xref: /titanic_41/usr/src/uts/common/fs/zfs/dsl_scan.c (revision 2208104ea4cd6d9d44c2a0c21cf2479b92aacf08)
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, 2014 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 *,
54     const zbookmark_phys_t *);
55 
56 static scan_cb_t dsl_scan_scrub_cb;
57 static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
58 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
59 
60 int zfs_top_maxinflight = 32;		/* maximum I/Os per top-level */
61 int zfs_resilver_delay = 2;		/* number of ticks to delay resilver */
62 int zfs_scrub_delay = 4;		/* number of ticks to delay scrub */
63 int zfs_scan_idle = 50;			/* idle window in clock ticks */
64 
65 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
68 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
69 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
70 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
71 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
72 /* max number of blocks to free in a single TXG */
73 uint64_t zfs_free_max_blocks = UINT64_MAX;
74 
75 #define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
76 	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
77 	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
78 
79 extern int zfs_txg_timeout;
80 
81 /* the order has to match pool_scan_type */
82 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
83 	NULL,
84 	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
85 	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
86 };
87 
88 int
89 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
90 {
91 	int err;
92 	dsl_scan_t *scn;
93 	spa_t *spa = dp->dp_spa;
94 	uint64_t f;
95 
96 	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
97 	scn->scn_dp = dp;
98 
99 	/*
100 	 * It's possible that we're resuming a scan after a reboot so
101 	 * make sure that the scan_async_destroying flag is initialized
102 	 * appropriately.
103 	 */
104 	ASSERT(!scn->scn_async_destroying);
105 	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
106 	    SPA_FEATURE_ASYNC_DESTROY);
107 
108 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
109 	    "scrub_func", sizeof (uint64_t), 1, &f);
110 	if (err == 0) {
111 		/*
112 		 * There was an old-style scrub in progress.  Restart a
113 		 * new-style scrub from the beginning.
114 		 */
115 		scn->scn_restart_txg = txg;
116 		zfs_dbgmsg("old-style scrub was in progress; "
117 		    "restarting new-style scrub in txg %llu",
118 		    scn->scn_restart_txg);
119 
120 		/*
121 		 * Load the queue obj from the old location so that it
122 		 * can be freed by dsl_scan_done().
123 		 */
124 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
125 		    "scrub_queue", sizeof (uint64_t), 1,
126 		    &scn->scn_phys.scn_queue_obj);
127 	} else {
128 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
129 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
130 		    &scn->scn_phys);
131 		if (err == ENOENT)
132 			return (0);
133 		else if (err)
134 			return (err);
135 
136 		if (scn->scn_phys.scn_state == DSS_SCANNING &&
137 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
138 			/*
139 			 * A new-type scrub was in progress on an old
140 			 * pool, and the pool was accessed by old
141 			 * software.  Restart from the beginning, since
142 			 * the old software may have changed the pool in
143 			 * the meantime.
144 			 */
145 			scn->scn_restart_txg = txg;
146 			zfs_dbgmsg("new-style scrub was modified "
147 			    "by old software; restarting in txg %llu",
148 			    scn->scn_restart_txg);
149 		}
150 	}
151 
152 	spa_scan_stat_init(spa);
153 	return (0);
154 }
155 
156 void
157 dsl_scan_fini(dsl_pool_t *dp)
158 {
159 	if (dp->dp_scan) {
160 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
161 		dp->dp_scan = NULL;
162 	}
163 }
164 
165 /* ARGSUSED */
166 static int
167 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
168 {
169 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
170 
171 	if (scn->scn_phys.scn_state == DSS_SCANNING)
172 		return (SET_ERROR(EBUSY));
173 
174 	return (0);
175 }
176 
177 static void
178 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
179 {
180 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
181 	pool_scan_func_t *funcp = arg;
182 	dmu_object_type_t ot = 0;
183 	dsl_pool_t *dp = scn->scn_dp;
184 	spa_t *spa = dp->dp_spa;
185 
186 	ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
187 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
188 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
189 	scn->scn_phys.scn_func = *funcp;
190 	scn->scn_phys.scn_state = DSS_SCANNING;
191 	scn->scn_phys.scn_min_txg = 0;
192 	scn->scn_phys.scn_max_txg = tx->tx_txg;
193 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
194 	scn->scn_phys.scn_start_time = gethrestime_sec();
195 	scn->scn_phys.scn_errors = 0;
196 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
197 	scn->scn_restart_txg = 0;
198 	scn->scn_done_txg = 0;
199 	spa_scan_stat_init(spa);
200 
201 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
202 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
203 
204 		/* rewrite all disk labels */
205 		vdev_config_dirty(spa->spa_root_vdev);
206 
207 		if (vdev_resilver_needed(spa->spa_root_vdev,
208 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
209 			spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
210 		} else {
211 			spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
212 		}
213 
214 		spa->spa_scrub_started = B_TRUE;
215 		/*
216 		 * If this is an incremental scrub, limit the DDT scrub phase
217 		 * to just the auto-ditto class (for correctness); the rest
218 		 * of the scrub should go faster using top-down pruning.
219 		 */
220 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
221 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
222 
223 	}
224 
225 	/* back to the generic stuff */
226 
227 	if (dp->dp_blkstats == NULL) {
228 		dp->dp_blkstats =
229 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
230 	}
231 	bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
232 
233 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
234 		ot = DMU_OT_ZAP_OTHER;
235 
236 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
237 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
238 
239 	dsl_scan_sync_state(scn, tx);
240 
241 	spa_history_log_internal(spa, "scan setup", tx,
242 	    "func=%u mintxg=%llu maxtxg=%llu",
243 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
244 }
245 
246 /* ARGSUSED */
247 static void
248 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
249 {
250 	static const char *old_names[] = {
251 		"scrub_bookmark",
252 		"scrub_ddt_bookmark",
253 		"scrub_ddt_class_max",
254 		"scrub_queue",
255 		"scrub_min_txg",
256 		"scrub_max_txg",
257 		"scrub_func",
258 		"scrub_errors",
259 		NULL
260 	};
261 
262 	dsl_pool_t *dp = scn->scn_dp;
263 	spa_t *spa = dp->dp_spa;
264 	int i;
265 
266 	/* Remove any remnants of an old-style scrub. */
267 	for (i = 0; old_names[i]; i++) {
268 		(void) zap_remove(dp->dp_meta_objset,
269 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
270 	}
271 
272 	if (scn->scn_phys.scn_queue_obj != 0) {
273 		VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
274 		    scn->scn_phys.scn_queue_obj, tx));
275 		scn->scn_phys.scn_queue_obj = 0;
276 	}
277 
278 	/*
279 	 * If we were "restarted" from a stopped state, don't bother
280 	 * with anything else.
281 	 */
282 	if (scn->scn_phys.scn_state != DSS_SCANNING)
283 		return;
284 
285 	if (complete)
286 		scn->scn_phys.scn_state = DSS_FINISHED;
287 	else
288 		scn->scn_phys.scn_state = DSS_CANCELED;
289 
290 	spa_history_log_internal(spa, "scan done", tx,
291 	    "complete=%u", complete);
292 
293 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
294 		mutex_enter(&spa->spa_scrub_lock);
295 		while (spa->spa_scrub_inflight > 0) {
296 			cv_wait(&spa->spa_scrub_io_cv,
297 			    &spa->spa_scrub_lock);
298 		}
299 		mutex_exit(&spa->spa_scrub_lock);
300 		spa->spa_scrub_started = B_FALSE;
301 		spa->spa_scrub_active = B_FALSE;
302 
303 		/*
304 		 * If the scrub/resilver completed, update all DTLs to
305 		 * reflect this.  Whether it succeeded or not, vacate
306 		 * all temporary scrub DTLs.
307 		 */
308 		vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
309 		    complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
310 		if (complete) {
311 			spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
312 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
313 		}
314 		spa_errlog_rotate(spa);
315 
316 		/*
317 		 * We may have finished replacing a device.
318 		 * Let the async thread assess this and handle the detach.
319 		 */
320 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
321 	}
322 
323 	scn->scn_phys.scn_end_time = gethrestime_sec();
324 }
325 
326 /* ARGSUSED */
327 static int
328 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
329 {
330 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
331 
332 	if (scn->scn_phys.scn_state != DSS_SCANNING)
333 		return (SET_ERROR(ENOENT));
334 	return (0);
335 }
336 
337 /* ARGSUSED */
338 static void
339 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
340 {
341 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
342 
343 	dsl_scan_done(scn, B_FALSE, tx);
344 	dsl_scan_sync_state(scn, tx);
345 }
346 
347 int
348 dsl_scan_cancel(dsl_pool_t *dp)
349 {
350 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
351 	    dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
352 }
353 
354 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
355     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
356     dmu_objset_type_t ostype, dmu_tx_t *tx);
357 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
358     dmu_objset_type_t ostype,
359     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
360 
361 void
362 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
363 {
364 	zio_free(dp->dp_spa, txg, bp);
365 }
366 
367 void
368 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
369 {
370 	ASSERT(dsl_pool_sync_context(dp));
371 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
372 }
373 
374 static uint64_t
375 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
376 {
377 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
378 	if (ds->ds_is_snapshot)
379 		return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
380 	return (smt);
381 }
382 
383 static void
384 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
385 {
386 	VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
387 	    DMU_POOL_DIRECTORY_OBJECT,
388 	    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
389 	    &scn->scn_phys, tx));
390 }
391 
392 extern int zfs_vdev_async_write_active_min_dirty_percent;
393 
394 static boolean_t
395 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
396 {
397 	/* we never skip user/group accounting objects */
398 	if (zb && (int64_t)zb->zb_object < 0)
399 		return (B_FALSE);
400 
401 	if (scn->scn_pausing)
402 		return (B_TRUE); /* we're already pausing */
403 
404 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
405 		return (B_FALSE); /* we're resuming */
406 
407 	/* We only know how to resume from level-0 blocks. */
408 	if (zb && zb->zb_level != 0)
409 		return (B_FALSE);
410 
411 	/*
412 	 * We pause if:
413 	 *  - we have scanned for the maximum time: an entire txg
414 	 *    timeout (default 5 sec)
415 	 *  or
416 	 *  - we have scanned for at least the minimum time (default 1 sec
417 	 *    for scrub, 3 sec for resilver), and either we have sufficient
418 	 *    dirty data that we are starting to write more quickly
419 	 *    (default 30%), or someone is explicitly waiting for this txg
420 	 *    to complete.
421 	 *  or
422 	 *  - the spa is shutting down because this pool is being exported
423 	 *    or the machine is rebooting.
424 	 */
425 	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
426 	    zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
427 	uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
428 	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
429 	if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
430 	    (NSEC2MSEC(elapsed_nanosecs) > mintime &&
431 	    (txg_sync_waiting(scn->scn_dp) ||
432 	    dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
433 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
434 		if (zb) {
435 			dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
436 			    (longlong_t)zb->zb_objset,
437 			    (longlong_t)zb->zb_object,
438 			    (longlong_t)zb->zb_level,
439 			    (longlong_t)zb->zb_blkid);
440 			scn->scn_phys.scn_bookmark = *zb;
441 		}
442 		dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
443 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
444 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
445 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
446 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
447 		scn->scn_pausing = B_TRUE;
448 		return (B_TRUE);
449 	}
450 	return (B_FALSE);
451 }
452 
453 typedef struct zil_scan_arg {
454 	dsl_pool_t	*zsa_dp;
455 	zil_header_t	*zsa_zh;
456 } zil_scan_arg_t;
457 
458 /* ARGSUSED */
459 static int
460 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
461 {
462 	zil_scan_arg_t *zsa = arg;
463 	dsl_pool_t *dp = zsa->zsa_dp;
464 	dsl_scan_t *scn = dp->dp_scan;
465 	zil_header_t *zh = zsa->zsa_zh;
466 	zbookmark_phys_t zb;
467 
468 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
469 		return (0);
470 
471 	/*
472 	 * One block ("stubby") can be allocated a long time ago; we
473 	 * want to visit that one because it has been allocated
474 	 * (on-disk) even if it hasn't been claimed (even though for
475 	 * scrub there's nothing to do to it).
476 	 */
477 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
478 		return (0);
479 
480 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
481 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
482 
483 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
484 	return (0);
485 }
486 
487 /* ARGSUSED */
488 static int
489 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
490 {
491 	if (lrc->lrc_txtype == TX_WRITE) {
492 		zil_scan_arg_t *zsa = arg;
493 		dsl_pool_t *dp = zsa->zsa_dp;
494 		dsl_scan_t *scn = dp->dp_scan;
495 		zil_header_t *zh = zsa->zsa_zh;
496 		lr_write_t *lr = (lr_write_t *)lrc;
497 		blkptr_t *bp = &lr->lr_blkptr;
498 		zbookmark_phys_t zb;
499 
500 		if (BP_IS_HOLE(bp) ||
501 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
502 			return (0);
503 
504 		/*
505 		 * birth can be < claim_txg if this record's txg is
506 		 * already txg sync'ed (but this log block contains
507 		 * other records that are not synced)
508 		 */
509 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
510 			return (0);
511 
512 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
513 		    lr->lr_foid, ZB_ZIL_LEVEL,
514 		    lr->lr_offset / BP_GET_LSIZE(bp));
515 
516 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
517 	}
518 	return (0);
519 }
520 
521 static void
522 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
523 {
524 	uint64_t claim_txg = zh->zh_claim_txg;
525 	zil_scan_arg_t zsa = { dp, zh };
526 	zilog_t *zilog;
527 
528 	/*
529 	 * We only want to visit blocks that have been claimed but not yet
530 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
531 	 */
532 	if (claim_txg == 0 && spa_writeable(dp->dp_spa))
533 		return;
534 
535 	zilog = zil_alloc(dp->dp_meta_objset, zh);
536 
537 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
538 	    claim_txg);
539 
540 	zil_free(zilog);
541 }
542 
543 /* ARGSUSED */
544 static void
545 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
546     uint64_t objset, uint64_t object, uint64_t blkid)
547 {
548 	zbookmark_phys_t czb;
549 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
550 
551 	if (zfs_no_scrub_prefetch)
552 		return;
553 
554 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
555 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
556 		return;
557 
558 	SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
559 
560 	(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
561 	    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
562 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
563 }
564 
565 static boolean_t
566 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
567     const zbookmark_phys_t *zb)
568 {
569 	/*
570 	 * We never skip over user/group accounting objects (obj<0)
571 	 */
572 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
573 	    (int64_t)zb->zb_object >= 0) {
574 		/*
575 		 * If we already visited this bp & everything below (in
576 		 * a prior txg sync), don't bother doing it again.
577 		 */
578 		if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
579 			return (B_TRUE);
580 
581 		/*
582 		 * If we found the block we're trying to resume from, or
583 		 * we went past it to a different object, zero it out to
584 		 * indicate that it's OK to start checking for pausing
585 		 * again.
586 		 */
587 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
588 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
589 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
590 			    (longlong_t)zb->zb_objset,
591 			    (longlong_t)zb->zb_object,
592 			    (longlong_t)zb->zb_level,
593 			    (longlong_t)zb->zb_blkid);
594 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
595 		}
596 	}
597 	return (B_FALSE);
598 }
599 
600 /*
601  * Return nonzero on i/o error.
602  * Return new buf to write out in *bufp.
603  */
604 static int
605 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
606     dnode_phys_t *dnp, const blkptr_t *bp,
607     const zbookmark_phys_t *zb, dmu_tx_t *tx)
608 {
609 	dsl_pool_t *dp = scn->scn_dp;
610 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
611 	int err;
612 
613 	if (BP_GET_LEVEL(bp) > 0) {
614 		arc_flags_t flags = ARC_FLAG_WAIT;
615 		int i;
616 		blkptr_t *cbp;
617 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
618 		arc_buf_t *buf;
619 
620 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
621 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
622 		if (err) {
623 			scn->scn_phys.scn_errors++;
624 			return (err);
625 		}
626 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
627 			dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
628 			    zb->zb_object, zb->zb_blkid * epb + i);
629 		}
630 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
631 			zbookmark_phys_t czb;
632 
633 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
634 			    zb->zb_level - 1,
635 			    zb->zb_blkid * epb + i);
636 			dsl_scan_visitbp(cbp, &czb, dnp,
637 			    ds, scn, ostype, tx);
638 		}
639 		(void) arc_buf_remove_ref(buf, &buf);
640 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
641 		arc_flags_t flags = ARC_FLAG_WAIT;
642 		dnode_phys_t *cdnp;
643 		int i, j;
644 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
645 		arc_buf_t *buf;
646 
647 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
648 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
649 		if (err) {
650 			scn->scn_phys.scn_errors++;
651 			return (err);
652 		}
653 		for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
654 			for (j = 0; j < cdnp->dn_nblkptr; j++) {
655 				blkptr_t *cbp = &cdnp->dn_blkptr[j];
656 				dsl_scan_prefetch(scn, buf, cbp,
657 				    zb->zb_objset, zb->zb_blkid * epb + i, j);
658 			}
659 		}
660 		for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
661 			dsl_scan_visitdnode(scn, ds, ostype,
662 			    cdnp, zb->zb_blkid * epb + i, tx);
663 		}
664 
665 		(void) arc_buf_remove_ref(buf, &buf);
666 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
667 		arc_flags_t flags = ARC_FLAG_WAIT;
668 		objset_phys_t *osp;
669 		arc_buf_t *buf;
670 
671 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
672 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
673 		if (err) {
674 			scn->scn_phys.scn_errors++;
675 			return (err);
676 		}
677 
678 		osp = buf->b_data;
679 
680 		dsl_scan_visitdnode(scn, ds, osp->os_type,
681 		    &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
682 
683 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
684 			/*
685 			 * We also always visit user/group accounting
686 			 * objects, and never skip them, even if we are
687 			 * pausing.  This is necessary so that the space
688 			 * deltas from this txg get integrated.
689 			 */
690 			dsl_scan_visitdnode(scn, ds, osp->os_type,
691 			    &osp->os_groupused_dnode,
692 			    DMU_GROUPUSED_OBJECT, tx);
693 			dsl_scan_visitdnode(scn, ds, osp->os_type,
694 			    &osp->os_userused_dnode,
695 			    DMU_USERUSED_OBJECT, tx);
696 		}
697 		(void) arc_buf_remove_ref(buf, &buf);
698 	}
699 
700 	return (0);
701 }
702 
703 static void
704 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
705     dmu_objset_type_t ostype, dnode_phys_t *dnp,
706     uint64_t object, dmu_tx_t *tx)
707 {
708 	int j;
709 
710 	for (j = 0; j < dnp->dn_nblkptr; j++) {
711 		zbookmark_phys_t czb;
712 
713 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
714 		    dnp->dn_nlevels - 1, j);
715 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
716 		    &czb, dnp, ds, scn, ostype, tx);
717 	}
718 
719 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
720 		zbookmark_phys_t czb;
721 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
722 		    0, DMU_SPILL_BLKID);
723 		dsl_scan_visitbp(&dnp->dn_spill,
724 		    &czb, dnp, ds, scn, ostype, tx);
725 	}
726 }
727 
728 /*
729  * The arguments are in this order because mdb can only print the
730  * first 5; we want them to be useful.
731  */
732 static void
733 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
734     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
735     dmu_objset_type_t ostype, dmu_tx_t *tx)
736 {
737 	dsl_pool_t *dp = scn->scn_dp;
738 	arc_buf_t *buf = NULL;
739 	blkptr_t bp_toread = *bp;
740 
741 	/* ASSERT(pbuf == NULL || arc_released(pbuf)); */
742 
743 	if (dsl_scan_check_pause(scn, zb))
744 		return;
745 
746 	if (dsl_scan_check_resume(scn, dnp, zb))
747 		return;
748 
749 	if (BP_IS_HOLE(bp))
750 		return;
751 
752 	scn->scn_visited_this_txg++;
753 
754 	dprintf_bp(bp,
755 	    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
756 	    ds, ds ? ds->ds_object : 0,
757 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
758 	    bp);
759 
760 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
761 		return;
762 
763 	if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
764 		return;
765 
766 	/*
767 	 * If dsl_scan_ddt() has aready visited this block, it will have
768 	 * already done any translations or scrubbing, so don't call the
769 	 * callback again.
770 	 */
771 	if (ddt_class_contains(dp->dp_spa,
772 	    scn->scn_phys.scn_ddt_class_max, bp)) {
773 		ASSERT(buf == NULL);
774 		return;
775 	}
776 
777 	/*
778 	 * If this block is from the future (after cur_max_txg), then we
779 	 * are doing this on behalf of a deleted snapshot, and we will
780 	 * revisit the future block on the next pass of this dataset.
781 	 * Don't scan it now unless we need to because something
782 	 * under it was modified.
783 	 */
784 	if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
785 		scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
786 	}
787 }
788 
789 static void
790 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
791     dmu_tx_t *tx)
792 {
793 	zbookmark_phys_t zb;
794 
795 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
796 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
797 	dsl_scan_visitbp(bp, &zb, NULL,
798 	    ds, scn, DMU_OST_NONE, tx);
799 
800 	dprintf_ds(ds, "finished scan%s", "");
801 }
802 
803 void
804 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
805 {
806 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
807 	dsl_scan_t *scn = dp->dp_scan;
808 	uint64_t mintxg;
809 
810 	if (scn->scn_phys.scn_state != DSS_SCANNING)
811 		return;
812 
813 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
814 		if (ds->ds_is_snapshot) {
815 			/* Note, scn_cur_{min,max}_txg stays the same. */
816 			scn->scn_phys.scn_bookmark.zb_objset =
817 			    dsl_dataset_phys(ds)->ds_next_snap_obj;
818 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
819 			    "reset zb_objset to %llu",
820 			    (u_longlong_t)ds->ds_object,
821 			    (u_longlong_t)dsl_dataset_phys(ds)->
822 			    ds_next_snap_obj);
823 			scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
824 		} else {
825 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
826 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
827 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
828 			    "reset bookmark to -1,0,0,0",
829 			    (u_longlong_t)ds->ds_object);
830 		}
831 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
832 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
833 		ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
834 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
835 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
836 		if (ds->ds_is_snapshot) {
837 			/*
838 			 * We keep the same mintxg; it could be >
839 			 * ds_creation_txg if the previous snapshot was
840 			 * deleted too.
841 			 */
842 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
843 			    scn->scn_phys.scn_queue_obj,
844 			    dsl_dataset_phys(ds)->ds_next_snap_obj,
845 			    mintxg, tx) == 0);
846 			zfs_dbgmsg("destroying ds %llu; in queue; "
847 			    "replacing with %llu",
848 			    (u_longlong_t)ds->ds_object,
849 			    (u_longlong_t)dsl_dataset_phys(ds)->
850 			    ds_next_snap_obj);
851 		} else {
852 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
853 			    (u_longlong_t)ds->ds_object);
854 		}
855 	} else {
856 		zfs_dbgmsg("destroying ds %llu; ignoring",
857 		    (u_longlong_t)ds->ds_object);
858 	}
859 
860 	/*
861 	 * dsl_scan_sync() should be called after this, and should sync
862 	 * out our changed state, but just to be safe, do it here.
863 	 */
864 	dsl_scan_sync_state(scn, tx);
865 }
866 
867 void
868 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
869 {
870 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
871 	dsl_scan_t *scn = dp->dp_scan;
872 	uint64_t mintxg;
873 
874 	if (scn->scn_phys.scn_state != DSS_SCANNING)
875 		return;
876 
877 	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
878 
879 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
880 		scn->scn_phys.scn_bookmark.zb_objset =
881 		    dsl_dataset_phys(ds)->ds_prev_snap_obj;
882 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
883 		    "reset zb_objset to %llu",
884 		    (u_longlong_t)ds->ds_object,
885 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
886 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
887 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
888 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
889 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
890 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
891 		    scn->scn_phys.scn_queue_obj,
892 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
893 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
894 		    "replacing with %llu",
895 		    (u_longlong_t)ds->ds_object,
896 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
897 	}
898 	dsl_scan_sync_state(scn, tx);
899 }
900 
901 void
902 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
903 {
904 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
905 	dsl_scan_t *scn = dp->dp_scan;
906 	uint64_t mintxg;
907 
908 	if (scn->scn_phys.scn_state != DSS_SCANNING)
909 		return;
910 
911 	if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
912 		scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
913 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
914 		    "reset zb_objset to %llu",
915 		    (u_longlong_t)ds1->ds_object,
916 		    (u_longlong_t)ds2->ds_object);
917 	} else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
918 		scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
919 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
920 		    "reset zb_objset to %llu",
921 		    (u_longlong_t)ds2->ds_object,
922 		    (u_longlong_t)ds1->ds_object);
923 	}
924 
925 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
926 	    ds1->ds_object, &mintxg) == 0) {
927 		int err;
928 
929 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
930 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
931 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
932 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
933 		err = zap_add_int_key(dp->dp_meta_objset,
934 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
935 		VERIFY(err == 0 || err == EEXIST);
936 		if (err == EEXIST) {
937 			/* Both were there to begin with */
938 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
939 			    scn->scn_phys.scn_queue_obj,
940 			    ds1->ds_object, mintxg, tx));
941 		}
942 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
943 		    "replacing with %llu",
944 		    (u_longlong_t)ds1->ds_object,
945 		    (u_longlong_t)ds2->ds_object);
946 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
947 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
948 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
949 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
950 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
951 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
952 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
953 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
954 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
955 		    "replacing with %llu",
956 		    (u_longlong_t)ds2->ds_object,
957 		    (u_longlong_t)ds1->ds_object);
958 	}
959 
960 	dsl_scan_sync_state(scn, tx);
961 }
962 
963 struct enqueue_clones_arg {
964 	dmu_tx_t *tx;
965 	uint64_t originobj;
966 };
967 
968 /* ARGSUSED */
969 static int
970 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
971 {
972 	struct enqueue_clones_arg *eca = arg;
973 	dsl_dataset_t *ds;
974 	int err;
975 	dsl_scan_t *scn = dp->dp_scan;
976 
977 	if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
978 		return (0);
979 
980 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
981 	if (err)
982 		return (err);
983 
984 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
985 		dsl_dataset_t *prev;
986 		err = dsl_dataset_hold_obj(dp,
987 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
988 
989 		dsl_dataset_rele(ds, FTAG);
990 		if (err)
991 			return (err);
992 		ds = prev;
993 	}
994 	VERIFY(zap_add_int_key(dp->dp_meta_objset,
995 	    scn->scn_phys.scn_queue_obj, ds->ds_object,
996 	    dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
997 	dsl_dataset_rele(ds, FTAG);
998 	return (0);
999 }
1000 
1001 static void
1002 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1003 {
1004 	dsl_pool_t *dp = scn->scn_dp;
1005 	dsl_dataset_t *ds;
1006 	objset_t *os;
1007 
1008 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1009 
1010 	if (dmu_objset_from_ds(ds, &os))
1011 		goto out;
1012 
1013 	/*
1014 	 * Only the ZIL in the head (non-snapshot) is valid.  Even though
1015 	 * snapshots can have ZIL block pointers (which may be the same
1016 	 * BP as in the head), they must be ignored.  So we traverse the
1017 	 * ZIL here, rather than in scan_recurse(), because the regular
1018 	 * snapshot block-sharing rules don't apply to it.
1019 	 */
1020 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1021 		dsl_scan_zil(dp, &os->os_zil_header);
1022 
1023 	/*
1024 	 * Iterate over the bps in this ds.
1025 	 */
1026 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1027 	dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1028 
1029 	char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1030 	dsl_dataset_name(ds, dsname);
1031 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1032 	    "pausing=%u",
1033 	    (longlong_t)dsobj, dsname,
1034 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
1035 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
1036 	    (int)scn->scn_pausing);
1037 	kmem_free(dsname, ZFS_MAXNAMELEN);
1038 
1039 	if (scn->scn_pausing)
1040 		goto out;
1041 
1042 	/*
1043 	 * We've finished this pass over this dataset.
1044 	 */
1045 
1046 	/*
1047 	 * If we did not completely visit this dataset, do another pass.
1048 	 */
1049 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1050 		zfs_dbgmsg("incomplete pass; visiting again");
1051 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1052 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1053 		    scn->scn_phys.scn_queue_obj, ds->ds_object,
1054 		    scn->scn_phys.scn_cur_max_txg, tx) == 0);
1055 		goto out;
1056 	}
1057 
1058 	/*
1059 	 * Add descendent datasets to work queue.
1060 	 */
1061 	if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1062 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1063 		    scn->scn_phys.scn_queue_obj,
1064 		    dsl_dataset_phys(ds)->ds_next_snap_obj,
1065 		    dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1066 	}
1067 	if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1068 		boolean_t usenext = B_FALSE;
1069 		if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1070 			uint64_t count;
1071 			/*
1072 			 * A bug in a previous version of the code could
1073 			 * cause upgrade_clones_cb() to not set
1074 			 * ds_next_snap_obj when it should, leading to a
1075 			 * missing entry.  Therefore we can only use the
1076 			 * next_clones_obj when its count is correct.
1077 			 */
1078 			int err = zap_count(dp->dp_meta_objset,
1079 			    dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1080 			if (err == 0 &&
1081 			    count == dsl_dataset_phys(ds)->ds_num_children - 1)
1082 				usenext = B_TRUE;
1083 		}
1084 
1085 		if (usenext) {
1086 			VERIFY0(zap_join_key(dp->dp_meta_objset,
1087 			    dsl_dataset_phys(ds)->ds_next_clones_obj,
1088 			    scn->scn_phys.scn_queue_obj,
1089 			    dsl_dataset_phys(ds)->ds_creation_txg, tx));
1090 		} else {
1091 			struct enqueue_clones_arg eca;
1092 			eca.tx = tx;
1093 			eca.originobj = ds->ds_object;
1094 
1095 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1096 			    enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1097 		}
1098 	}
1099 
1100 out:
1101 	dsl_dataset_rele(ds, FTAG);
1102 }
1103 
1104 /* ARGSUSED */
1105 static int
1106 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1107 {
1108 	dmu_tx_t *tx = arg;
1109 	dsl_dataset_t *ds;
1110 	int err;
1111 	dsl_scan_t *scn = dp->dp_scan;
1112 
1113 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1114 	if (err)
1115 		return (err);
1116 
1117 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1118 		dsl_dataset_t *prev;
1119 		err = dsl_dataset_hold_obj(dp,
1120 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1121 		if (err) {
1122 			dsl_dataset_rele(ds, FTAG);
1123 			return (err);
1124 		}
1125 
1126 		/*
1127 		 * If this is a clone, we don't need to worry about it for now.
1128 		 */
1129 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1130 			dsl_dataset_rele(ds, FTAG);
1131 			dsl_dataset_rele(prev, FTAG);
1132 			return (0);
1133 		}
1134 		dsl_dataset_rele(ds, FTAG);
1135 		ds = prev;
1136 	}
1137 
1138 	VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1139 	    ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1140 	dsl_dataset_rele(ds, FTAG);
1141 	return (0);
1142 }
1143 
1144 /*
1145  * Scrub/dedup interaction.
1146  *
1147  * If there are N references to a deduped block, we don't want to scrub it
1148  * N times -- ideally, we should scrub it exactly once.
1149  *
1150  * We leverage the fact that the dde's replication class (enum ddt_class)
1151  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1152  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1153  *
1154  * To prevent excess scrubbing, the scrub begins by walking the DDT
1155  * to find all blocks with refcnt > 1, and scrubs each of these once.
1156  * Since there are two replication classes which contain blocks with
1157  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1158  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1159  *
1160  * There would be nothing more to say if a block's refcnt couldn't change
1161  * during a scrub, but of course it can so we must account for changes
1162  * in a block's replication class.
1163  *
1164  * Here's an example of what can occur:
1165  *
1166  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1167  * when visited during the top-down scrub phase, it will be scrubbed twice.
1168  * This negates our scrub optimization, but is otherwise harmless.
1169  *
1170  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1171  * on each visit during the top-down scrub phase, it will never be scrubbed.
1172  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1173  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1174  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1175  * while a scrub is in progress, it scrubs the block right then.
1176  */
1177 static void
1178 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1179 {
1180 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1181 	ddt_entry_t dde = { 0 };
1182 	int error;
1183 	uint64_t n = 0;
1184 
1185 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1186 		ddt_t *ddt;
1187 
1188 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1189 			break;
1190 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1191 		    (longlong_t)ddb->ddb_class,
1192 		    (longlong_t)ddb->ddb_type,
1193 		    (longlong_t)ddb->ddb_checksum,
1194 		    (longlong_t)ddb->ddb_cursor);
1195 
1196 		/* There should be no pending changes to the dedup table */
1197 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1198 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1199 
1200 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1201 		n++;
1202 
1203 		if (dsl_scan_check_pause(scn, NULL))
1204 			break;
1205 	}
1206 
1207 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1208 	    (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1209 	    (int)scn->scn_pausing);
1210 
1211 	ASSERT(error == 0 || error == ENOENT);
1212 	ASSERT(error != ENOENT ||
1213 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1214 }
1215 
1216 /* ARGSUSED */
1217 void
1218 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1219     ddt_entry_t *dde, dmu_tx_t *tx)
1220 {
1221 	const ddt_key_t *ddk = &dde->dde_key;
1222 	ddt_phys_t *ddp = dde->dde_phys;
1223 	blkptr_t bp;
1224 	zbookmark_phys_t zb = { 0 };
1225 
1226 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1227 		return;
1228 
1229 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1230 		if (ddp->ddp_phys_birth == 0 ||
1231 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1232 			continue;
1233 		ddt_bp_create(checksum, ddk, ddp, &bp);
1234 
1235 		scn->scn_visited_this_txg++;
1236 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1237 	}
1238 }
1239 
1240 static void
1241 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1242 {
1243 	dsl_pool_t *dp = scn->scn_dp;
1244 	zap_cursor_t zc;
1245 	zap_attribute_t za;
1246 
1247 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1248 	    scn->scn_phys.scn_ddt_class_max) {
1249 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1250 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1251 		dsl_scan_ddt(scn, tx);
1252 		if (scn->scn_pausing)
1253 			return;
1254 	}
1255 
1256 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1257 		/* First do the MOS & ORIGIN */
1258 
1259 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1260 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1261 		dsl_scan_visit_rootbp(scn, NULL,
1262 		    &dp->dp_meta_rootbp, tx);
1263 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1264 		if (scn->scn_pausing)
1265 			return;
1266 
1267 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1268 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1269 			    enqueue_cb, tx, DS_FIND_CHILDREN));
1270 		} else {
1271 			dsl_scan_visitds(scn,
1272 			    dp->dp_origin_snap->ds_object, tx);
1273 		}
1274 		ASSERT(!scn->scn_pausing);
1275 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
1276 	    ZB_DESTROYED_OBJSET) {
1277 		/*
1278 		 * If we were paused, continue from here.  Note if the
1279 		 * ds we were paused on was deleted, the zb_objset may
1280 		 * be -1, so we will skip this and find a new objset
1281 		 * below.
1282 		 */
1283 		dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1284 		if (scn->scn_pausing)
1285 			return;
1286 	}
1287 
1288 	/*
1289 	 * In case we were paused right at the end of the ds, zero the
1290 	 * bookmark so we don't think that we're still trying to resume.
1291 	 */
1292 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1293 
1294 	/* keep pulling things out of the zap-object-as-queue */
1295 	while (zap_cursor_init(&zc, dp->dp_meta_objset,
1296 	    scn->scn_phys.scn_queue_obj),
1297 	    zap_cursor_retrieve(&zc, &za) == 0) {
1298 		dsl_dataset_t *ds;
1299 		uint64_t dsobj;
1300 
1301 		dsobj = strtonum(za.za_name, NULL);
1302 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1303 		    scn->scn_phys.scn_queue_obj, dsobj, tx));
1304 
1305 		/* Set up min/max txg */
1306 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1307 		if (za.za_first_integer != 0) {
1308 			scn->scn_phys.scn_cur_min_txg =
1309 			    MAX(scn->scn_phys.scn_min_txg,
1310 			    za.za_first_integer);
1311 		} else {
1312 			scn->scn_phys.scn_cur_min_txg =
1313 			    MAX(scn->scn_phys.scn_min_txg,
1314 			    dsl_dataset_phys(ds)->ds_prev_snap_txg);
1315 		}
1316 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1317 		dsl_dataset_rele(ds, FTAG);
1318 
1319 		dsl_scan_visitds(scn, dsobj, tx);
1320 		zap_cursor_fini(&zc);
1321 		if (scn->scn_pausing)
1322 			return;
1323 	}
1324 	zap_cursor_fini(&zc);
1325 }
1326 
1327 static boolean_t
1328 dsl_scan_free_should_pause(dsl_scan_t *scn)
1329 {
1330 	uint64_t elapsed_nanosecs;
1331 
1332 	if (zfs_recover)
1333 		return (B_FALSE);
1334 
1335 	if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1336 		return (B_TRUE);
1337 
1338 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1339 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1340 	    (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1341 	    txg_sync_waiting(scn->scn_dp)) ||
1342 	    spa_shutting_down(scn->scn_dp->dp_spa));
1343 }
1344 
1345 static int
1346 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1347 {
1348 	dsl_scan_t *scn = arg;
1349 
1350 	if (!scn->scn_is_bptree ||
1351 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1352 		if (dsl_scan_free_should_pause(scn))
1353 			return (SET_ERROR(ERESTART));
1354 	}
1355 
1356 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1357 	    dmu_tx_get_txg(tx), bp, 0));
1358 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1359 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1360 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1361 	scn->scn_visited_this_txg++;
1362 	return (0);
1363 }
1364 
1365 boolean_t
1366 dsl_scan_active(dsl_scan_t *scn)
1367 {
1368 	spa_t *spa = scn->scn_dp->dp_spa;
1369 	uint64_t used = 0, comp, uncomp;
1370 
1371 	if (spa->spa_load_state != SPA_LOAD_NONE)
1372 		return (B_FALSE);
1373 	if (spa_shutting_down(spa))
1374 		return (B_FALSE);
1375 	if (scn->scn_phys.scn_state == DSS_SCANNING ||
1376 	    (scn->scn_async_destroying && !scn->scn_async_stalled))
1377 		return (B_TRUE);
1378 
1379 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1380 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1381 		    &used, &comp, &uncomp);
1382 	}
1383 	return (used != 0);
1384 }
1385 
1386 void
1387 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1388 {
1389 	dsl_scan_t *scn = dp->dp_scan;
1390 	spa_t *spa = dp->dp_spa;
1391 	int err = 0;
1392 
1393 	/*
1394 	 * Check for scn_restart_txg before checking spa_load_state, so
1395 	 * that we can restart an old-style scan while the pool is being
1396 	 * imported (see dsl_scan_init).
1397 	 */
1398 	if (scn->scn_restart_txg != 0 &&
1399 	    scn->scn_restart_txg <= tx->tx_txg) {
1400 		pool_scan_func_t func = POOL_SCAN_SCRUB;
1401 		dsl_scan_done(scn, B_FALSE, tx);
1402 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1403 			func = POOL_SCAN_RESILVER;
1404 		zfs_dbgmsg("restarting scan func=%u txg=%llu",
1405 		    func, tx->tx_txg);
1406 		dsl_scan_setup_sync(&func, tx);
1407 	}
1408 
1409 	/*
1410 	 * If the scan is inactive due to a stalled async destroy, try again.
1411 	 */
1412 	if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1413 	    spa_sync_pass(dp->dp_spa) > 1)
1414 		return;
1415 
1416 	scn->scn_visited_this_txg = 0;
1417 	scn->scn_pausing = B_FALSE;
1418 	scn->scn_sync_start_time = gethrtime();
1419 	spa->spa_scrub_active = B_TRUE;
1420 
1421 	/*
1422 	 * First process the async destroys.  If we pause, don't do
1423 	 * any scrubbing or resilvering.  This ensures that there are no
1424 	 * async destroys while we are scanning, so the scan code doesn't
1425 	 * have to worry about traversing it.  It is also faster to free the
1426 	 * blocks than to scrub them.
1427 	 */
1428 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1429 		scn->scn_is_bptree = B_FALSE;
1430 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1431 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1432 		err = bpobj_iterate(&dp->dp_free_bpobj,
1433 		    dsl_scan_free_block_cb, scn, tx);
1434 		VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1435 
1436 		if (err != 0 && err != ERESTART)
1437 			zfs_panic_recover("error %u from bpobj_iterate()", err);
1438 	}
1439 
1440 	if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1441 		ASSERT(scn->scn_async_destroying);
1442 		scn->scn_is_bptree = B_TRUE;
1443 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1444 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1445 		err = bptree_iterate(dp->dp_meta_objset,
1446 		    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1447 		VERIFY0(zio_wait(scn->scn_zio_root));
1448 
1449 		if (err == EIO || err == ECKSUM) {
1450 			err = 0;
1451 		} else if (err != 0 && err != ERESTART) {
1452 			zfs_panic_recover("error %u from "
1453 			    "traverse_dataset_destroyed()", err);
1454 		}
1455 
1456 		if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1457 			/* finished; deactivate async destroy feature */
1458 			spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1459 			ASSERT(!spa_feature_is_active(spa,
1460 			    SPA_FEATURE_ASYNC_DESTROY));
1461 			VERIFY0(zap_remove(dp->dp_meta_objset,
1462 			    DMU_POOL_DIRECTORY_OBJECT,
1463 			    DMU_POOL_BPTREE_OBJ, tx));
1464 			VERIFY0(bptree_free(dp->dp_meta_objset,
1465 			    dp->dp_bptree_obj, tx));
1466 			dp->dp_bptree_obj = 0;
1467 			scn->scn_async_destroying = B_FALSE;
1468 			scn->scn_async_stalled = B_FALSE;
1469 		} else {
1470 			/*
1471 			 * If we didn't make progress, mark the async
1472 			 * destroy as stalled, so that we will not initiate
1473 			 * a spa_sync() on its behalf.  Note that we only
1474 			 * check this if we are not finished, because if the
1475 			 * bptree had no blocks for us to visit, we can
1476 			 * finish without "making progress".
1477 			 */
1478 			scn->scn_async_stalled =
1479 			    (scn->scn_visited_this_txg == 0);
1480 		}
1481 	}
1482 	if (scn->scn_visited_this_txg) {
1483 		zfs_dbgmsg("freed %llu blocks in %llums from "
1484 		    "free_bpobj/bptree txg %llu; err=%u",
1485 		    (longlong_t)scn->scn_visited_this_txg,
1486 		    (longlong_t)
1487 		    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1488 		    (longlong_t)tx->tx_txg, err);
1489 		scn->scn_visited_this_txg = 0;
1490 
1491 		/*
1492 		 * Write out changes to the DDT that may be required as a
1493 		 * result of the blocks freed.  This ensures that the DDT
1494 		 * is clean when a scrub/resilver runs.
1495 		 */
1496 		ddt_sync(spa, tx->tx_txg);
1497 	}
1498 	if (err != 0)
1499 		return;
1500 	if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1501 	    (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1502 	    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1503 	    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1504 		/*
1505 		 * We have finished background destroying, but there is still
1506 		 * some space left in the dp_free_dir. Transfer this leaked
1507 		 * space to the dp_leak_dir.
1508 		 */
1509 		if (dp->dp_leak_dir == NULL) {
1510 			rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1511 			(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1512 			    LEAK_DIR_NAME, tx);
1513 			VERIFY0(dsl_pool_open_special_dir(dp,
1514 			    LEAK_DIR_NAME, &dp->dp_leak_dir));
1515 			rrw_exit(&dp->dp_config_rwlock, FTAG);
1516 		}
1517 		dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1518 		    dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1519 		    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1520 		    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1521 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1522 		    -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1523 		    -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1524 		    -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1525 	}
1526 	if (!scn->scn_async_destroying) {
1527 		/* finished; verify that space accounting went to zero */
1528 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1529 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1530 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1531 	}
1532 
1533 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1534 		return;
1535 
1536 	if (scn->scn_done_txg == tx->tx_txg) {
1537 		ASSERT(!scn->scn_pausing);
1538 		/* finished with scan. */
1539 		zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1540 		dsl_scan_done(scn, B_TRUE, tx);
1541 		ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1542 		dsl_scan_sync_state(scn, tx);
1543 		return;
1544 	}
1545 
1546 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1547 	    scn->scn_phys.scn_ddt_class_max) {
1548 		zfs_dbgmsg("doing scan sync txg %llu; "
1549 		    "ddt bm=%llu/%llu/%llu/%llx",
1550 		    (longlong_t)tx->tx_txg,
1551 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1552 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1553 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1554 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1555 		ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1556 		ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1557 		ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1558 		ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1559 	} else {
1560 		zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1561 		    (longlong_t)tx->tx_txg,
1562 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1563 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1564 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1565 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1566 	}
1567 
1568 	scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1569 	    NULL, ZIO_FLAG_CANFAIL);
1570 	dsl_pool_config_enter(dp, FTAG);
1571 	dsl_scan_visit(scn, tx);
1572 	dsl_pool_config_exit(dp, FTAG);
1573 	(void) zio_wait(scn->scn_zio_root);
1574 	scn->scn_zio_root = NULL;
1575 
1576 	zfs_dbgmsg("visited %llu blocks in %llums",
1577 	    (longlong_t)scn->scn_visited_this_txg,
1578 	    (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1579 
1580 	if (!scn->scn_pausing) {
1581 		scn->scn_done_txg = tx->tx_txg + 1;
1582 		zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1583 		    tx->tx_txg, scn->scn_done_txg);
1584 	}
1585 
1586 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1587 		mutex_enter(&spa->spa_scrub_lock);
1588 		while (spa->spa_scrub_inflight > 0) {
1589 			cv_wait(&spa->spa_scrub_io_cv,
1590 			    &spa->spa_scrub_lock);
1591 		}
1592 		mutex_exit(&spa->spa_scrub_lock);
1593 	}
1594 
1595 	dsl_scan_sync_state(scn, tx);
1596 }
1597 
1598 /*
1599  * This will start a new scan, or restart an existing one.
1600  */
1601 void
1602 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1603 {
1604 	if (txg == 0) {
1605 		dmu_tx_t *tx;
1606 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1607 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1608 
1609 		txg = dmu_tx_get_txg(tx);
1610 		dp->dp_scan->scn_restart_txg = txg;
1611 		dmu_tx_commit(tx);
1612 	} else {
1613 		dp->dp_scan->scn_restart_txg = txg;
1614 	}
1615 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1616 }
1617 
1618 boolean_t
1619 dsl_scan_resilvering(dsl_pool_t *dp)
1620 {
1621 	return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1622 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1623 }
1624 
1625 /*
1626  * scrub consumers
1627  */
1628 
1629 static void
1630 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1631 {
1632 	int i;
1633 
1634 	/*
1635 	 * If we resume after a reboot, zab will be NULL; don't record
1636 	 * incomplete stats in that case.
1637 	 */
1638 	if (zab == NULL)
1639 		return;
1640 
1641 	for (i = 0; i < 4; i++) {
1642 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1643 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1644 		if (t & DMU_OT_NEWTYPE)
1645 			t = DMU_OT_OTHER;
1646 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
1647 		int equal;
1648 
1649 		zb->zb_count++;
1650 		zb->zb_asize += BP_GET_ASIZE(bp);
1651 		zb->zb_lsize += BP_GET_LSIZE(bp);
1652 		zb->zb_psize += BP_GET_PSIZE(bp);
1653 		zb->zb_gangs += BP_COUNT_GANG(bp);
1654 
1655 		switch (BP_GET_NDVAS(bp)) {
1656 		case 2:
1657 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1658 			    DVA_GET_VDEV(&bp->blk_dva[1]))
1659 				zb->zb_ditto_2_of_2_samevdev++;
1660 			break;
1661 		case 3:
1662 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1663 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
1664 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1665 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
1666 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1667 			    DVA_GET_VDEV(&bp->blk_dva[2]));
1668 			if (equal == 1)
1669 				zb->zb_ditto_2_of_3_samevdev++;
1670 			else if (equal == 3)
1671 				zb->zb_ditto_3_of_3_samevdev++;
1672 			break;
1673 		}
1674 	}
1675 }
1676 
1677 static void
1678 dsl_scan_scrub_done(zio_t *zio)
1679 {
1680 	spa_t *spa = zio->io_spa;
1681 
1682 	zio_data_buf_free(zio->io_data, zio->io_size);
1683 
1684 	mutex_enter(&spa->spa_scrub_lock);
1685 	spa->spa_scrub_inflight--;
1686 	cv_broadcast(&spa->spa_scrub_io_cv);
1687 
1688 	if (zio->io_error && (zio->io_error != ECKSUM ||
1689 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1690 		spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1691 	}
1692 	mutex_exit(&spa->spa_scrub_lock);
1693 }
1694 
1695 static int
1696 dsl_scan_scrub_cb(dsl_pool_t *dp,
1697     const blkptr_t *bp, const zbookmark_phys_t *zb)
1698 {
1699 	dsl_scan_t *scn = dp->dp_scan;
1700 	size_t size = BP_GET_PSIZE(bp);
1701 	spa_t *spa = dp->dp_spa;
1702 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1703 	boolean_t needs_io;
1704 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1705 	int scan_delay = 0;
1706 
1707 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
1708 	    phys_birth >= scn->scn_phys.scn_max_txg)
1709 		return (0);
1710 
1711 	count_block(dp->dp_blkstats, bp);
1712 
1713 	if (BP_IS_EMBEDDED(bp))
1714 		return (0);
1715 
1716 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1717 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1718 		zio_flags |= ZIO_FLAG_SCRUB;
1719 		needs_io = B_TRUE;
1720 		scan_delay = zfs_scrub_delay;
1721 	} else {
1722 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1723 		zio_flags |= ZIO_FLAG_RESILVER;
1724 		needs_io = B_FALSE;
1725 		scan_delay = zfs_resilver_delay;
1726 	}
1727 
1728 	/* If it's an intent log block, failure is expected. */
1729 	if (zb->zb_level == ZB_ZIL_LEVEL)
1730 		zio_flags |= ZIO_FLAG_SPECULATIVE;
1731 
1732 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1733 		vdev_t *vd = vdev_lookup_top(spa,
1734 		    DVA_GET_VDEV(&bp->blk_dva[d]));
1735 
1736 		/*
1737 		 * Keep track of how much data we've examined so that
1738 		 * zpool(1M) status can make useful progress reports.
1739 		 */
1740 		scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1741 		spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1742 
1743 		/* if it's a resilver, this may not be in the target range */
1744 		if (!needs_io) {
1745 			if (DVA_GET_GANG(&bp->blk_dva[d])) {
1746 				/*
1747 				 * Gang members may be spread across multiple
1748 				 * vdevs, so the best estimate we have is the
1749 				 * scrub range, which has already been checked.
1750 				 * XXX -- it would be better to change our
1751 				 * allocation policy to ensure that all
1752 				 * gang members reside on the same vdev.
1753 				 */
1754 				needs_io = B_TRUE;
1755 			} else {
1756 				needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1757 				    phys_birth, 1);
1758 			}
1759 		}
1760 	}
1761 
1762 	if (needs_io && !zfs_no_scrub_io) {
1763 		vdev_t *rvd = spa->spa_root_vdev;
1764 		uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1765 		void *data = zio_data_buf_alloc(size);
1766 
1767 		mutex_enter(&spa->spa_scrub_lock);
1768 		while (spa->spa_scrub_inflight >= maxinflight)
1769 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1770 		spa->spa_scrub_inflight++;
1771 		mutex_exit(&spa->spa_scrub_lock);
1772 
1773 		/*
1774 		 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1775 		 * then throttle our workload to limit the impact of a scan.
1776 		 */
1777 		if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1778 			delay(scan_delay);
1779 
1780 		zio_nowait(zio_read(NULL, spa, bp, data, size,
1781 		    dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1782 		    zio_flags, zb));
1783 	}
1784 
1785 	/* do not relocate this block */
1786 	return (0);
1787 }
1788 
1789 int
1790 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1791 {
1792 	spa_t *spa = dp->dp_spa;
1793 
1794 	/*
1795 	 * Purge all vdev caches and probe all devices.  We do this here
1796 	 * rather than in sync context because this requires a writer lock
1797 	 * on the spa_config lock, which we can't do from sync context.  The
1798 	 * spa_scrub_reopen flag indicates that vdev_open() should not
1799 	 * attempt to start another scrub.
1800 	 */
1801 	spa_vdev_state_enter(spa, SCL_NONE);
1802 	spa->spa_scrub_reopen = B_TRUE;
1803 	vdev_reopen(spa->spa_root_vdev);
1804 	spa->spa_scrub_reopen = B_FALSE;
1805 	(void) spa_vdev_state_exit(spa, NULL, 0);
1806 
1807 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1808 	    dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1809 }
1810