xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_scan.c (revision bbf21555)
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, 2019, Datto Inc. All rights reserved.
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 
115 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
116     const zbookmark_phys_t *);
117 
118 static scan_cb_t dsl_scan_scrub_cb;
119 
120 static int scan_ds_queue_compare(const void *a, const void *b);
121 static int scan_prefetch_queue_compare(const void *a, const void *b);
122 static void scan_ds_queue_clear(dsl_scan_t *scn);
123 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
124     uint64_t *txg);
125 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
126 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
127 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
128 
129 extern 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  */
137 int 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  */
144 int zfs_top_maxinflight = 32;		/* maximum I/Os per top-level */
145 unsigned int zfs_resilver_delay = 2;	/* number of ticks to delay resilver */
146 unsigned int zfs_scrub_delay = 4;	/* number of ticks to delay scrub */
147 unsigned 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  */
157 unsigned long zfs_scan_vdev_limit = 4 << 20;
158 
159 int zfs_scan_issue_strategy = 0;
160 int zfs_scan_legacy = B_FALSE;	/* don't queue & sort zios, go direct */
161 uint64_t zfs_scan_max_ext_gap = 2 << 20;	/* in bytes */
162 
163 unsigned 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  */
171 uint64_t zfs_scan_fill_weight = 3;
172 static uint64_t fill_weight;
173 
174 /* See dsl_scan_should_clear() for details on the memory limit tunables */
175 uint64_t zfs_scan_mem_lim_min = 16 << 20;	/* bytes */
176 uint64_t zfs_scan_mem_lim_soft_max = 128 << 20;	/* bytes */
177 int zfs_scan_mem_lim_fact = 20;		/* fraction of physmem */
178 int zfs_scan_mem_lim_soft_fact = 20;	/* fraction of mem lim above */
179 
180 unsigned int zfs_scrub_min_time_ms = 1000; /* min millisecs to scrub per txg */
181 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
182 /* min millisecs to obsolete per txg */
183 unsigned int zfs_obsolete_min_time_ms = 500;
184 /* min millisecs to resilver per txg */
185 unsigned int zfs_resilver_min_time_ms = 3000;
186 int zfs_scan_suspend_progress = 0; /* set to prevent scans from progressing */
187 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
188 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
189 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
190 /* max number of blocks to free in a single TXG */
191 uint64_t zfs_async_block_max_blocks = UINT64_MAX;
192 
193 int 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 
212 extern int zfs_txg_timeout;
213 
214 /*
215  * Enable/disable the processing of the free_bpobj object.
216  */
217 boolean_t zfs_free_bpobj_enabled = B_TRUE;
218 
219 /* the order has to match pool_scan_type */
220 static 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 */
227 typedef 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  */
241 typedef 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  */
253 typedef 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 
292 struct 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() */
315 typedef 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() */
324 typedef 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 
331 static 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);
333 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
334     scan_io_t *sio);
335 
336 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
337 static void scan_io_queues_destroy(dsl_scan_t *scn);
338 
339 static 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 */
342 static void
sio_free(scan_io_t * sio)343 sio_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 */
352 static scan_io_t *
sio_alloc(unsigned short nr_dvas)353 sio_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 
361 void
scan_init(void)362 scan_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 
384 void
scan_fini(void)385 scan_fini(void)
386 {
387 	for (int i = 0; i < SPA_DVAS_PER_BP; i++) {
388 		kmem_cache_destroy(sio_cache[i]);
389 	}
390 }
391 
392 static inline boolean_t
dsl_scan_is_running(const dsl_scan_t * scn)393 dsl_scan_is_running(const dsl_scan_t *scn)
394 {
395 	return (scn->scn_phys.scn_state == DSS_SCANNING);
396 }
397 
398 boolean_t
dsl_scan_resilvering(dsl_pool_t * dp)399 dsl_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 
405 static inline void
sio2bp(const scan_io_t * sio,blkptr_t * bp)406 sio2bp(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 
421 static inline void
bp2sio(const blkptr_t * bp,scan_io_t * sio,int dva_i)422 bp2sio(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 
441 int
dsl_scan_init(dsl_pool_t * dp,uint64_t txg)442 dsl_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 		} else if (dsl_scan_resilvering(dp)) {
553 			/*
554 			 * If a resilver is in progress and there are already
555 			 * errors, restart it instead of finishing this scan and
556 			 * then restarting it. If there haven't been any errors
557 			 * then remember that the incore DTL is valid.
558 			 */
559 			if (scn->scn_phys.scn_errors > 0) {
560 				scn->scn_restart_txg = txg;
561 				zfs_dbgmsg("resilver can't excise DTL_MISSING "
562 				    "when finished; restarting in txg %llu",
563 				    (u_longlong_t)scn->scn_restart_txg);
564 			} else {
565 				/* it's safe to excise DTL when finished */
566 				spa->spa_scrub_started = B_TRUE;
567 			}
568 		}
569 	}
570 
571 	bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
572 
573 	/* reload the queue into the in-core state */
574 	if (scn->scn_phys.scn_queue_obj != 0) {
575 		zap_cursor_t zc;
576 		zap_attribute_t za;
577 
578 		for (zap_cursor_init(&zc, dp->dp_meta_objset,
579 		    scn->scn_phys.scn_queue_obj);
580 		    zap_cursor_retrieve(&zc, &za) == 0;
581 		    (void) zap_cursor_advance(&zc)) {
582 			scan_ds_queue_insert(scn,
583 			    zfs_strtonum(za.za_name, NULL),
584 			    za.za_first_integer);
585 		}
586 		zap_cursor_fini(&zc);
587 	}
588 
589 	spa_scan_stat_init(spa);
590 	return (0);
591 }
592 
593 void
dsl_scan_fini(dsl_pool_t * dp)594 dsl_scan_fini(dsl_pool_t *dp)
595 {
596 	if (dp->dp_scan != NULL) {
597 		dsl_scan_t *scn = dp->dp_scan;
598 
599 		if (scn->scn_taskq != NULL)
600 			taskq_destroy(scn->scn_taskq);
601 		scan_ds_queue_clear(scn);
602 		avl_destroy(&scn->scn_queue);
603 		avl_destroy(&scn->scn_prefetch_queue);
604 
605 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
606 		dp->dp_scan = NULL;
607 	}
608 }
609 
610 static boolean_t
dsl_scan_restarting(dsl_scan_t * scn,dmu_tx_t * tx)611 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
612 {
613 	return (scn->scn_restart_txg != 0 &&
614 	    scn->scn_restart_txg <= tx->tx_txg);
615 }
616 
617 boolean_t
dsl_scan_resilver_scheduled(dsl_pool_t * dp)618 dsl_scan_resilver_scheduled(dsl_pool_t *dp)
619 {
620 	return ((dp->dp_scan && dp->dp_scan->scn_restart_txg != 0) ||
621 	    (spa_async_tasks(dp->dp_spa) & SPA_ASYNC_RESILVER));
622 }
623 
624 boolean_t
dsl_scan_scrubbing(const dsl_pool_t * dp)625 dsl_scan_scrubbing(const dsl_pool_t *dp)
626 {
627 	dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
628 
629 	return (scn_phys->scn_state == DSS_SCANNING &&
630 	    scn_phys->scn_func == POOL_SCAN_SCRUB);
631 }
632 
633 boolean_t
dsl_scan_is_paused_scrub(const dsl_scan_t * scn)634 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
635 {
636 	return (dsl_scan_scrubbing(scn->scn_dp) &&
637 	    scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
638 }
639 
640 /*
641  * Writes out a persistent dsl_scan_phys_t record to the pool directory.
642  * Because we can be running in the block sorting algorithm, we do not always
643  * want to write out the record, only when it is "safe" to do so. This safety
644  * condition is achieved by making sure that the sorting queues are empty
645  * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
646  * is inconsistent with how much actual scanning progress has been made. The
647  * kind of sync to be performed is specified by the sync_type argument. If the
648  * sync is optional, we only sync if the queues are empty. If the sync is
649  * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
650  * third possible state is a "cached" sync. This is done in response to:
651  * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
652  *	destroyed, so we wouldn't be able to restart scanning from it.
653  * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
654  *	superseded by a newer snapshot.
655  * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
656  *	swapped with its clone.
657  * In all cases, a cached sync simply rewrites the last record we've written,
658  * just slightly modified. For the modifications that are performed to the
659  * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
660  * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
661  */
662 static void
dsl_scan_sync_state(dsl_scan_t * scn,dmu_tx_t * tx,state_sync_type_t sync_type)663 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
664 {
665 	int i;
666 	spa_t *spa = scn->scn_dp->dp_spa;
667 
668 	ASSERT(sync_type != SYNC_MANDATORY || scn->scn_bytes_pending == 0);
669 	if (scn->scn_bytes_pending == 0) {
670 		for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
671 			vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
672 			dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
673 
674 			if (q == NULL)
675 				continue;
676 
677 			mutex_enter(&vd->vdev_scan_io_queue_lock);
678 			ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
679 			ASSERT3P(zfs_btree_first(&q->q_exts_by_size, NULL), ==,
680 			    NULL);
681 			ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
682 			mutex_exit(&vd->vdev_scan_io_queue_lock);
683 		}
684 
685 		if (scn->scn_phys.scn_queue_obj != 0)
686 			scan_ds_queue_sync(scn, tx);
687 		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
688 		    DMU_POOL_DIRECTORY_OBJECT,
689 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
690 		    &scn->scn_phys, tx));
691 		bcopy(&scn->scn_phys, &scn->scn_phys_cached,
692 		    sizeof (scn->scn_phys));
693 
694 		if (scn->scn_checkpointing)
695 			zfs_dbgmsg("finish scan checkpoint");
696 
697 		scn->scn_checkpointing = B_FALSE;
698 		scn->scn_last_checkpoint = ddi_get_lbolt();
699 	} else if (sync_type == SYNC_CACHED) {
700 		VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
701 		    DMU_POOL_DIRECTORY_OBJECT,
702 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
703 		    &scn->scn_phys_cached, tx));
704 	}
705 }
706 
707 /* ARGSUSED */
708 static int
dsl_scan_setup_check(void * arg,dmu_tx_t * tx)709 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
710 {
711 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
712 
713 	if (dsl_scan_is_running(scn))
714 		return (SET_ERROR(EBUSY));
715 
716 	return (0);
717 }
718 
719 static void
dsl_scan_setup_sync(void * arg,dmu_tx_t * tx)720 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
721 {
722 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
723 	pool_scan_func_t *funcp = arg;
724 	dmu_object_type_t ot = 0;
725 	dsl_pool_t *dp = scn->scn_dp;
726 	spa_t *spa = dp->dp_spa;
727 
728 	ASSERT(!dsl_scan_is_running(scn));
729 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
730 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
731 	scn->scn_phys.scn_func = *funcp;
732 	scn->scn_phys.scn_state = DSS_SCANNING;
733 	scn->scn_phys.scn_min_txg = 0;
734 	scn->scn_phys.scn_max_txg = tx->tx_txg;
735 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
736 	scn->scn_phys.scn_start_time = gethrestime_sec();
737 	scn->scn_phys.scn_errors = 0;
738 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
739 	scn->scn_issued_before_pass = 0;
740 	scn->scn_restart_txg = 0;
741 	scn->scn_done_txg = 0;
742 	scn->scn_last_checkpoint = 0;
743 	scn->scn_checkpointing = B_FALSE;
744 	spa_scan_stat_init(spa);
745 
746 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
747 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
748 
749 		/* rewrite all disk labels */
750 		vdev_config_dirty(spa->spa_root_vdev);
751 
752 		if (vdev_resilver_needed(spa->spa_root_vdev,
753 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
754 			spa_event_notify(spa, NULL, NULL,
755 			    ESC_ZFS_RESILVER_START);
756 		} else {
757 			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
758 		}
759 
760 		spa->spa_scrub_started = B_TRUE;
761 		/*
762 		 * If this is an incremental scrub, limit the DDT scrub phase
763 		 * to just the auto-ditto class (for correctness); the rest
764 		 * of the scrub should go faster using top-down pruning.
765 		 */
766 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
767 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
768 
769 	}
770 
771 	/* back to the generic stuff */
772 
773 	if (dp->dp_blkstats == NULL) {
774 		dp->dp_blkstats =
775 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
776 		mutex_init(&dp->dp_blkstats->zab_lock, NULL,
777 		    MUTEX_DEFAULT, NULL);
778 	}
779 	bzero(&dp->dp_blkstats->zab_type, sizeof (dp->dp_blkstats->zab_type));
780 
781 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
782 		ot = DMU_OT_ZAP_OTHER;
783 
784 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
785 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
786 
787 	bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
788 
789 	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
790 
791 	spa_history_log_internal(spa, "scan setup", tx,
792 	    "func=%u mintxg=%llu maxtxg=%llu",
793 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
794 }
795 
796 /*
797  * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
798  * Can also be called to resume a paused scrub.
799  */
800 int
dsl_scan(dsl_pool_t * dp,pool_scan_func_t func)801 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
802 {
803 	spa_t *spa = dp->dp_spa;
804 	dsl_scan_t *scn = dp->dp_scan;
805 
806 	/*
807 	 * Purge all vdev caches and probe all devices.  We do this here
808 	 * rather than in sync context because this requires a writer lock
809 	 * on the spa_config lock, which we can't do from sync context.  The
810 	 * spa_scrub_reopen flag indicates that vdev_open() should not
811 	 * attempt to start another scrub.
812 	 */
813 	spa_vdev_state_enter(spa, SCL_NONE);
814 	spa->spa_scrub_reopen = B_TRUE;
815 	vdev_reopen(spa->spa_root_vdev);
816 	spa->spa_scrub_reopen = B_FALSE;
817 	(void) spa_vdev_state_exit(spa, NULL, 0);
818 
819 	if (func == POOL_SCAN_RESILVER) {
820 		dsl_scan_restart_resilver(spa->spa_dsl_pool, 0);
821 		return (0);
822 	}
823 
824 	if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
825 		/* got scrub start cmd, resume paused scrub */
826 		int err = dsl_scrub_set_pause_resume(scn->scn_dp,
827 		    POOL_SCRUB_NORMAL);
828 		if (err == 0) {
829 			spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
830 			return (ECANCELED);
831 		}
832 		return (SET_ERROR(err));
833 	}
834 
835 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
836 	    dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
837 }
838 
839 /* ARGSUSED */
840 static void
dsl_scan_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)841 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
842 {
843 	static const char *old_names[] = {
844 		"scrub_bookmark",
845 		"scrub_ddt_bookmark",
846 		"scrub_ddt_class_max",
847 		"scrub_queue",
848 		"scrub_min_txg",
849 		"scrub_max_txg",
850 		"scrub_func",
851 		"scrub_errors",
852 		NULL
853 	};
854 
855 	dsl_pool_t *dp = scn->scn_dp;
856 	spa_t *spa = dp->dp_spa;
857 	int i;
858 
859 	/* Remove any remnants of an old-style scrub. */
860 	for (i = 0; old_names[i]; i++) {
861 		(void) zap_remove(dp->dp_meta_objset,
862 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
863 	}
864 
865 	if (scn->scn_phys.scn_queue_obj != 0) {
866 		VERIFY0(dmu_object_free(dp->dp_meta_objset,
867 		    scn->scn_phys.scn_queue_obj, tx));
868 		scn->scn_phys.scn_queue_obj = 0;
869 	}
870 	scan_ds_queue_clear(scn);
871 
872 	scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
873 
874 	/*
875 	 * If we were "restarted" from a stopped state, don't bother
876 	 * with anything else.
877 	 */
878 	if (!dsl_scan_is_running(scn)) {
879 		ASSERT(!scn->scn_is_sorted);
880 		return;
881 	}
882 
883 	if (scn->scn_is_sorted) {
884 		scan_io_queues_destroy(scn);
885 		scn->scn_is_sorted = B_FALSE;
886 
887 		if (scn->scn_taskq != NULL) {
888 			taskq_destroy(scn->scn_taskq);
889 			scn->scn_taskq = NULL;
890 		}
891 	}
892 
893 	scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
894 
895 	if (dsl_scan_restarting(scn, tx))
896 		spa_history_log_internal(spa, "scan aborted, restarting", tx,
897 		    "errors=%llu", spa_get_errlog_size(spa));
898 	else if (!complete)
899 		spa_history_log_internal(spa, "scan cancelled", tx,
900 		    "errors=%llu", spa_get_errlog_size(spa));
901 	else
902 		spa_history_log_internal(spa, "scan done", tx,
903 		    "errors=%llu", spa_get_errlog_size(spa));
904 
905 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
906 		spa->spa_scrub_active = B_FALSE;
907 
908 		/*
909 		 * If the scrub/resilver completed, update all DTLs to
910 		 * reflect this.  Whether it succeeded or not, vacate
911 		 * all temporary scrub DTLs.
912 		 *
913 		 * As the scrub does not currently support traversing
914 		 * data that have been freed but are part of a checkpoint,
915 		 * we don't mark the scrub as done in the DTLs as faults
916 		 * may still exist in those vdevs.
917 		 */
918 		if (complete &&
919 		    !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
920 			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
921 			    scn->scn_phys.scn_max_txg, B_TRUE);
922 
923 			spa_event_notify(spa, NULL, NULL,
924 			    scn->scn_phys.scn_min_txg ?
925 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
926 		} else {
927 			vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
928 			    0, B_TRUE);
929 		}
930 		spa_errlog_rotate(spa);
931 
932 		/*
933 		 * Don't clear flag until after vdev_dtl_reassess to ensure that
934 		 * DTL_MISSING will get updated when possible.
935 		 */
936 		spa->spa_scrub_started = B_FALSE;
937 
938 		/*
939 		 * We may have finished replacing a device.
940 		 * Let the async thread assess this and handle the detach.
941 		 */
942 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
943 
944 		/*
945 		 * Clear any resilver_deferred flags in the config.
946 		 * If there are drives that need resilvering, kick
947 		 * off an asynchronous request to start resilver.
948 		 * vdev_clear_resilver_deferred() may update the config
949 		 * before the resilver can restart. In the event of
950 		 * a crash during this period, the spa loading code
951 		 * will find the drives that need to be resilvered
952 		 * and start the resilver then.
953 		 */
954 		if (spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER) &&
955 		    vdev_clear_resilver_deferred(spa->spa_root_vdev, tx)) {
956 			spa_history_log_internal(spa,
957 			    "starting deferred resilver", tx, "errors=%llu",
958 			    (u_longlong_t)spa_get_errlog_size(spa));
959 			spa_async_request(spa, SPA_ASYNC_RESILVER);
960 		}
961 	}
962 
963 	scn->scn_phys.scn_end_time = gethrestime_sec();
964 
965 	ASSERT(!dsl_scan_is_running(scn));
966 }
967 
968 /* ARGSUSED */
969 static int
dsl_scan_cancel_check(void * arg,dmu_tx_t * tx)970 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
971 {
972 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
973 
974 	if (!dsl_scan_is_running(scn))
975 		return (SET_ERROR(ENOENT));
976 	return (0);
977 }
978 
979 /* ARGSUSED */
980 static void
dsl_scan_cancel_sync(void * arg,dmu_tx_t * tx)981 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
982 {
983 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
984 
985 	dsl_scan_done(scn, B_FALSE, tx);
986 	dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
987 	spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
988 }
989 
990 int
dsl_scan_cancel(dsl_pool_t * dp)991 dsl_scan_cancel(dsl_pool_t *dp)
992 {
993 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
994 	    dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
995 }
996 
997 static int
dsl_scrub_pause_resume_check(void * arg,dmu_tx_t * tx)998 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
999 {
1000 	pool_scrub_cmd_t *cmd = arg;
1001 	dsl_pool_t *dp = dmu_tx_pool(tx);
1002 	dsl_scan_t *scn = dp->dp_scan;
1003 
1004 	if (*cmd == POOL_SCRUB_PAUSE) {
1005 		/* can't pause a scrub when there is no in-progress scrub */
1006 		if (!dsl_scan_scrubbing(dp))
1007 			return (SET_ERROR(ENOENT));
1008 
1009 		/* can't pause a paused scrub */
1010 		if (dsl_scan_is_paused_scrub(scn))
1011 			return (SET_ERROR(EBUSY));
1012 	} else if (*cmd != POOL_SCRUB_NORMAL) {
1013 		return (SET_ERROR(ENOTSUP));
1014 	}
1015 
1016 	return (0);
1017 }
1018 
1019 static void
dsl_scrub_pause_resume_sync(void * arg,dmu_tx_t * tx)1020 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
1021 {
1022 	pool_scrub_cmd_t *cmd = arg;
1023 	dsl_pool_t *dp = dmu_tx_pool(tx);
1024 	spa_t *spa = dp->dp_spa;
1025 	dsl_scan_t *scn = dp->dp_scan;
1026 
1027 	if (*cmd == POOL_SCRUB_PAUSE) {
1028 		/* can't pause a scrub when there is no in-progress scrub */
1029 		spa->spa_scan_pass_scrub_pause = gethrestime_sec();
1030 		scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
1031 		scn->scn_phys_cached.scn_flags |= DSF_SCRUB_PAUSED;
1032 		dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1033 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
1034 	} else {
1035 		ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
1036 		if (dsl_scan_is_paused_scrub(scn)) {
1037 			/*
1038 			 * We need to keep track of how much time we spend
1039 			 * paused per pass so that we can adjust the scrub rate
1040 			 * shown in the output of 'zpool status'
1041 			 */
1042 			spa->spa_scan_pass_scrub_spent_paused +=
1043 			    gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
1044 			spa->spa_scan_pass_scrub_pause = 0;
1045 			scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
1046 			scn->scn_phys_cached.scn_flags &= ~DSF_SCRUB_PAUSED;
1047 			dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1048 		}
1049 	}
1050 }
1051 
1052 /*
1053  * Set scrub pause/resume state if it makes sense to do so
1054  */
1055 int
dsl_scrub_set_pause_resume(const dsl_pool_t * dp,pool_scrub_cmd_t cmd)1056 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
1057 {
1058 	return (dsl_sync_task(spa_name(dp->dp_spa),
1059 	    dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
1060 	    ZFS_SPACE_CHECK_RESERVED));
1061 }
1062 
1063 
1064 /* start a new scan, or restart an existing one. */
1065 void
dsl_scan_restart_resilver(dsl_pool_t * dp,uint64_t txg)1066 dsl_scan_restart_resilver(dsl_pool_t *dp, uint64_t txg)
1067 {
1068 	if (txg == 0) {
1069 		dmu_tx_t *tx;
1070 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1071 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1072 
1073 		txg = dmu_tx_get_txg(tx);
1074 		dp->dp_scan->scn_restart_txg = txg;
1075 		dmu_tx_commit(tx);
1076 	} else {
1077 		dp->dp_scan->scn_restart_txg = txg;
1078 	}
1079 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1080 }
1081 
1082 void
dsl_free(dsl_pool_t * dp,uint64_t txg,const blkptr_t * bp)1083 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
1084 {
1085 	zio_free(dp->dp_spa, txg, bp);
1086 }
1087 
1088 void
dsl_free_sync(zio_t * pio,dsl_pool_t * dp,uint64_t txg,const blkptr_t * bpp)1089 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
1090 {
1091 	ASSERT(dsl_pool_sync_context(dp));
1092 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
1093 }
1094 
1095 static int
scan_ds_queue_compare(const void * a,const void * b)1096 scan_ds_queue_compare(const void *a, const void *b)
1097 {
1098 	const scan_ds_t *sds_a = a, *sds_b = b;
1099 
1100 	if (sds_a->sds_dsobj < sds_b->sds_dsobj)
1101 		return (-1);
1102 	if (sds_a->sds_dsobj == sds_b->sds_dsobj)
1103 		return (0);
1104 	return (1);
1105 }
1106 
1107 static void
scan_ds_queue_clear(dsl_scan_t * scn)1108 scan_ds_queue_clear(dsl_scan_t *scn)
1109 {
1110 	void *cookie = NULL;
1111 	scan_ds_t *sds;
1112 	while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1113 		kmem_free(sds, sizeof (*sds));
1114 	}
1115 }
1116 
1117 static boolean_t
scan_ds_queue_contains(dsl_scan_t * scn,uint64_t dsobj,uint64_t * txg)1118 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1119 {
1120 	scan_ds_t srch, *sds;
1121 
1122 	srch.sds_dsobj = dsobj;
1123 	sds = avl_find(&scn->scn_queue, &srch, NULL);
1124 	if (sds != NULL && txg != NULL)
1125 		*txg = sds->sds_txg;
1126 	return (sds != NULL);
1127 }
1128 
1129 static void
scan_ds_queue_insert(dsl_scan_t * scn,uint64_t dsobj,uint64_t txg)1130 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1131 {
1132 	scan_ds_t *sds;
1133 	avl_index_t where;
1134 
1135 	sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1136 	sds->sds_dsobj = dsobj;
1137 	sds->sds_txg = txg;
1138 
1139 	VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1140 	avl_insert(&scn->scn_queue, sds, where);
1141 }
1142 
1143 static void
scan_ds_queue_remove(dsl_scan_t * scn,uint64_t dsobj)1144 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1145 {
1146 	scan_ds_t srch, *sds;
1147 
1148 	srch.sds_dsobj = dsobj;
1149 
1150 	sds = avl_find(&scn->scn_queue, &srch, NULL);
1151 	VERIFY(sds != NULL);
1152 	avl_remove(&scn->scn_queue, sds);
1153 	kmem_free(sds, sizeof (*sds));
1154 }
1155 
1156 static void
scan_ds_queue_sync(dsl_scan_t * scn,dmu_tx_t * tx)1157 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1158 {
1159 	dsl_pool_t *dp = scn->scn_dp;
1160 	spa_t *spa = dp->dp_spa;
1161 	dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1162 	    DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1163 
1164 	ASSERT0(scn->scn_bytes_pending);
1165 	ASSERT(scn->scn_phys.scn_queue_obj != 0);
1166 
1167 	VERIFY0(dmu_object_free(dp->dp_meta_objset,
1168 	    scn->scn_phys.scn_queue_obj, tx));
1169 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1170 	    DMU_OT_NONE, 0, tx);
1171 	for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1172 	    sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1173 		VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1174 		    scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1175 		    sds->sds_txg, tx));
1176 	}
1177 }
1178 
1179 /*
1180  * Computes the memory limit state that we're currently in. A sorted scan
1181  * needs quite a bit of memory to hold the sorting queue, so we need to
1182  * reasonably constrain the size so it doesn't impact overall system
1183  * performance. We compute two limits:
1184  * 1) Hard memory limit: if the amount of memory used by the sorting
1185  *	queues on a pool gets above this value, we stop the metadata
1186  *	scanning portion and start issuing the queued up and sorted
1187  *	I/Os to reduce memory usage.
1188  *	This limit is calculated as a fraction of physmem (by default 5%).
1189  *	We constrain the lower bound of the hard limit to an absolute
1190  *	minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1191  *	the upper bound to 5% of the total pool size - no chance we'll
1192  *	ever need that much memory, but just to keep the value in check.
1193  * 2) Soft memory limit: once we hit the hard memory limit, we start
1194  *	issuing I/O to reduce queue memory usage, but we don't want to
1195  *	completely empty out the queues, since we might be able to find I/Os
1196  *	that will fill in the gaps of our non-sequential IOs at some point
1197  *	in the future. So we stop the issuing of I/Os once the amount of
1198  *	memory used drops below the soft limit (at which point we stop issuing
1199  *	I/O and start scanning metadata again).
1200  *
1201  *	This limit is calculated by subtracting a fraction of the hard
1202  *	limit from the hard limit. By default this fraction is 5%, so
1203  *	the soft limit is 95% of the hard limit. We cap the size of the
1204  *	difference between the hard and soft limits at an absolute
1205  *	maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1206  *	sufficient to not cause too frequent switching between the
1207  *	metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1208  *	worth of queues is about 1.2 GiB of on-pool data, so scanning
1209  *	that should take at least a decent fraction of a second).
1210  */
1211 static boolean_t
dsl_scan_should_clear(dsl_scan_t * scn)1212 dsl_scan_should_clear(dsl_scan_t *scn)
1213 {
1214 	spa_t *spa = scn->scn_dp->dp_spa;
1215 	vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1216 	uint64_t alloc, mlim_hard, mlim_soft, mused;
1217 
1218 	alloc = metaslab_class_get_alloc(spa_normal_class(spa));
1219 	alloc += metaslab_class_get_alloc(spa_special_class(spa));
1220 	alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
1221 
1222 	mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1223 	    zfs_scan_mem_lim_min);
1224 	mlim_hard = MIN(mlim_hard, alloc / 20);
1225 	mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1226 	    zfs_scan_mem_lim_soft_max);
1227 	mused = 0;
1228 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1229 		vdev_t *tvd = rvd->vdev_child[i];
1230 		dsl_scan_io_queue_t *queue;
1231 
1232 		mutex_enter(&tvd->vdev_scan_io_queue_lock);
1233 		queue = tvd->vdev_scan_io_queue;
1234 		if (queue != NULL) {
1235 			/* # extents in exts_by_size = # in exts_by_addr */
1236 			mused += zfs_btree_numnodes(&queue->q_exts_by_size) *
1237 			    sizeof (range_seg_gap_t) + queue->q_sio_memused;
1238 		}
1239 		mutex_exit(&tvd->vdev_scan_io_queue_lock);
1240 	}
1241 
1242 	dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1243 
1244 	if (mused == 0)
1245 		ASSERT0(scn->scn_bytes_pending);
1246 
1247 	/*
1248 	 * If we are above our hard limit, we need to clear out memory.
1249 	 * If we are below our soft limit, we need to accumulate sequential IOs.
1250 	 * Otherwise, we should keep doing whatever we are currently doing.
1251 	 */
1252 	if (mused >= mlim_hard)
1253 		return (B_TRUE);
1254 	else if (mused < mlim_soft)
1255 		return (B_FALSE);
1256 	else
1257 		return (scn->scn_clearing);
1258 }
1259 
1260 static boolean_t
dsl_scan_check_suspend(dsl_scan_t * scn,const zbookmark_phys_t * zb)1261 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1262 {
1263 	/* we never skip user/group accounting objects */
1264 	if (zb && (int64_t)zb->zb_object < 0)
1265 		return (B_FALSE);
1266 
1267 	if (scn->scn_suspending)
1268 		return (B_TRUE); /* we're already suspending */
1269 
1270 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1271 		return (B_FALSE); /* we're resuming */
1272 
1273 	/* We only know how to resume from level-0 blocks. */
1274 	if (zb && zb->zb_level != 0)
1275 		return (B_FALSE);
1276 
1277 	/*
1278 	 * We suspend if:
1279 	 *  - we have scanned for at least the minimum time (default 1 sec
1280 	 *    for scrub, 3 sec for resilver), and either we have sufficient
1281 	 *    dirty data that we are starting to write more quickly
1282 	 *    (default 30%), or someone is explicitly waiting for this txg
1283 	 *    to complete.
1284 	 *  or
1285 	 *  - the spa is shutting down because this pool is being exported
1286 	 *    or the machine is rebooting.
1287 	 *  or
1288 	 *  - the scan queue has reached its memory use limit
1289 	 */
1290 	hrtime_t curr_time_ns = gethrtime();
1291 	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1292 	uint64_t sync_time_ns = curr_time_ns -
1293 	    scn->scn_dp->dp_spa->spa_sync_starttime;
1294 
1295 	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
1296 	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1297 	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1298 
1299 	if ((NSEC2MSEC(scan_time_ns) > mintime &&
1300 	    (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
1301 	    txg_sync_waiting(scn->scn_dp) ||
1302 	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1303 	    spa_shutting_down(scn->scn_dp->dp_spa) ||
1304 	    (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1305 		if (zb) {
1306 			dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1307 			    (longlong_t)zb->zb_objset,
1308 			    (longlong_t)zb->zb_object,
1309 			    (longlong_t)zb->zb_level,
1310 			    (longlong_t)zb->zb_blkid);
1311 			scn->scn_phys.scn_bookmark = *zb;
1312 		} else {
1313 			dsl_scan_phys_t *scnp = &scn->scn_phys;
1314 
1315 			dprintf("suspending at DDT bookmark "
1316 			    "%llx/%llx/%llx/%llx\n",
1317 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1318 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1319 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1320 			    (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1321 		}
1322 		scn->scn_suspending = B_TRUE;
1323 		return (B_TRUE);
1324 	}
1325 	return (B_FALSE);
1326 }
1327 
1328 typedef struct zil_scan_arg {
1329 	dsl_pool_t	*zsa_dp;
1330 	zil_header_t	*zsa_zh;
1331 } zil_scan_arg_t;
1332 
1333 /* ARGSUSED */
1334 static int
dsl_scan_zil_block(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)1335 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1336 {
1337 	zil_scan_arg_t *zsa = arg;
1338 	dsl_pool_t *dp = zsa->zsa_dp;
1339 	dsl_scan_t *scn = dp->dp_scan;
1340 	zil_header_t *zh = zsa->zsa_zh;
1341 	zbookmark_phys_t zb;
1342 
1343 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1344 		return (0);
1345 
1346 	/*
1347 	 * One block ("stubby") can be allocated a long time ago; we
1348 	 * want to visit that one because it has been allocated
1349 	 * (on-disk) even if it hasn't been claimed (even though for
1350 	 * scrub there's nothing to do to it).
1351 	 */
1352 	if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1353 		return (0);
1354 
1355 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1356 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1357 
1358 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1359 	return (0);
1360 }
1361 
1362 /* ARGSUSED */
1363 static int
dsl_scan_zil_record(zilog_t * zilog,lr_t * lrc,void * arg,uint64_t claim_txg)1364 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
1365 {
1366 	if (lrc->lrc_txtype == TX_WRITE) {
1367 		zil_scan_arg_t *zsa = arg;
1368 		dsl_pool_t *dp = zsa->zsa_dp;
1369 		dsl_scan_t *scn = dp->dp_scan;
1370 		zil_header_t *zh = zsa->zsa_zh;
1371 		lr_write_t *lr = (lr_write_t *)lrc;
1372 		blkptr_t *bp = &lr->lr_blkptr;
1373 		zbookmark_phys_t zb;
1374 
1375 		if (BP_IS_HOLE(bp) ||
1376 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1377 			return (0);
1378 
1379 		/*
1380 		 * birth can be < claim_txg if this record's txg is
1381 		 * already txg sync'ed (but this log block contains
1382 		 * other records that are not synced)
1383 		 */
1384 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
1385 			return (0);
1386 
1387 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1388 		    lr->lr_foid, ZB_ZIL_LEVEL,
1389 		    lr->lr_offset / BP_GET_LSIZE(bp));
1390 
1391 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1392 	}
1393 	return (0);
1394 }
1395 
1396 static void
dsl_scan_zil(dsl_pool_t * dp,zil_header_t * zh)1397 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1398 {
1399 	uint64_t claim_txg = zh->zh_claim_txg;
1400 	zil_scan_arg_t zsa = { dp, zh };
1401 	zilog_t *zilog;
1402 
1403 	ASSERT(spa_writeable(dp->dp_spa));
1404 
1405 	/*
1406 	 * We only want to visit blocks that have been claimed
1407 	 * but not yet replayed.
1408 	 */
1409 	if (claim_txg == 0)
1410 		return;
1411 
1412 	zilog = zil_alloc(dp->dp_meta_objset, zh);
1413 
1414 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1415 	    claim_txg, B_FALSE);
1416 
1417 	zil_free(zilog);
1418 }
1419 
1420 /*
1421  * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1422  * here is to sort the AVL tree by the order each block will be needed.
1423  */
1424 static int
scan_prefetch_queue_compare(const void * a,const void * b)1425 scan_prefetch_queue_compare(const void *a, const void *b)
1426 {
1427 	const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1428 	const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1429 	const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1430 
1431 	return (zbookmark_compare(spc_a->spc_datablkszsec,
1432 	    spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1433 	    spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1434 }
1435 
1436 static void
scan_prefetch_ctx_rele(scan_prefetch_ctx_t * spc,void * tag)1437 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, void *tag)
1438 {
1439 	if (zfs_refcount_remove(&spc->spc_refcnt, tag) == 0) {
1440 		zfs_refcount_destroy(&spc->spc_refcnt);
1441 		kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1442 	}
1443 }
1444 
1445 static scan_prefetch_ctx_t *
scan_prefetch_ctx_create(dsl_scan_t * scn,dnode_phys_t * dnp,void * tag)1446 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, void *tag)
1447 {
1448 	scan_prefetch_ctx_t *spc;
1449 
1450 	spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1451 	zfs_refcount_create(&spc->spc_refcnt);
1452 	zfs_refcount_add(&spc->spc_refcnt, tag);
1453 	spc->spc_scn = scn;
1454 	if (dnp != NULL) {
1455 		spc->spc_datablkszsec = dnp->dn_datablkszsec;
1456 		spc->spc_indblkshift = dnp->dn_indblkshift;
1457 		spc->spc_root = B_FALSE;
1458 	} else {
1459 		spc->spc_datablkszsec = 0;
1460 		spc->spc_indblkshift = 0;
1461 		spc->spc_root = B_TRUE;
1462 	}
1463 
1464 	return (spc);
1465 }
1466 
1467 static void
scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t * spc,void * tag)1468 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, void *tag)
1469 {
1470 	zfs_refcount_add(&spc->spc_refcnt, tag);
1471 }
1472 
1473 static boolean_t
dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t * spc,const zbookmark_phys_t * zb)1474 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1475     const zbookmark_phys_t *zb)
1476 {
1477 	zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1478 	dnode_phys_t tmp_dnp;
1479 	dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1480 
1481 	if (zb->zb_objset != last_zb->zb_objset)
1482 		return (B_TRUE);
1483 	if ((int64_t)zb->zb_object < 0)
1484 		return (B_FALSE);
1485 
1486 	tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1487 	tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1488 
1489 	if (zbookmark_subtree_completed(dnp, zb, last_zb))
1490 		return (B_TRUE);
1491 
1492 	return (B_FALSE);
1493 }
1494 
1495 static void
dsl_scan_prefetch(scan_prefetch_ctx_t * spc,blkptr_t * bp,zbookmark_phys_t * zb)1496 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1497 {
1498 	avl_index_t idx;
1499 	dsl_scan_t *scn = spc->spc_scn;
1500 	spa_t *spa = scn->scn_dp->dp_spa;
1501 	scan_prefetch_issue_ctx_t *spic;
1502 
1503 	if (zfs_no_scrub_prefetch)
1504 		return;
1505 
1506 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1507 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1508 	    BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1509 		return;
1510 
1511 	if (dsl_scan_check_prefetch_resume(spc, zb))
1512 		return;
1513 
1514 	scan_prefetch_ctx_add_ref(spc, scn);
1515 	spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1516 	spic->spic_spc = spc;
1517 	spic->spic_bp = *bp;
1518 	spic->spic_zb = *zb;
1519 
1520 	/*
1521 	 * Add the IO to the queue of blocks to prefetch. This allows us to
1522 	 * prioritize blocks that we will need first for the main traversal
1523 	 * thread.
1524 	 */
1525 	mutex_enter(&spa->spa_scrub_lock);
1526 	if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1527 		/* this block is already queued for prefetch */
1528 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1529 		scan_prefetch_ctx_rele(spc, scn);
1530 		mutex_exit(&spa->spa_scrub_lock);
1531 		return;
1532 	}
1533 
1534 	avl_insert(&scn->scn_prefetch_queue, spic, idx);
1535 	cv_broadcast(&spa->spa_scrub_io_cv);
1536 	mutex_exit(&spa->spa_scrub_lock);
1537 }
1538 
1539 static void
dsl_scan_prefetch_dnode(dsl_scan_t * scn,dnode_phys_t * dnp,uint64_t objset,uint64_t object)1540 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1541     uint64_t objset, uint64_t object)
1542 {
1543 	int i;
1544 	zbookmark_phys_t zb;
1545 	scan_prefetch_ctx_t *spc;
1546 
1547 	if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1548 		return;
1549 
1550 	SET_BOOKMARK(&zb, objset, object, 0, 0);
1551 
1552 	spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1553 
1554 	for (i = 0; i < dnp->dn_nblkptr; i++) {
1555 		zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1556 		zb.zb_blkid = i;
1557 		dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1558 	}
1559 
1560 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1561 		zb.zb_level = 0;
1562 		zb.zb_blkid = DMU_SPILL_BLKID;
1563 		dsl_scan_prefetch(spc, &dnp->dn_spill, &zb);
1564 	}
1565 
1566 	scan_prefetch_ctx_rele(spc, FTAG);
1567 }
1568 
1569 void
dsl_scan_prefetch_cb(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * private)1570 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1571     arc_buf_t *buf, void *private)
1572 {
1573 	scan_prefetch_ctx_t *spc = private;
1574 	dsl_scan_t *scn = spc->spc_scn;
1575 	spa_t *spa = scn->scn_dp->dp_spa;
1576 
1577 	/* broadcast that the IO has completed for rate limitting purposes */
1578 	mutex_enter(&spa->spa_scrub_lock);
1579 	ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1580 	spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1581 	cv_broadcast(&spa->spa_scrub_io_cv);
1582 	mutex_exit(&spa->spa_scrub_lock);
1583 
1584 	/* if there was an error or we are done prefetching, just cleanup */
1585 	if (buf == NULL || scn->scn_suspending)
1586 		goto out;
1587 
1588 	if (BP_GET_LEVEL(bp) > 0) {
1589 		int i;
1590 		blkptr_t *cbp;
1591 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1592 		zbookmark_phys_t czb;
1593 
1594 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1595 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1596 			    zb->zb_level - 1, zb->zb_blkid * epb + i);
1597 			dsl_scan_prefetch(spc, cbp, &czb);
1598 		}
1599 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1600 		dnode_phys_t *cdnp = buf->b_data;
1601 		int i;
1602 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1603 
1604 		for (i = 0, cdnp = buf->b_data; i < epb;
1605 		    i += cdnp->dn_extra_slots + 1,
1606 		    cdnp += cdnp->dn_extra_slots + 1) {
1607 			dsl_scan_prefetch_dnode(scn, cdnp,
1608 			    zb->zb_objset, zb->zb_blkid * epb + i);
1609 		}
1610 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1611 		objset_phys_t *osp = buf->b_data;
1612 
1613 		dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1614 		    zb->zb_objset, DMU_META_DNODE_OBJECT);
1615 
1616 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
1617 			dsl_scan_prefetch_dnode(scn,
1618 			    &osp->os_groupused_dnode, zb->zb_objset,
1619 			    DMU_GROUPUSED_OBJECT);
1620 			dsl_scan_prefetch_dnode(scn,
1621 			    &osp->os_userused_dnode, zb->zb_objset,
1622 			    DMU_USERUSED_OBJECT);
1623 		}
1624 	}
1625 
1626 out:
1627 	if (buf != NULL)
1628 		arc_buf_destroy(buf, private);
1629 	scan_prefetch_ctx_rele(spc, scn);
1630 }
1631 
1632 /* ARGSUSED */
1633 static void
dsl_scan_prefetch_thread(void * arg)1634 dsl_scan_prefetch_thread(void *arg)
1635 {
1636 	dsl_scan_t *scn = arg;
1637 	spa_t *spa = scn->scn_dp->dp_spa;
1638 	vdev_t *rvd = spa->spa_root_vdev;
1639 	uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1640 	scan_prefetch_issue_ctx_t *spic;
1641 
1642 	/* loop until we are told to stop */
1643 	while (!scn->scn_prefetch_stop) {
1644 		arc_flags_t flags = ARC_FLAG_NOWAIT |
1645 		    ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1646 		int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1647 
1648 		mutex_enter(&spa->spa_scrub_lock);
1649 
1650 		/*
1651 		 * Wait until we have an IO to issue and are not above our
1652 		 * maximum in flight limit.
1653 		 */
1654 		while (!scn->scn_prefetch_stop &&
1655 		    (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1656 		    spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1657 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1658 		}
1659 
1660 		/* recheck if we should stop since we waited for the cv */
1661 		if (scn->scn_prefetch_stop) {
1662 			mutex_exit(&spa->spa_scrub_lock);
1663 			break;
1664 		}
1665 
1666 		/* remove the prefetch IO from the tree */
1667 		spic = avl_first(&scn->scn_prefetch_queue);
1668 		spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1669 		avl_remove(&scn->scn_prefetch_queue, spic);
1670 
1671 		mutex_exit(&spa->spa_scrub_lock);
1672 
1673 		if (BP_IS_PROTECTED(&spic->spic_bp)) {
1674 			ASSERT(BP_GET_TYPE(&spic->spic_bp) == DMU_OT_DNODE ||
1675 			    BP_GET_TYPE(&spic->spic_bp) == DMU_OT_OBJSET);
1676 			ASSERT3U(BP_GET_LEVEL(&spic->spic_bp), ==, 0);
1677 			zio_flags |= ZIO_FLAG_RAW;
1678 		}
1679 
1680 		/* issue the prefetch asynchronously */
1681 		(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1682 		    &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1683 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1684 
1685 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1686 	}
1687 
1688 	ASSERT(scn->scn_prefetch_stop);
1689 
1690 	/* free any prefetches we didn't get to complete */
1691 	mutex_enter(&spa->spa_scrub_lock);
1692 	while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1693 		avl_remove(&scn->scn_prefetch_queue, spic);
1694 		scan_prefetch_ctx_rele(spic->spic_spc, scn);
1695 		kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1696 	}
1697 	ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1698 	mutex_exit(&spa->spa_scrub_lock);
1699 }
1700 
1701 static boolean_t
dsl_scan_check_resume(dsl_scan_t * scn,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)1702 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1703     const zbookmark_phys_t *zb)
1704 {
1705 	/*
1706 	 * We never skip over user/group accounting objects (obj<0)
1707 	 */
1708 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1709 	    (int64_t)zb->zb_object >= 0) {
1710 		/*
1711 		 * If we already visited this bp & everything below (in
1712 		 * a prior txg sync), don't bother doing it again.
1713 		 */
1714 		if (zbookmark_subtree_completed(dnp, zb,
1715 		    &scn->scn_phys.scn_bookmark))
1716 			return (B_TRUE);
1717 
1718 		/*
1719 		 * If we found the block we're trying to resume from, or
1720 		 * we went past it to a different object, zero it out to
1721 		 * indicate that it's OK to start checking for suspending
1722 		 * again.
1723 		 */
1724 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
1725 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
1726 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
1727 			    (longlong_t)zb->zb_objset,
1728 			    (longlong_t)zb->zb_object,
1729 			    (longlong_t)zb->zb_level,
1730 			    (longlong_t)zb->zb_blkid);
1731 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
1732 		}
1733 	}
1734 	return (B_FALSE);
1735 }
1736 
1737 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1738     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1739     dmu_objset_type_t ostype, dmu_tx_t *tx);
1740 static void dsl_scan_visitdnode(
1741     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1742     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1743 
1744 /*
1745  * Return nonzero on i/o error.
1746  * Return new buf to write out in *bufp.
1747  */
1748 static int
dsl_scan_recurse(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb,dmu_tx_t * tx)1749 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1750     dnode_phys_t *dnp, const blkptr_t *bp,
1751     const zbookmark_phys_t *zb, dmu_tx_t *tx)
1752 {
1753 	dsl_pool_t *dp = scn->scn_dp;
1754 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1755 	int err;
1756 
1757 	if (BP_GET_LEVEL(bp) > 0) {
1758 		arc_flags_t flags = ARC_FLAG_WAIT;
1759 		int i;
1760 		blkptr_t *cbp;
1761 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1762 		arc_buf_t *buf;
1763 
1764 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1765 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1766 		if (err) {
1767 			scn->scn_phys.scn_errors++;
1768 			return (err);
1769 		}
1770 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1771 			zbookmark_phys_t czb;
1772 
1773 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1774 			    zb->zb_level - 1,
1775 			    zb->zb_blkid * epb + i);
1776 			dsl_scan_visitbp(cbp, &czb, dnp,
1777 			    ds, scn, ostype, tx);
1778 		}
1779 		arc_buf_destroy(buf, &buf);
1780 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1781 		arc_flags_t flags = ARC_FLAG_WAIT;
1782 		dnode_phys_t *cdnp;
1783 		int i;
1784 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1785 		arc_buf_t *buf;
1786 
1787 		if (BP_IS_PROTECTED(bp)) {
1788 			ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF);
1789 			zio_flags |= ZIO_FLAG_RAW;
1790 		}
1791 
1792 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1793 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1794 		if (err) {
1795 			scn->scn_phys.scn_errors++;
1796 			return (err);
1797 		}
1798 		for (i = 0, cdnp = buf->b_data; i < epb;
1799 		    i += cdnp->dn_extra_slots + 1,
1800 		    cdnp += cdnp->dn_extra_slots + 1) {
1801 			dsl_scan_visitdnode(scn, ds, ostype,
1802 			    cdnp, zb->zb_blkid * epb + i, tx);
1803 		}
1804 
1805 		arc_buf_destroy(buf, &buf);
1806 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1807 		arc_flags_t flags = ARC_FLAG_WAIT;
1808 		objset_phys_t *osp;
1809 		arc_buf_t *buf;
1810 
1811 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1812 		    ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1813 		if (err) {
1814 			scn->scn_phys.scn_errors++;
1815 			return (err);
1816 		}
1817 
1818 		osp = buf->b_data;
1819 
1820 		dsl_scan_visitdnode(scn, ds, osp->os_type,
1821 		    &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1822 
1823 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
1824 			/*
1825 			 * We also always visit user/group/project accounting
1826 			 * objects, and never skip them, even if we are
1827 			 * suspending.  This is necessary so that the space
1828 			 * deltas from this txg get integrated.
1829 			 */
1830 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
1831 				dsl_scan_visitdnode(scn, ds, osp->os_type,
1832 				    &osp->os_projectused_dnode,
1833 				    DMU_PROJECTUSED_OBJECT, tx);
1834 			dsl_scan_visitdnode(scn, ds, osp->os_type,
1835 			    &osp->os_groupused_dnode,
1836 			    DMU_GROUPUSED_OBJECT, tx);
1837 			dsl_scan_visitdnode(scn, ds, osp->os_type,
1838 			    &osp->os_userused_dnode,
1839 			    DMU_USERUSED_OBJECT, tx);
1840 		}
1841 		arc_buf_destroy(buf, &buf);
1842 	}
1843 
1844 	return (0);
1845 }
1846 
1847 static void
dsl_scan_visitdnode(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,uint64_t object,dmu_tx_t * tx)1848 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1849     dmu_objset_type_t ostype, dnode_phys_t *dnp,
1850     uint64_t object, dmu_tx_t *tx)
1851 {
1852 	int j;
1853 
1854 	for (j = 0; j < dnp->dn_nblkptr; j++) {
1855 		zbookmark_phys_t czb;
1856 
1857 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1858 		    dnp->dn_nlevels - 1, j);
1859 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
1860 		    &czb, dnp, ds, scn, ostype, tx);
1861 	}
1862 
1863 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1864 		zbookmark_phys_t czb;
1865 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1866 		    0, DMU_SPILL_BLKID);
1867 		dsl_scan_visitbp(DN_SPILL_BLKPTR(dnp),
1868 		    &czb, dnp, ds, scn, ostype, tx);
1869 	}
1870 }
1871 
1872 /*
1873  * The arguments are in this order because mdb can only print the
1874  * first 5; we want them to be useful.
1875  */
1876 static void
dsl_scan_visitbp(blkptr_t * bp,const zbookmark_phys_t * zb,dnode_phys_t * dnp,dsl_dataset_t * ds,dsl_scan_t * scn,dmu_objset_type_t ostype,dmu_tx_t * tx)1877 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1878     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1879     dmu_objset_type_t ostype, dmu_tx_t *tx)
1880 {
1881 	dsl_pool_t *dp = scn->scn_dp;
1882 	blkptr_t *bp_toread = NULL;
1883 
1884 	if (dsl_scan_check_suspend(scn, zb))
1885 		return;
1886 
1887 	if (dsl_scan_check_resume(scn, dnp, zb))
1888 		return;
1889 
1890 	scn->scn_visited_this_txg++;
1891 
1892 	/*
1893 	 * This debugging is commented out to conserve stack space.  This
1894 	 * function is called recursively and the debugging addes several
1895 	 * bytes to the stack for each call.  It can be commented back in
1896 	 * if required to debug an issue in dsl_scan_visitbp().
1897 	 *
1898 	 * dprintf_bp(bp,
1899 	 *	"visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1900 	 *	ds, ds ? ds->ds_object : 0,
1901 	 *	zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1902 	 *	bp);
1903 	 */
1904 
1905 	if (BP_IS_HOLE(bp)) {
1906 		scn->scn_holes_this_txg++;
1907 		return;
1908 	}
1909 
1910 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
1911 		scn->scn_lt_min_this_txg++;
1912 		return;
1913 	}
1914 
1915 	bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
1916 	*bp_toread = *bp;
1917 
1918 	if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
1919 		goto out;
1920 
1921 	/*
1922 	 * If dsl_scan_ddt() has already visited this block, it will have
1923 	 * already done any translations or scrubbing, so don't call the
1924 	 * callback again.
1925 	 */
1926 	if (ddt_class_contains(dp->dp_spa,
1927 	    scn->scn_phys.scn_ddt_class_max, bp)) {
1928 		scn->scn_ddt_contained_this_txg++;
1929 		goto out;
1930 	}
1931 
1932 	/*
1933 	 * If this block is from the future (after cur_max_txg), then we
1934 	 * are doing this on behalf of a deleted snapshot, and we will
1935 	 * revisit the future block on the next pass of this dataset.
1936 	 * Don't scan it now unless we need to because something
1937 	 * under it was modified.
1938 	 */
1939 	if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
1940 		scn->scn_gt_max_this_txg++;
1941 		goto out;
1942 	}
1943 
1944 	scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
1945 
1946 out:
1947 	kmem_free(bp_toread, sizeof (blkptr_t));
1948 }
1949 
1950 static void
dsl_scan_visit_rootbp(dsl_scan_t * scn,dsl_dataset_t * ds,blkptr_t * bp,dmu_tx_t * tx)1951 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
1952     dmu_tx_t *tx)
1953 {
1954 	zbookmark_phys_t zb;
1955 	scan_prefetch_ctx_t *spc;
1956 
1957 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1958 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1959 
1960 	if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
1961 		SET_BOOKMARK(&scn->scn_prefetch_bookmark,
1962 		    zb.zb_objset, 0, 0, 0);
1963 	} else {
1964 		scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
1965 	}
1966 
1967 	scn->scn_objsets_visited_this_txg++;
1968 
1969 	spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
1970 	dsl_scan_prefetch(spc, bp, &zb);
1971 	scan_prefetch_ctx_rele(spc, FTAG);
1972 
1973 	dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
1974 
1975 	dprintf_ds(ds, "finished scan%s", "");
1976 }
1977 
1978 static void
ds_destroyed_scn_phys(dsl_dataset_t * ds,dsl_scan_phys_t * scn_phys)1979 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
1980 {
1981 	if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
1982 		if (ds->ds_is_snapshot) {
1983 			/*
1984 			 * Note:
1985 			 *  - scn_cur_{min,max}_txg stays the same.
1986 			 *  - Setting the flag is not really necessary if
1987 			 *    scn_cur_max_txg == scn_max_txg, because there
1988 			 *    is nothing after this snapshot that we care
1989 			 *    about.  However, we set it anyway and then
1990 			 *    ignore it when we retraverse it in
1991 			 *    dsl_scan_visitds().
1992 			 */
1993 			scn_phys->scn_bookmark.zb_objset =
1994 			    dsl_dataset_phys(ds)->ds_next_snap_obj;
1995 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
1996 			    "reset zb_objset to %llu",
1997 			    (u_longlong_t)ds->ds_object,
1998 			    (u_longlong_t)dsl_dataset_phys(ds)->
1999 			    ds_next_snap_obj);
2000 			scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
2001 		} else {
2002 			SET_BOOKMARK(&scn_phys->scn_bookmark,
2003 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
2004 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
2005 			    "reset bookmark to -1,0,0,0",
2006 			    (u_longlong_t)ds->ds_object);
2007 		}
2008 	}
2009 }
2010 
2011 /*
2012  * Invoked when a dataset is destroyed. We need to make sure that:
2013  *
2014  * 1) If it is the dataset that was currently being scanned, we write
2015  *	a new dsl_scan_phys_t and marking the objset reference in it
2016  *	as destroyed.
2017  * 2) Remove it from the work queue, if it was present.
2018  *
2019  * If the dataset was actually a snapshot, instead of marking the dataset
2020  * as destroyed, we instead substitute the next snapshot in line.
2021  */
2022 void
dsl_scan_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)2023 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
2024 {
2025 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2026 	dsl_scan_t *scn = dp->dp_scan;
2027 	uint64_t mintxg;
2028 
2029 	if (!dsl_scan_is_running(scn))
2030 		return;
2031 
2032 	ds_destroyed_scn_phys(ds, &scn->scn_phys);
2033 	ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
2034 
2035 	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2036 		scan_ds_queue_remove(scn, ds->ds_object);
2037 		if (ds->ds_is_snapshot)
2038 			scan_ds_queue_insert(scn,
2039 			    dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
2040 	}
2041 
2042 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2043 	    ds->ds_object, &mintxg) == 0) {
2044 		ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
2045 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2046 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2047 		if (ds->ds_is_snapshot) {
2048 			/*
2049 			 * We keep the same mintxg; it could be >
2050 			 * ds_creation_txg if the previous snapshot was
2051 			 * deleted too.
2052 			 */
2053 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
2054 			    scn->scn_phys.scn_queue_obj,
2055 			    dsl_dataset_phys(ds)->ds_next_snap_obj,
2056 			    mintxg, tx) == 0);
2057 			zfs_dbgmsg("destroying ds %llu; in queue; "
2058 			    "replacing with %llu",
2059 			    (u_longlong_t)ds->ds_object,
2060 			    (u_longlong_t)dsl_dataset_phys(ds)->
2061 			    ds_next_snap_obj);
2062 		} else {
2063 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
2064 			    (u_longlong_t)ds->ds_object);
2065 		}
2066 	}
2067 
2068 	/*
2069 	 * dsl_scan_sync() should be called after this, and should sync
2070 	 * out our changed state, but just to be safe, do it here.
2071 	 */
2072 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2073 }
2074 
2075 static void
ds_snapshotted_bookmark(dsl_dataset_t * ds,zbookmark_phys_t * scn_bookmark)2076 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
2077 {
2078 	if (scn_bookmark->zb_objset == ds->ds_object) {
2079 		scn_bookmark->zb_objset =
2080 		    dsl_dataset_phys(ds)->ds_prev_snap_obj;
2081 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
2082 		    "reset zb_objset to %llu",
2083 		    (u_longlong_t)ds->ds_object,
2084 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2085 	}
2086 }
2087 
2088 /*
2089  * Called when a dataset is snapshotted. If we were currently traversing
2090  * this snapshot, we reset our bookmark to point at the newly created
2091  * snapshot. We also modify our work queue to remove the old snapshot and
2092  * replace with the new one.
2093  */
2094 void
dsl_scan_ds_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)2095 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
2096 {
2097 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
2098 	dsl_scan_t *scn = dp->dp_scan;
2099 	uint64_t mintxg;
2100 
2101 	if (!dsl_scan_is_running(scn))
2102 		return;
2103 
2104 	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
2105 
2106 	ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
2107 	ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
2108 
2109 	if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
2110 		scan_ds_queue_remove(scn, ds->ds_object);
2111 		scan_ds_queue_insert(scn,
2112 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
2113 	}
2114 
2115 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2116 	    ds->ds_object, &mintxg) == 0) {
2117 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2118 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
2119 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
2120 		    scn->scn_phys.scn_queue_obj,
2121 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
2122 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
2123 		    "replacing with %llu",
2124 		    (u_longlong_t)ds->ds_object,
2125 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
2126 	}
2127 
2128 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2129 }
2130 
2131 static void
ds_clone_swapped_bookmark(dsl_dataset_t * ds1,dsl_dataset_t * ds2,zbookmark_phys_t * scn_bookmark)2132 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
2133     zbookmark_phys_t *scn_bookmark)
2134 {
2135 	if (scn_bookmark->zb_objset == ds1->ds_object) {
2136 		scn_bookmark->zb_objset = ds2->ds_object;
2137 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2138 		    "reset zb_objset to %llu",
2139 		    (u_longlong_t)ds1->ds_object,
2140 		    (u_longlong_t)ds2->ds_object);
2141 	} else if (scn_bookmark->zb_objset == ds2->ds_object) {
2142 		scn_bookmark->zb_objset = ds1->ds_object;
2143 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2144 		    "reset zb_objset to %llu",
2145 		    (u_longlong_t)ds2->ds_object,
2146 		    (u_longlong_t)ds1->ds_object);
2147 	}
2148 }
2149 
2150 /*
2151  * Called when a parent dataset and its clone are swapped. If we were
2152  * currently traversing the dataset, we need to switch to traversing the
2153  * newly promoted parent.
2154  */
2155 void
dsl_scan_ds_clone_swapped(dsl_dataset_t * ds1,dsl_dataset_t * ds2,dmu_tx_t * tx)2156 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2157 {
2158 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2159 	dsl_scan_t *scn = dp->dp_scan;
2160 	uint64_t mintxg;
2161 
2162 	if (!dsl_scan_is_running(scn))
2163 		return;
2164 
2165 	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2166 	ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2167 
2168 	if (scan_ds_queue_contains(scn, ds1->ds_object, &mintxg)) {
2169 		scan_ds_queue_remove(scn, ds1->ds_object);
2170 		scan_ds_queue_insert(scn, ds2->ds_object, mintxg);
2171 	}
2172 	if (scan_ds_queue_contains(scn, ds2->ds_object, &mintxg)) {
2173 		scan_ds_queue_remove(scn, ds2->ds_object);
2174 		scan_ds_queue_insert(scn, ds1->ds_object, mintxg);
2175 	}
2176 
2177 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2178 	    ds1->ds_object, &mintxg) == 0) {
2179 		int err;
2180 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2181 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2182 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2183 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2184 		err = zap_add_int_key(dp->dp_meta_objset,
2185 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
2186 		VERIFY(err == 0 || err == EEXIST);
2187 		if (err == EEXIST) {
2188 			/* Both were there to begin with */
2189 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2190 			    scn->scn_phys.scn_queue_obj,
2191 			    ds1->ds_object, mintxg, tx));
2192 		}
2193 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
2194 		    "replacing with %llu",
2195 		    (u_longlong_t)ds1->ds_object,
2196 		    (u_longlong_t)ds2->ds_object);
2197 	}
2198 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2199 	    ds2->ds_object, &mintxg) == 0) {
2200 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2201 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2202 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2203 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2204 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2205 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
2206 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
2207 		    "replacing with %llu",
2208 		    (u_longlong_t)ds2->ds_object,
2209 		    (u_longlong_t)ds1->ds_object);
2210 	}
2211 
2212 	dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2213 }
2214 
2215 /* ARGSUSED */
2216 static int
enqueue_clones_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)2217 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2218 {
2219 	uint64_t originobj = *(uint64_t *)arg;
2220 	dsl_dataset_t *ds;
2221 	int err;
2222 	dsl_scan_t *scn = dp->dp_scan;
2223 
2224 	if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2225 		return (0);
2226 
2227 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2228 	if (err)
2229 		return (err);
2230 
2231 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2232 		dsl_dataset_t *prev;
2233 		err = dsl_dataset_hold_obj(dp,
2234 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2235 
2236 		dsl_dataset_rele(ds, FTAG);
2237 		if (err)
2238 			return (err);
2239 		ds = prev;
2240 	}
2241 	scan_ds_queue_insert(scn, ds->ds_object,
2242 	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2243 	dsl_dataset_rele(ds, FTAG);
2244 	return (0);
2245 }
2246 
2247 static void
dsl_scan_visitds(dsl_scan_t * scn,uint64_t dsobj,dmu_tx_t * tx)2248 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2249 {
2250 	dsl_pool_t *dp = scn->scn_dp;
2251 	dsl_dataset_t *ds;
2252 
2253 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2254 
2255 	if (scn->scn_phys.scn_cur_min_txg >=
2256 	    scn->scn_phys.scn_max_txg) {
2257 		/*
2258 		 * This can happen if this snapshot was created after the
2259 		 * scan started, and we already completed a previous snapshot
2260 		 * that was created after the scan started.  This snapshot
2261 		 * only references blocks with:
2262 		 *
2263 		 *	birth < our ds_creation_txg
2264 		 *	cur_min_txg is no less than ds_creation_txg.
2265 		 *	We have already visited these blocks.
2266 		 * or
2267 		 *	birth > scn_max_txg
2268 		 *	The scan requested not to visit these blocks.
2269 		 *
2270 		 * Subsequent snapshots (and clones) can reference our
2271 		 * blocks, or blocks with even higher birth times.
2272 		 * Therefore we do not need to visit them either,
2273 		 * so we do not add them to the work queue.
2274 		 *
2275 		 * Note that checking for cur_min_txg >= cur_max_txg
2276 		 * is not sufficient, because in that case we may need to
2277 		 * visit subsequent snapshots.  This happens when min_txg > 0,
2278 		 * which raises cur_min_txg.  In this case we will visit
2279 		 * this dataset but skip all of its blocks, because the
2280 		 * rootbp's birth time is < cur_min_txg.  Then we will
2281 		 * add the next snapshots/clones to the work queue.
2282 		 */
2283 		char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
2284 		dsl_dataset_name(ds, dsname);
2285 		zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2286 		    "cur_min_txg (%llu) >= max_txg (%llu)",
2287 		    (longlong_t)dsobj, dsname,
2288 		    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2289 		    (longlong_t)scn->scn_phys.scn_max_txg);
2290 		kmem_free(dsname, MAXNAMELEN);
2291 
2292 		goto out;
2293 	}
2294 
2295 	/*
2296 	 * Only the ZIL in the head (non-snapshot) is valid. Even though
2297 	 * snapshots can have ZIL block pointers (which may be the same
2298 	 * BP as in the head), they must be ignored. In addition, $ORIGIN
2299 	 * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2300 	 * need to look for a ZIL in it either. So we traverse the ZIL here,
2301 	 * rather than in scan_recurse(), because the regular snapshot
2302 	 * block-sharing rules don't apply to it.
2303 	 */
2304 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds) &&
2305 	    (dp->dp_origin_snap == NULL ||
2306 	    ds->ds_dir != dp->dp_origin_snap->ds_dir)) {
2307 		objset_t *os;
2308 		if (dmu_objset_from_ds(ds, &os) != 0) {
2309 			goto out;
2310 		}
2311 		dsl_scan_zil(dp, &os->os_zil_header);
2312 	}
2313 
2314 	/*
2315 	 * Iterate over the bps in this ds.
2316 	 */
2317 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
2318 	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2319 	dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2320 	rrw_exit(&ds->ds_bp_rwlock, FTAG);
2321 
2322 	char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2323 	dsl_dataset_name(ds, dsname);
2324 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2325 	    "suspending=%u",
2326 	    (longlong_t)dsobj, dsname,
2327 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
2328 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
2329 	    (int)scn->scn_suspending);
2330 	kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2331 
2332 	if (scn->scn_suspending)
2333 		goto out;
2334 
2335 	/*
2336 	 * We've finished this pass over this dataset.
2337 	 */
2338 
2339 	/*
2340 	 * If we did not completely visit this dataset, do another pass.
2341 	 */
2342 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2343 		zfs_dbgmsg("incomplete pass; visiting again");
2344 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2345 		scan_ds_queue_insert(scn, ds->ds_object,
2346 		    scn->scn_phys.scn_cur_max_txg);
2347 		goto out;
2348 	}
2349 
2350 	/*
2351 	 * Add descendent datasets to work queue.
2352 	 */
2353 	if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2354 		scan_ds_queue_insert(scn,
2355 		    dsl_dataset_phys(ds)->ds_next_snap_obj,
2356 		    dsl_dataset_phys(ds)->ds_creation_txg);
2357 	}
2358 	if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2359 		boolean_t usenext = B_FALSE;
2360 		if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2361 			uint64_t count;
2362 			/*
2363 			 * A bug in a previous version of the code could
2364 			 * cause upgrade_clones_cb() to not set
2365 			 * ds_next_snap_obj when it should, leading to a
2366 			 * missing entry.  Therefore we can only use the
2367 			 * next_clones_obj when its count is correct.
2368 			 */
2369 			int err = zap_count(dp->dp_meta_objset,
2370 			    dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2371 			if (err == 0 &&
2372 			    count == dsl_dataset_phys(ds)->ds_num_children - 1)
2373 				usenext = B_TRUE;
2374 		}
2375 
2376 		if (usenext) {
2377 			zap_cursor_t zc;
2378 			zap_attribute_t za;
2379 			for (zap_cursor_init(&zc, dp->dp_meta_objset,
2380 			    dsl_dataset_phys(ds)->ds_next_clones_obj);
2381 			    zap_cursor_retrieve(&zc, &za) == 0;
2382 			    (void) zap_cursor_advance(&zc)) {
2383 				scan_ds_queue_insert(scn,
2384 				    zfs_strtonum(za.za_name, NULL),
2385 				    dsl_dataset_phys(ds)->ds_creation_txg);
2386 			}
2387 			zap_cursor_fini(&zc);
2388 		} else {
2389 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2390 			    enqueue_clones_cb, &ds->ds_object,
2391 			    DS_FIND_CHILDREN));
2392 		}
2393 	}
2394 
2395 out:
2396 	dsl_dataset_rele(ds, FTAG);
2397 }
2398 
2399 /* ARGSUSED */
2400 static int
enqueue_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)2401 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2402 {
2403 	dsl_dataset_t *ds;
2404 	int err;
2405 	dsl_scan_t *scn = dp->dp_scan;
2406 
2407 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2408 	if (err)
2409 		return (err);
2410 
2411 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2412 		dsl_dataset_t *prev;
2413 		err = dsl_dataset_hold_obj(dp,
2414 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2415 		if (err) {
2416 			dsl_dataset_rele(ds, FTAG);
2417 			return (err);
2418 		}
2419 
2420 		/*
2421 		 * If this is a clone, we don't need to worry about it for now.
2422 		 */
2423 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2424 			dsl_dataset_rele(ds, FTAG);
2425 			dsl_dataset_rele(prev, FTAG);
2426 			return (0);
2427 		}
2428 		dsl_dataset_rele(ds, FTAG);
2429 		ds = prev;
2430 	}
2431 
2432 	scan_ds_queue_insert(scn, ds->ds_object,
2433 	    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2434 	dsl_dataset_rele(ds, FTAG);
2435 	return (0);
2436 }
2437 
2438 /* ARGSUSED */
2439 void
dsl_scan_ddt_entry(dsl_scan_t * scn,enum zio_checksum checksum,ddt_entry_t * dde,dmu_tx_t * tx)2440 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2441     ddt_entry_t *dde, dmu_tx_t *tx)
2442 {
2443 	const ddt_key_t *ddk = &dde->dde_key;
2444 	ddt_phys_t *ddp = dde->dde_phys;
2445 	blkptr_t bp;
2446 	zbookmark_phys_t zb = { 0 };
2447 	int p;
2448 
2449 	if (scn->scn_phys.scn_state != DSS_SCANNING)
2450 		return;
2451 
2452 	/*
2453 	 * This function is special because it is the only thing
2454 	 * that can add scan_io_t's to the vdev scan queues from
2455 	 * outside dsl_scan_sync(). For the most part this is ok
2456 	 * as long as it is called from within syncing context.
2457 	 * However, dsl_scan_sync() expects that no new sio's will
2458 	 * be added between when all the work for a scan is done
2459 	 * and the next txg when the scan is actually marked as
2460 	 * completed. This check ensures we do not issue new sio's
2461 	 * during this period.
2462 	 */
2463 	if (scn->scn_done_txg != 0)
2464 		return;
2465 
2466 	for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2467 		if (ddp->ddp_phys_birth == 0 ||
2468 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2469 			continue;
2470 		ddt_bp_create(checksum, ddk, ddp, &bp);
2471 
2472 		scn->scn_visited_this_txg++;
2473 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2474 	}
2475 }
2476 
2477 /*
2478  * Scrub/dedup interaction.
2479  *
2480  * If there are N references to a deduped block, we don't want to scrub it
2481  * N times -- ideally, we should scrub it exactly once.
2482  *
2483  * We leverage the fact that the dde's replication class (enum ddt_class)
2484  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2485  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2486  *
2487  * To prevent excess scrubbing, the scrub begins by walking the DDT
2488  * to find all blocks with refcnt > 1, and scrubs each of these once.
2489  * Since there are two replication classes which contain blocks with
2490  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2491  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2492  *
2493  * There would be nothing more to say if a block's refcnt couldn't change
2494  * during a scrub, but of course it can so we must account for changes
2495  * in a block's replication class.
2496  *
2497  * Here's an example of what can occur:
2498  *
2499  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2500  * when visited during the top-down scrub phase, it will be scrubbed twice.
2501  * This negates our scrub optimization, but is otherwise harmless.
2502  *
2503  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2504  * on each visit during the top-down scrub phase, it will never be scrubbed.
2505  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2506  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2507  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2508  * while a scrub is in progress, it scrubs the block right then.
2509  */
2510 static void
dsl_scan_ddt(dsl_scan_t * scn,dmu_tx_t * tx)2511 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2512 {
2513 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2514 	ddt_entry_t dde = { 0 };
2515 	int error;
2516 	uint64_t n = 0;
2517 
2518 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2519 		ddt_t *ddt;
2520 
2521 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2522 			break;
2523 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2524 		    (longlong_t)ddb->ddb_class,
2525 		    (longlong_t)ddb->ddb_type,
2526 		    (longlong_t)ddb->ddb_checksum,
2527 		    (longlong_t)ddb->ddb_cursor);
2528 
2529 		/* There should be no pending changes to the dedup table */
2530 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2531 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2532 
2533 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2534 		n++;
2535 
2536 		if (dsl_scan_check_suspend(scn, NULL))
2537 			break;
2538 	}
2539 
2540 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2541 	    "suspending=%u", (longlong_t)n,
2542 	    (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2543 
2544 	ASSERT(error == 0 || error == ENOENT);
2545 	ASSERT(error != ENOENT ||
2546 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2547 }
2548 
2549 static uint64_t
dsl_scan_ds_maxtxg(dsl_dataset_t * ds)2550 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2551 {
2552 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2553 	if (ds->ds_is_snapshot)
2554 		return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2555 	return (smt);
2556 }
2557 
2558 static void
dsl_scan_visit(dsl_scan_t * scn,dmu_tx_t * tx)2559 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2560 {
2561 	scan_ds_t *sds;
2562 	dsl_pool_t *dp = scn->scn_dp;
2563 
2564 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2565 	    scn->scn_phys.scn_ddt_class_max) {
2566 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2567 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2568 		dsl_scan_ddt(scn, tx);
2569 		if (scn->scn_suspending)
2570 			return;
2571 	}
2572 
2573 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2574 		/* First do the MOS & ORIGIN */
2575 
2576 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2577 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2578 		dsl_scan_visit_rootbp(scn, NULL,
2579 		    &dp->dp_meta_rootbp, tx);
2580 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2581 		if (scn->scn_suspending)
2582 			return;
2583 
2584 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2585 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2586 			    enqueue_cb, NULL, DS_FIND_CHILDREN));
2587 		} else {
2588 			dsl_scan_visitds(scn,
2589 			    dp->dp_origin_snap->ds_object, tx);
2590 		}
2591 		ASSERT(!scn->scn_suspending);
2592 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
2593 	    ZB_DESTROYED_OBJSET) {
2594 		uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2595 		/*
2596 		 * If we were suspended, continue from here. Note if the
2597 		 * ds we were suspended on was deleted, the zb_objset may
2598 		 * be -1, so we will skip this and find a new objset
2599 		 * below.
2600 		 */
2601 		dsl_scan_visitds(scn, dsobj, tx);
2602 		if (scn->scn_suspending)
2603 			return;
2604 	}
2605 
2606 	/*
2607 	 * In case we suspended right at the end of the ds, zero the
2608 	 * bookmark so we don't think that we're still trying to resume.
2609 	 */
2610 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
2611 
2612 	/*
2613 	 * Keep pulling things out of the dataset avl queue. Updates to the
2614 	 * persistent zap-object-as-queue happen only at checkpoints.
2615 	 */
2616 	while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2617 		dsl_dataset_t *ds;
2618 		uint64_t dsobj = sds->sds_dsobj;
2619 		uint64_t txg = sds->sds_txg;
2620 
2621 		/* dequeue and free the ds from the queue */
2622 		scan_ds_queue_remove(scn, dsobj);
2623 		sds = NULL;	/* must not be touched after removal */
2624 
2625 		/* Set up min / max txg */
2626 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2627 		if (txg != 0) {
2628 			scn->scn_phys.scn_cur_min_txg =
2629 			    MAX(scn->scn_phys.scn_min_txg, txg);
2630 		} else {
2631 			scn->scn_phys.scn_cur_min_txg =
2632 			    MAX(scn->scn_phys.scn_min_txg,
2633 			    dsl_dataset_phys(ds)->ds_prev_snap_txg);
2634 		}
2635 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2636 		dsl_dataset_rele(ds, FTAG);
2637 
2638 		dsl_scan_visitds(scn, dsobj, tx);
2639 		if (scn->scn_suspending)
2640 			return;
2641 	}
2642 	/* No more objsets to fetch, we're done */
2643 	scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2644 	ASSERT0(scn->scn_suspending);
2645 }
2646 
2647 static uint64_t
dsl_scan_count_leaves(vdev_t * vd)2648 dsl_scan_count_leaves(vdev_t *vd)
2649 {
2650 	uint64_t i, leaves = 0;
2651 
2652 	/* we only count leaves that belong to the main pool and are readable */
2653 	if (vd->vdev_islog || vd->vdev_isspare ||
2654 	    vd->vdev_isl2cache || !vdev_readable(vd))
2655 		return (0);
2656 
2657 	if (vd->vdev_ops->vdev_op_leaf)
2658 		return (1);
2659 
2660 	for (i = 0; i < vd->vdev_children; i++) {
2661 		leaves += dsl_scan_count_leaves(vd->vdev_child[i]);
2662 	}
2663 
2664 	return (leaves);
2665 }
2666 
2667 
2668 static void
scan_io_queues_update_zio_stats(dsl_scan_io_queue_t * q,const blkptr_t * bp)2669 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2670 {
2671 	int i;
2672 	uint64_t cur_size = 0;
2673 
2674 	for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2675 		cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2676 	}
2677 
2678 	q->q_total_zio_size_this_txg += cur_size;
2679 	q->q_zios_this_txg++;
2680 }
2681 
2682 static void
scan_io_queues_update_seg_stats(dsl_scan_io_queue_t * q,uint64_t start,uint64_t end)2683 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2684     uint64_t end)
2685 {
2686 	q->q_total_seg_size_this_txg += end - start;
2687 	q->q_segs_this_txg++;
2688 }
2689 
2690 static boolean_t
scan_io_queue_check_suspend(dsl_scan_t * scn)2691 scan_io_queue_check_suspend(dsl_scan_t *scn)
2692 {
2693 	/* See comment in dsl_scan_check_suspend() */
2694 	uint64_t curr_time_ns = gethrtime();
2695 	uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2696 	uint64_t sync_time_ns = curr_time_ns -
2697 	    scn->scn_dp->dp_spa->spa_sync_starttime;
2698 	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
2699 	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2700 	    zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2701 
2702 	return ((NSEC2MSEC(scan_time_ns) > mintime &&
2703 	    (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
2704 	    txg_sync_waiting(scn->scn_dp) ||
2705 	    NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2706 	    spa_shutting_down(scn->scn_dp->dp_spa));
2707 }
2708 
2709 /*
2710  * Given a list of scan_io_t's in io_list, this issues the io's out to
2711  * disk. This consumes the io_list and frees the scan_io_t's. This is
2712  * called when emptying queues, either when we're up against the memory
2713  * limit or when we have finished scanning. Returns B_TRUE if we stopped
2714  * processing the list before we finished. Any zios that were not issued
2715  * will remain in the io_list.
2716  */
2717 static boolean_t
scan_io_queue_issue(dsl_scan_io_queue_t * queue,list_t * io_list)2718 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2719 {
2720 	dsl_scan_t *scn = queue->q_scn;
2721 	scan_io_t *sio;
2722 	int64_t bytes_issued = 0;
2723 	boolean_t suspended = B_FALSE;
2724 
2725 	while ((sio = list_head(io_list)) != NULL) {
2726 		blkptr_t bp;
2727 
2728 		if (scan_io_queue_check_suspend(scn)) {
2729 			suspended = B_TRUE;
2730 			break;
2731 		}
2732 
2733 		sio2bp(sio, &bp);
2734 		bytes_issued += SIO_GET_ASIZE(sio);
2735 		scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2736 		    &sio->sio_zb, queue);
2737 		(void) list_remove_head(io_list);
2738 		scan_io_queues_update_zio_stats(queue, &bp);
2739 		sio_free(sio);
2740 	}
2741 
2742 	atomic_add_64(&scn->scn_bytes_pending, -bytes_issued);
2743 
2744 	return (suspended);
2745 }
2746 
2747 /*
2748  * Given a range_seg_t (extent) and a list, this function passes over a
2749  * scan queue and gathers up the appropriate ios which fit into that
2750  * scan seg (starting from lowest LBA). At the end, we remove the segment
2751  * from the q_exts_by_addr range tree.
2752  */
2753 static boolean_t
scan_io_queue_gather(dsl_scan_io_queue_t * queue,range_seg_t * rs,list_t * list)2754 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2755 {
2756 	scan_io_t *srch_sio, *sio, *next_sio;
2757 	avl_index_t idx;
2758 	uint_t num_sios = 0;
2759 	int64_t bytes_issued = 0;
2760 
2761 	ASSERT(rs != NULL);
2762 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2763 
2764 	srch_sio = sio_alloc(1);
2765 	srch_sio->sio_nr_dvas = 1;
2766 	SIO_SET_OFFSET(srch_sio, rs_get_start(rs, queue->q_exts_by_addr));
2767 
2768 	/*
2769 	 * The exact start of the extent might not contain any matching zios,
2770 	 * so if that's the case, examine the next one in the tree.
2771 	 */
2772 	sio = avl_find(&queue->q_sios_by_addr, srch_sio, &idx);
2773 	sio_free(srch_sio);
2774 
2775 	if (sio == NULL)
2776 		sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2777 
2778 	while (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2779 	    queue->q_exts_by_addr) && num_sios <= 32) {
2780 		ASSERT3U(SIO_GET_OFFSET(sio), >=, rs_get_start(rs,
2781 		    queue->q_exts_by_addr));
2782 		ASSERT3U(SIO_GET_END_OFFSET(sio), <=, rs_get_end(rs,
2783 		    queue->q_exts_by_addr));
2784 
2785 		next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2786 		avl_remove(&queue->q_sios_by_addr, sio);
2787 		queue->q_sio_memused -= SIO_GET_MUSED(sio);
2788 
2789 		bytes_issued += SIO_GET_ASIZE(sio);
2790 		num_sios++;
2791 		list_insert_tail(list, sio);
2792 		sio = next_sio;
2793 	}
2794 
2795 	/*
2796 	 * We limit the number of sios we process at once to 32 to avoid
2797 	 * biting off more than we can chew. If we didn't take everything
2798 	 * in the segment we update it to reflect the work we were able to
2799 	 * complete. Otherwise, we remove it from the range tree entirely.
2800 	 */
2801 	if (sio != NULL && SIO_GET_OFFSET(sio) < rs_get_end(rs,
2802 	    queue->q_exts_by_addr)) {
2803 		range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2804 		    -bytes_issued);
2805 		range_tree_resize_segment(queue->q_exts_by_addr, rs,
2806 		    SIO_GET_OFFSET(sio), rs_get_end(rs,
2807 		    queue->q_exts_by_addr) - SIO_GET_OFFSET(sio));
2808 
2809 		return (B_TRUE);
2810 	} else {
2811 		uint64_t rstart = rs_get_start(rs, queue->q_exts_by_addr);
2812 		uint64_t rend = rs_get_end(rs, queue->q_exts_by_addr);
2813 		range_tree_remove(queue->q_exts_by_addr, rstart, rend - rstart);
2814 		return (B_FALSE);
2815 	}
2816 }
2817 
2818 
2819 /*
2820  * This is called from the queue emptying thread and selects the next
2821  * extent from which we are to issue io's. The behavior of this function
2822  * depends on the state of the scan, the current memory consumption and
2823  * whether or not we are performing a scan shutdown.
2824  * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2825  *	needs to perform a checkpoint
2826  * 2) We select the largest available extent if we are up against the
2827  *	memory limit.
2828  * 3) Otherwise we don't select any extents.
2829  */
2830 static const range_seg_t *
scan_io_queue_fetch_ext(dsl_scan_io_queue_t * queue)2831 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2832 {
2833 	dsl_scan_t *scn = queue->q_scn;
2834 	range_tree_t *rt = queue->q_exts_by_addr;
2835 
2836 	ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2837 	ASSERT(scn->scn_is_sorted);
2838 
2839 	/* handle tunable overrides */
2840 	if (scn->scn_checkpointing || scn->scn_clearing) {
2841 		if (zfs_scan_issue_strategy == 1) {
2842 			return (range_tree_first(rt));
2843 		} else if (zfs_scan_issue_strategy == 2) {
2844 			/*
2845 			 * We need to get the original entry in the by_addr
2846 			 * tree so we can modify it.
2847 			 */
2848 			range_seg_t *size_rs =
2849 			    zfs_btree_first(&queue->q_exts_by_size, NULL);
2850 			if (size_rs == NULL)
2851 				return (NULL);
2852 			uint64_t start = rs_get_start(size_rs, rt);
2853 			uint64_t size = rs_get_end(size_rs, rt) - start;
2854 			range_seg_t *addr_rs = range_tree_find(rt, start,
2855 			    size);
2856 			ASSERT3P(addr_rs, !=, NULL);
2857 			ASSERT3U(rs_get_start(size_rs, rt), ==,
2858 			    rs_get_start(addr_rs, rt));
2859 			ASSERT3U(rs_get_end(size_rs, rt), ==,
2860 			    rs_get_end(addr_rs, rt));
2861 			return (addr_rs);
2862 		}
2863 	}
2864 
2865 	/*
2866 	 * During normal clearing, we want to issue our largest segments
2867 	 * first, keeping IO as sequential as possible, and leaving the
2868 	 * smaller extents for later with the hope that they might eventually
2869 	 * grow to larger sequential segments. However, when the scan is
2870 	 * checkpointing, no new extents will be added to the sorting queue,
2871 	 * so the way we are sorted now is as good as it will ever get.
2872 	 * In this case, we instead switch to issuing extents in LBA order.
2873 	 */
2874 	if (scn->scn_checkpointing) {
2875 		return (range_tree_first(rt));
2876 	} else if (scn->scn_clearing) {
2877 		/*
2878 		 * We need to get the original entry in the by_addr
2879 		 * tree so we can modify it.
2880 		 */
2881 		range_seg_t *size_rs = zfs_btree_first(&queue->q_exts_by_size,
2882 		    NULL);
2883 		if (size_rs == NULL)
2884 			return (NULL);
2885 		uint64_t start = rs_get_start(size_rs, rt);
2886 		uint64_t size = rs_get_end(size_rs, rt) - start;
2887 		range_seg_t *addr_rs = range_tree_find(rt, start, size);
2888 		ASSERT3P(addr_rs, !=, NULL);
2889 		ASSERT3U(rs_get_start(size_rs, rt), ==, rs_get_start(addr_rs,
2890 		    rt));
2891 		ASSERT3U(rs_get_end(size_rs, rt), ==, rs_get_end(addr_rs, rt));
2892 		return (addr_rs);
2893 	} else {
2894 		return (NULL);
2895 	}
2896 }
2897 
2898 static void
scan_io_queues_run_one(void * arg)2899 scan_io_queues_run_one(void *arg)
2900 {
2901 	dsl_scan_io_queue_t *queue = arg;
2902 	kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
2903 	boolean_t suspended = B_FALSE;
2904 	range_seg_t *rs = NULL;
2905 	scan_io_t *sio = NULL;
2906 	list_t sio_list;
2907 	uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
2908 	uint64_t nr_leaves = dsl_scan_count_leaves(queue->q_vd);
2909 
2910 	ASSERT(queue->q_scn->scn_is_sorted);
2911 
2912 	list_create(&sio_list, sizeof (scan_io_t),
2913 	    offsetof(scan_io_t, sio_nodes.sio_list_node));
2914 	mutex_enter(q_lock);
2915 
2916 	/* calculate maximum in-flight bytes for this txg (min 1MB) */
2917 	queue->q_maxinflight_bytes =
2918 	    MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
2919 
2920 	/* reset per-queue scan statistics for this txg */
2921 	queue->q_total_seg_size_this_txg = 0;
2922 	queue->q_segs_this_txg = 0;
2923 	queue->q_total_zio_size_this_txg = 0;
2924 	queue->q_zios_this_txg = 0;
2925 
2926 	/* loop until we have run out of time or sios */
2927 	while ((rs = (range_seg_t *)scan_io_queue_fetch_ext(queue)) != NULL) {
2928 		uint64_t seg_start = 0, seg_end = 0;
2929 		boolean_t more_left = B_TRUE;
2930 
2931 		ASSERT(list_is_empty(&sio_list));
2932 
2933 		/* loop while we still have sios left to process in this rs */
2934 		while (more_left) {
2935 			scan_io_t *first_sio, *last_sio;
2936 
2937 			/*
2938 			 * We have selected which extent needs to be
2939 			 * processed next. Gather up the corresponding sios.
2940 			 */
2941 			more_left = scan_io_queue_gather(queue, rs, &sio_list);
2942 			ASSERT(!list_is_empty(&sio_list));
2943 			first_sio = list_head(&sio_list);
2944 			last_sio = list_tail(&sio_list);
2945 
2946 			seg_end = SIO_GET_END_OFFSET(last_sio);
2947 			if (seg_start == 0)
2948 				seg_start = SIO_GET_OFFSET(first_sio);
2949 
2950 			/*
2951 			 * Issuing sios can take a long time so drop the
2952 			 * queue lock. The sio queue won't be updated by
2953 			 * other threads since we're in syncing context so
2954 			 * we can be sure that our trees will remain exactly
2955 			 * as we left them.
2956 			 */
2957 			mutex_exit(q_lock);
2958 			suspended = scan_io_queue_issue(queue, &sio_list);
2959 			mutex_enter(q_lock);
2960 
2961 			if (suspended)
2962 				break;
2963 		}
2964 		/* update statistics for debugging purposes */
2965 		scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
2966 
2967 		if (suspended)
2968 			break;
2969 	}
2970 
2971 
2972 	/*
2973 	 * If we were suspended in the middle of processing,
2974 	 * requeue any unfinished sios and exit.
2975 	 */
2976 	while ((sio = list_head(&sio_list)) != NULL) {
2977 		list_remove(&sio_list, sio);
2978 		scan_io_queue_insert_impl(queue, sio);
2979 	}
2980 
2981 	mutex_exit(q_lock);
2982 	list_destroy(&sio_list);
2983 }
2984 
2985 /*
2986  * Performs an emptying run on all scan queues in the pool. This just
2987  * punches out one thread per top-level vdev, each of which processes
2988  * only that vdev's scan queue. We can parallelize the I/O here because
2989  * we know that each queue's io's only affect its own top-level vdev.
2990  *
2991  * This function waits for the queue runs to complete, and must be
2992  * called from dsl_scan_sync (or in general, syncing context).
2993  */
2994 static void
scan_io_queues_run(dsl_scan_t * scn)2995 scan_io_queues_run(dsl_scan_t *scn)
2996 {
2997 	spa_t *spa = scn->scn_dp->dp_spa;
2998 
2999 	ASSERT(scn->scn_is_sorted);
3000 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
3001 
3002 	if (scn->scn_bytes_pending == 0)
3003 		return;
3004 
3005 	if (scn->scn_taskq == NULL) {
3006 		char *tq_name = kmem_zalloc(ZFS_MAX_DATASET_NAME_LEN + 16,
3007 		    KM_SLEEP);
3008 		int nthreads = spa->spa_root_vdev->vdev_children;
3009 
3010 		/*
3011 		 * We need to make this taskq *always* execute as many
3012 		 * threads in parallel as we have top-level vdevs and no
3013 		 * less, otherwise strange serialization of the calls to
3014 		 * scan_io_queues_run_one can occur during spa_sync runs
3015 		 * and that significantly impacts performance.
3016 		 */
3017 		(void) snprintf(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16,
3018 		    "dsl_scan_tq_%s", spa->spa_name);
3019 		scn->scn_taskq = taskq_create(tq_name, nthreads, minclsyspri,
3020 		    nthreads, nthreads, TASKQ_PREPOPULATE);
3021 		kmem_free(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16);
3022 	}
3023 
3024 	for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
3025 		vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
3026 
3027 		mutex_enter(&vd->vdev_scan_io_queue_lock);
3028 		if (vd->