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, 2017 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 * Note on space map block size:
42 *
43 * The data for a given space map can be kept on blocks of any size.
44 * Larger blocks entail fewer I/O operations, but they also cause the
45 * DMU to keep more data in-core, and also to waste more I/O bandwidth
46 * when only a few blocks have changed since the last transaction group.
47 */
48
49 /*
50 * Enabled whenever we want to stress test the use of double-word
51 * space map entries.
52 */
53 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
54
55 /*
56 * Override the default indirect block size of 128K, instead using 16K for
57 * spacemaps (2^14 bytes). This dramatically reduces write inflation since
58 * appending to a spacemap typically has to write one data block (4KB) and one
59 * or two indirect blocks (16K-32K, rather than 128K).
60 */
61 int space_map_ibs = 14;
62
63 boolean_t
64 sm_entry_is_debug(uint64_t e)
65 {
66 return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX);
67 }
68
69 boolean_t
70 sm_entry_is_single_word(uint64_t e)
71 {
72 uint8_t prefix = SM_PREFIX_DECODE(e);
73 return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX);
74 }
75
76 boolean_t
77 sm_entry_is_double_word(uint64_t e)
78 {
79 return (SM_PREFIX_DECODE(e) == SM2_PREFIX);
80 }
81
82 /*
83 * Iterate through the space map, invoking the callback on each (non-debug)
84 * space map entry.
85 */
86 int
87 space_map_iterate(space_map_t *sm, sm_cb_t callback, void *arg)
88 {
89 uint64_t sm_len = space_map_length(sm);
90 ASSERT3U(sm->sm_blksz, !=, 0);
91
92 dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, sm_len,
93 ZIO_PRIORITY_SYNC_READ);
94
95 uint64_t blksz = sm->sm_blksz;
96 int error = 0;
97 for (uint64_t block_base = 0; block_base < sm_len && error == 0;
98 block_base += blksz) {
99 dmu_buf_t *db;
100 error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
101 block_base, FTAG, &db, DMU_READ_PREFETCH);
102 if (error != 0)
103 return (error);
104
105 uint64_t *block_start = db->db_data;
106 uint64_t block_length = MIN(sm_len - block_base, blksz);
107 uint64_t *block_end = block_start +
108 (block_length / sizeof (uint64_t));
109
110 VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
111 VERIFY3U(block_length, !=, 0);
112 ASSERT3U(blksz, ==, db->db_size);
113
114 for (uint64_t *block_cursor = block_start;
115 block_cursor < block_end && error == 0; block_cursor++) {
116 uint64_t e = *block_cursor;
117
118 if (sm_entry_is_debug(e)) /* Skip debug entries */
119 continue;
120
121 uint64_t raw_offset, raw_run, vdev_id;
122 maptype_t type;
123 if (sm_entry_is_single_word(e)) {
124 type = SM_TYPE_DECODE(e);
125 vdev_id = SM_NO_VDEVID;
126 raw_offset = SM_OFFSET_DECODE(e);
127 raw_run = SM_RUN_DECODE(e);
128 } else {
129 /* it is a two-word entry */
130 ASSERT(sm_entry_is_double_word(e));
131 raw_run = SM2_RUN_DECODE(e);
132 vdev_id = SM2_VDEV_DECODE(e);
133
134 /* move on to the second word */
135 block_cursor++;
136 e = *block_cursor;
137 VERIFY3P(block_cursor, <=, block_end);
138
139 type = SM2_TYPE_DECODE(e);
140 raw_offset = SM2_OFFSET_DECODE(e);
141 }
142
143 uint64_t entry_offset = (raw_offset << sm->sm_shift) +
144 sm->sm_start;
145 uint64_t entry_run = raw_run << sm->sm_shift;
146
147 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
148 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
149 ASSERT3U(entry_offset, >=, sm->sm_start);
150 ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
151 ASSERT3U(entry_run, <=, sm->sm_size);
152 ASSERT3U(entry_offset + entry_run, <=,
153 sm->sm_start + sm->sm_size);
154
155 space_map_entry_t sme = {
156 .sme_type = type,
157 .sme_vdev = vdev_id,
158 .sme_offset = entry_offset,
159 .sme_run = entry_run
160 };
161 error = callback(&sme, arg);
162 }
163 dmu_buf_rele(db, FTAG);
164 }
165 return (error);
166 }
167
168 /*
169 * Reads the entries from the last block of the space map into
170 * buf in reverse order. Populates nwords with number of words
171 * in the last block.
172 *
173 * Refer to block comment within space_map_incremental_destroy()
174 * to understand why this function is needed.
175 */
176 static int
177 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
178 uint64_t bufsz, uint64_t *nwords)
179 {
180 int error = 0;
181 dmu_buf_t *db;
182
183 /*
184 * Find the offset of the last word in the space map and use
185 * that to read the last block of the space map with
186 * dmu_buf_hold().
187 */
188 uint64_t last_word_offset =
189 sm->sm_phys->smp_objsize - sizeof (uint64_t);
190 error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
191 FTAG, &db, DMU_READ_NO_PREFETCH);
192 if (error != 0)
193 return (error);
194
195 ASSERT3U(sm->sm_object, ==, db->db_object);
196 ASSERT3U(sm->sm_blksz, ==, db->db_size);
197 ASSERT3U(bufsz, >=, db->db_size);
198 ASSERT(nwords != NULL);
199
200 uint64_t *words = db->db_data;
201 *nwords =
202 (sm->sm_phys->smp_objsize - db->db_offset) / sizeof (uint64_t);
203
204 ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
205
206 uint64_t n = *nwords;
207 uint64_t j = n - 1;
208 for (uint64_t i = 0; i < n; i++) {
209 uint64_t entry = words[i];
210 if (sm_entry_is_double_word(entry)) {
211 /*
212 * Since we are populating the buffer backwards
213 * we have to be extra careful and add the two
214 * words of the double-word entry in the right
215 * order.
216 */
217 ASSERT3U(j, >, 0);
218 buf[j - 1] = entry;
219
220 i++;
221 ASSERT3U(i, <, n);
222 entry = words[i];
223 buf[j] = entry;
224 j -= 2;
225 } else {
226 ASSERT(sm_entry_is_debug(entry) ||
227 sm_entry_is_single_word(entry));
228 buf[j] = entry;
229 j--;
230 }
231 }
232
233 /*
234 * Assert that we wrote backwards all the
235 * way to the beginning of the buffer.
236 */
237 ASSERT3S(j, ==, -1);
238
239 dmu_buf_rele(db, FTAG);
240 return (error);
241 }
242
243 /*
244 * Note: This function performs destructive actions - specifically
245 * it deletes entries from the end of the space map. Thus, callers
246 * should ensure that they are holding the appropriate locks for
247 * the space map that they provide.
248 */
249 int
250 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
251 dmu_tx_t *tx)
252 {
253 uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
254 uint64_t *buf = zio_buf_alloc(bufsz);
255
256 dmu_buf_will_dirty(sm->sm_dbuf, tx);
257
258 /*
259 * Ideally we would want to iterate from the beginning of the
260 * space map to the end in incremental steps. The issue with this
261 * approach is that we don't have any field on-disk that points
262 * us where to start between each step. We could try zeroing out
263 * entries that we've destroyed, but this doesn't work either as
264 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
265 *
266 * As a result, we destroy its entries incrementally starting from
267 * the end after applying the callback to each of them.
268 *
269 * The problem with this approach is that we cannot literally
270 * iterate through the words in the space map backwards as we
271 * can't distinguish two-word space map entries from their second
272 * word. Thus we do the following:
273 *
274 * 1] We get all the entries from the last block of the space map
275 * and put them into a buffer in reverse order. This way the
276 * last entry comes first in the buffer, the second to last is
277 * second, etc.
278 * 2] We iterate through the entries in the buffer and we apply
279 * the callback to each one. As we move from entry to entry we
280 * we decrease the size of the space map, deleting effectively
281 * each entry.
282 * 3] If there are no more entries in the space map or the callback
283 * returns a value other than 0, we stop iterating over the
284 * space map. If there are entries remaining and the callback
285 * returned 0, we go back to step [1].
286 */
287 int error = 0;
288 while (space_map_length(sm) > 0 && error == 0) {
289 uint64_t nwords = 0;
290 error = space_map_reversed_last_block_entries(sm, buf, bufsz,
291 &nwords);
292 if (error != 0)
293 break;
294
295 ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
296
297 for (uint64_t i = 0; i < nwords; i++) {
298 uint64_t e = buf[i];
299
300 if (sm_entry_is_debug(e)) {
301 sm->sm_phys->smp_objsize -= sizeof (uint64_t);
302 space_map_update(sm);
303 continue;
304 }
305
306 int words = 1;
307 uint64_t raw_offset, raw_run, vdev_id;
308 maptype_t type;
309 if (sm_entry_is_single_word(e)) {
310 type = SM_TYPE_DECODE(e);
311 vdev_id = SM_NO_VDEVID;
312 raw_offset = SM_OFFSET_DECODE(e);
313 raw_run = SM_RUN_DECODE(e);
314 } else {
315 ASSERT(sm_entry_is_double_word(e));
316 words = 2;
317
318 raw_run = SM2_RUN_DECODE(e);
319 vdev_id = SM2_VDEV_DECODE(e);
320
321 /* move to the second word */
322 i++;
323 e = buf[i];
324
325 ASSERT3P(i, <=, nwords);
326
327 type = SM2_TYPE_DECODE(e);
328 raw_offset = SM2_OFFSET_DECODE(e);
329 }
330
331 uint64_t entry_offset =
332 (raw_offset << sm->sm_shift) + sm->sm_start;
333 uint64_t entry_run = raw_run << sm->sm_shift;
334
335 VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
336 VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
337 VERIFY3U(entry_offset, >=, sm->sm_start);
338 VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
339 VERIFY3U(entry_run, <=, sm->sm_size);
340 VERIFY3U(entry_offset + entry_run, <=,
341 sm->sm_start + sm->sm_size);
342
343 space_map_entry_t sme = {
344 .sme_type = type,
345 .sme_vdev = vdev_id,
346 .sme_offset = entry_offset,
347 .sme_run = entry_run
348 };
349 error = callback(&sme, arg);
350 if (error != 0)
351 break;
352
353 if (type == SM_ALLOC)
354 sm->sm_phys->smp_alloc -= entry_run;
355 else
356 sm->sm_phys->smp_alloc += entry_run;
357 sm->sm_phys->smp_objsize -= words * sizeof (uint64_t);
358 space_map_update(sm);
359 }
360 }
361
362 if (space_map_length(sm) == 0) {
363 ASSERT0(error);
364 ASSERT0(sm->sm_phys->smp_objsize);
365 ASSERT0(sm->sm_alloc);
366 }
367
368 zio_buf_free(buf, bufsz);
369 return (error);
370 }
371
372 typedef struct space_map_load_arg {
373 space_map_t *smla_sm;
374 range_tree_t *smla_rt;
375 maptype_t smla_type;
376 } space_map_load_arg_t;
377
378 static int
379 space_map_load_callback(space_map_entry_t *sme, void *arg)
380 {
381 space_map_load_arg_t *smla = arg;
382 if (sme->sme_type == smla->smla_type) {
383 VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
384 smla->smla_sm->sm_size);
385 range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
386 } else {
387 range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
388 }
389
390 return (0);
391 }
392
393 /*
394 * Load the space map disk into the specified range tree. Segments of maptype
395 * are added to the range tree, other segment types are removed.
396 */
397 int
398 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
399 {
400 uint64_t space;
401 int err;
402 space_map_load_arg_t smla;
403
404 VERIFY0(range_tree_space(rt));
405 space = space_map_allocated(sm);
406
407 if (maptype == SM_FREE) {
408 range_tree_add(rt, sm->sm_start, sm->sm_size);
409 space = sm->sm_size - space;
410 }
411
412 smla.smla_rt = rt;
413 smla.smla_sm = sm;
414 smla.smla_type = maptype;
415 err = space_map_iterate(sm, space_map_load_callback, &smla);
416
417 if (err == 0) {
418 VERIFY3U(range_tree_space(rt), ==, space);
419 } else {
420 range_tree_vacate(rt, NULL, NULL);
421 }
422
423 return (err);
424 }
425
426 void
427 space_map_histogram_clear(space_map_t *sm)
428 {
429 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
430 return;
431
432 bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
433 }
434
435 boolean_t
436 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
437 {
438 /*
439 * Verify that the in-core range tree does not have any
440 * ranges smaller than our sm_shift size.
441 */
442 for (int i = 0; i < sm->sm_shift; i++) {
443 if (rt->rt_histogram[i] != 0)
444 return (B_FALSE);
445 }
446 return (B_TRUE);
447 }
448
449 void
450 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
451 {
452 int idx = 0;
453
454 ASSERT(dmu_tx_is_syncing(tx));
455 VERIFY3U(space_map_object(sm), !=, 0);
456
457 if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
458 return;
459
460 dmu_buf_will_dirty(sm->sm_dbuf, tx);
461
462 ASSERT(space_map_histogram_verify(sm, rt));
463 /*
464 * Transfer the content of the range tree histogram to the space
465 * map histogram. The space map histogram contains 32 buckets ranging
466 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
467 * however, can represent ranges from 2^0 to 2^63. Since the space
468 * map only cares about allocatable blocks (minimum of sm_shift) we
469 * can safely ignore all ranges in the range tree smaller than sm_shift.
470 */
471 for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
472
473 /*
474 * Since the largest histogram bucket in the space map is
475 * 2^(32+sm_shift-1), we need to normalize the values in
476 * the range tree for any bucket larger than that size. For
477 * example given an sm_shift of 9, ranges larger than 2^40
478 * would get normalized as if they were 1TB ranges. Assume
479 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
480 * the calculation below would normalize this to 5 * 2^4 (16).
481 */
482 ASSERT3U(i, >=, idx + sm->sm_shift);
483 sm->sm_phys->smp_histogram[idx] +=
484 rt->rt_histogram[i] << (i - idx - sm->sm_shift);
485
486 /*
487 * Increment the space map's index as long as we haven't
488 * reached the maximum bucket size. Accumulate all ranges
489 * larger than the max bucket size into the last bucket.
490 */
491 if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
492 ASSERT3U(idx + sm->sm_shift, ==, i);
493 idx++;
494 ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
495 }
496 }
497 }
498
499 static void
500 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
501 {
502 dmu_buf_will_dirty(sm->sm_dbuf, tx);
503
504 uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
505 SM_DEBUG_ACTION_ENCODE(maptype) |
506 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
507 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
508
509 dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_objsize,
510 sizeof (dentry), &dentry, tx);
511
512 sm->sm_phys->smp_objsize += sizeof (dentry);
513 }
514
515 /*
516 * Writes one or more entries given a segment.
517 *
518 * Note: The function may release the dbuf from the pointer initially
519 * passed to it, and return a different dbuf. Also, the space map's
520 * dbuf must be dirty for the changes in sm_phys to take effect.
521 */
522 static void
523 space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype,
524 uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx)
525 {
526 ASSERT3U(words, !=, 0);
527 ASSERT3U(words, <=, 2);
528
529 /* ensure the vdev_id can be represented by the space map */
530 ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
531
532 /*
533 * if this is a single word entry, ensure that no vdev was
534 * specified.
535 */
536 IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
537
538 dmu_buf_t *db = *dbp;
539 ASSERT3U(db->db_size, ==, sm->sm_blksz);
540
541 uint64_t *block_base = db->db_data;
542 uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
543 uint64_t *block_cursor = block_base +
544 (sm->sm_phys->smp_objsize - db->db_offset) / sizeof (uint64_t);
545
546 ASSERT3P(block_cursor, <=, block_end);
547
548 uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
549 uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
550 uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
551
552 ASSERT3U(rs->rs_start, >=, sm->sm_start);
553 ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size);
554 ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size);
555 ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size);
556
557 while (size != 0) {
558 ASSERT3P(block_cursor, <=, block_end);
559
560 /*
561 * If we are at the end of this block, flush it and start
562 * writing again from the beginning.
563 */
564 if (block_cursor == block_end) {
565 dmu_buf_rele(db, tag);
566
567 uint64_t next_word_offset = sm->sm_phys->smp_objsize;
568 VERIFY0(dmu_buf_hold(sm->sm_os,
569 space_map_object(sm), next_word_offset,
570 tag, &db, DMU_READ_PREFETCH));
571 dmu_buf_will_dirty(db, tx);
572
573 /* update caller's dbuf */
574 *dbp = db;
575
576 ASSERT3U(db->db_size, ==, sm->sm_blksz);
577
578 block_base = db->db_data;
579 block_cursor = block_base;
580 block_end = block_base +
581 (db->db_size / sizeof (uint64_t));
582 }
583
584 /*
585 * If we are writing a two-word entry and we only have one
586 * word left on this block, just pad it with an empty debug
587 * entry and write the two-word entry in the next block.
588 */
589 uint64_t *next_entry = block_cursor + 1;
590 if (next_entry == block_end && words > 1) {
591 ASSERT3U(words, ==, 2);
592 *block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
593 SM_DEBUG_ACTION_ENCODE(0) |
594 SM_DEBUG_SYNCPASS_ENCODE(0) |
595 SM_DEBUG_TXG_ENCODE(0);
596 block_cursor++;
597 sm->sm_phys->smp_objsize += sizeof (uint64_t);
598 ASSERT3P(block_cursor, ==, block_end);
599 continue;
600 }
601
602 uint64_t run_len = MIN(size, run_max);
603 switch (words) {
604 case 1:
605 *block_cursor = SM_OFFSET_ENCODE(start) |
606 SM_TYPE_ENCODE(maptype) |
607 SM_RUN_ENCODE(run_len);
608 block_cursor++;
609 break;
610 case 2:
611 /* write the first word of the entry */
612 *block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
613 SM2_RUN_ENCODE(run_len) |
614 SM2_VDEV_ENCODE(vdev_id);
615 block_cursor++;
616
617 /* move on to the second word of the entry */
618 ASSERT3P(block_cursor, <, block_end);
619 *block_cursor = SM2_TYPE_ENCODE(maptype) |
620 SM2_OFFSET_ENCODE(start);
621 block_cursor++;
622 break;
623 default:
624 panic("%d-word space map entries are not supported",
625 words);
626 break;
627 }
628 sm->sm_phys->smp_objsize += words * sizeof (uint64_t);
629
630 start += run_len;
631 size -= run_len;
632 }
633 ASSERT0(size);
634
635 }
636
637 /*
638 * Note: The space map's dbuf must be dirty for the changes in sm_phys to
639 * take effect.
640 */
641 static void
642 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
643 uint64_t vdev_id, dmu_tx_t *tx)
644 {
645 spa_t *spa = tx->tx_pool->dp_spa;
646 dmu_buf_t *db;
647
648 space_map_write_intro_debug(sm, maptype, tx);
649
650 #ifdef DEBUG
651 /*
652 * We do this right after we write the intro debug entry
653 * because the estimate does not take it into account.
654 */
655 uint64_t initial_objsize = sm->sm_phys->smp_objsize;
656 uint64_t estimated_growth =
657 space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
658 uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
659 #endif
660
661 /*
662 * Find the offset right after the last word in the space map
663 * and use that to get a hold of the last block, so we can
664 * start appending to it.
665 */
666 uint64_t next_word_offset = sm->sm_phys->smp_objsize;
667 VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
668 next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
669 ASSERT3U(db->db_size, ==, sm->sm_blksz);
670
671 dmu_buf_will_dirty(db, tx);
672
673 avl_tree_t *t = &rt->rt_root;
674 for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
675 uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
676 uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
677 uint8_t words = 1;
678
679 /*
680 * We only write two-word entries when both of the following
681 * are true:
682 *
683 * [1] The feature is enabled.
684 * [2] The offset or run is too big for a single-word entry,
685 * or the vdev_id is set (meaning not equal to
686 * SM_NO_VDEVID).
687 *
688 * Note that for purposes of testing we've added the case that
689 * we write two-word entries occasionally when the feature is
690 * enabled and zfs_force_some_double_word_sm_entries has been
691 * set.
692 */
693 if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
694 (offset >= (1ULL << SM_OFFSET_BITS) ||
695 length > SM_RUN_MAX ||
696 vdev_id != SM_NO_VDEVID ||
697 (zfs_force_some_double_word_sm_entries &&
698 spa_get_random(100) == 0)))
699 words = 2;
700
701 space_map_write_seg(sm, rs, maptype, vdev_id, words,
702 &db, FTAG, tx);
703 }
704
705 dmu_buf_rele(db, FTAG);
706
707 #ifdef DEBUG
708 /*
709 * We expect our estimation to be based on the worst case
710 * scenario [see comment in space_map_estimate_optimal_size()].
711 * Therefore we expect the actual objsize to be equal or less
712 * than whatever we estimated it to be.
713 */
714 ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_objsize);
715 #endif
716 }
717
718 /*
719 * Note: This function manipulates the state of the given space map but
720 * does not hold any locks implicitly. Thus the caller is responsible
721 * for synchronizing writes to the space map.
722 */
723 void
724 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
725 uint64_t vdev_id, dmu_tx_t *tx)
726 {
727 objset_t *os = sm->sm_os;
728
729 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
730 VERIFY3U(space_map_object(sm), !=, 0);
731
732 dmu_buf_will_dirty(sm->sm_dbuf, tx);
733
734 /*
735 * This field is no longer necessary since the in-core space map
736 * now contains the object number but is maintained for backwards
737 * compatibility.
738 */
739 sm->sm_phys->smp_object = sm->sm_object;
740
741 if (range_tree_is_empty(rt)) {
742 VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
743 return;
744 }
745
746 if (maptype == SM_ALLOC)
747 sm->sm_phys->smp_alloc += range_tree_space(rt);
748 else
749 sm->sm_phys->smp_alloc -= range_tree_space(rt);
750
751 uint64_t nodes = avl_numnodes(&rt->rt_root);
752 uint64_t rt_space = range_tree_space(rt);
753
754 space_map_write_impl(sm, rt, maptype, vdev_id, tx);
755
756 /*
757 * Ensure that the space_map's accounting wasn't changed
758 * while we were in the middle of writing it out.
759 */
760 VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
761 VERIFY3U(range_tree_space(rt), ==, rt_space);
762 }
763
764 static int
765 space_map_open_impl(space_map_t *sm)
766 {
767 int error;
768 u_longlong_t blocks;
769
770 error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
771 if (error)
772 return (error);
773
774 dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
775 sm->sm_phys = sm->sm_dbuf->db_data;
776 return (0);
777 }
778
779 int
780 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
781 uint64_t start, uint64_t size, uint8_t shift)
782 {
783 space_map_t *sm;
784 int error;
785
786 ASSERT(*smp == NULL);
787 ASSERT(os != NULL);
788 ASSERT(object != 0);
789
790 sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
791
792 sm->sm_start = start;
793 sm->sm_size = size;
794 sm->sm_shift = shift;
795 sm->sm_os = os;
796 sm->sm_object = object;
797
798 error = space_map_open_impl(sm);
799 if (error != 0) {
800 space_map_close(sm);
801 return (error);
802 }
803 *smp = sm;
804
805 return (0);
806 }
807
808 void
809 space_map_close(space_map_t *sm)
810 {
811 if (sm == NULL)
812 return;
813
814 if (sm->sm_dbuf != NULL)
815 dmu_buf_rele(sm->sm_dbuf, sm);
816 sm->sm_dbuf = NULL;
817 sm->sm_phys = NULL;
818
819 kmem_free(sm, sizeof (*sm));
820 }
821
822 void
823 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
824 {
825 objset_t *os = sm->sm_os;
826 spa_t *spa = dmu_objset_spa(os);
827 dmu_object_info_t doi;
828
829 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
830 ASSERT(dmu_tx_is_syncing(tx));
831 VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
832
833 dmu_object_info_from_db(sm->sm_dbuf, &doi);
834
835 /*
836 * If the space map has the wrong bonus size (because
837 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
838 * the wrong block size (because space_map_blksz has changed),
839 * free and re-allocate its object with the updated sizes.
840 *
841 * Otherwise, just truncate the current object.
842 */
843 if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
844 doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
845 doi.doi_data_block_size != blocksize ||
846 doi.doi_metadata_block_size != 1 << space_map_ibs) {
847 zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
848 "object[%llu]: old bonus %u, old blocksz %u",
849 dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
850 doi.doi_bonus_size, doi.doi_data_block_size);
851
852 space_map_free(sm, tx);
853 dmu_buf_rele(sm->sm_dbuf, sm);
854
855 sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
856 VERIFY0(space_map_open_impl(sm));
857 } else {
858 VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
859
860 /*
861 * If the spacemap is reallocated, its histogram
862 * will be reset. Do the same in the common case so that
863 * bugs related to the uncommon case do not go unnoticed.
864 */
865 bzero(sm->sm_phys->smp_histogram,
866 sizeof (sm->sm_phys->smp_histogram));
867 }
868
869 dmu_buf_will_dirty(sm->sm_dbuf, tx);
870 sm->sm_phys->smp_objsize = 0;
871 sm->sm_phys->smp_alloc = 0;
872 }
873
874 /*
875 * Update the in-core space_map allocation and length values.
876 */
877 void
878 space_map_update(space_map_t *sm)
879 {
880 if (sm == NULL)
881 return;
882
883 sm->sm_alloc = sm->sm_phys->smp_alloc;
884 sm->sm_length = sm->sm_phys->smp_objsize;
885 }
886
887 uint64_t
888 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
889 {
890 spa_t *spa = dmu_objset_spa(os);
891 uint64_t object;
892 int bonuslen;
893
894 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
895 spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
896 bonuslen = sizeof (space_map_phys_t);
897 ASSERT3U(bonuslen, <=, dmu_bonus_max());
898 } else {
899 bonuslen = SPACE_MAP_SIZE_V0;
900 }
901
902 object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
903 space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
904
905 return (object);
906 }
907
908 void
909 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx)
910 {
911 spa_t *spa = dmu_objset_spa(os);
912 if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
913 dmu_object_info_t doi;
914
915 VERIFY0(dmu_object_info(os, smobj, &doi));
916 if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
917 spa_feature_decr(spa,
918 SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
919 }
920 }
921
922 VERIFY0(dmu_object_free(os, smobj, tx));
923 }
924
925 void
926 space_map_free(space_map_t *sm, dmu_tx_t *tx)
927 {
928 if (sm == NULL)
929 return;
930
931 space_map_free_obj(sm->sm_os, space_map_object(sm), tx);
932 sm->sm_object = 0;
933 }
934
935 /*
936 * Given a range tree, it makes a worst-case estimate of how much
937 * space would the tree's segments take if they were written to
938 * the given space map.
939 */
940 uint64_t
941 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
942 uint64_t vdev_id)
943 {
944 spa_t *spa = dmu_objset_spa(sm->sm_os);
945 uint64_t shift = sm->sm_shift;
946 uint64_t *histogram = rt->rt_histogram;
947 uint64_t entries_for_seg = 0;
948
949 /*
950 * In order to get a quick estimate of the optimal size that this
951 * range tree would have on-disk as a space map, we iterate through
952 * its histogram buckets instead of iterating through its nodes.
953 *
954 * Note that this is a highest-bound/worst-case estimate for the
955 * following reasons:
956 *
957 * 1] We assume that we always add a debug padding for each block
958 * we write and we also assume that we start at the last word
959 * of a block attempting to write a two-word entry.
960 * 2] Rounding up errors due to the way segments are distributed
961 * in the buckets of the range tree's histogram.
962 * 3] The activation of zfs_force_some_double_word_sm_entries
963 * (tunable) when testing.
964 *
965 * = Math and Rounding Errors =
966 *
967 * rt_histogram[i] bucket of a range tree represents the number
968 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
969 * that, we want to divide the buckets into groups: Buckets that
970 * can be represented using a single-word entry, ones that can
971 * be represented with a double-word entry, and ones that can
972 * only be represented with multiple two-word entries.
973 *
974 * [Note that if the new encoding feature is not enabled there
975 * are only two groups: single-word entry buckets and multiple
976 * single-word entry buckets. The information below assumes
977 * two-word entries enabled, but it can easily applied when
978 * the feature is not enabled]
979 *
980 * To find the highest bucket that can be represented with a
981 * single-word entry we look at the maximum run that such entry
982 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
983 * the run of a space map entry is shifted by sm_shift, thus we
984 * add it to the exponent]. This way, excluding the value of the
985 * maximum run that can be represented by a single-word entry,
986 * all runs that are smaller exist in buckets 0 to
987 * SM_RUN_BITS + shift - 1.
988 *
989 * To find the highest bucket that can be represented with a
990 * double-word entry, we follow the same approach. Finally, any
991 * bucket higher than that are represented with multiple two-word
992 * entries. To be more specific, if the highest bucket whose
993 * segments can be represented with a single two-word entry is X,
994 * then bucket X+1 will need 2 two-word entries for each of its
995 * segments, X+2 will need 4, X+3 will need 8, ...etc.
996 *
997 * With all of the above we make our estimation based on bucket
998 * groups. There is a rounding error though. As we mentioned in
999 * the example with the one-word entry, the maximum run that can
1000 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
1001 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
1002 * that length fall into the next bucket (and bucket group) where
1003 * we start counting two-word entries and this is one more reason
1004 * why the estimated size may end up being bigger than the actual
1005 * size written.
1006 */
1007 uint64_t size = 0;
1008 uint64_t idx = 0;
1009
1010 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
1011 (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
1012
1013 /*
1014 * If we are trying to force some double word entries just
1015 * assume the worst-case of every single word entry being
1016 * written as a double word entry.
1017 */
1018 uint64_t entry_size =
1019 (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
1020 zfs_force_some_double_word_sm_entries) ?
1021 (2 * sizeof (uint64_t)) : sizeof (uint64_t);
1022
1023 uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
1024 for (; idx <= single_entry_max_bucket; idx++)
1025 size += histogram[idx] * entry_size;
1026
1027 if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
1028 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1029 ASSERT3U(idx, >=, single_entry_max_bucket);
1030 entries_for_seg =
1031 1ULL << (idx - single_entry_max_bucket);
1032 size += histogram[idx] *
1033 entries_for_seg * entry_size;
1034 }
1035 return (size);
1036 }
1037 }
1038
1039 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
1040
1041 uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
1042 for (; idx <= double_entry_max_bucket; idx++)
1043 size += histogram[idx] * 2 * sizeof (uint64_t);
1044
1045 for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1046 ASSERT3U(idx, >=, double_entry_max_bucket);
1047 entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
1048 size += histogram[idx] *
1049 entries_for_seg * 2 * sizeof (uint64_t);
1050 }
1051
1052 /*
1053 * Assume the worst case where we start with the padding at the end
1054 * of the current block and we add an extra padding entry at the end
1055 * of all subsequent blocks.
1056 */
1057 size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
1058
1059 return (size);
1060 }
1061
1062 uint64_t
1063 space_map_object(space_map_t *sm)
1064 {
1065 return (sm != NULL ? sm->sm_object : 0);
1066 }
1067
1068 /*
1069 * Returns the already synced, on-disk allocated space.
1070 */
1071 uint64_t
1072 space_map_allocated(space_map_t *sm)
1073 {
1074 return (sm != NULL ? sm->sm_alloc : 0);
1075 }
1076
1077 /*
1078 * Returns the already synced, on-disk length;
1079 */
1080 uint64_t
1081 space_map_length(space_map_t *sm)
1082 {
1083 return (sm != NULL ? sm->sm_length : 0);
1084 }
1085
1086 /*
1087 * Returns the allocated space that is currently syncing.
1088 */
1089 int64_t
1090 space_map_alloc_delta(space_map_t *sm)
1091 {
1092 if (sm == NULL)
1093 return (0);
1094 ASSERT(sm->sm_dbuf != NULL);
1095 return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
1096 }