm->sm_phys->smp_histogram[idx] +=
201 rt->rt_histogram[i] << (i - idx - sm->sm_shift);
202
203 /*
204 * Increment the space map's index as long as we haven't
205 * reached the maximum bucket size. Accumulate all ranges
206 * larger than the max bucket size into the last bucket.
207 */
208 if (idx < SPACE_MAP_HISTOGRAM_SIZE(sm) - 1) {
209 ASSERT3U(idx + sm->sm_shift, ==, i);
210 idx++;
211 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE(sm));
212 }
213 }
214 }
215
216 uint64_t
217 space_map_entries(space_map_t *sm, range_tree_t *rt)
218 {
219 avl_tree_t *t = &rt->rt_root;
220 range_seg_t *rs;
221 uint64_t size, entries;
222
223 /*
224 * All space_maps always have a debug entry so account for it here.
225 */
226 entries = 1;
227
228 /*
229 * Traverse the range tree and calculate the number of space map
230 * entries that would be required to write out the range tree.
231 */
232 for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
233 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
234 entries += howmany(size, SM_RUN_MAX);
235 }
236 return (entries);
237 }
238
239 void
240 space_map_set_blocksize(space_map_t *sm, uint64_t size, dmu_tx_t *tx)
241 {
242 uint32_t blksz;
243 u_longlong_t blocks;
244
245 ASSERT3U(sm->sm_blksz, !=, 0);
246 ASSERT3U(space_map_object(sm), !=, 0);
247 ASSERT(sm->sm_dbuf != NULL);
248 VERIFY(ISP2(space_map_max_blksz));
249
250 if (sm->sm_blksz >= space_map_max_blksz)
251 return;
252
253 /*
254 * The object contains more than one block so we can't adjust
255 * its size.
256 */
257 if (sm->sm_phys->smp_objsize > sm->sm_blksz)
258 return;
259
260 if (size > sm->sm_blksz) {
261 uint64_t newsz;
262
263 /*
264 * Older software versions treat space map blocks as fixed
265 * entities. The DMU is capable of handling different block
266 * sizes making it possible for us to increase the
267 * block size and maintain backwards compatibility. The
268 * caveat is that the new block sizes must be a
269 * power of 2 so that old software can append to the file,
270 * adding more blocks. The block size can grow until it
271 * reaches space_map_max_blksz.
272 */
273 newsz = ISP2(size) ? size : 1ULL << highbit(size);
274 if (newsz > space_map_max_blksz)
275 newsz = space_map_max_blksz;
276
277 VERIFY0(dmu_object_set_blocksize(sm->sm_os,
278 space_map_object(sm), newsz, 0, tx));
279 dmu_object_size_from_db(sm->sm_dbuf, &blksz, &blocks);
280
281 zfs_dbgmsg("txg %llu, spa %s, increasing blksz from %d to %d",
282 dmu_tx_get_txg(tx), spa_name(dmu_objset_spa(sm->sm_os)),
283 sm->sm_blksz, blksz);
284
285 VERIFY3U(newsz, ==, blksz);
286 VERIFY3U(sm->sm_blksz, <, blksz);
287 sm->sm_blksz = blksz;
288 }
289 }
290
291 /*
292 * Note: space_map_write() will drop sm_lock across dmu_write() calls.
293 */
294 void
295 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
296 dmu_tx_t *tx)
297 {
298 objset_t *os = sm->sm_os;
299 spa_t *spa = dmu_objset_spa(os);
300 avl_tree_t *t = &rt->rt_root;
301 range_seg_t *rs;
302 uint64_t size, total, rt_space, nodes;
303 uint64_t *entry, *entry_map, *entry_map_end;
304 uint64_t newsz, expected_entries, actual_entries = 1;
305
306 ASSERT(MUTEX_HELD(rt->rt_lock));
307 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
308 VERIFY3U(space_map_object(sm), !=, 0);
309 dmu_buf_will_dirty(sm->sm_dbuf, tx);
310
311 /*
312 * This field is no longer necessary since the in-core space map
313 * now contains the object number but is maintained for backwards
314 * compatibility.
315 */
316 sm->sm_phys->smp_object = sm->sm_object;
317
318 if (range_tree_space(rt) == 0) {
319 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
320 return;
321 }
322
323 if (maptype == SM_ALLOC)
324 sm->sm_phys->smp_alloc += range_tree_space(rt);
325 else
326 sm->sm_phys->smp_alloc -= range_tree_space(rt);
327
328 expected_entries = space_map_entries(sm, rt);
329
330 /*
331 * Calculate the new size for the space map on-disk and see if
332 * we can grow the block size to accommodate the new size.
333 */
334 newsz = sm->sm_phys->smp_objsize + expected_entries * sizeof (uint64_t);
335 space_map_set_blocksize(sm, newsz, tx);
336
337 entry_map = zio_buf_alloc(sm->sm_blksz);
338 entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
339 entry = entry_map;
340
341 *entry++ = SM_DEBUG_ENCODE(1) |
342 SM_DEBUG_ACTION_ENCODE(maptype) |
343 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
344 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
345
346 total = 0;
347 nodes = avl_numnodes(&rt->rt_root);
348 rt_space = range_tree_space(rt);
349 for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
350 uint64_t start;
351
352 size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
353 start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
354
355 total += size << sm->sm_shift;
356
357 while (size != 0) {
358 uint64_t run_len;
359
360 run_len = MIN(size, SM_RUN_MAX);
361
362 if (entry == entry_map_end) {
363 mutex_exit(rt->rt_lock);
364 dmu_write(os, space_map_object(sm),
365 sm->sm_phys->smp_objsize, sm->sm_blksz,
366 entry_map, tx);
367 mutex_enter(rt->rt_lock);
368 sm->sm_phys->smp_objsize += sm->sm_blksz;
369 entry = entry_map;
370 }
371
372 *entry++ = SM_OFFSET_ENCODE(start) |
373 SM_TYPE_ENCODE(maptype) |
374 SM_RUN_ENCODE(run_len);
375
376 start += run_len;
377 size -= run_len;
378 actual_entries++;
379 }
380 }
381
382 if (entry != entry_map) {
383 size = (entry - entry_map) * sizeof (uint64_t);
384 mutex_exit(rt->rt_lock);
385 dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
386 size, entry_map, tx);
387 mutex_enter(rt->rt_lock);
388 sm->sm_phys->smp_objsize += size;
389 }
390 ASSERT3U(expected_entries, ==, actual_entries);
391
392 /*
393 * Ensure that the space_map's accounting wasn't changed
394 * while we were in the middle of writing it out.
395 */
396 VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
397 VERIFY3U(range_tree_space(rt), ==, rt_space);
398 VERIFY3U(range_tree_space(rt), ==, total);
399
400 zio_buf_free(entry_map, sm->sm_blksz);
401 }
402
403 static int
404 space_map_open_impl(space_map_t *sm)
405 {
406 int error;
407 u_longlong_t blocks;
408
409 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
410 if (error)
411 return (error);
412
413 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
414 sm->sm_phys = sm->sm_dbuf->db_data;
415 return (0);
416 }
417
418 int
419 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
420 uint64_t start, uint64_t size, uint8_t shift, kmutex_t *lp)
421 {
422 space_map_t *sm;
423 int error;
424
425 ASSERT(*smp == NULL);
426 ASSERT(os != NULL);
427 ASSERT(object != 0);
428
429 sm = kmem_alloc(sizeof (space_map_t), KM_PUSHPAGE);
430
431 sm->sm_start = start;
432 sm->sm_size = size;
433 sm->sm_shift = shift;
434 sm->sm_lock = lp;
435 sm->sm_os = os;
436 sm->sm_object = object;
437 sm->sm_length = 0;
438 sm->sm_alloc = 0;
439 sm->sm_blksz = 0;
440 sm->sm_dbuf = NULL;
441 sm->sm_phys = NULL;
442
443 error = space_map_open_impl(sm);
444 if (error != 0) {
445 space_map_close(sm);
446 return (error);
447 }
448
449 *smp = sm;
450
451 return (0);
452 }
453
454 void
455 space_map_close(space_map_t *sm)
456 {
457 if (sm == NULL)
458 return;
459
460 if (sm->sm_dbuf != NULL)
461 dmu_buf_rele(sm->sm_dbuf, sm);
462 sm->sm_dbuf = NULL;
463 sm->sm_phys = NULL;
464
465 kmem_free(sm, sizeof (*sm));
466 }
467
468 static void
469 space_map_reallocate(space_map_t *sm, dmu_tx_t *tx)
470 {
471 ASSERT(dmu_tx_is_syncing(tx));
472
473 space_map_free(sm, tx);
474 dmu_buf_rele(sm->sm_dbuf, sm);
475
476 sm->sm_object = space_map_alloc(sm->sm_os, tx);
477 VERIFY0(space_map_open_impl(sm));
478 }
479
480 void
481 space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
482 {
483 objset_t *os = sm->sm_os;
484 spa_t *spa = dmu_objset_spa(os);
485 dmu_object_info_t doi;
486 int bonuslen;
487
488 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
489 ASSERT(dmu_tx_is_syncing(tx));
490
491 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
492 dmu_object_info_from_db(sm->sm_dbuf, &doi);
493
494 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
495 bonuslen = sizeof (space_map_phys_t);
496 ASSERT3U(bonuslen, <=, dmu_bonus_max());
497 } else {
498 bonuslen = SPACE_MAP_SIZE_V0;
499 }
500
501 if (bonuslen != doi.doi_bonus_size ||
502 doi.doi_data_block_size != SPACE_MAP_INITIAL_BLOCKSIZE) {
503 zfs_dbgmsg("txg %llu, spa %s, reallocating: "
504 "old bonus %u, old blocksz %u", dmu_tx_get_txg(tx),
505 spa_name(spa), doi.doi_bonus_size, doi.doi_data_block_size);
506 space_map_reallocate(sm, tx);
507 VERIFY3U(sm->sm_blksz, ==, SPACE_MAP_INITIAL_BLOCKSIZE);
508 }
509
510 dmu_buf_will_dirty(sm->sm_dbuf, tx);
511 sm->sm_phys->smp_objsize = 0;
512 sm->sm_phys->smp_alloc = 0;
513 }
514
515 /*
516 * Update the in-core space_map allocation and length values.
517 */
518 void
519 space_map_update(space_map_t *sm)
520 {
521 if (sm == NULL)
522 return;
523
524 ASSERT(MUTEX_HELD(sm->sm_lock));
525
526 sm->sm_alloc = sm->sm_phys->smp_alloc;
527 sm->sm_length = sm->sm_phys->smp_objsize;
528 }
529
530 uint64_t
531 space_map_alloc(objset_t *os, dmu_tx_t *tx)
532 {
533 spa_t *spa = dmu_objset_spa(os);
534 uint64_t object;
535 int bonuslen;
536
537 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
538 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
539 bonuslen = sizeof (space_map_phys_t);
540 ASSERT3U(bonuslen, <=, dmu_bonus_max());
541 } else {
542 bonuslen = SPACE_MAP_SIZE_V0;
543 }
544
545 object = dmu_object_alloc(os,
546 DMU_OT_SPACE_MAP, SPACE_MAP_INITIAL_BLOCKSIZE,
547 DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
548
549 return (object);
550 }
551
552 void
553 space_map_free(space_map_t *sm, dmu_tx_t *tx)
554 {
555 spa_t *spa;
556
557 if (sm == NULL)
558 return;
559
560 spa = dmu_objset_spa(sm->sm_os);
561 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
562 dmu_object_info_t doi;
563
564 dmu_object_info_from_db(sm->sm_dbuf, &doi);
565 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
566 VERIFY(spa_feature_is_active(spa,
567 SPA_FEATURE_SPACEMAP_HISTOGRAM));
568 spa_feature_decr(spa,
569 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
570 }
571 }
572
573 VERIFY3U(dmu_object_free(sm->sm_os, space_map_object(sm), tx), ==, 0);
574 sm->sm_object = 0;
575 }
576
577 uint64_t
578 space_map_object(space_map_t *sm)
579 {
580 return (sm != NULL ? sm->sm_object : 0);
581 }
582
583 /*
584 * Returns the already synced, on-disk allocated space.
585 */
586 uint64_t
587 space_map_allocated(space_map_t *sm)
588 {
589 return (sm != NULL ? sm->sm_alloc : 0);
590 }
591
592 /*
593 * Returns the already synced, on-disk length;
594 */
595 uint64_t
596 space_map_length(space_map_t *sm)
597 {
598 return (sm != NULL ? sm->sm_length : 0);
599 }
600
601 /*
602 * Returns the allocated space that is currently syncing.
603 */
604 int64_t
605 space_map_alloc_delta(space_map_t *sm)
606 {
607 if (sm == NULL)
608 return (0);
609 ASSERT(sm->sm_dbuf != NULL);
610 return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
611 }