xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_scan.c (revision 3b2aab18808792cbd248a12f1edf139b89833c13)
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) 2012 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/dsl_scan.h>
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dnode.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/arc.h>
36 #include <sys/zap.h>
37 #include <sys/zio.h>
38 #include <sys/zfs_context.h>
39 #include <sys/fs/zfs.h>
40 #include <sys/zfs_znode.h>
41 #include <sys/spa_impl.h>
42 #include <sys/vdev_impl.h>
43 #include <sys/zil_impl.h>
44 #include <sys/zio_checksum.h>
45 #include <sys/ddt.h>
46 #include <sys/sa.h>
47 #include <sys/sa_impl.h>
48 #include <sys/zfeature.h>
49 #ifdef _KERNEL
50 #include <sys/zfs_vfsops.h>
51 #endif
52 
53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
54 
55 static scan_cb_t dsl_scan_defrag_cb;
56 static scan_cb_t dsl_scan_scrub_cb;
57 static scan_cb_t dsl_scan_remove_cb;
58 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
59 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
60 
61 int zfs_top_maxinflight = 32;		/* maximum I/Os per top-level */
62 int zfs_resilver_delay = 2;		/* number of ticks to delay resilver */
63 int zfs_scrub_delay = 4;		/* number of ticks to delay scrub */
64 int zfs_scan_idle = 50;			/* idle window in clock ticks */
65 
66 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
67 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
68 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
69 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
70 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */
71 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
72 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
73 
74 #define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
75 	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
76 	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
77 
78 extern int zfs_txg_timeout;
79 
80 /* the order has to match pool_scan_type */
81 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
82 	NULL,
83 	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
84 	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
85 };
86 
87 int
88 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
89 {
90 	int err;
91 	dsl_scan_t *scn;
92 	spa_t *spa = dp->dp_spa;
93 	uint64_t f;
94 
95 	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
96 	scn->scn_dp = dp;
97 
98 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
99 	    "scrub_func", sizeof (uint64_t), 1, &f);
100 	if (err == 0) {
101 		/*
102 		 * There was an old-style scrub in progress.  Restart a
103 		 * new-style scrub from the beginning.
104 		 */
105 		scn->scn_restart_txg = txg;
106 		zfs_dbgmsg("old-style scrub was in progress; "
107 		    "restarting new-style scrub in txg %llu",
108 		    scn->scn_restart_txg);
109 
110 		/*
111 		 * Load the queue obj from the old location so that it
112 		 * can be freed by dsl_scan_done().
113 		 */
114 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
115 		    "scrub_queue", sizeof (uint64_t), 1,
116 		    &scn->scn_phys.scn_queue_obj);
117 	} else {
118 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
119 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
120 		    &scn->scn_phys);
121 		if (err == ENOENT)
122 			return (0);
123 		else if (err)
124 			return (err);
125 
126 		if (scn->scn_phys.scn_state == DSS_SCANNING &&
127 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
128 			/*
129 			 * A new-type scrub was in progress on an old
130 			 * pool, and the pool was accessed by old
131 			 * software.  Restart from the beginning, since
132 			 * the old software may have changed the pool in
133 			 * the meantime.
134 			 */
135 			scn->scn_restart_txg = txg;
136 			zfs_dbgmsg("new-style scrub was modified "
137 			    "by old software; restarting in txg %llu",
138 			    scn->scn_restart_txg);
139 		}
140 	}
141 
142 	spa_scan_stat_init(spa);
143 	return (0);
144 }
145 
146 void
147 dsl_scan_fini(dsl_pool_t *dp)
148 {
149 	if (dp->dp_scan) {
150 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
151 		dp->dp_scan = NULL;
152 	}
153 }
154 
155 /* ARGSUSED */
156 static int
157 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
158 {
159 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
160 
161 	if (scn->scn_phys.scn_state == DSS_SCANNING)
162 		return (EBUSY);
163 
164 	return (0);
165 }
166 
167 static void
168 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
169 {
170 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
171 	pool_scan_func_t *funcp = arg;
172 	dmu_object_type_t ot = 0;
173 	dsl_pool_t *dp = scn->scn_dp;
174 	spa_t *spa = dp->dp_spa;
175 
176 	ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
177 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
178 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
179 	scn->scn_phys.scn_func = *funcp;
180 	scn->scn_phys.scn_state = DSS_SCANNING;
181 	scn->scn_phys.scn_min_txg = 0;
182 	scn->scn_phys.scn_max_txg = tx->tx_txg;
183 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
184 	scn->scn_phys.scn_start_time = gethrestime_sec();
185 	scn->scn_phys.scn_errors = 0;
186 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
187 	scn->scn_restart_txg = 0;
188 	spa_scan_stat_init(spa);
189 
190 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
191 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
192 
193 		/* rewrite all disk labels */
194 		vdev_config_dirty(spa->spa_root_vdev);
195 
196 		if (vdev_resilver_needed(spa->spa_root_vdev,
197 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
198 			spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
199 		} else {
200 			spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
201 		}
202 
203 		spa->spa_scrub_started = B_TRUE;
204 		/*
205 		 * If this is an incremental scrub, limit the DDT scrub phase
206 		 * to just the auto-ditto class (for correctness); the rest
207 		 * of the scrub should go faster using top-down pruning.
208 		 */
209 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
210 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
211 
212 	}
213 
214 	/* back to the generic stuff */
215 
216 	if (dp->dp_blkstats == NULL) {
217 		dp->dp_blkstats =
218 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
219 	}
220 	bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
221 
222 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
223 		ot = DMU_OT_ZAP_OTHER;
224 
225 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
226 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
227 
228 	dsl_scan_sync_state(scn, tx);
229 
230 	spa_history_log_internal(spa, "scan setup", tx,
231 	    "func=%u mintxg=%llu maxtxg=%llu",
232 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
233 }
234 
235 /* ARGSUSED */
236 static void
237 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
238 {
239 	static const char *old_names[] = {
240 		"scrub_bookmark",
241 		"scrub_ddt_bookmark",
242 		"scrub_ddt_class_max",
243 		"scrub_queue",
244 		"scrub_min_txg",
245 		"scrub_max_txg",
246 		"scrub_func",
247 		"scrub_errors",
248 		NULL
249 	};
250 
251 	dsl_pool_t *dp = scn->scn_dp;
252 	spa_t *spa = dp->dp_spa;
253 	int i;
254 
255 	/* Remove any remnants of an old-style scrub. */
256 	for (i = 0; old_names[i]; i++) {
257 		(void) zap_remove(dp->dp_meta_objset,
258 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
259 	}
260 
261 	if (scn->scn_phys.scn_queue_obj != 0) {
262 		VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
263 		    scn->scn_phys.scn_queue_obj, tx));
264 		scn->scn_phys.scn_queue_obj = 0;
265 	}
266 
267 	/*
268 	 * If we were "restarted" from a stopped state, don't bother
269 	 * with anything else.
270 	 */
271 	if (scn->scn_phys.scn_state != DSS_SCANNING)
272 		return;
273 
274 	if (complete)
275 		scn->scn_phys.scn_state = DSS_FINISHED;
276 	else
277 		scn->scn_phys.scn_state = DSS_CANCELED;
278 
279 	spa_history_log_internal(spa, "scan done", tx,
280 	    "complete=%u", complete);
281 
282 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
283 		mutex_enter(&spa->spa_scrub_lock);
284 		while (spa->spa_scrub_inflight > 0) {
285 			cv_wait(&spa->spa_scrub_io_cv,
286 			    &spa->spa_scrub_lock);
287 		}
288 		mutex_exit(&spa->spa_scrub_lock);
289 		spa->spa_scrub_started = B_FALSE;
290 		spa->spa_scrub_active = B_FALSE;
291 
292 		/*
293 		 * If the scrub/resilver completed, update all DTLs to
294 		 * reflect this.  Whether it succeeded or not, vacate
295 		 * all temporary scrub DTLs.
296 		 */
297 		vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
298 		    complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
299 		if (complete) {
300 			spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
301 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
302 		}
303 		spa_errlog_rotate(spa);
304 
305 		/*
306 		 * We may have finished replacing a device.
307 		 * Let the async thread assess this and handle the detach.
308 		 */
309 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
310 	}
311 
312 	scn->scn_phys.scn_end_time = gethrestime_sec();
313 }
314 
315 /* ARGSUSED */
316 static int
317 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
318 {
319 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
320 
321 	if (scn->scn_phys.scn_state != DSS_SCANNING)
322 		return (ENOENT);
323 	return (0);
324 }
325 
326 /* ARGSUSED */
327 static void
328 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
329 {
330 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
331 
332 	dsl_scan_done(scn, B_FALSE, tx);
333 	dsl_scan_sync_state(scn, tx);
334 }
335 
336 int
337 dsl_scan_cancel(dsl_pool_t *dp)
338 {
339 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
340 	    dsl_scan_cancel_sync, NULL, 3));
341 }
342 
343 static void dsl_scan_visitbp(blkptr_t *bp,
344     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
345     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
346     dmu_tx_t *tx);
347 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
348     dmu_objset_type_t ostype,
349     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
350 
351 void
352 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
353 {
354 	zio_free(dp->dp_spa, txg, bp);
355 }
356 
357 void
358 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
359 {
360 	ASSERT(dsl_pool_sync_context(dp));
361 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
362 }
363 
364 static uint64_t
365 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
366 {
367 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
368 	if (dsl_dataset_is_snapshot(ds))
369 		return (MIN(smt, ds->ds_phys->ds_creation_txg));
370 	return (smt);
371 }
372 
373 static void
374 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
375 {
376 	VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
377 	    DMU_POOL_DIRECTORY_OBJECT,
378 	    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
379 	    &scn->scn_phys, tx));
380 }
381 
382 static boolean_t
383 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
384 {
385 	uint64_t elapsed_nanosecs;
386 	int mintime;
387 
388 	/* we never skip user/group accounting objects */
389 	if (zb && (int64_t)zb->zb_object < 0)
390 		return (B_FALSE);
391 
392 	if (scn->scn_pausing)
393 		return (B_TRUE); /* we're already pausing */
394 
395 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
396 		return (B_FALSE); /* we're resuming */
397 
398 	/* We only know how to resume from level-0 blocks. */
399 	if (zb && zb->zb_level != 0)
400 		return (B_FALSE);
401 
402 	mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
403 	    zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
404 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
405 	if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
406 	    (elapsed_nanosecs / MICROSEC > mintime &&
407 	    txg_sync_waiting(scn->scn_dp)) ||
408 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
409 		if (zb) {
410 			dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
411 			    (longlong_t)zb->zb_objset,
412 			    (longlong_t)zb->zb_object,
413 			    (longlong_t)zb->zb_level,
414 			    (longlong_t)zb->zb_blkid);
415 			scn->scn_phys.scn_bookmark = *zb;
416 		}
417 		dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
418 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
419 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
420 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
421 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
422 		scn->scn_pausing = B_TRUE;
423 		return (B_TRUE);
424 	}
425 	return (B_FALSE);
426 }
427 
428 typedef struct zil_scan_arg {
429 	dsl_pool_t	*zsa_dp;
430 	zil_header_t	*zsa_zh;
431 } zil_scan_arg_t;
432 
433 /* ARGSUSED */
434 static int
435 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
436 {
437 	zil_scan_arg_t *zsa = arg;
438 	dsl_pool_t *dp = zsa->zsa_dp;
439 	dsl_scan_t *scn = dp->dp_scan;
440 	zil_header_t *zh = zsa->zsa_zh;
441 	zbookmark_t zb;
442 
443 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
444 		return (0);
445 
446 	/*
447 	 * One block ("stubby") can be allocated a long time ago; we
448 	 * want to visit that one because it has been allocated
449 	 * (on-disk) even if it hasn't been claimed (even though for
450 	 * scrub there's nothing to do to it).
451 	 */
452 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
453 		return (0);
454 
455 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
456 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
457 
458 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
459 	return (0);
460 }
461 
462 /* ARGSUSED */
463 static int
464 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
465 {
466 	if (lrc->lrc_txtype == TX_WRITE) {
467 		zil_scan_arg_t *zsa = arg;
468 		dsl_pool_t *dp = zsa->zsa_dp;
469 		dsl_scan_t *scn = dp->dp_scan;
470 		zil_header_t *zh = zsa->zsa_zh;
471 		lr_write_t *lr = (lr_write_t *)lrc;
472 		blkptr_t *bp = &lr->lr_blkptr;
473 		zbookmark_t zb;
474 
475 		if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
476 			return (0);
477 
478 		/*
479 		 * birth can be < claim_txg if this record's txg is
480 		 * already txg sync'ed (but this log block contains
481 		 * other records that are not synced)
482 		 */
483 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
484 			return (0);
485 
486 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
487 		    lr->lr_foid, ZB_ZIL_LEVEL,
488 		    lr->lr_offset / BP_GET_LSIZE(bp));
489 
490 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
491 	}
492 	return (0);
493 }
494 
495 static void
496 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
497 {
498 	uint64_t claim_txg = zh->zh_claim_txg;
499 	zil_scan_arg_t zsa = { dp, zh };
500 	zilog_t *zilog;
501 
502 	/*
503 	 * We only want to visit blocks that have been claimed but not yet
504 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
505 	 */
506 	if (claim_txg == 0 && spa_writeable(dp->dp_spa))
507 		return;
508 
509 	zilog = zil_alloc(dp->dp_meta_objset, zh);
510 
511 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
512 	    claim_txg);
513 
514 	zil_free(zilog);
515 }
516 
517 /* ARGSUSED */
518 static void
519 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
520     uint64_t objset, uint64_t object, uint64_t blkid)
521 {
522 	zbookmark_t czb;
523 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
524 
525 	if (zfs_no_scrub_prefetch)
526 		return;
527 
528 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
529 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
530 		return;
531 
532 	SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
533 
534 	(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
535 	    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
536 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
537 }
538 
539 static boolean_t
540 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
541     const zbookmark_t *zb)
542 {
543 	/*
544 	 * We never skip over user/group accounting objects (obj<0)
545 	 */
546 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
547 	    (int64_t)zb->zb_object >= 0) {
548 		/*
549 		 * If we already visited this bp & everything below (in
550 		 * a prior txg sync), don't bother doing it again.
551 		 */
552 		if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
553 			return (B_TRUE);
554 
555 		/*
556 		 * If we found the block we're trying to resume from, or
557 		 * we went past it to a different object, zero it out to
558 		 * indicate that it's OK to start checking for pausing
559 		 * again.
560 		 */
561 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
562 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
563 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
564 			    (longlong_t)zb->zb_objset,
565 			    (longlong_t)zb->zb_object,
566 			    (longlong_t)zb->zb_level,
567 			    (longlong_t)zb->zb_blkid);
568 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
569 		}
570 	}
571 	return (B_FALSE);
572 }
573 
574 /*
575  * Return nonzero on i/o error.
576  * Return new buf to write out in *bufp.
577  */
578 static int
579 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
580     dnode_phys_t *dnp, const blkptr_t *bp,
581     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
582 {
583 	dsl_pool_t *dp = scn->scn_dp;
584 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
585 	int err;
586 
587 	if (BP_GET_LEVEL(bp) > 0) {
588 		uint32_t flags = ARC_WAIT;
589 		int i;
590 		blkptr_t *cbp;
591 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
592 
593 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
594 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
595 		if (err) {
596 			scn->scn_phys.scn_errors++;
597 			return (err);
598 		}
599 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
600 			dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
601 			    zb->zb_object, zb->zb_blkid * epb + i);
602 		}
603 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
604 			zbookmark_t czb;
605 
606 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
607 			    zb->zb_level - 1,
608 			    zb->zb_blkid * epb + i);
609 			dsl_scan_visitbp(cbp, &czb, dnp,
610 			    *bufp, ds, scn, ostype, tx);
611 		}
612 	} else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
613 		uint32_t flags = ARC_WAIT;
614 
615 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
616 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
617 		if (err) {
618 			scn->scn_phys.scn_errors++;
619 			return (err);
620 		}
621 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
622 		uint32_t flags = ARC_WAIT;
623 		dnode_phys_t *cdnp;
624 		int i, j;
625 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
626 
627 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
628 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
629 		if (err) {
630 			scn->scn_phys.scn_errors++;
631 			return (err);
632 		}
633 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
634 			for (j = 0; j < cdnp->dn_nblkptr; j++) {
635 				blkptr_t *cbp = &cdnp->dn_blkptr[j];
636 				dsl_scan_prefetch(scn, *bufp, cbp,
637 				    zb->zb_objset, zb->zb_blkid * epb + i, j);
638 			}
639 		}
640 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
641 			dsl_scan_visitdnode(scn, ds, ostype,
642 			    cdnp, *bufp, zb->zb_blkid * epb + i, tx);
643 		}
644 
645 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
646 		uint32_t flags = ARC_WAIT;
647 		objset_phys_t *osp;
648 
649 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
650 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
651 		if (err) {
652 			scn->scn_phys.scn_errors++;
653 			return (err);
654 		}
655 
656 		osp = (*bufp)->b_data;
657 
658 		dsl_scan_visitdnode(scn, ds, osp->os_type,
659 		    &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
660 
661 		if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
662 			/*
663 			 * We also always visit user/group accounting
664 			 * objects, and never skip them, even if we are
665 			 * pausing.  This is necessary so that the space
666 			 * deltas from this txg get integrated.
667 			 */
668 			dsl_scan_visitdnode(scn, ds, osp->os_type,
669 			    &osp->os_groupused_dnode, *bufp,
670 			    DMU_GROUPUSED_OBJECT, tx);
671 			dsl_scan_visitdnode(scn, ds, osp->os_type,
672 			    &osp->os_userused_dnode, *bufp,
673 			    DMU_USERUSED_OBJECT, tx);
674 		}
675 	}
676 
677 	return (0);
678 }
679 
680 static void
681 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
682     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
683     uint64_t object, dmu_tx_t *tx)
684 {
685 	int j;
686 
687 	for (j = 0; j < dnp->dn_nblkptr; j++) {
688 		zbookmark_t czb;
689 
690 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
691 		    dnp->dn_nlevels - 1, j);
692 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
693 		    &czb, dnp, buf, ds, scn, ostype, tx);
694 	}
695 
696 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
697 		zbookmark_t czb;
698 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
699 		    0, DMU_SPILL_BLKID);
700 		dsl_scan_visitbp(&dnp->dn_spill,
701 		    &czb, dnp, buf, ds, scn, ostype, tx);
702 	}
703 }
704 
705 /*
706  * The arguments are in this order because mdb can only print the
707  * first 5; we want them to be useful.
708  */
709 static void
710 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
711     dnode_phys_t *dnp, arc_buf_t *pbuf,
712     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
713     dmu_tx_t *tx)
714 {
715 	dsl_pool_t *dp = scn->scn_dp;
716 	arc_buf_t *buf = NULL;
717 	blkptr_t bp_toread = *bp;
718 
719 	/* ASSERT(pbuf == NULL || arc_released(pbuf)); */
720 
721 	if (dsl_scan_check_pause(scn, zb))
722 		return;
723 
724 	if (dsl_scan_check_resume(scn, dnp, zb))
725 		return;
726 
727 	if (bp->blk_birth == 0)
728 		return;
729 
730 	scn->scn_visited_this_txg++;
731 
732 	dprintf_bp(bp,
733 	    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
734 	    ds, ds ? ds->ds_object : 0,
735 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
736 	    pbuf, bp);
737 
738 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
739 		return;
740 
741 	if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
742 	    &buf) != 0)
743 		return;
744 
745 	/*
746 	 * If dsl_scan_ddt() has aready visited this block, it will have
747 	 * already done any translations or scrubbing, so don't call the
748 	 * callback again.
749 	 */
750 	if (ddt_class_contains(dp->dp_spa,
751 	    scn->scn_phys.scn_ddt_class_max, bp)) {
752 		ASSERT(buf == NULL);
753 		return;
754 	}
755 
756 	/*
757 	 * If this block is from the future (after cur_max_txg), then we
758 	 * are doing this on behalf of a deleted snapshot, and we will
759 	 * revisit the future block on the next pass of this dataset.
760 	 * Don't scan it now unless we need to because something
761 	 * under it was modified.
762 	 */
763 	if (bp->blk_birth <= scn->scn_phys.scn_cur_max_txg) {
764 		scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
765 	}
766 	if (buf)
767 		(void) arc_buf_remove_ref(buf, &buf);
768 }
769 
770 static void
771 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
772     dmu_tx_t *tx)
773 {
774 	zbookmark_t zb;
775 
776 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
777 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
778 	dsl_scan_visitbp(bp, &zb, NULL, NULL,
779 	    ds, scn, DMU_OST_NONE, tx);
780 
781 	dprintf_ds(ds, "finished scan%s", "");
782 }
783 
784 void
785 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
786 {
787 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
788 	dsl_scan_t *scn = dp->dp_scan;
789 	uint64_t mintxg;
790 
791 	if (scn->scn_phys.scn_state != DSS_SCANNING)
792 		return;
793 
794 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
795 		if (dsl_dataset_is_snapshot(ds)) {
796 			/* Note, scn_cur_{min,max}_txg stays the same. */
797 			scn->scn_phys.scn_bookmark.zb_objset =
798 			    ds->ds_phys->ds_next_snap_obj;
799 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
800 			    "reset zb_objset to %llu",
801 			    (u_longlong_t)ds->ds_object,
802 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
803 			scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
804 		} else {
805 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
806 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
807 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
808 			    "reset bookmark to -1,0,0,0",
809 			    (u_longlong_t)ds->ds_object);
810 		}
811 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
812 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
813 		ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
814 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
815 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
816 		if (dsl_dataset_is_snapshot(ds)) {
817 			/*
818 			 * We keep the same mintxg; it could be >
819 			 * ds_creation_txg if the previous snapshot was
820 			 * deleted too.
821 			 */
822 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
823 			    scn->scn_phys.scn_queue_obj,
824 			    ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
825 			zfs_dbgmsg("destroying ds %llu; in queue; "
826 			    "replacing with %llu",
827 			    (u_longlong_t)ds->ds_object,
828 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
829 		} else {
830 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
831 			    (u_longlong_t)ds->ds_object);
832 		}
833 	} else {
834 		zfs_dbgmsg("destroying ds %llu; ignoring",
835 		    (u_longlong_t)ds->ds_object);
836 	}
837 
838 	/*
839 	 * dsl_scan_sync() should be called after this, and should sync
840 	 * out our changed state, but just to be safe, do it here.
841 	 */
842 	dsl_scan_sync_state(scn, tx);
843 }
844 
845 void
846 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
847 {
848 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
849 	dsl_scan_t *scn = dp->dp_scan;
850 	uint64_t mintxg;
851 
852 	if (scn->scn_phys.scn_state != DSS_SCANNING)
853 		return;
854 
855 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
856 
857 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
858 		scn->scn_phys.scn_bookmark.zb_objset =
859 		    ds->ds_phys->ds_prev_snap_obj;
860 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
861 		    "reset zb_objset to %llu",
862 		    (u_longlong_t)ds->ds_object,
863 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
864 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
865 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
866 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
867 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
868 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
869 		    scn->scn_phys.scn_queue_obj,
870 		    ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
871 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
872 		    "replacing with %llu",
873 		    (u_longlong_t)ds->ds_object,
874 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
875 	}
876 	dsl_scan_sync_state(scn, tx);
877 }
878 
879 void
880 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
881 {
882 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
883 	dsl_scan_t *scn = dp->dp_scan;
884 	uint64_t mintxg;
885 
886 	if (scn->scn_phys.scn_state != DSS_SCANNING)
887 		return;
888 
889 	if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
890 		scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
891 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
892 		    "reset zb_objset to %llu",
893 		    (u_longlong_t)ds1->ds_object,
894 		    (u_longlong_t)ds2->ds_object);
895 	} else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
896 		scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
897 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
898 		    "reset zb_objset to %llu",
899 		    (u_longlong_t)ds2->ds_object,
900 		    (u_longlong_t)ds1->ds_object);
901 	}
902 
903 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
904 	    ds1->ds_object, &mintxg) == 0) {
905 		int err;
906 
907 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
908 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
909 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
910 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
911 		err = zap_add_int_key(dp->dp_meta_objset,
912 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
913 		VERIFY(err == 0 || err == EEXIST);
914 		if (err == EEXIST) {
915 			/* Both were there to begin with */
916 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
917 			    scn->scn_phys.scn_queue_obj,
918 			    ds1->ds_object, mintxg, tx));
919 		}
920 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
921 		    "replacing with %llu",
922 		    (u_longlong_t)ds1->ds_object,
923 		    (u_longlong_t)ds2->ds_object);
924 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
925 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
926 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
927 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
928 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
929 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
930 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
931 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
932 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
933 		    "replacing with %llu",
934 		    (u_longlong_t)ds2->ds_object,
935 		    (u_longlong_t)ds1->ds_object);
936 	}
937 
938 	dsl_scan_sync_state(scn, tx);
939 }
940 
941 struct enqueue_clones_arg {
942 	dmu_tx_t *tx;
943 	uint64_t originobj;
944 };
945 
946 /* ARGSUSED */
947 static int
948 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
949 {
950 	struct enqueue_clones_arg *eca = arg;
951 	dsl_dataset_t *ds;
952 	int err;
953 	dsl_scan_t *scn = dp->dp_scan;
954 
955 	if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
956 		return (0);
957 
958 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
959 	if (err)
960 		return (err);
961 
962 	while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
963 		dsl_dataset_t *prev;
964 		err = dsl_dataset_hold_obj(dp,
965 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
966 
967 		dsl_dataset_rele(ds, FTAG);
968 		if (err)
969 			return (err);
970 		ds = prev;
971 	}
972 	VERIFY(zap_add_int_key(dp->dp_meta_objset,
973 	    scn->scn_phys.scn_queue_obj, ds->ds_object,
974 	    ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
975 	dsl_dataset_rele(ds, FTAG);
976 	return (0);
977 }
978 
979 static void
980 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
981 {
982 	dsl_pool_t *dp = scn->scn_dp;
983 	dsl_dataset_t *ds;
984 	objset_t *os;
985 
986 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
987 
988 	if (dmu_objset_from_ds(ds, &os))
989 		goto out;
990 
991 	/*
992 	 * Only the ZIL in the head (non-snapshot) is valid.  Even though
993 	 * snapshots can have ZIL block pointers (which may be the same
994 	 * BP as in the head), they must be ignored.  So we traverse the
995 	 * ZIL here, rather than in scan_recurse(), because the regular
996 	 * snapshot block-sharing rules don't apply to it.
997 	 */
998 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
999 		dsl_scan_zil(dp, &os->os_zil_header);
1000 
1001 	/*
1002 	 * Iterate over the bps in this ds.
1003 	 */
1004 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1005 	dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1006 
1007 	char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1008 	dsl_dataset_name(ds, dsname);
1009 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1010 	    "pausing=%u",
1011 	    (longlong_t)dsobj, dsname,
1012 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
1013 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
1014 	    (int)scn->scn_pausing);
1015 	kmem_free(dsname, ZFS_MAXNAMELEN);
1016 
1017 	if (scn->scn_pausing)
1018 		goto out;
1019 
1020 	/*
1021 	 * We've finished this pass over this dataset.
1022 	 */
1023 
1024 	/*
1025 	 * If we did not completely visit this dataset, do another pass.
1026 	 */
1027 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1028 		zfs_dbgmsg("incomplete pass; visiting again");
1029 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1030 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1031 		    scn->scn_phys.scn_queue_obj, ds->ds_object,
1032 		    scn->scn_phys.scn_cur_max_txg, tx) == 0);
1033 		goto out;
1034 	}
1035 
1036 	/*
1037 	 * Add descendent datasets to work queue.
1038 	 */
1039 	if (ds->ds_phys->ds_next_snap_obj != 0) {
1040 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1041 		    scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1042 		    ds->ds_phys->ds_creation_txg, tx) == 0);
1043 	}
1044 	if (ds->ds_phys->ds_num_children > 1) {
1045 		boolean_t usenext = B_FALSE;
1046 		if (ds->ds_phys->ds_next_clones_obj != 0) {
1047 			uint64_t count;
1048 			/*
1049 			 * A bug in a previous version of the code could
1050 			 * cause upgrade_clones_cb() to not set
1051 			 * ds_next_snap_obj when it should, leading to a
1052 			 * missing entry.  Therefore we can only use the
1053 			 * next_clones_obj when its count is correct.
1054 			 */
1055 			int err = zap_count(dp->dp_meta_objset,
1056 			    ds->ds_phys->ds_next_clones_obj, &count);
1057 			if (err == 0 &&
1058 			    count == ds->ds_phys->ds_num_children - 1)
1059 				usenext = B_TRUE;
1060 		}
1061 
1062 		if (usenext) {
1063 			VERIFY0(zap_join_key(dp->dp_meta_objset,
1064 			    ds->ds_phys->ds_next_clones_obj,
1065 			    scn->scn_phys.scn_queue_obj,
1066 			    ds->ds_phys->ds_creation_txg, tx));
1067 		} else {
1068 			struct enqueue_clones_arg eca;
1069 			eca.tx = tx;
1070 			eca.originobj = ds->ds_object;
1071 
1072 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1073 			    enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1074 		}
1075 	}
1076 
1077 out:
1078 	dsl_dataset_rele(ds, FTAG);
1079 }
1080 
1081 /* ARGSUSED */
1082 static int
1083 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1084 {
1085 	dmu_tx_t *tx = arg;
1086 	dsl_dataset_t *ds;
1087 	int err;
1088 	dsl_scan_t *scn = dp->dp_scan;
1089 
1090 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1091 	if (err)
1092 		return (err);
1093 
1094 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
1095 		dsl_dataset_t *prev;
1096 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1097 		    FTAG, &prev);
1098 		if (err) {
1099 			dsl_dataset_rele(ds, FTAG);
1100 			return (err);
1101 		}
1102 
1103 		/*
1104 		 * If this is a clone, we don't need to worry about it for now.
1105 		 */
1106 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1107 			dsl_dataset_rele(ds, FTAG);
1108 			dsl_dataset_rele(prev, FTAG);
1109 			return (0);
1110 		}
1111 		dsl_dataset_rele(ds, FTAG);
1112 		ds = prev;
1113 	}
1114 
1115 	VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1116 	    ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1117 	dsl_dataset_rele(ds, FTAG);
1118 	return (0);
1119 }
1120 
1121 /*
1122  * Scrub/dedup interaction.
1123  *
1124  * If there are N references to a deduped block, we don't want to scrub it
1125  * N times -- ideally, we should scrub it exactly once.
1126  *
1127  * We leverage the fact that the dde's replication class (enum ddt_class)
1128  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1129  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1130  *
1131  * To prevent excess scrubbing, the scrub begins by walking the DDT
1132  * to find all blocks with refcnt > 1, and scrubs each of these once.
1133  * Since there are two replication classes which contain blocks with
1134  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1135  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1136  *
1137  * There would be nothing more to say if a block's refcnt couldn't change
1138  * during a scrub, but of course it can so we must account for changes
1139  * in a block's replication class.
1140  *
1141  * Here's an example of what can occur:
1142  *
1143  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1144  * when visited during the top-down scrub phase, it will be scrubbed twice.
1145  * This negates our scrub optimization, but is otherwise harmless.
1146  *
1147  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1148  * on each visit during the top-down scrub phase, it will never be scrubbed.
1149  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1150  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1151  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1152  * while a scrub is in progress, it scrubs the block right then.
1153  */
1154 static void
1155 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1156 {
1157 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1158 	ddt_entry_t dde = { 0 };
1159 	int error;
1160 	uint64_t n = 0;
1161 
1162 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1163 		ddt_t *ddt;
1164 
1165 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1166 			break;
1167 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1168 		    (longlong_t)ddb->ddb_class,
1169 		    (longlong_t)ddb->ddb_type,
1170 		    (longlong_t)ddb->ddb_checksum,
1171 		    (longlong_t)ddb->ddb_cursor);
1172 
1173 		/* There should be no pending changes to the dedup table */
1174 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1175 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1176 
1177 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1178 		n++;
1179 
1180 		if (dsl_scan_check_pause(scn, NULL))
1181 			break;
1182 	}
1183 
1184 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1185 	    (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1186 	    (int)scn->scn_pausing);
1187 
1188 	ASSERT(error == 0 || error == ENOENT);
1189 	ASSERT(error != ENOENT ||
1190 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1191 }
1192 
1193 /* ARGSUSED */
1194 void
1195 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1196     ddt_entry_t *dde, dmu_tx_t *tx)
1197 {
1198 	const ddt_key_t *ddk = &dde->dde_key;
1199 	ddt_phys_t *ddp = dde->dde_phys;
1200 	blkptr_t bp;
1201 	zbookmark_t zb = { 0 };
1202 
1203 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1204 		return;
1205 
1206 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1207 		if (ddp->ddp_phys_birth == 0 ||
1208 		    ddp->ddp_phys_birth > scn->scn_phys.scn_cur_max_txg)
1209 			continue;
1210 		ddt_bp_create(checksum, ddk, ddp, &bp);
1211 
1212 		scn->scn_visited_this_txg++;
1213 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1214 	}
1215 }
1216 
1217 static void
1218 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1219 {
1220 	dsl_pool_t *dp = scn->scn_dp;
1221 	zap_cursor_t zc;
1222 	zap_attribute_t za;
1223 
1224 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1225 	    scn->scn_phys.scn_ddt_class_max) {
1226 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1227 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1228 		dsl_scan_ddt(scn, tx);
1229 		if (scn->scn_pausing)
1230 			return;
1231 	}
1232 
1233 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1234 		/* First do the MOS & ORIGIN */
1235 
1236 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1237 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1238 		dsl_scan_visit_rootbp(scn, NULL,
1239 		    &dp->dp_meta_rootbp, tx);
1240 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1241 		if (scn->scn_pausing)
1242 			return;
1243 
1244 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1245 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1246 			    enqueue_cb, tx, DS_FIND_CHILDREN));
1247 		} else {
1248 			dsl_scan_visitds(scn,
1249 			    dp->dp_origin_snap->ds_object, tx);
1250 		}
1251 		ASSERT(!scn->scn_pausing);
1252 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
1253 	    ZB_DESTROYED_OBJSET) {
1254 		/*
1255 		 * If we were paused, continue from here.  Note if the
1256 		 * ds we were paused on was deleted, the zb_objset may
1257 		 * be -1, so we will skip this and find a new objset
1258 		 * below.
1259 		 */
1260 		dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1261 		if (scn->scn_pausing)
1262 			return;
1263 	}
1264 
1265 	/*
1266 	 * In case we were paused right at the end of the ds, zero the
1267 	 * bookmark so we don't think that we're still trying to resume.
1268 	 */
1269 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1270 
1271 	/* keep pulling things out of the zap-object-as-queue */
1272 	while (zap_cursor_init(&zc, dp->dp_meta_objset,
1273 	    scn->scn_phys.scn_queue_obj),
1274 	    zap_cursor_retrieve(&zc, &za) == 0) {
1275 		dsl_dataset_t *ds;
1276 		uint64_t dsobj;
1277 
1278 		dsobj = strtonum(za.za_name, NULL);
1279 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1280 		    scn->scn_phys.scn_queue_obj, dsobj, tx));
1281 
1282 		/* Set up min/max txg */
1283 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1284 		if (za.za_first_integer != 0) {
1285 			scn->scn_phys.scn_cur_min_txg =
1286 			    MAX(scn->scn_phys.scn_min_txg,
1287 			    za.za_first_integer);
1288 		} else {
1289 			scn->scn_phys.scn_cur_min_txg =
1290 			    MAX(scn->scn_phys.scn_min_txg,
1291 			    ds->ds_phys->ds_prev_snap_txg);
1292 		}
1293 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1294 		dsl_dataset_rele(ds, FTAG);
1295 
1296 		dsl_scan_visitds(scn, dsobj, tx);
1297 		zap_cursor_fini(&zc);
1298 		if (scn->scn_pausing)
1299 			return;
1300 	}
1301 	zap_cursor_fini(&zc);
1302 }
1303 
1304 static boolean_t
1305 dsl_scan_free_should_pause(dsl_scan_t *scn)
1306 {
1307 	uint64_t elapsed_nanosecs;
1308 
1309 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1310 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1311 	    (elapsed_nanosecs / MICROSEC > zfs_free_min_time_ms &&
1312 	    txg_sync_waiting(scn->scn_dp)) ||
1313 	    spa_shutting_down(scn->scn_dp->dp_spa));
1314 }
1315 
1316 static int
1317 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1318 {
1319 	dsl_scan_t *scn = arg;
1320 
1321 	if (!scn->scn_is_bptree ||
1322 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1323 		if (dsl_scan_free_should_pause(scn))
1324 			return (ERESTART);
1325 	}
1326 
1327 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1328 	    dmu_tx_get_txg(tx), bp, 0));
1329 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1330 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1331 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1332 	scn->scn_visited_this_txg++;
1333 	return (0);
1334 }
1335 
1336 boolean_t
1337 dsl_scan_active(dsl_scan_t *scn)
1338 {
1339 	spa_t *spa = scn->scn_dp->dp_spa;
1340 	uint64_t used = 0, comp, uncomp;
1341 
1342 	if (spa->spa_load_state != SPA_LOAD_NONE)
1343 		return (B_FALSE);
1344 	if (spa_shutting_down(spa))
1345 		return (B_FALSE);
1346 
1347 	if (scn->scn_phys.scn_state == DSS_SCANNING)
1348 		return (B_TRUE);
1349 
1350 	if (spa_feature_is_active(spa,
1351 	    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1352 		return (B_TRUE);
1353 	}
1354 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1355 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1356 		    &used, &comp, &uncomp);
1357 	}
1358 	return (used != 0);
1359 }
1360 
1361 void
1362 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1363 {
1364 	dsl_scan_t *scn = dp->dp_scan;
1365 	spa_t *spa = dp->dp_spa;
1366 	int err;
1367 
1368 	/*
1369 	 * Check for scn_restart_txg before checking spa_load_state, so
1370 	 * that we can restart an old-style scan while the pool is being
1371 	 * imported (see dsl_scan_init).
1372 	 */
1373 	if (scn->scn_restart_txg != 0 &&
1374 	    scn->scn_restart_txg <= tx->tx_txg) {
1375 		pool_scan_func_t func = POOL_SCAN_SCRUB;
1376 		dsl_scan_done(scn, B_FALSE, tx);
1377 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1378 			func = POOL_SCAN_RESILVER;
1379 		zfs_dbgmsg("restarting scan func=%u txg=%llu",
1380 		    func, tx->tx_txg);
1381 		dsl_scan_setup_sync(&func, tx);
1382 	}
1383 
1384 	if (!dsl_scan_active(scn) ||
1385 	    spa_sync_pass(dp->dp_spa) > 1)
1386 		return;
1387 
1388 	scn->scn_visited_this_txg = 0;
1389 	scn->scn_pausing = B_FALSE;
1390 	scn->scn_sync_start_time = gethrtime();
1391 	spa->spa_scrub_active = B_TRUE;
1392 
1393 	/*
1394 	 * First process the free list.  If we pause the free, don't do
1395 	 * any scanning.  This ensures that there is no free list when
1396 	 * we are scanning, so the scan code doesn't have to worry about
1397 	 * traversing it.
1398 	 */
1399 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1400 		scn->scn_is_bptree = B_FALSE;
1401 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1402 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1403 		err = bpobj_iterate(&dp->dp_free_bpobj,
1404 		    dsl_scan_free_block_cb, scn, tx);
1405 		VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1406 
1407 		if (err == 0 && spa_feature_is_active(spa,
1408 		    &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
1409 			scn->scn_is_bptree = B_TRUE;
1410 			scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1411 			    NULL, ZIO_FLAG_MUSTSUCCEED);
1412 			err = bptree_iterate(dp->dp_meta_objset,
1413 			    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1414 			    scn, tx);
1415 			VERIFY0(zio_wait(scn->scn_zio_root));
1416 
1417 			if (err == 0) {
1418 				zfeature_info_t *feat = &spa_feature_table
1419 				    [SPA_FEATURE_ASYNC_DESTROY];
1420 				/* finished; deactivate async destroy feature */
1421 				spa_feature_decr(spa, feat, tx);
1422 				ASSERT(!spa_feature_is_active(spa, feat));
1423 				VERIFY0(zap_remove(dp->dp_meta_objset,
1424 				    DMU_POOL_DIRECTORY_OBJECT,
1425 				    DMU_POOL_BPTREE_OBJ, tx));
1426 				VERIFY0(bptree_free(dp->dp_meta_objset,
1427 				    dp->dp_bptree_obj, tx));
1428 				dp->dp_bptree_obj = 0;
1429 			}
1430 		}
1431 		if (scn->scn_visited_this_txg) {
1432 			zfs_dbgmsg("freed %llu blocks in %llums from "
1433 			    "free_bpobj/bptree txg %llu",
1434 			    (longlong_t)scn->scn_visited_this_txg,
1435 			    (longlong_t)
1436 			    (gethrtime() - scn->scn_sync_start_time) / MICROSEC,
1437 			    (longlong_t)tx->tx_txg);
1438 			scn->scn_visited_this_txg = 0;
1439 			/*
1440 			 * Re-sync the ddt so that we can further modify
1441 			 * it when doing bprewrite.
1442 			 */
1443 			ddt_sync(spa, tx->tx_txg);
1444 		}
1445 		if (err == ERESTART)
1446 			return;
1447 	}
1448 
1449 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1450 		return;
1451 
1452 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1453 	    scn->scn_phys.scn_ddt_class_max) {
1454 		zfs_dbgmsg("doing scan sync txg %llu; "
1455 		    "ddt bm=%llu/%llu/%llu/%llx",
1456 		    (longlong_t)tx->tx_txg,
1457 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1458 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1459 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1460 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1461 		ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1462 		ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1463 		ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1464 		ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1465 	} else {
1466 		zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1467 		    (longlong_t)tx->tx_txg,
1468 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1469 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1470 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1471 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1472 	}
1473 
1474 	scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1475 	    NULL, ZIO_FLAG_CANFAIL);
1476 	dsl_pool_config_enter(dp, FTAG);
1477 	dsl_scan_visit(scn, tx);
1478 	dsl_pool_config_exit(dp, FTAG);
1479 	(void) zio_wait(scn->scn_zio_root);
1480 	scn->scn_zio_root = NULL;
1481 
1482 	zfs_dbgmsg("visited %llu blocks in %llums",
1483 	    (longlong_t)scn->scn_visited_this_txg,
1484 	    (longlong_t)(gethrtime() - scn->scn_sync_start_time) / MICROSEC);
1485 
1486 	if (!scn->scn_pausing) {
1487 		/* finished with scan. */
1488 		zfs_dbgmsg("finished scan txg %llu", (longlong_t)tx->tx_txg);
1489 		dsl_scan_done(scn, B_TRUE, tx);
1490 	}
1491 
1492 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1493 		mutex_enter(&spa->spa_scrub_lock);
1494 		while (spa->spa_scrub_inflight > 0) {
1495 			cv_wait(&spa->spa_scrub_io_cv,
1496 			    &spa->spa_scrub_lock);
1497 		}
1498 		mutex_exit(&spa->spa_scrub_lock);
1499 	}
1500 
1501 	dsl_scan_sync_state(scn, tx);
1502 }
1503 
1504 /*
1505  * This will start a new scan, or restart an existing one.
1506  */
1507 void
1508 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1509 {
1510 	if (txg == 0) {
1511 		dmu_tx_t *tx;
1512 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1513 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1514 
1515 		txg = dmu_tx_get_txg(tx);
1516 		dp->dp_scan->scn_restart_txg = txg;
1517 		dmu_tx_commit(tx);
1518 	} else {
1519 		dp->dp_scan->scn_restart_txg = txg;
1520 	}
1521 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1522 }
1523 
1524 boolean_t
1525 dsl_scan_resilvering(dsl_pool_t *dp)
1526 {
1527 	return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1528 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1529 }
1530 
1531 /*
1532  * scrub consumers
1533  */
1534 
1535 static void
1536 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1537 {
1538 	int i;
1539 
1540 	/*
1541 	 * If we resume after a reboot, zab will be NULL; don't record
1542 	 * incomplete stats in that case.
1543 	 */
1544 	if (zab == NULL)
1545 		return;
1546 
1547 	for (i = 0; i < 4; i++) {
1548 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1549 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1550 		if (t & DMU_OT_NEWTYPE)
1551 			t = DMU_OT_OTHER;
1552 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
1553 		int equal;
1554 
1555 		zb->zb_count++;
1556 		zb->zb_asize += BP_GET_ASIZE(bp);
1557 		zb->zb_lsize += BP_GET_LSIZE(bp);
1558 		zb->zb_psize += BP_GET_PSIZE(bp);
1559 		zb->zb_gangs += BP_COUNT_GANG(bp);
1560 
1561 		switch (BP_GET_NDVAS(bp)) {
1562 		case 2:
1563 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1564 			    DVA_GET_VDEV(&bp->blk_dva[1]))
1565 				zb->zb_ditto_2_of_2_samevdev++;
1566 			break;
1567 		case 3:
1568 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1569 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
1570 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1571 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
1572 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1573 			    DVA_GET_VDEV(&bp->blk_dva[2]));
1574 			if (equal == 1)
1575 				zb->zb_ditto_2_of_3_samevdev++;
1576 			else if (equal == 3)
1577 				zb->zb_ditto_3_of_3_samevdev++;
1578 			break;
1579 		}
1580 	}
1581 }
1582 
1583 static void
1584 dsl_scan_scrub_done(zio_t *zio)
1585 {
1586 	spa_t *spa = zio->io_spa;
1587 
1588 	zio_data_buf_free(zio->io_data, zio->io_size);
1589 
1590 	mutex_enter(&spa->spa_scrub_lock);
1591 	spa->spa_scrub_inflight--;
1592 	cv_broadcast(&spa->spa_scrub_io_cv);
1593 
1594 	if (zio->io_error && (zio->io_error != ECKSUM ||
1595 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1596 		spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1597 	}
1598 	mutex_exit(&spa->spa_scrub_lock);
1599 }
1600 
1601 static int
1602 dsl_scan_scrub_cb(dsl_pool_t *dp,
1603     const blkptr_t *bp, const zbookmark_t *zb)
1604 {
1605 	dsl_scan_t *scn = dp->dp_scan;
1606 	size_t size = BP_GET_PSIZE(bp);
1607 	spa_t *spa = dp->dp_spa;
1608 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1609 	boolean_t needs_io;
1610 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1611 	int zio_priority;
1612 	int scan_delay = 0;
1613 
1614 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
1615 	    phys_birth >= scn->scn_phys.scn_max_txg)
1616 		return (0);
1617 
1618 	count_block(dp->dp_blkstats, bp);
1619 
1620 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1621 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1622 		zio_flags |= ZIO_FLAG_SCRUB;
1623 		zio_priority = ZIO_PRIORITY_SCRUB;
1624 		needs_io = B_TRUE;
1625 		scan_delay = zfs_scrub_delay;
1626 	} else {
1627 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1628 		zio_flags |= ZIO_FLAG_RESILVER;
1629 		zio_priority = ZIO_PRIORITY_RESILVER;
1630 		needs_io = B_FALSE;
1631 		scan_delay = zfs_resilver_delay;
1632 	}
1633 
1634 	/* If it's an intent log block, failure is expected. */
1635 	if (zb->zb_level == ZB_ZIL_LEVEL)
1636 		zio_flags |= ZIO_FLAG_SPECULATIVE;
1637 
1638 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1639 		vdev_t *vd = vdev_lookup_top(spa,
1640 		    DVA_GET_VDEV(&bp->blk_dva[d]));
1641 
1642 		/*
1643 		 * Keep track of how much data we've examined so that
1644 		 * zpool(1M) status can make useful progress reports.
1645 		 */
1646 		scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1647 		spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1648 
1649 		/* if it's a resilver, this may not be in the target range */
1650 		if (!needs_io) {
1651 			if (DVA_GET_GANG(&bp->blk_dva[d])) {
1652 				/*
1653 				 * Gang members may be spread across multiple
1654 				 * vdevs, so the best estimate we have is the
1655 				 * scrub range, which has already been checked.
1656 				 * XXX -- it would be better to change our
1657 				 * allocation policy to ensure that all
1658 				 * gang members reside on the same vdev.
1659 				 */
1660 				needs_io = B_TRUE;
1661 			} else {
1662 				needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1663 				    phys_birth, 1);
1664 			}
1665 		}
1666 	}
1667 
1668 	if (needs_io && !zfs_no_scrub_io) {
1669 		vdev_t *rvd = spa->spa_root_vdev;
1670 		uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1671 		void *data = zio_data_buf_alloc(size);
1672 
1673 		mutex_enter(&spa->spa_scrub_lock);
1674 		while (spa->spa_scrub_inflight >= maxinflight)
1675 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1676 		spa->spa_scrub_inflight++;
1677 		mutex_exit(&spa->spa_scrub_lock);
1678 
1679 		/*
1680 		 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1681 		 * then throttle our workload to limit the impact of a scan.
1682 		 */
1683 		if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1684 			delay(scan_delay);
1685 
1686 		zio_nowait(zio_read(NULL, spa, bp, data, size,
1687 		    dsl_scan_scrub_done, NULL, zio_priority,
1688 		    zio_flags, zb));
1689 	}
1690 
1691 	/* do not relocate this block */
1692 	return (0);
1693 }
1694 
1695 int
1696 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1697 {
1698 	spa_t *spa = dp->dp_spa;
1699 
1700 	/*
1701 	 * Purge all vdev caches and probe all devices.  We do this here
1702 	 * rather than in sync context because this requires a writer lock
1703 	 * on the spa_config lock, which we can't do from sync context.  The
1704 	 * spa_scrub_reopen flag indicates that vdev_open() should not
1705 	 * attempt to start another scrub.
1706 	 */
1707 	spa_vdev_state_enter(spa, SCL_NONE);
1708 	spa->spa_scrub_reopen = B_TRUE;
1709 	vdev_reopen(spa->spa_root_vdev);
1710 	spa->spa_scrub_reopen = B_FALSE;
1711 	(void) spa_vdev_state_exit(spa, NULL, 0);
1712 
1713 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1714 	    dsl_scan_setup_sync, &func, 0));
1715 }
1716