xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa_checkpoint.c (revision bc1f688b4872ace323eaddbb1a6365d054e7bf56)
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) 2017 by Delphix. All rights reserved.
24  */
25 
26 /*
27  * Storage Pool Checkpoint
28  *
29  * A storage pool checkpoint can be thought of as a pool-wide snapshot or
30  * a stable version of extreme rewind that guarantees no blocks from the
31  * checkpointed state will have been overwritten. It remembers the entire
32  * state of the storage pool (e.g. snapshots, dataset names, etc..) from the
33  * point that it was taken and the user can rewind back to that point even if
34  * they applied destructive operations on their datasets or even enabled new
35  * zpool on-disk features. If a pool has a checkpoint that is no longer
36  * needed, the user can discard it.
37  *
38  * == On disk data structures used ==
39  *
40  * - The pool has a new feature flag and a new entry in the MOS. The feature
41  *   flag is set to active when we create the checkpoint and remains active
42  *   until the checkpoint is fully discarded. The entry in the MOS config
43  *   (DMU_POOL_ZPOOL_CHECKPOINT) is populated with the uberblock that
44  *   references the state of the pool when we take the checkpoint. The entry
45  *   remains populated until we start discarding the checkpoint or we rewind
46  *   back to it.
47  *
48  * - Each vdev contains a vdev-wide space map while the pool has a checkpoint,
49  *   which persists until the checkpoint is fully discarded. The space map
50  *   contains entries that have been freed in the current state of the pool
51  *   but we want to keep around in case we decide to rewind to the checkpoint.
52  *   [see vdev_checkpoint_sm]
53  *
54  * - Each metaslab's ms_sm space map behaves the same as without the
55  *   checkpoint, with the only exception being the scenario when we free
56  *   blocks that belong to the checkpoint. In this case, these blocks remain
57  *   ALLOCATED in the metaslab's space map and they are added as FREE in the
58  *   vdev's checkpoint space map.
59  *
60  * - Each uberblock has a field (ub_checkpoint_txg) which holds the txg that
61  *   the uberblock was checkpointed. For normal uberblocks this field is 0.
62  *
63  * == Overview of operations ==
64  *
65  * - To create a checkpoint, we first wait for the current TXG to be synced,
66  *   so we can use the most recently synced uberblock (spa_ubsync) as the
67  *   checkpointed uberblock. Then we use an early synctask to place that
68  *   uberblock in MOS config, increment the feature flag for the checkpoint
69  *   (marking it active), and setting spa_checkpoint_txg (see its use below)
70  *   to the TXG of the checkpointed uberblock. We use an early synctask for
71  *   the aforementioned operations to ensure that no blocks were dirtied
72  *   between the current TXG and the TXG of the checkpointed uberblock
73  *   (e.g the previous txg).
74  *
75  * - When a checkpoint exists, we need to ensure that the blocks that
76  *   belong to the checkpoint are freed but never reused. This means that
77  *   these blocks should never end up in the ms_allocatable or the ms_freeing
78  *   trees of a metaslab. Therefore, whenever there is a checkpoint the new
79  *   ms_checkpointing tree is used in addition to the aforementioned ones.
80  *
81  *   Whenever a block is freed and we find out that it is referenced by the
82  *   checkpoint (we find out by comparing its birth to spa_checkpoint_txg),
83  *   we place it in the ms_checkpointing tree instead of the ms_freeingtree.
84  *   This way, we divide the blocks that are being freed into checkpointed
85  *   and not-checkpointed blocks.
86  *
87  *   In order to persist these frees, we write the extents from the
88  *   ms_freeingtree to the ms_sm as usual, and the extents from the
89  *   ms_checkpointing tree to the vdev_checkpoint_sm. This way, these
90  *   checkpointed extents will remain allocated in the metaslab's ms_sm space
91  *   map, and therefore won't be reused [see metaslab_sync()]. In addition,
92  *   when we discard the checkpoint, we can find the entries that have
93  *   actually been freed in vdev_checkpoint_sm.
94  *   [see spa_checkpoint_discard_thread_sync()]
95  *
96  * - To discard the checkpoint we use an early synctask to delete the
97  *   checkpointed uberblock from the MOS config, set spa_checkpoint_txg to 0,
98  *   and wakeup the discarding zthr thread (an open-context async thread).
99  *   We use an early synctask to ensure that the operation happens before any
100  *   new data end up in the checkpoint's data structures.
101  *
102  *   Once the synctask is done and the discarding zthr is awake, we discard
103  *   the checkpointed data over multiple TXGs by having the zthr prefetching
104  *   entries from vdev_checkpoint_sm and then starting a synctask that places
105  *   them as free blocks in to their respective ms_allocatable and ms_sm
106  *   structures.
107  *   [see spa_checkpoint_discard_thread()]
108  *
109  *   When there are no entries left in the vdev_checkpoint_sm of all
110  *   top-level vdevs, a final synctask runs that decrements the feature flag.
111  *
112  * - To rewind to the checkpoint, we first use the current uberblock and
113  *   open the MOS so we can access the checkpointed uberblock from the MOS
114  *   config. After we retrieve the checkpointed uberblock, we use it as the
115  *   current uberblock for the pool by writing it to disk with an updated
116  *   TXG, opening its version of the MOS, and moving on as usual from there.
117  *   [see spa_ld_checkpoint_rewind()]
118  *
119  *   An important note on rewinding to the checkpoint has to do with how we
120  *   handle ZIL blocks. In the scenario of a rewind, we clear out any ZIL
121  *   blocks that have not been claimed by the time we took the checkpoint
122  *   as they should no longer be valid.
123  *   [see comment in zil_claim()]
124  *
125  * == Miscellaneous information ==
126  *
127  * - In the hypothetical event that we take a checkpoint, remove a vdev,
128  *   and attempt to rewind, the rewind would fail as the checkpointed
129  *   uberblock would reference data in the removed device. For this reason
130  *   and others of similar nature, we disallow the following operations that
131  *   can change the config:
132  *   	vdev removal and attach/detach, mirror splitting, and pool reguid.
133  *
134  * - As most of the checkpoint logic is implemented in the SPA and doesn't
135  *   distinguish datasets when it comes to space accounting, having a
136  *   checkpoint can potentially break the boundaries set by dataset
137  *   reservations.
138  */
139 
140 #include <sys/dmu_tx.h>
141 #include <sys/dsl_dir.h>
142 #include <sys/dsl_synctask.h>
143 #include <sys/metaslab_impl.h>
144 #include <sys/spa.h>
145 #include <sys/spa_impl.h>
146 #include <sys/spa_checkpoint.h>
147 #include <sys/vdev_impl.h>
148 #include <sys/zap.h>
149 #include <sys/zfeature.h>
150 
151 /*
152  * The following parameter limits the amount of memory to be used for the
153  * prefetching of the checkpoint space map done on each vdev while
154  * discarding the checkpoint.
155  *
156  * The reason it exists is because top-level vdevs with long checkpoint
157  * space maps can potentially take up a lot of memory depending on the
158  * amount of checkpointed data that has been freed within them while
159  * the pool had a checkpoint.
160  */
161 uint64_t	zfs_spa_discard_memory_limit = 16 * 1024 * 1024;
162 
163 int
164 spa_checkpoint_get_stats(spa_t *spa, pool_checkpoint_stat_t *pcs)
165 {
166 	if (!spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT))
167 		return (SET_ERROR(ZFS_ERR_NO_CHECKPOINT));
168 
169 	bzero(pcs, sizeof (pool_checkpoint_stat_t));
170 
171 	int error = zap_contains(spa_meta_objset(spa),
172 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT);
173 	ASSERT(error == 0 || error == ENOENT);
174 
175 	if (error == ENOENT)
176 		pcs->pcs_state = CS_CHECKPOINT_DISCARDING;
177 	else
178 		pcs->pcs_state = CS_CHECKPOINT_EXISTS;
179 
180 	pcs->pcs_space = spa->spa_checkpoint_info.sci_dspace;
181 	pcs->pcs_start_time = spa->spa_checkpoint_info.sci_timestamp;
182 
183 	return (0);
184 }
185 
186 static void
187 spa_checkpoint_discard_complete_sync(void *arg, dmu_tx_t *tx)
188 {
189 	spa_t *spa = arg;
190 
191 	spa->spa_checkpoint_info.sci_timestamp = 0;
192 
193 	spa_feature_decr(spa, SPA_FEATURE_POOL_CHECKPOINT, tx);
194 
195 	spa_history_log_internal(spa, "spa discard checkpoint", tx,
196 	    "finished discarding checkpointed state from the pool");
197 }
198 
199 typedef struct spa_checkpoint_discard_sync_callback_arg {
200 	vdev_t *sdc_vd;
201 	uint64_t sdc_txg;
202 	uint64_t sdc_entry_limit;
203 } spa_checkpoint_discard_sync_callback_arg_t;
204 
205 static int
206 spa_checkpoint_discard_sync_callback(space_map_entry_t *sme, void *arg)
207 {
208 	spa_checkpoint_discard_sync_callback_arg_t *sdc = arg;
209 	vdev_t *vd = sdc->sdc_vd;
210 	metaslab_t *ms = vd->vdev_ms[sme->sme_offset >> vd->vdev_ms_shift];
211 	uint64_t end = sme->sme_offset + sme->sme_run;
212 
213 	if (sdc->sdc_entry_limit == 0)
214 		return (EINTR);
215 
216 	/*
217 	 * Since the space map is not condensed, we know that
218 	 * none of its entries is crossing the boundaries of
219 	 * its respective metaslab.
220 	 *
221 	 * That said, there is no fundamental requirement that
222 	 * the checkpoint's space map entries should not cross
223 	 * metaslab boundaries. So if needed we could add code
224 	 * that handles metaslab-crossing segments in the future.
225 	 */
226 	VERIFY3U(sme->sme_type, ==, SM_FREE);
227 	VERIFY3U(sme->sme_offset, >=, ms->ms_start);
228 	VERIFY3U(end, <=, ms->ms_start + ms->ms_size);
229 
230 	/*
231 	 * At this point we should not be processing any
232 	 * other frees concurrently, so the lock is technically
233 	 * unnecessary. We use the lock anyway though to
234 	 * potentially save ourselves from future headaches.
235 	 */
236 	mutex_enter(&ms->ms_lock);
237 	if (range_tree_is_empty(ms->ms_freeing))
238 		vdev_dirty(vd, VDD_METASLAB, ms, sdc->sdc_txg);
239 	range_tree_add(ms->ms_freeing, sme->sme_offset, sme->sme_run);
240 	mutex_exit(&ms->ms_lock);
241 
242 	ASSERT3U(vd->vdev_spa->spa_checkpoint_info.sci_dspace, >=,
243 	    sme->sme_run);
244 	ASSERT3U(vd->vdev_stat.vs_checkpoint_space, >=, sme->sme_run);
245 
246 	vd->vdev_spa->spa_checkpoint_info.sci_dspace -= sme->sme_run;
247 	vd->vdev_stat.vs_checkpoint_space -= sme->sme_run;
248 	sdc->sdc_entry_limit--;
249 
250 	return (0);
251 }
252 
253 static void
254 spa_checkpoint_accounting_verify(spa_t *spa)
255 {
256 	vdev_t *rvd = spa->spa_root_vdev;
257 	uint64_t ckpoint_sm_space_sum = 0;
258 	uint64_t vs_ckpoint_space_sum = 0;
259 
260 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
261 		vdev_t *vd = rvd->vdev_child[c];
262 
263 		if (vd->vdev_checkpoint_sm != NULL) {
264 			ckpoint_sm_space_sum +=
265 			    -vd->vdev_checkpoint_sm->sm_alloc;
266 			vs_ckpoint_space_sum +=
267 			    vd->vdev_stat.vs_checkpoint_space;
268 			ASSERT3U(ckpoint_sm_space_sum, ==,
269 			    vs_ckpoint_space_sum);
270 		} else {
271 			ASSERT0(vd->vdev_stat.vs_checkpoint_space);
272 		}
273 	}
274 	ASSERT3U(spa->spa_checkpoint_info.sci_dspace, ==, ckpoint_sm_space_sum);
275 }
276 
277 static void
278 spa_checkpoint_discard_thread_sync(void *arg, dmu_tx_t *tx)
279 {
280 	vdev_t *vd = arg;
281 	int error;
282 
283 	/*
284 	 * The space map callback is applied only to non-debug entries.
285 	 * Because the number of debug entries is less or equal to the
286 	 * number of non-debug entries, we want to ensure that we only
287 	 * read what we prefetched from open-context.
288 	 *
289 	 * Thus, we set the maximum entries that the space map callback
290 	 * will be applied to be half the entries that could fit in the
291 	 * imposed memory limit.
292 	 *
293 	 * Note that since this is a conservative estimate we also
294 	 * assume the worst case scenario in our computation where each
295 	 * entry is two-word.
296 	 */
297 	uint64_t max_entry_limit =
298 	    (zfs_spa_discard_memory_limit / (2 * sizeof (uint64_t))) >> 1;
299 
300 	/*
301 	 * Iterate from the end of the space map towards the beginning,
302 	 * placing its entries on ms_freeing and removing them from the
303 	 * space map. The iteration stops if one of the following
304 	 * conditions is true:
305 	 *
306 	 * 1] We reached the beginning of the space map. At this point
307 	 *    the space map should be completely empty and
308 	 *    space_map_incremental_destroy should have returned 0.
309 	 *    The next step would be to free and close the space map
310 	 *    and remove its entry from its vdev's top zap. This allows
311 	 *    spa_checkpoint_discard_thread() to move on to the next vdev.
312 	 *
313 	 * 2] We reached the memory limit (amount of memory used to hold
314 	 *    space map entries in memory) and space_map_incremental_destroy
315 	 *    returned EINTR. This means that there are entries remaining
316 	 *    in the space map that will be cleared in a future invocation
317 	 *    of this function by spa_checkpoint_discard_thread().
318 	 */
319 	spa_checkpoint_discard_sync_callback_arg_t sdc;
320 	sdc.sdc_vd = vd;
321 	sdc.sdc_txg = tx->tx_txg;
322 	sdc.sdc_entry_limit = max_entry_limit;
323 
324 	uint64_t words_before =
325 	    space_map_length(vd->vdev_checkpoint_sm) / sizeof (uint64_t);
326 
327 	error = space_map_incremental_destroy(vd->vdev_checkpoint_sm,
328 	    spa_checkpoint_discard_sync_callback, &sdc, tx);
329 
330 	uint64_t words_after =
331 	    space_map_length(vd->vdev_checkpoint_sm) / sizeof (uint64_t);
332 
333 #ifdef DEBUG
334 	spa_checkpoint_accounting_verify(vd->vdev_spa);
335 #endif
336 
337 	zfs_dbgmsg("discarding checkpoint: txg %llu, vdev id %d, "
338 	    "deleted %llu words - %llu words are left",
339 	    tx->tx_txg, vd->vdev_id, (words_before - words_after),
340 	    words_after);
341 
342 	if (error != EINTR) {
343 		if (error != 0) {
344 			zfs_panic_recover("zfs: error %d was returned "
345 			    "while incrementally destroying the checkpoint "
346 			    "space map of vdev %llu\n",
347 			    error, vd->vdev_id);
348 		}
349 		ASSERT0(words_after);
350 		ASSERT0(vd->vdev_checkpoint_sm->sm_alloc);
351 		ASSERT0(space_map_length(vd->vdev_checkpoint_sm));
352 
353 		space_map_free(vd->vdev_checkpoint_sm, tx);
354 		space_map_close(vd->vdev_checkpoint_sm);
355 		vd->vdev_checkpoint_sm = NULL;
356 
357 		VERIFY0(zap_remove(spa_meta_objset(vd->vdev_spa),
358 		    vd->vdev_top_zap, VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, tx));
359 	}
360 }
361 
362 static boolean_t
363 spa_checkpoint_discard_is_done(spa_t *spa)
364 {
365 	vdev_t *rvd = spa->spa_root_vdev;
366 
367 	ASSERT(!spa_has_checkpoint(spa));
368 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT));
369 
370 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
371 		if (rvd->vdev_child[c]->vdev_checkpoint_sm != NULL)
372 			return (B_FALSE);
373 		ASSERT0(rvd->vdev_child[c]->vdev_stat.vs_checkpoint_space);
374 	}
375 
376 	return (B_TRUE);
377 }
378 
379 /* ARGSUSED */
380 boolean_t
381 spa_checkpoint_discard_thread_check(void *arg, zthr_t *zthr)
382 {
383 	spa_t *spa = arg;
384 
385 	if (!spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT))
386 		return (B_FALSE);
387 
388 	if (spa_has_checkpoint(spa))
389 		return (B_FALSE);
390 
391 	return (B_TRUE);
392 }
393 
394 int
395 spa_checkpoint_discard_thread(void *arg, zthr_t *zthr)
396 {
397 	spa_t *spa = arg;
398 	vdev_t *rvd = spa->spa_root_vdev;
399 
400 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
401 		vdev_t *vd = rvd->vdev_child[c];
402 
403 		while (vd->vdev_checkpoint_sm != NULL) {
404 			space_map_t *checkpoint_sm = vd->vdev_checkpoint_sm;
405 			int numbufs;
406 			dmu_buf_t **dbp;
407 
408 			if (zthr_iscancelled(zthr))
409 				return (0);
410 
411 			ASSERT3P(vd->vdev_ops, !=, &vdev_indirect_ops);
412 
413 			uint64_t size = MIN(space_map_length(checkpoint_sm),
414 			    zfs_spa_discard_memory_limit);
415 			uint64_t offset =
416 			    space_map_length(checkpoint_sm) - size;
417 
418 			/*
419 			 * Ensure that the part of the space map that will
420 			 * be destroyed by the synctask, is prefetched in
421 			 * memory before the synctask runs.
422 			 */
423 			int error = dmu_buf_hold_array_by_bonus(
424 			    checkpoint_sm->sm_dbuf, offset, size,
425 			    B_TRUE, FTAG, &numbufs, &dbp);
426 			if (error != 0) {
427 				zfs_panic_recover("zfs: error %d was returned "
428 				    "while prefetching checkpoint space map "
429 				    "entries of vdev %llu\n",
430 				    error, vd->vdev_id);
431 			}
432 
433 			VERIFY0(dsl_sync_task(spa->spa_name, NULL,
434 			    spa_checkpoint_discard_thread_sync, vd,
435 			    0, ZFS_SPACE_CHECK_NONE));
436 
437 			dmu_buf_rele_array(dbp, numbufs, FTAG);
438 		}
439 	}
440 
441 	VERIFY(spa_checkpoint_discard_is_done(spa));
442 	VERIFY0(spa->spa_checkpoint_info.sci_dspace);
443 	VERIFY0(dsl_sync_task(spa->spa_name, NULL,
444 	    spa_checkpoint_discard_complete_sync, spa,
445 	    0, ZFS_SPACE_CHECK_NONE));
446 
447 	return (0);
448 }
449 
450 
451 /* ARGSUSED */
452 static int
453 spa_checkpoint_check(void *arg, dmu_tx_t *tx)
454 {
455 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
456 
457 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_POOL_CHECKPOINT))
458 		return (SET_ERROR(ENOTSUP));
459 
460 	if (!spa_top_vdevs_spacemap_addressable(spa))
461 		return (SET_ERROR(ZFS_ERR_VDEV_TOO_BIG));
462 
463 	if (spa->spa_vdev_removal != NULL)
464 		return (SET_ERROR(ZFS_ERR_DEVRM_IN_PROGRESS));
465 
466 	if (spa->spa_checkpoint_txg != 0)
467 		return (SET_ERROR(ZFS_ERR_CHECKPOINT_EXISTS));
468 
469 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT))
470 		return (SET_ERROR(ZFS_ERR_DISCARDING_CHECKPOINT));
471 
472 	return (0);
473 }
474 
475 /* ARGSUSED */
476 static void
477 spa_checkpoint_sync(void *arg, dmu_tx_t *tx)
478 {
479 	dsl_pool_t *dp = dmu_tx_pool(tx);
480 	spa_t *spa = dp->dp_spa;
481 	uberblock_t checkpoint = spa->spa_ubsync;
482 
483 	/*
484 	 * At this point, there should not be a checkpoint in the MOS.
485 	 */
486 	ASSERT3U(zap_contains(spa_meta_objset(spa), DMU_POOL_DIRECTORY_OBJECT,
487 	    DMU_POOL_ZPOOL_CHECKPOINT), ==, ENOENT);
488 
489 	ASSERT0(spa->spa_checkpoint_info.sci_timestamp);
490 	ASSERT0(spa->spa_checkpoint_info.sci_dspace);
491 
492 	/*
493 	 * Since the checkpointed uberblock is the one that just got synced
494 	 * (we use spa_ubsync), its txg must be equal to the txg number of
495 	 * the txg we are syncing, minus 1.
496 	 */
497 	ASSERT3U(checkpoint.ub_txg, ==, spa->spa_syncing_txg - 1);
498 
499 	/*
500 	 * Once the checkpoint is in place, we need to ensure that none of
501 	 * its blocks will be marked for reuse after it has been freed.
502 	 * When there is a checkpoint and a block is freed, we compare its
503 	 * birth txg to the txg of the checkpointed uberblock to see if the
504 	 * block is part of the checkpoint or not. Therefore, we have to set
505 	 * spa_checkpoint_txg before any frees happen in this txg (which is
506 	 * why this is done as an early_synctask as explained in the comment
507 	 * in spa_checkpoint()).
508 	 */
509 	spa->spa_checkpoint_txg = checkpoint.ub_txg;
510 	spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp;
511 
512 	checkpoint.ub_checkpoint_txg = checkpoint.ub_txg;
513 	VERIFY0(zap_add(spa->spa_dsl_pool->dp_meta_objset,
514 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT,
515 	    sizeof (uint64_t), sizeof (uberblock_t) / sizeof (uint64_t),
516 	    &checkpoint, tx));
517 
518 	/*
519 	 * Increment the feature refcount and thus activate the feature.
520 	 * Note that the feature will be deactivated when we've
521 	 * completely discarded all checkpointed state (both vdev
522 	 * space maps and uberblock).
523 	 */
524 	spa_feature_incr(spa, SPA_FEATURE_POOL_CHECKPOINT, tx);
525 
526 	spa_history_log_internal(spa, "spa checkpoint", tx,
527 	    "checkpointed uberblock txg=%llu", checkpoint.ub_txg);
528 }
529 
530 /*
531  * Create a checkpoint for the pool.
532  */
533 int
534 spa_checkpoint(const char *pool)
535 {
536 	int error;
537 	spa_t *spa;
538 
539 	error = spa_open(pool, &spa, FTAG);
540 	if (error != 0)
541 		return (error);
542 
543 	mutex_enter(&spa->spa_vdev_top_lock);
544 
545 	/*
546 	 * Wait for current syncing txg to finish so the latest synced
547 	 * uberblock (spa_ubsync) has all the changes that we expect
548 	 * to see if we were to revert later to the checkpoint. In other
549 	 * words we want the checkpointed uberblock to include/reference
550 	 * all the changes that were pending at the time that we issued
551 	 * the checkpoint command.
552 	 */
553 	txg_wait_synced(spa_get_dsl(spa), 0);
554 
555 	/*
556 	 * As the checkpointed uberblock references blocks from the previous
557 	 * txg (spa_ubsync) we want to ensure that are not freeing any of
558 	 * these blocks in the same txg that the following synctask will
559 	 * run. Thus, we run it as an early synctask, so the dirty changes
560 	 * that are synced to disk afterwards during zios and other synctasks
561 	 * do not reuse checkpointed blocks.
562 	 */
563 	error = dsl_early_sync_task(pool, spa_checkpoint_check,
564 	    spa_checkpoint_sync, NULL, 0, ZFS_SPACE_CHECK_NORMAL);
565 
566 	mutex_exit(&spa->spa_vdev_top_lock);
567 
568 	spa_close(spa, FTAG);
569 	return (error);
570 }
571 
572 /* ARGSUSED */
573 static int
574 spa_checkpoint_discard_check(void *arg, dmu_tx_t *tx)
575 {
576 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
577 
578 	if (!spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT))
579 		return (SET_ERROR(ZFS_ERR_NO_CHECKPOINT));
580 
581 	if (spa->spa_checkpoint_txg == 0)
582 		return (SET_ERROR(ZFS_ERR_DISCARDING_CHECKPOINT));
583 
584 	VERIFY0(zap_contains(spa_meta_objset(spa),
585 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT));
586 
587 	return (0);
588 }
589 
590 /* ARGSUSED */
591 static void
592 spa_checkpoint_discard_sync(void *arg, dmu_tx_t *tx)
593 {
594 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
595 
596 	VERIFY0(zap_remove(spa_meta_objset(spa), DMU_POOL_DIRECTORY_OBJECT,
597 	    DMU_POOL_ZPOOL_CHECKPOINT, tx));
598 
599 	spa->spa_checkpoint_txg = 0;
600 
601 	zthr_wakeup(spa->spa_checkpoint_discard_zthr);
602 
603 	spa_history_log_internal(spa, "spa discard checkpoint", tx,
604 	    "started discarding checkpointed state from the pool");
605 }
606 
607 /*
608  * Discard the checkpoint from a pool.
609  */
610 int
611 spa_checkpoint_discard(const char *pool)
612 {
613 	/*
614 	 * Similarly to spa_checkpoint(), we want our synctask to run
615 	 * before any pending dirty data are written to disk so they
616 	 * won't end up in the checkpoint's data structures (e.g.
617 	 * ms_checkpointing and vdev_checkpoint_sm) and re-create any
618 	 * space maps that the discarding open-context thread has
619 	 * deleted.
620 	 * [see spa_discard_checkpoint_sync and spa_discard_checkpoint_thread]
621 	 */
622 	return (dsl_early_sync_task(pool, spa_checkpoint_discard_check,
623 	    spa_checkpoint_discard_sync, NULL, 0,
624 	    ZFS_SPACE_CHECK_DISCARD_CHECKPOINT));
625 }
626