xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa_log_spacemap.c (revision af1d63aba5cec023f92214c1f1faec9b489ac517)
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 /*
23  * Copyright (c) 2018, 2019 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/dmu_objset.h>
27 #include <sys/metaslab.h>
28 #include <sys/metaslab_impl.h>
29 #include <sys/spa.h>
30 #include <sys/spa_impl.h>
31 #include <sys/spa_log_spacemap.h>
32 #include <sys/vdev_impl.h>
33 #include <sys/zap.h>
34 
35 /*
36  * Log Space Maps
37  *
38  * Log space maps are an optimization in ZFS metadata allocations for pools
39  * whose workloads are primarily random-writes. Random-write workloads are also
40  * typically random-free, meaning that they are freeing from locations scattered
41  * throughout the pool. This means that each TXG we will have to append some
42  * FREE records to almost every metaslab. With log space maps, we hold their
43  * changes in memory and log them altogether in one pool-wide space map on-disk
44  * for persistence. As more blocks are accumulated in the log space maps and
45  * more unflushed changes are accounted in memory, we flush a selected group
46  * of metaslabs every TXG to relieve memory pressure and potential overheads
47  * when loading the pool. Flushing a metaslab to disk relieves memory as we
48  * flush any unflushed changes from memory to disk (i.e. the metaslab's space
49  * map) and saves import time by making old log space maps obsolete and
50  * eventually destroying them. [A log space map is said to be obsolete when all
51  * its entries have made it to their corresponding metaslab space maps].
52  *
53  * == On disk data structures used ==
54  *
55  * - The pool has a new feature flag and a new entry in the MOS. The feature
56  *   is activated when we create the first log space map and remains active
57  *   for the lifetime of the pool. The new entry in the MOS Directory [refer
58  *   to DMU_POOL_LOG_SPACEMAP_ZAP] is populated with a ZAP whose key-value
59  *   pairs are of the form <key: txg, value: log space map object for that txg>.
60  *   This entry is our on-disk reference of the log space maps that exist in
61  *   the pool for each TXG and it is used during import to load all the
62  *   metaslab unflushed changes in memory. To see how this structure is first
63  *   created and later populated refer to spa_generate_syncing_log_sm(). To see
64  *   how it is used during import time refer to spa_ld_log_sm_metadata().
65  *
66  * - Each vdev has a new entry in its vdev_top_zap (see field
67  *   VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS) which holds the msp_unflushed_txg of
68  *   each metaslab in this vdev. This field is the on-disk counterpart of the
69  *   in-memory field ms_unflushed_txg which tells us from which TXG and onwards
70  *   the metaslab haven't had its changes flushed. During import, we use this
71  *   to ignore any entries in the space map log that are for this metaslab but
72  *   from a TXG before msp_unflushed_txg. At that point, we also populate its
73  *   in-memory counterpart and from there both fields are updated every time
74  *   we flush that metaslab.
75  *
76  * - A space map is created every TXG and, during that TXG, it is used to log
77  *   all incoming changes (the log space map). When created, the log space map
78  *   is referenced in memory by spa_syncing_log_sm and its object ID is inserted
79  *   to the space map ZAP mentioned above. The log space map is closed at the
80  *   end of the TXG and will be destroyed when it becomes fully obsolete. We
81  *   know when a log space map has become obsolete by looking at the oldest
82  *   (and smallest) ms_unflushed_txg in the pool. If the value of that is bigger
83  *   than the log space map's TXG, then it means that there is no metaslab who
84  *   doesn't have the changes from that log and we can therefore destroy it.
85  *   [see spa_cleanup_old_sm_logs()].
86  *
87  * == Important in-memory structures ==
88  *
89  * - The per-spa field spa_metaslabs_by_flushed sorts all the metaslabs in
90  *   the pool by their ms_unflushed_txg field. It is primarily used for three
91  *   reasons. First of all, it is used during flushing where we try to flush
92  *   metaslabs in-order from the oldest-flushed to the most recently flushed
93  *   every TXG. Secondly, it helps us to lookup the ms_unflushed_txg of the
94  *   oldest flushed metaslab to distinguish which log space maps have become
95  *   obsolete and which ones are still relevant. Finally it tells us which
96  *   metaslabs have unflushed changes in a pool where this feature was just
97  *   enabled, as we don't immediately add all of the pool's metaslabs but we
98  *   add them over time as they go through metaslab_sync(). The reason that
99  *   we do that is to ease these pools into the behavior of the flushing
100  *   algorithm (described later on).
101  *
102  * - The per-spa field spa_sm_logs_by_txg can be thought as the in-memory
103  *   counterpart of the space map ZAP mentioned above. It's an AVL tree whose
104  *   nodes represent the log space maps in the pool. This in-memory
105  *   representation of log space maps in the pool sorts the log space maps by
106  *   the TXG that they were created (which is also the TXG of their unflushed
107  *   changes). It also contains the following extra information for each
108  *   space map:
109  *   [1] The number of metaslabs that were last flushed on that TXG. This is
110  *       important because if that counter is zero and this is the oldest
111  *       log then it means that it is also obsolete.
112  *   [2] The number of blocks of that space map. This field is used by the
113  *       block heuristic of our flushing algorithm (described later on).
114  *       It represents how many blocks of metadata changes ZFS had to write
115  *       to disk for that TXG.
116  *
117  * - The per-spa field spa_log_summary is a list of entries that summarizes
118  *   the metaslab and block counts of all the nodes of the spa_sm_logs_by_txg
119  *   AVL tree mentioned above. The reason this exists is that our flushing
120  *   algorithm (described later) tries to estimate how many metaslabs to flush
121  *   in each TXG by iterating over all the log space maps and looking at their
122  *   block counts. Summarizing that information means that don't have to
123  *   iterate through each space map, minimizing the runtime overhead of the
124  *   flushing algorithm which would be induced in syncing context. In terms of
125  *   implementation the log summary is used as a queue:
126  *   * we modify or pop entries from its head when we flush metaslabs
127  *   * we modify or append entries to its tail when we sync changes.
128  *
129  * - Each metaslab has two new range trees that hold its unflushed changes,
130  *   ms_unflushed_allocs and ms_unflushed_frees. These are always disjoint.
131  *
132  * == Flushing algorithm ==
133  *
134  * The decision of how many metaslabs to flush on a give TXG is guided by
135  * two heuristics:
136  *
137  * [1] The memory heuristic -
138  * We keep track of the memory used by the unflushed trees from all the
139  * metaslabs [see sus_memused of spa_unflushed_stats] and we ensure that it
140  * stays below a certain threshold which is determined by an arbitrary hard
141  * limit and an arbitrary percentage of the system's memory [see
142  * spa_log_exceeds_memlimit()]. When we see that the memory usage of the
143  * unflushed changes are passing that threshold, we flush metaslabs, which
144  * empties their unflushed range trees, reducing the memory used.
145  *
146  * [2] The block heuristic -
147  * We try to keep the total number of blocks in the log space maps in check
148  * so the log doesn't grow indefinitely and we don't induce a lot of overhead
149  * when loading the pool. At the same time we don't want to flush a lot of
150  * metaslabs too often as this would defeat the purpose of the log space map.
151  * As a result we set a limit in the amount of blocks that we think it's
152  * acceptable for the log space maps to have and try not to cross it.
153  * [see sus_blocklimit from spa_unflushed_stats].
154  *
155  * In order to stay below the block limit every TXG we have to estimate how
156  * many metaslabs we need to flush based on the current rate of incoming blocks
157  * and our history of log space map blocks. The main idea here is to answer
158  * the question of how many metaslabs do we need to flush in order to get rid
159  * at least an X amount of log space map blocks. We can answer this question
160  * by iterating backwards from the oldest log space map to the newest one
161  * and looking at their metaslab and block counts. At this point the log summary
162  * mentioned above comes handy as it reduces the amount of things that we have
163  * to iterate (even though it may reduce the preciseness of our estimates due
164  * to its aggregation of data). So with that in mind, we project the incoming
165  * rate of the current TXG into the future and attempt to approximate how many
166  * metaslabs would we need to flush from now in order to avoid exceeding our
167  * block limit in different points in the future (granted that we would keep
168  * flushing the same number of metaslabs for every TXG). Then we take the
169  * maximum number from all these estimates to be on the safe side. For the
170  * exact implementation details of algorithm refer to
171  * spa_estimate_metaslabs_to_flush.
172  */
173 
174 /*
175  * This is used as the block size for the space maps used for the
176  * log space map feature. These space maps benefit from a bigger
177  * block size as we expect to be writing a lot of data to them at
178  * once.
179  */
180 unsigned long zfs_log_sm_blksz = 1ULL << 17;
181 
182 /*
183  * Percentage of the overall system’s memory that ZFS allows to be
184  * used for unflushed changes (e.g. the sum of size of all the nodes
185  * in the unflushed trees).
186  *
187  * Note that this value is calculated over 1000000 for finer granularity
188  * (thus the _ppm suffix; reads as "parts per million"). As an example,
189  * the default of 1000 allows 0.1% of memory to be used.
190  */
191 unsigned long zfs_unflushed_max_mem_ppm = 1000;
192 
193 /*
194  * Specific hard-limit in memory that ZFS allows to be used for
195  * unflushed changes.
196  */
197 unsigned long zfs_unflushed_max_mem_amt = 1ULL << 30;
198 
199 /*
200  * The following tunable determines the number of blocks that can be used for
201  * the log space maps. It is expressed as a percentage of the total number of
202  * metaslabs in the pool (i.e. the default of 400 means that the number of log
203  * blocks is capped at 4 times the number of metaslabs).
204  *
205  * This value exists to tune our flushing algorithm, with higher values
206  * flushing metaslabs less often (doing less I/Os) per TXG versus lower values
207  * flushing metaslabs more aggressively with the upside of saving overheads
208  * when loading the pool. Another factor in this tradeoff is that flushing
209  * less often can potentially lead to better utilization of the metaslab space
210  * map's block size as we accumulate more changes per flush.
211  *
212  * Given that this tunable indirectly controls the flush rate (metaslabs
213  * flushed per txg) and that's why making it a percentage in terms of the
214  * number of metaslabs in the pool makes sense here.
215  *
216  * As a rule of thumb we default this tunable to 400% based on the following:
217  *
218  * 1] Assuming a constant flush rate and a constant incoming rate of log blocks
219  *    it is reasonable to expect that the amount of obsolete entries changes
220  *    linearly from txg to txg (e.g. the oldest log should have the most
221  *    obsolete entries, and the most recent one the least). With this we could
222  *    say that, at any given time, about half of the entries in the whole space
223  *    map log are obsolete. Thus for every two entries for a metaslab in the
224  *    log space map, only one of them is valid and actually makes it to the
225  *    metaslab's space map.
226  *    [factor of 2]
227  * 2] Each entry in the log space map is guaranteed to be two words while
228  *    entries in metaslab space maps are generally single-word.
229  *    [an extra factor of 2 - 400% overall]
230  * 3] Even if [1] and [2] are slightly less than 2 each, we haven't taken into
231  *    account any consolidation of segments from the log space map to the
232  *    unflushed range trees nor their history (e.g. a segment being allocated,
233  *    then freed, then allocated again means 3 log space map entries but 0
234  *    metaslab space map entries). Depending on the workload, we've seen ~1.8
235  *    non-obsolete log space map entries per metaslab entry, for a total of
236  *    ~600%. Since most of these estimates though are workload dependent, we
237  *    default on 400% to be conservative.
238  *
239  *    Thus we could say that even in the worst
240  *    case of [1] and [2], the factor should end up being 4.
241  *
242  * That said, regardless of the number of metaslabs in the pool we need to
243  * provide upper and lower bounds for the log block limit.
244  * [see zfs_unflushed_log_block_{min,max}]
245  */
246 unsigned long zfs_unflushed_log_block_pct = 400;
247 
248 /*
249  * If the number of metaslabs is small and our incoming rate is high, we could
250  * get into a situation that we are flushing all our metaslabs every TXG. Thus
251  * we always allow at least this many log blocks.
252  */
253 unsigned long zfs_unflushed_log_block_min = 1000;
254 
255 /*
256  * If the log becomes too big, the import time of the pool can take a hit in
257  * terms of performance. Thus we have a hard limit in the size of the log in
258  * terms of blocks.
259  */
260 unsigned long zfs_unflushed_log_block_max = (1ULL << 18);
261 
262 /*
263  * Max # of rows allowed for the log_summary. The tradeoff here is accuracy and
264  * stability of the flushing algorithm (longer summary) vs its runtime overhead
265  * (smaller summary is faster to traverse).
266  */
267 unsigned long zfs_max_logsm_summary_length = 10;
268 
269 /*
270  * Tunable that sets the lower bound on the metaslabs to flush every TXG.
271  *
272  * Setting this to 0 has no effect since if the pool is idle we won't even be
273  * creating log space maps and therefore we won't be flushing. On the other
274  * hand if the pool has any incoming workload our block heuristic will start
275  * flushing metaslabs anyway.
276  *
277  * The point of this tunable is to be used in extreme cases where we really
278  * want to flush more metaslabs than our adaptable heuristic plans to flush.
279  */
280 unsigned long zfs_min_metaslabs_to_flush = 1;
281 
282 /*
283  * Tunable that specifies how far in the past do we want to look when trying to
284  * estimate the incoming log blocks for the current TXG.
285  *
286  * Setting this too high may not only increase runtime but also minimize the
287  * effect of the incoming rates from the most recent TXGs as we take the
288  * average over all the blocks that we walk
289  * [see spa_estimate_incoming_log_blocks].
290  */
291 unsigned long zfs_max_log_walking = 5;
292 
293 /*
294  * This tunable exists solely for testing purposes. It ensures that the log
295  * spacemaps are not flushed and destroyed during export in order for the
296  * relevant log spacemap import code paths to be tested (effectively simulating
297  * a crash).
298  */
299 int zfs_keep_log_spacemaps_at_export = 0;
300 
301 static uint64_t
spa_estimate_incoming_log_blocks(spa_t * spa)302 spa_estimate_incoming_log_blocks(spa_t *spa)
303 {
304 	ASSERT3U(spa_sync_pass(spa), ==, 1);
305 	uint64_t steps = 0, sum = 0;
306 
307 	for (spa_log_sm_t *sls = avl_last(&spa->spa_sm_logs_by_txg);
308 	    sls != NULL && steps < zfs_max_log_walking;
309 	    sls = AVL_PREV(&spa->spa_sm_logs_by_txg, sls)) {
310 		if (sls->sls_txg == spa_syncing_txg(spa)) {
311 			/*
312 			 * skip the log created in this TXG as this would
313 			 * make our estimations inaccurate.
314 			 */
315 			continue;
316 		}
317 		sum += sls->sls_nblocks;
318 		steps++;
319 	}
320 	return ((steps > 0) ? DIV_ROUND_UP(sum, steps) : 0);
321 }
322 
323 uint64_t
spa_log_sm_blocklimit(spa_t * spa)324 spa_log_sm_blocklimit(spa_t *spa)
325 {
326 	return (spa->spa_unflushed_stats.sus_blocklimit);
327 }
328 
329 void
spa_log_sm_set_blocklimit(spa_t * spa)330 spa_log_sm_set_blocklimit(spa_t *spa)
331 {
332 	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
333 		ASSERT0(spa_log_sm_blocklimit(spa));
334 		return;
335 	}
336 
337 	uint64_t calculated_limit =
338 	    (spa_total_metaslabs(spa) * zfs_unflushed_log_block_pct) / 100;
339 	spa->spa_unflushed_stats.sus_blocklimit = MIN(MAX(calculated_limit,
340 	    zfs_unflushed_log_block_min), zfs_unflushed_log_block_max);
341 }
342 
343 uint64_t
spa_log_sm_nblocks(spa_t * spa)344 spa_log_sm_nblocks(spa_t *spa)
345 {
346 	return (spa->spa_unflushed_stats.sus_nblocks);
347 }
348 
349 /*
350  * Ensure that the in-memory log space map structures and the summary
351  * have the same block and metaslab counts.
352  */
353 static void
spa_log_summary_verify_counts(spa_t * spa)354 spa_log_summary_verify_counts(spa_t *spa)
355 {
356 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
357 
358 	if ((zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) == 0)
359 		return;
360 
361 	uint64_t ms_in_avl = avl_numnodes(&spa->spa_metaslabs_by_flushed);
362 
363 	uint64_t ms_in_summary = 0, blk_in_summary = 0;
364 	for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
365 	    e; e = list_next(&spa->spa_log_summary, e)) {
366 		ms_in_summary += e->lse_mscount;
367 		blk_in_summary += e->lse_blkcount;
368 	}
369 
370 	uint64_t ms_in_logs = 0, blk_in_logs = 0;
371 	for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
372 	    sls; sls = AVL_NEXT(&spa->spa_sm_logs_by_txg, sls)) {
373 		ms_in_logs += sls->sls_mscount;
374 		blk_in_logs += sls->sls_nblocks;
375 	}
376 
377 	VERIFY3U(ms_in_logs, ==, ms_in_summary);
378 	VERIFY3U(ms_in_logs, ==, ms_in_avl);
379 	VERIFY3U(blk_in_logs, ==, blk_in_summary);
380 	VERIFY3U(blk_in_logs, ==, spa_log_sm_nblocks(spa));
381 }
382 
383 static boolean_t
summary_entry_is_full(spa_t * spa,log_summary_entry_t * e)384 summary_entry_is_full(spa_t *spa, log_summary_entry_t *e)
385 {
386 	uint64_t blocks_per_row = MAX(1,
387 	    DIV_ROUND_UP(spa_log_sm_blocklimit(spa),
388 	    zfs_max_logsm_summary_length));
389 
390 	return (blocks_per_row <= e->lse_blkcount);
391 }
392 
393 /*
394  * Update the log summary information to reflect the fact that a metaslab
395  * was flushed or destroyed (e.g due to device removal or pool export/destroy).
396  *
397  * We typically flush the oldest flushed metaslab so the first (and oldest)
398  * entry of the summary is updated. However if that metaslab is getting loaded
399  * we may flush the second oldest one which may be part of an entry later in
400  * the summary. Moreover, if we call into this function from metaslab_fini()
401  * the metaslabs probably won't be ordered by ms_unflushed_txg. Thus we ask
402  * for a txg as an argument so we can locate the appropriate summary entry for
403  * the metaslab.
404  */
405 void
spa_log_summary_decrement_mscount(spa_t * spa,uint64_t txg)406 spa_log_summary_decrement_mscount(spa_t *spa, uint64_t txg)
407 {
408 	/*
409 	 * We don't track summary data for read-only pools and this function
410 	 * can be called from metaslab_fini(). In that case return immediately.
411 	 */
412 	if (!spa_writeable(spa))
413 		return;
414 
415 	log_summary_entry_t *target = NULL;
416 	for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
417 	    e != NULL; e = list_next(&spa->spa_log_summary, e)) {
418 		if (e->lse_start > txg)
419 			break;
420 		target = e;
421 	}
422 
423 	if (target == NULL || target->lse_mscount == 0) {
424 		/*
425 		 * We didn't find a summary entry for this metaslab. We must be
426 		 * at the teardown of a spa_load() attempt that got an error
427 		 * while reading the log space maps.
428 		 */
429 		VERIFY3S(spa_load_state(spa), ==, SPA_LOAD_ERROR);
430 		return;
431 	}
432 
433 	target->lse_mscount--;
434 }
435 
436 /*
437  * Update the log summary information to reflect the fact that we destroyed
438  * old log space maps. Since we can only destroy the oldest log space maps,
439  * we decrement the block count of the oldest summary entry and potentially
440  * destroy it when that count hits 0.
441  *
442  * This function is called after a metaslab is flushed and typically that
443  * metaslab is the oldest flushed, which means that this function will
444  * typically decrement the block count of the first entry of the summary and
445  * potentially free it if the block count gets to zero (its metaslab count
446  * should be zero too at that point).
447  *
448  * There are certain scenarios though that don't work exactly like that so we
449  * need to account for them:
450  *
451  * Scenario [1]: It is possible that after we flushed the oldest flushed
452  * metaslab and we destroyed the oldest log space map, more recent logs had 0
453  * metaslabs pointing to them so we got rid of them too. This can happen due
454  * to metaslabs being destroyed through device removal, or because the oldest
455  * flushed metaslab was loading but we kept flushing more recently flushed
456  * metaslabs due to the memory pressure of unflushed changes. Because of that,
457  * we always iterate from the beginning of the summary and if blocks_gone is
458  * bigger than the block_count of the current entry we free that entry (we
459  * expect its metaslab count to be zero), we decrement blocks_gone and on to
460  * the next entry repeating this procedure until blocks_gone gets decremented
461  * to 0. Doing this also works for the typical case mentioned above.
462  *
463  * Scenario [2]: The oldest flushed metaslab isn't necessarily accounted by
464  * the first (and oldest) entry in the summary. If the first few entries of
465  * the summary were only accounting metaslabs from a device that was just
466  * removed, then the current oldest flushed metaslab could be accounted by an
467  * entry somewhere in the middle of the summary. Moreover flushing that
468  * metaslab will destroy all the log space maps older than its ms_unflushed_txg
469  * because they became obsolete after the removal. Thus, iterating as we did
470  * for scenario [1] works out for this case too.
471  *
472  * Scenario [3]: At times we decide to flush all the metaslabs in the pool
473  * in one TXG (either because we are exporting the pool or because our flushing
474  * heuristics decided to do so). When that happens all the log space maps get
475  * destroyed except the one created for the current TXG which doesn't have
476  * any log blocks yet. As log space maps get destroyed with every metaslab that
477  * we flush, entries in the summary are also destroyed. This brings a weird
478  * corner-case when we flush the last metaslab and the log space map of the
479  * current TXG is in the same summary entry with other log space maps that
480  * are older. When that happens we are eventually left with this one last
481  * summary entry whose blocks are gone (blocks_gone equals the entry's block
482  * count) but its metaslab count is non-zero (because it accounts all the
483  * metaslabs in the pool as they all got flushed). Under this scenario we can't
484  * free this last summary entry as it's referencing all the metaslabs in the
485  * pool and its block count will get incremented at the end of this sync (when
486  * we close the syncing log space map). Thus we just decrement its current
487  * block count and leave it alone. In the case that the pool gets exported,
488  * its metaslab count will be decremented over time as we call metaslab_fini()
489  * for all the metaslabs in the pool and the entry will be freed at
490  * spa_unload_log_sm_metadata().
491  */
492 void
spa_log_summary_decrement_blkcount(spa_t * spa,uint64_t blocks_gone)493 spa_log_summary_decrement_blkcount(spa_t *spa, uint64_t blocks_gone)
494 {
495 	for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
496 	    e != NULL; e = list_head(&spa->spa_log_summary)) {
497 		if (e->lse_blkcount > blocks_gone) {
498 			/*
499 			 * Assert that we stopped at an entry that is not
500 			 * obsolete.
501 			 */
502 			ASSERT(e->lse_mscount != 0);
503 
504 			e->lse_blkcount -= blocks_gone;
505 			blocks_gone = 0;
506 			break;
507 		} else if (e->lse_mscount == 0) {
508 			/* remove obsolete entry */
509 			blocks_gone -= e->lse_blkcount;
510 			list_remove(&spa->spa_log_summary, e);
511 			kmem_free(e, sizeof (log_summary_entry_t));
512 		} else {
513 			/* Verify that this is scenario [3] mentioned above. */
514 			VERIFY3U(blocks_gone, ==, e->lse_blkcount);
515 
516 			/*
517 			 * Assert that this is scenario [3] further by ensuring
518 			 * that this is the only entry in the summary.
519 			 */
520 			VERIFY3P(e, ==, list_tail(&spa->spa_log_summary));
521 			ASSERT3P(e, ==, list_head(&spa->spa_log_summary));
522 
523 			blocks_gone = e->lse_blkcount = 0;
524 			break;
525 		}
526 	}
527 
528 	/*
529 	 * Ensure that there is no way we are trying to remove more blocks
530 	 * than the # of blocks in the summary.
531 	 */
532 	ASSERT0(blocks_gone);
533 }
534 
535 void
spa_log_sm_decrement_mscount(spa_t * spa,uint64_t txg)536 spa_log_sm_decrement_mscount(spa_t *spa, uint64_t txg)
537 {
538 	spa_log_sm_t target = { .sls_txg = txg };
539 	spa_log_sm_t *sls = avl_find(&spa->spa_sm_logs_by_txg,
540 	    &target, NULL);
541 
542 	if (sls == NULL) {
543 		/*
544 		 * We must be at the teardown of a spa_load() attempt that
545 		 * got an error while reading the log space maps.
546 		 */
547 		VERIFY3S(spa_load_state(spa), ==, SPA_LOAD_ERROR);
548 		return;
549 	}
550 
551 	ASSERT(sls->sls_mscount > 0);
552 	sls->sls_mscount--;
553 }
554 
555 void
spa_log_sm_increment_current_mscount(spa_t * spa)556 spa_log_sm_increment_current_mscount(spa_t *spa)
557 {
558 	spa_log_sm_t *last_sls = avl_last(&spa->spa_sm_logs_by_txg);
559 
560 	ASSERT3U(last_sls->sls_txg, ==, spa_syncing_txg(spa));
561 	last_sls->sls_mscount++;
562 }
563 
564 static void
summary_add_data(spa_t * spa,uint64_t txg,uint64_t metaslabs_flushed,uint64_t nblocks)565 summary_add_data(spa_t *spa, uint64_t txg, uint64_t metaslabs_flushed,
566     uint64_t nblocks)
567 {
568 	log_summary_entry_t *e = list_tail(&spa->spa_log_summary);
569 
570 	if (e == NULL || summary_entry_is_full(spa, e)) {
571 		e = kmem_zalloc(sizeof (log_summary_entry_t), KM_SLEEP);
572 		e->lse_start = txg;
573 		list_insert_tail(&spa->spa_log_summary, e);
574 	}
575 
576 	ASSERT3U(e->lse_start, <=, txg);
577 	e->lse_mscount += metaslabs_flushed;
578 	e->lse_blkcount += nblocks;
579 }
580 
581 static void
spa_log_summary_add_incoming_blocks(spa_t * spa,uint64_t nblocks)582 spa_log_summary_add_incoming_blocks(spa_t *spa, uint64_t nblocks)
583 {
584 	summary_add_data(spa, spa_syncing_txg(spa), 0, nblocks);
585 }
586 
587 void
spa_log_summary_add_flushed_metaslab(spa_t * spa)588 spa_log_summary_add_flushed_metaslab(spa_t *spa)
589 {
590 	summary_add_data(spa, spa_syncing_txg(spa), 1, 0);
591 }
592 
593 /*
594  * This function attempts to estimate how many metaslabs should
595  * we flush to satisfy our block heuristic for the log spacemap
596  * for the upcoming TXGs.
597  *
598  * Specifically, it first tries to estimate the number of incoming
599  * blocks in this TXG. Then by projecting that incoming rate to
600  * future TXGs and using the log summary, it figures out how many
601  * flushes we would need to do for future TXGs individually to
602  * stay below our block limit and returns the maximum number of
603  * flushes from those estimates.
604  */
605 static uint64_t
spa_estimate_metaslabs_to_flush(spa_t * spa)606 spa_estimate_metaslabs_to_flush(spa_t *spa)
607 {
608 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
609 	ASSERT3U(spa_sync_pass(spa), ==, 1);
610 	ASSERT(spa_log_sm_blocklimit(spa) != 0);
611 
612 	/*
613 	 * This variable contains the incoming rate that will be projected
614 	 * and used for our flushing estimates in the future.
615 	 */
616 	uint64_t incoming = spa_estimate_incoming_log_blocks(spa);
617 
618 	/*
619 	 * At any point in time this variable tells us how many
620 	 * TXGs in the future we are so we can make our estimations.
621 	 */
622 	uint64_t txgs_in_future = 1;
623 
624 	/*
625 	 * This variable tells us how much room do we have until we hit
626 	 * our limit. When it goes negative, it means that we've exceeded
627 	 * our limit and we need to flush.
628 	 *
629 	 * Note that since we start at the first TXG in the future (i.e.
630 	 * txgs_in_future starts from 1) we already decrement this
631 	 * variable by the incoming rate.
632 	 */
633 	int64_t available_blocks =
634 	    spa_log_sm_blocklimit(spa) - spa_log_sm_nblocks(spa) - incoming;
635 
636 	/*
637 	 * This variable tells us the total number of flushes needed to
638 	 * keep the log size within the limit when we reach txgs_in_future.
639 	 */
640 	uint64_t total_flushes = 0;
641 
642 	/* Holds the current maximum of our estimates so far. */
643 	uint64_t max_flushes_pertxg =
644 	    MIN(avl_numnodes(&spa->spa_metaslabs_by_flushed),
645 	    zfs_min_metaslabs_to_flush);
646 
647 	/*
648 	 * For our estimations we only look as far in the future
649 	 * as the summary allows us.
650 	 */
651 	for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
652 	    e; e = list_next(&spa->spa_log_summary, e)) {
653 
654 		/*
655 		 * If there is still room before we exceed our limit
656 		 * then keep skipping TXGs accumulating more blocks
657 		 * based on the incoming rate until we exceed it.
658 		 */
659 		if (available_blocks >= 0) {
660 			uint64_t skip_txgs = (available_blocks / incoming) + 1;
661 			available_blocks -= (skip_txgs * incoming);
662 			txgs_in_future += skip_txgs;
663 			ASSERT3S(available_blocks, >=, -incoming);
664 		}
665 
666 		/*
667 		 * At this point we're far enough into the future where
668 		 * the limit was just exceeded and we flush metaslabs
669 		 * based on the current entry in the summary, updating
670 		 * our available_blocks.
671 		 */
672 		ASSERT3S(available_blocks, <, 0);
673 		available_blocks += e->lse_blkcount;
674 		total_flushes += e->lse_mscount;
675 
676 		/*
677 		 * Keep the running maximum of the total_flushes that
678 		 * we've done so far over the number of TXGs in the
679 		 * future that we are. The idea here is to estimate
680 		 * the average number of flushes that we should do
681 		 * every TXG so that when we are that many TXGs in the
682 		 * future we stay under the limit.
683 		 */
684 		max_flushes_pertxg = MAX(max_flushes_pertxg,
685 		    DIV_ROUND_UP(total_flushes, txgs_in_future));
686 		ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=,
687 		    max_flushes_pertxg);
688 	}
689 	return (max_flushes_pertxg);
690 }
691 
692 uint64_t
spa_log_sm_memused(spa_t * spa)693 spa_log_sm_memused(spa_t *spa)
694 {
695 	return (spa->spa_unflushed_stats.sus_memused);
696 }
697 
698 static boolean_t
spa_log_exceeds_memlimit(spa_t * spa)699 spa_log_exceeds_memlimit(spa_t *spa)
700 {
701 	if (spa_log_sm_memused(spa) > zfs_unflushed_max_mem_amt)
702 		return (B_TRUE);
703 
704 	uint64_t system_mem_allowed = ((physmem * PAGESIZE) *
705 	    zfs_unflushed_max_mem_ppm) / 1000000;
706 	if (spa_log_sm_memused(spa) > system_mem_allowed)
707 		return (B_TRUE);
708 
709 	return (B_FALSE);
710 }
711 
712 boolean_t
spa_flush_all_logs_requested(spa_t * spa)713 spa_flush_all_logs_requested(spa_t *spa)
714 {
715 	return (spa->spa_log_flushall_txg != 0);
716 }
717 
718 void
spa_flush_metaslabs(spa_t * spa,dmu_tx_t * tx)719 spa_flush_metaslabs(spa_t *spa, dmu_tx_t *tx)
720 {
721 	uint64_t txg = dmu_tx_get_txg(tx);
722 
723 	if (spa_sync_pass(spa) != 1)
724 		return;
725 
726 	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
727 		return;
728 
729 	/*
730 	 * If we don't have any metaslabs with unflushed changes
731 	 * return immediately.
732 	 */
733 	if (avl_numnodes(&spa->spa_metaslabs_by_flushed) == 0)
734 		return;
735 
736 	/*
737 	 * During SPA export we leave a few empty TXGs to go by [see
738 	 * spa_final_dirty_txg() to understand why]. For this specific
739 	 * case, it is important to not flush any metaslabs as that
740 	 * would dirty this TXG.
741 	 *
742 	 * That said, during one of these dirty TXGs that is less or
743 	 * equal to spa_final_dirty(), spa_unload() will request that
744 	 * we try to flush all the metaslabs for that TXG before
745 	 * exporting the pool, thus we ensure that we didn't get a
746 	 * request of flushing everything before we attempt to return
747 	 * immediately.
748 	 */
749 	if (spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
750 	    !dmu_objset_is_dirty(spa_meta_objset(spa), txg) &&
751 	    !spa_flush_all_logs_requested(spa))
752 		return;
753 
754 	/*
755 	 * We need to generate a log space map before flushing because this
756 	 * will set up the in-memory data (i.e. node in spa_sm_logs_by_txg)
757 	 * for this TXG's flushed metaslab count (aka sls_mscount which is
758 	 * manipulated in many ways down the metaslab_flush() codepath).
759 	 *
760 	 * That is not to say that we may generate a log space map when we
761 	 * don't need it. If we are flushing metaslabs, that means that we
762 	 * were going to write changes to disk anyway, so even if we were
763 	 * not flushing, a log space map would have been created anyway in
764 	 * metaslab_sync().
765 	 */
766 	spa_generate_syncing_log_sm(spa, tx);
767 
768 	/*
769 	 * This variable tells us how many metaslabs we want to flush based
770 	 * on the block-heuristic of our flushing algorithm (see block comment
771 	 * of log space map feature). We also decrement this as we flush
772 	 * metaslabs and attempt to destroy old log space maps.
773 	 */
774 	uint64_t want_to_flush;
775 	if (spa_flush_all_logs_requested(spa)) {
776 		ASSERT3S(spa_state(spa), ==, POOL_STATE_EXPORTED);
777 		want_to_flush = avl_numnodes(&spa->spa_metaslabs_by_flushed);
778 	} else {
779 		want_to_flush = spa_estimate_metaslabs_to_flush(spa);
780 	}
781 
782 	ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=,
783 	    want_to_flush);
784 
785 	/* Used purely for verification purposes */
786 	uint64_t visited = 0;
787 
788 	/*
789 	 * Ideally we would only iterate through spa_metaslabs_by_flushed
790 	 * using only one variable (curr). We can't do that because
791 	 * metaslab_flush() mutates position of curr in the AVL when
792 	 * it flushes that metaslab by moving it to the end of the tree.
793 	 * Thus we always keep track of the original next node of the
794 	 * current node (curr) in another variable (next).
795 	 */
796 	metaslab_t *next = NULL;
797 	for (metaslab_t *curr = avl_first(&spa->spa_metaslabs_by_flushed);
798 	    curr != NULL; curr = next) {
799 		next = AVL_NEXT(&spa->spa_metaslabs_by_flushed, curr);
800 
801 		/*
802 		 * If this metaslab has been flushed this txg then we've done
803 		 * a full circle over the metaslabs.
804 		 */
805 		if (metaslab_unflushed_txg(curr) == txg)
806 			break;
807 
808 		/*
809 		 * If we are done flushing for the block heuristic and the
810 		 * unflushed changes don't exceed the memory limit just stop.
811 		 */
812 		if (want_to_flush == 0 && !spa_log_exceeds_memlimit(spa))
813 			break;
814 
815 		mutex_enter(&curr->ms_sync_lock);
816 		mutex_enter(&curr->ms_lock);
817 		boolean_t flushed = metaslab_flush(curr, tx);
818 		mutex_exit(&curr->ms_lock);
819 		mutex_exit(&curr->ms_sync_lock);
820 
821 		/*
822 		 * If we failed to flush a metaslab (because it was loading),
823 		 * then we are done with the block heuristic as it's not
824 		 * possible to destroy any log space maps once you've skipped
825 		 * a metaslab. In that case we just set our counter to 0 but
826 		 * we continue looping in case there is still memory pressure
827 		 * due to unflushed changes. Note that, flushing a metaslab
828 		 * that is not the oldest flushed in the pool, will never
829 		 * destroy any log space maps [see spa_cleanup_old_sm_logs()].
830 		 */
831 		if (!flushed) {
832 			want_to_flush = 0;
833 		} else if (want_to_flush > 0) {
834 			want_to_flush--;
835 		}
836 
837 		visited++;
838 	}
839 	ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=, visited);
840 }
841 
842 /*
843  * Close the log space map for this TXG and update the block counts
844  * for the the log's in-memory structure and the summary.
845  */
846 void
spa_sync_close_syncing_log_sm(spa_t * spa)847 spa_sync_close_syncing_log_sm(spa_t *spa)
848 {
849 	if (spa_syncing_log_sm(spa) == NULL)
850 		return;
851 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
852 
853 	spa_log_sm_t *sls = avl_last(&spa->spa_sm_logs_by_txg);
854 	ASSERT3U(sls->sls_txg, ==, spa_syncing_txg(spa));
855 
856 	sls->sls_nblocks = space_map_nblocks(spa_syncing_log_sm(spa));
857 	spa->spa_unflushed_stats.sus_nblocks += sls->sls_nblocks;
858 
859 	/*
860 	 * Note that we can't assert that sls_mscount is not 0,
861 	 * because there is the case where the first metaslab
862 	 * in spa_metaslabs_by_flushed is loading and we were
863 	 * not able to flush any metaslabs the current TXG.
864 	 */
865 	ASSERT(sls->sls_nblocks != 0);
866 
867 	spa_log_summary_add_incoming_blocks(spa, sls->sls_nblocks);
868 	spa_log_summary_verify_counts(spa);
869 
870 	space_map_close(spa->spa_syncing_log_sm);
871 	spa->spa_syncing_log_sm = NULL;
872 
873 	/*
874 	 * At this point we tried to flush as many metaslabs as we
875 	 * can as the pool is getting exported. Reset the "flush all"
876 	 * so the last few TXGs before closing the pool can be empty
877 	 * (e.g. not dirty).
878 	 */
879 	if (spa_flush_all_logs_requested(spa)) {
880 		ASSERT3S(spa_state(spa), ==, POOL_STATE_EXPORTED);
881 		spa->spa_log_flushall_txg = 0;
882 	}
883 }
884 
885 void
spa_cleanup_old_sm_logs(spa_t * spa,dmu_tx_t * tx)886 spa_cleanup_old_sm_logs(spa_t *spa, dmu_tx_t *tx)
887 {
888 	objset_t *mos = spa_meta_objset(spa);
889 
890 	uint64_t spacemap_zap;
891 	int error = zap_lookup(mos, DMU_POOL_DIRECTORY_OBJECT,
892 	    DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
893 	if (error == ENOENT) {
894 		ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
895 		return;
896 	}
897 	VERIFY0(error);
898 
899 	metaslab_t *oldest = avl_first(&spa->spa_metaslabs_by_flushed);
900 	uint64_t oldest_flushed_txg = metaslab_unflushed_txg(oldest);
901 
902 	/* Free all log space maps older than the oldest_flushed_txg. */
903 	for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
904 	    sls && sls->sls_txg < oldest_flushed_txg;
905 	    sls = avl_first(&spa->spa_sm_logs_by_txg)) {
906 		ASSERT0(sls->sls_mscount);
907 		avl_remove(&spa->spa_sm_logs_by_txg, sls);
908 		space_map_free_obj(mos, sls->sls_sm_obj, tx);
909 		VERIFY0(zap_remove_int(mos, spacemap_zap, sls->sls_txg, tx));
910 		spa->spa_unflushed_stats.sus_nblocks -= sls->sls_nblocks;
911 		kmem_free(sls, sizeof (spa_log_sm_t));
912 	}
913 }
914 
915 static spa_log_sm_t *
spa_log_sm_alloc(uint64_t sm_obj,uint64_t txg)916 spa_log_sm_alloc(uint64_t sm_obj, uint64_t txg)
917 {
918 	spa_log_sm_t *sls = kmem_zalloc(sizeof (*sls), KM_SLEEP);
919 
920 	sls->sls_sm_obj = sm_obj;
921 	sls->sls_txg = txg;
922 	return (sls);
923 }
924 
925 void
spa_generate_syncing_log_sm(spa_t * spa,dmu_tx_t * tx)926 spa_generate_syncing_log_sm(spa_t *spa, dmu_tx_t *tx)
927 {
928 	uint64_t txg = dmu_tx_get_txg(tx);
929 	objset_t *mos = spa_meta_objset(spa);
930 
931 	if (spa_syncing_log_sm(spa) != NULL)
932 		return;
933 
934 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP))
935 		return;
936 
937 	uint64_t spacemap_zap;
938 	int error = zap_lookup(mos, DMU_POOL_DIRECTORY_OBJECT,
939 	    DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
940 	if (error == ENOENT) {
941 		ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
942 
943 		error = 0;
944 		spacemap_zap = zap_create(mos,
945 		    DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
946 		VERIFY0(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
947 		    DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1,
948 		    &spacemap_zap, tx));
949 		spa_feature_incr(spa, SPA_FEATURE_LOG_SPACEMAP, tx);
950 	}
951 	VERIFY0(error);
952 
953 	uint64_t sm_obj;
954 	ASSERT3U(zap_lookup_int_key(mos, spacemap_zap, txg, &sm_obj),
955 	    ==, ENOENT);
956 	sm_obj = space_map_alloc(mos, zfs_log_sm_blksz, tx);
957 	VERIFY0(zap_add_int_key(mos, spacemap_zap, txg, sm_obj, tx));
958 	avl_add(&spa->spa_sm_logs_by_txg, spa_log_sm_alloc(sm_obj, txg));
959 
960 	/*
961 	 * We pass UINT64_MAX as the space map's representation size
962 	 * and SPA_MINBLOCKSHIFT as the shift, to make the space map
963 	 * accept any sorts of segments since there's no real advantage
964 	 * to being more restrictive (given that we're already going
965 	 * to be using 2-word entries).
966 	 */
967 	VERIFY0(space_map_open(&spa->spa_syncing_log_sm, mos, sm_obj,
968 	    0, UINT64_MAX, SPA_MINBLOCKSHIFT));
969 
970 	/*
971 	 * If the log space map feature was just enabled, the blocklimit
972 	 * has not yet been set.
973 	 */
974 	if (spa_log_sm_blocklimit(spa) == 0)
975 		spa_log_sm_set_blocklimit(spa);
976 }
977 
978 /*
979  * Find all the log space maps stored in the space map ZAP and sort
980  * them by their TXG in spa_sm_logs_by_txg.
981  */
982 static int
spa_ld_log_sm_metadata(spa_t * spa)983 spa_ld_log_sm_metadata(spa_t *spa)
984 {
985 	int error;
986 	uint64_t spacemap_zap;
987 
988 	ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
989 
990 	error = zap_lookup(spa_meta_objset(spa), DMU_POOL_DIRECTORY_OBJECT,
991 	    DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
992 	if (error == ENOENT) {
993 		/* the space map ZAP doesn't exist yet */
994 		return (0);
995 	} else if (error != 0) {
996 		spa_load_failed(spa, "spa_ld_log_sm_metadata(): failed at "
997 		    "zap_lookup(DMU_POOL_DIRECTORY_OBJECT) [error %d]",
998 		    error);
999 		return (error);
1000 	}
1001 
1002 	zap_cursor_t zc;
1003 	zap_attribute_t za;
1004 	for (zap_cursor_init(&zc, spa_meta_objset(spa), spacemap_zap);
1005 	    (error = zap_cursor_retrieve(&zc, &za)) == 0;
1006 	    zap_cursor_advance(&zc)) {
1007 		uint64_t log_txg = zfs_strtonum(za.za_name, NULL);
1008 		spa_log_sm_t *sls =
1009 		    spa_log_sm_alloc(za.za_first_integer, log_txg);
1010 		avl_add(&spa->spa_sm_logs_by_txg, sls);
1011 	}
1012 	zap_cursor_fini(&zc);
1013 	if (error != ENOENT) {
1014 		spa_load_failed(spa, "spa_ld_log_sm_metadata(): failed at "
1015 		    "zap_cursor_retrieve(spacemap_zap) [error %d]",
1016 		    error);
1017 		return (error);
1018 	}
1019 
1020 	for (metaslab_t *m = avl_first(&spa->spa_metaslabs_by_flushed);
1021 	    m; m = AVL_NEXT(&spa->spa_metaslabs_by_flushed, m)) {
1022 		spa_log_sm_t target = { .sls_txg = metaslab_unflushed_txg(m) };
1023 		spa_log_sm_t *sls = avl_find(&spa->spa_sm_logs_by_txg,
1024 		    &target, NULL);
1025 
1026 		/*
1027 		 * At this point if sls is zero it means that a bug occurred
1028 		 * in ZFS the last time the pool was open or earlier in the
1029 		 * import code path. In general, we would have placed a
1030 		 * VERIFY() here or in this case just let the kernel panic
1031 		 * with NULL pointer dereference when incrementing sls_mscount,
1032 		 * but since this is the import code path we can be a bit more
1033 		 * lenient. Thus, for DEBUG bits we always cause a panic, while
1034 		 * in production we log the error and just fail the import.
1035 		 */
1036 		ASSERT(sls != NULL);
1037 		if (sls == NULL) {
1038 			spa_load_failed(spa, "spa_ld_log_sm_metadata(): bug "
1039 			    "encountered: could not find log spacemap for "
1040 			    "TXG %ld [error %d]",
1041 			    metaslab_unflushed_txg(m), ENOENT);
1042 			return (ENOENT);
1043 		}
1044 		sls->sls_mscount++;
1045 	}
1046 
1047 	return (0);
1048 }
1049 
1050 typedef struct spa_ld_log_sm_arg {
1051 	spa_t *slls_spa;
1052 	uint64_t slls_txg;
1053 } spa_ld_log_sm_arg_t;
1054 
1055 static int
spa_ld_log_sm_cb(space_map_entry_t * sme,void * arg)1056 spa_ld_log_sm_cb(space_map_entry_t *sme, void *arg)
1057 {
1058 	uint64_t offset = sme->sme_offset;
1059 	uint64_t size = sme->sme_run;
1060 	uint32_t vdev_id = sme->sme_vdev;
1061 	spa_ld_log_sm_arg_t *slls = arg;
1062 	spa_t *spa = slls->slls_spa;
1063 
1064 	vdev_t *vd = vdev_lookup_top(spa, vdev_id);
1065 
1066 	/*
1067 	 * If the vdev has been removed (i.e. it is indirect or a hole)
1068 	 * skip this entry. The contents of this vdev have already moved
1069 	 * elsewhere.
1070 	 */
1071 	if (!vdev_is_concrete(vd))
1072 		return (0);
1073 
1074 	metaslab_t *ms = vd->vdev_ms[offset >> vd->vdev_ms_shift];
1075 	ASSERT(!ms->ms_loaded);
1076 
1077 	/*
1078 	 * If we have already flushed entries for this TXG to this
1079 	 * metaslab's space map, then ignore it. Note that we flush
1080 	 * before processing any allocations/frees for that TXG, so
1081 	 * the metaslab's space map only has entries from *before*
1082 	 * the unflushed TXG.
1083 	 */
1084 	if (slls->slls_txg < metaslab_unflushed_txg(ms))
1085 		return (0);
1086 
1087 	switch (sme->sme_type) {
1088 	case SM_ALLOC:
1089 		range_tree_remove_xor_add_segment(offset, offset + size,
1090 		    ms->ms_unflushed_frees, ms->ms_unflushed_allocs);
1091 		break;
1092 	case SM_FREE:
1093 		range_tree_remove_xor_add_segment(offset, offset + size,
1094 		    ms->ms_unflushed_allocs, ms->ms_unflushed_frees);
1095 		break;
1096 	default:
1097 		panic("invalid maptype_t");
1098 		break;
1099 	}
1100 	return (0);
1101 }
1102 
1103 static int
spa_ld_log_sm_data(spa_t * spa)1104 spa_ld_log_sm_data(spa_t *spa)
1105 {
1106 	int error = 0;
1107 
1108 	/*
1109 	 * If we are not going to do any writes there is no need
1110 	 * to read the log space maps.
1111 	 */
1112 	if (!spa_writeable(spa))
1113 		return (0);
1114 
1115 	ASSERT0(spa->spa_unflushed_stats.sus_nblocks);
1116 	ASSERT0(spa->spa_unflushed_stats.sus_memused);
1117 
1118 	hrtime_t read_logs_starttime = gethrtime();
1119 	/* this is a no-op when we don't have space map logs */
1120 	for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
1121 	    sls; sls = AVL_NEXT(&spa->spa_sm_logs_by_txg, sls)) {
1122 		space_map_t *sm = NULL;
1123 		error = space_map_open(&sm, spa_meta_objset(spa),
1124 		    sls->sls_sm_obj, 0, UINT64_MAX, SPA_MINBLOCKSHIFT);
1125 		if (error != 0) {
1126 			spa_load_failed(spa, "spa_ld_log_sm_data(): failed at "
1127 			    "space_map_open(obj=%llu) [error %d]",
1128 			    (u_longlong_t)sls->sls_sm_obj, error);
1129 			goto out;
1130 		}
1131 
1132 		struct spa_ld_log_sm_arg vla = {
1133 			.slls_spa = spa,
1134 			.slls_txg = sls->sls_txg
1135 		};
1136 		error = space_map_iterate(sm, space_map_length(sm),
1137 		    spa_ld_log_sm_cb, &vla);
1138 		if (error != 0) {
1139 			space_map_close(sm);
1140 			spa_load_failed(spa, "spa_ld_log_sm_data(): failed "
1141 			    "at space_map_iterate(obj=%llu) [error %d]",
1142 			    (u_longlong_t)sls->sls_sm_obj, error);
1143 			goto out;
1144 		}
1145 
1146 		ASSERT0(sls->sls_nblocks);
1147 		sls->sls_nblocks = space_map_nblocks(sm);
1148 		spa->spa_unflushed_stats.sus_nblocks += sls->sls_nblocks;
1149 		summary_add_data(spa, sls->sls_txg,
1150 		    sls->sls_mscount, sls->sls_nblocks);
1151 
1152 		space_map_close(sm);
1153 	}
1154 	hrtime_t read_logs_endtime = gethrtime();
1155 	spa_load_note(spa,
1156 	    "read %llu log space maps (%llu total blocks - blksz = %llu bytes) "
1157 	    "in %lld ms", (u_longlong_t)avl_numnodes(&spa->spa_sm_logs_by_txg),
1158 	    (u_longlong_t)spa_log_sm_nblocks(spa),
1159 	    (u_longlong_t)zfs_log_sm_blksz,
1160 	    (longlong_t)((read_logs_endtime - read_logs_starttime) / 1000000));
1161 
1162 out:
1163 	/*
1164 	 * Now that the metaslabs contain their unflushed changes:
1165 	 * [1] recalculate their actual allocated space
1166 	 * [2] recalculate their weights
1167 	 * [3] sum up the memory usage of their unflushed range trees
1168 	 * [4] optionally load them, if debug_load is set
1169 	 *
1170 	 * Note that even in the case where we get here because of an
1171 	 * error (e.g. error != 0), we still want to update the fields
1172 	 * below in order to have a proper teardown in spa_unload().
1173 	 */
1174 	for (metaslab_t *m = avl_first(&spa->spa_metaslabs_by_flushed);
1175 	    m != NULL; m = AVL_NEXT(&spa->spa_metaslabs_by_flushed, m)) {
1176 		mutex_enter(&m->ms_lock);
1177 		m->ms_allocated_space = space_map_allocated(m->ms_sm) +
1178 		    range_tree_space(m->ms_unflushed_allocs) -
1179 		    range_tree_space(m->ms_unflushed_frees);
1180 
1181 		vdev_t *vd = m->ms_group->mg_vd;
1182 		metaslab_space_update(vd, m->ms_group->mg_class,
1183 		    range_tree_space(m->ms_unflushed_allocs), 0, 0);
1184 		metaslab_space_update(vd, m->ms_group->mg_class,
1185 		    -range_tree_space(m->ms_unflushed_frees), 0, 0);
1186 
1187 		ASSERT0(m->ms_weight & METASLAB_ACTIVE_MASK);
1188 		metaslab_recalculate_weight_and_sort(m);
1189 
1190 		spa->spa_unflushed_stats.sus_memused +=
1191 		    metaslab_unflushed_changes_memused(m);
1192 
1193 		if (metaslab_debug_load && m->ms_sm != NULL) {
1194 			VERIFY0(metaslab_load(m));
1195 			metaslab_set_selected_txg(m, 0);
1196 		}
1197 		mutex_exit(&m->ms_lock);
1198 	}
1199 
1200 	return (error);
1201 }
1202 
1203 static int
spa_ld_unflushed_txgs(vdev_t * vd)1204 spa_ld_unflushed_txgs(vdev_t *vd)
1205 {
1206 	spa_t *spa = vd->vdev_spa;
1207 	objset_t *mos = spa_meta_objset(spa);
1208 
1209 	if (vd->vdev_top_zap == 0)
1210 		return (0);
1211 
1212 	uint64_t object = 0;
1213 	int error = zap_lookup(mos, vd->vdev_top_zap,
1214 	    VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS,
1215 	    sizeof (uint64_t), 1, &object);
1216 	if (error == ENOENT)
1217 		return (0);
1218 	else if (error != 0) {
1219 		spa_load_failed(spa, "spa_ld_unflushed_txgs(): failed at "
1220 		    "zap_lookup(vdev_top_zap=%llu) [error %d]",
1221 		    (u_longlong_t)vd->vdev_top_zap, error);
1222 		return (error);
1223 	}
1224 
1225 	for (uint64_t m = 0; m < vd->vdev_ms_count; m++) {
1226 		metaslab_t *ms = vd->vdev_ms[m];
1227 		ASSERT(ms != NULL);
1228 
1229 		metaslab_unflushed_phys_t entry;
1230 		uint64_t entry_size = sizeof (entry);
1231 		uint64_t entry_offset = ms->ms_id * entry_size;
1232 
1233 		error = dmu_read(mos, object,
1234 		    entry_offset, entry_size, &entry, 0);
1235 		if (error != 0) {
1236 			spa_load_failed(spa, "spa_ld_unflushed_txgs(): "
1237 			    "failed at dmu_read(obj=%llu) [error %d]",
1238 			    (u_longlong_t)object, error);
1239 			return (error);
1240 		}
1241 
1242 		ms->ms_unflushed_txg = entry.msp_unflushed_txg;
1243 		if (ms->ms_unflushed_txg != 0) {
1244 			mutex_enter(&spa->spa_flushed_ms_lock);
1245 			avl_add(&spa->spa_metaslabs_by_flushed, ms);
1246 			mutex_exit(&spa->spa_flushed_ms_lock);
1247 		}
1248 	}
1249 	return (0);
1250 }
1251 
1252 /*
1253  * Read all the log space map entries into their respective
1254  * metaslab unflushed trees and keep them sorted by TXG in the
1255  * SPA's metadata. In addition, setup all the metadata for the
1256  * memory and the block heuristics.
1257  */
1258 int
spa_ld_log_spacemaps(spa_t * spa)1259 spa_ld_log_spacemaps(spa_t *spa)
1260 {
1261 	int error;
1262 
1263 	spa_log_sm_set_blocklimit(spa);
1264 
1265 	for (uint64_t c = 0; c < spa->spa_root_vdev->vdev_children; c++) {
1266 		vdev_t *vd = spa->spa_root_vdev->vdev_child[c];
1267 		error = spa_ld_unflushed_txgs(vd);
1268 		if (error != 0)
1269 			return (error);
1270 	}
1271 
1272 	error = spa_ld_log_sm_metadata(spa);
1273 	if (error != 0)
1274 		return (error);
1275 
1276 	/*
1277 	 * Note: we don't actually expect anything to change at this point
1278 	 * but we grab the config lock so we don't fail any assertions
1279 	 * when using vdev_lookup_top().
1280 	 */
1281 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
1282 	error = spa_ld_log_sm_data(spa);
1283 	spa_config_exit(spa, SCL_CONFIG, FTAG);
1284 
1285 	return (error);
1286 }
1287