xref: /illumos-gate/usr/src/uts/common/fs/zfs/space_map.c (revision 555d674d)
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, 2018 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. Stop after reading 'end' bytes of the space map.
85  */
86 int
87 space_map_iterate(space_map_t *sm, uint64_t end, sm_cb_t callback, void *arg)
88 {
89 	uint64_t blksz = sm->sm_blksz;
90 
91 	ASSERT3U(blksz, !=, 0);
92 	ASSERT3U(end, <=, space_map_length(sm));
93 	ASSERT0(P2PHASE(end, sizeof (uint64_t)));
94 
95 	dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, end,
96 	    ZIO_PRIORITY_SYNC_READ);
97 
98 	int error = 0;
99 	for (uint64_t block_base = 0; block_base < end && error == 0;
100 	    block_base += blksz) {
101 		dmu_buf_t *db;
102 		error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
103 		    block_base, FTAG, &db, DMU_READ_PREFETCH);
104 		if (error != 0)
105 			return (error);
106 
107 		uint64_t *block_start = db->db_data;
108 		uint64_t block_length = MIN(end - block_base, blksz);
109 		uint64_t *block_end = block_start +
110 		    (block_length / sizeof (uint64_t));
111 
112 		VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
113 		VERIFY3U(block_length, !=, 0);
114 		ASSERT3U(blksz, ==, db->db_size);
115 
116 		for (uint64_t *block_cursor = block_start;
117 		    block_cursor < block_end && error == 0; block_cursor++) {
118 			uint64_t e = *block_cursor;
119 
120 			if (sm_entry_is_debug(e)) /* Skip debug entries */
121 				continue;
122 
123 			uint64_t raw_offset, raw_run, vdev_id;
124 			maptype_t type;
125 			if (sm_entry_is_single_word(e)) {
126 				type = SM_TYPE_DECODE(e);
127 				vdev_id = SM_NO_VDEVID;
128 				raw_offset = SM_OFFSET_DECODE(e);
129 				raw_run = SM_RUN_DECODE(e);
130 			} else {
131 				/* it is a two-word entry */
132 				ASSERT(sm_entry_is_double_word(e));
133 				raw_run = SM2_RUN_DECODE(e);
134 				vdev_id = SM2_VDEV_DECODE(e);
135 
136 				/* move on to the second word */
137 				block_cursor++;
138 				e = *block_cursor;
139 				VERIFY3P(block_cursor, <=, block_end);
140 
141 				type = SM2_TYPE_DECODE(e);
142 				raw_offset = SM2_OFFSET_DECODE(e);
143 			}
144 
145 			uint64_t entry_offset = (raw_offset << sm->sm_shift) +
146 			    sm->sm_start;
147 			uint64_t entry_run = raw_run << sm->sm_shift;
148 
149 			VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
150 			VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
151 			ASSERT3U(entry_offset, >=, sm->sm_start);
152 			ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
153 			ASSERT3U(entry_run, <=, sm->sm_size);
154 			ASSERT3U(entry_offset + entry_run, <=,
155 			    sm->sm_start + sm->sm_size);
156 
157 			space_map_entry_t sme = {
158 			    .sme_type = type,
159 			    .sme_vdev = vdev_id,
160 			    .sme_offset = entry_offset,
161 			    .sme_run = entry_run
162 			};
163 			error = callback(&sme, arg);
164 		}
165 		dmu_buf_rele(db, FTAG);
166 	}
167 	return (error);
168 }
169 
170 /*
171  * Reads the entries from the last block of the space map into
172  * buf in reverse order. Populates nwords with number of words
173  * in the last block.
174  *
175  * Refer to block comment within space_map_incremental_destroy()
176  * to understand why this function is needed.
177  */
178 static int
179 space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
180     uint64_t bufsz, uint64_t *nwords)
181 {
182 	int error = 0;
183 	dmu_buf_t *db;
184 
185 	/*
186 	 * Find the offset of the last word in the space map and use
187 	 * that to read the last block of the space map with
188 	 * dmu_buf_hold().
189 	 */
190 	uint64_t last_word_offset =
191 	    sm->sm_phys->smp_length - sizeof (uint64_t);
192 	error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
193 	    FTAG, &db, DMU_READ_NO_PREFETCH);
194 	if (error != 0)
195 		return (error);
196 
197 	ASSERT3U(sm->sm_object, ==, db->db_object);
198 	ASSERT3U(sm->sm_blksz, ==, db->db_size);
199 	ASSERT3U(bufsz, >=, db->db_size);
200 	ASSERT(nwords != NULL);
201 
202 	uint64_t *words = db->db_data;
203 	*nwords =
204 	    (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
205 
206 	ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
207 
208 	uint64_t n = *nwords;
209 	uint64_t j = n - 1;
210 	for (uint64_t i = 0; i < n; i++) {
211 		uint64_t entry = words[i];
212 		if (sm_entry_is_double_word(entry)) {
213 			/*
214 			 * Since we are populating the buffer backwards
215 			 * we have to be extra careful and add the two
216 			 * words of the double-word entry in the right
217 			 * order.
218 			 */
219 			ASSERT3U(j, >, 0);
220 			buf[j - 1] = entry;
221 
222 			i++;
223 			ASSERT3U(i, <, n);
224 			entry = words[i];
225 			buf[j] = entry;
226 			j -= 2;
227 		} else {
228 			ASSERT(sm_entry_is_debug(entry) ||
229 			    sm_entry_is_single_word(entry));
230 			buf[j] = entry;
231 			j--;
232 		}
233 	}
234 
235 	/*
236 	 * Assert that we wrote backwards all the
237 	 * way to the beginning of the buffer.
238 	 */
239 	ASSERT3S(j, ==, -1);
240 
241 	dmu_buf_rele(db, FTAG);
242 	return (error);
243 }
244 
245 /*
246  * Note: This function performs destructive actions - specifically
247  * it deletes entries from the end of the space map. Thus, callers
248  * should ensure that they are holding the appropriate locks for
249  * the space map that they provide.
250  */
251 int
252 space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
253     dmu_tx_t *tx)
254 {
255 	uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
256 	uint64_t *buf = zio_buf_alloc(bufsz);
257 
258 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
259 
260 	/*
261 	 * Ideally we would want to iterate from the beginning of the
262 	 * space map to the end in incremental steps. The issue with this
263 	 * approach is that we don't have any field on-disk that points
264 	 * us where to start between each step. We could try zeroing out
265 	 * entries that we've destroyed, but this doesn't work either as
266 	 * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
267 	 *
268 	 * As a result, we destroy its entries incrementally starting from
269 	 * the end after applying the callback to each of them.
270 	 *
271 	 * The problem with this approach is that we cannot literally
272 	 * iterate through the words in the space map backwards as we
273 	 * can't distinguish two-word space map entries from their second
274 	 * word. Thus we do the following:
275 	 *
276 	 * 1] We get all the entries from the last block of the space map
277 	 *    and put them into a buffer in reverse order. This way the
278 	 *    last entry comes first in the buffer, the second to last is
279 	 *    second, etc.
280 	 * 2] We iterate through the entries in the buffer and we apply
281 	 *    the callback to each one. As we move from entry to entry we
282 	 *    we decrease the size of the space map, deleting effectively
283 	 *    each entry.
284 	 * 3] If there are no more entries in the space map or the callback
285 	 *    returns a value other than 0, we stop iterating over the
286 	 *    space map. If there are entries remaining and the callback
287 	 *    returned 0, we go back to step [1].
288 	 */
289 	int error = 0;
290 	while (space_map_length(sm) > 0 && error == 0) {
291 		uint64_t nwords = 0;
292 		error = space_map_reversed_last_block_entries(sm, buf, bufsz,
293 		    &nwords);
294 		if (error != 0)
295 			break;
296 
297 		ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
298 
299 		for (uint64_t i = 0; i < nwords; i++) {
300 			uint64_t e = buf[i];
301 
302 			if (sm_entry_is_debug(e)) {
303 				sm->sm_phys->smp_length -= sizeof (uint64_t);
304 				continue;
305 			}
306 
307 			int words = 1;
308 			uint64_t raw_offset, raw_run, vdev_id;
309 			maptype_t type;
310 			if (sm_entry_is_single_word(e)) {
311 				type = SM_TYPE_DECODE(e);
312 				vdev_id = SM_NO_VDEVID;
313 				raw_offset = SM_OFFSET_DECODE(e);
314 				raw_run = SM_RUN_DECODE(e);
315 			} else {
316 				ASSERT(sm_entry_is_double_word(e));
317 				words = 2;
318 
319 				raw_run = SM2_RUN_DECODE(e);
320 				vdev_id = SM2_VDEV_DECODE(e);
321 
322 				/* move to the second word */
323 				i++;
324 				e = buf[i];
325 
326 				ASSERT3P(i, <=, nwords);
327 
328 				type = SM2_TYPE_DECODE(e);
329 				raw_offset = SM2_OFFSET_DECODE(e);
330 			}
331 
332 			uint64_t entry_offset =
333 			    (raw_offset << sm->sm_shift) + sm->sm_start;
334 			uint64_t entry_run = raw_run << sm->sm_shift;
335 
336 			VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
337 			VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
338 			VERIFY3U(entry_offset, >=, sm->sm_start);
339 			VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
340 			VERIFY3U(entry_run, <=, sm->sm_size);
341 			VERIFY3U(entry_offset + entry_run, <=,
342 			    sm->sm_start + sm->sm_size);
343 
344 			space_map_entry_t sme = {
345 			    .sme_type = type,
346 			    .sme_vdev = vdev_id,
347 			    .sme_offset = entry_offset,
348 			    .sme_run = entry_run
349 			};
350 			error = callback(&sme, arg);
351 			if (error != 0)
352 				break;
353 
354 			if (type == SM_ALLOC)
355 				sm->sm_phys->smp_alloc -= entry_run;
356 			else
357 				sm->sm_phys->smp_alloc += entry_run;
358 			sm->sm_phys->smp_length -= words * sizeof (uint64_t);
359 		}
360 	}
361 
362 	if (space_map_length(sm) == 0) {
363 		ASSERT0(error);
364 		ASSERT0(space_map_allocated(sm));
365 	}
366 
367 	zio_buf_free(buf, bufsz);
368 	return (error);
369 }
370 
371 typedef struct space_map_load_arg {
372 	space_map_t	*smla_sm;
373 	range_tree_t	*smla_rt;
374 	maptype_t	smla_type;
375 } space_map_load_arg_t;
376 
377 static int
378 space_map_load_callback(space_map_entry_t *sme, void *arg)
379 {
380 	space_map_load_arg_t *smla = arg;
381 	if (sme->sme_type == smla->smla_type) {
382 		VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
383 		    smla->smla_sm->sm_size);
384 		range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
385 	} else {
386 		range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
387 	}
388 
389 	return (0);
390 }
391 
392 /*
393  * Load the spacemap into the rangetree, like space_map_load. But only
394  * read the first 'length' bytes of the spacemap.
395  */
396 int
397 space_map_load_length(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
398     uint64_t length)
399 {
400 	space_map_load_arg_t smla;
401 
402 	VERIFY0(range_tree_space(rt));
403 
404 	if (maptype == SM_FREE)
405 		range_tree_add(rt, sm->sm_start, sm->sm_size);
406 
407 	smla.smla_rt = rt;
408 	smla.smla_sm = sm;
409 	smla.smla_type = maptype;
410 	int err = space_map_iterate(sm, length,
411 	    space_map_load_callback, &smla);
412 
413 	if (err != 0)
414 		range_tree_vacate(rt, NULL, NULL);
415 
416 	return (err);
417 }
418 
419 /*
420  * Load the space map disk into the specified range tree. Segments of maptype
421  * are added to the range tree, other segment types are removed.
422  */
423 int
424 space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
425 {
426 	return (space_map_load_length(sm, rt, maptype, space_map_length(sm)));
427 }
428 
429 void
430 space_map_histogram_clear(space_map_t *sm)
431 {
432 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
433 		return;
434 
435 	bzero(sm->sm_phys->smp_histogram, sizeof (sm->sm_phys->smp_histogram));
436 }
437 
438 boolean_t
439 space_map_histogram_verify(space_map_t *sm, range_tree_t *rt)
440 {
441 	/*
442 	 * Verify that the in-core range tree does not have any
443 	 * ranges smaller than our sm_shift size.
444 	 */
445 	for (int i = 0; i < sm->sm_shift; i++) {
446 		if (rt->rt_histogram[i] != 0)
447 			return (B_FALSE);
448 	}
449 	return (B_TRUE);
450 }
451 
452 void
453 space_map_histogram_add(space_map_t *sm, range_tree_t *rt, dmu_tx_t *tx)
454 {
455 	int idx = 0;
456 
457 	ASSERT(dmu_tx_is_syncing(tx));
458 	VERIFY3U(space_map_object(sm), !=, 0);
459 
460 	if (sm->sm_dbuf->db_size != sizeof (space_map_phys_t))
461 		return;
462 
463 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
464 
465 	ASSERT(space_map_histogram_verify(sm, rt));
466 	/*
467 	 * Transfer the content of the range tree histogram to the space
468 	 * map histogram. The space map histogram contains 32 buckets ranging
469 	 * between 2^sm_shift to 2^(32+sm_shift-1). The range tree,
470 	 * however, can represent ranges from 2^0 to 2^63. Since the space
471 	 * map only cares about allocatable blocks (minimum of sm_shift) we
472 	 * can safely ignore all ranges in the range tree smaller than sm_shift.
473 	 */
474 	for (int i = sm->sm_shift; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
475 
476 		/*
477 		 * Since the largest histogram bucket in the space map is
478 		 * 2^(32+sm_shift-1), we need to normalize the values in
479 		 * the range tree for any bucket larger than that size. For
480 		 * example given an sm_shift of 9, ranges larger than 2^40
481 		 * would get normalized as if they were 1TB ranges. Assume
482 		 * the range tree had a count of 5 in the 2^44 (16TB) bucket,
483 		 * the calculation below would normalize this to 5 * 2^4 (16).
484 		 */
485 		ASSERT3U(i, >=, idx + sm->sm_shift);
486 		sm->sm_phys->smp_histogram[idx] +=
487 		    rt->rt_histogram[i] << (i - idx - sm->sm_shift);
488 
489 		/*
490 		 * Increment the space map's index as long as we haven't
491 		 * reached the maximum bucket size. Accumulate all ranges
492 		 * larger than the max bucket size into the last bucket.
493 		 */
494 		if (idx < SPACE_MAP_HISTOGRAM_SIZE - 1) {
495 			ASSERT3U(idx + sm->sm_shift, ==, i);
496 			idx++;
497 			ASSERT3U(idx, <, SPACE_MAP_HISTOGRAM_SIZE);
498 		}
499 	}
500 }
501 
502 static void
503 space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
504 {
505 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
506 
507 	uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
508 	    SM_DEBUG_ACTION_ENCODE(maptype) |
509 	    SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
510 	    SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
511 
512 	dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_length,
513 	    sizeof (dentry), &dentry, tx);
514 
515 	sm->sm_phys->smp_length += sizeof (dentry);
516 }
517 
518 /*
519  * Writes one or more entries given a segment.
520  *
521  * Note: The function may release the dbuf from the pointer initially
522  * passed to it, and return a different dbuf. Also, the space map's
523  * dbuf must be dirty for the changes in sm_phys to take effect.
524  */
525 static void
526 space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype,
527     uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx)
528 {
529 	ASSERT3U(words, !=, 0);
530 	ASSERT3U(words, <=, 2);
531 
532 	/* ensure the vdev_id can be represented by the space map */
533 	ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
534 
535 	/*
536 	 * if this is a single word entry, ensure that no vdev was
537 	 * specified.
538 	 */
539 	IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
540 
541 	dmu_buf_t *db = *dbp;
542 	ASSERT3U(db->db_size, ==, sm->sm_blksz);
543 
544 	uint64_t *block_base = db->db_data;
545 	uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
546 	uint64_t *block_cursor = block_base +
547 	    (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
548 
549 	ASSERT3P(block_cursor, <=, block_end);
550 
551 	uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
552 	uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
553 	uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
554 
555 	ASSERT3U(rs->rs_start, >=, sm->sm_start);
556 	ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size);
557 	ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size);
558 	ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size);
559 
560 	while (size != 0) {
561 		ASSERT3P(block_cursor, <=, block_end);
562 
563 		/*
564 		 * If we are at the end of this block, flush it and start
565 		 * writing again from the beginning.
566 		 */
567 		if (block_cursor == block_end) {
568 			dmu_buf_rele(db, tag);
569 
570 			uint64_t next_word_offset = sm->sm_phys->smp_length;
571 			VERIFY0(dmu_buf_hold(sm->sm_os,
572 			    space_map_object(sm), next_word_offset,
573 			    tag, &db, DMU_READ_PREFETCH));
574 			dmu_buf_will_dirty(db, tx);
575 
576 			/* update caller's dbuf */
577 			*dbp = db;
578 
579 			ASSERT3U(db->db_size, ==, sm->sm_blksz);
580 
581 			block_base = db->db_data;
582 			block_cursor = block_base;
583 			block_end = block_base +
584 			    (db->db_size / sizeof (uint64_t));
585 		}
586 
587 		/*
588 		 * If we are writing a two-word entry and we only have one
589 		 * word left on this block, just pad it with an empty debug
590 		 * entry and write the two-word entry in the next block.
591 		 */
592 		uint64_t *next_entry = block_cursor + 1;
593 		if (next_entry == block_end && words > 1) {
594 			ASSERT3U(words, ==, 2);
595 			*block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
596 			    SM_DEBUG_ACTION_ENCODE(0) |
597 			    SM_DEBUG_SYNCPASS_ENCODE(0) |
598 			    SM_DEBUG_TXG_ENCODE(0);
599 			block_cursor++;
600 			sm->sm_phys->smp_length += sizeof (uint64_t);
601 			ASSERT3P(block_cursor, ==, block_end);
602 			continue;
603 		}
604 
605 		uint64_t run_len = MIN(size, run_max);
606 		switch (words) {
607 		case 1:
608 			*block_cursor = SM_OFFSET_ENCODE(start) |
609 			    SM_TYPE_ENCODE(maptype) |
610 			    SM_RUN_ENCODE(run_len);
611 			block_cursor++;
612 			break;
613 		case 2:
614 			/* write the first word of the entry */
615 			*block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
616 			    SM2_RUN_ENCODE(run_len) |
617 			    SM2_VDEV_ENCODE(vdev_id);
618 			block_cursor++;
619 
620 			/* move on to the second word of the entry */
621 			ASSERT3P(block_cursor, <, block_end);
622 			*block_cursor = SM2_TYPE_ENCODE(maptype) |
623 			    SM2_OFFSET_ENCODE(start);
624 			block_cursor++;
625 			break;
626 		default:
627 			panic("%d-word space map entries are not supported",
628 			    words);
629 			break;
630 		}
631 		sm->sm_phys->smp_length += words * sizeof (uint64_t);
632 
633 		start += run_len;
634 		size -= run_len;
635 	}
636 	ASSERT0(size);
637 
638 }
639 
640 /*
641  * Note: The space map's dbuf must be dirty for the changes in sm_phys to
642  * take effect.
643  */
644 static void
645 space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
646     uint64_t vdev_id, dmu_tx_t *tx)
647 {
648 	spa_t *spa = tx->tx_pool->dp_spa;
649 	dmu_buf_t *db;
650 
651 	space_map_write_intro_debug(sm, maptype, tx);
652 
653 #ifdef DEBUG
654 	/*
655 	 * We do this right after we write the intro debug entry
656 	 * because the estimate does not take it into account.
657 	 */
658 	uint64_t initial_objsize = sm->sm_phys->smp_length;
659 	uint64_t estimated_growth =
660 	    space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
661 	uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
662 #endif
663 
664 	/*
665 	 * Find the offset right after the last word in the space map
666 	 * and use that to get a hold of the last block, so we can
667 	 * start appending to it.
668 	 */
669 	uint64_t next_word_offset = sm->sm_phys->smp_length;
670 	VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
671 	    next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
672 	ASSERT3U(db->db_size, ==, sm->sm_blksz);
673 
674 	dmu_buf_will_dirty(db, tx);
675 
676 	avl_tree_t *t = &rt->rt_root;
677 	for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
678 		uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
679 		uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
680 		uint8_t words = 1;
681 
682 		/*
683 		 * We only write two-word entries when both of the following
684 		 * are true:
685 		 *
686 		 * [1] The feature is enabled.
687 		 * [2] The offset or run is too big for a single-word entry,
688 		 *	or the vdev_id is set (meaning not equal to
689 		 *	SM_NO_VDEVID).
690 		 *
691 		 * Note that for purposes of testing we've added the case that
692 		 * we write two-word entries occasionally when the feature is
693 		 * enabled and zfs_force_some_double_word_sm_entries has been
694 		 * set.
695 		 */
696 		if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
697 		    (offset >= (1ULL << SM_OFFSET_BITS) ||
698 		    length > SM_RUN_MAX ||
699 		    vdev_id != SM_NO_VDEVID ||
700 		    (zfs_force_some_double_word_sm_entries &&
701 		    spa_get_random(100) == 0)))
702 			words = 2;
703 
704 		space_map_write_seg(sm, rs, maptype, vdev_id, words,
705 		    &db, FTAG, tx);
706 	}
707 
708 	dmu_buf_rele(db, FTAG);
709 
710 #ifdef DEBUG
711 	/*
712 	 * We expect our estimation to be based on the worst case
713 	 * scenario [see comment in space_map_estimate_optimal_size()].
714 	 * Therefore we expect the actual objsize to be equal or less
715 	 * than whatever we estimated it to be.
716 	 */
717 	ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_length);
718 #endif
719 }
720 
721 /*
722  * Note: This function manipulates the state of the given space map but
723  * does not hold any locks implicitly. Thus the caller is responsible
724  * for synchronizing writes to the space map.
725  */
726 void
727 space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
728     uint64_t vdev_id, dmu_tx_t *tx)
729 {
730 	objset_t *os = sm->sm_os;
731 
732 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
733 	VERIFY3U(space_map_object(sm), !=, 0);
734 
735 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
736 
737 	/*
738 	 * This field is no longer necessary since the in-core space map
739 	 * now contains the object number but is maintained for backwards
740 	 * compatibility.
741 	 */
742 	sm->sm_phys->smp_object = sm->sm_object;
743 
744 	if (range_tree_is_empty(rt)) {
745 		VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
746 		return;
747 	}
748 
749 	if (maptype == SM_ALLOC)
750 		sm->sm_phys->smp_alloc += range_tree_space(rt);
751 	else
752 		sm->sm_phys->smp_alloc -= range_tree_space(rt);
753 
754 	uint64_t nodes = avl_numnodes(&rt->rt_root);
755 	uint64_t rt_space = range_tree_space(rt);
756 
757 	space_map_write_impl(sm, rt, maptype, vdev_id, tx);
758 
759 	/*
760 	 * Ensure that the space_map's accounting wasn't changed
761 	 * while we were in the middle of writing it out.
762 	 */
763 	VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
764 	VERIFY3U(range_tree_space(rt), ==, rt_space);
765 }
766 
767 static int
768 space_map_open_impl(space_map_t *sm)
769 {
770 	int error;
771 	u_longlong_t blocks;
772 
773 	error = dmu_bonus_hold(sm->sm_os, sm->sm_object, sm, &sm->sm_dbuf);
774 	if (error)
775 		return (error);
776 
777 	dmu_object_size_from_db(sm->sm_dbuf, &sm->sm_blksz, &blocks);
778 	sm->sm_phys = sm->sm_dbuf->db_data;
779 	return (0);
780 }
781 
782 int
783 space_map_open(space_map_t **smp, objset_t *os, uint64_t object,
784     uint64_t start, uint64_t size, uint8_t shift)
785 {
786 	space_map_t *sm;
787 	int error;
788 
789 	ASSERT(*smp == NULL);
790 	ASSERT(os != NULL);
791 	ASSERT(object != 0);
792 
793 	sm = kmem_zalloc(sizeof (space_map_t), KM_SLEEP);
794 
795 	sm->sm_start = start;
796 	sm->sm_size = size;
797 	sm->sm_shift = shift;
798 	sm->sm_os = os;
799 	sm->sm_object = object;
800 
801 	error = space_map_open_impl(sm);
802 	if (error != 0) {
803 		space_map_close(sm);
804 		return (error);
805 	}
806 	*smp = sm;
807 
808 	return (0);
809 }
810 
811 void
812 space_map_close(space_map_t *sm)
813 {
814 	if (sm == NULL)
815 		return;
816 
817 	if (sm->sm_dbuf != NULL)
818 		dmu_buf_rele(sm->sm_dbuf, sm);
819 	sm->sm_dbuf = NULL;
820 	sm->sm_phys = NULL;
821 
822 	kmem_free(sm, sizeof (*sm));
823 }
824 
825 void
826 space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
827 {
828 	objset_t *os = sm->sm_os;
829 	spa_t *spa = dmu_objset_spa(os);
830 	dmu_object_info_t doi;
831 
832 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
833 	ASSERT(dmu_tx_is_syncing(tx));
834 	VERIFY3U(dmu_tx_get_txg(tx), <=, spa_final_dirty_txg(spa));
835 
836 	dmu_object_info_from_db(sm->sm_dbuf, &doi);
837 
838 	/*
839 	 * If the space map has the wrong bonus size (because
840 	 * SPA_FEATURE_SPACEMAP_HISTOGRAM has recently been enabled), or
841 	 * the wrong block size (because space_map_blksz has changed),
842 	 * free and re-allocate its object with the updated sizes.
843 	 *
844 	 * Otherwise, just truncate the current object.
845 	 */
846 	if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
847 	    doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
848 	    doi.doi_data_block_size != blocksize ||
849 	    doi.doi_metadata_block_size != 1 << space_map_ibs) {
850 		zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
851 		    "object[%llu]: old bonus %u, old blocksz %u",
852 		    dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
853 		    doi.doi_bonus_size, doi.doi_data_block_size);
854 
855 		space_map_free(sm, tx);
856 		dmu_buf_rele(sm->sm_dbuf, sm);
857 
858 		sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
859 		VERIFY0(space_map_open_impl(sm));
860 	} else {
861 		VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
862 
863 		/*
864 		 * If the spacemap is reallocated, its histogram
865 		 * will be reset.  Do the same in the common case so that
866 		 * bugs related to the uncommon case do not go unnoticed.
867 		 */
868 		bzero(sm->sm_phys->smp_histogram,
869 		    sizeof (sm->sm_phys->smp_histogram));
870 	}
871 
872 	dmu_buf_will_dirty(sm->sm_dbuf, tx);
873 	sm->sm_phys->smp_length = 0;
874 	sm->sm_phys->smp_alloc = 0;
875 }
876 
877 uint64_t
878 space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
879 {
880 	spa_t *spa = dmu_objset_spa(os);
881 	uint64_t object;
882 	int bonuslen;
883 
884 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
885 		spa_feature_incr(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
886 		bonuslen = sizeof (space_map_phys_t);
887 		ASSERT3U(bonuslen, <=, dmu_bonus_max());
888 	} else {
889 		bonuslen = SPACE_MAP_SIZE_V0;
890 	}
891 
892 	object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
893 	    space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
894 
895 	return (object);
896 }
897 
898 void
899 space_map_free_obj(objset_t *os, uint64_t smobj, dmu_tx_t *tx)
900 {
901 	spa_t *spa = dmu_objset_spa(os);
902 	if (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM)) {
903 		dmu_object_info_t doi;
904 
905 		VERIFY0(dmu_object_info(os, smobj, &doi));
906 		if (doi.doi_bonus_size != SPACE_MAP_SIZE_V0) {
907 			spa_feature_decr(spa,
908 			    SPA_FEATURE_SPACEMAP_HISTOGRAM, tx);
909 		}
910 	}
911 
912 	VERIFY0(dmu_object_free(os, smobj, tx));
913 }
914 
915 void
916 space_map_free(space_map_t *sm, dmu_tx_t *tx)
917 {
918 	if (sm == NULL)
919 		return;
920 
921 	space_map_free_obj(sm->sm_os, space_map_object(sm), tx);
922 	sm->sm_object = 0;
923 }
924 
925 /*
926  * Given a range tree, it makes a worst-case estimate of how much
927  * space would the tree's segments take if they were written to
928  * the given space map.
929  */
930 uint64_t
931 space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
932     uint64_t vdev_id)
933 {
934 	spa_t *spa = dmu_objset_spa(sm->sm_os);
935 	uint64_t shift = sm->sm_shift;
936 	uint64_t *histogram = rt->rt_histogram;
937 	uint64_t entries_for_seg = 0;
938 
939 	/*
940 	 * In order to get a quick estimate of the optimal size that this
941 	 * range tree would have on-disk as a space map, we iterate through
942 	 * its histogram buckets instead of iterating through its nodes.
943 	 *
944 	 * Note that this is a highest-bound/worst-case estimate for the
945 	 * following reasons:
946 	 *
947 	 * 1] We assume that we always add a debug padding for each block
948 	 *    we write and we also assume that we start at the last word
949 	 *    of a block attempting to write a two-word entry.
950 	 * 2] Rounding up errors due to the way segments are distributed
951 	 *    in the buckets of the range tree's histogram.
952 	 * 3] The activation of zfs_force_some_double_word_sm_entries
953 	 *    (tunable) when testing.
954 	 *
955 	 * = Math and Rounding Errors =
956 	 *
957 	 * rt_histogram[i] bucket of a range tree represents the number
958 	 * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
959 	 * that, we want to divide the buckets into groups: Buckets that
960 	 * can be represented using a single-word entry, ones that can
961 	 * be represented with a double-word entry, and ones that can
962 	 * only be represented with multiple two-word entries.
963 	 *
964 	 * [Note that if the new encoding feature is not enabled there
965 	 * are only two groups: single-word entry buckets and multiple
966 	 * single-word entry buckets. The information below assumes
967 	 * two-word entries enabled, but it can easily applied when
968 	 * the feature is not enabled]
969 	 *
970 	 * To find the highest bucket that can be represented with a
971 	 * single-word entry we look at the maximum run that such entry
972 	 * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
973 	 * the run of a space map entry is shifted by sm_shift, thus we
974 	 * add it to the exponent]. This way, excluding the value of the
975 	 * maximum run that can be represented by a single-word entry,
976 	 * all runs that are smaller exist in buckets 0 to
977 	 * SM_RUN_BITS + shift - 1.
978 	 *
979 	 * To find the highest bucket that can be represented with a
980 	 * double-word entry, we follow the same approach. Finally, any
981 	 * bucket higher than that are represented with multiple two-word
982 	 * entries. To be more specific, if the highest bucket whose
983 	 * segments can be represented with a single two-word entry is X,
984 	 * then bucket X+1 will need 2 two-word entries for each of its
985 	 * segments, X+2 will need 4, X+3 will need 8, ...etc.
986 	 *
987 	 * With all of the above we make our estimation based on bucket
988 	 * groups. There is a rounding error though. As we mentioned in
989 	 * the example with the one-word entry, the maximum run that can
990 	 * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
991 	 * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
992 	 * that length fall into the next bucket (and bucket group) where
993 	 * we start counting two-word entries and this is one more reason
994 	 * why the estimated size may end up being bigger than the actual
995 	 * size written.
996 	 */
997 	uint64_t size = 0;
998 	uint64_t idx = 0;
999 
1000 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
1001 	    (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
1002 
1003 		/*
1004 		 * If we are trying to force some double word entries just
1005 		 * assume the worst-case of every single word entry being
1006 		 * written as a double word entry.
1007 		 */
1008 		uint64_t entry_size =
1009 		    (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
1010 		    zfs_force_some_double_word_sm_entries) ?
1011 		    (2 * sizeof (uint64_t)) : sizeof (uint64_t);
1012 
1013 		uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
1014 		for (; idx <= single_entry_max_bucket; idx++)
1015 			size += histogram[idx] * entry_size;
1016 
1017 		if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
1018 			for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1019 				ASSERT3U(idx, >=, single_entry_max_bucket);
1020 				entries_for_seg =
1021 				    1ULL << (idx - single_entry_max_bucket);
1022 				size += histogram[idx] *
1023 				    entries_for_seg * entry_size;
1024 			}
1025 			return (size);
1026 		}
1027 	}
1028 
1029 	ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
1030 
1031 	uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
1032 	for (; idx <= double_entry_max_bucket; idx++)
1033 		size += histogram[idx] * 2 * sizeof (uint64_t);
1034 
1035 	for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
1036 		ASSERT3U(idx, >=, double_entry_max_bucket);
1037 		entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
1038 		size += histogram[idx] *
1039 		    entries_for_seg * 2 * sizeof (uint64_t);
1040 	}
1041 
1042 	/*
1043 	 * Assume the worst case where we start with the padding at the end
1044 	 * of the current block and we add an extra padding entry at the end
1045 	 * of all subsequent blocks.
1046 	 */
1047 	size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
1048 
1049 	return (size);
1050 }
1051 
1052 uint64_t
1053 space_map_object(space_map_t *sm)
1054 {
1055 	return (sm != NULL ? sm->sm_object : 0);
1056 }
1057 
1058 int64_t
1059 space_map_allocated(space_map_t *sm)
1060 {
1061 	return (sm != NULL ? sm->sm_phys->smp_alloc : 0);
1062 }
1063 
1064 uint64_t
1065 space_map_length(space_map_t *sm)
1066 {
1067 	return (sm != NULL ? sm->sm_phys->smp_length : 0);
1068 }
1069