xref: /illumos-gate/usr/src/uts/common/fs/zfs/space_map.c (revision b5acee43ccb19bbf950ab21d1d87ef9d8d47ca9e)
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, 2019 by Delphix. All rights reserved.
27  * Copyright 2019 Joyent, Inc.
28  */
29 
30 #include <sys/zfs_context.h>
31 #include <sys/spa.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dnode.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/zio.h>
37 #include <sys/space_map.h>
38 #include <sys/spa_log_spacemap.h>
39 #include <sys/refcount.h>
40 #include <sys/zfeature.h>
41 
42 /*
43  * Note on space map block size:
44  *
45  * The data for a given space map can be kept on blocks of any size.
46  * Larger blocks entail fewer I/O operations, but they also cause the
47  * DMU to keep more data in-core, and also to waste more I/O bandwidth
48  * when only a few blocks have changed since the last transaction group.
49  */
50 
51 /*
52  * Enabled whenever we want to stress test the use of double-word
53  * space map entries.
54  */
55 boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
56 
57 /*
58  * Override the default indirect block size of 128K, instead using 16K for
59  * spacemaps (2^14 bytes).  This dramatically reduces write inflation since
60  * appending to a spacemap typically has to write one data block (4KB) and one
61  * or two indirect blocks (16K-32K, rather than 128K).
62  */
63 int space_map_ibs = 14;
64 
65 boolean_t
66 sm_entry_is_debug(uint64_t e)
67 {
68 	return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX);
69 }
70 
71 boolean_t
72 sm_entry_is_single_word(uint64_t e)
73 {
74 	uint8_t prefix = SM_PREFIX_DECODE(e);
75 	return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX);
76 }
77 
78 boolean_t
79 sm_entry_is_double_word(uint64_t e)
80 {
81 	return (SM_PREFIX_DECODE(e) == SM2_PREFIX);
82 }
83 
84 /*
85  * Iterate through the space map, invoking the callback on each (non-debug)
86  * space map entry. Stop after reading 'end' bytes of the space map.
87  */
88 int
89 space_map_iterate(space_map_t *sm, uint64_t end, sm_cb_t callback, void *arg)
90 {
91 	uint64_t blksz = sm->sm_blksz;
92 
93 	ASSERT3U(blksz, !=, 0);
94 	ASSERT3U(end, <=, space_map_length(sm));
95 	ASSERT0(P2PHASE(end, sizeof (uint64_t)));
96 
97 	dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, end,
98 	    ZIO_PRIORITY_SYNC_READ);
99 
100 	int error = 0;
101 	for (uint64_t block_base = 0; block_base < end && error == 0;
102 	    block_base += blksz) {
103 		dmu_buf_t *db;
104 		error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
105 		    block_base, FTAG, &db, DMU_READ_PREFETCH);
106 		if (error != 0)
107 			return (error);
108 
109 		uint64_t *block_start = db->db_data;
110 		uint64_t block_length = MIN(end - block_base, blksz);
111 		uint64_t *block_end = block_start +
112 		    (block_length / sizeof (uint64_t));
113 
114 		VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
115 		VERIFY3U(block_length, !=, 0);
116 		ASSERT3U(blksz, ==, db->db_size);
117 
118 		for (uint64_t *block_cursor = block_start;
119 		    block_cursor < block_end && error == 0; block_cursor++) {
120 			uint64_t e = *block_cursor;
121 
122 			if (sm_entry_is_debug(e)) /* Skip debug entries */
123 				continue;
124 
125 			uint64_t raw_offset, raw_run, vdev_id;
126 			maptype_t type;
127 			if (sm_entry_is_single_word(e)) {
128 				type = SM_TYPE_DECODE(e);
129 				vdev_id = SM_NO_VDEVID;
130 				raw_offset = SM_OFFSET_DECODE(e);
131 				raw_run = SM_RUN_DECODE(e);
132 			} else {
133 				/* it is a two-word entry */
134 				ASSERT(sm_entry_is_double_word(e));
135 				raw_run = SM2_RUN_DECODE(e);
136 				vdev_id = SM2_VDEV_DECODE(e);
137 
138 				/* move on to the second word */
139 				block_cursor++;
140 				e = *block_cursor;
141 				VERIFY3P(block_cursor, <=, block_end);
142 
143 				type = SM2_TYPE_DECODE(e);
144 				raw_offset = SM2_OFFSET_DECODE(e);
145 			}
146 
147 			uint64_t entry_offset = (raw_offset << sm->sm_shift) +
148 			    sm->sm_start;
149 			uint64_t entry_run = raw_run << sm->sm_shift;
150 
151 			VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
152 			VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
153 			ASSERT3U(entry_offset, >=, sm->sm_start);
154 			ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
155 			ASSERT3U(entry_run, <=, sm->sm_size);
156 			ASSERT3U(entry_offset + entry_run, <=,
157 			    sm->sm_start + sm->sm_size);
158 
159 			space_map_entry_t sme = {
160 			    .sme_type = type,
161 			    .sme_vdev = vdev_id,
162 			    .sme_offset = entry_offset,
163 			    .sme_run = entry_run
164 			};
165 			error = callback(&sme, arg);
166 		}
167 		dmu_buf_rele(db, FTAG);
168 	}
169 	return (error);
170 }
171 
172 /*
173  * Reads the entries from the last block of the space map into
174  * buf in reverse order. Populates nwords with number of words
175  * in the last block.
176  *
177  * Refer to block comment within space_map_incremental_destroy()
178  * to understand why this function is needed.
179  */
180 static int
181 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
182     uint64_t bufsz, uint64_t *nwords)
183 {
184 	int error = 0;
185 	dmu_buf_t *db;
186 
187 	/*
188 	 * Find the offset of the last word in the space map and use
189 	 * that to read the last block of the space map with
190 	 * dmu_buf_hold().
191 	 */
192 	uint64_t last_word_offset =
193 	    sm->sm_phys->smp_length - sizeof (uint64_t);
194 	error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
195 	    FTAG, &db, DMU_READ_NO_PREFETCH);
196 	if (error != 0)
197 		return (error);
198 
199 	ASSERT3U(sm->sm_object, ==, db->db_object);
200 	ASSERT3U(sm->sm_blksz, ==, db->db_size);
201 	ASSERT3U(bufsz, >=, db->db_size);
202 	ASSERT(nwords != NULL);
203 
204 	uint64_t *words = db->db_data;
205 	*nwords =
206 	    (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
207 
208 	ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
209 
210 	uint64_t n = *nwords;
211 	uint64_t j = n - 1;
212 	for (uint64_t i = 0; i < n; i++) {
213 		uint64_t entry = words[i];
214 		if (sm_entry_is_double_word(entry)) {
215 			/*
216 			 * Since we are populating the buffer backwards
217 			 * we have to be extra careful and add the two
218 			 * words of the double-word entry in the right
219 			 * order.
220 			 */
221 			ASSERT3U(j, >, 0);
222 			buf[j - 1] = entry;
223 
224 			i++;
225 			ASSERT3U(i, <, n);
226 			entry = words[i];
227 			buf[j] = entry;
228 			j -= 2;
229 		} else {
230 			ASSERT(sm_entry_is_debug(entry) ||
231 			    sm_entry_is_single_word(entry));
232 			buf[j] = entry;
233 			j--;
234 		}
235 	}
236 
237 	/*
238 	 * Assert that we wrote backwards all the
239 	 * way to the beginning of the buffer.
240 	 */
241 	ASSERT3S(j, ==, -1);
242 
243 	dmu_buf_rele(db, FTAG);
244 	return (error);
245 }
246 
247 /*
248  * Note: This function performs destructive actions - specifically
249  * it deletes entries from the end of the space map. Thus, callers
250  * should ensure that they are holding the appropriate locks for
251  * the space map that they provide.
252  */
253 int
254 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
255     dmu_tx_t *tx)
256 {
257 	uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
258 	uint64_t *buf = zio_buf_alloc(bufsz);
259 
260 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
261 
262 	/*
263 	 * Ideally we would want to iterate from the beginning of the
264 	 * space map to the end in incremental steps. The issue with this
265 	 * approach is that we don't have any field on-disk that points
266 	 * us where to start between each step. We could try zeroing out
267 	 * entries that we've destroyed, but this doesn't work either as
268 	 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
269 	 *
270 	 * As a result, we destroy its entries incrementally starting from
271 	 * the end after applying the callback to each of them.
272 	 *
273 	 * The problem with this approach is that we cannot literally
274 	 * iterate through the words in the space map backwards as we
275 	 * can't distinguish two-word space map entries from their second
276 	 * word. Thus we do the following:
277 	 *
278 	 * 1] We get all the entries from the last block of the space map
279 	 *    and put them into a buffer in reverse order. This way the
280 	 *    last entry comes first in the buffer, the second to last is
281 	 *    second, etc.
282 	 * 2] We iterate through the entries in the buffer and we apply
283 	 *    the callback to each one. As we move from entry to entry we
284 	 *    we decrease the size of the space map, deleting effectively
285 	 *    each entry.
286 	 * 3] If there are no more entries in the space map or the callback
287 	 *    returns a value other than 0, we stop iterating over the
288 	 *    space map. If there are entries remaining and the callback
289 	 *    returned 0, we go back to step [1].
290 	 */
291 	int error = 0;
292 	while (space_map_length(sm) > 0 && error == 0) {
293 		uint64_t nwords = 0;
294 		error = space_map_reversed_last_block_entries(sm, buf, bufsz,
295 		    &nwords);
296 		if (error != 0)
297 			break;
298 
299 		ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
300 
301 		for (uint64_t i = 0; i < nwords; i++) {
302 			uint64_t e = buf[i];
303 
304 			if (sm_entry_is_debug(e)) {
305 				sm->sm_phys->smp_length -= sizeof (uint64_t);
306 				continue;
307 			}
308 
309 			int words = 1;
310 			uint64_t raw_offset, raw_run, vdev_id;
311 			maptype_t type;
312 			if (sm_entry_is_single_word(e)) {
313 				type = SM_TYPE_DECODE(e);
314 				vdev_id = SM_NO_VDEVID;
315 				raw_offset = SM_OFFSET_DECODE(e);
316 				raw_run = SM_RUN_DECODE(e);
317 			} else {
318 				ASSERT(sm_entry_is_double_word(e));
319 				words = 2;
320 
321 				raw_run = SM2_RUN_DECODE(e);
322 				vdev_id = SM2_VDEV_DECODE(e);
323 
324 				/* move to the second word */
325 				i++;
326 				e = buf[i];
327 
328 				ASSERT3P(i, <=, nwords);
329 
330 				type = SM2_TYPE_DECODE(e);
331 				raw_offset = SM2_OFFSET_DECODE(e);
332 			}
333 
334 			uint64_t entry_offset =
335 			    (raw_offset << sm->sm_shift) + sm->sm_start;
336 			uint64_t entry_run = raw_run << sm->sm_shift;
337 
338 			VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
339 			VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
340 			VERIFY3U(entry_offset, >=, sm->sm_start);
341 			VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
342 			VERIFY3U(entry_run, <=, sm->sm_size);
343 			VERIFY3U(entry_offset + entry_run, <=,
344 			    sm->sm_start + sm->sm_size);
345 
346 			space_map_entry_t sme = {
347 			    .sme_type = type,
348 			    .sme_vdev = vdev_id,
349 			    .sme_offset = entry_offset,
350 			    .sme_run = entry_run
351 			};
352 			error = callback(&sme, arg);
353 			if (error != 0)
354 				break;
355 
356 			if (type == SM_ALLOC)
357 				sm->sm_phys->smp_alloc -= entry_run;
358 			else
359 				sm->sm_phys->smp_alloc += entry_run;
360 			sm->sm_phys->smp_length -= words * sizeof (uint64_t);
361 		}
362 	}
363 
364 	if (space_map_length(sm) == 0) {
365 		ASSERT0(error);
366 		ASSERT0(space_map_allocated(sm));
367 	}
368 
369 	zio_buf_free(buf, bufsz);
370 	return (error);
371 }
372 
373 typedef struct space_map_load_arg {
374 	space_map_t	*smla_sm;
375 	range_tree_t	*smla_rt;
376 	maptype_t	smla_type;
377 } space_map_load_arg_t;
378 
379 static int
380 space_map_load_callback(space_map_entry_t *sme, void *arg)
381 {
382 	space_map_load_arg_t *smla = arg;
383 	if (sme->sme_type == smla->smla_type) {
384 		VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
385 		    smla->smla_sm->sm_size);
386 		range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
387 	} else {
388 		range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
389 	}
390 
391 	return (0);
392 }
393 
394 /*
395  * Load the spacemap into the rangetree, like space_map_load. But only
396  * read the first 'length' bytes of the spacemap.
397  */
398 int
399 space_map_load_length(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
400     uint64_t length)
401 {
402 	space_map_load_arg_t smla;
403 
404 	VERIFY0(range_tree_space(rt));
405 
406 	if (maptype == SM_FREE)
407 		range_tree_add(rt, sm->sm_start, sm->sm_size);
408 
409 	smla.smla_rt = rt;
410 	smla.smla_sm = sm;
411 	smla.smla_type = maptype;
412 	int err = space_map_iterate(sm, length,
413 	    space_map_load_callback, &smla);
414 
415 	if (err != 0)
416 		range_tree_vacate(rt, NULL, NULL);
417 
418 	return (err);
419 }
420 
421 /*
422  * Load the space map disk into the specified range tree. Segments of maptype
423  * are added to the range tree, other segment types are removed.
424  */
425 int
426 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
427 {
428 	return (space_map_load_length(sm, rt, maptype, space_map_length(sm)));
429 }
430 
431 void
432 space_map_histogram_clear(space_map_t *sm)
433 {
434 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
435 		return;
436 
437 	bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
438 }
439 
440 boolean_t
441 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
442 {
443 	/*
444 	 * Verify that the in-core range tree does not have any
445 	 * ranges smaller than our sm_shift size.
446 	 */
447 	for (int i = 0; i < sm->sm_shift; i++) {
448 		if (rt->rt_histogram[i] != 0)
449 			return (B_FALSE);
450 	}
451 	return (B_TRUE);
452 }
453 
454 void
455 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
456 {
457 	int idx = 0;
458 
459 	ASSERT(dmu_tx_is_syncing(tx));
460 	VERIFY3U(space_map_object(sm), !=, 0);
461 
462 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
463 		return;
464 
465 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
466 
467 	ASSERT(space_map_histogram_verify(sm, rt));
468 	/*
469 	 * Transfer the content of the range tree histogram to the space
470 	 * map histogram. The space map histogram contains 32 buckets ranging
471 	 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
472 	 * however, can represent ranges from 2^0 to 2^63. Since the space
473 	 * map only cares about allocatable blocks (minimum of sm_shift) we
474 	 * can safely ignore all ranges in the range tree smaller than sm_shift.
475 	 */
476 	for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
477 
478 		/*
479 		 * Since the largest histogram bucket in the space map is
480 		 * 2^(32+sm_shift-1), we need to normalize the values in
481 		 * the range tree for any bucket larger than that size. For
482 		 * example given an sm_shift of 9, ranges larger than 2^40
483 		 * would get normalized as if they were 1TB ranges. Assume
484 		 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
485 		 * the calculation below would normalize this to 5 * 2^4 (16).
486 		 */
487 		ASSERT3U(i, >=, idx + sm->sm_shift);
488 		sm->sm_phys->smp_histogram[idx] +=
489 		    rt->rt_histogram[i] << (i - idx - sm->sm_shift);
490 
491 		/*
492 		 * Increment the space map's index as long as we haven't
493 		 * reached the maximum bucket size. Accumulate all ranges
494 		 * larger than the max bucket size into the last bucket.
495 		 */
496 		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
497 			ASSERT3U(idx + sm->sm_shift, ==, i);
498 			idx++;
499 			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
500 		}
501 	}
502 }
503 
504 static void
505 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
506 {
507 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
508 
509 	uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
510 	    SM_DEBUG_ACTION_ENCODE(maptype) |
511 	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
512 	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
513 
514 	dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_length,
515 	    sizeof (dentry), &dentry, tx);
516 
517 	sm->sm_phys->smp_length += sizeof (dentry);
518 }
519 
520 /*
521  * Writes one or more entries given a segment.
522  *
523  * Note: The function may release the dbuf from the pointer initially
524  * passed to it, and return a different dbuf. Also, the space map's
525  * dbuf must be dirty for the changes in sm_phys to take effect.
526  */
527 static void
528 space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype,
529     uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx)
530 {
531 	ASSERT3U(words, !=, 0);
532 	ASSERT3U(words, <=, 2);
533 
534 	/* ensure the vdev_id can be represented by the space map */
535 	ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
536 
537 	/*
538 	 * if this is a single word entry, ensure that no vdev was
539 	 * specified.
540 	 */
541 	IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
542 
543 	dmu_buf_t *db = *dbp;
544 	ASSERT3U(db->db_size, ==, sm->sm_blksz);
545 
546 	uint64_t *block_base = db->db_data;
547 	uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
548 	uint64_t *block_cursor = block_base +
549 	    (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
550 
551 	ASSERT3P(block_cursor, <=, block_end);
552 
553 	uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
554 	uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
555 	uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
556 
557 	ASSERT3U(rs->rs_start, >=, sm->sm_start);
558 	ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size);
559 	ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size);
560 	ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size);
561 
562 	while (size != 0) {
563 		ASSERT3P(block_cursor, <=, block_end);
564 
565 		/*
566 		 * If we are at the end of this block, flush it and start
567 		 * writing again from the beginning.
568 		 */
569 		if (block_cursor == block_end) {
570 			dmu_buf_rele(db, tag);
571 
572 			uint64_t next_word_offset = sm->sm_phys->smp_length;
573 			VERIFY0(dmu_buf_hold(sm->sm_os,
574 			    space_map_object(sm), next_word_offset,
575 			    tag, &db, DMU_READ_PREFETCH));
576 			dmu_buf_will_dirty(db, tx);
577 
578 			/* update caller's dbuf */
579 			*dbp = db;
580 
581 			ASSERT3U(db->db_size, ==, sm->sm_blksz);
582 
583 			block_base = db->db_data;
584 			block_cursor = block_base;
585 			block_end = block_base +
586 			    (db->db_size / sizeof (uint64_t));
587 		}
588 
589 		/*
590 		 * If we are writing a two-word entry and we only have one
591 		 * word left on this block, just pad it with an empty debug
592 		 * entry and write the two-word entry in the next block.
593 		 */
594 		uint64_t *next_entry = block_cursor + 1;
595 		if (next_entry == block_end && words > 1) {
596 			ASSERT3U(words, ==, 2);
597 			*block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
598 			    SM_DEBUG_ACTION_ENCODE(0) |
599 			    SM_DEBUG_SYNCPASS_ENCODE(0) |
600 			    SM_DEBUG_TXG_ENCODE(0);
601 			block_cursor++;
602 			sm->sm_phys->smp_length += sizeof (uint64_t);
603 			ASSERT3P(block_cursor, ==, block_end);
604 			continue;
605 		}
606 
607 		uint64_t run_len = MIN(size, run_max);
608 		switch (words) {
609 		case 1:
610 			*block_cursor = SM_OFFSET_ENCODE(start) |
611 			    SM_TYPE_ENCODE(maptype) |
612 			    SM_RUN_ENCODE(run_len);
613 			block_cursor++;
614 			break;
615 		case 2:
616 			/* write the first word of the entry */
617 			*block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
618 			    SM2_RUN_ENCODE(run_len) |
619 			    SM2_VDEV_ENCODE(vdev_id);
620 			block_cursor++;
621 
622 			/* move on to the second word of the entry */
623 			ASSERT3P(block_cursor, <, block_end);
624 			*block_cursor = SM2_TYPE_ENCODE(maptype) |
625 			    SM2_OFFSET_ENCODE(start);
626 			block_cursor++;
627 			break;
628 		default:
629 			panic("%d-word space map entries are not supported",
630 			    words);
631 			break;
632 		}
633 		sm->sm_phys->smp_length += words * sizeof (uint64_t);
634 
635 		start += run_len;
636 		size -= run_len;
637 	}
638 	ASSERT0(size);
639 
640 }
641 
642 /*
643  * Note: The space map's dbuf must be dirty for the changes in sm_phys to
644  * take effect.
645  */
646 static void
647 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
648     uint64_t vdev_id, dmu_tx_t *tx)
649 {
650 	spa_t *spa = tx->tx_pool->dp_spa;
651 	dmu_buf_t *db;
652 
653 	space_map_write_intro_debug(sm, maptype, tx);
654 
655 #ifdef DEBUG
656 	/*
657 	 * We do this right after we write the intro debug entry
658 	 * because the estimate does not take it into account.
659 	 */
660 	uint64_t initial_objsize = sm->sm_phys->smp_length;
661 	uint64_t estimated_growth =
662 	    space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
663 	uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
664 #endif
665 
666 	/*
667 	 * Find the offset right after the last word in the space map
668 	 * and use that to get a hold of the last block, so we can
669 	 * start appending to it.
670 	 */
671 	uint64_t next_word_offset = sm->sm_phys->smp_length;
672 	VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
673 	    next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
674 	ASSERT3U(db->db_size, ==, sm->sm_blksz);
675 
676 	dmu_buf_will_dirty(db, tx);
677 
678 	avl_tree_t *t = &rt->rt_root;
679 	for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
680 		uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
681 		uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
682 		uint8_t words = 1;
683 
684 		/*
685 		 * We only write two-word entries when both of the following
686 		 * are true:
687 		 *
688 		 * [1] The feature is enabled.
689 		 * [2] The offset or run is too big for a single-word entry,
690 		 *	or the vdev_id is set (meaning not equal to
691 		 *	SM_NO_VDEVID).
692 		 *
693 		 * Note that for purposes of testing we've added the case that
694 		 * we write two-word entries occasionally when the feature is
695 		 * enabled and zfs_force_some_double_word_sm_entries has been
696 		 * set.
697 		 */
698 		if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
699 		    (offset >= (1ULL << SM_OFFSET_BITS) ||
700 		    length > SM_RUN_MAX ||
701 		    vdev_id != SM_NO_VDEVID ||
702 		    (zfs_force_some_double_word_sm_entries &&
703 		    spa_get_random(100) == 0)))
704 			words = 2;
705 
706 		space_map_write_seg(sm, rs, maptype, vdev_id, words,
707 		    &db, FTAG, tx);
708 	}
709 
710 	dmu_buf_rele(db, FTAG);
711 
712 #ifdef DEBUG
713 	/*
714 	 * We expect our estimation to be based on the worst case
715 	 * scenario [see comment in space_map_estimate_optimal_size()].
716 	 * Therefore we expect the actual objsize to be equal or less
717 	 * than whatever we estimated it to be.
718 	 */
719 	ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_length);
720 #endif
721 }
722 
723 /*
724  * Note: This function manipulates the state of the given space map but
725  * does not hold any locks implicitly. Thus the caller is responsible
726  * for synchronizing writes to the space map.
727  */
728 void
729 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
730     uint64_t vdev_id, dmu_tx_t *tx)
731 {
732 	objset_t *os = sm->sm_os;
733 
734 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
735 	VERIFY3U(space_map_object(sm), !=, 0);
736 
737 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
738 
739 	/*
740 	 * This field is no longer necessary since the in-core space map
741 	 * now contains the object number but is maintained for backwards
742 	 * compatibility.
743 	 */
744 	sm->sm_phys->smp_object = sm->sm_object;
745 
746 	if (range_tree_is_empty(rt)) {
747 		VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
748 		return;
749 	}
750 
751 	if (maptype == SM_ALLOC)
752 		sm->sm_phys->smp_alloc += range_tree_space(rt);
753 	else
754 		sm->sm_phys->smp_alloc -= range_tree_space(rt);
755 
756 	uint64_t nodes = avl_numnodes(&rt->rt_root);
757 	uint64_t rt_space = range_tree_space(rt);
758 
759 	space_map_write_impl(sm, rt, maptype, vdev_id, tx);
760 
761 	/*
762 	 * Ensure that the space_map's accounting wasn't changed
763 	 * while we were in the middle of writing it out.
764 	 */
765 	VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
766 	VERIFY3U(range_tree_space(rt), ==, rt_space);
767 }
768 
769 static int
770 space_map_open_impl(space_map_t *sm)
771 {
772 	int error;
773 	u_longlong_t blocks;
774 
775 	error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
776 	if (error)
777 		return (error);
778 
779 	dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
780 	sm->sm_phys = sm->sm_dbuf->db_data;
781 	return (0);
782 }
783 
784 int
785 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
786     uint64_t start, uint64_t size, uint8_t shift)
787 {
788 	space_map_t *sm;
789 	int error;
790 
791 	ASSERT(*smp == NULL);
792 	ASSERT(os != NULL);
793 	ASSERT(object != 0);
794 
795 	sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
796 
797 	sm->sm_start = start;
798 	sm->sm_size = size;
799 	sm->sm_shift = shift;
800 	sm->sm_os = os;
801 	sm->sm_object = object;
802 
803 	error = space_map_open_impl(sm);
804 	if (error != 0) {
805 		space_map_close(sm);
806 		return (error);
807 	}
808 	*smp = sm;
809 
810 	return (0);
811 }
812 
813 void
814 space_map_close(space_map_t *sm)
815 {
816 	if (sm == NULL)
817 		return;
818 
819 	if (sm->sm_dbuf != NULL)
820 		dmu_buf_rele(sm->sm_dbuf, sm);
821 	sm->sm_dbuf = NULL;
822 	sm->sm_phys = NULL;
823 
824 	kmem_free(sm, sizeof (*sm));
825 }
826 
827 void
828 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
829 {
830 	objset_t *os = sm->sm_os;
831 	spa_t *spa = dmu_objset_spa(os);
832 	dmu_object_info_t doi;
833 
834 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
835 	ASSERT(dmu_tx_is_syncing(tx));
836 	VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
837 
838 	dmu_object_info_from_db(sm->sm_dbuf, &doi);
839 
840 	/*
841 	 * If the space map has the wrong bonus size (because
842 	 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
843 	 * the wrong block size (because space_map_blksz has changed),
844 	 * free and re-allocate its object with the updated sizes.
845 	 *
846 	 * Otherwise, just truncate the current object.
847 	 */
848 	if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
849 	    doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
850 	    doi.doi_data_block_size != blocksize ||
851 	    doi.doi_metadata_block_size != 1 << space_map_ibs) {
852 		zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
853 		    "object[%llu]: old bonus %u, old blocksz %u",
854 		    dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
855 		    doi.doi_bonus_size, doi.doi_data_block_size);
856 
857 		space_map_free(sm, tx);
858 		dmu_buf_rele(sm->sm_dbuf, sm);
859 
860 		sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
861 		VERIFY0(space_map_open_impl(sm));
862 	} else {
863 		VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
864 
865 		/*
866 		 * If the spacemap is reallocated, its histogram
867 		 * will be reset.  Do the same in the common case so that
868 		 * bugs related to the uncommon case do not go unnoticed.
869 		 */
870 		bzero(sm->sm_phys->smp_histogram,
871 		    sizeof (sm->sm_phys->smp_histogram));
872 	}
873 
874 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
875 	sm->sm_phys->smp_length = 0;
876 	sm->sm_phys->smp_alloc = 0;
877 }
878 
879 uint64_t
880 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
881 {
882 	spa_t *spa = dmu_objset_spa(os);
883 	uint64_t object;
884 	int bonuslen;
885 
886 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
887 		spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
888 		bonuslen = sizeof (space_map_phys_t);
889 		ASSERT3U(bonuslen, <=, dmu_bonus_max());
890 	} else {
891 		bonuslen = SPACE_MAP_SIZE_V0;
892 	}
893 
894 	object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
895 	    space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
896 
897 	return (object);
898 }
899 
900 void
901 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx)
902 {
903 	spa_t *spa = dmu_objset_spa(os);
904 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
905 		dmu_object_info_t doi;
906 
907 		VERIFY0(dmu_object_info(os, smobj, &doi));
908 		if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
909 			spa_feature_decr(spa,
910 			    SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
911 		}
912 	}
913 
914 	VERIFY0(dmu_object_free(os, smobj, tx));
915 }
916 
917 void
918 space_map_free(space_map_t *sm, dmu_tx_t *tx)
919 {
920 	if (sm == NULL)
921 		return;
922 
923 	space_map_free_obj(sm->sm_os, space_map_object(sm), tx);
924 	sm->sm_object = 0;
925 }
926 
927 /*
928  * Given a range tree, it makes a worst-case estimate of how much
929  * space would the tree's segments take if they were written to
930  * the given space map.
931  */
932 uint64_t
933 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
934     uint64_t vdev_id)
935 {
936 	spa_t *spa = dmu_objset_spa(sm->sm_os);
937 	uint64_t shift = sm->sm_shift;
938 	uint64_t *histogram = rt->rt_histogram;
939 	uint64_t entries_for_seg = 0;
940 
941 	/*
942 	 * In order to get a quick estimate of the optimal size that this
943 	 * range tree would have on-disk as a space map, we iterate through
944 	 * its histogram buckets instead of iterating through its nodes.
945 	 *
946 	 * Note that this is a highest-bound/worst-case estimate for the
947 	 * following reasons:
948 	 *
949 	 * 1] We assume that we always add a debug padding for each block
950 	 *    we write and we also assume that we start at the last word
951 	 *    of a block attempting to write a two-word entry.
952 	 * 2] Rounding up errors due to the way segments are distributed
953 	 *    in the buckets of the range tree's histogram.
954 	 * 3] The activation of zfs_force_some_double_word_sm_entries
955 	 *    (tunable) when testing.
956 	 *
957 	 * = Math and Rounding Errors =
958 	 *
959 	 * rt_histogram[i] bucket of a range tree represents the number
960 	 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
961 	 * that, we want to divide the buckets into groups: Buckets that
962 	 * can be represented using a single-word entry, ones that can
963 	 * be represented with a double-word entry, and ones that can
964 	 * only be represented with multiple two-word entries.
965 	 *
966 	 * [Note that if the new encoding feature is not enabled there
967 	 * are only two groups: single-word entry buckets and multiple
968 	 * single-word entry buckets. The information below assumes
969 	 * two-word entries enabled, but it can easily applied when
970 	 * the feature is not enabled]
971 	 *
972 	 * To find the highest bucket that can be represented with a
973 	 * single-word entry we look at the maximum run that such entry
974 	 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
975 	 * the run of a space map entry is shifted by sm_shift, thus we
976 	 * add it to the exponent]. This way, excluding the value of the
977 	 * maximum run that can be represented by a single-word entry,
978 	 * all runs that are smaller exist in buckets 0 to
979 	 * SM_RUN_BITS + shift - 1.
980 	 *
981 	 * To find the highest bucket that can be represented with a
982 	 * double-word entry, we follow the same approach. Finally, any
983 	 * bucket higher than that are represented with multiple two-word
984 	 * entries. To be more specific, if the highest bucket whose
985 	 * segments can be represented with a single two-word entry is X,
986 	 * then bucket X+1 will need 2 two-word entries for each of its
987 	 * segments, X+2 will need 4, X+3 will need 8, ...etc.
988 	 *
989 	 * With all of the above we make our estimation based on bucket
990 	 * groups. There is a rounding error though. As we mentioned in
991 	 * the example with the one-word entry, the maximum run that can
992 	 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
993 	 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
994 	 * that length fall into the next bucket (and bucket group) where
995 	 * we start counting two-word entries and this is one more reason
996 	 * why the estimated size may end up being bigger than the actual
997 	 * size written.
998 	 */
999 	uint64_t size = 0;
1000 	uint64_t idx = 0;
1001 
1002 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
1003 	    (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
1004 
1005 		/*
1006 		 * If we are trying to force some double word entries just
1007 		 * assume the worst-case of every single word entry being
1008 		 * written as a double word entry.
1009 		 */
1010 		uint64_t entry_size =
1011 		    (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
1012 		    zfs_force_some_double_word_sm_entries) ?
1013 		    (2 * sizeof (uint64_t)) : sizeof (uint64_t);
1014 
1015 		uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
1016 		for (; idx <= single_entry_max_bucket; idx++)
1017 			size += histogram[idx] * entry_size;
1018 
1019 		if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
1020 			for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1021 				ASSERT3U(idx, >=, single_entry_max_bucket);
1022 				entries_for_seg =
1023 				    1ULL << (idx - single_entry_max_bucket);
1024 				size += histogram[idx] *
1025 				    entries_for_seg * entry_size;
1026 			}
1027 			return (size);
1028 		}
1029 	}
1030 
1031 	ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
1032 
1033 	uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
1034 	for (; idx <= double_entry_max_bucket; idx++)
1035 		size += histogram[idx] * 2 * sizeof (uint64_t);
1036 
1037 	for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1038 		ASSERT3U(idx, >=, double_entry_max_bucket);
1039 		entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
1040 		size += histogram[idx] *
1041 		    entries_for_seg * 2 * sizeof (uint64_t);
1042 	}
1043 
1044 	/*
1045 	 * Assume the worst case where we start with the padding at the end
1046 	 * of the current block and we add an extra padding entry at the end
1047 	 * of all subsequent blocks.
1048 	 */
1049 	size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
1050 
1051 	return (size);
1052 }
1053 
1054 uint64_t
1055 space_map_object(space_map_t *sm)
1056 {
1057 	return (sm != NULL ? sm->sm_object : 0);
1058 }
1059 
1060 int64_t
1061 space_map_allocated(space_map_t *sm)
1062 {
1063 	return (sm != NULL ? sm->sm_phys->smp_alloc : 0);
1064 }
1065 
1066 uint64_t
1067 space_map_length(space_map_t *sm)
1068 {
1069 	return (sm != NULL ? sm->sm_phys->smp_length : 0);
1070 }
1071 
1072 uint64_t
1073 space_map_nblocks(space_map_t *sm)
1074 {
1075 	if (sm == NULL)
1076 		return (0);
1077 	return (DIV_ROUND_UP(space_map_length(sm), sm->sm_blksz));
1078 }
1079