xref: /freebsd/sys/contrib/openzfs/module/zfs/dsl_scan.c (revision 258a0d760aa8b42899a000e30f610f900a402556)
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 https://opensource.org/licenses/CDDL-1.0.
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 (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2021 by Delphix. All rights reserved.
24  * Copyright 2016 Gary Mills
25  * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
26  * Copyright (c) 2015, Nexenta Systems, Inc. All rights reserved.
27  * Copyright 2019 Joyent, Inc.
28  */
29 
30 #include <sys/dsl_scan.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_prop.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_synctask.h>
36 #include <sys/dnode.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dmu_objset.h>
39 #include <sys/arc.h>
40 #include <sys/arc_impl.h>
41 #include <sys/zap.h>
42 #include <sys/zio.h>
43 #include <sys/zfs_context.h>
44 #include <sys/fs/zfs.h>
45 #include <sys/zfs_znode.h>
46 #include <sys/spa_impl.h>
47 #include <sys/vdev_impl.h>
48 #include <sys/zil_impl.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/brt.h>
51 #include <sys/ddt.h>
52 #include <sys/sa.h>
53 #include <sys/sa_impl.h>
54 #include <sys/zfeature.h>
55 #include <sys/abd.h>
56 #include <sys/range_tree.h>
57 #include <sys/dbuf.h>
58 #ifdef _KERNEL
59 #include <sys/zfs_vfsops.h>
60 #endif
61 
62 /*
63  * Grand theory statement on scan queue sorting
64  *
65  * Scanning is implemented by recursively traversing all indirection levels
66  * in an object and reading all blocks referenced from said objects. This
67  * results in us approximately traversing the object from lowest logical
68  * offset to the highest. For best performance, we would want the logical
69  * blocks to be physically contiguous. However, this is frequently not the
70  * case with pools given the allocation patterns of copy-on-write filesystems.
71  * So instead, we put the I/Os into a reordering queue and issue them in a
72  * way that will most benefit physical disks (LBA-order).
73  *
74  * Queue management:
75  *
76  * Ideally, we would want to scan all metadata and queue up all block I/O
77  * prior to starting to issue it, because that allows us to do an optimal
78  * sorting job. This can however consume large amounts of memory. Therefore
79  * we continuously monitor the size of the queues and constrain them to 5%
80  * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
81  * limit, we clear out a few of the largest extents at the head of the queues
82  * to make room for more scanning. Hopefully, these extents will be fairly
83  * large and contiguous, allowing us to approach sequential I/O throughput
84  * even without a fully sorted tree.
85  *
86  * Metadata scanning takes place in dsl_scan_visit(), which is called from
87  * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
88  * metadata on the pool, or we need to make room in memory because our
89  * queues are too large, dsl_scan_visit() is postponed and
90  * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
91  * that metadata scanning and queued I/O issuing are mutually exclusive. This
92  * allows us to provide maximum sequential I/O throughput for the majority of
93  * I/O's issued since sequential I/O performance is significantly negatively
94  * impacted if it is interleaved with random I/O.
95  *
96  * Implementation Notes
97  *
98  * One side effect of the queued scanning algorithm is that the scanning code
99  * needs to be notified whenever a block is freed. This is needed to allow
100  * the scanning code to remove these I/Os from the issuing queue. Additionally,
101  * we do not attempt to queue gang blocks to be issued sequentially since this
102  * is very hard to do and would have an extremely limited performance benefit.
103  * Instead, we simply issue gang I/Os as soon as we find them using the legacy
104  * algorithm.
105  *
106  * Backwards compatibility
107  *
108  * This new algorithm is backwards compatible with the legacy on-disk data
109  * structures (and therefore does not require a new feature flag).
110  * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
111  * will stop scanning metadata (in logical order) and wait for all outstanding
112  * sorted I/O to complete. Once this is done, we write out a checkpoint
113  * bookmark, indicating that we have scanned everything logically before it.
114  * If the pool is imported on a machine without the new sorting algorithm,
115  * the scan simply resumes from the last checkpoint using the legacy algorithm.
116  */
117 
118 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
119     const zbookmark_phys_t *);
120 
121 static scan_cb_t dsl_scan_scrub_cb;
122 
123 static int scan_ds_queue_compare(const void *a, const void *b);
124 static int scan_prefetch_queue_compare(const void *a, const void *b);
125 static void scan_ds_queue_clear(dsl_scan_t *scn);
126 static void scan_ds_prefetch_queue_clear(dsl_scan_t *scn);
127 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
128     uint64_t *txg);
129 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
130 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
131 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
132 static uint64_t dsl_scan_count_data_disks(spa_t *spa);
133 static void read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb);
134 
135 extern uint_t zfs_vdev_async_write_active_min_dirty_percent;
136 static int zfs_scan_blkstats = 0;
137 
138 /*
139  * 'zpool status' uses bytes processed per pass to report throughput and
140  * estimate time remaining.  We define a pass to start when the scanning
141  * phase completes for a sequential resilver.  Optionally, this value
142  * may be used to reset the pass statistics every N txgs to provide an
143  * estimated completion time based on currently observed performance.
144  */
145 static uint_t zfs_scan_report_txgs = 0;
146 
147 /*
148  * By default zfs will check to ensure it is not over the hard memory
149  * limit before each txg. If finer-grained control of this is needed
150  * this value can be set to 1 to enable checking before scanning each
151  * block.
152  */
153 static int zfs_scan_strict_mem_lim = B_FALSE;
154 
155 /*
156  * Maximum number of parallelly executed bytes per leaf vdev. We attempt
157  * to strike a balance here between keeping the vdev queues full of I/Os
158  * at all times and not overflowing the queues to cause long latency,
159  * which would cause long txg sync times. No matter what, we will not
160  * overload the drives with I/O, since that is protected by
161  * zfs_vdev_scrub_max_active.
162  */
163 static uint64_t zfs_scan_vdev_limit = 16 << 20;
164 
165 static uint_t zfs_scan_issue_strategy = 0;
166 
167 /* don't queue & sort zios, go direct */
168 static int zfs_scan_legacy = B_FALSE;
169 static uint64_t zfs_scan_max_ext_gap = 2 << 20; /* in bytes */
170 
171 /*
172  * fill_weight is non-tunable at runtime, so we copy it at module init from
173  * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
174  * break queue sorting.
175  */
176 static uint_t zfs_scan_fill_weight = 3;
177 static uint64_t fill_weight;
178 
179 /* See dsl_scan_should_clear() for details on the memory limit tunables */
180 static const uint64_t zfs_scan_mem_lim_min = 16 << 20;	/* bytes */
181 static const uint64_t zfs_scan_mem_lim_soft_max = 128 << 20;	/* bytes */
182 
183 
184 /* fraction of physmem */
185 static uint_t zfs_scan_mem_lim_fact = 20;
186 
187 /* fraction of mem lim above */
188 static uint_t zfs_scan_mem_lim_soft_fact = 20;
189 
190 /* minimum milliseconds to scrub per txg */
191 static uint_t zfs_scrub_min_time_ms = 1000;
192 
193 /* minimum milliseconds to obsolete per txg */
194 static uint_t zfs_obsolete_min_time_ms = 500;
195 
196 /* minimum milliseconds to free per txg */
197 static uint_t zfs_free_min_time_ms = 1000;
198 
199 /* minimum milliseconds to resilver per txg */
200 static uint_t zfs_resilver_min_time_ms = 3000;
201 
202 static uint_t zfs_scan_checkpoint_intval = 7200; /* in seconds */
203 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
204 static int zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
205 static int zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
206 static const enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
207 /* max number of blocks to free in a single TXG */
208 static uint64_t zfs_async_block_max_blocks = UINT64_MAX;
209 /* max number of dedup blocks to free in a single TXG */
210 static uint64_t zfs_max_async_dedup_frees = 100000;
211 
212 /* set to disable resilver deferring */
213 static int zfs_resilver_disable_defer = B_FALSE;
214 
215 /*
216  * We wait a few txgs after importing a pool to begin scanning so that
217  * the import / mounting code isn't held up by scrub / resilver IO.
218  * Unfortunately, it is a bit difficult to determine exactly how long
219  * this will take since userspace will trigger fs mounts asynchronously
220  * and the kernel will create zvol minors asynchronously. As a result,
221  * the value provided here is a bit arbitrary, but represents a
222  * reasonable estimate of how many txgs it will take to finish fully
223  * importing a pool
224  */
225 #define	SCAN_IMPORT_WAIT_TXGS 		5
226 
227 #define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
228 	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
229 	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
230 
231 /*
232  * Enable/disable the processing of the free_bpobj object.
233  */
234 static int zfs_free_bpobj_enabled = 1;
235 
236 /* Error blocks to be scrubbed in one txg. */
237 static uint_t zfs_scrub_error_blocks_per_txg = 1 << 12;
238 
239 /* the order has to match pool_scan_type */
240 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
241 	NULL,
242 	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
243 	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
244 };
245 
246 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
247 typedef struct {
248 	uint64_t	sds_dsobj;
249 	uint64_t	sds_txg;
250 	avl_node_t	sds_node;
251 } scan_ds_t;
252 
253 /*
254  * This controls what conditions are placed on dsl_scan_sync_state():
255  * SYNC_OPTIONAL) write out scn_phys iff scn_queues_pending == 0
256  * SYNC_MANDATORY) write out scn_phys always. scn_queues_pending must be 0.
257  * SYNC_CACHED) if scn_queues_pending == 0, write out scn_phys. Otherwise
258  *	write out the scn_phys_cached version.
259  * See dsl_scan_sync_state for details.
260  */
261 typedef enum {
262 	SYNC_OPTIONAL,
263 	SYNC_MANDATORY,
264 	SYNC_CACHED
265 } state_sync_type_t;
266 
267 /*
268  * This struct represents the minimum information needed to reconstruct a
269  * zio for sequential scanning. This is useful because many of these will
270  * accumulate in the sequential IO queues before being issued, so saving
271  * memory matters here.
272  */
273 typedef struct scan_io {
274 	/* fields from blkptr_t */
275 	uint64_t		sio_blk_prop;
276 	uint64_t		sio_phys_birth;
277 	uint64_t		sio_birth;
278 	zio_cksum_t		sio_cksum;
279 	uint32_t		sio_nr_dvas;
280 
281 	/* fields from zio_t */
282 	uint32_t		sio_flags;
283 	zbookmark_phys_t	sio_zb;
284 
285 	/* members for queue sorting */
286 	union {
287 		avl_node_t	sio_addr_node; /* link into issuing queue */
288 		list_node_t	sio_list_node; /* link for issuing to disk */
289 	} sio_nodes;
290 
291 	/*
292 	 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
293 	 * depending on how many were in the original bp. Only the
294 	 * first DVA is really used for sorting and issuing purposes.
295 	 * The other DVAs (if provided) simply exist so that the zio
296 	 * layer can find additional copies to repair from in the
297 	 * event of an error. This array must go at the end of the
298 	 * struct to allow this for the variable number of elements.
299 	 */
300 	dva_t			sio_dva[];
301 } scan_io_t;
302 
303 #define	SIO_SET_OFFSET(sio, x)		DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
304 #define	SIO_SET_ASIZE(sio, x)		DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
305 #define	SIO_GET_OFFSET(sio)		DVA_GET_OFFSET(&(sio)->sio_dva[0])
306 #define	SIO_GET_ASIZE(sio)		DVA_GET_ASIZE(&(sio)->sio_dva[0])
307 #define	SIO_GET_END_OFFSET(sio)		\
308 	(SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
309 #define	SIO_GET_MUSED(sio)		\
310 	(sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
311 
312 struct dsl_scan_io_queue {
313 	dsl_scan_t	*q_scn; /* associated dsl_scan_t */
314 	vdev_t		*q_vd; /* top-level vdev that this queue represents */
315 	zio_t		*q_zio; /* scn_zio_root child for waiting on IO */
316 
317 	/* trees used for sorting I/Os and extents of I/Os */
318 	range_tree_t	*q_exts_by_addr;
319 	zfs_btree_t	q_exts_by_size;
320 	avl_tree_t	q_sios_by_addr;
321 	uint64_t	q_sio_memused;
322 	uint64_t	q_last_ext_addr;
323 
324 	/* members for zio rate limiting */
325 	uint64_t	q_maxinflight_bytes;
326 	uint64_t	q_inflight_bytes;
327 	kcondvar_t	q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
328 
329 	/* per txg statistics */
330 	uint64_t	q_total_seg_size_this_txg;
331 	uint64_t	q_segs_this_txg;
332 	uint64_t	q_total_zio_size_this_txg;
333 	uint64_t	q_zios_this_txg;
334 };
335 
336 /* private data for dsl_scan_prefetch_cb() */
337 typedef struct scan_prefetch_ctx {
338 	zfs_refcount_t spc_refcnt;	/* refcount for memory management */
339 	dsl_scan_t *spc_scn;		/* dsl_scan_t for the pool */
340 	boolean_t spc_root;		/* is this prefetch for an objset? */
341 	uint8_t spc_indblkshift;	/* dn_indblkshift of current dnode */
342 	uint16_t spc_datablkszsec;	/* dn_idatablkszsec of current dnode */
343 } scan_prefetch_ctx_t;
344 
345 /* private data for dsl_scan_prefetch() */
346 typedef struct scan_prefetch_issue_ctx {
347 	avl_node_t spic_avl_node;	/* link into scn->scn_prefetch_queue */
348 	scan_prefetch_ctx_t *spic_spc;	/* spc for the callback */
349 	blkptr_t spic_bp;		/* bp to prefetch */
350 	zbookmark_phys_t spic_zb;	/* bookmark to prefetch */
351 } scan_prefetch_issue_ctx_t;
352 
353 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
354     const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
355 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
356     scan_io_t *sio);
357 
358 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
359 static void scan_io_queues_destroy(dsl_scan_t *scn);
360 
361 static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
362 
363 /* sio->sio_nr_dvas must be set so we know which cache to free from */
364 static void
365 sio_free(scan_io_t *sio)
366 {
367 	ASSERT3U(sio->sio_nr_dvas, >, 0);
368 	ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
369 
370 	kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
371 }
372 
373 /* It is up to the caller to set sio->sio_nr_dvas for freeing */
374 static scan_io_t *
375 sio_alloc(unsigned short nr_dvas)
376 {
377 	ASSERT3U(nr_dvas, >, 0);
378 	ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
379 
380 	return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
381 }
382 
383 void
384 scan_init(void)
385 {
386 	/*
387 	 * This is used in ext_size_compare() to weight segments
388 	 * based on how sparse they are. This cannot be changed
389 	 * mid-scan and the tree comparison functions don't currently
390 	 * have a mechanism for passing additional context to the
391 	 * compare functions. Thus we store this value globally and
392 	 * we only allow it to be set at module initialization time
393 	 */
394 	fill_weight = zfs_scan_fill_weight;
395 
396 	for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
397 		char name[36];
398 
399 		(void) snprintf(name, sizeof (name), "sio_cache_%d", i);
400 		sio_cache[i] = kmem_cache_create(name,
401 		    (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
402 		    0, NULL, NULL, NULL, NULL, NULL, 0);
403 	}
404 }
405 
406 void
407 scan_fini(void)
408 {
409 	for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
410 		kmem_cache_destroy(sio_cache[i]);
411 	}
412 }
413 
414 static inline boolean_t
415 dsl_scan_is_running(const dsl_scan_t *scn)
416 {
417 	return (scn->scn_phys.scn_state == DSS_SCANNING);
418 }
419 
420 boolean_t
421 dsl_scan_resilvering(dsl_pool_t *dp)
422 {
423 	return (dsl_scan_is_running(dp->dp_scan) &&
424 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
425 }
426 
427 static inline void
428 sio2bp(const scan_io_t *sio, blkptr_t *bp)
429 {
430 	memset(bp, 0, sizeof (*bp));
431 	bp->blk_prop = sio->sio_blk_prop;
432 	bp->blk_phys_birth = sio->sio_phys_birth;
433 	bp->blk_birth = sio->sio_birth;
434 	bp->blk_fill = 1;	/* we always only work with data pointers */
435 	bp->blk_cksum = sio->sio_cksum;
436 
437 	ASSERT3U(sio->sio_nr_dvas, >, 0);
438 	ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
439 
440 	memcpy(bp->blk_dva, sio->sio_dva, sio->sio_nr_dvas * sizeof (dva_t));
441 }
442 
443 static inline void
444 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
445 {
446 	sio->sio_blk_prop = bp->blk_prop;
447 	sio->sio_phys_birth = bp->blk_phys_birth;
448 	sio->sio_birth = bp->blk_birth;
449 	sio->sio_cksum = bp->blk_cksum;
450 	sio->sio_nr_dvas = BP_GET_NDVAS(bp);
451 
452 	/*
453 	 * Copy the DVAs to the sio. We need all copies of the block so
454 	 * that the self healing code can use the alternate copies if the
455 	 * first is corrupted. We want the DVA at index dva_i to be first
456 	 * in the sio since this is the primary one that we want to issue.
457 	 */
458 	for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
459 		sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
460 	}
461 }
462 
463 int
464 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
465 {
466 	int err;
467 	dsl_scan_t *scn;
468 	spa_t *spa = dp->dp_spa;
469 	uint64_t f;
470 
471 	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
472 	scn->scn_dp = dp;
473 
474 	/*
475 	 * It's possible that we're resuming a scan after a reboot so
476 	 * make sure that the scan_async_destroying flag is initialized
477 	 * appropriately.
478 	 */
479 	ASSERT(!scn->scn_async_destroying);
480 	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
481 	    SPA_FEATURE_ASYNC_DESTROY);
482 
483 	/*
484 	 * Calculate the max number of in-flight bytes for pool-wide
485 	 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max).
486 	 * Limits for the issuing phase are done per top-level vdev and
487 	 * are handled separately.
488 	 */
489 	scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20,
490 	    zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa)));
491 
492 	avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
493 	    offsetof(scan_ds_t, sds_node));
494 	avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
495 	    sizeof (scan_prefetch_issue_ctx_t),
496 	    offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
497 
498 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
499 	    "scrub_func", sizeof (uint64_t), 1, &f);
500 	if (err == 0) {
501 		/*
502 		 * There was an old-style scrub in progress.  Restart a
503 		 * new-style scrub from the beginning.
504 		 */
505 		scn->scn_restart_txg = txg;
506 		zfs_dbgmsg("old-style scrub was in progress for %s; "
507 		    "restarting new-style scrub in txg %llu",
508 		    spa->spa_name,
509 		    (longlong_t)scn->scn_restart_txg);
510 
511 		/*
512 		 * Load the queue obj from the old location so that it
513 		 * can be freed by dsl_scan_done().
514 		 */
515 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
516 		    "scrub_queue", sizeof (uint64_t), 1,
517 		    &scn->scn_phys.scn_queue_obj);
518 	} else {
519 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
520 		    DMU_POOL_ERRORSCRUB, sizeof (uint64_t),
521 		    ERRORSCRUB_PHYS_NUMINTS, &scn->errorscrub_phys);
522 
523 		if (err != 0 && err != ENOENT)
524 			return (err);
525 
526 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
527 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
528 		    &scn->scn_phys);
529 
530 		/*
531 		 * Detect if the pool contains the signature of #2094.  If it
532 		 * does properly update the scn->scn_phys structure and notify
533 		 * the administrator by setting an errata for the pool.
534 		 */
535 		if (err == EOVERFLOW) {
536 			uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
537 			VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
538 			VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
539 			    (23 * sizeof (uint64_t)));
540 
541 			err = zap_lookup(dp->dp_meta_objset,
542 			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
543 			    sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
544 			if (err == 0) {
545 				uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
546 
547 				if (overflow & ~DSL_SCAN_FLAGS_MASK ||
548 				    scn->scn_async_destroying) {
549 					spa->spa_errata =
550 					    ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
551 					return (EOVERFLOW);
552 				}
553 
554 				memcpy(&scn->scn_phys, zaptmp,
555 				    SCAN_PHYS_NUMINTS * sizeof (uint64_t));
556 				scn->scn_phys.scn_flags = overflow;
557 
558 				/* Required scrub already in progress. */
559 				if (scn->scn_phys.scn_state == DSS_FINISHED ||
560 				    scn->scn_phys.scn_state == DSS_CANCELED)
561 					spa->spa_errata =
562 					    ZPOOL_ERRATA_ZOL_2094_SCRUB;
563 			}
564 		}
565 
566 		if (err == ENOENT)
567 			return (0);
568 		else if (err)
569 			return (err);
570 
571 		/*
572 		 * We might be restarting after a reboot, so jump the issued
573 		 * counter to how far we've scanned. We know we're consistent
574 		 * up to here.
575 		 */
576 		scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
577 
578 		if (dsl_scan_is_running(scn) &&
579 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
580 			/*
581 			 * A new-type scrub was in progress on an old
582 			 * pool, and the pool was accessed by old
583 			 * software.  Restart from the beginning, since
584 			 * the old software may have changed the pool in
585 			 * the meantime.
586 			 */
587 			scn->scn_restart_txg = txg;
588 			zfs_dbgmsg("new-style scrub for %s was modified "
589 			    "by old software; restarting in txg %llu",
590 			    spa->spa_name,
591 			    (longlong_t)scn->scn_restart_txg);
592 		} else if (dsl_scan_resilvering(dp)) {
593 			/*
594 			 * If a resilver is in progress and there are already
595 			 * errors, restart it instead of finishing this scan and
596 			 * then restarting it. If there haven't been any errors
597 			 * then remember that the incore DTL is valid.
598 			 */
599 			if (scn->scn_phys.scn_errors > 0) {
600 				scn->scn_restart_txg = txg;
601 				zfs_dbgmsg("resilver can't excise DTL_MISSING "
602 				    "when finished; restarting on %s in txg "
603 				    "%llu",
604 				    spa->spa_name,
605 				    (u_longlong_t)scn->scn_restart_txg);
606 			} else {
607 				/* it's safe to excise DTL when finished */
608 				spa->spa_scrub_started = B_TRUE;
609 			}
610 		}
611 	}
612 
613 	memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
614 
615 	/* reload the queue into the in-core state */
616 	if (scn->scn_phys.scn_queue_obj != 0) {
617 		zap_cursor_t zc;
618 		zap_attribute_t za;
619 
620 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
621 		    scn->scn_phys.scn_queue_obj);
622 		    zap_cursor_retrieve(&zc, &za) == 0;
623 		    (void) zap_cursor_advance(&zc)) {
624 			scan_ds_queue_insert(scn,
625 			    zfs_strtonum(za.za_name, NULL),
626 			    za.za_first_integer);
627 		}
628 		zap_cursor_fini(&zc);
629 	}
630 
631 	spa_scan_stat_init(spa);
632 	vdev_scan_stat_init(spa->spa_root_vdev);
633 
634 	return (0);
635 }
636 
637 void
638 dsl_scan_fini(dsl_pool_t *dp)
639 {
640 	if (dp->dp_scan != NULL) {
641 		dsl_scan_t *scn = dp->dp_scan;
642 
643 		if (scn->scn_taskq != NULL)
644 			taskq_destroy(scn->scn_taskq);
645 
646 		scan_ds_queue_clear(scn);
647 		avl_destroy(&scn->scn_queue);
648 		scan_ds_prefetch_queue_clear(scn);
649 		avl_destroy(&scn->scn_prefetch_queue);
650 
651 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
652 		dp->dp_scan = NULL;
653 	}
654 }
655 
656 static boolean_t
657 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
658 {
659 	return (scn->scn_restart_txg != 0 &&
660 	    scn->scn_restart_txg <= tx->tx_txg);
661 }
662 
663 boolean_t
664 dsl_scan_resilver_scheduled(dsl_pool_t *dp)
665 {
666 	return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) ||
667 	    (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER));
668 }
669 
670 boolean_t
671 dsl_scan_scrubbing(const dsl_pool_t *dp)
672 {
673 	dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
674 
675 	return (scn_phys->scn_state == DSS_SCANNING &&
676 	    scn_phys->scn_func == POOL_SCAN_SCRUB);
677 }
678 
679 boolean_t
680 dsl_errorscrubbing(const dsl_pool_t *dp)
681 {
682 	dsl_errorscrub_phys_t *errorscrub_phys = &dp->dp_scan->errorscrub_phys;
683 
684 	return (errorscrub_phys->dep_state == DSS_ERRORSCRUBBING &&
685 	    errorscrub_phys->dep_func == POOL_SCAN_ERRORSCRUB);
686 }
687 
688 boolean_t
689 dsl_errorscrub_is_paused(const dsl_scan_t *scn)
690 {
691 	return (dsl_errorscrubbing(scn->scn_dp) &&
692 	    scn->errorscrub_phys.dep_paused_flags);
693 }
694 
695 boolean_t
696 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
697 {
698 	return (dsl_scan_scrubbing(scn->scn_dp) &&
699 	    scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
700 }
701 
702 static void
703 dsl_errorscrub_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
704 {
705 	scn->errorscrub_phys.dep_cursor =
706 	    zap_cursor_serialize(&scn->errorscrub_cursor);
707 
708 	VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
709 	    DMU_POOL_DIRECTORY_OBJECT,
710 	    DMU_POOL_ERRORSCRUB, sizeof (uint64_t), ERRORSCRUB_PHYS_NUMINTS,
711 	    &scn->errorscrub_phys, tx));
712 }
713 
714 static void
715 dsl_errorscrub_setup_sync(void *arg, dmu_tx_t *tx)
716 {
717 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
718 	pool_scan_func_t *funcp = arg;
719 	dsl_pool_t *dp = scn->scn_dp;
720 	spa_t *spa = dp->dp_spa;
721 
722 	ASSERT(!dsl_scan_is_running(scn));
723 	ASSERT(!dsl_errorscrubbing(scn->scn_dp));
724 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
725 
726 	memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys));
727 	scn->errorscrub_phys.dep_func = *funcp;
728 	scn->errorscrub_phys.dep_state = DSS_ERRORSCRUBBING;
729 	scn->errorscrub_phys.dep_start_time = gethrestime_sec();
730 	scn->errorscrub_phys.dep_to_examine = spa_get_last_errlog_size(spa);
731 	scn->errorscrub_phys.dep_examined = 0;
732 	scn->errorscrub_phys.dep_errors = 0;
733 	scn->errorscrub_phys.dep_cursor = 0;
734 	zap_cursor_init_serialized(&scn->errorscrub_cursor,
735 	    spa->spa_meta_objset, spa->spa_errlog_last,
736 	    scn->errorscrub_phys.dep_cursor);
737 
738 	vdev_config_dirty(spa->spa_root_vdev);
739 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_START);
740 
741 	dsl_errorscrub_sync_state(scn, tx);
742 
743 	spa_history_log_internal(spa, "error scrub setup", tx,
744 	    "func=%u mintxg=%u maxtxg=%llu",
745 	    *funcp, 0, (u_longlong_t)tx->tx_txg);
746 }
747 
748 static int
749 dsl_errorscrub_setup_check(void *arg, dmu_tx_t *tx)
750 {
751 	(void) arg;
752 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
753 
754 	if (dsl_scan_is_running(scn) || (dsl_errorscrubbing(scn->scn_dp))) {
755 		return (SET_ERROR(EBUSY));
756 	}
757 
758 	if (spa_get_last_errlog_size(scn->scn_dp->dp_spa) == 0) {
759 		return (ECANCELED);
760 	}
761 	return (0);
762 }
763 
764 /*
765  * Writes out a persistent dsl_scan_phys_t record to the pool directory.
766  * Because we can be running in the block sorting algorithm, we do not always
767  * want to write out the record, only when it is "safe" to do so. This safety
768  * condition is achieved by making sure that the sorting queues are empty
769  * (scn_queues_pending == 0). When this condition is not true, the sync'd state
770  * is inconsistent with how much actual scanning progress has been made. The
771  * kind of sync to be performed is specified by the sync_type argument. If the
772  * sync is optional, we only sync if the queues are empty. If the sync is
773  * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
774  * third possible state is a "cached" sync. This is done in response to:
775  * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
776  *	destroyed, so we wouldn't be able to restart scanning from it.
777  * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
778  *	superseded by a newer snapshot.
779  * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
780  *	swapped with its clone.
781  * In all cases, a cached sync simply rewrites the last record we've written,
782  * just slightly modified. For the modifications that are performed to the
783  * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
784  * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
785  */
786 static void
787 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
788 {
789 	int i;
790 	spa_t *spa = scn->scn_dp->dp_spa;
791 
792 	ASSERT(sync_type != SYNC_MANDATORY || scn->scn_queues_pending == 0);
793 	if (scn->scn_queues_pending == 0) {
794 		for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
795 			vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
796 			dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
797 
798 			if (q == NULL)
799 				continue;
800 
801 			mutex_enter(&vd->vdev_scan_io_queue_lock);
802 			ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
803 			ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
804 			    NULL);
805 			ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
806 			mutex_exit(&vd->vdev_scan_io_queue_lock);
807 		}
808 
809 		if (scn->scn_phys.scn_queue_obj != 0)
810 			scan_ds_queue_sync(scn, tx);
811 		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
812 		    DMU_POOL_DIRECTORY_OBJECT,
813 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
814 		    &scn->scn_phys, tx));
815 		memcpy(&scn->scn_phys_cached, &scn->scn_phys,
816 		    sizeof (scn->scn_phys));
817 
818 		if (scn->scn_checkpointing)
819 			zfs_dbgmsg("finish scan checkpoint for %s",
820 			    spa->spa_name);
821 
822 		scn->scn_checkpointing = B_FALSE;
823 		scn->scn_last_checkpoint = ddi_get_lbolt();
824 	} else if (sync_type == SYNC_CACHED) {
825 		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
826 		    DMU_POOL_DIRECTORY_OBJECT,
827 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
828 		    &scn->scn_phys_cached, tx));
829 	}
830 }
831 
832 int
833 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
834 {
835 	(void) arg;
836 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
837 	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
838 
839 	if (dsl_scan_is_running(scn) || vdev_rebuild_active(rvd) ||
840 	    dsl_errorscrubbing(scn->scn_dp))
841 		return (SET_ERROR(EBUSY));
842 
843 	return (0);
844 }
845 
846 void
847 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
848 {
849 	(void) arg;
850 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
851 	pool_scan_func_t *funcp = arg;
852 	dmu_object_type_t ot = 0;
853 	dsl_pool_t *dp = scn->scn_dp;
854 	spa_t *spa = dp->dp_spa;
855 
856 	ASSERT(!dsl_scan_is_running(scn));
857 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
858 	memset(&scn->scn_phys, 0, sizeof (scn->scn_phys));
859 
860 	/*
861 	 * If we are starting a fresh scrub, we erase the error scrub
862 	 * information from disk.
863 	 */
864 	memset(&scn->errorscrub_phys, 0, sizeof (scn->errorscrub_phys));
865 	dsl_errorscrub_sync_state(scn, tx);
866 
867 	scn->scn_phys.scn_func = *funcp;
868 	scn->scn_phys.scn_state = DSS_SCANNING;
869 	scn->scn_phys.scn_min_txg = 0;
870 	scn->scn_phys.scn_max_txg = tx->tx_txg;
871 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
872 	scn->scn_phys.scn_start_time = gethrestime_sec();
873 	scn->scn_phys.scn_errors = 0;
874 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
875 	scn->scn_issued_before_pass = 0;
876 	scn->scn_restart_txg = 0;
877 	scn->scn_done_txg = 0;
878 	scn->scn_last_checkpoint = 0;
879 	scn->scn_checkpointing = B_FALSE;
880 	spa_scan_stat_init(spa);
881 	vdev_scan_stat_init(spa->spa_root_vdev);
882 
883 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
884 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
885 
886 		/* rewrite all disk labels */
887 		vdev_config_dirty(spa->spa_root_vdev);
888 
889 		if (vdev_resilver_needed(spa->spa_root_vdev,
890 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
891 			nvlist_t *aux = fnvlist_alloc();
892 			fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
893 			    "healing");
894 			spa_event_notify(spa, NULL, aux,
895 			    ESC_ZFS_RESILVER_START);
896 			nvlist_free(aux);
897 		} else {
898 			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
899 		}
900 
901 		spa->spa_scrub_started = B_TRUE;
902 		/*
903 		 * If this is an incremental scrub, limit the DDT scrub phase
904 		 * to just the auto-ditto class (for correctness); the rest
905 		 * of the scrub should go faster using top-down pruning.
906 		 */
907 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
908 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
909 
910 		/*
911 		 * When starting a resilver clear any existing rebuild state.
912 		 * This is required to prevent stale rebuild status from
913 		 * being reported when a rebuild is run, then a resilver and
914 		 * finally a scrub.  In which case only the scrub status
915 		 * should be reported by 'zpool status'.
916 		 */
917 		if (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) {
918 			vdev_t *rvd = spa->spa_root_vdev;
919 			for (uint64_t i = 0; i < rvd->vdev_children; i++) {
920 				vdev_t *vd = rvd->vdev_child[i];
921 				vdev_rebuild_clear_sync(
922 				    (void *)(uintptr_t)vd->vdev_id, tx);
923 			}
924 		}
925 	}
926 
927 	/* back to the generic stuff */
928 
929 	if (zfs_scan_blkstats) {
930 		if (dp->dp_blkstats == NULL) {
931 			dp->dp_blkstats =
932 			    vmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
933 		}
934 		memset(&dp->dp_blkstats->zab_type, 0,
935 		    sizeof (dp->dp_blkstats->zab_type));
936 	} else {
937 		if (dp->dp_blkstats) {
938 			vmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
939 			dp->dp_blkstats = NULL;
940 		}
941 	}
942 
943 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
944 		ot = DMU_OT_ZAP_OTHER;
945 
946 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
947 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
948 
949 	memcpy(&scn->scn_phys_cached, &scn->scn_phys, sizeof (scn->scn_phys));
950 
951 	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
952 
953 	spa_history_log_internal(spa, "scan setup", tx,
954 	    "func=%u mintxg=%llu maxtxg=%llu",
955 	    *funcp, (u_longlong_t)scn->scn_phys.scn_min_txg,
956 	    (u_longlong_t)scn->scn_phys.scn_max_txg);
957 }
958 
959 /*
960  * Called by ZFS_IOC_POOL_SCRUB and ZFS_IOC_POOL_SCAN ioctl to start a scrub,
961  * error scrub or resilver. Can also be called to resume a paused scrub or
962  * error scrub.
963  */
964 int
965 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
966 {
967 	spa_t *spa = dp->dp_spa;
968 	dsl_scan_t *scn = dp->dp_scan;
969 
970 	/*
971 	 * Purge all vdev caches and probe all devices.  We do this here
972 	 * rather than in sync context because this requires a writer lock
973 	 * on the spa_config lock, which we can't do from sync context.  The
974 	 * spa_scrub_reopen flag indicates that vdev_open() should not
975 	 * attempt to start another scrub.
976 	 */
977 	spa_vdev_state_enter(spa, SCL_NONE);
978 	spa->spa_scrub_reopen = B_TRUE;
979 	vdev_reopen(spa->spa_root_vdev);
980 	spa->spa_scrub_reopen = B_FALSE;
981 	(void) spa_vdev_state_exit(spa, NULL, 0);
982 
983 	if (func == POOL_SCAN_RESILVER) {
984 		dsl_scan_restart_resilver(spa->spa_dsl_pool, 0);
985 		return (0);
986 	}
987 
988 	if (func == POOL_SCAN_ERRORSCRUB) {
989 		if (dsl_errorscrub_is_paused(dp->dp_scan)) {
990 			/*
991 			 * got error scrub start cmd, resume paused error scrub.
992 			 */
993 			int err = dsl_scrub_set_pause_resume(scn->scn_dp,
994 			    POOL_SCRUB_NORMAL);
995 			if (err == 0) {
996 				spa_event_notify(spa, NULL, NULL,
997 				    ESC_ZFS_ERRORSCRUB_RESUME);
998 				return (ECANCELED);
999 			}
1000 			return (SET_ERROR(err));
1001 		}
1002 
1003 		return (dsl_sync_task(spa_name(dp->dp_spa),
1004 		    dsl_errorscrub_setup_check, dsl_errorscrub_setup_sync,
1005 		    &func, 0, ZFS_SPACE_CHECK_RESERVED));
1006 	}
1007 
1008 	if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
1009 		/* got scrub start cmd, resume paused scrub */
1010 		int err = dsl_scrub_set_pause_resume(scn->scn_dp,
1011 		    POOL_SCRUB_NORMAL);
1012 		if (err == 0) {
1013 			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
1014 			return (SET_ERROR(ECANCELED));
1015 		}
1016 		return (SET_ERROR(err));
1017 	}
1018 
1019 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1020 	    dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
1021 }
1022 
1023 static void
1024 dsl_errorscrub_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
1025 {
1026 	dsl_pool_t *dp = scn->scn_dp;
1027 	spa_t *spa = dp->dp_spa;
1028 
1029 	if (complete) {
1030 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_FINISH);
1031 		spa_history_log_internal(spa, "error scrub done", tx,
1032 		    "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1033 	} else {
1034 		spa_history_log_internal(spa, "error scrub canceled", tx,
1035 		    "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1036 	}
1037 
1038 	scn->errorscrub_phys.dep_state = complete ? DSS_FINISHED : DSS_CANCELED;
1039 	spa->spa_scrub_active = B_FALSE;
1040 	spa_errlog_rotate(spa);
1041 	scn->errorscrub_phys.dep_end_time = gethrestime_sec();
1042 	zap_cursor_fini(&scn->errorscrub_cursor);
1043 
1044 	if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1045 		spa->spa_errata = 0;
1046 
1047 	ASSERT(!dsl_errorscrubbing(scn->scn_dp));
1048 }
1049 
1050 static void
1051 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
1052 {
1053 	static const char *old_names[] = {
1054 		"scrub_bookmark",
1055 		"scrub_ddt_bookmark",
1056 		"scrub_ddt_class_max",
1057 		"scrub_queue",
1058 		"scrub_min_txg",
1059 		"scrub_max_txg",
1060 		"scrub_func",
1061 		"scrub_errors",
1062 		NULL
1063 	};
1064 
1065 	dsl_pool_t *dp = scn->scn_dp;
1066 	spa_t *spa = dp->dp_spa;
1067 	int i;
1068 
1069 	/* Remove any remnants of an old-style scrub. */
1070 	for (i = 0; old_names[i]; i++) {
1071 		(void) zap_remove(dp->dp_meta_objset,
1072 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
1073 	}
1074 
1075 	if (scn->scn_phys.scn_queue_obj != 0) {
1076 		VERIFY0(dmu_object_free(dp->dp_meta_objset,
1077 		    scn->scn_phys.scn_queue_obj, tx));
1078 		scn->scn_phys.scn_queue_obj = 0;
1079 	}
1080 	scan_ds_queue_clear(scn);
1081 	scan_ds_prefetch_queue_clear(scn);
1082 
1083 	scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1084 
1085 	/*
1086 	 * If we were "restarted" from a stopped state, don't bother
1087 	 * with anything else.
1088 	 */
1089 	if (!dsl_scan_is_running(scn)) {
1090 		ASSERT(!scn->scn_is_sorted);
1091 		return;
1092 	}
1093 
1094 	if (scn->scn_is_sorted) {
1095 		scan_io_queues_destroy(scn);
1096 		scn->scn_is_sorted = B_FALSE;
1097 
1098 		if (scn->scn_taskq != NULL) {
1099 			taskq_destroy(scn->scn_taskq);
1100 			scn->scn_taskq = NULL;
1101 		}
1102 	}
1103 
1104 	scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
1105 
1106 	spa_notify_waiters(spa);
1107 
1108 	if (dsl_scan_restarting(scn, tx))
1109 		spa_history_log_internal(spa, "scan aborted, restarting", tx,
1110 		    "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1111 	else if (!complete)
1112 		spa_history_log_internal(spa, "scan cancelled", tx,
1113 		    "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1114 	else
1115 		spa_history_log_internal(spa, "scan done", tx,
1116 		    "errors=%llu", (u_longlong_t)spa_approx_errlog_size(spa));
1117 
1118 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1119 		spa->spa_scrub_active = B_FALSE;
1120 
1121 		/*
1122 		 * If the scrub/resilver completed, update all DTLs to
1123 		 * reflect this.  Whether it succeeded or not, vacate
1124 		 * all temporary scrub DTLs.
1125 		 *
1126 		 * As the scrub does not currently support traversing
1127 		 * data that have been freed but are part of a checkpoint,
1128 		 * we don't mark the scrub as done in the DTLs as faults
1129 		 * may still exist in those vdevs.
1130 		 */
1131 		if (complete &&
1132 		    !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
1133 			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
1134 			    scn->scn_phys.scn_max_txg, B_TRUE, B_FALSE);
1135 
1136 			if (scn->scn_phys.scn_min_txg) {
1137 				nvlist_t *aux = fnvlist_alloc();
1138 				fnvlist_add_string(aux, ZFS_EV_RESILVER_TYPE,
1139 				    "healing");
1140 				spa_event_notify(spa, NULL, aux,
1141 				    ESC_ZFS_RESILVER_FINISH);
1142 				nvlist_free(aux);
1143 			} else {
1144 				spa_event_notify(spa, NULL, NULL,
1145 				    ESC_ZFS_SCRUB_FINISH);
1146 			}
1147 		} else {
1148 			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
1149 			    0, B_TRUE, B_FALSE);
1150 		}
1151 		spa_errlog_rotate(spa);
1152 
1153 		/*
1154 		 * Don't clear flag until after vdev_dtl_reassess to ensure that
1155 		 * DTL_MISSING will get updated when possible.
1156 		 */
1157 		spa->spa_scrub_started = B_FALSE;
1158 
1159 		/*
1160 		 * We may have finished replacing a device.
1161 		 * Let the async thread assess this and handle the detach.
1162 		 */
1163 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
1164 
1165 		/*
1166 		 * Clear any resilver_deferred flags in the config.
1167 		 * If there are drives that need resilvering, kick
1168 		 * off an asynchronous request to start resilver.
1169 		 * vdev_clear_resilver_deferred() may update the config
1170 		 * before the resilver can restart. In the event of
1171 		 * a crash during this period, the spa loading code
1172 		 * will find the drives that need to be resilvered
1173 		 * and start the resilver then.
1174 		 */
1175 		if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) &&
1176 		    vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) {
1177 			spa_history_log_internal(spa,
1178 			    "starting deferred resilver", tx, "errors=%llu",
1179 			    (u_longlong_t)spa_approx_errlog_size(spa));
1180 			spa_async_request(spa, SPA_ASYNC_RESILVER);
1181 		}
1182 
1183 		/* Clear recent error events (i.e. duplicate events tracking) */
1184 		if (complete)
1185 			zfs_ereport_clear(spa, NULL);
1186 	}
1187 
1188 	scn->scn_phys.scn_end_time = gethrestime_sec();
1189 
1190 	if (spa->spa_errata == ZPOOL_ERRATA_ZOL_2094_SCRUB)
1191 		spa->spa_errata = 0;
1192 
1193 	ASSERT(!dsl_scan_is_running(scn));
1194 }
1195 
1196 static int
1197 dsl_errorscrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1198 {
1199 	pool_scrub_cmd_t *cmd = arg;
1200 	dsl_pool_t *dp = dmu_tx_pool(tx);
1201 	dsl_scan_t *scn = dp->dp_scan;
1202 
1203 	if (*cmd == POOL_SCRUB_PAUSE) {
1204 		/*
1205 		 * can't pause a error scrub when there is no in-progress
1206 		 * error scrub.
1207 		 */
1208 		if (!dsl_errorscrubbing(dp))
1209 			return (SET_ERROR(ENOENT));
1210 
1211 		/* can't pause a paused error scrub */
1212 		if (dsl_errorscrub_is_paused(scn))
1213 			return (SET_ERROR(EBUSY));
1214 	} else if (*cmd != POOL_SCRUB_NORMAL) {
1215 		return (SET_ERROR(ENOTSUP));
1216 	}
1217 
1218 	return (0);
1219 }
1220 
1221 static void
1222 dsl_errorscrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1223 {
1224 	pool_scrub_cmd_t *cmd = arg;
1225 	dsl_pool_t *dp = dmu_tx_pool(tx);
1226 	spa_t *spa = dp->dp_spa;
1227 	dsl_scan_t *scn = dp->dp_scan;
1228 
1229 	if (*cmd == POOL_SCRUB_PAUSE) {
1230 		spa->spa_scan_pass_errorscrub_pause = gethrestime_sec();
1231 		scn->errorscrub_phys.dep_paused_flags = B_TRUE;
1232 		dsl_errorscrub_sync_state(scn, tx);
1233 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_ERRORSCRUB_PAUSED);
1234 	} else {
1235 		ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1236 		if (dsl_errorscrub_is_paused(scn)) {
1237 			/*
1238 			 * We need to keep track of how much time we spend
1239 			 * paused per pass so that we can adjust the error scrub
1240 			 * rate shown in the output of 'zpool status'.
1241 			 */
1242 			spa->spa_scan_pass_errorscrub_spent_paused +=
1243 			    gethrestime_sec() -
1244 			    spa->spa_scan_pass_errorscrub_pause;
1245 
1246 			spa->spa_scan_pass_errorscrub_pause = 0;
1247 			scn->errorscrub_phys.dep_paused_flags = B_FALSE;
1248 
1249 			zap_cursor_init_serialized(
1250 			    &scn->errorscrub_cursor,
1251 			    spa->spa_meta_objset, spa->spa_errlog_last,
1252 			    scn->errorscrub_phys.dep_cursor);
1253 
1254 			dsl_errorscrub_sync_state(scn, tx);
1255 		}
1256 	}
1257 }
1258 
1259 static int
1260 dsl_errorscrub_cancel_check(void *arg, dmu_tx_t *tx)
1261 {
1262 	(void) arg;
1263 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1264 	/* can't cancel a error scrub when there is no one in-progress */
1265 	if (!dsl_errorscrubbing(scn->scn_dp))
1266 		return (SET_ERROR(ENOENT));
1267 	return (0);
1268 }
1269 
1270 static void
1271 dsl_errorscrub_cancel_sync(void *arg, dmu_tx_t *tx)
1272 {
1273 	(void) arg;
1274 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1275 
1276 	dsl_errorscrub_done(scn, B_FALSE, tx);
1277 	dsl_errorscrub_sync_state(scn, tx);
1278 	spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL,
1279 	    ESC_ZFS_ERRORSCRUB_ABORT);
1280 }
1281 
1282 static int
1283 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
1284 {
1285 	(void) arg;
1286 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1287 
1288 	if (!dsl_scan_is_running(scn))
1289 		return (SET_ERROR(ENOENT));
1290 	return (0);
1291 }
1292 
1293 static void
1294 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
1295 {
1296 	(void) arg;
1297 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1298 
1299 	dsl_scan_done(scn, B_FALSE, tx);
1300 	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
1301 	spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
1302 }
1303 
1304 int
1305 dsl_scan_cancel(dsl_pool_t *dp)
1306 {
1307 	if (dsl_errorscrubbing(dp)) {
1308 		return (dsl_sync_task(spa_name(dp->dp_spa),
1309 		    dsl_errorscrub_cancel_check, dsl_errorscrub_cancel_sync,
1310 		    NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1311 	}
1312 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1313 	    dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1314 }
1315 
1316 static int
1317 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1318 {
1319 	pool_scrub_cmd_t *cmd = arg;
1320 	dsl_pool_t *dp = dmu_tx_pool(tx);
1321 	dsl_scan_t *scn = dp->dp_scan;
1322 
1323 	if (*cmd == POOL_SCRUB_PAUSE) {
1324 		/* can't pause a scrub when there is no in-progress scrub */
1325 		if (!dsl_scan_scrubbing(dp))
1326 			return (SET_ERROR(ENOENT));
1327 
1328 		/* can't pause a paused scrub */
1329 		if (dsl_scan_is_paused_scrub(scn))
1330 			return (SET_ERROR(EBUSY));
1331 	} else if (*cmd != POOL_SCRUB_NORMAL) {
1332 		return (SET_ERROR(ENOTSUP));
1333 	}
1334 
1335 	return (0);
1336 }
1337 
1338 static void
1339 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1340 {
1341 	pool_scrub_cmd_t *cmd = arg;
1342 	dsl_pool_t *dp = dmu_tx_pool(tx);
1343 	spa_t *spa = dp->dp_spa;
1344 	dsl_scan_t *scn = dp->dp_scan;
1345 
1346 	if (*cmd == POOL_SCRUB_PAUSE) {
1347 		/* can't pause a scrub when there is no in-progress scrub */
1348 		spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1349 		scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1350 		scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1351 		dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1352 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1353 		spa_notify_waiters(spa);
1354 	} else {
1355 		ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1356 		if (dsl_scan_is_paused_scrub(scn)) {
1357 			/*
1358 			 * We need to keep track of how much time we spend
1359 			 * paused per pass so that we can adjust the scrub rate
1360 			 * shown in the output of 'zpool status'
1361 			 */
1362 			spa->spa_scan_pass_scrub_spent_paused +=
1363 			    gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1364 			spa->spa_scan_pass_scrub_pause = 0;
1365 			scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1366 			scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1367 			dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1368 		}
1369 	}
1370 }
1371 
1372 /*
1373  * Set scrub pause/resume state if it makes sense to do so
1374  */
1375 int
1376 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1377 {
1378 	if (dsl_errorscrubbing(dp)) {
1379 		return (dsl_sync_task(spa_name(dp->dp_spa),
1380 		    dsl_errorscrub_pause_resume_check,
1381 		    dsl_errorscrub_pause_resume_sync, &cmd, 3,
1382 		    ZFS_SPACE_CHECK_RESERVED));
1383 	}
1384 	return (dsl_sync_task(spa_name(dp->dp_spa),
1385 	    dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1386 	    ZFS_SPACE_CHECK_RESERVED));
1387 }
1388 
1389 
1390 /* start a new scan, or restart an existing one. */
1391 void
1392 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg)
1393 {
1394 	if (txg == 0) {
1395 		dmu_tx_t *tx;
1396 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1397 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1398 
1399 		txg = dmu_tx_get_txg(tx);
1400 		dp->dp_scan->scn_restart_txg = txg;
1401 		dmu_tx_commit(tx);
1402 	} else {
1403 		dp->dp_scan->scn_restart_txg = txg;
1404 	}
1405 	zfs_dbgmsg("restarting resilver for %s at txg=%llu",
1406 	    dp->dp_spa->spa_name, (longlong_t)txg);
1407 }
1408 
1409 void
1410 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1411 {
1412 	zio_free(dp->dp_spa, txg, bp);
1413 }
1414 
1415 void
1416 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1417 {
1418 	ASSERT(dsl_pool_sync_context(dp));
1419 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1420 }
1421 
1422 static int
1423 scan_ds_queue_compare(const void *a, const void *b)
1424 {
1425 	const scan_ds_t *sds_a = a, *sds_b = b;
1426 
1427 	if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1428 		return (-1);
1429 	if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1430 		return (0);
1431 	return (1);
1432 }
1433 
1434 static void
1435 scan_ds_queue_clear(dsl_scan_t *scn)
1436 {
1437 	void *cookie = NULL;
1438 	scan_ds_t *sds;
1439 	while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1440 		kmem_free(sds, sizeof (*sds));
1441 	}
1442 }
1443 
1444 static boolean_t
1445 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1446 {
1447 	scan_ds_t srch, *sds;
1448 
1449 	srch.sds_dsobj = dsobj;
1450 	sds = avl_find(&scn->scn_queue, &srch, NULL);
1451 	if (sds != NULL && txg != NULL)
1452 		*txg = sds->sds_txg;
1453 	return (sds != NULL);
1454 }
1455 
1456 static void
1457 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1458 {
1459 	scan_ds_t *sds;
1460 	avl_index_t where;
1461 
1462 	sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1463 	sds->sds_dsobj = dsobj;
1464 	sds->sds_txg = txg;
1465 
1466 	VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1467 	avl_insert(&scn->scn_queue, sds, where);
1468 }
1469 
1470 static void
1471 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1472 {
1473 	scan_ds_t srch, *sds;
1474 
1475 	srch.sds_dsobj = dsobj;
1476 
1477 	sds = avl_find(&scn->scn_queue, &srch, NULL);
1478 	VERIFY(sds != NULL);
1479 	avl_remove(&scn->scn_queue, sds);
1480 	kmem_free(sds, sizeof (*sds));
1481 }
1482 
1483 static void
1484 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1485 {
1486 	dsl_pool_t *dp = scn->scn_dp;
1487 	spa_t *spa = dp->dp_spa;
1488 	dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1489 	    DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1490 
1491 	ASSERT0(scn->scn_queues_pending);
1492 	ASSERT(scn->scn_phys.scn_queue_obj != 0);
1493 
1494 	VERIFY0(dmu_object_free(dp->dp_meta_objset,
1495 	    scn->scn_phys.scn_queue_obj, tx));
1496 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1497 	    DMU_OT_NONE, 0, tx);
1498 	for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1499 	    sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1500 		VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1501 		    scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1502 		    sds->sds_txg, tx));
1503 	}
1504 }
1505 
1506 /*
1507  * Computes the memory limit state that we're currently in. A sorted scan
1508  * needs quite a bit of memory to hold the sorting queue, so we need to
1509  * reasonably constrain the size so it doesn't impact overall system
1510  * performance. We compute two limits:
1511  * 1) Hard memory limit: if the amount of memory used by the sorting
1512  *	queues on a pool gets above this value, we stop the metadata
1513  *	scanning portion and start issuing the queued up and sorted
1514  *	I/Os to reduce memory usage.
1515  *	This limit is calculated as a fraction of physmem (by default 5%).
1516  *	We constrain the lower bound of the hard limit to an absolute
1517  *	minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1518  *	the upper bound to 5% of the total pool size - no chance we'll
1519  *	ever need that much memory, but just to keep the value in check.
1520  * 2) Soft memory limit: once we hit the hard memory limit, we start
1521  *	issuing I/O to reduce queue memory usage, but we don't want to
1522  *	completely empty out the queues, since we might be able to find I/Os
1523  *	that will fill in the gaps of our non-sequential IOs at some point
1524  *	in the future. So we stop the issuing of I/Os once the amount of
1525  *	memory used drops below the soft limit (at which point we stop issuing
1526  *	I/O and start scanning metadata again).
1527  *
1528  *	This limit is calculated by subtracting a fraction of the hard
1529  *	limit from the hard limit. By default this fraction is 5%, so
1530  *	the soft limit is 95% of the hard limit. We cap the size of the
1531  *	difference between the hard and soft limits at an absolute
1532  *	maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1533  *	sufficient to not cause too frequent switching between the
1534  *	metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1535  *	worth of queues is about 1.2 GiB of on-pool data, so scanning
1536  *	that should take at least a decent fraction of a second).
1537  */
1538 static boolean_t
1539 dsl_scan_should_clear(dsl_scan_t *scn)
1540 {
1541 	spa_t *spa = scn->scn_dp->dp_spa;
1542 	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1543 	uint64_t alloc, mlim_hard, mlim_soft, mused;
1544 
1545 	alloc = metaslab_class_get_alloc(spa_normal_class(spa));
1546 	alloc += metaslab_class_get_alloc(spa_special_class(spa));
1547 	alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
1548 
1549 	mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1550 	    zfs_scan_mem_lim_min);
1551 	mlim_hard = MIN(mlim_hard, alloc / 20);
1552 	mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1553 	    zfs_scan_mem_lim_soft_max);
1554 	mused = 0;
1555 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1556 		vdev_t *tvd = rvd->vdev_child[i];
1557 		dsl_scan_io_queue_t *queue;
1558 
1559 		mutex_enter(&tvd->vdev_scan_io_queue_lock);
1560 		queue = tvd->vdev_scan_io_queue;
1561 		if (queue != NULL) {
1562 			/*
1563 			 * # of extents in exts_by_addr = # in exts_by_size.
1564 			 * B-tree efficiency is ~75%, but can be as low as 50%.
1565 			 */
1566 			mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1567 			    ((sizeof (range_seg_gap_t) + sizeof (uint64_t)) *
1568 			    3 / 2) + queue->q_sio_memused;
1569 		}
1570 		mutex_exit(&tvd->vdev_scan_io_queue_lock);
1571 	}
1572 
1573 	dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1574 
1575 	if (mused == 0)
1576 		ASSERT0(scn->scn_queues_pending);
1577 
1578 	/*
1579 	 * If we are above our hard limit, we need to clear out memory.
1580 	 * If we are below our soft limit, we need to accumulate sequential IOs.
1581 	 * Otherwise, we should keep doing whatever we are currently doing.
1582 	 */
1583 	if (mused >= mlim_hard)
1584 		return (B_TRUE);
1585 	else if (mused < mlim_soft)
1586 		return (B_FALSE);
1587 	else
1588 		return (scn->scn_clearing);
1589 }
1590 
1591 static boolean_t
1592 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1593 {
1594 	/* we never skip user/group accounting objects */
1595 	if (zb && (int64_t)zb->zb_object < 0)
1596 		return (B_FALSE);
1597 
1598 	if (scn->scn_suspending)
1599 		return (B_TRUE); /* we're already suspending */
1600 
1601 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1602 		return (B_FALSE); /* we're resuming */
1603 
1604 	/* We only know how to resume from level-0 and objset blocks. */
1605 	if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL))
1606 		return (B_FALSE);
1607 
1608 	/*
1609 	 * We suspend if:
1610 	 *  - we have scanned for at least the minimum time (default 1 sec
1611 	 *    for scrub, 3 sec for resilver), and either we have sufficient
1612 	 *    dirty data that we are starting to write more quickly
1613 	 *    (default 30%), someone is explicitly waiting for this txg
1614 	 *    to complete, or we have used up all of the time in the txg
1615 	 *    timeout (default 5 sec).
1616 	 *  or
1617 	 *  - the spa is shutting down because this pool is being exported
1618 	 *    or the machine is rebooting.
1619 	 *  or
1620 	 *  - the scan queue has reached its memory use limit
1621 	 */
1622 	uint64_t curr_time_ns = gethrtime();
1623 	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1624 	uint64_t sync_time_ns = curr_time_ns -
1625 	    scn->scn_dp->dp_spa->spa_sync_starttime;
1626 	uint64_t dirty_min_bytes = zfs_dirty_data_max *
1627 	    zfs_vdev_async_write_active_min_dirty_percent / 100;
1628 	uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1629 	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1630 
1631 	if ((NSEC2MSEC(scan_time_ns) > mintime &&
1632 	    (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
1633 	    txg_sync_waiting(scn->scn_dp) ||
1634 	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1635 	    spa_shutting_down(scn->scn_dp->dp_spa) ||
1636 	    (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1637 		if (zb && zb->zb_level == ZB_ROOT_LEVEL) {
1638 			dprintf("suspending at first available bookmark "
1639 			    "%llx/%llx/%llx/%llx\n",
1640 			    (longlong_t)zb->zb_objset,
1641 			    (longlong_t)zb->zb_object,
1642 			    (longlong_t)zb->zb_level,
1643 			    (longlong_t)zb->zb_blkid);
1644 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
1645 			    zb->zb_objset, 0, 0, 0);
1646 		} else if (zb != NULL) {
1647 			dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1648 			    (longlong_t)zb->zb_objset,
1649 			    (longlong_t)zb->zb_object,
1650 			    (longlong_t)zb->zb_level,
1651 			    (longlong_t)zb->zb_blkid);
1652 			scn->scn_phys.scn_bookmark = *zb;
1653 		} else {
1654 #ifdef ZFS_DEBUG
1655 			dsl_scan_phys_t *scnp = &scn->scn_phys;
1656 			dprintf("suspending at at DDT bookmark "
1657 			    "%llx/%llx/%llx/%llx\n",
1658 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1659 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1660 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1661 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1662 #endif
1663 		}
1664 		scn->scn_suspending = B_TRUE;
1665 		return (B_TRUE);
1666 	}
1667 	return (B_FALSE);
1668 }
1669 
1670 static boolean_t
1671 dsl_error_scrub_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1672 {
1673 	/*
1674 	 * We suspend if:
1675 	 *  - we have scrubbed for at least the minimum time (default 1 sec
1676 	 *    for error scrub), someone is explicitly waiting for this txg
1677 	 *    to complete, or we have used up all of the time in the txg
1678 	 *    timeout (default 5 sec).
1679 	 *  or
1680 	 *  - the spa is shutting down because this pool is being exported
1681 	 *    or the machine is rebooting.
1682 	 */
1683 	uint64_t curr_time_ns = gethrtime();
1684 	uint64_t error_scrub_time_ns = curr_time_ns - scn->scn_sync_start_time;
1685 	uint64_t sync_time_ns = curr_time_ns -
1686 	    scn->scn_dp->dp_spa->spa_sync_starttime;
1687 	int mintime = zfs_scrub_min_time_ms;
1688 
1689 	if ((NSEC2MSEC(error_scrub_time_ns) > mintime &&
1690 	    (txg_sync_waiting(scn->scn_dp) ||
1691 	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1692 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
1693 		if (zb) {
1694 			dprintf("error scrub suspending at bookmark "
1695 			    "%llx/%llx/%llx/%llx\n",
1696 			    (longlong_t)zb->zb_objset,
1697 			    (longlong_t)zb->zb_object,
1698 			    (longlong_t)zb->zb_level,
1699 			    (longlong_t)zb->zb_blkid);
1700 		}
1701 		return (B_TRUE);
1702 	}
1703 	return (B_FALSE);
1704 }
1705 
1706 typedef struct zil_scan_arg {
1707 	dsl_pool_t	*zsa_dp;
1708 	zil_header_t	*zsa_zh;
1709 } zil_scan_arg_t;
1710 
1711 static int
1712 dsl_scan_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
1713     uint64_t claim_txg)
1714 {
1715 	(void) zilog;
1716 	zil_scan_arg_t *zsa = arg;
1717 	dsl_pool_t *dp = zsa->zsa_dp;
1718 	dsl_scan_t *scn = dp->dp_scan;
1719 	zil_header_t *zh = zsa->zsa_zh;
1720 	zbookmark_phys_t zb;
1721 
1722 	ASSERT(!BP_IS_REDACTED(bp));
1723 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1724 		return (0);
1725 
1726 	/*
1727 	 * One block ("stubby") can be allocated a long time ago; we
1728 	 * want to visit that one because it has been allocated
1729 	 * (on-disk) even if it hasn't been claimed (even though for
1730 	 * scrub there's nothing to do to it).
1731 	 */
1732 	if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1733 		return (0);
1734 
1735 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1736 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1737 
1738 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1739 	return (0);
1740 }
1741 
1742 static int
1743 dsl_scan_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
1744     uint64_t claim_txg)
1745 {
1746 	(void) zilog;
1747 	if (lrc->lrc_txtype == TX_WRITE) {
1748 		zil_scan_arg_t *zsa = arg;
1749 		dsl_pool_t *dp = zsa->zsa_dp;
1750 		dsl_scan_t *scn = dp->dp_scan;
1751 		zil_header_t *zh = zsa->zsa_zh;
1752 		const lr_write_t *lr = (const lr_write_t *)lrc;
1753 		const blkptr_t *bp = &lr->lr_blkptr;
1754 		zbookmark_phys_t zb;
1755 
1756 		ASSERT(!BP_IS_REDACTED(bp));
1757 		if (BP_IS_HOLE(bp) ||
1758 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1759 			return (0);
1760 
1761 		/*
1762 		 * birth can be < claim_txg if this record's txg is
1763 		 * already txg sync'ed (but this log block contains
1764 		 * other records that are not synced)
1765 		 */
1766 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
1767 			return (0);
1768 
1769 		ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
1770 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1771 		    lr->lr_foid, ZB_ZIL_LEVEL,
1772 		    lr->lr_offset / BP_GET_LSIZE(bp));
1773 
1774 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1775 	}
1776 	return (0);
1777 }
1778 
1779 static void
1780 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1781 {
1782 	uint64_t claim_txg = zh->zh_claim_txg;
1783 	zil_scan_arg_t zsa = { dp, zh };
1784 	zilog_t *zilog;
1785 
1786 	ASSERT(spa_writeable(dp->dp_spa));
1787 
1788 	/*
1789 	 * We only want to visit blocks that have been claimed but not yet
1790 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
1791 	 */
1792 	if (claim_txg == 0)
1793 		return;
1794 
1795 	zilog = zil_alloc(dp->dp_meta_objset, zh);
1796 
1797 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1798 	    claim_txg, B_FALSE);
1799 
1800 	zil_free(zilog);
1801 }
1802 
1803 /*
1804  * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1805  * here is to sort the AVL tree by the order each block will be needed.
1806  */
1807 static int
1808 scan_prefetch_queue_compare(const void *a, const void *b)
1809 {
1810 	const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1811 	const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1812 	const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1813 
1814 	return (zbookmark_compare(spc_a->spc_datablkszsec,
1815 	    spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1816 	    spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1817 }
1818 
1819 static void
1820 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, const void *tag)
1821 {
1822 	if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1823 		zfs_refcount_destroy(&spc->spc_refcnt);
1824 		kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1825 	}
1826 }
1827 
1828 static scan_prefetch_ctx_t *
1829 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, const void *tag)
1830 {
1831 	scan_prefetch_ctx_t *spc;
1832 
1833 	spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1834 	zfs_refcount_create(&spc->spc_refcnt);
1835 	zfs_refcount_add(&spc->spc_refcnt, tag);
1836 	spc->spc_scn = scn;
1837 	if (dnp != NULL) {
1838 		spc->spc_datablkszsec = dnp->dn_datablkszsec;
1839 		spc->spc_indblkshift = dnp->dn_indblkshift;
1840 		spc->spc_root = B_FALSE;
1841 	} else {
1842 		spc->spc_datablkszsec = 0;
1843 		spc->spc_indblkshift = 0;
1844 		spc->spc_root = B_TRUE;
1845 	}
1846 
1847 	return (spc);
1848 }
1849 
1850 static void
1851 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, const void *tag)
1852 {
1853 	zfs_refcount_add(&spc->spc_refcnt, tag);
1854 }
1855 
1856 static void
1857 scan_ds_prefetch_queue_clear(dsl_scan_t *scn)
1858 {
1859 	spa_t *spa = scn->scn_dp->dp_spa;
1860 	void *cookie = NULL;
1861 	scan_prefetch_issue_ctx_t *spic = NULL;
1862 
1863 	mutex_enter(&spa->spa_scrub_lock);
1864 	while ((spic = avl_destroy_nodes(&scn->scn_prefetch_queue,
1865 	    &cookie)) != NULL) {
1866 		scan_prefetch_ctx_rele(spic->spic_spc, scn);
1867 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1868 	}
1869 	mutex_exit(&spa->spa_scrub_lock);
1870 }
1871 
1872 static boolean_t
1873 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1874     const zbookmark_phys_t *zb)
1875 {
1876 	zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1877 	dnode_phys_t tmp_dnp;
1878 	dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1879 
1880 	if (zb->zb_objset != last_zb->zb_objset)
1881 		return (B_TRUE);
1882 	if ((int64_t)zb->zb_object < 0)
1883 		return (B_FALSE);
1884 
1885 	tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1886 	tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1887 
1888 	if (zbookmark_subtree_completed(dnp, zb, last_zb))
1889 		return (B_TRUE);
1890 
1891 	return (B_FALSE);
1892 }
1893 
1894 static void
1895 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1896 {
1897 	avl_index_t idx;
1898 	dsl_scan_t *scn = spc->spc_scn;
1899 	spa_t *spa = scn->scn_dp->dp_spa;
1900 	scan_prefetch_issue_ctx_t *spic;
1901 
1902 	if (zfs_no_scrub_prefetch || BP_IS_REDACTED(bp))
1903 		return;
1904 
1905 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1906 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1907 	    BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1908 		return;
1909 
1910 	if (dsl_scan_check_prefetch_resume(spc, zb))
1911 		return;
1912 
1913 	scan_prefetch_ctx_add_ref(spc, scn);
1914 	spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1915 	spic->spic_spc = spc;
1916 	spic->spic_bp = *bp;
1917 	spic->spic_zb = *zb;
1918 
1919 	/*
1920 	 * Add the IO to the queue of blocks to prefetch. This allows us to
1921 	 * prioritize blocks that we will need first for the main traversal
1922 	 * thread.
1923 	 */
1924 	mutex_enter(&spa->spa_scrub_lock);
1925 	if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1926 		/* this block is already queued for prefetch */
1927 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1928 		scan_prefetch_ctx_rele(spc, scn);
1929 		mutex_exit(&spa->spa_scrub_lock);
1930 		return;
1931 	}
1932 
1933 	avl_insert(&scn->scn_prefetch_queue, spic, idx);
1934 	cv_broadcast(&spa->spa_scrub_io_cv);
1935 	mutex_exit(&spa->spa_scrub_lock);
1936 }
1937 
1938 static void
1939 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1940     uint64_t objset, uint64_t object)
1941 {
1942 	int i;
1943 	zbookmark_phys_t zb;
1944 	scan_prefetch_ctx_t *spc;
1945 
1946 	if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1947 		return;
1948 
1949 	SET_BOOKMARK(&zb, objset, object, 0, 0);
1950 
1951 	spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1952 
1953 	for (i = 0; i < dnp->dn_nblkptr; i++) {
1954 		zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1955 		zb.zb_blkid = i;
1956 		dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1957 	}
1958 
1959 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1960 		zb.zb_level = 0;
1961 		zb.zb_blkid = DMU_SPILL_BLKID;
1962 		dsl_scan_prefetch(spc, DN_SPILL_BLKPTR(dnp), &zb);
1963 	}
1964 
1965 	scan_prefetch_ctx_rele(spc, FTAG);
1966 }
1967 
1968 static void
1969 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1970     arc_buf_t *buf, void *private)
1971 {
1972 	(void) zio;
1973 	scan_prefetch_ctx_t *spc = private;
1974 	dsl_scan_t *scn = spc->spc_scn;
1975 	spa_t *spa = scn->scn_dp->dp_spa;
1976 
1977 	/* broadcast that the IO has completed for rate limiting purposes */
1978 	mutex_enter(&spa->spa_scrub_lock);
1979 	ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1980 	spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1981 	cv_broadcast(&spa->spa_scrub_io_cv);
1982 	mutex_exit(&spa->spa_scrub_lock);
1983 
1984 	/* if there was an error or we are done prefetching, just cleanup */
1985 	if (buf == NULL || scn->scn_prefetch_stop)
1986 		goto out;
1987 
1988 	if (BP_GET_LEVEL(bp) > 0) {
1989 		int i;
1990 		blkptr_t *cbp;
1991 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1992 		zbookmark_phys_t czb;
1993 
1994 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1995 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1996 			    zb->zb_level - 1, zb->zb_blkid * epb + i);
1997 			dsl_scan_prefetch(spc, cbp, &czb);
1998 		}
1999 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
2000 		dnode_phys_t *cdnp;
2001 		int i;
2002 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
2003 
2004 		for (i = 0, cdnp = buf->b_data; i < epb;
2005 		    i += cdnp->dn_extra_slots + 1,
2006 		    cdnp += cdnp->dn_extra_slots + 1) {
2007 			dsl_scan_prefetch_dnode(scn, cdnp,
2008 			    zb->zb_objset, zb->zb_blkid * epb + i);
2009 		}
2010 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
2011 		objset_phys_t *osp = buf->b_data;
2012 
2013 		dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
2014 		    zb->zb_objset, DMU_META_DNODE_OBJECT);
2015 
2016 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
2017 			dsl_scan_prefetch_dnode(scn,
2018 			    &osp->os_groupused_dnode, zb->zb_objset,
2019 			    DMU_GROUPUSED_OBJECT);
2020 			dsl_scan_prefetch_dnode(scn,
2021 			    &osp->os_userused_dnode, zb->zb_objset,
2022 			    DMU_USERUSED_OBJECT);
2023 		}
2024 	}
2025 
2026 out:
2027 	if (buf != NULL)
2028 		arc_buf_destroy(buf, private);
2029 	scan_prefetch_ctx_rele(spc, scn);
2030 }
2031 
2032 static void
2033 dsl_scan_prefetch_thread(void *arg)
2034 {
2035 	dsl_scan_t *scn = arg;
2036 	spa_t *spa = scn->scn_dp->dp_spa;
2037 	scan_prefetch_issue_ctx_t *spic;
2038 
2039 	/* loop until we are told to stop */
2040 	while (!scn->scn_prefetch_stop) {
2041 		arc_flags_t flags = ARC_FLAG_NOWAIT |
2042 		    ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
2043 		int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
2044 
2045 		mutex_enter(&spa->spa_scrub_lock);
2046 
2047 		/*
2048 		 * Wait until we have an IO to issue and are not above our
2049 		 * maximum in flight limit.
2050 		 */
2051 		while (!scn->scn_prefetch_stop &&
2052 		    (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
2053 		    spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
2054 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2055 		}
2056 
2057 		/* recheck if we should stop since we waited for the cv */
2058 		if (scn->scn_prefetch_stop) {
2059 			mutex_exit(&spa->spa_scrub_lock);
2060 			break;
2061 		}
2062 
2063 		/* remove the prefetch IO from the tree */
2064 		spic = avl_first(&scn->scn_prefetch_queue);
2065 		spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
2066 		avl_remove(&scn->scn_prefetch_queue, spic);
2067 
2068 		mutex_exit(&spa->spa_scrub_lock);
2069 
2070 		if (BP_IS_PROTECTED(&spic->spic_bp)) {
2071 			ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
2072 			    BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
2073 			ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
2074 			zio_flags |= ZIO_FLAG_RAW;
2075 		}
2076 
2077 		/* issue the prefetch asynchronously */
2078 		(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
2079 		    &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
2080 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
2081 
2082 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
2083 	}
2084 
2085 	ASSERT(scn->scn_prefetch_stop);
2086 
2087 	/* free any prefetches we didn't get to complete */
2088 	mutex_enter(&spa->spa_scrub_lock);
2089 	while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
2090 		avl_remove(&scn->scn_prefetch_queue, spic);
2091 		scan_prefetch_ctx_rele(spic->spic_spc, scn);
2092 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
2093 	}
2094 	ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
2095 	mutex_exit(&spa->spa_scrub_lock);
2096 }
2097 
2098 static boolean_t
2099 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
2100     const zbookmark_phys_t *zb)
2101 {
2102 	/*
2103 	 * We never skip over user/group accounting objects (obj<0)
2104 	 */
2105 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
2106 	    (int64_t)zb->zb_object >= 0) {
2107 		/*
2108 		 * If we already visited this bp & everything below (in
2109 		 * a prior txg sync), don't bother doing it again.
2110 		 */
2111 		if (zbookmark_subtree_completed(dnp, zb,
2112 		    &scn->scn_phys.scn_bookmark))
2113 			return (B_TRUE);
2114 
2115 		/*
2116 		 * If we found the block we're trying to resume from, or
2117 		 * we went past it, zero it out to indicate that it's OK
2118 		 * to start checking for suspending again.
2119 		 */
2120 		if (zbookmark_subtree_tbd(dnp, zb,
2121 		    &scn->scn_phys.scn_bookmark)) {
2122 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
2123 			    (longlong_t)zb->zb_objset,
2124 			    (longlong_t)zb->zb_object,
2125 			    (longlong_t)zb->zb_level,
2126 			    (longlong_t)zb->zb_blkid);
2127 			memset(&scn->scn_phys.scn_bookmark, 0, sizeof (*zb));
2128 		}
2129 	}
2130 	return (B_FALSE);
2131 }
2132 
2133 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
2134     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
2135     dmu_objset_type_t ostype, dmu_tx_t *tx);
2136 inline __attribute__((always_inline)) static void dsl_scan_visitdnode(
2137     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
2138     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
2139 
2140 /*
2141  * Return nonzero on i/o error.
2142  * Return new buf to write out in *bufp.
2143  */
2144 inline __attribute__((always_inline)) static int
2145 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
2146     dnode_phys_t *dnp, const blkptr_t *bp,
2147     const zbookmark_phys_t *zb, dmu_tx_t *tx)
2148 {
2149 	dsl_pool_t *dp = scn->scn_dp;
2150 	spa_t *spa = dp->dp_spa;
2151 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
2152 	int err;
2153 
2154 	ASSERT(!BP_IS_REDACTED(bp));
2155 
2156 	/*
2157 	 * There is an unlikely case of encountering dnodes with contradicting
2158 	 * dn_bonuslen and DNODE_FLAG_SPILL_BLKPTR flag before in files created
2159 	 * or modified before commit 4254acb was merged. As it is not possible
2160 	 * to know which of the two is correct, report an error.
2161 	 */
2162 	if (dnp != NULL &&
2163 	    dnp->dn_bonuslen > DN_MAX_BONUS_LEN(dnp)) {
2164 		scn->scn_phys.scn_errors++;
2165 		spa_log_error(spa, zb, &bp->blk_birth);
2166 		return (SET_ERROR(EINVAL));
2167 	}
2168 
2169 	if (BP_GET_LEVEL(bp) > 0) {
2170 		arc_flags_t flags = ARC_FLAG_WAIT;
2171 		int i;
2172 		blkptr_t *cbp;
2173 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
2174 		arc_buf_t *buf;
2175 
2176 		err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2177 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2178 		if (err) {
2179 			scn->scn_phys.scn_errors++;
2180 			return (err);
2181 		}
2182 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
2183 			zbookmark_phys_t czb;
2184 
2185 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
2186 			    zb->zb_level - 1,
2187 			    zb->zb_blkid * epb + i);
2188 			dsl_scan_visitbp(cbp, &czb, dnp,
2189 			    ds, scn, ostype, tx);
2190 		}
2191 		arc_buf_destroy(buf, &buf);
2192 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
2193 		arc_flags_t flags = ARC_FLAG_WAIT;
2194 		dnode_phys_t *cdnp;
2195 		int i;
2196 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
2197 		arc_buf_t *buf;
2198 
2199 		if (BP_IS_PROTECTED(bp)) {
2200 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
2201 			zio_flags |= ZIO_FLAG_RAW;
2202 		}
2203 
2204 		err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2205 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2206 		if (err) {
2207 			scn->scn_phys.scn_errors++;
2208 			return (err);
2209 		}
2210 		for (i = 0, cdnp = buf->b_data; i < epb;
2211 		    i += cdnp->dn_extra_slots + 1,
2212 		    cdnp += cdnp->dn_extra_slots + 1) {
2213 			dsl_scan_visitdnode(scn, ds, ostype,
2214 			    cdnp, zb->zb_blkid * epb + i, tx);
2215 		}
2216 
2217 		arc_buf_destroy(buf, &buf);
2218 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
2219 		arc_flags_t flags = ARC_FLAG_WAIT;
2220 		objset_phys_t *osp;
2221 		arc_buf_t *buf;
2222 
2223 		err = arc_read(NULL, spa, bp, arc_getbuf_func, &buf,
2224 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
2225 		if (err) {
2226 			scn->scn_phys.scn_errors++;
2227 			return (err);
2228 		}
2229 
2230 		osp = buf->b_data;
2231 
2232 		dsl_scan_visitdnode(scn, ds, osp->os_type,
2233 		    &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
2234 
2235 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
2236 			/*
2237 			 * We also always visit user/group/project accounting
2238 			 * objects, and never skip them, even if we are
2239 			 * suspending. This is necessary so that the
2240 			 * space deltas from this txg get integrated.
2241 			 */
2242 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
2243 				dsl_scan_visitdnode(scn, ds, osp->os_type,
2244 				    &osp->os_projectused_dnode,
2245 				    DMU_PROJECTUSED_OBJECT, tx);
2246 			dsl_scan_visitdnode(scn, ds, osp->os_type,
2247 			    &osp->os_groupused_dnode,
2248 			    DMU_GROUPUSED_OBJECT, tx);
2249 			dsl_scan_visitdnode(scn, ds, osp->os_type,
2250 			    &osp->os_userused_dnode,
2251 			    DMU_USERUSED_OBJECT, tx);
2252 		}
2253 		arc_buf_destroy(buf, &buf);
2254 	} else if (!zfs_blkptr_verify(spa, bp,
2255 	    BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
2256 		/*
2257 		 * Sanity check the block pointer contents, this is handled
2258 		 * by arc_read() for the cases above.
2259 		 */
2260 		scn->scn_phys.scn_errors++;
2261 		spa_log_error(spa, zb, &bp->blk_birth);
2262 		return (SET_ERROR(EINVAL));
2263 	}
2264 
2265 	return (0);
2266 }
2267 
2268 inline __attribute__((always_inline)) static void
2269 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
2270     dmu_objset_type_t ostype, dnode_phys_t *dnp,
2271     uint64_t object, dmu_tx_t *tx)
2272 {
2273 	int j;
2274 
2275 	for (j = 0; j < dnp->dn_nblkptr; j++) {
2276 		zbookmark_phys_t czb;
2277 
2278 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
2279 		    dnp->dn_nlevels - 1, j);
2280 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
2281 		    &czb, dnp, ds, scn, ostype, tx);
2282 	}
2283 
2284 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
2285 		zbookmark_phys_t czb;
2286 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
2287 		    0, DMU_SPILL_BLKID);
2288 		dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
2289 		    &czb, dnp, ds, scn, ostype, tx);
2290 	}
2291 }
2292 
2293 /*
2294  * The arguments are in this order because mdb can only print the
2295  * first 5; we want them to be useful.
2296  */
2297 static void
2298 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
2299     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
2300     dmu_objset_type_t ostype, dmu_tx_t *tx)
2301 {
2302 	dsl_pool_t *dp = scn->scn_dp;
2303 	blkptr_t *bp_toread = NULL;
2304 
2305 	if (dsl_scan_check_suspend(scn, zb))
2306 		return;
2307 
2308 	if (dsl_scan_check_resume(scn, dnp, zb))
2309 		return;
2310 
2311 	scn->scn_visited_this_txg++;
2312 
2313 	if (BP_IS_HOLE(bp)) {
2314 		scn->scn_holes_this_txg++;
2315 		return;
2316 	}
2317 
2318 	if (BP_IS_REDACTED(bp)) {
2319 		ASSERT(dsl_dataset_feature_is_active(ds,
2320 		    SPA_FEATURE_REDACTED_DATASETS));
2321 		return;
2322 	}
2323 
2324 	/*
2325 	 * Check if this block contradicts any filesystem flags.
2326 	 */
2327 	spa_feature_t f = SPA_FEATURE_LARGE_BLOCKS;
2328 	if (BP_GET_LSIZE(bp) > SPA_OLD_MAXBLOCKSIZE)
2329 		ASSERT(dsl_dataset_feature_is_active(ds, f));
2330 
2331 	f = zio_checksum_to_feature(BP_GET_CHECKSUM(bp));
2332 	if (f != SPA_FEATURE_NONE)
2333 		ASSERT(dsl_dataset_feature_is_active(ds, f));
2334 
2335 	f = zio_compress_to_feature(BP_GET_COMPRESS(bp));
2336 	if (f != SPA_FEATURE_NONE)
2337 		ASSERT(dsl_dataset_feature_is_active(ds, f));
2338 
2339 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
2340 		scn->scn_lt_min_this_txg++;
2341 		return;
2342 	}
2343 
2344 	bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
2345 	*bp_toread = *bp;
2346 
2347 	if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
2348 		goto out;
2349 
2350 	/*
2351 	 * If dsl_scan_ddt() has already visited this block, it will have
2352 	 * already done any translations or scrubbing, so don't call the
2353 	 * callback again.
2354 	 */
2355 	if (ddt_class_contains(dp->dp_spa,
2356 	    scn->scn_phys.scn_ddt_class_max, bp)) {
2357 		scn->scn_ddt_contained_this_txg++;
2358 		goto out;
2359 	}
2360 
2361 	/*
2362 	 * If this block is from the future (after cur_max_txg), then we
2363 	 * are doing this on behalf of a deleted snapshot, and we will
2364 	 * revisit the future block on the next pass of this dataset.
2365 	 * Don't scan it now unless we need to because something
2366 	 * under it was modified.
2367 	 */
2368 	if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
2369 		scn->scn_gt_max_this_txg++;
2370 		goto out;
2371 	}
2372 
2373 	scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
2374 
2375 out:
2376 	kmem_free(bp_toread, sizeof (blkptr_t));
2377 }
2378 
2379 static void
2380 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
2381     dmu_tx_t *tx)
2382 {
2383 	zbookmark_phys_t zb;
2384 	scan_prefetch_ctx_t *spc;
2385 
2386 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2387 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
2388 
2389 	if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
2390 		SET_BOOKMARK(&scn->scn_prefetch_bookmark,
2391 		    zb.zb_objset, 0, 0, 0);
2392 	} else {
2393 		scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
2394 	}
2395 
2396 	scn->scn_objsets_visited_this_txg++;
2397 
2398 	spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
2399 	dsl_scan_prefetch(spc, bp, &zb);
2400 	scan_prefetch_ctx_rele(spc, FTAG);
2401 
2402 	dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
2403 
2404 	dprintf_ds(ds, "finished scan%s", "");
2405 }
2406 
2407 static void
2408 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
2409 {
2410 	if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
2411 		if (ds->ds_is_snapshot) {
2412 			/*
2413 			 * Note:
2414 			 *  - scn_cur_{min,max}_txg stays the same.
2415 			 *  - Setting the flag is not really necessary if
2416 			 *    scn_cur_max_txg == scn_max_txg, because there
2417 			 *    is nothing after this snapshot that we care
2418 			 *    about.  However, we set it anyway and then
2419 			 *    ignore it when we retraverse it in
2420 			 *    dsl_scan_visitds().
2421 			 */
2422 			scn_phys->scn_bookmark.zb_objset =
2423 			    dsl_dataset_phys(ds)->ds_next_snap_obj;
2424 			zfs_dbgmsg("destroying ds %llu on %s; currently "
2425 			    "traversing; reset zb_objset to %llu",
2426 			    (u_longlong_t)ds->ds_object,
2427 			    ds->ds_dir->dd_pool->dp_spa->spa_name,
2428 			    (u_longlong_t)dsl_dataset_phys(ds)->
2429 			    ds_next_snap_obj);
2430 			scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2431 		} else {
2432 			SET_BOOKMARK(&scn_phys->scn_bookmark,
2433 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
2434 			zfs_dbgmsg("destroying ds %llu on %s; currently "
2435 			    "traversing; reset bookmark to -1,0,0,0",
2436 			    (u_longlong_t)ds->ds_object,
2437 			    ds->ds_dir->dd_pool->dp_spa->spa_name);
2438 		}
2439 	}
2440 }
2441 
2442 /*
2443  * Invoked when a dataset is destroyed. We need to make sure that:
2444  *
2445  * 1) If it is the dataset that was currently being scanned, we write
2446  *	a new dsl_scan_phys_t and marking the objset reference in it
2447  *	as destroyed.
2448  * 2) Remove it from the work queue, if it was present.
2449  *
2450  * If the dataset was actually a snapshot, instead of marking the dataset
2451  * as destroyed, we instead substitute the next snapshot in line.
2452  */
2453 void
2454 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2455 {
2456 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2457 	dsl_scan_t *scn = dp->dp_scan;
2458 	uint64_t mintxg;
2459 
2460 	if (!dsl_scan_is_running(scn))
2461 		return;
2462 
2463 	ds_destroyed_scn_phys(ds, &scn->scn_phys);
2464 	ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2465 
2466 	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2467 		scan_ds_queue_remove(scn, ds->ds_object);
2468 		if (ds->ds_is_snapshot)
2469 			scan_ds_queue_insert(scn,
2470 			    dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2471 	}
2472 
2473 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2474 	    ds->ds_object, &mintxg) == 0) {
2475 		ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2476 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2477 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2478 		if (ds->ds_is_snapshot) {
2479 			/*
2480 			 * We keep the same mintxg; it could be >
2481 			 * ds_creation_txg if the previous snapshot was
2482 			 * deleted too.
2483 			 */
2484 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
2485 			    scn->scn_phys.scn_queue_obj,
2486 			    dsl_dataset_phys(ds)->ds_next_snap_obj,
2487 			    mintxg, tx) == 0);
2488 			zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2489 			    "replacing with %llu",
2490 			    (u_longlong_t)ds->ds_object,
2491 			    dp->dp_spa->spa_name,
2492 			    (u_longlong_t)dsl_dataset_phys(ds)->
2493 			    ds_next_snap_obj);
2494 		} else {
2495 			zfs_dbgmsg("destroying ds %llu on %s; in queue; "
2496 			    "removing",
2497 			    (u_longlong_t)ds->ds_object,
2498 			    dp->dp_spa->spa_name);
2499 		}
2500 	}
2501 
2502 	/*
2503 	 * dsl_scan_sync() should be called after this, and should sync
2504 	 * out our changed state, but just to be safe, do it here.
2505 	 */
2506 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2507 }
2508 
2509 static void
2510 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2511 {
2512 	if (scn_bookmark->zb_objset == ds->ds_object) {
2513 		scn_bookmark->zb_objset =
2514 		    dsl_dataset_phys(ds)->ds_prev_snap_obj;
2515 		zfs_dbgmsg("snapshotting ds %llu on %s; currently traversing; "
2516 		    "reset zb_objset to %llu",
2517 		    (u_longlong_t)ds->ds_object,
2518 		    ds->ds_dir->dd_pool->dp_spa->spa_name,
2519 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2520 	}
2521 }
2522 
2523 /*
2524  * Called when a dataset is snapshotted. If we were currently traversing
2525  * this snapshot, we reset our bookmark to point at the newly created
2526  * snapshot. We also modify our work queue to remove the old snapshot and
2527  * replace with the new one.
2528  */
2529 void
2530 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2531 {
2532 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2533 	dsl_scan_t *scn = dp->dp_scan;
2534 	uint64_t mintxg;
2535 
2536 	if (!dsl_scan_is_running(scn))
2537 		return;
2538 
2539 	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2540 
2541 	ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2542 	ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2543 
2544 	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2545 		scan_ds_queue_remove(scn, ds->ds_object);
2546 		scan_ds_queue_insert(scn,
2547 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2548 	}
2549 
2550 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2551 	    ds->ds_object, &mintxg) == 0) {
2552 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2553 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2554 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
2555 		    scn->scn_phys.scn_queue_obj,
2556 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2557 		zfs_dbgmsg("snapshotting ds %llu on %s; in queue; "
2558 		    "replacing with %llu",
2559 		    (u_longlong_t)ds->ds_object,
2560 		    dp->dp_spa->spa_name,
2561 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2562 	}
2563 
2564 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2565 }
2566 
2567 static void
2568 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2569     zbookmark_phys_t *scn_bookmark)
2570 {
2571 	if (scn_bookmark->zb_objset == ds1->ds_object) {
2572 		scn_bookmark->zb_objset = ds2->ds_object;
2573 		zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2574 		    "reset zb_objset to %llu",
2575 		    (u_longlong_t)ds1->ds_object,
2576 		    ds1->ds_dir->dd_pool->dp_spa->spa_name,
2577 		    (u_longlong_t)ds2->ds_object);
2578 	} else if (scn_bookmark->zb_objset == ds2->ds_object) {
2579 		scn_bookmark->zb_objset = ds1->ds_object;
2580 		zfs_dbgmsg("clone_swap ds %llu on %s; currently traversing; "
2581 		    "reset zb_objset to %llu",
2582 		    (u_longlong_t)ds2->ds_object,
2583 		    ds2->ds_dir->dd_pool->dp_spa->spa_name,
2584 		    (u_longlong_t)ds1->ds_object);
2585 	}
2586 }
2587 
2588 /*
2589  * Called when an origin dataset and its clone are swapped.  If we were
2590  * currently traversing the dataset, we need to switch to traversing the
2591  * newly promoted clone.
2592  */
2593 void
2594 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2595 {
2596 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2597 	dsl_scan_t *scn = dp->dp_scan;
2598 	uint64_t mintxg1, mintxg2;
2599 	boolean_t ds1_queued, ds2_queued;
2600 
2601 	if (!dsl_scan_is_running(scn))
2602 		return;
2603 
2604 	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2605 	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2606 
2607 	/*
2608 	 * Handle the in-memory scan queue.
2609 	 */
2610 	ds1_queued = scan_ds_queue_contains(scn, ds1->ds_object, &mintxg1);
2611 	ds2_queued = scan_ds_queue_contains(scn, ds2->ds_object, &mintxg2);
2612 
2613 	/* Sanity checking. */
2614 	if (ds1_queued) {
2615 		ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2616 		ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2617 	}
2618 	if (ds2_queued) {
2619 		ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2620 		ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2621 	}
2622 
2623 	if (ds1_queued && ds2_queued) {
2624 		/*
2625 		 * If both are queued, we don't need to do anything.
2626 		 * The swapping code below would not handle this case correctly,
2627 		 * since we can't insert ds2 if it is already there. That's
2628 		 * because scan_ds_queue_insert() prohibits a duplicate insert
2629 		 * and panics.
2630 		 */
2631 	} else if (ds1_queued) {
2632 		scan_ds_queue_remove(scn, ds1->ds_object);
2633 		scan_ds_queue_insert(scn, ds2->ds_object, mintxg1);
2634 	} else if (ds2_queued) {
2635 		scan_ds_queue_remove(scn, ds2->ds_object);
2636 		scan_ds_queue_insert(scn, ds1->ds_object, mintxg2);
2637 	}
2638 
2639 	/*
2640 	 * Handle the on-disk scan queue.
2641 	 * The on-disk state is an out-of-date version of the in-memory state,
2642 	 * so the in-memory and on-disk values for ds1_queued and ds2_queued may
2643 	 * be different. Therefore we need to apply the swap logic to the
2644 	 * on-disk state independently of the in-memory state.
2645 	 */
2646 	ds1_queued = zap_lookup_int_key(dp->dp_meta_objset,
2647 	    scn->scn_phys.scn_queue_obj, ds1->ds_object, &mintxg1) == 0;
2648 	ds2_queued = zap_lookup_int_key(dp->dp_meta_objset,
2649 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg2) == 0;
2650 
2651 	/* Sanity checking. */
2652 	if (ds1_queued) {
2653 		ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2654 		ASSERT3U(mintxg1, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2655 	}
2656 	if (ds2_queued) {
2657 		ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2658 		ASSERT3U(mintxg2, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2659 	}
2660 
2661 	if (ds1_queued && ds2_queued) {
2662 		/*
2663 		 * If both are queued, we don't need to do anything.
2664 		 * Alternatively, we could check for EEXIST from
2665 		 * zap_add_int_key() and back out to the original state, but
2666 		 * that would be more work than checking for this case upfront.
2667 		 */
2668 	} else if (ds1_queued) {
2669 		VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2670 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2671 		VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2672 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg1, tx));
2673 		zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2674 		    "replacing with %llu",
2675 		    (u_longlong_t)ds1->ds_object,
2676 		    dp->dp_spa->spa_name,
2677 		    (u_longlong_t)ds2->ds_object);
2678 	} else if (ds2_queued) {
2679 		VERIFY3S(0, ==, zap_remove_int(dp->dp_meta_objset,
2680 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2681 		VERIFY3S(0, ==, zap_add_int_key(dp->dp_meta_objset,
2682 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg2, tx));
2683 		zfs_dbgmsg("clone_swap ds %llu on %s; in queue; "
2684 		    "replacing with %llu",
2685 		    (u_longlong_t)ds2->ds_object,
2686 		    dp->dp_spa->spa_name,
2687 		    (u_longlong_t)ds1->ds_object);
2688 	}
2689 
2690 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2691 }
2692 
2693 static int
2694 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2695 {
2696 	uint64_t originobj = *(uint64_t *)arg;
2697 	dsl_dataset_t *ds;
2698 	int err;
2699 	dsl_scan_t *scn = dp->dp_scan;
2700 
2701 	if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2702 		return (0);
2703 
2704 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2705 	if (err)
2706 		return (err);
2707 
2708 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2709 		dsl_dataset_t *prev;
2710 		err = dsl_dataset_hold_obj(dp,
2711 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2712 
2713 		dsl_dataset_rele(ds, FTAG);
2714 		if (err)
2715 			return (err);
2716 		ds = prev;
2717 	}
2718 	scan_ds_queue_insert(scn, ds->ds_object,
2719 	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2720 	dsl_dataset_rele(ds, FTAG);
2721 	return (0);
2722 }
2723 
2724 static void
2725 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2726 {
2727 	dsl_pool_t *dp = scn->scn_dp;
2728 	dsl_dataset_t *ds;
2729 
2730 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2731 
2732 	if (scn->scn_phys.scn_cur_min_txg >=
2733 	    scn->scn_phys.scn_max_txg) {
2734 		/*
2735 		 * This can happen if this snapshot was created after the
2736 		 * scan started, and we already completed a previous snapshot
2737 		 * that was created after the scan started.  This snapshot
2738 		 * only references blocks with:
2739 		 *
2740 		 *	birth < our ds_creation_txg
2741 		 *	cur_min_txg is no less than ds_creation_txg.
2742 		 *	We have already visited these blocks.
2743 		 * or
2744 		 *	birth > scn_max_txg
2745 		 *	The scan requested not to visit these blocks.
2746 		 *
2747 		 * Subsequent snapshots (and clones) can reference our
2748 		 * blocks, or blocks with even higher birth times.
2749 		 * Therefore we do not need to visit them either,
2750 		 * so we do not add them to the work queue.
2751 		 *
2752 		 * Note that checking for cur_min_txg >= cur_max_txg
2753 		 * is not sufficient, because in that case we may need to
2754 		 * visit subsequent snapshots.  This happens when min_txg > 0,
2755 		 * which raises cur_min_txg.  In this case we will visit
2756 		 * this dataset but skip all of its blocks, because the
2757 		 * rootbp's birth time is < cur_min_txg.  Then we will
2758 		 * add the next snapshots/clones to the work queue.
2759 		 */
2760 		char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2761 		dsl_dataset_name(ds, dsname);
2762 		zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2763 		    "cur_min_txg (%llu) >= max_txg (%llu)",
2764 		    (longlong_t)dsobj, dsname,
2765 		    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2766 		    (longlong_t)scn->scn_phys.scn_max_txg);
2767 		kmem_free(dsname, MAXNAMELEN);
2768 
2769 		goto out;
2770 	}
2771 
2772 	/*
2773 	 * Only the ZIL in the head (non-snapshot) is valid. Even though
2774 	 * snapshots can have ZIL block pointers (which may be the same
2775 	 * BP as in the head), they must be ignored. In addition, $ORIGIN
2776 	 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2777 	 * need to look for a ZIL in it either. So we traverse the ZIL here,
2778 	 * rather than in scan_recurse(), because the regular snapshot
2779 	 * block-sharing rules don't apply to it.
2780 	 */
2781 	if (!dsl_dataset_is_snapshot(ds) &&
2782 	    (dp->dp_origin_snap == NULL ||
2783 	    ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2784 		objset_t *os;
2785 		if (dmu_objset_from_ds(ds, &os) != 0) {
2786 			goto out;
2787 		}
2788 		dsl_scan_zil(dp, &os->os_zil_header);
2789 	}
2790 
2791 	/*
2792 	 * Iterate over the bps in this ds.
2793 	 */
2794 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2795 	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2796 	dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2797 	rrw_exit(&ds->ds_bp_rwlock, FTAG);
2798 
2799 	char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2800 	dsl_dataset_name(ds, dsname);
2801 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2802 	    "suspending=%u",
2803 	    (longlong_t)dsobj, dsname,
2804 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2805 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
2806 	    (int)scn->scn_suspending);
2807 	kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2808 
2809 	if (scn->scn_suspending)
2810 		goto out;
2811 
2812 	/*
2813 	 * We've finished this pass over this dataset.
2814 	 */
2815 
2816 	/*
2817 	 * If we did not completely visit this dataset, do another pass.
2818 	 */
2819 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2820 		zfs_dbgmsg("incomplete pass on %s; visiting again",
2821 		    dp->dp_spa->spa_name);
2822 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2823 		scan_ds_queue_insert(scn, ds->ds_object,
2824 		    scn->scn_phys.scn_cur_max_txg);
2825 		goto out;
2826 	}
2827 
2828 	/*
2829 	 * Add descendant datasets to work queue.
2830 	 */
2831 	if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2832 		scan_ds_queue_insert(scn,
2833 		    dsl_dataset_phys(ds)->ds_next_snap_obj,
2834 		    dsl_dataset_phys(ds)->ds_creation_txg);
2835 	}
2836 	if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2837 		boolean_t usenext = B_FALSE;
2838 		if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2839 			uint64_t count;
2840 			/*
2841 			 * A bug in a previous version of the code could
2842 			 * cause upgrade_clones_cb() to not set
2843 			 * ds_next_snap_obj when it should, leading to a
2844 			 * missing entry.  Therefore we can only use the
2845 			 * next_clones_obj when its count is correct.
2846 			 */
2847 			int err = zap_count(dp->dp_meta_objset,
2848 			    dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2849 			if (err == 0 &&
2850 			    count == dsl_dataset_phys(ds)->ds_num_children - 1)
2851 				usenext = B_TRUE;
2852 		}
2853 
2854 		if (usenext) {
2855 			zap_cursor_t zc;
2856 			zap_attribute_t za;
2857 			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2858 			    dsl_dataset_phys(ds)->ds_next_clones_obj);
2859 			    zap_cursor_retrieve(&zc, &za) == 0;
2860 			    (void) zap_cursor_advance(&zc)) {
2861 				scan_ds_queue_insert(scn,
2862 				    zfs_strtonum(za.za_name, NULL),
2863 				    dsl_dataset_phys(ds)->ds_creation_txg);
2864 			}
2865 			zap_cursor_fini(&zc);
2866 		} else {
2867 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2868 			    enqueue_clones_cb, &ds->ds_object,
2869 			    DS_FIND_CHILDREN));
2870 		}
2871 	}
2872 
2873 out:
2874 	dsl_dataset_rele(ds, FTAG);
2875 }
2876 
2877 static int
2878 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2879 {
2880 	(void) arg;
2881 	dsl_dataset_t *ds;
2882 	int err;
2883 	dsl_scan_t *scn = dp->dp_scan;
2884 
2885 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2886 	if (err)
2887 		return (err);
2888 
2889 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2890 		dsl_dataset_t *prev;
2891 		err = dsl_dataset_hold_obj(dp,
2892 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2893 		if (err) {
2894 			dsl_dataset_rele(ds, FTAG);
2895 			return (err);
2896 		}
2897 
2898 		/*
2899 		 * If this is a clone, we don't need to worry about it for now.
2900 		 */
2901 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2902 			dsl_dataset_rele(ds, FTAG);
2903 			dsl_dataset_rele(prev, FTAG);
2904 			return (0);
2905 		}
2906 		dsl_dataset_rele(ds, FTAG);
2907 		ds = prev;
2908 	}
2909 
2910 	scan_ds_queue_insert(scn, ds->ds_object,
2911 	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2912 	dsl_dataset_rele(ds, FTAG);
2913 	return (0);
2914 }
2915 
2916 void
2917 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2918     ddt_entry_t *dde, dmu_tx_t *tx)
2919 {
2920 	(void) tx;
2921 	const ddt_key_t *ddk = &dde->dde_key;
2922 	ddt_phys_t *ddp = dde->dde_phys;
2923 	blkptr_t bp;
2924 	zbookmark_phys_t zb = { 0 };
2925 
2926 	if (!dsl_scan_is_running(scn))
2927 		return;
2928 
2929 	/*
2930 	 * This function is special because it is the only thing
2931 	 * that can add scan_io_t's to the vdev scan queues from
2932 	 * outside dsl_scan_sync(). For the most part this is ok
2933 	 * as long as it is called from within syncing context.
2934 	 * However, dsl_scan_sync() expects that no new sio's will
2935 	 * be added between when all the work for a scan is done
2936 	 * and the next txg when the scan is actually marked as
2937 	 * completed. This check ensures we do not issue new sio's
2938 	 * during this period.
2939 	 */
2940 	if (scn->scn_done_txg != 0)
2941 		return;
2942 
2943 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2944 		if (ddp->ddp_phys_birth == 0 ||
2945 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2946 			continue;
2947 		ddt_bp_create(checksum, ddk, ddp, &bp);
2948 
2949 		scn->scn_visited_this_txg++;
2950 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2951 	}
2952 }
2953 
2954 /*
2955  * Scrub/dedup interaction.
2956  *
2957  * If there are N references to a deduped block, we don't want to scrub it
2958  * N times -- ideally, we should scrub it exactly once.
2959  *
2960  * We leverage the fact that the dde's replication class (enum ddt_class)
2961  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2962  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2963  *
2964  * To prevent excess scrubbing, the scrub begins by walking the DDT
2965  * to find all blocks with refcnt > 1, and scrubs each of these once.
2966  * Since there are two replication classes which contain blocks with
2967  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2968  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2969  *
2970  * There would be nothing more to say if a block's refcnt couldn't change
2971  * during a scrub, but of course it can so we must account for changes
2972  * in a block's replication class.
2973  *
2974  * Here's an example of what can occur:
2975  *
2976  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2977  * when visited during the top-down scrub phase, it will be scrubbed twice.
2978  * This negates our scrub optimization, but is otherwise harmless.
2979  *
2980  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2981  * on each visit during the top-down scrub phase, it will never be scrubbed.
2982  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2983  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2984  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2985  * while a scrub is in progress, it scrubs the block right then.
2986  */
2987 static void
2988 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2989 {
2990 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2991 	ddt_entry_t dde = {{{{0}}}};
2992 	int error;
2993 	uint64_t n = 0;
2994 
2995 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2996 		ddt_t *ddt;
2997 
2998 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2999 			break;
3000 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
3001 		    (longlong_t)ddb->ddb_class,
3002 		    (longlong_t)ddb->ddb_type,
3003 		    (longlong_t)ddb->ddb_checksum,
3004 		    (longlong_t)ddb->ddb_cursor);
3005 
3006 		/* There should be no pending changes to the dedup table */
3007 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
3008 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
3009 
3010 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
3011 		n++;
3012 
3013 		if (dsl_scan_check_suspend(scn, NULL))
3014 			break;
3015 	}
3016 
3017 	zfs_dbgmsg("scanned %llu ddt entries on %s with class_max = %u; "
3018 	    "suspending=%u", (longlong_t)n, scn->scn_dp->dp_spa->spa_name,
3019 	    (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
3020 
3021 	ASSERT(error == 0 || error == ENOENT);
3022 	ASSERT(error != ENOENT ||
3023 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
3024 }
3025 
3026 static uint64_t
3027 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
3028 {
3029 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
3030 	if (ds->ds_is_snapshot)
3031 		return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
3032 	return (smt);
3033 }
3034 
3035 static void
3036 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
3037 {
3038 	scan_ds_t *sds;
3039 	dsl_pool_t *dp = scn->scn_dp;
3040 
3041 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
3042 	    scn->scn_phys.scn_ddt_class_max) {
3043 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
3044 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
3045 		dsl_scan_ddt(scn, tx);
3046 		if (scn->scn_suspending)
3047 			return;
3048 	}
3049 
3050 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
3051 		/* First do the MOS & ORIGIN */
3052 
3053 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
3054 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
3055 		dsl_scan_visit_rootbp(scn, NULL,
3056 		    &dp->dp_meta_rootbp, tx);
3057 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
3058 		if (scn->scn_suspending)
3059 			return;
3060 
3061 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
3062 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
3063 			    enqueue_cb, NULL, DS_FIND_CHILDREN));
3064 		} else {
3065 			dsl_scan_visitds(scn,
3066 			    dp->dp_origin_snap->ds_object, tx);
3067 		}
3068 		ASSERT(!scn->scn_suspending);
3069 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
3070 	    ZB_DESTROYED_OBJSET) {
3071 		uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
3072 		/*
3073 		 * If we were suspended, continue from here. Note if the
3074 		 * ds we were suspended on was deleted, the zb_objset may
3075 		 * be -1, so we will skip this and find a new objset
3076 		 * below.
3077 		 */
3078 		dsl_scan_visitds(scn, dsobj, tx);
3079 		if (scn->scn_suspending)
3080 			return;
3081 	}
3082 
3083 	/*
3084 	 * In case we suspended right at the end of the ds, zero the
3085 	 * bookmark so we don't think that we're still trying to resume.
3086 	 */
3087 	memset(&scn->scn_phys.scn_bookmark, 0, sizeof (zbookmark_phys_t));
3088 
3089 	/*
3090 	 * Keep pulling things out of the dataset avl queue. Updates to the
3091 	 * persistent zap-object-as-queue happen only at checkpoints.
3092 	 */
3093 	while ((sds = avl_first(&scn->scn_queue)) != NULL) {
3094 		dsl_dataset_t *ds;
3095 		uint64_t dsobj = sds->sds_dsobj;
3096 		uint64_t txg = sds->sds_txg;
3097 
3098 		/* dequeue and free the ds from the queue */
3099 		scan_ds_queue_remove(scn, dsobj);
3100 		sds = NULL;
3101 
3102 		/* set up min / max txg */
3103 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
3104 		if (txg != 0) {
3105 			scn->scn_phys.scn_cur_min_txg =
3106 			    MAX(scn->scn_phys.scn_min_txg, txg);
3107 		} else {
3108 			scn->scn_phys.scn_cur_min_txg =
3109 			    MAX(scn->scn_phys.scn_min_txg,
3110 			    dsl_dataset_phys(ds)->ds_prev_snap_txg);
3111 		}
3112 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
3113 		dsl_dataset_rele(ds, FTAG);
3114 
3115 		dsl_scan_visitds(scn, dsobj, tx);
3116 		if (scn->scn_suspending)
3117 			return;
3118 	}
3119 
3120 	/* No more objsets to fetch, we're done */
3121 	scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
3122 	ASSERT0(scn->scn_suspending);
3123 }
3124 
3125 static uint64_t
3126 dsl_scan_count_data_disks(spa_t *spa)
3127 {
3128 	vdev_t *rvd = spa->spa_root_vdev;
3129 	uint64_t i, leaves = 0;
3130 
3131 	for (i = 0; i < rvd->vdev_children; i++) {
3132 		vdev_t *vd = rvd->vdev_child[i];
3133 		if (vd->vdev_islog || vd->vdev_isspare || vd->vdev_isl2cache)
3134 			continue;
3135 		leaves += vdev_get_ndisks(vd) - vdev_get_nparity(vd);
3136 	}
3137 	return (leaves);
3138 }
3139 
3140 static void
3141 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
3142 {
3143 	int i;
3144 	uint64_t cur_size = 0;
3145 
3146 	for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3147 		cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
3148 	}
3149 
3150 	q->q_total_zio_size_this_txg += cur_size;
3151 	q->q_zios_this_txg++;
3152 }
3153 
3154 static void
3155 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
3156     uint64_t end)
3157 {
3158 	q->q_total_seg_size_this_txg += end - start;
3159 	q->q_segs_this_txg++;
3160 }
3161 
3162 static boolean_t
3163 scan_io_queue_check_suspend(dsl_scan_t *scn)
3164 {
3165 	/* See comment in dsl_scan_check_suspend() */
3166 	uint64_t curr_time_ns = gethrtime();
3167 	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
3168 	uint64_t sync_time_ns = curr_time_ns -
3169 	    scn->scn_dp->dp_spa->spa_sync_starttime;
3170 	uint64_t dirty_min_bytes = zfs_dirty_data_max *
3171 	    zfs_vdev_async_write_active_min_dirty_percent / 100;
3172 	uint_t mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
3173 	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
3174 
3175 	return ((NSEC2MSEC(scan_time_ns) > mintime &&
3176 	    (scn->scn_dp->dp_dirty_total >= dirty_min_bytes ||
3177 	    txg_sync_waiting(scn->scn_dp) ||
3178 	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
3179 	    spa_shutting_down(scn->scn_dp->dp_spa));
3180 }
3181 
3182 /*
3183  * Given a list of scan_io_t's in io_list, this issues the I/Os out to
3184  * disk. This consumes the io_list and frees the scan_io_t's. This is
3185  * called when emptying queues, either when we're up against the memory
3186  * limit or when we have finished scanning. Returns B_TRUE if we stopped
3187  * processing the list before we finished. Any sios that were not issued
3188  * will remain in the io_list.
3189  */
3190 static boolean_t
3191 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
3192 {
3193 	dsl_scan_t *scn = queue->q_scn;
3194 	scan_io_t *sio;
3195 	boolean_t suspended = B_FALSE;
3196 
3197 	while ((sio = list_head(io_list)) != NULL) {
3198 		blkptr_t bp;
3199 
3200 		if (scan_io_queue_check_suspend(scn)) {
3201 			suspended = B_TRUE;
3202 			break;
3203 		}
3204 
3205 		sio2bp(sio, &bp);
3206 		scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
3207 		    &sio->sio_zb, queue);
3208 		(void) list_remove_head(io_list);
3209 		scan_io_queues_update_zio_stats(queue, &bp);
3210 		sio_free(sio);
3211 	}
3212 	return (suspended);
3213 }
3214 
3215 /*
3216  * This function removes sios from an IO queue which reside within a given
3217  * range_seg_t and inserts them (in offset order) into a list. Note that
3218  * we only ever return a maximum of 32 sios at once. If there are more sios
3219  * to process within this segment that did not make it onto the list we
3220  * return B_TRUE and otherwise B_FALSE.
3221  */
3222 static boolean_t
3223 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
3224 {
3225 	scan_io_t *srch_sio, *sio, *next_sio;
3226 	avl_index_t idx;
3227 	uint_t num_sios = 0;
3228 	int64_t bytes_issued = 0;
3229 
3230 	ASSERT(rs != NULL);
3231 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3232 
3233 	srch_sio = sio_alloc(1);
3234 	srch_sio->sio_nr_dvas = 1;
3235 	SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
3236 
3237 	/*
3238 	 * The exact start of the extent might not contain any matching zios,
3239 	 * so if that's the case, examine the next one in the tree.
3240 	 */
3241 	sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
3242 	sio_free(srch_sio);
3243 
3244 	if (sio == NULL)
3245 		sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
3246 
3247 	while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
3248 	    queue->q_exts_by_addr) && num_sios <= 32) {
3249 		ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
3250 		    queue->q_exts_by_addr));
3251 		ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
3252 		    queue->q_exts_by_addr));
3253 
3254 		next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
3255 		avl_remove(&queue->q_sios_by_addr, sio);
3256 		if (avl_is_empty(&queue->q_sios_by_addr))
3257 			atomic_add_64(&queue->q_scn->scn_queues_pending, -1);
3258 		queue->q_sio_memused -= SIO_GET_MUSED(sio);
3259 
3260 		bytes_issued += SIO_GET_ASIZE(sio);
3261 		num_sios++;
3262 		list_insert_tail(list, sio);
3263 		sio = next_sio;
3264 	}
3265 
3266 	/*
3267 	 * We limit the number of sios we process at once to 32 to avoid
3268 	 * biting off more than we can chew. If we didn't take everything
3269 	 * in the segment we update it to reflect the work we were able to
3270 	 * complete. Otherwise, we remove it from the range tree entirely.
3271 	 */
3272 	if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
3273 	    queue->q_exts_by_addr)) {
3274 		range_tree_adjust_fill(queue->q_exts_by_addr, rs,
3275 		    -bytes_issued);
3276 		range_tree_resize_segment(queue->q_exts_by_addr, rs,
3277 		    SIO_GET_OFFSET(sio), rs_get_end(rs,
3278 		    queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
3279 		queue->q_last_ext_addr = SIO_GET_OFFSET(sio);
3280 		return (B_TRUE);
3281 	} else {
3282 		uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
3283 		uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
3284 		range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
3285 		queue->q_last_ext_addr = -1;
3286 		return (B_FALSE);
3287 	}
3288 }
3289 
3290 /*
3291  * This is called from the queue emptying thread and selects the next
3292  * extent from which we are to issue I/Os. The behavior of this function
3293  * depends on the state of the scan, the current memory consumption and
3294  * whether or not we are performing a scan shutdown.
3295  * 1) We select extents in an elevator algorithm (LBA-order) if the scan
3296  * 	needs to perform a checkpoint
3297  * 2) We select the largest available extent if we are up against the
3298  * 	memory limit.
3299  * 3) Otherwise we don't select any extents.
3300  */
3301 static range_seg_t *
3302 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
3303 {
3304 	dsl_scan_t *scn = queue->q_scn;
3305 	range_tree_t *rt = queue->q_exts_by_addr;
3306 
3307 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3308 	ASSERT(scn->scn_is_sorted);
3309 
3310 	if (!scn->scn_checkpointing && !scn->scn_clearing)
3311 		return (NULL);
3312 
3313 	/*
3314 	 * During normal clearing, we want to issue our largest segments
3315 	 * first, keeping IO as sequential as possible, and leaving the
3316 	 * smaller extents for later with the hope that they might eventually
3317 	 * grow to larger sequential segments. However, when the scan is
3318 	 * checkpointing, no new extents will be added to the sorting queue,
3319 	 * so the way we are sorted now is as good as it will ever get.
3320 	 * In this case, we instead switch to issuing extents in LBA order.
3321 	 */
3322 	if ((zfs_scan_issue_strategy < 1 && scn->scn_checkpointing) ||
3323 	    zfs_scan_issue_strategy == 1)
3324 		return (range_tree_first(rt));
3325 
3326 	/*
3327 	 * Try to continue previous extent if it is not completed yet.  After
3328 	 * shrink in scan_io_queue_gather() it may no longer be the best, but
3329 	 * otherwise we leave shorter remnant every txg.
3330 	 */
3331 	uint64_t start;
3332 	uint64_t size = 1ULL << rt->rt_shift;
3333 	range_seg_t *addr_rs;
3334 	if (queue->q_last_ext_addr != -1) {
3335 		start = queue->q_last_ext_addr;
3336 		addr_rs = range_tree_find(rt, start, size);
3337 		if (addr_rs != NULL)
3338 			return (addr_rs);
3339 	}
3340 
3341 	/*
3342 	 * Nothing to continue, so find new best extent.
3343 	 */
3344 	uint64_t *v = zfs_btree_first(&queue->q_exts_by_size, NULL);
3345 	if (v == NULL)
3346 		return (NULL);
3347 	queue->q_last_ext_addr = start = *v << rt->rt_shift;
3348 
3349 	/*
3350 	 * We need to get the original entry in the by_addr tree so we can
3351 	 * modify it.
3352 	 */
3353 	addr_rs = range_tree_find(rt, start, size);
3354 	ASSERT3P(addr_rs, !=, NULL);
3355 	ASSERT3U(rs_get_start(addr_rs, rt), ==, start);
3356 	ASSERT3U(rs_get_end(addr_rs, rt), >, start);
3357 	return (addr_rs);
3358 }
3359 
3360 static void
3361 scan_io_queues_run_one(void *arg)
3362 {
3363 	dsl_scan_io_queue_t *queue = arg;
3364 	kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3365 	boolean_t suspended = B_FALSE;
3366 	range_seg_t *rs;
3367 	scan_io_t *sio;
3368 	zio_t *zio;
3369 	list_t sio_list;
3370 
3371 	ASSERT(queue->q_scn->scn_is_sorted);
3372 
3373 	list_create(&sio_list, sizeof (scan_io_t),
3374 	    offsetof(scan_io_t, sio_nodes.sio_list_node));
3375 	zio = zio_null(queue->q_scn->scn_zio_root, queue->q_scn->scn_dp->dp_spa,
3376 	    NULL, NULL, NULL, ZIO_FLAG_CANFAIL);
3377 	mutex_enter(q_lock);
3378 	queue->q_zio = zio;
3379 
3380 	/* Calculate maximum in-flight bytes for this vdev. */
3381 	queue->q_maxinflight_bytes = MAX(1, zfs_scan_vdev_limit *
3382 	    (vdev_get_ndisks(queue->q_vd) - vdev_get_nparity(queue->q_vd)));
3383 
3384 	/* reset per-queue scan statistics for this txg */
3385 	queue->q_total_seg_size_this_txg = 0;
3386 	queue->q_segs_this_txg = 0;
3387 	queue->q_total_zio_size_this_txg = 0;
3388 	queue->q_zios_this_txg = 0;
3389 
3390 	/* loop until we run out of time or sios */
3391 	while ((rs = scan_io_queue_fetch_ext(queue)) != NULL) {
3392 		uint64_t seg_start = 0, seg_end = 0;
3393 		boolean_t more_left;
3394 
3395 		ASSERT(list_is_empty(&sio_list));
3396 
3397 		/* loop while we still have sios left to process in this rs */
3398 		do {
3399 			scan_io_t *first_sio, *last_sio;
3400 
3401 			/*
3402 			 * We have selected which extent needs to be
3403 			 * processed next. Gather up the corresponding sios.
3404 			 */
3405 			more_left = scan_io_queue_gather(queue, rs, &sio_list);
3406 			ASSERT(!list_is_empty(&sio_list));
3407 			first_sio = list_head(&sio_list);
3408 			last_sio = list_tail(&sio_list);
3409 
3410 			seg_end = SIO_GET_END_OFFSET(last_sio);
3411 			if (seg_start == 0)
3412 				seg_start = SIO_GET_OFFSET(first_sio);
3413 
3414 			/*
3415 			 * Issuing sios can take a long time so drop the
3416 			 * queue lock. The sio queue won't be updated by
3417 			 * other threads since we're in syncing context so
3418 			 * we can be sure that our trees will remain exactly
3419 			 * as we left them.
3420 			 */
3421 			mutex_exit(q_lock);
3422 			suspended = scan_io_queue_issue(queue, &sio_list);
3423 			mutex_enter(q_lock);
3424 
3425 			if (suspended)
3426 				break;
3427 		} while (more_left);
3428 
3429 		/* update statistics for debugging purposes */
3430 		scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
3431 
3432 		if (suspended)
3433 			break;
3434 	}
3435 
3436 	/*
3437 	 * If we were suspended in the middle of processing,
3438 	 * requeue any unfinished sios and exit.
3439 	 */
3440 	while ((sio = list_remove_head(&sio_list)) != NULL)
3441 		scan_io_queue_insert_impl(queue, sio);
3442 
3443 	queue->q_zio = NULL;
3444 	mutex_exit(q_lock);
3445 	zio_nowait(zio);
3446 	list_destroy(&sio_list);
3447 }
3448 
3449 /*
3450  * Performs an emptying run on all scan queues in the pool. This just
3451  * punches out one thread per top-level vdev, each of which processes
3452  * only that vdev's scan queue. We can parallelize the I/O here because
3453  * we know that each queue's I/Os only affect its own top-level vdev.
3454  *
3455  * This function waits for the queue runs to complete, and must be
3456  * called from dsl_scan_sync (or in general, syncing context).
3457  */
3458 static void
3459 scan_io_queues_run(dsl_scan_t *scn)
3460 {
3461 	spa_t *spa = scn->scn_dp->dp_spa;
3462 
3463 	ASSERT(scn->scn_is_sorted);
3464 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3465 
3466 	if (scn->scn_queues_pending == 0)
3467 		return;
3468 
3469 	if (scn->scn_taskq == NULL) {
3470 		int nthreads = spa->spa_root_vdev->vdev_children;
3471 
3472 		/*
3473 		 * We need to make this taskq *always* execute as many
3474 		 * threads in parallel as we have top-level vdevs and no
3475 		 * less, otherwise strange serialization of the calls to
3476 		 * scan_io_queues_run_one can occur during spa_sync runs
3477 		 * and that significantly impacts performance.
3478 		 */
3479 		scn->scn_taskq = taskq_create("dsl_scan_iss", nthreads,
3480 		    minclsyspri, nthreads, nthreads, TASKQ_PREPOPULATE);
3481 	}
3482 
3483 	for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3484 		vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3485 
3486 		mutex_enter(&vd->vdev_scan_io_queue_lock);
3487 		if (vd->vdev_scan_io_queue != NULL) {
3488 			VERIFY(taskq_dispatch(scn->scn_taskq,
3489 			    scan_io_queues_run_one, vd->vdev_scan_io_queue,
3490 			    TQ_SLEEP) != TASKQID_INVALID);
3491 		}
3492 		mutex_exit(&vd->vdev_scan_io_queue_lock);
3493 	}
3494 
3495 	/*
3496 	 * Wait for the queues to finish issuing their IOs for this run
3497 	 * before we return. There may still be IOs in flight at this
3498 	 * point.
3499 	 */
3500 	taskq_wait(scn->scn_taskq);
3501 }
3502 
3503 static boolean_t
3504 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3505 {
3506 	uint64_t elapsed_nanosecs;
3507 
3508 	if (zfs_recover)
3509 		return (B_FALSE);
3510 
3511 	if (zfs_async_block_max_blocks != 0 &&
3512 	    scn->scn_visited_this_txg >= zfs_async_block_max_blocks) {
3513 		return (B_TRUE);
3514 	}
3515 
3516 	if (zfs_max_async_dedup_frees != 0 &&
3517 	    scn->scn_dedup_frees_this_txg >= zfs_max_async_dedup_frees) {
3518 		return (B_TRUE);
3519 	}
3520 
3521 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3522 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3523 	    (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3524 	    txg_sync_waiting(scn->scn_dp)) ||
3525 	    spa_shutting_down(scn->scn_dp->dp_spa));
3526 }
3527 
3528 static int
3529 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3530 {
3531 	dsl_scan_t *scn = arg;
3532 
3533 	if (!scn->scn_is_bptree ||
3534 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3535 		if (dsl_scan_async_block_should_pause(scn))
3536 			return (SET_ERROR(ERESTART));
3537 	}
3538 
3539 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3540 	    dmu_tx_get_txg(tx), bp, 0));
3541 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3542 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3543 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3544 	scn->scn_visited_this_txg++;
3545 	if (BP_GET_DEDUP(bp))
3546 		scn->scn_dedup_frees_this_txg++;
3547 	return (0);
3548 }
3549 
3550 static void
3551 dsl_scan_update_stats(dsl_scan_t *scn)
3552 {
3553 	spa_t *spa = scn->scn_dp->dp_spa;
3554 	uint64_t i;
3555 	uint64_t seg_size_total = 0, zio_size_total = 0;
3556 	uint64_t seg_count_total = 0, zio_count_total = 0;
3557 
3558 	for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3559 		vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3560 		dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3561 
3562 		if (queue == NULL)
3563 			continue;
3564 
3565 		seg_size_total += queue->q_total_seg_size_this_txg;
3566 		zio_size_total += queue->q_total_zio_size_this_txg;
3567 		seg_count_total += queue->q_segs_this_txg;
3568 		zio_count_total += queue->q_zios_this_txg;
3569 	}
3570 
3571 	if (seg_count_total == 0 || zio_count_total == 0) {
3572 		scn->scn_avg_seg_size_this_txg = 0;
3573 		scn->scn_avg_zio_size_this_txg = 0;
3574 		scn->scn_segs_this_txg = 0;
3575 		scn->scn_zios_this_txg = 0;
3576 		return;
3577 	}
3578 
3579 	scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3580 	scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3581 	scn->scn_segs_this_txg = seg_count_total;
3582 	scn->scn_zios_this_txg = zio_count_total;
3583 }
3584 
3585 static int
3586 bpobj_dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3587     dmu_tx_t *tx)
3588 {
3589 	ASSERT(!bp_freed);
3590 	return (dsl_scan_free_block_cb(arg, bp, tx));
3591 }
3592 
3593 static int
3594 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
3595     dmu_tx_t *tx)
3596 {
3597 	ASSERT(!bp_freed);
3598 	dsl_scan_t *scn = arg;
3599 	const dva_t *dva = &bp->blk_dva[0];
3600 
3601 	if (dsl_scan_async_block_should_pause(scn))
3602 		return (SET_ERROR(ERESTART));
3603 
3604 	spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3605 	    DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3606 	    DVA_GET_ASIZE(dva), tx);
3607 	scn->scn_visited_this_txg++;
3608 	return (0);
3609 }
3610 
3611 boolean_t
3612 dsl_scan_active(dsl_scan_t *scn)
3613 {
3614 	spa_t *spa = scn->scn_dp->dp_spa;
3615 	uint64_t used = 0, comp, uncomp;
3616 	boolean_t clones_left;
3617 
3618 	if (spa->spa_load_state != SPA_LOAD_NONE)
3619 		return (B_FALSE);
3620 	if (spa_shutting_down(spa))
3621 		return (B_FALSE);
3622 	if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3623 	    (scn->scn_async_destroying && !scn->scn_async_stalled))
3624 		return (B_TRUE);
3625 
3626 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3627 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3628 		    &used, &comp, &uncomp);
3629 	}
3630 	clones_left = spa_livelist_delete_check(spa);
3631 	return ((used != 0) || (clones_left));
3632 }
3633 
3634 boolean_t
3635 dsl_errorscrub_active(dsl_scan_t *scn)
3636 {
3637 	spa_t *spa = scn->scn_dp->dp_spa;
3638 	if (spa->spa_load_state != SPA_LOAD_NONE)
3639 		return (B_FALSE);
3640 	if (spa_shutting_down(spa))
3641 		return (B_FALSE);
3642 	if (dsl_errorscrubbing(scn->scn_dp))
3643 		return (B_TRUE);
3644 	return (B_FALSE);
3645 }
3646 
3647 static boolean_t
3648 dsl_scan_check_deferred(vdev_t *vd)
3649 {
3650 	boolean_t need_resilver = B_FALSE;
3651 
3652 	for (int c = 0; c < vd->vdev_children; c++) {
3653 		need_resilver |=
3654 		    dsl_scan_check_deferred(vd->vdev_child[c]);
3655 	}
3656 
3657 	if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3658 	    !vd->vdev_ops->vdev_op_leaf)
3659 		return (need_resilver);
3660 
3661 	if (!vd->vdev_resilver_deferred)
3662 		need_resilver = B_TRUE;
3663 
3664 	return (need_resilver);
3665 }
3666 
3667 static boolean_t
3668 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3669     uint64_t phys_birth)
3670 {
3671 	vdev_t *vd;
3672 
3673 	vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3674 
3675 	if (vd->vdev_ops == &vdev_indirect_ops) {
3676 		/*
3677 		 * The indirect vdev can point to multiple
3678 		 * vdevs.  For simplicity, always create
3679 		 * the resilver zio_t. zio_vdev_io_start()
3680 		 * will bypass the child resilver i/o's if
3681 		 * they are on vdevs that don't have DTL's.
3682 		 */
3683 		return (B_TRUE);
3684 	}
3685 
3686 	if (DVA_GET_GANG(dva)) {
3687 		/*
3688 		 * Gang members may be spread across multiple
3689 		 * vdevs, so the best estimate we have is the
3690 		 * scrub range, which has already been checked.
3691 		 * XXX -- it would be better to change our
3692 		 * allocation policy to ensure that all
3693 		 * gang members reside on the same vdev.
3694 		 */
3695 		return (B_TRUE);
3696 	}
3697 
3698 	/*
3699 	 * Check if the top-level vdev must resilver this offset.
3700 	 * When the offset does not intersect with a dirty leaf DTL
3701 	 * then it may be possible to skip the resilver IO.  The psize
3702 	 * is provided instead of asize to simplify the check for RAIDZ.
3703 	 */
3704 	if (!vdev_dtl_need_resilver(vd, dva, psize, phys_birth))
3705 		return (B_FALSE);
3706 
3707 	/*
3708 	 * Check that this top-level vdev has a device under it which
3709 	 * is resilvering and is not deferred.
3710 	 */
3711 	if (!dsl_scan_check_deferred(vd))
3712 		return (B_FALSE);
3713 
3714 	return (B_TRUE);
3715 }
3716 
3717 static int
3718 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3719 {
3720 	dsl_scan_t *scn = dp->dp_scan;
3721 	spa_t *spa = dp->dp_spa;
3722 	int err = 0;
3723 
3724 	if (spa_suspend_async_destroy(spa))
3725 		return (0);
3726 
3727 	if (zfs_free_bpobj_enabled &&
3728 	    spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3729 		scn->scn_is_bptree = B_FALSE;
3730 		scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3731 		scn->scn_zio_root = zio_root(spa, NULL,
3732 		    NULL, ZIO_FLAG_MUSTSUCCEED);
3733 		err = bpobj_iterate(&dp->dp_free_bpobj,
3734 		    bpobj_dsl_scan_free_block_cb, scn, tx);
3735 		VERIFY0(zio_wait(scn->scn_zio_root));
3736 		scn->scn_zio_root = NULL;
3737 
3738 		if (err != 0 && err != ERESTART)
3739 			zfs_panic_recover("error %u from bpobj_iterate()", err);
3740 	}
3741 
3742 	if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3743 		ASSERT(scn->scn_async_destroying);
3744 		scn->scn_is_bptree = B_TRUE;
3745 		scn->scn_zio_root = zio_root(spa, NULL,
3746 		    NULL, ZIO_FLAG_MUSTSUCCEED);
3747 		err = bptree_iterate(dp->dp_meta_objset,
3748 		    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3749 		VERIFY0(zio_wait(scn->scn_zio_root));
3750 		scn->scn_zio_root = NULL;
3751 
3752 		if (err == EIO || err == ECKSUM) {
3753 			err = 0;
3754 		} else if (err != 0 && err != ERESTART) {
3755 			zfs_panic_recover("error %u from "
3756 			    "traverse_dataset_destroyed()", err);
3757 		}
3758 
3759 		if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3760 			/* finished; deactivate async destroy feature */
3761 			spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3762 			ASSERT(!spa_feature_is_active(spa,
3763 			    SPA_FEATURE_ASYNC_DESTROY));
3764 			VERIFY0(zap_remove(dp->dp_meta_objset,
3765 			    DMU_POOL_DIRECTORY_OBJECT,
3766 			    DMU_POOL_BPTREE_OBJ, tx));
3767 			VERIFY0(bptree_free(dp->dp_meta_objset,
3768 			    dp->dp_bptree_obj, tx));
3769 			dp->dp_bptree_obj = 0;
3770 			scn->scn_async_destroying = B_FALSE;
3771 			scn->scn_async_stalled = B_FALSE;
3772 		} else {
3773 			/*
3774 			 * If we didn't make progress, mark the async
3775 			 * destroy as stalled, so that we will not initiate
3776 			 * a spa_sync() on its behalf.  Note that we only
3777 			 * check this if we are not finished, because if the
3778 			 * bptree had no blocks for us to visit, we can
3779 			 * finish without "making progress".
3780 			 */
3781 			scn->scn_async_stalled =
3782 			    (scn->scn_visited_this_txg == 0);
3783 		}
3784 	}
3785 	if (scn->scn_visited_this_txg) {
3786 		zfs_dbgmsg("freed %llu blocks in %llums from "
3787 		    "free_bpobj/bptree on %s in txg %llu; err=%u",
3788 		    (longlong_t)scn->scn_visited_this_txg,
3789 		    (longlong_t)
3790 		    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3791 		    spa->spa_name, (longlong_t)tx->tx_txg, err);
3792 		scn->scn_visited_this_txg = 0;
3793 		scn->scn_dedup_frees_this_txg = 0;
3794 
3795 		/*
3796 		 * Write out changes to the DDT and the BRT that may be required
3797 		 * as a result of the blocks freed.  This ensures that the DDT
3798 		 * and the BRT are clean when a scrub/resilver runs.
3799 		 */
3800 		ddt_sync(spa, tx->tx_txg);
3801 		brt_sync(spa, tx->tx_txg);
3802 	}
3803 	if (err != 0)
3804 		return (err);
3805 	if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3806 	    zfs_free_leak_on_eio &&
3807 	    (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3808 	    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3809 	    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3810 		/*
3811 		 * We have finished background destroying, but there is still
3812 		 * some space left in the dp_free_dir. Transfer this leaked
3813 		 * space to the dp_leak_dir.
3814 		 */
3815 		if (dp->dp_leak_dir == NULL) {
3816 			rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3817 			(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3818 			    LEAK_DIR_NAME, tx);
3819 			VERIFY0(dsl_pool_open_special_dir(dp,
3820 			    LEAK_DIR_NAME, &dp->dp_leak_dir));
3821 			rrw_exit(&dp->dp_config_rwlock, FTAG);
3822 		}
3823 		dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3824 		    dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3825 		    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3826 		    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3827 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3828 		    -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3829 		    -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3830 		    -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3831 	}
3832 
3833 	if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3834 	    !spa_livelist_delete_check(spa)) {
3835 		/* finished; verify that space accounting went to zero */
3836 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3837 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3838 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3839 	}
3840 
3841 	spa_notify_waiters(spa);
3842 
3843 	EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3844 	    0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3845 	    DMU_POOL_OBSOLETE_BPOBJ));
3846 	if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3847 		ASSERT(spa_feature_is_active(dp->dp_spa,
3848 		    SPA_FEATURE_OBSOLETE_COUNTS));
3849 
3850 		scn->scn_is_bptree = B_FALSE;
3851 		scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3852 		err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3853 		    dsl_scan_obsolete_block_cb, scn, tx);
3854 		if (err != 0 && err != ERESTART)
3855 			zfs_panic_recover("error %u from bpobj_iterate()", err);
3856 
3857 		if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3858 			dsl_pool_destroy_obsolete_bpobj(dp, tx);
3859 	}
3860 	return (0);
3861 }
3862 
3863 static void
3864 name_to_bookmark(char *buf, zbookmark_phys_t *zb)
3865 {
3866 	zb->zb_objset = zfs_strtonum(buf, &buf);
3867 	ASSERT(*buf == ':');
3868 	zb->zb_object = zfs_strtonum(buf + 1, &buf);
3869 	ASSERT(*buf == ':');
3870 	zb->zb_level = (int)zfs_strtonum(buf + 1, &buf);
3871 	ASSERT(*buf == ':');
3872 	zb->zb_blkid = zfs_strtonum(buf + 1, &buf);
3873 	ASSERT(*buf == '\0');
3874 }
3875 
3876 static void
3877 name_to_object(char *buf, uint64_t *obj)
3878 {
3879 	*obj = zfs_strtonum(buf, &buf);
3880 	ASSERT(*buf == '\0');
3881 }
3882 
3883 static void
3884 read_by_block_level(dsl_scan_t *scn, zbookmark_phys_t zb)
3885 {
3886 	dsl_pool_t *dp = scn->scn_dp;
3887 	dsl_dataset_t *ds;
3888 	objset_t *os;
3889 	if (dsl_dataset_hold_obj(dp, zb.zb_objset, FTAG, &ds) != 0)
3890 		return;
3891 
3892 	if (dmu_objset_from_ds(ds, &os) != 0) {
3893 		dsl_dataset_rele(ds, FTAG);
3894 		return;
3895 	}
3896 
3897 	/*
3898 	 * If the key is not loaded dbuf_dnode_findbp() will error out with
3899 	 * EACCES. However in that case dnode_hold() will eventually call
3900 	 * dbuf_read()->zio_wait() which may call spa_log_error(). This will
3901 	 * lead to a deadlock due to us holding the mutex spa_errlist_lock.
3902 	 * Avoid this by checking here if the keys are loaded, if not return.
3903 	 * If the keys are not loaded the head_errlog feature is meaningless
3904 	 * as we cannot figure out the birth txg of the block pointer.
3905 	 */
3906 	if (dsl_dataset_get_keystatus(ds->ds_dir) ==
3907 	    ZFS_KEYSTATUS_UNAVAILABLE) {
3908 		dsl_dataset_rele(ds, FTAG);
3909 		return;
3910 	}
3911 
3912 	dnode_t *dn;
3913 	blkptr_t bp;
3914 
3915 	if (dnode_hold(os, zb.zb_object, FTAG, &dn) != 0) {
3916 		dsl_dataset_rele(ds, FTAG);
3917 		return;
3918 	}
3919 
3920 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
3921 	int error = dbuf_dnode_findbp(dn, zb.zb_level, zb.zb_blkid, &bp, NULL,
3922 	    NULL);
3923 
3924 	if (error) {
3925 		rw_exit(&dn->dn_struct_rwlock);
3926 		dnode_rele(dn, FTAG);
3927 		dsl_dataset_rele(ds, FTAG);
3928 		return;
3929 	}
3930 
3931 	if (!error && BP_IS_HOLE(&bp)) {
3932 		rw_exit(&dn->dn_struct_rwlock);
3933 		dnode_rele(dn, FTAG);
3934 		dsl_dataset_rele(ds, FTAG);
3935 		return;
3936 	}
3937 
3938 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW |
3939 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCRUB;
3940 
3941 	/* If it's an intent log block, failure is expected. */
3942 	if (zb.zb_level == ZB_ZIL_LEVEL)
3943 		zio_flags |= ZIO_FLAG_SPECULATIVE;
3944 
3945 	ASSERT(!BP_IS_EMBEDDED(&bp));
3946 	scan_exec_io(dp, &bp, zio_flags, &zb, NULL);
3947 	rw_exit(&dn->dn_struct_rwlock);
3948 	dnode_rele(dn, FTAG);
3949 	dsl_dataset_rele(ds, FTAG);
3950 }
3951 
3952 /*
3953  * We keep track of the scrubbed error blocks in "count". This will be used
3954  * when deciding whether we exceeded zfs_scrub_error_blocks_per_txg. This
3955  * function is modelled after check_filesystem().
3956  */
3957 static int
3958 scrub_filesystem(spa_t *spa, uint64_t fs, zbookmark_err_phys_t *zep,
3959     int *count)
3960 {
3961 	dsl_dataset_t *ds;
3962 	dsl_pool_t *dp = spa->spa_dsl_pool;
3963 	dsl_scan_t *scn = dp->dp_scan;
3964 
3965 	int error = dsl_dataset_hold_obj(dp, fs, FTAG, &ds);
3966 	if (error != 0)
3967 		return (error);
3968 
3969 	uint64_t latest_txg;
3970 	uint64_t txg_to_consider = spa->spa_syncing_txg;
3971 	boolean_t check_snapshot = B_TRUE;
3972 
3973 	error = find_birth_txg(ds, zep, &latest_txg);
3974 
3975 	/*
3976 	 * If find_birth_txg() errors out, then err on the side of caution and
3977 	 * proceed. In worst case scenario scrub all objects. If zep->zb_birth
3978 	 * is 0 (e.g. in case of encryption with unloaded keys) also proceed to
3979 	 * scrub all objects.
3980 	 */
3981 	if (error == 0 && zep->zb_birth == latest_txg) {
3982 		/* Block neither free nor re written. */
3983 		zbookmark_phys_t zb;
3984 		zep_to_zb(fs, zep, &zb);
3985 		scn->scn_zio_root = zio_root(spa, NULL, NULL,
3986 		    ZIO_FLAG_CANFAIL);
3987 		/* We have already acquired the config lock for spa */
3988 		read_by_block_level(scn, zb);
3989 
3990 		(void) zio_wait(scn->scn_zio_root);
3991 		scn->scn_zio_root = NULL;
3992 
3993 		scn->errorscrub_phys.dep_examined++;
3994 		scn->errorscrub_phys.dep_to_examine--;
3995 		(*count)++;
3996 		if ((*count) == zfs_scrub_error_blocks_per_txg ||
3997 		    dsl_error_scrub_check_suspend(scn, &zb)) {
3998 			dsl_dataset_rele(ds, FTAG);
3999 			return (SET_ERROR(EFAULT));
4000 		}
4001 
4002 		check_snapshot = B_FALSE;
4003 	} else if (error == 0) {
4004 		txg_to_consider = latest_txg;
4005 	}
4006 
4007 	/*
4008 	 * Retrieve the number of snapshots if the dataset is not a snapshot.
4009 	 */
4010 	uint64_t snap_count = 0;
4011 	if (dsl_dataset_phys(ds)->ds_snapnames_zapobj != 0) {
4012 
4013 		error = zap_count(spa->spa_meta_objset,
4014 		    dsl_dataset_phys(ds)->ds_snapnames_zapobj, &snap_count);
4015 
4016 		if (error != 0) {
4017 			dsl_dataset_rele(ds, FTAG);
4018 			return (error);
4019 		}
4020 	}
4021 
4022 	if (snap_count == 0) {
4023 		/* Filesystem without snapshots. */
4024 		dsl_dataset_rele(ds, FTAG);
4025 		return (0);
4026 	}
4027 
4028 	uint64_t snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4029 	uint64_t snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4030 
4031 	dsl_dataset_rele(ds, FTAG);
4032 
4033 	/* Check only snapshots created from this file system. */
4034 	while (snap_obj != 0 && zep->zb_birth < snap_obj_txg &&
4035 	    snap_obj_txg <= txg_to_consider) {
4036 
4037 		error = dsl_dataset_hold_obj(dp, snap_obj, FTAG, &ds);
4038 		if (error != 0)
4039 			return (error);
4040 
4041 		if (dsl_dir_phys(ds->ds_dir)->dd_head_dataset_obj != fs) {
4042 			snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4043 			snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4044 			dsl_dataset_rele(ds, FTAG);
4045 			continue;
4046 		}
4047 
4048 		boolean_t affected = B_TRUE;
4049 		if (check_snapshot) {
4050 			uint64_t blk_txg;
4051 			error = find_birth_txg(ds, zep, &blk_txg);
4052 
4053 			/*
4054 			 * Scrub the snapshot also when zb_birth == 0 or when
4055 			 * find_birth_txg() returns an error.
4056 			 */
4057 			affected = (error == 0 && zep->zb_birth == blk_txg) ||
4058 			    (error != 0) || (zep->zb_birth == 0);
4059 		}
4060 
4061 		/* Scrub snapshots. */
4062 		if (affected) {
4063 			zbookmark_phys_t zb;
4064 			zep_to_zb(snap_obj, zep, &zb);
4065 			scn->scn_zio_root = zio_root(spa, NULL, NULL,
4066 			    ZIO_FLAG_CANFAIL);
4067 			/* We have already acquired the config lock for spa */
4068 			read_by_block_level(scn, zb);
4069 
4070 			(void) zio_wait(scn->scn_zio_root);
4071 			scn->scn_zio_root = NULL;
4072 
4073 			scn->errorscrub_phys.dep_examined++;
4074 			scn->errorscrub_phys.dep_to_examine--;
4075 			(*count)++;
4076 			if ((*count) == zfs_scrub_error_blocks_per_txg ||
4077 			    dsl_error_scrub_check_suspend(scn, &zb)) {
4078 				dsl_dataset_rele(ds, FTAG);
4079 				return (EFAULT);
4080 			}
4081 		}
4082 		snap_obj_txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
4083 		snap_obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
4084 		dsl_dataset_rele(ds, FTAG);
4085 	}
4086 	return (0);
4087 }
4088 
4089 void
4090 dsl_errorscrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
4091 {
4092 	spa_t *spa = dp->dp_spa;
4093 	dsl_scan_t *scn = dp->dp_scan;
4094 
4095 	/*
4096 	 * Only process scans in sync pass 1.
4097 	 */
4098 
4099 	if (spa_sync_pass(spa) > 1)
4100 		return;
4101 
4102 	/*
4103 	 * If the spa is shutting down, then stop scanning. This will
4104 	 * ensure that the scan does not dirty any new data during the
4105 	 * shutdown phase.
4106 	 */
4107 	if (spa_shutting_down(spa))
4108 		return;
4109 
4110 	if (!dsl_errorscrub_active(scn) || dsl_errorscrub_is_paused(scn)) {
4111 		return;
4112 	}
4113 
4114 	if (dsl_scan_resilvering(scn->scn_dp)) {
4115 		/* cancel the error scrub if resilver started */
4116 		dsl_scan_cancel(scn->scn_dp);
4117 		return;
4118 	}
4119 
4120 	spa->spa_scrub_active = B_TRUE;
4121 	scn->scn_sync_start_time = gethrtime();
4122 
4123 	/*
4124 	 * zfs_scan_suspend_progress can be set to disable scrub progress.
4125 	 * See more detailed comment in dsl_scan_sync().
4126 	 */
4127 	if (zfs_scan_suspend_progress) {
4128 		uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4129 		int mintime = zfs_scrub_min_time_ms;
4130 
4131 		while (zfs_scan_suspend_progress &&
4132 		    !txg_sync_waiting(scn->scn_dp) &&
4133 		    !spa_shutting_down(scn->scn_dp->dp_spa) &&
4134 		    NSEC2MSEC(scan_time_ns) < mintime) {
4135 			delay(hz);
4136 			scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4137 		}
4138 		return;
4139 	}
4140 
4141 	int i = 0;
4142 	zap_attribute_t *za;
4143 	zbookmark_phys_t *zb;
4144 	boolean_t limit_exceeded = B_FALSE;
4145 
4146 	za = kmem_zalloc(sizeof (zap_attribute_t), KM_SLEEP);
4147 	zb = kmem_zalloc(sizeof (zbookmark_phys_t), KM_SLEEP);
4148 
4149 	if (!spa_feature_is_enabled(spa, SPA_FEATURE_HEAD_ERRLOG)) {
4150 		for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0;
4151 		    zap_cursor_advance(&scn->errorscrub_cursor)) {
4152 			name_to_bookmark(za->za_name, zb);
4153 
4154 			scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4155 			    NULL, ZIO_FLAG_CANFAIL);
4156 			dsl_pool_config_enter(dp, FTAG);
4157 			read_by_block_level(scn, *zb);
4158 			dsl_pool_config_exit(dp, FTAG);
4159 
4160 			(void) zio_wait(scn->scn_zio_root);
4161 			scn->scn_zio_root = NULL;
4162 
4163 			scn->errorscrub_phys.dep_examined += 1;
4164 			scn->errorscrub_phys.dep_to_examine -= 1;
4165 			i++;
4166 			if (i == zfs_scrub_error_blocks_per_txg ||
4167 			    dsl_error_scrub_check_suspend(scn, zb)) {
4168 				limit_exceeded = B_TRUE;
4169 				break;
4170 			}
4171 		}
4172 
4173 		if (!limit_exceeded)
4174 			dsl_errorscrub_done(scn, B_TRUE, tx);
4175 
4176 		dsl_errorscrub_sync_state(scn, tx);
4177 		kmem_free(za, sizeof (*za));
4178 		kmem_free(zb, sizeof (*zb));
4179 		return;
4180 	}
4181 
4182 	int error = 0;
4183 	for (; zap_cursor_retrieve(&scn->errorscrub_cursor, za) == 0;
4184 	    zap_cursor_advance(&scn->errorscrub_cursor)) {
4185 
4186 		zap_cursor_t *head_ds_cursor;
4187 		zap_attribute_t *head_ds_attr;
4188 		zbookmark_err_phys_t head_ds_block;
4189 
4190 		head_ds_cursor = kmem_zalloc(sizeof (zap_cursor_t), KM_SLEEP);
4191 		head_ds_attr = kmem_zalloc(sizeof (zap_attribute_t), KM_SLEEP);
4192 
4193 		uint64_t head_ds_err_obj = za->za_first_integer;
4194 		uint64_t head_ds;
4195 		name_to_object(za->za_name, &head_ds);
4196 		boolean_t config_held = B_FALSE;
4197 		uint64_t top_affected_fs;
4198 
4199 		for (zap_cursor_init(head_ds_cursor, spa->spa_meta_objset,
4200 		    head_ds_err_obj); zap_cursor_retrieve(head_ds_cursor,
4201 		    head_ds_attr) == 0; zap_cursor_advance(head_ds_cursor)) {
4202 
4203 			name_to_errphys(head_ds_attr->za_name, &head_ds_block);
4204 
4205 			/*
4206 			 * In case we are called from spa_sync the pool
4207 			 * config is already held.
4208 			 */
4209 			if (!dsl_pool_config_held(dp)) {
4210 				dsl_pool_config_enter(dp, FTAG);
4211 				config_held = B_TRUE;
4212 			}
4213 
4214 			error = find_top_affected_fs(spa,
4215 			    head_ds, &head_ds_block, &top_affected_fs);
4216 			if (error)
4217 				break;
4218 
4219 			error = scrub_filesystem(spa, top_affected_fs,
4220 			    &head_ds_block, &i);
4221 
4222 			if (error == SET_ERROR(EFAULT)) {
4223 				limit_exceeded = B_TRUE;
4224 				break;
4225 			}
4226 		}
4227 
4228 		zap_cursor_fini(head_ds_cursor);
4229 		kmem_free(head_ds_cursor, sizeof (*head_ds_cursor));
4230 		kmem_free(head_ds_attr, sizeof (*head_ds_attr));
4231 
4232 		if (config_held)
4233 			dsl_pool_config_exit(dp, FTAG);
4234 	}
4235 
4236 	kmem_free(za, sizeof (*za));
4237 	kmem_free(zb, sizeof (*zb));
4238 	if (!limit_exceeded)
4239 		dsl_errorscrub_done(scn, B_TRUE, tx);
4240 
4241 	dsl_errorscrub_sync_state(scn, tx);
4242 }
4243 
4244 /*
4245  * This is the primary entry point for scans that is called from syncing
4246  * context. Scans must happen entirely during syncing context so that we
4247  * can guarantee that blocks we are currently scanning will not change out
4248  * from under us. While a scan is active, this function controls how quickly
4249  * transaction groups proceed, instead of the normal handling provided by
4250  * txg_sync_thread().
4251  */
4252 void
4253 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
4254 {
4255 	int err = 0;
4256 	dsl_scan_t *scn = dp->dp_scan;
4257 	spa_t *spa = dp->dp_spa;
4258 	state_sync_type_t sync_type = SYNC_OPTIONAL;
4259 
4260 	if (spa->spa_resilver_deferred &&
4261 	    !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
4262 		spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
4263 
4264 	/*
4265 	 * Check for scn_restart_txg before checking spa_load_state, so
4266 	 * that we can restart an old-style scan while the pool is being
4267 	 * imported (see dsl_scan_init). We also restart scans if there
4268 	 * is a deferred resilver and the user has manually disabled
4269 	 * deferred resilvers via the tunable.
4270 	 */
4271 	if (dsl_scan_restarting(scn, tx) ||
4272 	    (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
4273 		pool_scan_func_t func = POOL_SCAN_SCRUB;
4274 		dsl_scan_done(scn, B_FALSE, tx);
4275 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
4276 			func = POOL_SCAN_RESILVER;
4277 		zfs_dbgmsg("restarting scan func=%u on %s txg=%llu",
4278 		    func, dp->dp_spa->spa_name, (longlong_t)tx->tx_txg);
4279 		dsl_scan_setup_sync(&func, tx);
4280 	}
4281 
4282 	/*
4283 	 * Only process scans in sync pass 1.
4284 	 */
4285 	if (spa_sync_pass(spa) > 1)
4286 		return;
4287 
4288 	/*
4289 	 * If the spa is shutting down, then stop scanning. This will
4290 	 * ensure that the scan does not dirty any new data during the
4291 	 * shutdown phase.
4292 	 */
4293 	if (spa_shutting_down(spa))
4294 		return;
4295 
4296 	/*
4297 	 * If the scan is inactive due to a stalled async destroy, try again.
4298 	 */
4299 	if (!scn->scn_async_stalled && !dsl_scan_active(scn))
4300 		return;
4301 
4302 	/* reset scan statistics */
4303 	scn->scn_visited_this_txg = 0;
4304 	scn->scn_dedup_frees_this_txg = 0;
4305 	scn->scn_holes_this_txg = 0;
4306 	scn->scn_lt_min_this_txg = 0;
4307 	scn->scn_gt_max_this_txg = 0;
4308 	scn->scn_ddt_contained_this_txg = 0;
4309 	scn->scn_objsets_visited_this_txg = 0;
4310 	scn->scn_avg_seg_size_this_txg = 0;
4311 	scn->scn_segs_this_txg = 0;
4312 	scn->scn_avg_zio_size_this_txg = 0;
4313 	scn->scn_zios_this_txg = 0;
4314 	scn->scn_suspending = B_FALSE;
4315 	scn->scn_sync_start_time = gethrtime();
4316 	spa->spa_scrub_active = B_TRUE;
4317 
4318 	/*
4319 	 * First process the async destroys.  If we suspend, don't do
4320 	 * any scrubbing or resilvering.  This ensures that there are no
4321 	 * async destroys while we are scanning, so the scan code doesn't
4322 	 * have to worry about traversing it.  It is also faster to free the
4323 	 * blocks than to scrub them.
4324 	 */
4325 	err = dsl_process_async_destroys(dp, tx);
4326 	if (err != 0)
4327 		return;
4328 
4329 	if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
4330 		return;
4331 
4332 	/*
4333 	 * Wait a few txgs after importing to begin scanning so that
4334 	 * we can get the pool imported quickly.
4335 	 */
4336 	if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
4337 		return;
4338 
4339 	/*
4340 	 * zfs_scan_suspend_progress can be set to disable scan progress.
4341 	 * We don't want to spin the txg_sync thread, so we add a delay
4342 	 * here to simulate the time spent doing a scan. This is mostly
4343 	 * useful for testing and debugging.
4344 	 */
4345 	if (zfs_scan_suspend_progress) {
4346 		uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4347 		uint_t mintime = (scn->scn_phys.scn_func ==
4348 		    POOL_SCAN_RESILVER) ? zfs_resilver_min_time_ms :
4349 		    zfs_scrub_min_time_ms;
4350 
4351 		while (zfs_scan_suspend_progress &&
4352 		    !txg_sync_waiting(scn->scn_dp) &&
4353 		    !spa_shutting_down(scn->scn_dp->dp_spa) &&
4354 		    NSEC2MSEC(scan_time_ns) < mintime) {
4355 			delay(hz);
4356 			scan_time_ns = gethrtime() - scn->scn_sync_start_time;
4357 		}
4358 		return;
4359 	}
4360 
4361 	/*
4362 	 * Disabled by default, set zfs_scan_report_txgs to report
4363 	 * average performance over the last zfs_scan_report_txgs TXGs.
4364 	 */
4365 	if (!dsl_scan_is_paused_scrub(scn) && zfs_scan_report_txgs != 0 &&
4366 	    tx->tx_txg % zfs_scan_report_txgs == 0) {
4367 		scn->scn_issued_before_pass += spa->spa_scan_pass_issued;
4368 		spa_scan_stat_init(spa);
4369 	}
4370 
4371 	/*
4372 	 * It is possible to switch from unsorted to sorted at any time,
4373 	 * but afterwards the scan will remain sorted unless reloaded from
4374 	 * a checkpoint after a reboot.
4375 	 */
4376 	if (!zfs_scan_legacy) {
4377 		scn->scn_is_sorted = B_TRUE;
4378 		if (scn->scn_last_checkpoint == 0)
4379 			scn->scn_last_checkpoint = ddi_get_lbolt();
4380 	}
4381 
4382 	/*
4383 	 * For sorted scans, determine what kind of work we will be doing
4384 	 * this txg based on our memory limitations and whether or not we
4385 	 * need to perform a checkpoint.
4386 	 */
4387 	if (scn->scn_is_sorted) {
4388 		/*
4389 		 * If we are over our checkpoint interval, set scn_clearing
4390 		 * so that we can begin checkpointing immediately. The
4391 		 * checkpoint allows us to save a consistent bookmark
4392 		 * representing how much data we have scrubbed so far.
4393 		 * Otherwise, use the memory limit to determine if we should
4394 		 * scan for metadata or start issue scrub IOs. We accumulate
4395 		 * metadata until we hit our hard memory limit at which point
4396 		 * we issue scrub IOs until we are at our soft memory limit.
4397 		 */
4398 		if (scn->scn_checkpointing ||
4399 		    ddi_get_lbolt() - scn->scn_last_checkpoint >
4400 		    SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
4401 			if (!scn->scn_checkpointing)
4402 				zfs_dbgmsg("begin scan checkpoint for %s",
4403 				    spa->spa_name);
4404 
4405 			scn->scn_checkpointing = B_TRUE;
4406 			scn->scn_clearing = B_TRUE;
4407 		} else {
4408 			boolean_t should_clear = dsl_scan_should_clear(scn);
4409 			if (should_clear && !scn->scn_clearing) {
4410 				zfs_dbgmsg("begin scan clearing for %s",
4411 				    spa->spa_name);
4412 				scn->scn_clearing = B_TRUE;
4413 			} else if (!should_clear && scn->scn_clearing) {
4414 				zfs_dbgmsg("finish scan clearing for %s",
4415 				    spa->spa_name);
4416 				scn->scn_clearing = B_FALSE;
4417 			}
4418 		}
4419 	} else {
4420 		ASSERT0(scn->scn_checkpointing);
4421 		ASSERT0(scn->scn_clearing);
4422 	}
4423 
4424 	if (!scn->scn_clearing && scn->scn_done_txg == 0) {
4425 		/* Need to scan metadata for more blocks to scrub */
4426 		dsl_scan_phys_t *scnp = &scn->scn_phys;
4427 		taskqid_t prefetch_tqid;
4428 
4429 		/*
4430 		 * Calculate the max number of in-flight bytes for pool-wide
4431 		 * scanning operations (minimum 1MB, maximum 1/4 of arc_c_max).
4432 		 * Limits for the issuing phase are done per top-level vdev and
4433 		 * are handled separately.
4434 		 */
4435 		scn->scn_maxinflight_bytes = MIN(arc_c_max / 4, MAX(1ULL << 20,
4436 		    zfs_scan_vdev_limit * dsl_scan_count_data_disks(spa)));
4437 
4438 		if (scnp->scn_ddt_bookmark.ddb_class <=
4439 		    scnp->scn_ddt_class_max) {
4440 			ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
4441 			zfs_dbgmsg("doing scan sync for %s txg %llu; "
4442 			    "ddt bm=%llu/%llu/%llu/%llx",
4443 			    spa->spa_name,
4444 			    (longlong_t)tx->tx_txg,
4445 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
4446 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
4447 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
4448 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
4449 		} else {
4450 			zfs_dbgmsg("doing scan sync for %s txg %llu; "
4451 			    "bm=%llu/%llu/%llu/%llu",
4452 			    spa->spa_name,
4453 			    (longlong_t)tx->tx_txg,
4454 			    (longlong_t)scnp->scn_bookmark.zb_objset,
4455 			    (longlong_t)scnp->scn_bookmark.zb_object,
4456 			    (longlong_t)scnp->scn_bookmark.zb_level,
4457 			    (longlong_t)scnp->scn_bookmark.zb_blkid);
4458 		}
4459 
4460 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4461 		    NULL, ZIO_FLAG_CANFAIL);
4462 
4463 		scn->scn_prefetch_stop = B_FALSE;
4464 		prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
4465 		    dsl_scan_prefetch_thread, scn, TQ_SLEEP);
4466 		ASSERT(prefetch_tqid != TASKQID_INVALID);
4467 
4468 		dsl_pool_config_enter(dp, FTAG);
4469 		dsl_scan_visit(scn, tx);
4470 		dsl_pool_config_exit(dp, FTAG);
4471 
4472 		mutex_enter(&dp->dp_spa->spa_scrub_lock);
4473 		scn->scn_prefetch_stop = B_TRUE;
4474 		cv_broadcast(&spa->spa_scrub_io_cv);
4475 		mutex_exit(&dp->dp_spa->spa_scrub_lock);
4476 
4477 		taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
4478 		(void) zio_wait(scn->scn_zio_root);
4479 		scn->scn_zio_root = NULL;
4480 
4481 		zfs_dbgmsg("scan visited %llu blocks of %s in %llums "
4482 		    "(%llu os's, %llu holes, %llu < mintxg, "
4483 		    "%llu in ddt, %llu > maxtxg)",
4484 		    (longlong_t)scn->scn_visited_this_txg,
4485 		    spa->spa_name,
4486 		    (longlong_t)NSEC2MSEC(gethrtime() -
4487 		    scn->scn_sync_start_time),
4488 		    (longlong_t)scn->scn_objsets_visited_this_txg,
4489 		    (longlong_t)scn->scn_holes_this_txg,
4490 		    (longlong_t)scn->scn_lt_min_this_txg,
4491 		    (longlong_t)scn->scn_ddt_contained_this_txg,
4492 		    (longlong_t)scn->scn_gt_max_this_txg);
4493 
4494 		if (!scn->scn_suspending) {
4495 			ASSERT0(avl_numnodes(&scn->scn_queue));
4496 			scn->scn_done_txg = tx->tx_txg + 1;
4497 			if (scn->scn_is_sorted) {
4498 				scn->scn_checkpointing = B_TRUE;
4499 				scn->scn_clearing = B_TRUE;
4500 				scn->scn_issued_before_pass +=
4501 				    spa->spa_scan_pass_issued;
4502 				spa_scan_stat_init(spa);
4503 			}
4504 			zfs_dbgmsg("scan complete for %s txg %llu",
4505 			    spa->spa_name,
4506 			    (longlong_t)tx->tx_txg);
4507 		}
4508 	} else if (scn->scn_is_sorted && scn->scn_queues_pending != 0) {
4509 		ASSERT(scn->scn_clearing);
4510 
4511 		/* need to issue scrubbing IOs from per-vdev queues */
4512 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
4513 		    NULL, ZIO_FLAG_CANFAIL);
4514 		scan_io_queues_run(scn);
4515 		(void) zio_wait(scn->scn_zio_root);
4516 		scn->scn_zio_root = NULL;
4517 
4518 		/* calculate and dprintf the current memory usage */
4519 		(void) dsl_scan_should_clear(scn);
4520 		dsl_scan_update_stats(scn);
4521 
4522 		zfs_dbgmsg("scan issued %llu blocks for %s (%llu segs) "
4523 		    "in %llums (avg_block_size = %llu, avg_seg_size = %llu)",
4524 		    (longlong_t)scn->scn_zios_this_txg,
4525 		    spa->spa_name,
4526 		    (longlong_t)scn->scn_segs_this_txg,
4527 		    (longlong_t)NSEC2MSEC(gethrtime() -
4528 		    scn->scn_sync_start_time),
4529 		    (longlong_t)scn->scn_avg_zio_size_this_txg,
4530 		    (longlong_t)scn->scn_avg_seg_size_this_txg);
4531 	} else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
4532 		/* Finished with everything. Mark the scrub as complete */
4533 		zfs_dbgmsg("scan issuing complete txg %llu for %s",
4534 		    (longlong_t)tx->tx_txg,
4535 		    spa->spa_name);
4536 		ASSERT3U(scn->scn_done_txg, !=, 0);
4537 		ASSERT0(spa->spa_scrub_inflight);
4538 		ASSERT0(scn->scn_queues_pending);
4539 		dsl_scan_done(scn, B_TRUE, tx);
4540 		sync_type = SYNC_MANDATORY;
4541 	}
4542 
4543 	dsl_scan_sync_state(scn, tx, sync_type);
4544 }
4545 
4546 static void
4547 count_block_issued(spa_t *spa, const blkptr_t *bp, boolean_t all)
4548 {
4549 	/*
4550 	 * Don't count embedded bp's, since we already did the work of
4551 	 * scanning these when we scanned the containing block.
4552 	 */
4553 	if (BP_IS_EMBEDDED(bp))
4554 		return;
4555 
4556 	/*
4557 	 * Update the spa's stats on how many bytes we have issued.
4558 	 * Sequential scrubs create a zio for each DVA of the bp. Each
4559 	 * of these will include all DVAs for repair purposes, but the
4560 	 * zio code will only try the first one unless there is an issue.
4561 	 * Therefore, we should only count the first DVA for these IOs.
4562 	 */
4563 	atomic_add_64(&spa->spa_scan_pass_issued,
4564 	    all ? BP_GET_ASIZE(bp) : DVA_GET_ASIZE(&bp->blk_dva[0]));
4565 }
4566 
4567 static void
4568 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
4569 {
4570 	/*
4571 	 * If we resume after a reboot, zab will be NULL; don't record
4572 	 * incomplete stats in that case.
4573 	 */
4574 	if (zab == NULL)
4575 		return;
4576 
4577 	for (int i = 0; i < 4; i++) {
4578 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
4579 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
4580 
4581 		if (t & DMU_OT_NEWTYPE)
4582 			t = DMU_OT_OTHER;
4583 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
4584 		int equal;
4585 
4586 		zb->zb_count++;
4587 		zb->zb_asize += BP_GET_ASIZE(bp);
4588 		zb->zb_lsize += BP_GET_LSIZE(bp);
4589 		zb->zb_psize += BP_GET_PSIZE(bp);
4590 		zb->zb_gangs += BP_COUNT_GANG(bp);
4591 
4592 		switch (BP_GET_NDVAS(bp)) {
4593 		case 2:
4594 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4595 			    DVA_GET_VDEV(&bp->blk_dva[1]))
4596 				zb->zb_ditto_2_of_2_samevdev++;
4597 			break;
4598 		case 3:
4599 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4600 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
4601 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
4602 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
4603 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
4604 			    DVA_GET_VDEV(&bp->blk_dva[2]));
4605 			if (equal == 1)
4606 				zb->zb_ditto_2_of_3_samevdev++;
4607 			else if (equal == 3)
4608 				zb->zb_ditto_3_of_3_samevdev++;
4609 			break;
4610 		}
4611 	}
4612 }
4613 
4614 static void
4615 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
4616 {
4617 	avl_index_t idx;
4618 	dsl_scan_t *scn = queue->q_scn;
4619 
4620 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4621 
4622 	if (unlikely(avl_is_empty(&queue->q_sios_by_addr)))
4623 		atomic_add_64(&scn->scn_queues_pending, 1);
4624 	if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
4625 		/* block is already scheduled for reading */
4626 		sio_free(sio);
4627 		return;
4628 	}
4629 	avl_insert(&queue->q_sios_by_addr, sio, idx);
4630 	queue->q_sio_memused += SIO_GET_MUSED(sio);
4631 	range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio),
4632 	    SIO_GET_ASIZE(sio));
4633 }
4634 
4635 /*
4636  * Given all the info we got from our metadata scanning process, we
4637  * construct a scan_io_t and insert it into the scan sorting queue. The
4638  * I/O must already be suitable for us to process. This is controlled
4639  * by dsl_scan_enqueue().
4640  */
4641 static void
4642 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
4643     int zio_flags, const zbookmark_phys_t *zb)
4644 {
4645 	scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
4646 
4647 	ASSERT0(BP_IS_GANG(bp));
4648 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4649 
4650 	bp2sio(bp, sio, dva_i);
4651 	sio->sio_flags = zio_flags;
4652 	sio->sio_zb = *zb;
4653 
4654 	queue->q_last_ext_addr = -1;
4655 	scan_io_queue_insert_impl(queue, sio);
4656 }
4657 
4658 /*
4659  * Given a set of I/O parameters as discovered by the metadata traversal
4660  * process, attempts to place the I/O into the sorted queues (if allowed),
4661  * or immediately executes the I/O.
4662  */
4663 static void
4664 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4665     const zbookmark_phys_t *zb)
4666 {
4667 	spa_t *spa = dp->dp_spa;
4668 
4669 	ASSERT(!BP_IS_EMBEDDED(bp));
4670 
4671 	/*
4672 	 * Gang blocks are hard to issue sequentially, so we just issue them
4673 	 * here immediately instead of queuing them.
4674 	 */
4675 	if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
4676 		scan_exec_io(dp, bp, zio_flags, zb, NULL);
4677 		return;
4678 	}
4679 
4680 	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
4681 		dva_t dva;
4682 		vdev_t *vdev;
4683 
4684 		dva = bp->blk_dva[i];
4685 		vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
4686 		ASSERT(vdev != NULL);
4687 
4688 		mutex_enter(&vdev->vdev_scan_io_queue_lock);
4689 		if (vdev->vdev_scan_io_queue == NULL)
4690 			vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
4691 		ASSERT(dp->dp_scan != NULL);
4692 		scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
4693 		    i, zio_flags, zb);
4694 		mutex_exit(&vdev->vdev_scan_io_queue_lock);
4695 	}
4696 }
4697 
4698 static int
4699 dsl_scan_scrub_cb(dsl_pool_t *dp,
4700     const blkptr_t *bp, const zbookmark_phys_t *zb)
4701 {
4702 	dsl_scan_t *scn = dp->dp_scan;
4703 	spa_t *spa = dp->dp_spa;
4704 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
4705 	size_t psize = BP_GET_PSIZE(bp);
4706 	boolean_t needs_io = B_FALSE;
4707 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
4708 
4709 	count_block(dp->dp_blkstats, bp);
4710 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
4711 	    phys_birth >= scn->scn_phys.scn_max_txg) {
4712 		count_block_issued(spa, bp, B_TRUE);
4713 		return (0);
4714 	}
4715 
4716 	/* Embedded BP's have phys_birth==0, so we reject them above. */
4717 	ASSERT(!BP_IS_EMBEDDED(bp));
4718 
4719 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
4720 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
4721 		zio_flags |= ZIO_FLAG_SCRUB;
4722 		needs_io = B_TRUE;
4723 	} else {
4724 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
4725 		zio_flags |= ZIO_FLAG_RESILVER;
4726 		needs_io = B_FALSE;
4727 	}
4728 
4729 	/* If it's an intent log block, failure is expected. */
4730 	if (zb->zb_level == ZB_ZIL_LEVEL)
4731 		zio_flags |= ZIO_FLAG_SPECULATIVE;
4732 
4733 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
4734 		const dva_t *dva = &bp->blk_dva[d];
4735 
4736 		/*
4737 		 * Keep track of how much data we've examined so that
4738 		 * zpool(8) status can make useful progress reports.
4739 		 */
4740 		uint64_t asize = DVA_GET_ASIZE(dva);
4741 		scn->scn_phys.scn_examined += asize;
4742 		spa->spa_scan_pass_exam += asize;
4743 
4744 		/* if it's a resilver, this may not be in the target range */
4745 		if (!needs_io)
4746 			needs_io = dsl_scan_need_resilver(spa, dva, psize,
4747 			    phys_birth);
4748 	}
4749 
4750 	if (needs_io && !zfs_no_scrub_io) {
4751 		dsl_scan_enqueue(dp, bp, zio_flags, zb);
4752 	} else {
4753 		count_block_issued(spa, bp, B_TRUE);
4754 	}
4755 
4756 	/* do not relocate this block */
4757 	return (0);
4758 }
4759 
4760 static void
4761 dsl_scan_scrub_done(zio_t *zio)
4762 {
4763 	spa_t *spa = zio->io_spa;
4764 	blkptr_t *bp = zio->io_bp;
4765 	dsl_scan_io_queue_t *queue = zio->io_private;
4766 
4767 	abd_free(zio->io_abd);
4768 
4769 	if (queue == NULL) {
4770 		mutex_enter(&spa->spa_scrub_lock);
4771 		ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
4772 		spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
4773 		cv_broadcast(&spa->spa_scrub_io_cv);
4774 		mutex_exit(&spa->spa_scrub_lock);
4775 	} else {
4776 		mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
4777 		ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
4778 		queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
4779 		cv_broadcast(&queue->q_zio_cv);
4780 		mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
4781 	}
4782 
4783 	if (zio->io_error && (zio->io_error != ECKSUM ||
4784 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
4785 		if (dsl_errorscrubbing(spa->spa_dsl_pool) &&
4786 		    !dsl_errorscrub_is_paused(spa->spa_dsl_pool->dp_scan)) {
4787 			atomic_inc_64(&spa->spa_dsl_pool->dp_scan
4788 			    ->errorscrub_phys.dep_errors);
4789 		} else {
4790 			atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys
4791 			    .scn_errors);
4792 		}
4793 	}
4794 }
4795 
4796 /*
4797  * Given a scanning zio's information, executes the zio. The zio need
4798  * not necessarily be only sortable, this function simply executes the
4799  * zio, no matter what it is. The optional queue argument allows the
4800  * caller to specify that they want per top level vdev IO rate limiting
4801  * instead of the legacy global limiting.
4802  */
4803 static void
4804 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
4805     const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
4806 {
4807 	spa_t *spa = dp->dp_spa;
4808 	dsl_scan_t *scn = dp->dp_scan;
4809 	size_t size = BP_GET_PSIZE(bp);
4810 	abd_t *data = abd_alloc_for_io(size, B_FALSE);
4811 	zio_t *pio;
4812 
4813 	if (queue == NULL) {
4814 		ASSERT3U(scn->scn_maxinflight_bytes, >, 0);
4815 		mutex_enter(&spa->spa_scrub_lock);
4816 		while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
4817 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
4818 		spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
4819 		mutex_exit(&spa->spa_scrub_lock);
4820 		pio = scn->scn_zio_root;
4821 	} else {
4822 		kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
4823 
4824 		ASSERT3U(queue->q_maxinflight_bytes, >, 0);
4825 		mutex_enter(q_lock);
4826 		while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
4827 			cv_wait(&queue->q_zio_cv, q_lock);
4828 		queue->q_inflight_bytes += BP_GET_PSIZE(bp);
4829 		pio = queue->q_zio;
4830 		mutex_exit(q_lock);
4831 	}
4832 
4833 	ASSERT(pio != NULL);
4834 	count_block_issued(spa, bp, queue == NULL);
4835 	zio_nowait(zio_read(pio, spa, bp, data, size, dsl_scan_scrub_done,
4836 	    queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
4837 }
4838 
4839 /*
4840  * This is the primary extent sorting algorithm. We balance two parameters:
4841  * 1) how many bytes of I/O are in an extent
4842  * 2) how well the extent is filled with I/O (as a fraction of its total size)
4843  * Since we allow extents to have gaps between their constituent I/Os, it's
4844  * possible to have a fairly large extent that contains the same amount of
4845  * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
4846  * The algorithm sorts based on a score calculated from the extent's size,
4847  * the relative fill volume (in %) and a "fill weight" parameter that controls
4848  * the split between whether we prefer larger extents or more well populated
4849  * extents:
4850  *
4851  * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
4852  *
4853  * Example:
4854  * 1) assume extsz = 64 MiB
4855  * 2) assume fill = 32 MiB (extent is half full)
4856  * 3) assume fill_weight = 3
4857  * 4)	SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
4858  *	SCORE = 32M + (50 * 3 * 32M) / 100
4859  *	SCORE = 32M + (4800M / 100)
4860  *	SCORE = 32M + 48M
4861  *	         ^     ^
4862  *	         |     +--- final total relative fill-based score
4863  *	         +--------- final total fill-based score
4864  *	SCORE = 80M
4865  *
4866  * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
4867  * extents that are more completely filled (in a 3:2 ratio) vs just larger.
4868  * Note that as an optimization, we replace multiplication and division by
4869  * 100 with bitshifting by 7 (which effectively multiplies and divides by 128).
4870  *
4871  * Since we do not care if one extent is only few percent better than another,
4872  * compress the score into 6 bits via binary logarithm AKA highbit64() and
4873  * put into otherwise unused due to ashift high bits of offset.  This allows
4874  * to reduce q_exts_by_size B-tree elements to only 64 bits and compare them
4875  * with single operation.  Plus it makes scrubs more sequential and reduces
4876  * chances that minor extent change move it within the B-tree.
4877  */
4878 __attribute__((always_inline)) inline
4879 static int
4880 ext_size_compare(const void *x, const void *y)
4881 {
4882 	const uint64_t *a = x, *b = y;
4883 
4884 	return (TREE_CMP(*a, *b));
4885 }
4886 
4887 ZFS_BTREE_FIND_IN_BUF_FUNC(ext_size_find_in_buf, uint64_t,
4888     ext_size_compare)
4889 
4890 static void
4891 ext_size_create(range_tree_t *rt, void *arg)
4892 {
4893 	(void) rt;
4894 	zfs_btree_t *size_tree = arg;
4895 
4896 	zfs_btree_create(size_tree, ext_size_compare, ext_size_find_in_buf,
4897 	    sizeof (uint64_t));
4898 }
4899 
4900 static void
4901 ext_size_destroy(range_tree_t *rt, void *arg)
4902 {
4903 	(void) rt;
4904 	zfs_btree_t *size_tree = arg;
4905 	ASSERT0(zfs_btree_numnodes(size_tree));
4906 
4907 	zfs_btree_destroy(size_tree);
4908 }
4909 
4910 static uint64_t
4911 ext_size_value(range_tree_t *rt, range_seg_gap_t *rsg)
4912 {
4913 	(void) rt;
4914 	uint64_t size = rsg->rs_end - rsg->rs_start;
4915 	uint64_t score = rsg->rs_fill + ((((rsg->rs_fill << 7) / size) *
4916 	    fill_weight * rsg->rs_fill) >> 7);
4917 	ASSERT3U(rt->rt_shift, >=, 8);
4918 	return (((uint64_t)(64 - highbit64(score)) << 56) | rsg->rs_start);
4919 }
4920 
4921 static void
4922 ext_size_add(range_tree_t *rt, range_seg_t *rs, void *arg)
4923 {
4924 	zfs_btree_t *size_tree = arg;
4925 	ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4926 	uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4927 	zfs_btree_add(size_tree, &v);
4928 }
4929 
4930 static void
4931 ext_size_remove(range_tree_t *rt, range_seg_t *rs, void *arg)
4932 {
4933 	zfs_btree_t *size_tree = arg;
4934 	ASSERT3U(rt->rt_type, ==, RANGE_SEG_GAP);
4935 	uint64_t v = ext_size_value(rt, (range_seg_gap_t *)rs);
4936 	zfs_btree_remove(size_tree, &v);
4937 }
4938 
4939 static void
4940 ext_size_vacate(range_tree_t *rt, void *arg)
4941 {
4942 	zfs_btree_t *size_tree = arg;
4943 	zfs_btree_clear(size_tree);
4944 	zfs_btree_destroy(size_tree);
4945 
4946 	ext_size_create(rt, arg);
4947 }
4948 
4949 static const range_tree_ops_t ext_size_ops = {
4950 	.rtop_create = ext_size_create,
4951 	.rtop_destroy = ext_size_destroy,
4952 	.rtop_add = ext_size_add,
4953 	.rtop_remove = ext_size_remove,
4954 	.rtop_vacate = ext_size_vacate
4955 };
4956 
4957 /*
4958  * Comparator for the q_sios_by_addr tree. Sorting is simply performed
4959  * based on LBA-order (from lowest to highest).
4960  */
4961 static int
4962 sio_addr_compare(const void *x, const void *y)
4963 {
4964 	const scan_io_t *a = x, *b = y;
4965 
4966 	return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4967 }
4968 
4969 /* IO queues are created on demand when they are needed. */
4970 static dsl_scan_io_queue_t *
4971 scan_io_queue_create(vdev_t *vd)
4972 {
4973 	dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4974 	dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4975 
4976 	q->q_scn = scn;
4977 	q->q_vd = vd;
4978 	q->q_sio_memused = 0;
4979 	q->q_last_ext_addr = -1;
4980 	cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
4981 	q->q_exts_by_addr = range_tree_create_gap(&ext_size_ops, RANGE_SEG_GAP,
4982 	    &q->q_exts_by_size, 0, vd->vdev_ashift, zfs_scan_max_ext_gap);
4983 	avl_create(&q->q_sios_by_addr, sio_addr_compare,
4984 	    sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
4985 
4986 	return (q);
4987 }
4988 
4989 /*
4990  * Destroys a scan queue and all segments and scan_io_t's contained in it.
4991  * No further execution of I/O occurs, anything pending in the queue is
4992  * simply freed without being executed.
4993  */
4994 void
4995 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
4996 {
4997 	dsl_scan_t *scn = queue->q_scn;
4998 	scan_io_t *sio;
4999 	void *cookie = NULL;
5000 
5001 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
5002 
5003 	if (!avl_is_empty(&queue->q_sios_by_addr))
5004 		atomic_add_64(&scn->scn_queues_pending, -1);
5005 	while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
5006 	    NULL) {
5007 		ASSERT(range_tree_contains(queue->q_exts_by_addr,
5008 		    SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
5009 		queue->q_sio_memused -= SIO_GET_MUSED(sio);
5010 		sio_free(sio);
5011 	}
5012 
5013 	ASSERT0(queue->q_sio_memused);
5014 	range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
5015 	range_tree_destroy(queue->q_exts_by_addr);
5016 	avl_destroy(&queue->q_sios_by_addr);
5017 	cv_destroy(&queue->q_zio_cv);
5018 
5019 	kmem_free(queue, sizeof (*queue));
5020 }
5021 
5022 /*
5023  * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
5024  * called on behalf of vdev_top_transfer when creating or destroying
5025  * a mirror vdev due to zpool attach/detach.
5026  */
5027 void
5028 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
5029 {
5030 	mutex_enter(&svd->vdev_scan_io_queue_lock);
5031 	mutex_enter(&tvd->vdev_scan_io_queue_lock);
5032 
5033 	VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
5034 	tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
5035 	svd->vdev_scan_io_queue = NULL;
5036 	if (tvd->vdev_scan_io_queue != NULL)
5037 		tvd->vdev_scan_io_queue->q_vd = tvd;
5038 
5039 	mutex_exit(&tvd->vdev_scan_io_queue_lock);
5040 	mutex_exit(&svd->vdev_scan_io_queue_lock);
5041 }
5042 
5043 static void
5044 scan_io_queues_destroy(dsl_scan_t *scn)
5045 {
5046 	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
5047 
5048 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
5049 		vdev_t *tvd = rvd->vdev_child[i];
5050 
5051 		mutex_enter(&tvd->vdev_scan_io_queue_lock);
5052 		if (tvd->vdev_scan_io_queue != NULL)
5053 			dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
5054 		tvd->vdev_scan_io_queue = NULL;
5055 		mutex_exit(&tvd->vdev_scan_io_queue_lock);
5056 	}
5057 }
5058 
5059 static void
5060 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
5061 {
5062 	dsl_pool_t *dp = spa->spa_dsl_pool;
5063 	dsl_scan_t *scn = dp->dp_scan;
5064 	vdev_t *vdev;
5065 	kmutex_t *q_lock;
5066 	dsl_scan_io_queue_t *queue;
5067 	scan_io_t *srch_sio, *sio;
5068 	avl_index_t idx;
5069 	uint64_t start, size;
5070 
5071 	vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
5072 	ASSERT(vdev != NULL);
5073 	q_lock = &vdev->vdev_scan_io_queue_lock;
5074 	queue = vdev->vdev_scan_io_queue;
5075 
5076 	mutex_enter(q_lock);
5077 	if (queue == NULL) {
5078 		mutex_exit(q_lock);
5079 		return;
5080 	}
5081 
5082 	srch_sio = sio_alloc(BP_GET_NDVAS(bp));
5083 	bp2sio(bp, srch_sio, dva_i);
5084 	start = SIO_GET_OFFSET(srch_sio);
5085 	size = SIO_GET_ASIZE(srch_sio);
5086 
5087 	/*
5088 	 * We can find the zio in two states:
5089 	 * 1) Cold, just sitting in the queue of zio's to be issued at
5090 	 *	some point in the future. In this case, all we do is
5091 	 *	remove the zio from the q_sios_by_addr tree, decrement
5092 	 *	its data volume from the containing range_seg_t and
5093 	 *	resort the q_exts_by_size tree to reflect that the
5094 	 *	range_seg_t has lost some of its 'fill'. We don't shorten
5095 	 *	the range_seg_t - this is usually rare enough not to be
5096 	 *	worth the extra hassle of trying keep track of precise
5097 	 *	extent boundaries.
5098 	 * 2) Hot, where the zio is currently in-flight in
5099 	 *	dsl_scan_issue_ios. In this case, we can't simply
5100 	 *	reach in and stop the in-flight zio's, so we instead
5101 	 *	block the caller. Eventually, dsl_scan_issue_ios will
5102 	 *	be done with issuing the zio's it gathered and will
5103 	 *	signal us.
5104 	 */
5105 	sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
5106 	sio_free(srch_sio);
5107 
5108 	if (sio != NULL) {
5109 		blkptr_t tmpbp;
5110 
5111 		/* Got it while it was cold in the queue */
5112 		ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
5113 		ASSERT3U(size, ==, SIO_GET_ASIZE(sio));
5114 		avl_remove(&queue->q_sios_by_addr, sio);
5115 		if (avl_is_empty(&queue->q_sios_by_addr))
5116 			atomic_add_64(&scn->scn_queues_pending, -1);
5117 		queue->q_sio_memused -= SIO_GET_MUSED(sio);
5118 
5119 		ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
5120 		range_tree_remove_fill(queue->q_exts_by_addr, start, size);
5121 
5122 		/* count the block as though we issued it */
5123 		sio2bp(sio, &tmpbp);
5124 		count_block_issued(spa, &tmpbp, B_FALSE);
5125 
5126 		sio_free(sio);
5127 	}
5128 	mutex_exit(q_lock);
5129 }
5130 
5131 /*
5132  * Callback invoked when a zio_free() zio is executing. This needs to be
5133  * intercepted to prevent the zio from deallocating a particular portion
5134  * of disk space and it then getting reallocated and written to, while we
5135  * still have it queued up for processing.
5136  */
5137 void
5138 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
5139 {
5140 	dsl_pool_t *dp = spa->spa_dsl_pool;
5141 	dsl_scan_t *scn = dp->dp_scan;
5142 
5143 	ASSERT(!BP_IS_EMBEDDED(bp));
5144 	ASSERT(scn != NULL);
5145 	if (!dsl_scan_is_running(scn))
5146 		return;
5147 
5148 	for (int i = 0; i < BP_GET_NDVAS(bp); i++)
5149 		dsl_scan_freed_dva(spa, bp, i);
5150 }
5151 
5152 /*
5153  * Check if a vdev needs resilvering (non-empty DTL), if so, and resilver has
5154  * not started, start it. Otherwise, only restart if max txg in DTL range is
5155  * greater than the max txg in the current scan. If the DTL max is less than
5156  * the scan max, then the vdev has not missed any new data since the resilver
5157  * started, so a restart is not needed.
5158  */
5159 void
5160 dsl_scan_assess_vdev(dsl_pool_t *dp, vdev_t *vd)
5161 {
5162 	uint64_t min, max;
5163 
5164 	if (!vdev_resilver_needed(vd, &min, &max))
5165 		return;
5166 
5167 	if (!dsl_scan_resilvering(dp)) {
5168 		spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
5169 		return;
5170 	}
5171 
5172 	if (max <= dp->dp_scan->scn_phys.scn_max_txg)
5173 		return;
5174 
5175 	/* restart is needed, check if it can be deferred */
5176 	if (spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
5177 		vdev_defer_resilver(vd);
5178 	else
5179 		spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER);
5180 }
5181 
5182 ZFS_MODULE_PARAM(zfs, zfs_, scan_vdev_limit, U64, ZMOD_RW,
5183 	"Max bytes in flight per leaf vdev for scrubs and resilvers");
5184 
5185 ZFS_MODULE_PARAM(zfs, zfs_, scrub_min_time_ms, UINT, ZMOD_RW,
5186 	"Min millisecs to scrub per txg");
5187 
5188 ZFS_MODULE_PARAM(zfs, zfs_, obsolete_min_time_ms, UINT, ZMOD_RW,
5189 	"Min millisecs to obsolete per txg");
5190 
5191 ZFS_MODULE_PARAM(zfs, zfs_, free_min_time_ms, UINT, ZMOD_RW,
5192 	"Min millisecs to free per txg");
5193 
5194 ZFS_MODULE_PARAM(zfs, zfs_, resilver_min_time_ms, UINT, ZMOD_RW,
5195 	"Min millisecs to resilver per txg");
5196 
5197 ZFS_MODULE_PARAM(zfs, zfs_, scan_suspend_progress, INT, ZMOD_RW,
5198 	"Set to prevent scans from progressing");
5199 
5200 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_io, INT, ZMOD_RW,
5201 	"Set to disable scrub I/O");
5202 
5203 ZFS_MODULE_PARAM(zfs, zfs_, no_scrub_prefetch, INT, ZMOD_RW,
5204 	"Set to disable scrub prefetching");
5205 
5206 ZFS_MODULE_PARAM(zfs, zfs_, async_block_max_blocks, U64, ZMOD_RW,
5207 	"Max number of blocks freed in one txg");
5208 
5209 ZFS_MODULE_PARAM(zfs, zfs_, max_async_dedup_frees, U64, ZMOD_RW,
5210 	"Max number of dedup blocks freed in one txg");
5211 
5212 ZFS_MODULE_PARAM(zfs, zfs_, free_bpobj_enabled, INT, ZMOD_RW,
5213 	"Enable processing of the free_bpobj");
5214 
5215 ZFS_MODULE_PARAM(zfs, zfs_, scan_blkstats, INT, ZMOD_RW,
5216 	"Enable block statistics calculation during scrub");
5217 
5218 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_fact, UINT, ZMOD_RW,
5219 	"Fraction of RAM for scan hard limit");
5220 
5221 ZFS_MODULE_PARAM(zfs, zfs_, scan_issue_strategy, UINT, ZMOD_RW,
5222 	"IO issuing strategy during scrubbing. 0 = default, 1 = LBA, 2 = size");
5223 
5224 ZFS_MODULE_PARAM(zfs, zfs_, scan_legacy, INT, ZMOD_RW,
5225 	"Scrub using legacy non-sequential method");
5226 
5227 ZFS_MODULE_PARAM(zfs, zfs_, scan_checkpoint_intval, UINT, ZMOD_RW,
5228 	"Scan progress on-disk checkpointing interval");
5229 
5230 ZFS_MODULE_PARAM(zfs, zfs_, scan_max_ext_gap, U64, ZMOD_RW,
5231 	"Max gap in bytes between sequential scrub / resilver I/Os");
5232 
5233 ZFS_MODULE_PARAM(zfs, zfs_, scan_mem_lim_soft_fact, UINT, ZMOD_RW,
5234 	"Fraction of hard limit used as soft limit");
5235 
5236 ZFS_MODULE_PARAM(zfs, zfs_, scan_strict_mem_lim, INT, ZMOD_RW,
5237 	"Tunable to attempt to reduce lock contention");
5238 
5239 ZFS_MODULE_PARAM(zfs, zfs_, scan_fill_weight, UINT, ZMOD_RW,
5240 	"Tunable to adjust bias towards more filled segments during scans");
5241 
5242 ZFS_MODULE_PARAM(zfs, zfs_, scan_report_txgs, UINT, ZMOD_RW,
5243 	"Tunable to report resilver performance over the last N txgs");
5244 
5245 ZFS_MODULE_PARAM(zfs, zfs_, resilver_disable_defer, INT, ZMOD_RW,
5246 	"Process all resilvers immediately");
5247 
5248 ZFS_MODULE_PARAM(zfs, zfs_, scrub_error_blocks_per_txg, UINT, ZMOD_RW,
5249 	"Error blocks to be scrubbed in one txg");
5250 /* END CSTYLED */
5251