1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24 * Copyright 2016 Gary Mills
25 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
26 * Copyright 2019 Joyent, Inc.
27 * Copyright (c) 2017 Datto 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/zap.h>
41#include <sys/zio.h>
42#include <sys/zfs_context.h>
43#include <sys/fs/zfs.h>
44#include <sys/zfs_znode.h>
45#include <sys/spa_impl.h>
46#include <sys/vdev_impl.h>
47#include <sys/zil_impl.h>
48#include <sys/zio_checksum.h>
49#include <sys/ddt.h>
50#include <sys/sa.h>
51#include <sys/sa_impl.h>
52#include <sys/zfeature.h>
53#include <sys/abd.h>
54#include <sys/range_tree.h>
55#ifdef _KERNEL
56#include <sys/zfs_vfsops.h>
57#endif
58
59/*
60 * Grand theory statement on scan queue sorting
61 *
62 * Scanning is implemented by recursively traversing all indirection levels
63 * in an object and reading all blocks referenced from said objects. This
64 * results in us approximately traversing the object from lowest logical
65 * offset to the highest. For best performance, we would want the logical
66 * blocks to be physically contiguous. However, this is frequently not the
67 * case with pools given the allocation patterns of copy-on-write filesystems.
68 * So instead, we put the I/Os into a reordering queue and issue them in a
69 * way that will most benefit physical disks (LBA-order).
70 *
71 * Queue management:
72 *
73 * Ideally, we would want to scan all metadata and queue up all block I/O
74 * prior to starting to issue it, because that allows us to do an optimal
75 * sorting job. This can however consume large amounts of memory. Therefore
76 * we continuously monitor the size of the queues and constrain them to 5%
77 * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
78 * limit, we clear out a few of the largest extents at the head of the queues
79 * to make room for more scanning. Hopefully, these extents will be fairly
80 * large and contiguous, allowing us to approach sequential I/O throughput
81 * even without a fully sorted tree.
82 *
83 * Metadata scanning takes place in dsl_scan_visit(), which is called from
84 * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
85 * metadata on the pool, or we need to make room in memory because our
86 * queues are too large, dsl_scan_visit() is postponed and
87 * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
88 * that metadata scanning and queued I/O issuing are mutually exclusive. This
89 * allows us to provide maximum sequential I/O throughput for the majority of
90 * I/O's issued since sequential I/O performance is significantly negatively
91 * impacted if it is interleaved with random I/O.
92 *
93 * Implementation Notes
94 *
95 * One side effect of the queued scanning algorithm is that the scanning code
96 * needs to be notified whenever a block is freed. This is needed to allow
97 * the scanning code to remove these I/Os from the issuing queue. Additionally,
98 * we do not attempt to queue gang blocks to be issued sequentially since this
99 * is very hard to do and would have an extremely limited performance benefit.
100 * Instead, we simply issue gang I/Os as soon as we find them using the legacy
101 * algorithm.
102 *
103 * Backwards compatibility
104 *
105 * This new algorithm is backwards compatible with the legacy on-disk data
106 * structures (and therefore does not require a new feature flag).
107 * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
108 * will stop scanning metadata (in logical order) and wait for all outstanding
109 * sorted I/O to complete. Once this is done, we write out a checkpoint
110 * bookmark, indicating that we have scanned everything logically before it.
111 * If the pool is imported on a machine without the new sorting algorithm,
112 * the scan simply resumes from the last checkpoint using the legacy algorithm.
113 */
114
115typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
116    const zbookmark_phys_t *);
117
118static scan_cb_t dsl_scan_scrub_cb;
119
120static int scan_ds_queue_compare(const void *a, const void *b);
121static int scan_prefetch_queue_compare(const void *a, const void *b);
122static void scan_ds_queue_clear(dsl_scan_t *scn);
123static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
124    uint64_t *txg);
125static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
126static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
127static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
128
129extern int zfs_vdev_async_write_active_min_dirty_percent;
130
131/*
132 * By default zfs will check to ensure it is not over the hard memory
133 * limit before each txg. If finer-grained control of this is needed
134 * this value can be set to 1 to enable checking before scanning each
135 * block.
136 */
137int zfs_scan_strict_mem_lim = B_FALSE;
138
139/*
140 * Maximum number of parallelly executing I/Os per top-level vdev.
141 * Tune with care. Very high settings (hundreds) are known to trigger
142 * some firmware bugs and resets on certain SSDs.
143 */
144int zfs_top_maxinflight = 32;		/* maximum I/Os per top-level */
145unsigned int zfs_resilver_delay = 2;	/* number of ticks to delay resilver */
146unsigned int zfs_scrub_delay = 4;	/* number of ticks to delay scrub */
147unsigned int zfs_scan_idle = 50;	/* idle window in clock ticks */
148
149/*
150 * Maximum number of parallelly executed bytes per leaf vdev. We attempt
151 * to strike a balance here between keeping the vdev queues full of I/Os
152 * at all times and not overflowing the queues to cause long latency,
153 * which would cause long txg sync times. No matter what, we will not
154 * overload the drives with I/O, since that is protected by
155 * zfs_vdev_scrub_max_active.
156 */
157unsigned long zfs_scan_vdev_limit = 4 << 20;
158
159int zfs_scan_issue_strategy = 0;
160int zfs_scan_legacy = B_FALSE;	/* don't queue & sort zios, go direct */
161uint64_t zfs_scan_max_ext_gap = 2 << 20;	/* in bytes */
162
163unsigned int zfs_scan_checkpoint_intval = 7200;	/* seconds */
164#define	ZFS_SCAN_CHECKPOINT_INTVAL	SEC_TO_TICK(zfs_scan_checkpoint_intval)
165
166/*
167 * fill_weight is non-tunable at runtime, so we copy it at module init from
168 * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
169 * break queue sorting.
170 */
171uint64_t zfs_scan_fill_weight = 3;
172static uint64_t fill_weight;
173
174/* See dsl_scan_should_clear() for details on the memory limit tunables */
175uint64_t zfs_scan_mem_lim_min = 16 << 20;	/* bytes */
176uint64_t zfs_scan_mem_lim_soft_max = 128 << 20;	/* bytes */
177int zfs_scan_mem_lim_fact = 20;		/* fraction of physmem */
178int zfs_scan_mem_lim_soft_fact = 20;	/* fraction of mem lim above */
179
180unsigned int zfs_scrub_min_time_ms = 1000; /* min millisecs to scrub per txg */
181unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
182/* min millisecs to obsolete per txg */
183unsigned int zfs_obsolete_min_time_ms = 500;
184/* min millisecs to resilver per txg */
185unsigned int zfs_resilver_min_time_ms = 3000;
186int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
187boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
188boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
189enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
190/* max number of blocks to free in a single TXG */
191uint64_t zfs_async_block_max_blocks = UINT64_MAX;
192
193int zfs_resilver_disable_defer = 0; /* set to disable resilver deferring */
194
195/*
196 * We wait a few txgs after importing a pool to begin scanning so that
197 * the import / mounting code isn't held up by scrub / resilver IO.
198 * Unfortunately, it is a bit difficult to determine exactly how long
199 * this will take since userspace will trigger fs mounts asynchronously
200 * and the kernel will create zvol minors asynchronously. As a result,
201 * the value provided here is a bit arbitrary, but represents a
202 * reasonable estimate of how many txgs it will take to finish fully
203 * importing a pool
204 */
205#define	SCAN_IMPORT_WAIT_TXGS		5
206
207
208#define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
209	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
210	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
211
212extern int zfs_txg_timeout;
213
214/*
215 * Enable/disable the processing of the free_bpobj object.
216 */
217boolean_t zfs_free_bpobj_enabled = B_TRUE;
218
219/* the order has to match pool_scan_type */
220static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
221	NULL,
222	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
223	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
224};
225
226/* In core node for the scn->scn_queue. Represents a dataset to be scanned */
227typedef struct {
228	uint64_t	sds_dsobj;
229	uint64_t	sds_txg;
230	avl_node_t	sds_node;
231} scan_ds_t;
232
233/*
234 * This controls what conditions are placed on dsl_scan_sync_state():
235 * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
236 * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
237 * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
238 *	write out the scn_phys_cached version.
239 * See dsl_scan_sync_state for details.
240 */
241typedef enum {
242	SYNC_OPTIONAL,
243	SYNC_MANDATORY,
244	SYNC_CACHED
245} state_sync_type_t;
246
247/*
248 * This struct represents the minimum information needed to reconstruct a
249 * zio for sequential scanning. This is useful because many of these will
250 * accumulate in the sequential IO queues before being issued, so saving
251 * memory matters here.
252 */
253typedef struct scan_io {
254	/* fields from blkptr_t */
255	uint64_t		sio_blk_prop;
256	uint64_t		sio_phys_birth;
257	uint64_t		sio_birth;
258	zio_cksum_t		sio_cksum;
259	uint32_t		sio_nr_dvas;
260
261	/* fields from zio_t */
262	uint32_t		sio_flags;
263	zbookmark_phys_t	sio_zb;
264
265	/* members for queue sorting */
266	union {
267		avl_node_t	sio_addr_node; /* link into issuing queue */
268		list_node_t	sio_list_node; /* link for issuing to disk */
269	} sio_nodes;
270
271	/*
272	 * There may be up to SPA_DVAS_PER_BP DVAs here from the bp,
273	 * depending on how many were in the original bp. Only the
274	 * first DVA is really used for sorting and issuing purposes.
275	 * The other DVAs (if provided) simply exist so that the zio
276	 * layer can find additional copies to repair from in the
277	 * event of an error. This array must go at the end of the
278	 * struct to allow this for the variable number of elements.
279	 */
280	dva_t			sio_dva[0];
281} scan_io_t;
282
283#define	SIO_SET_OFFSET(sio, x)		DVA_SET_OFFSET(&(sio)->sio_dva[0], x)
284#define	SIO_SET_ASIZE(sio, x)		DVA_SET_ASIZE(&(sio)->sio_dva[0], x)
285#define	SIO_GET_OFFSET(sio)		DVA_GET_OFFSET(&(sio)->sio_dva[0])
286#define	SIO_GET_ASIZE(sio)		DVA_GET_ASIZE(&(sio)->sio_dva[0])
287#define	SIO_GET_END_OFFSET(sio)		\
288	(SIO_GET_OFFSET(sio) + SIO_GET_ASIZE(sio))
289#define	SIO_GET_MUSED(sio)		\
290	(sizeof (scan_io_t) + ((sio)->sio_nr_dvas * sizeof (dva_t)))
291
292struct dsl_scan_io_queue {
293	dsl_scan_t	*q_scn; /* associated dsl_scan_t */
294	vdev_t		*q_vd; /* top-level vdev that this queue represents */
295
296	/* trees used for sorting I/Os and extents of I/Os */
297	range_tree_t	*q_exts_by_addr;
298	zfs_btree_t		q_exts_by_size;
299	avl_tree_t	q_sios_by_addr;
300	uint64_t	q_sio_memused;
301
302	/* members for zio rate limiting */
303	uint64_t	q_maxinflight_bytes;
304	uint64_t	q_inflight_bytes;
305	kcondvar_t	q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
306
307	/* per txg statistics */
308	uint64_t	q_total_seg_size_this_txg;
309	uint64_t	q_segs_this_txg;
310	uint64_t	q_total_zio_size_this_txg;
311	uint64_t	q_zios_this_txg;
312};
313
314/* private data for dsl_scan_prefetch_cb() */
315typedef struct scan_prefetch_ctx {
316	zfs_refcount_t spc_refcnt;	/* refcount for memory management */
317	dsl_scan_t *spc_scn;		/* dsl_scan_t for the pool */
318	boolean_t spc_root;		/* is this prefetch for an objset? */
319	uint8_t spc_indblkshift;	/* dn_indblkshift of current dnode */
320	uint16_t spc_datablkszsec;	/* dn_idatablkszsec of current dnode */
321} scan_prefetch_ctx_t;
322
323/* private data for dsl_scan_prefetch() */
324typedef struct scan_prefetch_issue_ctx {
325	avl_node_t spic_avl_node;	/* link into scn->scn_prefetch_queue */
326	scan_prefetch_ctx_t *spic_spc;	/* spc for the callback */
327	blkptr_t spic_bp;		/* bp to prefetch */
328	zbookmark_phys_t spic_zb;	/* bookmark to prefetch */
329} scan_prefetch_issue_ctx_t;
330
331static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
332    const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
333static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
334    scan_io_t *sio);
335
336static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
337static void scan_io_queues_destroy(dsl_scan_t *scn);
338
339static kmem_cache_t *sio_cache[SPA_DVAS_PER_BP];
340
341/* sio->sio_nr_dvas must be set so we know which cache to free from */
342static void
343sio_free(scan_io_t *sio)
344{
345	ASSERT3U(sio->sio_nr_dvas, >, 0);
346	ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
347
348	kmem_cache_free(sio_cache[sio->sio_nr_dvas - 1], sio);
349}
350
351/* It is up to the caller to set sio->sio_nr_dvas for freeing */
352static scan_io_t *
353sio_alloc(unsigned short nr_dvas)
354{
355	ASSERT3U(nr_dvas, >, 0);
356	ASSERT3U(nr_dvas, <=, SPA_DVAS_PER_BP);
357
358	return (kmem_cache_alloc(sio_cache[nr_dvas - 1], KM_SLEEP));
359}
360
361void
362scan_init(void)
363{
364	/*
365	 * This is used in ext_size_compare() to weight segments
366	 * based on how sparse they are. This cannot be changed
367	 * mid-scan and the tree comparison functions don't currently
368	 * have a mechansim for passing additional context to the
369	 * compare functions. Thus we store this value globally and
370	 * we only allow it to be set at module intiailization time
371	 */
372	fill_weight = zfs_scan_fill_weight;
373
374	for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
375		char name[36];
376
377		(void) sprintf(name, "sio_cache_%d", i);
378		sio_cache[i] = kmem_cache_create(name,
379		    (sizeof (scan_io_t) + ((i + 1) * sizeof (dva_t))),
380		    0, NULL, NULL, NULL, NULL, NULL, 0);
381	}
382}
383
384void
385scan_fini(void)
386{
387	for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
388		kmem_cache_destroy(sio_cache[i]);
389	}
390}
391
392static inline boolean_t
393dsl_scan_is_running(const dsl_scan_t *scn)
394{
395	return (scn->scn_phys.scn_state == DSS_SCANNING);
396}
397
398boolean_t
399dsl_scan_resilvering(dsl_pool_t *dp)
400{
401	return (dsl_scan_is_running(dp->dp_scan) &&
402	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
403}
404
405static inline void
406sio2bp(const scan_io_t *sio, blkptr_t *bp)
407{
408	bzero(bp, sizeof (*bp));
409	bp->blk_prop = sio->sio_blk_prop;
410	bp->blk_phys_birth = sio->sio_phys_birth;
411	bp->blk_birth = sio->sio_birth;
412	bp->blk_fill = 1;	/* we always only work with data pointers */
413	bp->blk_cksum = sio->sio_cksum;
414
415	ASSERT3U(sio->sio_nr_dvas, >, 0);
416	ASSERT3U(sio->sio_nr_dvas, <=, SPA_DVAS_PER_BP);
417
418	bcopy(sio->sio_dva, bp->blk_dva, sio->sio_nr_dvas * sizeof (dva_t));
419}
420
421static inline void
422bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
423{
424	sio->sio_blk_prop = bp->blk_prop;
425	sio->sio_phys_birth = bp->blk_phys_birth;
426	sio->sio_birth = bp->blk_birth;
427	sio->sio_cksum = bp->blk_cksum;
428	sio->sio_nr_dvas = BP_GET_NDVAS(bp);
429
430	/*
431	 * Copy the DVAs to the sio. We need all copies of the block so
432	 * that the self healing code can use the alternate copies if the
433	 * first is corrupted. We want the DVA at index dva_i to be first
434	 * in the sio since this is the primary one that we want to issue.
435	 */
436	for (int i = 0, j = dva_i; i < sio->sio_nr_dvas; i++, j++) {
437		sio->sio_dva[i] = bp->blk_dva[j % sio->sio_nr_dvas];
438	}
439}
440
441int
442dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
443{
444	int err;
445	dsl_scan_t *scn;
446	spa_t *spa = dp->dp_spa;
447	uint64_t f;
448
449	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
450	scn->scn_dp = dp;
451
452	/*
453	 * It's possible that we're resuming a scan after a reboot so
454	 * make sure that the scan_async_destroying flag is initialized
455	 * appropriately.
456	 */
457	ASSERT(!scn->scn_async_destroying);
458	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
459	    SPA_FEATURE_ASYNC_DESTROY);
460
461	avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
462	    offsetof(scan_ds_t, sds_node));
463	avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
464	    sizeof (scan_prefetch_issue_ctx_t),
465	    offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
466
467	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
468	    "scrub_func", sizeof (uint64_t), 1, &f);
469	if (err == 0) {
470		/*
471		 * There was an old-style scrub in progress.  Restart a
472		 * new-style scrub from the beginning.
473		 */
474		scn->scn_restart_txg = txg;
475		zfs_dbgmsg("old-style scrub was in progress; "
476		    "restarting new-style scrub in txg %llu",
477		    (longlong_t)scn->scn_restart_txg);
478
479		/*
480		 * Load the queue obj from the old location so that it
481		 * can be freed by dsl_scan_done().
482		 */
483		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
484		    "scrub_queue", sizeof (uint64_t), 1,
485		    &scn->scn_phys.scn_queue_obj);
486	} else {
487		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
488		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
489		    &scn->scn_phys);
490
491		/*
492		 * Detect if the pool contains the signature of #2094.  If it
493		 * does properly update the scn->scn_phys structure and notify
494		 * the administrator by setting an errata for the pool.
495		 */
496		if (err == EOVERFLOW) {
497			uint64_t zaptmp[SCAN_PHYS_NUMINTS + 1];
498			VERIFY3S(SCAN_PHYS_NUMINTS, ==, 24);
499			VERIFY3S(offsetof(dsl_scan_phys_t, scn_flags), ==,
500			    (23 * sizeof (uint64_t)));
501
502			err = zap_lookup(dp->dp_meta_objset,
503			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SCAN,
504			    sizeof (uint64_t), SCAN_PHYS_NUMINTS + 1, &zaptmp);
505			if (err == 0) {
506				uint64_t overflow = zaptmp[SCAN_PHYS_NUMINTS];
507
508				if (overflow & ~DSF_VISIT_DS_AGAIN ||
509				    scn->scn_async_destroying) {
510					spa->spa_errata =
511					    ZPOOL_ERRATA_ZOL_2094_ASYNC_DESTROY;
512					return (EOVERFLOW);
513				}
514
515				bcopy(zaptmp, &scn->scn_phys,
516				    SCAN_PHYS_NUMINTS * sizeof (uint64_t));
517				scn->scn_phys.scn_flags = overflow;
518
519				/* Required scrub already in progress. */
520				if (scn->scn_phys.scn_state == DSS_FINISHED ||
521				    scn->scn_phys.scn_state == DSS_CANCELED)
522					spa->spa_errata =
523					    ZPOOL_ERRATA_ZOL_2094_SCRUB;
524			}
525		}
526
527		if (err == ENOENT)
528			return (0);
529		else if (err)
530			return (err);
531
532		/*
533		 * We might be restarting after a reboot, so jump the issued
534		 * counter to how far we've scanned. We know we're consistent
535		 * up to here.
536		 */
537		scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
538
539		if (dsl_scan_is_running(scn) &&
540		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
541			/*
542			 * A new-type scrub was in progress on an old
543			 * pool, and the pool was accessed by old
544			 * software.  Restart from the beginning, since
545			 * the old software may have changed the pool in
546			 * the meantime.
547			 */
548			scn->scn_restart_txg = txg;
549			zfs_dbgmsg("new-style scrub was modified "
550			    "by old software; restarting in txg %llu",
551			    (longlong_t)scn->scn_restart_txg);
552		}
553	}
554
555	bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
556
557	/* reload the queue into the in-core state */
558	if (scn->scn_phys.scn_queue_obj != 0) {
559		zap_cursor_t zc;
560		zap_attribute_t za;
561
562		for (zap_cursor_init(&zc, dp->dp_meta_objset,
563		    scn->scn_phys.scn_queue_obj);
564		    zap_cursor_retrieve(&zc, &za) == 0;
565		    (void) zap_cursor_advance(&zc)) {
566			scan_ds_queue_insert(scn,
567			    zfs_strtonum(za.za_name, NULL),
568			    za.za_first_integer);
569		}
570		zap_cursor_fini(&zc);
571	}
572
573	spa_scan_stat_init(spa);
574	return (0);
575}
576
577void
578dsl_scan_fini(dsl_pool_t *dp)
579{
580	if (dp->dp_scan != NULL) {
581		dsl_scan_t *scn = dp->dp_scan;
582
583		if (scn->scn_taskq != NULL)
584			taskq_destroy(scn->scn_taskq);
585		scan_ds_queue_clear(scn);
586		avl_destroy(&scn->scn_queue);
587		avl_destroy(&scn->scn_prefetch_queue);
588
589		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
590		dp->dp_scan = NULL;
591	}
592}
593
594static boolean_t
595dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
596{
597	return (scn->scn_restart_txg != 0 &&
598	    scn->scn_restart_txg <= tx->tx_txg);
599}
600
601boolean_t
602dsl_scan_scrubbing(const dsl_pool_t *dp)
603{
604	dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
605
606	return (scn_phys->scn_state == DSS_SCANNING &&
607	    scn_phys->scn_func == POOL_SCAN_SCRUB);
608}
609
610boolean_t
611dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
612{
613	return (dsl_scan_scrubbing(scn->scn_dp) &&
614	    scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
615}
616
617/*
618 * Writes out a persistent dsl_scan_phys_t record to the pool directory.
619 * Because we can be running in the block sorting algorithm, we do not always
620 * want to write out the record, only when it is "safe" to do so. This safety
621 * condition is achieved by making sure that the sorting queues are empty
622 * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
623 * is inconsistent with how much actual scanning progress has been made. The
624 * kind of sync to be performed is specified by the sync_type argument. If the
625 * sync is optional, we only sync if the queues are empty. If the sync is
626 * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
627 * third possible state is a "cached" sync. This is done in response to:
628 * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
629 *	destroyed, so we wouldn't be able to restart scanning from it.
630 * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
631 *	superseded by a newer snapshot.
632 * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
633 *	swapped with its clone.
634 * In all cases, a cached sync simply rewrites the last record we've written,
635 * just slightly modified. For the modifications that are performed to the
636 * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
637 * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
638 */
639static void
640dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
641{
642	int i;
643	spa_t *spa = scn->scn_dp->dp_spa;
644
645	ASSERT(sync_type != SYNC_MANDATORY || scn->scn_bytes_pending == 0);
646	if (scn->scn_bytes_pending == 0) {
647		for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
648			vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
649			dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
650
651			if (q == NULL)
652				continue;
653
654			mutex_enter(&vd->vdev_scan_io_queue_lock);
655			ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
656			ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
657			    NULL);
658			ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
659			mutex_exit(&vd->vdev_scan_io_queue_lock);
660		}
661
662		if (scn->scn_phys.scn_queue_obj != 0)
663			scan_ds_queue_sync(scn, tx);
664		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
665		    DMU_POOL_DIRECTORY_OBJECT,
666		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
667		    &scn->scn_phys, tx));
668		bcopy(&scn->scn_phys, &scn->scn_phys_cached,
669		    sizeof (scn->scn_phys));
670
671		if (scn->scn_checkpointing)
672			zfs_dbgmsg("finish scan checkpoint");
673
674		scn->scn_checkpointing = B_FALSE;
675		scn->scn_last_checkpoint = ddi_get_lbolt();
676	} else if (sync_type == SYNC_CACHED) {
677		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
678		    DMU_POOL_DIRECTORY_OBJECT,
679		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
680		    &scn->scn_phys_cached, tx));
681	}
682}
683
684/* ARGSUSED */
685static int
686dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
687{
688	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
689
690	if (dsl_scan_is_running(scn))
691		return (SET_ERROR(EBUSY));
692
693	return (0);
694}
695
696static void
697dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
698{
699	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
700	pool_scan_func_t *funcp = arg;
701	dmu_object_type_t ot = 0;
702	dsl_pool_t *dp = scn->scn_dp;
703	spa_t *spa = dp->dp_spa;
704
705	ASSERT(!dsl_scan_is_running(scn));
706	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
707	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
708	scn->scn_phys.scn_func = *funcp;
709	scn->scn_phys.scn_state = DSS_SCANNING;
710	scn->scn_phys.scn_min_txg = 0;
711	scn->scn_phys.scn_max_txg = tx->tx_txg;
712	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
713	scn->scn_phys.scn_start_time = gethrestime_sec();
714	scn->scn_phys.scn_errors = 0;
715	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
716	scn->scn_issued_before_pass = 0;
717	scn->scn_restart_txg = 0;
718	scn->scn_done_txg = 0;
719	scn->scn_last_checkpoint = 0;
720	scn->scn_checkpointing = B_FALSE;
721	spa_scan_stat_init(spa);
722
723	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
724		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
725
726		/* rewrite all disk labels */
727		vdev_config_dirty(spa->spa_root_vdev);
728
729		if (vdev_resilver_needed(spa->spa_root_vdev,
730		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
731			spa_event_notify(spa, NULL, NULL,
732			    ESC_ZFS_RESILVER_START);
733		} else {
734			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
735		}
736
737		spa->spa_scrub_started = B_TRUE;
738		/*
739		 * If this is an incremental scrub, limit the DDT scrub phase
740		 * to just the auto-ditto class (for correctness); the rest
741		 * of the scrub should go faster using top-down pruning.
742		 */
743		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
744			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
745
746	}
747
748	/* back to the generic stuff */
749
750	if (dp->dp_blkstats == NULL) {
751		dp->dp_blkstats =
752		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
753		mutex_init(&dp->dp_blkstats->zab_lock, NULL,
754		    MUTEX_DEFAULT, NULL);
755	}
756	bzero(&dp->dp_blkstats->zab_type, sizeof (dp->dp_blkstats->zab_type));
757
758	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
759		ot = DMU_OT_ZAP_OTHER;
760
761	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
762	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
763
764	bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
765
766	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
767
768	spa_history_log_internal(spa, "scan setup", tx,
769	    "func=%u mintxg=%llu maxtxg=%llu",
770	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
771}
772
773/*
774 * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
775 * Can also be called to resume a paused scrub.
776 */
777int
778dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
779{
780	spa_t *spa = dp->dp_spa;
781	dsl_scan_t *scn = dp->dp_scan;
782
783	/*
784	 * Purge all vdev caches and probe all devices.  We do this here
785	 * rather than in sync context because this requires a writer lock
786	 * on the spa_config lock, which we can't do from sync context.  The
787	 * spa_scrub_reopen flag indicates that vdev_open() should not
788	 * attempt to start another scrub.
789	 */
790	spa_vdev_state_enter(spa, SCL_NONE);
791	spa->spa_scrub_reopen = B_TRUE;
792	vdev_reopen(spa->spa_root_vdev);
793	spa->spa_scrub_reopen = B_FALSE;
794	(void) spa_vdev_state_exit(spa, NULL, 0);
795
796	if (func == POOL_SCAN_RESILVER) {
797		dsl_resilver_restart(spa->spa_dsl_pool, 0);
798		return (0);
799	}
800
801	if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
802		/* got scrub start cmd, resume paused scrub */
803		int err = dsl_scrub_set_pause_resume(scn->scn_dp,
804		    POOL_SCRUB_NORMAL);
805		if (err == 0) {
806			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
807			return (ECANCELED);
808		}
809		return (SET_ERROR(err));
810	}
811
812	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
813	    dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
814}
815
816/*
817 * Sets the resilver defer flag to B_FALSE on all leaf devs under vd. Returns
818 * B_TRUE if we have devices that need to be resilvered and are available to
819 * accept resilver I/Os.
820 */
821static boolean_t
822dsl_scan_clear_deferred(vdev_t *vd, dmu_tx_t *tx)
823{
824	boolean_t resilver_needed = B_FALSE;
825	spa_t *spa = vd->vdev_spa;
826
827	for (int c = 0; c < vd->vdev_children; c++) {
828		resilver_needed |=
829		    dsl_scan_clear_deferred(vd->vdev_child[c], tx);
830	}
831
832	if (vd == spa->spa_root_vdev &&
833	    spa_feature_is_active(spa, SPA_FEATURE_RESILVER_DEFER)) {
834		spa_feature_decr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
835		vdev_config_dirty(vd);
836		spa->spa_resilver_deferred = B_FALSE;
837		return (resilver_needed);
838	}
839
840	if (!vdev_is_concrete(vd) || vd->vdev_aux ||
841	    !vd->vdev_ops->vdev_op_leaf)
842		return (resilver_needed);
843
844	if (vd->vdev_resilver_deferred)
845		vd->vdev_resilver_deferred = B_FALSE;
846
847	return (!vdev_is_dead(vd) && !vd->vdev_offline &&
848	    vdev_resilver_needed(vd, NULL, NULL));
849}
850
851/* ARGSUSED */
852static void
853dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
854{
855	static const char *old_names[] = {
856		"scrub_bookmark",
857		"scrub_ddt_bookmark",
858		"scrub_ddt_class_max",
859		"scrub_queue",
860		"scrub_min_txg",
861		"scrub_max_txg",
862		"scrub_func",
863		"scrub_errors",
864		NULL
865	};
866
867	dsl_pool_t *dp = scn->scn_dp;
868	spa_t *spa = dp->dp_spa;
869	int i;
870
871	/* Remove any remnants of an old-style scrub. */
872	for (i = 0; old_names[i]; i++) {
873		(void) zap_remove(dp->dp_meta_objset,
874		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
875	}
876
877	if (scn->scn_phys.scn_queue_obj != 0) {
878		VERIFY0(dmu_object_free(dp->dp_meta_objset,
879		    scn->scn_phys.scn_queue_obj, tx));
880		scn->scn_phys.scn_queue_obj = 0;
881	}
882	scan_ds_queue_clear(scn);
883
884	scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
885
886	/*
887	 * If we were "restarted" from a stopped state, don't bother
888	 * with anything else.
889	 */
890	if (!dsl_scan_is_running(scn)) {
891		ASSERT(!scn->scn_is_sorted);
892		return;
893	}
894
895	if (scn->scn_is_sorted) {
896		scan_io_queues_destroy(scn);
897		scn->scn_is_sorted = B_FALSE;
898
899		if (scn->scn_taskq != NULL) {
900			taskq_destroy(scn->scn_taskq);
901			scn->scn_taskq = NULL;
902		}
903	}
904
905	scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
906
907	if (dsl_scan_restarting(scn, tx))
908		spa_history_log_internal(spa, "scan aborted, restarting", tx,
909		    "errors=%llu", spa_get_errlog_size(spa));
910	else if (!complete)
911		spa_history_log_internal(spa, "scan cancelled", tx,
912		    "errors=%llu", spa_get_errlog_size(spa));
913	else
914		spa_history_log_internal(spa, "scan done", tx,
915		    "errors=%llu", spa_get_errlog_size(spa));
916
917	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
918		spa->spa_scrub_started = B_FALSE;
919		spa->spa_scrub_active = B_FALSE;
920
921		/*
922		 * If the scrub/resilver completed, update all DTLs to
923		 * reflect this.  Whether it succeeded or not, vacate
924		 * all temporary scrub DTLs.
925		 *
926		 * As the scrub does not currently support traversing
927		 * data that have been freed but are part of a checkpoint,
928		 * we don't mark the scrub as done in the DTLs as faults
929		 * may still exist in those vdevs.
930		 */
931		if (complete &&
932		    !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
933			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
934			    scn->scn_phys.scn_max_txg, B_TRUE);
935
936			spa_event_notify(spa, NULL, NULL,
937			    scn->scn_phys.scn_min_txg ?
938			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
939		} else {
940			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
941			    0, B_TRUE);
942		}
943		spa_errlog_rotate(spa);
944
945		/*
946		 * We may have finished replacing a device.
947		 * Let the async thread assess this and handle the detach.
948		 */
949		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
950
951		/*
952		 * Clear any deferred_resilver flags in the config.
953		 * If there are drives that need resilvering, kick
954		 * off an asynchronous request to start resilver.
955		 * dsl_scan_clear_deferred() may update the config
956		 * before the resilver can restart. In the event of
957		 * a crash during this period, the spa loading code
958		 * will find the drives that need to be resilvered
959		 * when the machine reboots and start the resilver then.
960		 */
961		if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) {
962			boolean_t resilver_needed =
963			    dsl_scan_clear_deferred(spa->spa_root_vdev, tx);
964			if (resilver_needed) {
965				spa_history_log_internal(spa,
966				    "starting deferred resilver", tx,
967				    "errors=%llu", spa_get_errlog_size(spa));
968				spa_async_request(spa, SPA_ASYNC_RESILVER);
969			}
970		}
971	}
972
973	scn->scn_phys.scn_end_time = gethrestime_sec();
974
975	ASSERT(!dsl_scan_is_running(scn));
976}
977
978/* ARGSUSED */
979static int
980dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
981{
982	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
983
984	if (!dsl_scan_is_running(scn))
985		return (SET_ERROR(ENOENT));
986	return (0);
987}
988
989/* ARGSUSED */
990static void
991dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
992{
993	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
994
995	dsl_scan_done(scn, B_FALSE, tx);
996	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
997	spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
998}
999
1000int
1001dsl_scan_cancel(dsl_pool_t *dp)
1002{
1003	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
1004	    dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
1005}
1006
1007static int
1008dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
1009{
1010	pool_scrub_cmd_t *cmd = arg;
1011	dsl_pool_t *dp = dmu_tx_pool(tx);
1012	dsl_scan_t *scn = dp->dp_scan;
1013
1014	if (*cmd == POOL_SCRUB_PAUSE) {
1015		/* can't pause a scrub when there is no in-progress scrub */
1016		if (!dsl_scan_scrubbing(dp))
1017			return (SET_ERROR(ENOENT));
1018
1019		/* can't pause a paused scrub */
1020		if (dsl_scan_is_paused_scrub(scn))
1021			return (SET_ERROR(EBUSY));
1022	} else if (*cmd != POOL_SCRUB_NORMAL) {
1023		return (SET_ERROR(ENOTSUP));
1024	}
1025
1026	return (0);
1027}
1028
1029static void
1030dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1031{
1032	pool_scrub_cmd_t *cmd = arg;
1033	dsl_pool_t *dp = dmu_tx_pool(tx);
1034	spa_t *spa = dp->dp_spa;
1035	dsl_scan_t *scn = dp->dp_scan;
1036
1037	if (*cmd == POOL_SCRUB_PAUSE) {
1038		/* can't pause a scrub when there is no in-progress scrub */
1039		spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1040		scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1041		scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1042		dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1043		spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1044	} else {
1045		ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1046		if (dsl_scan_is_paused_scrub(scn)) {
1047			/*
1048			 * We need to keep track of how much time we spend
1049			 * paused per pass so that we can adjust the scrub rate
1050			 * shown in the output of 'zpool status'
1051			 */
1052			spa->spa_scan_pass_scrub_spent_paused +=
1053			    gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1054			spa->spa_scan_pass_scrub_pause = 0;
1055			scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1056			scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1057			dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1058		}
1059	}
1060}
1061
1062/*
1063 * Set scrub pause/resume state if it makes sense to do so
1064 */
1065int
1066dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1067{
1068	return (dsl_sync_task(spa_name(dp->dp_spa),
1069	    dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1070	    ZFS_SPACE_CHECK_RESERVED));
1071}
1072
1073
1074/* start a new scan, or restart an existing one. */
1075void
1076dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1077{
1078	if (txg == 0) {
1079		dmu_tx_t *tx;
1080		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1081		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1082
1083		txg = dmu_tx_get_txg(tx);
1084		dp->dp_scan->scn_restart_txg = txg;
1085		dmu_tx_commit(tx);
1086	} else {
1087		dp->dp_scan->scn_restart_txg = txg;
1088	}
1089	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1090}
1091
1092void
1093dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1094{
1095	zio_free(dp->dp_spa, txg, bp);
1096}
1097
1098void
1099dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1100{
1101	ASSERT(dsl_pool_sync_context(dp));
1102	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1103}
1104
1105static int
1106scan_ds_queue_compare(const void *a, const void *b)
1107{
1108	const scan_ds_t *sds_a = a, *sds_b = b;
1109
1110	if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1111		return (-1);
1112	if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1113		return (0);
1114	return (1);
1115}
1116
1117static void
1118scan_ds_queue_clear(dsl_scan_t *scn)
1119{
1120	void *cookie = NULL;
1121	scan_ds_t *sds;
1122	while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1123		kmem_free(sds, sizeof (*sds));
1124	}
1125}
1126
1127static boolean_t
1128scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1129{
1130	scan_ds_t srch, *sds;
1131
1132	srch.sds_dsobj = dsobj;
1133	sds = avl_find(&scn->scn_queue, &srch, NULL);
1134	if (sds != NULL && txg != NULL)
1135		*txg = sds->sds_txg;
1136	return (sds != NULL);
1137}
1138
1139static void
1140scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1141{
1142	scan_ds_t *sds;
1143	avl_index_t where;
1144
1145	sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1146	sds->sds_dsobj = dsobj;
1147	sds->sds_txg = txg;
1148
1149	VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1150	avl_insert(&scn->scn_queue, sds, where);
1151}
1152
1153static void
1154scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1155{
1156	scan_ds_t srch, *sds;
1157
1158	srch.sds_dsobj = dsobj;
1159
1160	sds = avl_find(&scn->scn_queue, &srch, NULL);
1161	VERIFY(sds != NULL);
1162	avl_remove(&scn->scn_queue, sds);
1163	kmem_free(sds, sizeof (*sds));
1164}
1165
1166static void
1167scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1168{
1169	dsl_pool_t *dp = scn->scn_dp;
1170	spa_t *spa = dp->dp_spa;
1171	dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1172	    DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1173
1174	ASSERT0(scn->scn_bytes_pending);
1175	ASSERT(scn->scn_phys.scn_queue_obj != 0);
1176
1177	VERIFY0(dmu_object_free(dp->dp_meta_objset,
1178	    scn->scn_phys.scn_queue_obj, tx));
1179	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1180	    DMU_OT_NONE, 0, tx);
1181	for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1182	    sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1183		VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1184		    scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1185		    sds->sds_txg, tx));
1186	}
1187}
1188
1189/*
1190 * Computes the memory limit state that we're currently in. A sorted scan
1191 * needs quite a bit of memory to hold the sorting queue, so we need to
1192 * reasonably constrain the size so it doesn't impact overall system
1193 * performance. We compute two limits:
1194 * 1) Hard memory limit: if the amount of memory used by the sorting
1195 *	queues on a pool gets above this value, we stop the metadata
1196 *	scanning portion and start issuing the queued up and sorted
1197 *	I/Os to reduce memory usage.
1198 *	This limit is calculated as a fraction of physmem (by default 5%).
1199 *	We constrain the lower bound of the hard limit to an absolute
1200 *	minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1201 *	the upper bound to 5% of the total pool size - no chance we'll
1202 *	ever need that much memory, but just to keep the value in check.
1203 * 2) Soft memory limit: once we hit the hard memory limit, we start
1204 *	issuing I/O to reduce queue memory usage, but we don't want to
1205 *	completely empty out the queues, since we might be able to find I/Os
1206 *	that will fill in the gaps of our non-sequential IOs at some point
1207 *	in the future. So we stop the issuing of I/Os once the amount of
1208 *	memory used drops below the soft limit (at which point we stop issuing
1209 *	I/O and start scanning metadata again).
1210 *
1211 *	This limit is calculated by subtracting a fraction of the hard
1212 *	limit from the hard limit. By default this fraction is 5%, so
1213 *	the soft limit is 95% of the hard limit. We cap the size of the
1214 *	difference between the hard and soft limits at an absolute
1215 *	maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1216 *	sufficient to not cause too frequent switching between the
1217 *	metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1218 *	worth of queues is about 1.2 GiB of on-pool data, so scanning
1219 *	that should take at least a decent fraction of a second).
1220 */
1221static boolean_t
1222dsl_scan_should_clear(dsl_scan_t *scn)
1223{
1224	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1225	uint64_t mlim_hard, mlim_soft, mused;
1226	uint64_t alloc = metaslab_class_get_alloc(spa_normal_class(
1227	    scn->scn_dp->dp_spa));
1228
1229	mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1230	    zfs_scan_mem_lim_min);
1231	mlim_hard = MIN(mlim_hard, alloc / 20);
1232	mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1233	    zfs_scan_mem_lim_soft_max);
1234	mused = 0;
1235	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1236		vdev_t *tvd = rvd->vdev_child[i];
1237		dsl_scan_io_queue_t *queue;
1238
1239		mutex_enter(&tvd->vdev_scan_io_queue_lock);
1240		queue = tvd->vdev_scan_io_queue;
1241		if (queue != NULL) {
1242			/* # extents in exts_by_size = # in exts_by_addr */
1243			mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1244			    sizeof (range_seg_t) + queue->q_sio_memused;
1245		}
1246		mutex_exit(&tvd->vdev_scan_io_queue_lock);
1247	}
1248
1249	dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1250
1251	if (mused == 0)
1252		ASSERT0(scn->scn_bytes_pending);
1253
1254	/*
1255	 * If we are above our hard limit, we need to clear out memory.
1256	 * If we are below our soft limit, we need to accumulate sequential IOs.
1257	 * Otherwise, we should keep doing whatever we are currently doing.
1258	 */
1259	if (mused >= mlim_hard)
1260		return (B_TRUE);
1261	else if (mused < mlim_soft)
1262		return (B_FALSE);
1263	else
1264		return (scn->scn_clearing);
1265}
1266
1267static boolean_t
1268dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1269{
1270	/* we never skip user/group accounting objects */
1271	if (zb && (int64_t)zb->zb_object < 0)
1272		return (B_FALSE);
1273
1274	if (scn->scn_suspending)
1275		return (B_TRUE); /* we're already suspending */
1276
1277	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1278		return (B_FALSE); /* we're resuming */
1279
1280	/* We only know how to resume from level-0 blocks. */
1281	if (zb && zb->zb_level != 0)
1282		return (B_FALSE);
1283
1284	/*
1285	 * We suspend if:
1286	 *  - we have scanned for at least the minimum time (default 1 sec
1287	 *    for scrub, 3 sec for resilver), and either we have sufficient
1288	 *    dirty data that we are starting to write more quickly
1289	 *    (default 30%), or someone is explicitly waiting for this txg
1290	 *    to complete.
1291	 *  or
1292	 *  - the spa is shutting down because this pool is being exported
1293	 *    or the machine is rebooting.
1294	 *  or
1295	 *  - the scan queue has reached its memory use limit
1296	 */
1297	hrtime_t curr_time_ns = gethrtime();
1298	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1299	uint64_t sync_time_ns = curr_time_ns -
1300	    scn->scn_dp->dp_spa->spa_sync_starttime;
1301
1302	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
1303	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1304	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1305
1306	if ((NSEC2MSEC(scan_time_ns) > mintime &&
1307	    (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
1308	    txg_sync_waiting(scn->scn_dp) ||
1309	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1310	    spa_shutting_down(scn->scn_dp->dp_spa) ||
1311	    (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1312		if (zb) {
1313			dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1314			    (longlong_t)zb->zb_objset,
1315			    (longlong_t)zb->zb_object,
1316			    (longlong_t)zb->zb_level,
1317			    (longlong_t)zb->zb_blkid);
1318			scn->scn_phys.scn_bookmark = *zb;
1319		} else {
1320			dsl_scan_phys_t *scnp = &scn->scn_phys;
1321
1322			dprintf("suspending at DDT bookmark "
1323			    "%llx/%llx/%llx/%llx\n",
1324			    (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1325			    (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1326			    (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1327			    (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1328		}
1329		scn->scn_suspending = B_TRUE;
1330		return (B_TRUE);
1331	}
1332	return (B_FALSE);
1333}
1334
1335typedef struct zil_scan_arg {
1336	dsl_pool_t	*zsa_dp;
1337	zil_header_t	*zsa_zh;
1338} zil_scan_arg_t;
1339
1340/* ARGSUSED */
1341static int
1342dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1343{
1344	zil_scan_arg_t *zsa = arg;
1345	dsl_pool_t *dp = zsa->zsa_dp;
1346	dsl_scan_t *scn = dp->dp_scan;
1347	zil_header_t *zh = zsa->zsa_zh;
1348	zbookmark_phys_t zb;
1349
1350	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1351		return (0);
1352
1353	/*
1354	 * One block ("stubby") can be allocated a long time ago; we
1355	 * want to visit that one because it has been allocated
1356	 * (on-disk) even if it hasn't been claimed (even though for
1357	 * scrub there's nothing to do to it).
1358	 */
1359	if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1360		return (0);
1361
1362	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1363	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1364
1365	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1366	return (0);
1367}
1368
1369/* ARGSUSED */
1370static int
1371dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
1372{
1373	if (lrc->lrc_txtype == TX_WRITE) {
1374		zil_scan_arg_t *zsa = arg;
1375		dsl_pool_t *dp = zsa->zsa_dp;
1376		dsl_scan_t *scn = dp->dp_scan;
1377		zil_header_t *zh = zsa->zsa_zh;
1378		lr_write_t *lr = (lr_write_t *)lrc;
1379		blkptr_t *bp = &lr->lr_blkptr;
1380		zbookmark_phys_t zb;
1381
1382		if (BP_IS_HOLE(bp) ||
1383		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1384			return (0);
1385
1386		/*
1387		 * birth can be < claim_txg if this record's txg is
1388		 * already txg sync'ed (but this log block contains
1389		 * other records that are not synced)
1390		 */
1391		if (claim_txg == 0 || bp->blk_birth < claim_txg)
1392			return (0);
1393
1394		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1395		    lr->lr_foid, ZB_ZIL_LEVEL,
1396		    lr->lr_offset / BP_GET_LSIZE(bp));
1397
1398		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1399	}
1400	return (0);
1401}
1402
1403static void
1404dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1405{
1406	uint64_t claim_txg = zh->zh_claim_txg;
1407	zil_scan_arg_t zsa = { dp, zh };
1408	zilog_t *zilog;
1409
1410	ASSERT(spa_writeable(dp->dp_spa));
1411
1412	/*
1413	 * We only want to visit blocks that have been claimed
1414	 * but not yet replayed.
1415	 */
1416	if (claim_txg == 0)
1417		return;
1418
1419	zilog = zil_alloc(dp->dp_meta_objset, zh);
1420
1421	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1422	    claim_txg, B_FALSE);
1423
1424	zil_free(zilog);
1425}
1426
1427/*
1428 * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1429 * here is to sort the AVL tree by the order each block will be needed.
1430 */
1431static int
1432scan_prefetch_queue_compare(const void *a, const void *b)
1433{
1434	const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1435	const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1436	const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1437
1438	return (zbookmark_compare(spc_a->spc_datablkszsec,
1439	    spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1440	    spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1441}
1442
1443static void
1444scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, void *tag)
1445{
1446	if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1447		zfs_refcount_destroy(&spc->spc_refcnt);
1448		kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1449	}
1450}
1451
1452static scan_prefetch_ctx_t *
1453scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, void *tag)
1454{
1455	scan_prefetch_ctx_t *spc;
1456
1457	spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1458	zfs_refcount_create(&spc->spc_refcnt);
1459	zfs_refcount_add(&spc->spc_refcnt, tag);
1460	spc->spc_scn = scn;
1461	if (dnp != NULL) {
1462		spc->spc_datablkszsec = dnp->dn_datablkszsec;
1463		spc->spc_indblkshift = dnp->dn_indblkshift;
1464		spc->spc_root = B_FALSE;
1465	} else {
1466		spc->spc_datablkszsec = 0;
1467		spc->spc_indblkshift = 0;
1468		spc->spc_root = B_TRUE;
1469	}
1470
1471	return (spc);
1472}
1473
1474static void
1475scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, void *tag)
1476{
1477	zfs_refcount_add(&spc->spc_refcnt, tag);
1478}
1479
1480static boolean_t
1481dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1482    const zbookmark_phys_t *zb)
1483{
1484	zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1485	dnode_phys_t tmp_dnp;
1486	dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1487
1488	if (zb->zb_objset != last_zb->zb_objset)
1489		return (B_TRUE);
1490	if ((int64_t)zb->zb_object < 0)
1491		return (B_FALSE);
1492
1493	tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1494	tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1495
1496	if (zbookmark_subtree_completed(dnp, zb, last_zb))
1497		return (B_TRUE);
1498
1499	return (B_FALSE);
1500}
1501
1502static void
1503dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1504{
1505	avl_index_t idx;
1506	dsl_scan_t *scn = spc->spc_scn;
1507	spa_t *spa = scn->scn_dp->dp_spa;
1508	scan_prefetch_issue_ctx_t *spic;
1509
1510	if (zfs_no_scrub_prefetch)
1511		return;
1512
1513	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1514	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1515	    BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1516		return;
1517
1518	if (dsl_scan_check_prefetch_resume(spc, zb))
1519		return;
1520
1521	scan_prefetch_ctx_add_ref(spc, scn);
1522	spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1523	spic->spic_spc = spc;
1524	spic->spic_bp = *bp;
1525	spic->spic_zb = *zb;
1526
1527	/*
1528	 * Add the IO to the queue of blocks to prefetch. This allows us to
1529	 * prioritize blocks that we will need first for the main traversal
1530	 * thread.
1531	 */
1532	mutex_enter(&spa->spa_scrub_lock);
1533	if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1534		/* this block is already queued for prefetch */
1535		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1536		scan_prefetch_ctx_rele(spc, scn);
1537		mutex_exit(&spa->spa_scrub_lock);
1538		return;
1539	}
1540
1541	avl_insert(&scn->scn_prefetch_queue, spic, idx);
1542	cv_broadcast(&spa->spa_scrub_io_cv);
1543	mutex_exit(&spa->spa_scrub_lock);
1544}
1545
1546static void
1547dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1548    uint64_t objset, uint64_t object)
1549{
1550	int i;
1551	zbookmark_phys_t zb;
1552	scan_prefetch_ctx_t *spc;
1553
1554	if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1555		return;
1556
1557	SET_BOOKMARK(&zb, objset, object, 0, 0);
1558
1559	spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1560
1561	for (i = 0; i < dnp->dn_nblkptr; i++) {
1562		zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1563		zb.zb_blkid = i;
1564		dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1565	}
1566
1567	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1568		zb.zb_level = 0;
1569		zb.zb_blkid = DMU_SPILL_BLKID;
1570		dsl_scan_prefetch(spc, &dnp->dn_spill, &zb);
1571	}
1572
1573	scan_prefetch_ctx_rele(spc, FTAG);
1574}
1575
1576void
1577dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1578    arc_buf_t *buf, void *private)
1579{
1580	scan_prefetch_ctx_t *spc = private;
1581	dsl_scan_t *scn = spc->spc_scn;
1582	spa_t *spa = scn->scn_dp->dp_spa;
1583
1584	/* broadcast that the IO has completed for rate limitting purposes */
1585	mutex_enter(&spa->spa_scrub_lock);
1586	ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1587	spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1588	cv_broadcast(&spa->spa_scrub_io_cv);
1589	mutex_exit(&spa->spa_scrub_lock);
1590
1591	/* if there was an error or we are done prefetching, just cleanup */
1592	if (buf == NULL || scn->scn_suspending)
1593		goto out;
1594
1595	if (BP_GET_LEVEL(bp) > 0) {
1596		int i;
1597		blkptr_t *cbp;
1598		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1599		zbookmark_phys_t czb;
1600
1601		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1602			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1603			    zb->zb_level - 1, zb->zb_blkid * epb + i);
1604			dsl_scan_prefetch(spc, cbp, &czb);
1605		}
1606	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1607		dnode_phys_t *cdnp = buf->b_data;
1608		int i;
1609		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1610
1611		for (i = 0, cdnp = buf->b_data; i < epb;
1612		    i += cdnp->dn_extra_slots + 1,
1613		    cdnp += cdnp->dn_extra_slots + 1) {
1614			dsl_scan_prefetch_dnode(scn, cdnp,
1615			    zb->zb_objset, zb->zb_blkid * epb + i);
1616		}
1617	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1618		objset_phys_t *osp = buf->b_data;
1619
1620		dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1621		    zb->zb_objset, DMU_META_DNODE_OBJECT);
1622
1623		if (OBJSET_BUF_HAS_USERUSED(buf)) {
1624			dsl_scan_prefetch_dnode(scn,
1625			    &osp->os_groupused_dnode, zb->zb_objset,
1626			    DMU_GROUPUSED_OBJECT);
1627			dsl_scan_prefetch_dnode(scn,
1628			    &osp->os_userused_dnode, zb->zb_objset,
1629			    DMU_USERUSED_OBJECT);
1630		}
1631	}
1632
1633out:
1634	if (buf != NULL)
1635		arc_buf_destroy(buf, private);
1636	scan_prefetch_ctx_rele(spc, scn);
1637}
1638
1639/* ARGSUSED */
1640static void
1641dsl_scan_prefetch_thread(void *arg)
1642{
1643	dsl_scan_t *scn = arg;
1644	spa_t *spa = scn->scn_dp->dp_spa;
1645	vdev_t *rvd = spa->spa_root_vdev;
1646	uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1647	scan_prefetch_issue_ctx_t *spic;
1648
1649	/* loop until we are told to stop */
1650	while (!scn->scn_prefetch_stop) {
1651		arc_flags_t flags = ARC_FLAG_NOWAIT |
1652		    ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1653		int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1654
1655		mutex_enter(&spa->spa_scrub_lock);
1656
1657		/*
1658		 * Wait until we have an IO to issue and are not above our
1659		 * maximum in flight limit.
1660		 */
1661		while (!scn->scn_prefetch_stop &&
1662		    (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1663		    spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1664			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1665		}
1666
1667		/* recheck if we should stop since we waited for the cv */
1668		if (scn->scn_prefetch_stop) {
1669			mutex_exit(&spa->spa_scrub_lock);
1670			break;
1671		}
1672
1673		/* remove the prefetch IO from the tree */
1674		spic = avl_first(&scn->scn_prefetch_queue);
1675		spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1676		avl_remove(&scn->scn_prefetch_queue, spic);
1677
1678		mutex_exit(&spa->spa_scrub_lock);
1679
1680		if (BP_IS_PROTECTED(&spic->spic_bp)) {
1681			ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
1682			    BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
1683			ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
1684			zio_flags |= ZIO_FLAG_RAW;
1685		}
1686
1687		/* issue the prefetch asynchronously */
1688		(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1689		    &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1690		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1691
1692		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1693	}
1694
1695	ASSERT(scn->scn_prefetch_stop);
1696
1697	/* free any prefetches we didn't get to complete */
1698	mutex_enter(&spa->spa_scrub_lock);
1699	while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1700		avl_remove(&scn->scn_prefetch_queue, spic);
1701		scan_prefetch_ctx_rele(spic->spic_spc, scn);
1702		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1703	}
1704	ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1705	mutex_exit(&spa->spa_scrub_lock);
1706}
1707
1708static boolean_t
1709dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1710    const zbookmark_phys_t *zb)
1711{
1712	/*
1713	 * We never skip over user/group accounting objects (obj<0)
1714	 */
1715	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1716	    (int64_t)zb->zb_object >= 0) {
1717		/*
1718		 * If we already visited this bp & everything below (in
1719		 * a prior txg sync), don't bother doing it again.
1720		 */
1721		if (zbookmark_subtree_completed(dnp, zb,
1722		    &scn->scn_phys.scn_bookmark))
1723			return (B_TRUE);
1724
1725		/*
1726		 * If we found the block we're trying to resume from, or
1727		 * we went past it to a different object, zero it out to
1728		 * indicate that it's OK to start checking for suspending
1729		 * again.
1730		 */
1731		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
1732		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
1733			dprintf("resuming at %llx/%llx/%llx/%llx\n",
1734			    (longlong_t)zb->zb_objset,
1735			    (longlong_t)zb->zb_object,
1736			    (longlong_t)zb->zb_level,
1737			    (longlong_t)zb->zb_blkid);
1738			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
1739		}
1740	}
1741	return (B_FALSE);
1742}
1743
1744static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1745    dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1746    dmu_objset_type_t ostype, dmu_tx_t *tx);
1747static void dsl_scan_visitdnode(
1748    dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1749    dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1750
1751/*
1752 * Return nonzero on i/o error.
1753 * Return new buf to write out in *bufp.
1754 */
1755static int
1756dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1757    dnode_phys_t *dnp, const blkptr_t *bp,
1758    const zbookmark_phys_t *zb, dmu_tx_t *tx)
1759{
1760	dsl_pool_t *dp = scn->scn_dp;
1761	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1762	int err;
1763
1764	if (BP_GET_LEVEL(bp) > 0) {
1765		arc_flags_t flags = ARC_FLAG_WAIT;
1766		int i;
1767		blkptr_t *cbp;
1768		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1769		arc_buf_t *buf;
1770
1771		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1772		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1773		if (err) {
1774			scn->scn_phys.scn_errors++;
1775			return (err);
1776		}
1777		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1778			zbookmark_phys_t czb;
1779
1780			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1781			    zb->zb_level - 1,
1782			    zb->zb_blkid * epb + i);
1783			dsl_scan_visitbp(cbp, &czb, dnp,
1784			    ds, scn, ostype, tx);
1785		}
1786		arc_buf_destroy(buf, &buf);
1787	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1788		arc_flags_t flags = ARC_FLAG_WAIT;
1789		dnode_phys_t *cdnp;
1790		int i;
1791		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1792		arc_buf_t *buf;
1793
1794		if (BP_IS_PROTECTED(bp)) {
1795			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
1796			zio_flags |= ZIO_FLAG_RAW;
1797		}
1798
1799		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1800		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1801		if (err) {
1802			scn->scn_phys.scn_errors++;
1803			return (err);
1804		}
1805		for (i = 0, cdnp = buf->b_data; i < epb;
1806		    i += cdnp->dn_extra_slots + 1,
1807		    cdnp += cdnp->dn_extra_slots + 1) {
1808			dsl_scan_visitdnode(scn, ds, ostype,
1809			    cdnp, zb->zb_blkid * epb + i, tx);
1810		}
1811
1812		arc_buf_destroy(buf, &buf);
1813	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1814		arc_flags_t flags = ARC_FLAG_WAIT;
1815		objset_phys_t *osp;
1816		arc_buf_t *buf;
1817
1818		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1819		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1820		if (err) {
1821			scn->scn_phys.scn_errors++;
1822			return (err);
1823		}
1824
1825		osp = buf->b_data;
1826
1827		dsl_scan_visitdnode(scn, ds, osp->os_type,
1828		    &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1829
1830		if (OBJSET_BUF_HAS_USERUSED(buf)) {
1831			/*
1832			 * We also always visit user/group/project accounting
1833			 * objects, and never skip them, even if we are
1834			 * suspending.  This is necessary so that the space
1835			 * deltas from this txg get integrated.
1836			 */
1837			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
1838				dsl_scan_visitdnode(scn, ds, osp->os_type,
1839				    &osp->os_projectused_dnode,
1840				    DMU_PROJECTUSED_OBJECT, tx);
1841			dsl_scan_visitdnode(scn, ds, osp->os_type,
1842			    &osp->os_groupused_dnode,
1843			    DMU_GROUPUSED_OBJECT, tx);
1844			dsl_scan_visitdnode(scn, ds, osp->os_type,
1845			    &osp->os_userused_dnode,
1846			    DMU_USERUSED_OBJECT, tx);
1847		}
1848		arc_buf_destroy(buf, &buf);
1849	}
1850
1851	return (0);
1852}
1853
1854static void
1855dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1856    dmu_objset_type_t ostype, dnode_phys_t *dnp,
1857    uint64_t object, dmu_tx_t *tx)
1858{
1859	int j;
1860
1861	for (j = 0; j < dnp->dn_nblkptr; j++) {
1862		zbookmark_phys_t czb;
1863
1864		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1865		    dnp->dn_nlevels - 1, j);
1866		dsl_scan_visitbp(&dnp->dn_blkptr[j],
1867		    &czb, dnp, ds, scn, ostype, tx);
1868	}
1869
1870	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1871		zbookmark_phys_t czb;
1872		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1873		    0, DMU_SPILL_BLKID);
1874		dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
1875		    &czb, dnp, ds, scn, ostype, tx);
1876	}
1877}
1878
1879/*
1880 * The arguments are in this order because mdb can only print the
1881 * first 5; we want them to be useful.
1882 */
1883static void
1884dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1885    dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1886    dmu_objset_type_t ostype, dmu_tx_t *tx)
1887{
1888	dsl_pool_t *dp = scn->scn_dp;
1889	blkptr_t *bp_toread = NULL;
1890
1891	if (dsl_scan_check_suspend(scn, zb))
1892		return;
1893
1894	if (dsl_scan_check_resume(scn, dnp, zb))
1895		return;
1896
1897	scn->scn_visited_this_txg++;
1898
1899	/*
1900	 * This debugging is commented out to conserve stack space.  This
1901	 * function is called recursively and the debugging addes several
1902	 * bytes to the stack for each call.  It can be commented back in
1903	 * if required to debug an issue in dsl_scan_visitbp().
1904	 *
1905	 * dprintf_bp(bp,
1906	 *	"visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1907	 *	ds, ds ? ds->ds_object : 0,
1908	 *	zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1909	 *	bp);
1910	 */
1911
1912	if (BP_IS_HOLE(bp)) {
1913		scn->scn_holes_this_txg++;
1914		return;
1915	}
1916
1917	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
1918		scn->scn_lt_min_this_txg++;
1919		return;
1920	}
1921
1922	bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
1923	*bp_toread = *bp;
1924
1925	if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
1926		goto out;
1927
1928	/*
1929	 * If dsl_scan_ddt() has already visited this block, it will have
1930	 * already done any translations or scrubbing, so don't call the
1931	 * callback again.
1932	 */
1933	if (ddt_class_contains(dp->dp_spa,
1934	    scn->scn_phys.scn_ddt_class_max, bp)) {
1935		scn->scn_ddt_contained_this_txg++;
1936		goto out;
1937	}
1938
1939	/*
1940	 * If this block is from the future (after cur_max_txg), then we
1941	 * are doing this on behalf of a deleted snapshot, and we will
1942	 * revisit the future block on the next pass of this dataset.
1943	 * Don't scan it now unless we need to because something
1944	 * under it was modified.
1945	 */
1946	if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
1947		scn->scn_gt_max_this_txg++;
1948		goto out;
1949	}
1950
1951	scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
1952
1953out:
1954	kmem_free(bp_toread, sizeof (blkptr_t));
1955}
1956
1957static void
1958dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
1959    dmu_tx_t *tx)
1960{
1961	zbookmark_phys_t zb;
1962	scan_prefetch_ctx_t *spc;
1963
1964	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1965	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1966
1967	if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
1968		SET_BOOKMARK(&scn->scn_prefetch_bookmark,
1969		    zb.zb_objset, 0, 0, 0);
1970	} else {
1971		scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
1972	}
1973
1974	scn->scn_objsets_visited_this_txg++;
1975
1976	spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
1977	dsl_scan_prefetch(spc, bp, &zb);
1978	scan_prefetch_ctx_rele(spc, FTAG);
1979
1980	dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
1981
1982	dprintf_ds(ds, "finished scan%s", "");
1983}
1984
1985static void
1986ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
1987{
1988	if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
1989		if (ds->ds_is_snapshot) {
1990			/*
1991			 * Note:
1992			 *  - scn_cur_{min,max}_txg stays the same.
1993			 *  - Setting the flag is not really necessary if
1994			 *    scn_cur_max_txg == scn_max_txg, because there
1995			 *    is nothing after this snapshot that we care
1996			 *    about.  However, we set it anyway and then
1997			 *    ignore it when we retraverse it in
1998			 *    dsl_scan_visitds().
1999			 */
2000			scn_phys->scn_bookmark.zb_objset =
2001			    dsl_dataset_phys(ds)->ds_next_snap_obj;
2002			zfs_dbgmsg("destroying ds %llu; currently traversing; "
2003			    "reset zb_objset to %llu",
2004			    (u_longlong_t)ds->ds_object,
2005			    (u_longlong_t)dsl_dataset_phys(ds)->
2006			    ds_next_snap_obj);
2007			scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2008		} else {
2009			SET_BOOKMARK(&scn_phys->scn_bookmark,
2010			    ZB_DESTROYED_OBJSET, 0, 0, 0);
2011			zfs_dbgmsg("destroying ds %llu; currently traversing; "
2012			    "reset bookmark to -1,0,0,0",
2013			    (u_longlong_t)ds->ds_object);
2014		}
2015	}
2016}
2017
2018/*
2019 * Invoked when a dataset is destroyed. We need to make sure that:
2020 *
2021 * 1) If it is the dataset that was currently being scanned, we write
2022 *	a new dsl_scan_phys_t and marking the objset reference in it
2023 *	as destroyed.
2024 * 2) Remove it from the work queue, if it was present.
2025 *
2026 * If the dataset was actually a snapshot, instead of marking the dataset
2027 * as destroyed, we instead substitute the next snapshot in line.
2028 */
2029void
2030dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2031{
2032	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2033	dsl_scan_t *scn = dp->dp_scan;
2034	uint64_t mintxg;
2035
2036	if (!dsl_scan_is_running(scn))
2037		return;
2038
2039	ds_destroyed_scn_phys(ds, &scn->scn_phys);
2040	ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2041
2042	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2043		scan_ds_queue_remove(scn, ds->ds_object);
2044		if (ds->ds_is_snapshot)
2045			scan_ds_queue_insert(scn,
2046			    dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2047	}
2048
2049	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2050	    ds->ds_object, &mintxg) == 0) {
2051		ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2052		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2053		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2054		if (ds->ds_is_snapshot) {
2055			/*
2056			 * We keep the same mintxg; it could be >
2057			 * ds_creation_txg if the previous snapshot was
2058			 * deleted too.
2059			 */
2060			VERIFY(zap_add_int_key(dp->dp_meta_objset,
2061			    scn->scn_phys.scn_queue_obj,
2062			    dsl_dataset_phys(ds)->ds_next_snap_obj,
2063			    mintxg, tx) == 0);
2064			zfs_dbgmsg("destroying ds %llu; in queue; "
2065			    "replacing with %llu",
2066			    (u_longlong_t)ds->ds_object,
2067			    (u_longlong_t)dsl_dataset_phys(ds)->
2068			    ds_next_snap_obj);
2069		} else {
2070			zfs_dbgmsg("destroying ds %llu; in queue; removing",
2071			    (u_longlong_t)ds->ds_object);
2072		}
2073	}
2074
2075	/*
2076	 * dsl_scan_sync() should be called after this, and should sync
2077	 * out our changed state, but just to be safe, do it here.
2078	 */
2079	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2080}
2081
2082static void
2083ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2084{
2085	if (scn_bookmark->zb_objset == ds->ds_object) {
2086		scn_bookmark->zb_objset =
2087		    dsl_dataset_phys(ds)->ds_prev_snap_obj;
2088		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2089		    "reset zb_objset to %llu",
2090		    (u_longlong_t)ds->ds_object,
2091		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2092	}
2093}
2094
2095/*
2096 * Called when a dataset is snapshotted. If we were currently traversing
2097 * this snapshot, we reset our bookmark to point at the newly created
2098 * snapshot. We also modify our work queue to remove the old snapshot and
2099 * replace with the new one.
2100 */
2101void
2102dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2103{
2104	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2105	dsl_scan_t *scn = dp->dp_scan;
2106	uint64_t mintxg;
2107
2108	if (!dsl_scan_is_running(scn))
2109		return;
2110
2111	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2112
2113	ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2114	ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2115
2116	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2117		scan_ds_queue_remove(scn, ds->ds_object);
2118		scan_ds_queue_insert(scn,
2119		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2120	}
2121
2122	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2123	    ds->ds_object, &mintxg) == 0) {
2124		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2125		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2126		VERIFY(zap_add_int_key(dp->dp_meta_objset,
2127		    scn->scn_phys.scn_queue_obj,
2128		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2129		zfs_dbgmsg("snapshotting ds %llu; in queue; "
2130		    "replacing with %llu",
2131		    (u_longlong_t)ds->ds_object,
2132		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2133	}
2134
2135	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2136}
2137
2138static void
2139ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2140    zbookmark_phys_t *scn_bookmark)
2141{
2142	if (scn_bookmark->zb_objset == ds1->ds_object) {
2143		scn_bookmark->zb_objset = ds2->ds_object;
2144		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2145		    "reset zb_objset to %llu",
2146		    (u_longlong_t)ds1->ds_object,
2147		    (u_longlong_t)ds2->ds_object);
2148	} else if (scn_bookmark->zb_objset == ds2->ds_object) {
2149		scn_bookmark->zb_objset = ds1->ds_object;
2150		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2151		    "reset zb_objset to %llu",
2152		    (u_longlong_t)ds2->ds_object,
2153		    (u_longlong_t)ds1->ds_object);
2154	}
2155}
2156
2157/*
2158 * Called when a parent dataset and its clone are swapped. If we were
2159 * currently traversing the dataset, we need to switch to traversing the
2160 * newly promoted parent.
2161 */
2162void
2163dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2164{
2165	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2166	dsl_scan_t *scn = dp->dp_scan;
2167	uint64_t mintxg;
2168
2169	if (!dsl_scan_is_running(scn))
2170		return;
2171
2172	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2173	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2174
2175	if (scan_ds_queue_contains(scn, ds1->ds_object, &mintxg)) {
2176		scan_ds_queue_remove(scn, ds1->ds_object);
2177		scan_ds_queue_insert(scn, ds2->ds_object, mintxg);
2178	}
2179	if (scan_ds_queue_contains(scn, ds2->ds_object, &mintxg)) {
2180		scan_ds_queue_remove(scn, ds2->ds_object);
2181		scan_ds_queue_insert(scn, ds1->ds_object, mintxg);
2182	}
2183
2184	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2185	    ds1->ds_object, &mintxg) == 0) {
2186		int err;
2187		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2188		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2189		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2190		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2191		err = zap_add_int_key(dp->dp_meta_objset,
2192		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
2193		VERIFY(err == 0 || err == EEXIST);
2194		if (err == EEXIST) {
2195			/* Both were there to begin with */
2196			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2197			    scn->scn_phys.scn_queue_obj,
2198			    ds1->ds_object, mintxg, tx));
2199		}
2200		zfs_dbgmsg("clone_swap ds %llu; in queue; "
2201		    "replacing with %llu",
2202		    (u_longlong_t)ds1->ds_object,
2203		    (u_longlong_t)ds2->ds_object);
2204	}
2205	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2206	    ds2->ds_object, &mintxg) == 0) {
2207		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2208		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2209		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2210		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2211		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2212		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
2213		zfs_dbgmsg("clone_swap ds %llu; in queue; "
2214		    "replacing with %llu",
2215		    (u_longlong_t)ds2->ds_object,
2216		    (u_longlong_t)ds1->ds_object);
2217	}
2218
2219	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2220}
2221
2222/* ARGSUSED */
2223static int
2224enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2225{
2226	uint64_t originobj = *(uint64_t *)arg;
2227	dsl_dataset_t *ds;
2228	int err;
2229	dsl_scan_t *scn = dp->dp_scan;
2230
2231	if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2232		return (0);
2233
2234	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2235	if (err)
2236		return (err);
2237
2238	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2239		dsl_dataset_t *prev;
2240		err = dsl_dataset_hold_obj(dp,
2241		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2242
2243		dsl_dataset_rele(ds, FTAG);
2244		if (err)
2245			return (err);
2246		ds = prev;
2247	}
2248	scan_ds_queue_insert(scn, ds->ds_object,
2249	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2250	dsl_dataset_rele(ds, FTAG);
2251	return (0);
2252}
2253
2254static void
2255dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2256{
2257	dsl_pool_t *dp = scn->scn_dp;
2258	dsl_dataset_t *ds;
2259
2260	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2261
2262	if (scn->scn_phys.scn_cur_min_txg >=
2263	    scn->scn_phys.scn_max_txg) {
2264		/*
2265		 * This can happen if this snapshot was created after the
2266		 * scan started, and we already completed a previous snapshot
2267		 * that was created after the scan started.  This snapshot
2268		 * only references blocks with:
2269		 *
2270		 *	birth < our ds_creation_txg
2271		 *	cur_min_txg is no less than ds_creation_txg.
2272		 *	We have already visited these blocks.
2273		 * or
2274		 *	birth > scn_max_txg
2275		 *	The scan requested not to visit these blocks.
2276		 *
2277		 * Subsequent snapshots (and clones) can reference our
2278		 * blocks, or blocks with even higher birth times.
2279		 * Therefore we do not need to visit them either,
2280		 * so we do not add them to the work queue.
2281		 *
2282		 * Note that checking for cur_min_txg >= cur_max_txg
2283		 * is not sufficient, because in that case we may need to
2284		 * visit subsequent snapshots.  This happens when min_txg > 0,
2285		 * which raises cur_min_txg.  In this case we will visit
2286		 * this dataset but skip all of its blocks, because the
2287		 * rootbp's birth time is < cur_min_txg.  Then we will
2288		 * add the next snapshots/clones to the work queue.
2289		 */
2290		char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
2291		dsl_dataset_name(ds, dsname);
2292		zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2293		    "cur_min_txg (%llu) >= max_txg (%llu)",
2294		    (longlong_t)dsobj, dsname,
2295		    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2296		    (longlong_t)scn->scn_phys.scn_max_txg);
2297		kmem_free(dsname, MAXNAMELEN);
2298
2299		goto out;
2300	}
2301
2302	/*
2303	 * Only the ZIL in the head (non-snapshot) is valid. Even though
2304	 * snapshots can have ZIL block pointers (which may be the same
2305	 * BP as in the head), they must be ignored. In addition, $ORIGIN
2306	 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2307	 * need to look for a ZIL in it either. So we traverse the ZIL here,
2308	 * rather than in scan_recurse(), because the regular snapshot
2309	 * block-sharing rules don't apply to it.
2310	 */
2311	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds) &&
2312	    (dp->dp_origin_snap == NULL ||
2313	    ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2314		objset_t *os;
2315		if (dmu_objset_from_ds(ds, &os) != 0) {
2316			goto out;
2317		}
2318		dsl_scan_zil(dp, &os->os_zil_header);
2319	}
2320
2321	/*
2322	 * Iterate over the bps in this ds.
2323	 */
2324	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2325	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2326	dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2327	rrw_exit(&ds->ds_bp_rwlock, FTAG);
2328
2329	char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2330	dsl_dataset_name(ds, dsname);
2331	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2332	    "suspending=%u",
2333	    (longlong_t)dsobj, dsname,
2334	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2335	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
2336	    (int)scn->scn_suspending);
2337	kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2338
2339	if (scn->scn_suspending)
2340		goto out;
2341
2342	/*
2343	 * We've finished this pass over this dataset.
2344	 */
2345
2346	/*
2347	 * If we did not completely visit this dataset, do another pass.
2348	 */
2349	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2350		zfs_dbgmsg("incomplete pass; visiting again");
2351		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2352		scan_ds_queue_insert(scn, ds->ds_object,
2353		    scn->scn_phys.scn_cur_max_txg);
2354		goto out;
2355	}
2356
2357	/*
2358	 * Add descendent datasets to work queue.
2359	 */
2360	if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2361		scan_ds_queue_insert(scn,
2362		    dsl_dataset_phys(ds)->ds_next_snap_obj,
2363		    dsl_dataset_phys(ds)->ds_creation_txg);
2364	}
2365	if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2366		boolean_t usenext = B_FALSE;
2367		if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2368			uint64_t count;
2369			/*
2370			 * A bug in a previous version of the code could
2371			 * cause upgrade_clones_cb() to not set
2372			 * ds_next_snap_obj when it should, leading to a
2373			 * missing entry.  Therefore we can only use the
2374			 * next_clones_obj when its count is correct.
2375			 */
2376			int err = zap_count(dp->dp_meta_objset,
2377			    dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2378			if (err == 0 &&
2379			    count == dsl_dataset_phys(ds)->ds_num_children - 1)
2380				usenext = B_TRUE;
2381		}
2382
2383		if (usenext) {
2384			zap_cursor_t zc;
2385			zap_attribute_t za;
2386			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2387			    dsl_dataset_phys(ds)->ds_next_clones_obj);
2388			    zap_cursor_retrieve(&zc, &za) == 0;
2389			    (void) zap_cursor_advance(&zc)) {
2390				scan_ds_queue_insert(scn,
2391				    zfs_strtonum(za.za_name, NULL),
2392				    dsl_dataset_phys(ds)->ds_creation_txg);
2393			}
2394			zap_cursor_fini(&zc);
2395		} else {
2396			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2397			    enqueue_clones_cb, &ds->ds_object,
2398			    DS_FIND_CHILDREN));
2399		}
2400	}
2401
2402out:
2403	dsl_dataset_rele(ds, FTAG);
2404}
2405
2406/* ARGSUSED */
2407static int
2408enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2409{
2410	dsl_dataset_t *ds;
2411	int err;
2412	dsl_scan_t *scn = dp->dp_scan;
2413
2414	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2415	if (err)
2416		return (err);
2417
2418	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2419		dsl_dataset_t *prev;
2420		err = dsl_dataset_hold_obj(dp,
2421		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2422		if (err) {
2423			dsl_dataset_rele(ds, FTAG);
2424			return (err);
2425		}
2426
2427		/*
2428		 * If this is a clone, we don't need to worry about it for now.
2429		 */
2430		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2431			dsl_dataset_rele(ds, FTAG);
2432			dsl_dataset_rele(prev, FTAG);
2433			return (0);
2434		}
2435		dsl_dataset_rele(ds, FTAG);
2436		ds = prev;
2437	}
2438
2439	scan_ds_queue_insert(scn, ds->ds_object,
2440	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2441	dsl_dataset_rele(ds, FTAG);
2442	return (0);
2443}
2444
2445/* ARGSUSED */
2446void
2447dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2448    ddt_entry_t *dde, dmu_tx_t *tx)
2449{
2450	const ddt_key_t *ddk = &dde->dde_key;
2451	ddt_phys_t *ddp = dde->dde_phys;
2452	blkptr_t bp;
2453	zbookmark_phys_t zb = { 0 };
2454	int p;
2455
2456	if (scn->scn_phys.scn_state != DSS_SCANNING)
2457		return;
2458
2459	/*
2460	 * This function is special because it is the only thing
2461	 * that can add scan_io_t's to the vdev scan queues from
2462	 * outside dsl_scan_sync(). For the most part this is ok
2463	 * as long as it is called from within syncing context.
2464	 * However, dsl_scan_sync() expects that no new sio's will
2465	 * be added between when all the work for a scan is done
2466	 * and the next txg when the scan is actually marked as
2467	 * completed. This check ensures we do not issue new sio's
2468	 * during this period.
2469	 */
2470	if (scn->scn_done_txg != 0)
2471		return;
2472
2473	for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2474		if (ddp->ddp_phys_birth == 0 ||
2475		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2476			continue;
2477		ddt_bp_create(checksum, ddk, ddp, &bp);
2478
2479		scn->scn_visited_this_txg++;
2480		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2481	}
2482}
2483
2484/*
2485 * Scrub/dedup interaction.
2486 *
2487 * If there are N references to a deduped block, we don't want to scrub it
2488 * N times -- ideally, we should scrub it exactly once.
2489 *
2490 * We leverage the fact that the dde's replication class (enum ddt_class)
2491 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2492 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2493 *
2494 * To prevent excess scrubbing, the scrub begins by walking the DDT
2495 * to find all blocks with refcnt > 1, and scrubs each of these once.
2496 * Since there are two replication classes which contain blocks with
2497 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2498 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2499 *
2500 * There would be nothing more to say if a block's refcnt couldn't change
2501 * during a scrub, but of course it can so we must account for changes
2502 * in a block's replication class.
2503 *
2504 * Here's an example of what can occur:
2505 *
2506 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2507 * when visited during the top-down scrub phase, it will be scrubbed twice.
2508 * This negates our scrub optimization, but is otherwise harmless.
2509 *
2510 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2511 * on each visit during the top-down scrub phase, it will never be scrubbed.
2512 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2513 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2514 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2515 * while a scrub is in progress, it scrubs the block right then.
2516 */
2517static void
2518dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2519{
2520	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2521	ddt_entry_t dde = { 0 };
2522	int error;
2523	uint64_t n = 0;
2524
2525	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2526		ddt_t *ddt;
2527
2528		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2529			break;
2530		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2531		    (longlong_t)ddb->ddb_class,
2532		    (longlong_t)ddb->ddb_type,
2533		    (longlong_t)ddb->ddb_checksum,
2534		    (longlong_t)ddb->ddb_cursor);
2535
2536		/* There should be no pending changes to the dedup table */
2537		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2538		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2539
2540		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2541		n++;
2542
2543		if (dsl_scan_check_suspend(scn, NULL))
2544			break;
2545	}
2546
2547	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2548	    "suspending=%u", (longlong_t)n,
2549	    (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2550
2551	ASSERT(error == 0 || error == ENOENT);
2552	ASSERT(error != ENOENT ||
2553	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2554}
2555
2556static uint64_t
2557dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2558{
2559	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2560	if (ds->ds_is_snapshot)
2561		return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2562	return (smt);
2563}
2564
2565static void
2566dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2567{
2568	scan_ds_t *sds;
2569	dsl_pool_t *dp = scn->scn_dp;
2570
2571	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2572	    scn->scn_phys.scn_ddt_class_max) {
2573		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2574		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2575		dsl_scan_ddt(scn, tx);
2576		if (scn->scn_suspending)
2577			return;
2578	}
2579
2580	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2581		/* First do the MOS & ORIGIN */
2582
2583		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2584		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2585		dsl_scan_visit_rootbp(scn, NULL,
2586		    &dp->dp_meta_rootbp, tx);
2587		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2588		if (scn->scn_suspending)
2589			return;
2590
2591		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2592			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2593			    enqueue_cb, NULL, DS_FIND_CHILDREN));
2594		} else {
2595			dsl_scan_visitds(scn,
2596			    dp->dp_origin_snap->ds_object, tx);
2597		}
2598		ASSERT(!scn->scn_suspending);
2599	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
2600	    ZB_DESTROYED_OBJSET) {
2601		uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2602		/*
2603		 * If we were suspended, continue from here. Note if the
2604		 * ds we were suspended on was deleted, the zb_objset may
2605		 * be -1, so we will skip this and find a new objset
2606		 * below.
2607		 */
2608		dsl_scan_visitds(scn, dsobj, tx);
2609		if (scn->scn_suspending)
2610			return;
2611	}
2612
2613	/*
2614	 * In case we suspended right at the end of the ds, zero the
2615	 * bookmark so we don't think that we're still trying to resume.
2616	 */
2617	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
2618
2619	/*
2620	 * Keep pulling things out of the dataset avl queue. Updates to the
2621	 * persistent zap-object-as-queue happen only at checkpoints.
2622	 */
2623	while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2624		dsl_dataset_t *ds;
2625		uint64_t dsobj = sds->sds_dsobj;
2626		uint64_t txg = sds->sds_txg;
2627
2628		/* dequeue and free the ds from the queue */
2629		scan_ds_queue_remove(scn, dsobj);
2630		sds = NULL;	/* must not be touched after removal */
2631
2632		/* Set up min / max txg */
2633		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2634		if (txg != 0) {
2635			scn->scn_phys.scn_cur_min_txg =
2636			    MAX(scn->scn_phys.scn_min_txg, txg);
2637		} else {
2638			scn->scn_phys.scn_cur_min_txg =
2639			    MAX(scn->scn_phys.scn_min_txg,
2640			    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2641		}
2642		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2643		dsl_dataset_rele(ds, FTAG);
2644
2645		dsl_scan_visitds(scn, dsobj, tx);
2646		if (scn->scn_suspending)
2647			return;
2648	}
2649	/* No more objsets to fetch, we're done */
2650	scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2651	ASSERT0(scn->scn_suspending);
2652}
2653
2654static uint64_t
2655dsl_scan_count_leaves(vdev_t *vd)
2656{
2657	uint64_t i, leaves = 0;
2658
2659	/* we only count leaves that belong to the main pool and are readable */
2660	if (vd->vdev_islog || vd->vdev_isspare ||
2661	    vd->vdev_isl2cache || !vdev_readable(vd))
2662		return (0);
2663
2664	if (vd->vdev_ops->vdev_op_leaf)
2665		return (1);
2666
2667	for (i = 0; i < vd->vdev_children; i++) {
2668		leaves += dsl_scan_count_leaves(vd->vdev_child[i]);
2669	}
2670
2671	return (leaves);
2672}
2673
2674
2675static void
2676scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2677{
2678	int i;
2679	uint64_t cur_size = 0;
2680
2681	for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2682		cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2683	}
2684
2685	q->q_total_zio_size_this_txg += cur_size;
2686	q->q_zios_this_txg++;
2687}
2688
2689static void
2690scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2691    uint64_t end)
2692{
2693	q->q_total_seg_size_this_txg += end - start;
2694	q->q_segs_this_txg++;
2695}
2696
2697static boolean_t
2698scan_io_queue_check_suspend(dsl_scan_t *scn)
2699{
2700	/* See comment in dsl_scan_check_suspend() */
2701	uint64_t curr_time_ns = gethrtime();
2702	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2703	uint64_t sync_time_ns = curr_time_ns -
2704	    scn->scn_dp->dp_spa->spa_sync_starttime;
2705	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
2706	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2707	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2708
2709	return ((NSEC2MSEC(scan_time_ns) > mintime &&
2710	    (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
2711	    txg_sync_waiting(scn->scn_dp) ||
2712	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2713	    spa_shutting_down(scn->scn_dp->dp_spa));
2714}
2715
2716/*
2717 * Given a list of scan_io_t's in io_list, this issues the io's out to
2718 * disk. This consumes the io_list and frees the scan_io_t's. This is
2719 * called when emptying queues, either when we're up against the memory
2720 * limit or when we have finished scanning. Returns B_TRUE if we stopped
2721 * processing the list before we finished. Any zios that were not issued
2722 * will remain in the io_list.
2723 */
2724static boolean_t
2725scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2726{
2727	dsl_scan_t *scn = queue->q_scn;
2728	scan_io_t *sio;
2729	int64_t bytes_issued = 0;
2730	boolean_t suspended = B_FALSE;
2731
2732	while ((sio = list_head(io_list)) != NULL) {
2733		blkptr_t bp;
2734
2735		if (scan_io_queue_check_suspend(scn)) {
2736			suspended = B_TRUE;
2737			break;
2738		}
2739
2740		sio2bp(sio, &bp);
2741		bytes_issued += SIO_GET_ASIZE(sio);
2742		scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2743		    &sio->sio_zb, queue);
2744		(void) list_remove_head(io_list);
2745		scan_io_queues_update_zio_stats(queue, &bp);
2746		sio_free(sio);
2747	}
2748
2749	atomic_add_64(&scn->scn_bytes_pending, -bytes_issued);
2750
2751	return (suspended);
2752}
2753
2754/*
2755 * Given a range_seg_t (extent) and a list, this function passes over a
2756 * scan queue and gathers up the appropriate ios which fit into that
2757 * scan seg (starting from lowest LBA). At the end, we remove the segment
2758 * from the q_exts_by_addr range tree.
2759 */
2760static boolean_t
2761scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2762{
2763	scan_io_t *srch_sio, *sio, *next_sio;
2764	avl_index_t idx;
2765	uint_t num_sios = 0;
2766	int64_t bytes_issued = 0;
2767
2768	ASSERT(rs != NULL);
2769	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2770
2771	srch_sio = sio_alloc(1);
2772	srch_sio->sio_nr_dvas = 1;
2773	SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
2774
2775	/*
2776	 * The exact start of the extent might not contain any matching zios,
2777	 * so if that's the case, examine the next one in the tree.
2778	 */
2779	sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
2780	sio_free(srch_sio);
2781
2782	if (sio == NULL)
2783		sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2784
2785	while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2786	    queue->q_exts_by_addr) && num_sios <= 32) {
2787		ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
2788		    queue->q_exts_by_addr));
2789		ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
2790		    queue->q_exts_by_addr));
2791
2792		next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2793		avl_remove(&queue->q_sios_by_addr, sio);
2794		queue->q_sio_memused -= SIO_GET_MUSED(sio);
2795
2796		bytes_issued += SIO_GET_ASIZE(sio);
2797		num_sios++;
2798		list_insert_tail(list, sio);
2799		sio = next_sio;
2800	}
2801
2802	/*
2803	 * We limit the number of sios we process at once to 32 to avoid
2804	 * biting off more than we can chew. If we didn't take everything
2805	 * in the segment we update it to reflect the work we were able to
2806	 * complete. Otherwise, we remove it from the range tree entirely.
2807	 */
2808	if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2809	    queue->q_exts_by_addr)) {
2810		range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2811		    -bytes_issued);
2812		range_tree_resize_segment(queue->q_exts_by_addr, rs,
2813		    SIO_GET_OFFSET(sio), rs_get_end(rs,
2814		    queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
2815
2816		return (B_TRUE);
2817	} else {
2818		uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
2819		uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
2820		range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
2821		return (B_FALSE);
2822	}
2823}
2824
2825
2826/*
2827 * This is called from the queue emptying thread and selects the next
2828 * extent from which we are to issue io's. The behavior of this function
2829 * depends on the state of the scan, the current memory consumption and
2830 * whether or not we are performing a scan shutdown.
2831 * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2832 *	needs to perform a checkpoint
2833 * 2) We select the largest available extent if we are up against the
2834 *	memory limit.
2835 * 3) Otherwise we don't select any extents.
2836 */
2837static const range_seg_t *
2838scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2839{
2840	dsl_scan_t *scn = queue->q_scn;
2841	range_tree_t *rt = queue->q_exts_by_addr;
2842
2843	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2844	ASSERT(scn->scn_is_sorted);
2845
2846	/* handle tunable overrides */
2847	if (scn->scn_checkpointing || scn->scn_clearing) {
2848		if (zfs_scan_issue_strategy == 1) {
2849			return (range_tree_first(rt));
2850		} else if (zfs_scan_issue_strategy == 2) {
2851			range_seg_t *size_rs =
2852			    zfs_btree_first(&queue->q_exts_by_size, NULL);
2853			uint64_t start = rs_get_start(size_rs, rt);
2854			uint64_t size = rs_get_end(size_rs, rt) - start;
2855			range_seg_t *addr_rs = range_tree_find(rt, start,
2856			    size);
2857			ASSERT3P(addr_rs, !=, NULL);
2858			return (addr_rs);
2859		}
2860	}
2861
2862	/*
2863	 * During normal clearing, we want to issue our largest segments
2864	 * first, keeping IO as sequential as possible, and leaving the
2865	 * smaller extents for later with the hope that they might eventually
2866	 * grow to larger sequential segments. However, when the scan is
2867	 * checkpointing, no new extents will be added to the sorting queue,
2868	 * so the way we are sorted now is as good as it will ever get.
2869	 * In this case, we instead switch to issuing extents in LBA order.
2870	 */
2871	if (scn->scn_checkpointing) {
2872		return (range_tree_first(rt));
2873	} else if (scn->scn_clearing) {
2874		range_seg_t *size_rs = zfs_btree_first(&queue->q_exts_by_size,
2875		    NULL);
2876		uint64_t start = rs_get_start(size_rs, rt);
2877		uint64_t size = rs_get_end(size_rs, rt) - start;
2878		range_seg_t *addr_rs = range_tree_find(rt, start, size);
2879		ASSERT3P(addr_rs, !=, NULL);
2880		return (addr_rs);
2881	} else {
2882		return (NULL);
2883	}
2884}
2885
2886static void
2887scan_io_queues_run_one(void *arg)
2888{
2889	dsl_scan_io_queue_t *queue = arg;
2890	kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
2891	boolean_t suspended = B_FALSE;
2892	range_seg_t *rs = NULL;
2893	scan_io_t *sio = NULL;
2894	list_t sio_list;
2895	uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
2896	uint64_t nr_leaves = dsl_scan_count_leaves(queue->q_vd);
2897
2898	ASSERT(queue->q_scn->scn_is_sorted);
2899
2900	list_create(&sio_list, sizeof (scan_io_t),
2901	    offsetof(scan_io_t, sio_nodes.sio_list_node));
2902	mutex_enter(q_lock);
2903
2904	/* calculate maximum in-flight bytes for this txg (min 1MB) */
2905	queue->q_maxinflight_bytes =
2906	    MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
2907
2908	/* reset per-queue scan statistics for this txg */
2909	queue->q_total_seg_size_this_txg = 0;
2910	queue->q_segs_this_txg = 0;
2911	queue->q_total_zio_size_this_txg = 0;
2912	queue->q_zios_this_txg = 0;
2913
2914	/* loop until we have run out of time or sios */
2915	while ((rs = (range_seg_t *)scan_io_queue_fetch_ext(queue)) != NULL) {
2916		uint64_t seg_start = 0, seg_end = 0;
2917		boolean_t more_left = B_TRUE;
2918
2919		ASSERT(list_is_empty(&sio_list));
2920
2921		/* loop while we still have sios left to process in this rs */
2922		while (more_left) {
2923			scan_io_t *first_sio, *last_sio;
2924
2925			/*
2926			 * We have selected which extent needs to be
2927			 * processed next. Gather up the corresponding sios.
2928			 */
2929			more_left = scan_io_queue_gather(queue, rs, &sio_list);
2930			ASSERT(!list_is_empty(&sio_list));
2931			first_sio = list_head(&sio_list);
2932			last_sio = list_tail(&sio_list);
2933
2934			seg_end = SIO_GET_END_OFFSET(last_sio);
2935			if (seg_start == 0)
2936				seg_start = SIO_GET_OFFSET(first_sio);
2937
2938			/*
2939			 * Issuing sios can take a long time so drop the
2940			 * queue lock. The sio queue won't be updated by
2941			 * other threads since we're in syncing context so
2942			 * we can be sure that our trees will remain exactly
2943			 * as we left them.
2944			 */
2945			mutex_exit(q_lock);
2946			suspended = scan_io_queue_issue(queue, &sio_list);
2947			mutex_enter(q_lock);
2948
2949			if (suspended)
2950				break;
2951		}
2952		/* update statistics for debugging purposes */
2953		scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
2954
2955		if (suspended)
2956			break;
2957	}
2958
2959
2960	/*
2961	 * If we were suspended in the middle of processing,
2962	 * requeue any unfinished sios and exit.
2963	 */
2964	while ((sio = list_head(&sio_list)) != NULL) {
2965		list_remove(&sio_list, sio);
2966		scan_io_queue_insert_impl(queue, sio);
2967	}
2968
2969	mutex_exit(q_lock);
2970	list_destroy(&sio_list);
2971}
2972
2973/*
2974 * Performs an emptying run on all scan queues in the pool. This just
2975 * punches out one thread per top-level vdev, each of which processes
2976 * only that vdev's scan queue. We can parallelize the I/O here because
2977 * we know that each queue's io's only affect its own top-level vdev.
2978 *
2979 * This function waits for the queue runs to complete, and must be
2980 * called from dsl_scan_sync (or in general, syncing context).
2981 */
2982static void
2983scan_io_queues_run(dsl_scan_t *scn)
2984{
2985	spa_t *spa = scn->scn_dp->dp_spa;
2986
2987	ASSERT(scn->scn_is_sorted);
2988	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
2989
2990	if (scn->scn_bytes_pending == 0)
2991		return;
2992
2993	if (scn->scn_taskq == NULL) {
2994		char *tq_name = kmem_zalloc(ZFS_MAX_DATASET_NAME_LEN + 16,
2995		    KM_SLEEP);
2996		int nthreads = spa->spa_root_vdev->vdev_children;
2997
2998		/*
2999		 * We need to make this taskq *always* execute as many
3000		 * threads in parallel as we have top-level vdevs and no
3001		 * less, otherwise strange serialization of the calls to
3002		 * scan_io_queues_run_one can occur during spa_sync runs
3003		 * and that significantly impacts performance.
3004		 */
3005		(void) snprintf(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16,
3006		    "dsl_scan_tq_%s", spa->spa_name);
3007		scn->scn_taskq = taskq_create(tq_name, nthreads, minclsyspri,
3008		    nthreads, nthreads, TASKQ_PREPOPULATE);
3009		kmem_free(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16);
3010	}
3011
3012	for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3013		vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3014
3015		mutex_enter(&vd->vdev_scan_io_queue_lock);
3016		if (vd->vdev_scan_io_queue != NULL) {
3017			VERIFY(taskq_dispatch(scn->scn_taskq,
3018			    scan_io_queues_run_one, vd->vdev_scan_io_queue,
3019			    TQ_SLEEP) != TASKQID_INVALID);
3020		}
3021		mutex_exit(&vd->vdev_scan_io_queue_lock);
3022	}
3023
3024	/*
3025	 * Wait for the queues to finish issuing thir IOs for this run
3026	 * before we return. There may still be IOs in flight at this
3027	 * point.
3028	 */
3029	taskq_wait(scn->scn_taskq);
3030}
3031
3032static boolean_t
3033dsl_scan_async_block_should_pause(dsl_scan_t *scn)
3034{
3035	uint64_t elapsed_nanosecs;
3036
3037	if (zfs_recover)
3038		return (B_FALSE);
3039
3040	if (scn->scn_visited_this_txg >= zfs_async_block_max_blocks)
3041		return (B_TRUE);
3042
3043	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
3044	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
3045	    (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
3046	    txg_sync_waiting(scn->scn_dp)) ||
3047	    spa_shutting_down(scn->scn_dp->dp_spa));
3048}
3049
3050static int
3051dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3052{
3053	dsl_scan_t *scn = arg;
3054
3055	if (!scn->scn_is_bptree ||
3056	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
3057		if (dsl_scan_async_block_should_pause(scn))
3058			return (SET_ERROR(ERESTART));
3059	}
3060
3061	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
3062	    dmu_tx_get_txg(tx), bp, 0));
3063	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
3064	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
3065	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
3066	scn->scn_visited_this_txg++;
3067	return (0);
3068}
3069
3070static void
3071dsl_scan_update_stats(dsl_scan_t *scn)
3072{
3073	spa_t *spa = scn->scn_dp->dp_spa;
3074	uint64_t i;
3075	uint64_t seg_size_total = 0, zio_size_total = 0;
3076	uint64_t seg_count_total = 0, zio_count_total = 0;
3077
3078	for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3079		vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3080		dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
3081
3082		if (queue == NULL)
3083			continue;
3084
3085		seg_size_total += queue->q_total_seg_size_this_txg;
3086		zio_size_total += queue->q_total_zio_size_this_txg;
3087		seg_count_total += queue->q_segs_this_txg;
3088		zio_count_total += queue->q_zios_this_txg;
3089	}
3090
3091	if (seg_count_total == 0 || zio_count_total == 0) {
3092		scn->scn_avg_seg_size_this_txg = 0;
3093		scn->scn_avg_zio_size_this_txg = 0;
3094		scn->scn_segs_this_txg = 0;
3095		scn->scn_zios_this_txg = 0;
3096		return;
3097	}
3098
3099	scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
3100	scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
3101	scn->scn_segs_this_txg = seg_count_total;
3102	scn->scn_zios_this_txg = zio_count_total;
3103}
3104
3105static int
3106dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
3107{
3108	dsl_scan_t *scn = arg;
3109	const dva_t *dva = &bp->blk_dva[0];
3110
3111	if (dsl_scan_async_block_should_pause(scn))
3112		return (SET_ERROR(ERESTART));
3113
3114	spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
3115	    DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
3116	    DVA_GET_ASIZE(dva), tx);
3117	scn->scn_visited_this_txg++;
3118	return (0);
3119}
3120
3121boolean_t
3122dsl_scan_active(dsl_scan_t *scn)
3123{
3124	spa_t *spa = scn->scn_dp->dp_spa;
3125	uint64_t used = 0, comp, uncomp;
3126
3127	if (spa->spa_load_state != SPA_LOAD_NONE)
3128		return (B_FALSE);
3129	if (spa_shutting_down(spa))
3130		return (B_FALSE);
3131	if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
3132	    (scn->scn_async_destroying && !scn->scn_async_stalled))
3133		return (B_TRUE);
3134
3135	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
3136		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
3137		    &used, &comp, &uncomp);
3138	}
3139	return (used != 0);
3140}
3141
3142static boolean_t
3143dsl_scan_check_deferred(vdev_t *vd)
3144{
3145	boolean_t need_resilver = B_FALSE;
3146
3147	for (int c = 0; c < vd->vdev_children; c++) {
3148		need_resilver |=
3149		    dsl_scan_check_deferred(vd->vdev_child[c]);
3150	}
3151
3152	if (!vdev_is_concrete(vd) || vd->vdev_aux ||
3153	    !vd->vdev_ops->vdev_op_leaf)
3154		return (need_resilver);
3155
3156	if (!vd->vdev_resilver_deferred)
3157		need_resilver = B_TRUE;
3158
3159	return (need_resilver);
3160}
3161
3162static boolean_t
3163dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
3164    uint64_t phys_birth)
3165{
3166	vdev_t *vd;
3167
3168	vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3169
3170	if (vd->vdev_ops == &vdev_indirect_ops) {
3171		/*
3172		 * The indirect vdev can point to multiple
3173		 * vdevs.  For simplicity, always create
3174		 * the resilver zio_t. zio_vdev_io_start()
3175		 * will bypass the child resilver i/o's if
3176		 * they are on vdevs that don't have DTL's.
3177		 */
3178		return (B_TRUE);
3179	}
3180
3181	if (DVA_GET_GANG(dva)) {
3182		/*
3183		 * Gang members may be spread across multiple
3184		 * vdevs, so the best estimate we have is the
3185		 * scrub range, which has already been checked.
3186		 * XXX -- it would be better to change our
3187		 * allocation policy to ensure that all
3188		 * gang members reside on the same vdev.
3189		 */
3190		return (B_TRUE);
3191	}
3192
3193	/*
3194	 * Check if the txg falls within the range which must be
3195	 * resilvered.  DVAs outside this range can always be skipped.
3196	 */
3197	if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
3198		return (B_FALSE);
3199
3200	/*
3201	 * Check if the top-level vdev must resilver this offset.
3202	 * When the offset does not intersect with a dirty leaf DTL
3203	 * then it may be possible to skip the resilver IO.  The psize
3204	 * is provided instead of asize to simplify the check for RAIDZ.
3205	 */
3206	if (!vdev_dtl_need_resilver(vd, DVA_GET_OFFSET(dva), psize))
3207		return (B_FALSE);
3208
3209	/*
3210	 * Check that this top-level vdev has a device under it which
3211	 * is resilvering and is not deferred.
3212	 */
3213	if (!dsl_scan_check_deferred(vd))
3214		return (B_FALSE);
3215
3216	return (B_TRUE);
3217}
3218
3219static int
3220dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3221{
3222	int err = 0;
3223	dsl_scan_t *scn = dp->dp_scan;
3224	spa_t *spa = dp->dp_spa;
3225
3226	if (spa_suspend_async_destroy(spa))
3227		return (0);
3228
3229	if (zfs_free_bpobj_enabled &&
3230	    spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3231		scn->scn_is_bptree = B_FALSE;
3232		scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3233		scn->scn_zio_root = zio_root(spa, NULL,
3234		    NULL, ZIO_FLAG_MUSTSUCCEED);
3235		err = bpobj_iterate(&dp->dp_free_bpobj,
3236		    dsl_scan_free_block_cb, scn, tx);
3237		VERIFY0(zio_wait(scn->scn_zio_root));
3238		scn->scn_zio_root = NULL;
3239
3240		if (err != 0 && err != ERESTART)
3241			zfs_panic_recover("error %u from bpobj_iterate()", err);
3242	}
3243
3244	if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3245		ASSERT(scn->scn_async_destroying);
3246		scn->scn_is_bptree = B_TRUE;
3247		scn->scn_zio_root = zio_root(spa, NULL,
3248		    NULL, ZIO_FLAG_MUSTSUCCEED);
3249		err = bptree_iterate(dp->dp_meta_objset,
3250		    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3251		VERIFY0(zio_wait(scn->scn_zio_root));
3252		scn->scn_zio_root = NULL;
3253
3254		if (err == EIO || err == ECKSUM) {
3255			err = 0;
3256		} else if (err != 0 && err != ERESTART) {
3257			zfs_panic_recover("error %u from "
3258			    "traverse_dataset_destroyed()", err);
3259		}
3260
3261		if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3262			/* finished; deactivate async destroy feature */
3263			spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3264			ASSERT(!spa_feature_is_active(spa,
3265			    SPA_FEATURE_ASYNC_DESTROY));
3266			VERIFY0(zap_remove(dp->dp_meta_objset,
3267			    DMU_POOL_DIRECTORY_OBJECT,
3268			    DMU_POOL_BPTREE_OBJ, tx));
3269			VERIFY0(bptree_free(dp->dp_meta_objset,
3270			    dp->dp_bptree_obj, tx));
3271			dp->dp_bptree_obj = 0;
3272			scn->scn_async_destroying = B_FALSE;
3273			scn->scn_async_stalled = B_FALSE;
3274		} else {
3275			/*
3276			 * If we didn't make progress, mark the async
3277			 * destroy as stalled, so that we will not initiate
3278			 * a spa_sync() on its behalf.  Note that we only
3279			 * check this if we are not finished, because if the
3280			 * bptree had no blocks for us to visit, we can
3281			 * finish without "making progress".
3282			 */
3283			scn->scn_async_stalled =
3284			    (scn->scn_visited_this_txg == 0);
3285		}
3286	}
3287	if (scn->scn_visited_this_txg) {
3288		zfs_dbgmsg("freed %llu blocks in %llums from "
3289		    "free_bpobj/bptree txg %llu; err=%d",
3290		    (longlong_t)scn->scn_visited_this_txg,
3291		    (longlong_t)
3292		    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3293		    (longlong_t)tx->tx_txg, err);
3294		scn->scn_visited_this_txg = 0;
3295
3296		/*
3297		 * Write out changes to the DDT that may be required as a
3298		 * result of the blocks freed.  This ensures that the DDT
3299		 * is clean when a scrub/resilver runs.
3300		 */
3301		ddt_sync(spa, tx->tx_txg);
3302	}
3303	if (err != 0)
3304		return (err);
3305	if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3306	    zfs_free_leak_on_eio &&
3307	    (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3308	    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3309	    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3310		/*
3311		 * We have finished background destroying, but there is still
3312		 * some space left in the dp_free_dir. Transfer this leaked
3313		 * space to the dp_leak_dir.
3314		 */
3315		if (dp->dp_leak_dir == NULL) {
3316			rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3317			(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3318			    LEAK_DIR_NAME, tx);
3319			VERIFY0(dsl_pool_open_special_dir(dp,
3320			    LEAK_DIR_NAME, &dp->dp_leak_dir));
3321			rrw_exit(&dp->dp_config_rwlock, FTAG);
3322		}
3323		dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3324		    dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3325		    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3326		    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3327		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3328		    -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3329		    -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3330		    -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3331	}
3332
3333	if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
3334		/* finished; verify that space accounting went to zero */
3335		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3336		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3337		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3338	}
3339
3340	EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3341	    0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3342	    DMU_POOL_OBSOLETE_BPOBJ));
3343	if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3344		ASSERT(spa_feature_is_active(dp->dp_spa,
3345		    SPA_FEATURE_OBSOLETE_COUNTS));
3346
3347		scn->scn_is_bptree = B_FALSE;
3348		scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3349		err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3350		    dsl_scan_obsolete_block_cb, scn, tx);
3351		if (err != 0 && err != ERESTART)
3352			zfs_panic_recover("error %u from bpobj_iterate()", err);
3353
3354		if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3355			dsl_pool_destroy_obsolete_bpobj(dp, tx);
3356	}
3357
3358	return (0);
3359}
3360
3361/*
3362 * This is the primary entry point for scans that is called from syncing
3363 * context. Scans must happen entirely during syncing context so that we
3364 * cna guarantee that blocks we are currently scanning will not change out
3365 * from under us. While a scan is active, this funciton controls how quickly
3366 * transaction groups proceed, instead of the normal handling provided by
3367 * txg_sync_thread().
3368 */
3369void
3370dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
3371{
3372	dsl_scan_t *scn = dp->dp_scan;
3373	spa_t *spa = dp->dp_spa;
3374	int err = 0;
3375	state_sync_type_t sync_type = SYNC_OPTIONAL;
3376
3377	if (spa->spa_resilver_deferred &&
3378	    !spa_feature_is_active(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))
3379		spa_feature_incr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
3380
3381	/*
3382	 * Check for scn_restart_txg before checking spa_load_state, so
3383	 * that we can restart an old-style scan while the pool is being
3384	 * imported (see dsl_scan_init). We also restart scans if there
3385	 * is a deferred resilver and the user has manually disabled
3386	 * deferred resilvers via the tunable.
3387	 */
3388	if (dsl_scan_restarting(scn, tx) ||
3389	    (spa->spa_resilver_deferred && zfs_resilver_disable_defer)) {
3390		pool_scan_func_t func = POOL_SCAN_SCRUB;
3391		dsl_scan_done(scn, B_FALSE, tx);
3392		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
3393			func = POOL_SCAN_RESILVER;
3394		zfs_dbgmsg("restarting scan func=%u txg=%llu",
3395		    func, (longlong_t)tx->tx_txg);
3396		dsl_scan_setup_sync(&func, tx);
3397	}
3398
3399	/*
3400	 * Only process scans in sync pass 1.
3401	 */
3402	if (spa_sync_pass(dp->dp_spa) > 1)
3403		return;
3404
3405	/*
3406	 * If the spa is shutting down, then stop scanning. This will
3407	 * ensure that the scan does not dirty any new data during the
3408	 * shutdown phase.
3409	 */
3410	if (spa_shutting_down(spa))
3411		return;
3412
3413	/*
3414	 * If the scan is inactive due to a stalled async destroy, try again.
3415	 */
3416	if (!scn->scn_async_stalled && !dsl_scan_active(scn))
3417		return;
3418
3419	/* reset scan statistics */
3420	scn->scn_visited_this_txg = 0;
3421	scn->scn_holes_this_txg = 0;
3422	scn->scn_lt_min_this_txg = 0;
3423	scn->scn_gt_max_this_txg = 0;
3424	scn->scn_ddt_contained_this_txg = 0;
3425	scn->scn_objsets_visited_this_txg = 0;
3426	scn->scn_avg_seg_size_this_txg = 0;
3427	scn->scn_segs_this_txg = 0;
3428	scn->scn_avg_zio_size_this_txg = 0;
3429	scn->scn_zios_this_txg = 0;
3430	scn->scn_suspending = B_FALSE;
3431	scn->scn_sync_start_time = gethrtime();
3432	spa->spa_scrub_active = B_TRUE;
3433
3434	/*
3435	 * First process the async destroys.  If we pause, don't do
3436	 * any scrubbing or resilvering.  This ensures that there are no
3437	 * async destroys while we are scanning, so the scan code doesn't
3438	 * have to worry about traversing it.  It is also faster to free the
3439	 * blocks than to scrub them.
3440	 */
3441	err = dsl_process_async_destroys(dp, tx);
3442	if (err != 0)
3443		return;
3444
3445	if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
3446		return;
3447
3448	/*
3449	 * Wait a few txgs after importing to begin scanning so that
3450	 * we can get the pool imported quickly.
3451	 */
3452	if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
3453		return;
3454
3455	/*
3456	 * zfs_scan_suspend_progress can be set to disable scan progress.
3457	 * We don't want to spin the txg_sync thread, so we add a delay
3458	 * here to simulate the time spent doing a scan. This is mostly
3459	 * useful for testing and debugging.
3460	 */
3461	if (zfs_scan_suspend_progress) {
3462		uint64_t scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3463		int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
3464		    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
3465
3466		while (zfs_scan_suspend_progress &&
3467		    !txg_sync_waiting(scn->scn_dp) &&
3468		    !spa_shutting_down(scn->scn_dp->dp_spa) &&
3469		    NSEC2MSEC(scan_time_ns) < mintime) {
3470			delay(hz);
3471			scan_time_ns = gethrtime() - scn->scn_sync_start_time;
3472		}
3473		return;
3474	}
3475
3476	/*
3477	 * It is possible to switch from unsorted to sorted at any time,
3478	 * but afterwards the scan will remain sorted unless reloaded from
3479	 * a checkpoint after a reboot.
3480	 */
3481	if (!zfs_scan_legacy) {
3482		scn->scn_is_sorted = B_TRUE;
3483		if (scn->scn_last_checkpoint == 0)
3484			scn->scn_last_checkpoint = ddi_get_lbolt();
3485	}
3486
3487	/*
3488	 * For sorted scans, determine what kind of work we will be doing
3489	 * this txg based on our memory limitations and whether or not we
3490	 * need to perform a checkpoint.
3491	 */
3492	if (scn->scn_is_sorted) {
3493		/*
3494		 * If we are over our checkpoint interval, set scn_clearing
3495		 * so that we can begin checkpointing immediately. The
3496		 * checkpoint allows us to save a consisent bookmark
3497		 * representing how much data we have scrubbed so far.
3498		 * Otherwise, use the memory limit to determine if we should
3499		 * scan for metadata or start issue scrub IOs. We accumulate
3500		 * metadata until we hit our hard memory limit at which point
3501		 * we issue scrub IOs until we are at our soft memory limit.
3502		 */
3503		if (scn->scn_checkpointing ||
3504		    ddi_get_lbolt() - scn->scn_last_checkpoint >
3505		    SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
3506			if (!scn->scn_checkpointing)
3507				zfs_dbgmsg("begin scan checkpoint");
3508
3509			scn->scn_checkpointing = B_TRUE;
3510			scn->scn_clearing = B_TRUE;
3511		} else {
3512			boolean_t should_clear = dsl_scan_should_clear(scn);
3513			if (should_clear && !scn->scn_clearing) {
3514				zfs_dbgmsg("begin scan clearing");
3515				scn->scn_clearing = B_TRUE;
3516			} else if (!should_clear && scn->scn_clearing) {
3517				zfs_dbgmsg("finish scan clearing");
3518				scn->scn_clearing = B_FALSE;
3519			}
3520		}
3521	} else {
3522		ASSERT0(scn->scn_checkpointing);
3523		ASSERT0(scn->scn_clearing);
3524	}
3525
3526	if (!scn->scn_clearing && scn->scn_done_txg == 0) {
3527		/* Need to scan metadata for more blocks to scrub */
3528		dsl_scan_phys_t *scnp = &scn->scn_phys;
3529		taskqid_t prefetch_tqid;
3530		uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
3531		uint64_t nr_leaves = dsl_scan_count_leaves(spa->spa_root_vdev);
3532
3533		/*
3534		 * Calculate the max number of in-flight bytes for pool-wide
3535		 * scanning operations (minimum 1MB). Limits for the issuing
3536		 * phase are done per top-level vdev and are handled separately.
3537		 */
3538		scn->scn_maxinflight_bytes =
3539		    MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
3540
3541		if (scnp->scn_ddt_bookmark.ddb_class <=
3542		    scnp->scn_ddt_class_max) {
3543			ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
3544			zfs_dbgmsg("doing scan sync txg %llu; "
3545			    "ddt bm=%llu/%llu/%llu/%llx",
3546			    (longlong_t)tx->tx_txg,
3547			    (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
3548			    (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
3549			    (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
3550			    (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
3551		} else {
3552			zfs_dbgmsg("doing scan sync txg %llu; "
3553			    "bm=%llu/%llu/%llu/%llu",
3554			    (longlong_t)tx->tx_txg,
3555			    (longlong_t)scnp->scn_bookmark.zb_objset,
3556			    (longlong_t)scnp->scn_bookmark.zb_object,
3557			    (longlong_t)scnp->scn_bookmark.zb_level,
3558			    (longlong_t)scnp->scn_bookmark.zb_blkid);
3559		}
3560
3561		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3562		    NULL, ZIO_FLAG_CANFAIL);
3563
3564		scn->scn_prefetch_stop = B_FALSE;
3565		prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
3566		    dsl_scan_prefetch_thread, scn, TQ_SLEEP);
3567		ASSERT(prefetch_tqid != TASKQID_INVALID);
3568
3569		dsl_pool_config_enter(dp, FTAG);
3570		dsl_scan_visit(scn, tx);
3571		dsl_pool_config_exit(dp, FTAG);
3572
3573		mutex_enter(&dp->dp_spa->spa_scrub_lock);
3574		scn->scn_prefetch_stop = B_TRUE;
3575		cv_broadcast(&spa->spa_scrub_io_cv);
3576		mutex_exit(&dp->dp_spa->spa_scrub_lock);
3577
3578		taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
3579		(void) zio_wait(scn->scn_zio_root);
3580		scn->scn_zio_root = NULL;
3581
3582		zfs_dbgmsg("scan visited %llu blocks in %llums "
3583		    "(%llu os's, %llu holes, %llu < mintxg, "
3584		    "%llu in ddt, %llu > maxtxg)",
3585		    (longlong_t)scn->scn_visited_this_txg,
3586		    (longlong_t)NSEC2MSEC(gethrtime() -
3587		    scn->scn_sync_start_time),
3588		    (longlong_t)scn->scn_objsets_visited_this_txg,
3589		    (longlong_t)scn->scn_holes_this_txg,
3590		    (longlong_t)scn->scn_lt_min_this_txg,
3591		    (longlong_t)scn->scn_ddt_contained_this_txg,
3592		    (longlong_t)scn->scn_gt_max_this_txg);
3593
3594		if (!scn->scn_suspending) {
3595			ASSERT0(avl_numnodes(&scn->scn_queue));
3596			scn->scn_done_txg = tx->tx_txg + 1;
3597			if (scn->scn_is_sorted) {
3598				scn->scn_checkpointing = B_TRUE;
3599				scn->scn_clearing = B_TRUE;
3600			}
3601			zfs_dbgmsg("scan complete txg %llu",
3602			    (longlong_t)tx->tx_txg);
3603		}
3604	} else if (scn->scn_is_sorted && scn->scn_bytes_pending != 0) {
3605		ASSERT(scn->scn_clearing);
3606
3607		/* need to issue scrubbing IOs from per-vdev queues */
3608		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3609		    NULL, ZIO_FLAG_CANFAIL);
3610		scan_io_queues_run(scn);
3611		(void) zio_wait(scn->scn_zio_root);
3612		scn->scn_zio_root = NULL;
3613
3614		/* calculate and dprintf the current memory usage */
3615		(void) dsl_scan_should_clear(scn);
3616		dsl_scan_update_stats(scn);
3617
3618		zfs_dbgmsg("scrubbed %llu blocks (%llu segs) in %llums "
3619		    "(avg_block_size = %llu, avg_seg_size = %llu)",
3620		    (longlong_t)scn->scn_zios_this_txg,
3621		    (longlong_t)scn->scn_segs_this_txg,
3622		    (longlong_t)NSEC2MSEC(gethrtime() -
3623		    scn->scn_sync_start_time),
3624		    (longlong_t)scn->scn_avg_zio_size_this_txg,
3625		    (longlong_t)scn->scn_avg_seg_size_this_txg);
3626	} else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
3627		/* Finished with everything. Mark the scrub as complete */
3628		zfs_dbgmsg("scan issuing complete txg %llu",
3629		    (longlong_t)tx->tx_txg);
3630		ASSERT3U(scn->scn_done_txg, !=, 0);
3631		ASSERT0(spa->spa_scrub_inflight);
3632		ASSERT0(scn->scn_bytes_pending);
3633		dsl_scan_done(scn, B_TRUE, tx);
3634		sync_type = SYNC_MANDATORY;
3635	}
3636
3637	dsl_scan_sync_state(scn, tx, sync_type);
3638}
3639
3640static void
3641count_block(dsl_scan_t *scn, zfs_all_blkstats_t *zab, const blkptr_t *bp)
3642{
3643	int i;
3644
3645	/*
3646	 * Don't count embedded bp's, since we already did the work of
3647	 * scanning these when we scanned the containing block.
3648	 */
3649	if (BP_IS_EMBEDDED(bp))
3650		return;
3651
3652	/*
3653	 * Update the spa's stats on how many bytes we have issued.
3654	 * Sequential scrubs create a zio for each DVA of the bp. Each
3655	 * of these will include all DVAs for repair purposes, but the
3656	 * zio code will only try the first one unless there is an issue.
3657	 * Therefore, we should only count the first DVA for these IOs.
3658	 */
3659	if (scn->scn_is_sorted) {
3660		atomic_add_64(&scn->scn_dp->dp_spa->spa_scan_pass_issued,
3661		    DVA_GET_ASIZE(&bp->blk_dva[0]));
3662	} else {
3663		spa_t *spa = scn->scn_dp->dp_spa;
3664
3665		for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3666			atomic_add_64(&spa->spa_scan_pass_issued,
3667			    DVA_GET_ASIZE(&bp->blk_dva[i]));
3668		}
3669	}
3670
3671	/*
3672	 * If we resume after a reboot, zab will be NULL; don't record
3673	 * incomplete stats in that case.
3674	 */
3675	if (zab == NULL)
3676		return;
3677
3678	mutex_enter(&zab->zab_lock);
3679
3680	for (i = 0; i < 4; i++) {
3681		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
3682		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
3683		if (t & DMU_OT_NEWTYPE)
3684			t = DMU_OT_OTHER;
3685		zfs_blkstat_t *zb = &zab->zab_type[l][t];
3686		int equal;
3687
3688		zb->zb_count++;
3689		zb->zb_asize += BP_GET_ASIZE(bp);
3690		zb->zb_lsize += BP_GET_LSIZE(bp);
3691		zb->zb_psize += BP_GET_PSIZE(bp);
3692		zb->zb_gangs += BP_COUNT_GANG(bp);
3693
3694		switch (BP_GET_NDVAS(bp)) {
3695		case 2:
3696			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3697			    DVA_GET_VDEV(&bp->blk_dva[1]))
3698				zb->zb_ditto_2_of_2_samevdev++;
3699			break;
3700		case 3:
3701			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3702			    DVA_GET_VDEV(&bp->blk_dva[1])) +
3703			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3704			    DVA_GET_VDEV(&bp->blk_dva[2])) +
3705			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
3706			    DVA_GET_VDEV(&bp->blk_dva[2]));
3707			if (equal == 1)
3708				zb->zb_ditto_2_of_3_samevdev++;
3709			else if (equal == 3)
3710				zb->zb_ditto_3_of_3_samevdev++;
3711			break;
3712		}
3713	}
3714
3715	mutex_exit(&zab->zab_lock);
3716}
3717
3718static void
3719scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
3720{
3721	avl_index_t idx;
3722	int64_t asize = SIO_GET_ASIZE(sio);
3723	dsl_scan_t *scn = queue->q_scn;
3724
3725	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3726
3727	if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
3728		/* block is already scheduled for reading */
3729		atomic_add_64(&scn->scn_bytes_pending, -asize);
3730		sio_free(sio);
3731		return;
3732	}
3733	avl_insert(&queue->q_sios_by_addr, sio, idx);
3734	queue->q_sio_memused += SIO_GET_MUSED(sio);
3735	range_tree_add(queue->q_exts_by_addr, SIO_GET_OFFSET(sio), asize);
3736}
3737
3738/*
3739 * Given all the info we got from our metadata scanning process, we
3740 * construct a scan_io_t and insert it into the scan sorting queue. The
3741 * I/O must already be suitable for us to process. This is controlled
3742 * by dsl_scan_enqueue().
3743 */
3744static void
3745scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
3746    int zio_flags, const zbookmark_phys_t *zb)
3747{
3748	dsl_scan_t *scn = queue->q_scn;
3749	scan_io_t *sio = sio_alloc(BP_GET_NDVAS(bp));
3750
3751	ASSERT0(BP_IS_GANG(bp));
3752	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3753
3754	bp2sio(bp, sio, dva_i);
3755	sio->sio_flags = zio_flags;
3756	sio->sio_zb = *zb;
3757
3758	/*
3759	 * Increment the bytes pending counter now so that we can't
3760	 * get an integer underflow in case the worker processes the
3761	 * zio before we get to incrementing this counter.
3762	 */
3763	atomic_add_64(&scn->scn_bytes_pending, SIO_GET_ASIZE(sio));
3764
3765	scan_io_queue_insert_impl(queue, sio);
3766}
3767
3768/*
3769 * Given a set of I/O parameters as discovered by the metadata traversal
3770 * process, attempts to place the I/O into the sorted queues (if allowed),
3771 * or immediately executes the I/O.
3772 */
3773static void
3774dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3775    const zbookmark_phys_t *zb)
3776{
3777	spa_t *spa = dp->dp_spa;
3778
3779	ASSERT(!BP_IS_EMBEDDED(bp));
3780
3781	/*
3782	 * Gang blocks are hard to issue sequentially, so we just issue them
3783	 * here immediately instead of queuing them.
3784	 */
3785	if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
3786		scan_exec_io(dp, bp, zio_flags, zb, NULL);
3787		return;
3788	}
3789	for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3790		dva_t dva;
3791		vdev_t *vdev;
3792
3793		dva = bp->blk_dva[i];
3794		vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
3795		ASSERT(vdev != NULL);
3796
3797		mutex_enter(&vdev->vdev_scan_io_queue_lock);
3798		if (vdev->vdev_scan_io_queue == NULL)
3799			vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
3800		ASSERT(dp->dp_scan != NULL);
3801		scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
3802		    i, zio_flags, zb);
3803		mutex_exit(&vdev->vdev_scan_io_queue_lock);
3804	}
3805}
3806
3807static int
3808dsl_scan_scrub_cb(dsl_pool_t *dp,
3809    const blkptr_t *bp, const zbookmark_phys_t *zb)
3810{
3811	dsl_scan_t *scn = dp->dp_scan;
3812	spa_t *spa = dp->dp_spa;
3813	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
3814	size_t psize = BP_GET_PSIZE(bp);
3815	boolean_t needs_io;
3816	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
3817	int d;
3818
3819	if (phys_birth <= scn->scn_phys.scn_min_txg ||
3820	    phys_birth >= scn->scn_phys.scn_max_txg) {
3821		count_block(scn, dp->dp_blkstats, bp);
3822		return (0);
3823	}
3824
3825	/* Embedded BP's have phys_birth==0, so we reject them above. */
3826	ASSERT(!BP_IS_EMBEDDED(bp));
3827
3828	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
3829	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
3830		zio_flags |= ZIO_FLAG_SCRUB;
3831		needs_io = B_TRUE;
3832	} else {
3833		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
3834		zio_flags |= ZIO_FLAG_RESILVER;
3835		needs_io = B_FALSE;
3836	}
3837
3838	/* If it's an intent log block, failure is expected. */
3839	if (zb->zb_level == ZB_ZIL_LEVEL)
3840		zio_flags |= ZIO_FLAG_SPECULATIVE;
3841
3842	for (d = 0; d < BP_GET_NDVAS(bp); d++) {
3843		const dva_t *dva = &bp->blk_dva[d];
3844
3845		/*
3846		 * Keep track of how much data we've examined so that
3847		 * zpool(1M) status can make useful progress reports.
3848		 */
3849		scn->scn_phys.scn_examined += DVA_GET_ASIZE(dva);
3850		spa->spa_scan_pass_exam += DVA_GET_ASIZE(dva);
3851
3852		/* if it's a resilver, this may not be in the target range */
3853		if (!needs_io)
3854			needs_io = dsl_scan_need_resilver(spa, dva, psize,
3855			    phys_birth);
3856	}
3857
3858	if (needs_io && !zfs_no_scrub_io) {
3859		dsl_scan_enqueue(dp, bp, zio_flags, zb);
3860	} else {
3861		count_block(scn, dp->dp_blkstats, bp);
3862	}
3863
3864	/* do not relocate this block */
3865	return (0);
3866}
3867
3868static void
3869dsl_scan_scrub_done(zio_t *zio)
3870{
3871	spa_t *spa = zio->io_spa;
3872	blkptr_t *bp = zio->io_bp;
3873	dsl_scan_io_queue_t *queue = zio->io_private;
3874
3875	abd_free(zio->io_abd);
3876
3877	if (queue == NULL) {
3878		mutex_enter(&spa->spa_scrub_lock);
3879		ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
3880		spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
3881		cv_broadcast(&spa->spa_scrub_io_cv);
3882		mutex_exit(&spa->spa_scrub_lock);
3883	} else {
3884		mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
3885		ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
3886		queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
3887		cv_broadcast(&queue->q_zio_cv);
3888		mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
3889	}
3890
3891	if (zio->io_error && (zio->io_error != ECKSUM ||
3892	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
3893		atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors);
3894	}
3895}
3896
3897/*
3898 * Given a scanning zio's information, executes the zio. The zio need
3899 * not necessarily be only sortable, this function simply executes the
3900 * zio, no matter what it is. The optional queue argument allows the
3901 * caller to specify that they want per top level vdev IO rate limiting
3902 * instead of the legacy global limiting.
3903 */
3904static void
3905scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3906    const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
3907{
3908	spa_t *spa = dp->dp_spa;
3909	dsl_scan_t *scn = dp->dp_scan;
3910	size_t size = BP_GET_PSIZE(bp);
3911	abd_t *data = abd_alloc_for_io(size, B_FALSE);
3912
3913	if (queue == NULL) {
3914		mutex_enter(&spa->spa_scrub_lock);
3915		while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
3916			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
3917		spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
3918		mutex_exit(&spa->spa_scrub_lock);
3919	} else {
3920		kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3921
3922		mutex_enter(q_lock);
3923		while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
3924			cv_wait(&queue->q_zio_cv, q_lock);
3925		queue->q_inflight_bytes += BP_GET_PSIZE(bp);
3926		mutex_exit(q_lock);
3927	}
3928
3929	count_block(dp->dp_scan, dp->dp_blkstats, bp);
3930	zio_nowait(zio_read(dp->dp_scan->scn_zio_root, spa, bp, data, size,
3931	    dsl_scan_scrub_done, queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
3932}
3933
3934/*
3935 * This is the primary extent sorting algorithm. We balance two parameters:
3936 * 1) how many bytes of I/O are in an extent
3937 * 2) how well the extent is filled with I/O (as a fraction of its total size)
3938 * Since we allow extents to have gaps between their constituent I/Os, it's
3939 * possible to have a fairly large extent that contains the same amount of
3940 * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
3941 * The algorithm sorts based on a score calculated from the extent's size,
3942 * the relative fill volume (in %) and a "fill weight" parameter that controls
3943 * the split between whether we prefer larger extents or more well populated
3944 * extents:
3945 *
3946 * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
3947 *
3948 * Example:
3949 * 1) assume extsz = 64 MiB
3950 * 2) assume fill = 32 MiB (extent is half full)
3951 * 3) assume fill_weight = 3
3952 * 4)	SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
3953 *	SCORE = 32M + (50 * 3 * 32M) / 100
3954 *	SCORE = 32M + (4800M / 100)
3955 *	SCORE = 32M + 48M
3956 *		^	^
3957 *		|	+--- final total relative fill-based score
3958 *		+--------- final total fill-based score
3959 *	SCORE = 80M
3960 *
3961 * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
3962 * extents that are more completely filled (in a 3:2 ratio) vs just larger.
3963 * Note that as an optimization, we replace multiplication and division by
3964 * 100 with bitshifting by 7 (which effecitvely multiplies and divides by 128).
3965 */
3966static int
3967ext_size_compare(const void *x, const void *y)
3968{
3969	const range_seg_gap_t *rsa = x, *rsb = y;
3970
3971	uint64_t sa = rsa->rs_end - rsa->rs_start;
3972	uint64_t sb = rsb->rs_end - rsb->rs_start;
3973	uint64_t score_a, score_b;
3974
3975	score_a = rsa->rs_fill + ((((rsa->rs_fill << 7) / sa) *
3976	    fill_weight * rsa->rs_fill) >> 7);
3977	score_b = rsb->rs_fill + ((((rsb->rs_fill << 7) / sb) *
3978	    fill_weight * rsb->rs_fill) >> 7);
3979
3980	if (score_a > score_b)
3981		return (-1);
3982	if (score_a == score_b) {
3983		if (rsa->rs_start < rsb->rs_start)
3984			return (-1);
3985		if (rsa->rs_start == rsb->rs_start)
3986			return (0);
3987		return (1);
3988	}
3989	return (1);
3990}
3991
3992/*
3993 * Comparator for the q_sios_by_addr tree. Sorting is simply performed
3994 * based on LBA-order (from lowest to highest).
3995 */
3996static int
3997sio_addr_compare(const void *x, const void *y)
3998{
3999	const scan_io_t *a = x, *b = y;
4000
4001	return (TREE_CMP(SIO_GET_OFFSET(a), SIO_GET_OFFSET(b)));
4002}
4003
4004/* IO queues are created on demand when they are needed. */
4005static dsl_scan_io_queue_t *
4006scan_io_queue_create(vdev_t *vd)
4007{
4008	dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
4009	dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
4010
4011	q->q_scn = scn;
4012	q->q_vd = vd;
4013	q->q_sio_memused = 0;
4014	cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
4015	q->q_exts_by_addr = range_tree_create_impl(&rt_btree_ops, RANGE_SEG_GAP,
4016	    &q->q_exts_by_size, 0, 0, ext_size_compare, zfs_scan_max_ext_gap);
4017	avl_create(&q->q_sios_by_addr, sio_addr_compare,
4018	    sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
4019
4020	return (q);
4021}
4022
4023/*
4024 * Destroys a scan queue and all segments and scan_io_t's contained in it.
4025 * No further execution of I/O occurs, anything pending in the queue is
4026 * simply freed without being executed.
4027 */
4028void
4029dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
4030{
4031	dsl_scan_t *scn = queue->q_scn;
4032	scan_io_t *sio;
4033	void *cookie = NULL;
4034	int64_t bytes_dequeued = 0;
4035
4036	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
4037
4038	while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
4039	    NULL) {
4040		ASSERT(range_tree_contains(queue->q_exts_by_addr,
4041		    SIO_GET_OFFSET(sio), SIO_GET_ASIZE(sio)));
4042		bytes_dequeued += SIO_GET_ASIZE(sio);
4043		queue->q_sio_memused -= SIO_GET_MUSED(sio);
4044		sio_free(sio);
4045	}
4046
4047	ASSERT0(queue->q_sio_memused);
4048	atomic_add_64(&scn->scn_bytes_pending, -bytes_dequeued);
4049	range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
4050	range_tree_destroy(queue->q_exts_by_addr);
4051	avl_destroy(&queue->q_sios_by_addr);
4052	cv_destroy(&queue->q_zio_cv);
4053
4054	kmem_free(queue, sizeof (*queue));
4055}
4056
4057/*
4058 * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
4059 * called on behalf of vdev_top_transfer when creating or destroying
4060 * a mirror vdev due to zpool attach/detach.
4061 */
4062void
4063dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
4064{
4065	mutex_enter(&svd->vdev_scan_io_queue_lock);
4066	mutex_enter(&tvd->vdev_scan_io_queue_lock);
4067
4068	VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
4069	tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
4070	svd->vdev_scan_io_queue = NULL;
4071	if (tvd->vdev_scan_io_queue != NULL)
4072		tvd->vdev_scan_io_queue->q_vd = tvd;
4073
4074	mutex_exit(&tvd->vdev_scan_io_queue_lock);
4075	mutex_exit(&svd->vdev_scan_io_queue_lock);
4076}
4077
4078static void
4079scan_io_queues_destroy(dsl_scan_t *scn)
4080{
4081	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
4082
4083	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
4084		vdev_t *tvd = rvd->vdev_child[i];
4085
4086		mutex_enter(&tvd->vdev_scan_io_queue_lock);
4087		if (tvd->vdev_scan_io_queue != NULL)
4088			dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
4089		tvd->vdev_scan_io_queue = NULL;
4090		mutex_exit(&tvd->vdev_scan_io_queue_lock);
4091	}
4092}
4093
4094static void
4095dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
4096{
4097	dsl_pool_t *dp = spa->spa_dsl_pool;
4098	dsl_scan_t *scn = dp->dp_scan;
4099	vdev_t *vdev;
4100	kmutex_t *q_lock;
4101	dsl_scan_io_queue_t *queue;
4102	scan_io_t *srch_sio, *sio;
4103	avl_index_t idx;
4104	uint64_t start, size;
4105
4106	vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
4107	ASSERT(vdev != NULL);
4108	q_lock = &vdev->vdev_scan_io_queue_lock;
4109	queue = vdev->vdev_scan_io_queue;
4110
4111	mutex_enter(q_lock);
4112	if (queue == NULL) {
4113		mutex_exit(q_lock);
4114		return;
4115	}
4116
4117	srch_sio = sio_alloc(BP_GET_NDVAS(bp));
4118	bp2sio(bp, srch_sio, dva_i);
4119	start = SIO_GET_OFFSET(srch_sio);
4120	size = SIO_GET_ASIZE(srch_sio);
4121
4122	/*
4123	 * We can find the zio in two states:
4124	 * 1) Cold, just sitting in the queue of zio's to be issued at
4125	 *	some point in the future. In this case, all we do is
4126	 *	remove the zio from the q_sios_by_addr tree, decrement
4127	 *	its data volume from the containing range_seg_t and
4128	 *	resort the q_exts_by_size tree to reflect that the
4129	 *	range_seg_t has lost some of its 'fill'. We don't shorten
4130	 *	the range_seg_t - this is usually rare enough not to be
4131	 *	worth the extra hassle of trying keep track of precise
4132	 *	extent boundaries.
4133	 * 2) Hot, where the zio is currently in-flight in
4134	 *	dsl_scan_issue_ios. In this case, we can't simply
4135	 *	reach in and stop the in-flight zio's, so we instead
4136	 *	block the caller. Eventually, dsl_scan_issue_ios will
4137	 *	be done with issuing the zio's it gathered and will
4138	 *	signal us.
4139	 */
4140	sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
4141	sio_free(srch_sio);
4142
4143	if (sio != NULL) {
4144		int64_t asize = SIO_GET_ASIZE(sio);
4145		blkptr_t tmpbp;
4146
4147		/* Got it while it was cold in the queue */
4148		ASSERT3U(start, ==, SIO_GET_OFFSET(sio));
4149		ASSERT3U(size, ==, asize);
4150		avl_remove(&queue->q_sios_by_addr, sio);
4151		queue->q_sio_memused -= SIO_GET_MUSED(sio);
4152
4153		ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
4154		range_tree_remove_fill(queue->q_exts_by_addr, start, size);
4155
4156		/*
4157		 * We only update scn_bytes_pending in the cold path,
4158		 * otherwise it will already have been accounted for as
4159		 * part of the zio's execution.
4160		 */
4161		atomic_add_64(&scn->scn_bytes_pending, -asize);
4162
4163		/* count the block as though we issued it */
4164		sio2bp(sio, &tmpbp);
4165		count_block(scn, dp->dp_blkstats, &tmpbp);
4166
4167		sio_free(sio);
4168	}
4169	mutex_exit(q_lock);
4170}
4171
4172/*
4173 * Callback invoked when a zio_free() zio is executing. This needs to be
4174 * intercepted to prevent the zio from deallocating a particular portion
4175 * of disk space and it then getting reallocated and written to, while we
4176 * still have it queued up for processing.
4177 */
4178void
4179dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
4180{
4181	dsl_pool_t *dp = spa->spa_dsl_pool;
4182	dsl_scan_t *scn = dp->dp_scan;
4183
4184	ASSERT(!BP_IS_EMBEDDED(bp));
4185	ASSERT(scn != NULL);
4186	if (!dsl_scan_is_running(scn))
4187		return;
4188
4189	for (int i = 0; i < BP_GET_NDVAS(bp); i++)
4190		dsl_scan_freed_dva(spa, bp, i);
4191}
4192