xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_scan.c (revision 5d7b4d438c4a51eccc95e77a83a437b4d48380eb)
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) 2013 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 	/*
99 	 * It's possible that we're resuming a scan after a reboot so
100 	 * make sure that the scan_async_destroying flag is initialized
101 	 * appropriately.
102 	 */
103 	ASSERT(!scn->scn_async_destroying);
104 	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
105 	    SPA_FEATURE_ASYNC_DESTROY);
106 
107 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
108 	    "scrub_func", sizeof (uint64_t), 1, &f);
109 	if (err == 0) {
110 		/*
111 		 * There was an old-style scrub in progress.  Restart a
112 		 * new-style scrub from the beginning.
113 		 */
114 		scn->scn_restart_txg = txg;
115 		zfs_dbgmsg("old-style scrub was in progress; "
116 		    "restarting new-style scrub in txg %llu",
117 		    scn->scn_restart_txg);
118 
119 		/*
120 		 * Load the queue obj from the old location so that it
121 		 * can be freed by dsl_scan_done().
122 		 */
123 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
124 		    "scrub_queue", sizeof (uint64_t), 1,
125 		    &scn->scn_phys.scn_queue_obj);
126 	} else {
127 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
128 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
129 		    &scn->scn_phys);
130 		if (err == ENOENT)
131 			return (0);
132 		else if (err)
133 			return (err);
134 
135 		if (scn->scn_phys.scn_state == DSS_SCANNING &&
136 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
137 			/*
138 			 * A new-type scrub was in progress on an old
139 			 * pool, and the pool was accessed by old
140 			 * software.  Restart from the beginning, since
141 			 * the old software may have changed the pool in
142 			 * the meantime.
143 			 */
144 			scn->scn_restart_txg = txg;
145 			zfs_dbgmsg("new-style scrub was modified "
146 			    "by old software; restarting in txg %llu",
147 			    scn->scn_restart_txg);
148 		}
149 	}
150 
151 	spa_scan_stat_init(spa);
152 	return (0);
153 }
154 
155 void
156 dsl_scan_fini(dsl_pool_t *dp)
157 {
158 	if (dp->dp_scan) {
159 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
160 		dp->dp_scan = NULL;
161 	}
162 }
163 
164 /* ARGSUSED */
165 static int
166 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
167 {
168 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
169 
170 	if (scn->scn_phys.scn_state == DSS_SCANNING)
171 		return (SET_ERROR(EBUSY));
172 
173 	return (0);
174 }
175 
176 static void
177 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
178 {
179 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
180 	pool_scan_func_t *funcp = arg;
181 	dmu_object_type_t ot = 0;
182 	dsl_pool_t *dp = scn->scn_dp;
183 	spa_t *spa = dp->dp_spa;
184 
185 	ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
186 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
187 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
188 	scn->scn_phys.scn_func = *funcp;
189 	scn->scn_phys.scn_state = DSS_SCANNING;
190 	scn->scn_phys.scn_min_txg = 0;
191 	scn->scn_phys.scn_max_txg = tx->tx_txg;
192 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
193 	scn->scn_phys.scn_start_time = gethrestime_sec();
194 	scn->scn_phys.scn_errors = 0;
195 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
196 	scn->scn_restart_txg = 0;
197 	scn->scn_done_txg = 0;
198 	spa_scan_stat_init(spa);
199 
200 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
201 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
202 
203 		/* rewrite all disk labels */
204 		vdev_config_dirty(spa->spa_root_vdev);
205 
206 		if (vdev_resilver_needed(spa->spa_root_vdev,
207 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
208 			spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
209 		} else {
210 			spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
211 		}
212 
213 		spa->spa_scrub_started = B_TRUE;
214 		/*
215 		 * If this is an incremental scrub, limit the DDT scrub phase
216 		 * to just the auto-ditto class (for correctness); the rest
217 		 * of the scrub should go faster using top-down pruning.
218 		 */
219 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
220 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
221 
222 	}
223 
224 	/* back to the generic stuff */
225 
226 	if (dp->dp_blkstats == NULL) {
227 		dp->dp_blkstats =
228 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
229 	}
230 	bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
231 
232 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
233 		ot = DMU_OT_ZAP_OTHER;
234 
235 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
236 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
237 
238 	dsl_scan_sync_state(scn, tx);
239 
240 	spa_history_log_internal(spa, "scan setup", tx,
241 	    "func=%u mintxg=%llu maxtxg=%llu",
242 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
243 }
244 
245 /* ARGSUSED */
246 static void
247 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
248 {
249 	static const char *old_names[] = {
250 		"scrub_bookmark",
251 		"scrub_ddt_bookmark",
252 		"scrub_ddt_class_max",
253 		"scrub_queue",
254 		"scrub_min_txg",
255 		"scrub_max_txg",
256 		"scrub_func",
257 		"scrub_errors",
258 		NULL
259 	};
260 
261 	dsl_pool_t *dp = scn->scn_dp;
262 	spa_t *spa = dp->dp_spa;
263 	int i;
264 
265 	/* Remove any remnants of an old-style scrub. */
266 	for (i = 0; old_names[i]; i++) {
267 		(void) zap_remove(dp->dp_meta_objset,
268 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
269 	}
270 
271 	if (scn->scn_phys.scn_queue_obj != 0) {
272 		VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
273 		    scn->scn_phys.scn_queue_obj, tx));
274 		scn->scn_phys.scn_queue_obj = 0;
275 	}
276 
277 	/*
278 	 * If we were "restarted" from a stopped state, don't bother
279 	 * with anything else.
280 	 */
281 	if (scn->scn_phys.scn_state != DSS_SCANNING)
282 		return;
283 
284 	if (complete)
285 		scn->scn_phys.scn_state = DSS_FINISHED;
286 	else
287 		scn->scn_phys.scn_state = DSS_CANCELED;
288 
289 	spa_history_log_internal(spa, "scan done", tx,
290 	    "complete=%u", complete);
291 
292 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
293 		mutex_enter(&spa->spa_scrub_lock);
294 		while (spa->spa_scrub_inflight > 0) {
295 			cv_wait(&spa->spa_scrub_io_cv,
296 			    &spa->spa_scrub_lock);
297 		}
298 		mutex_exit(&spa->spa_scrub_lock);
299 		spa->spa_scrub_started = B_FALSE;
300 		spa->spa_scrub_active = B_FALSE;
301 
302 		/*
303 		 * If the scrub/resilver completed, update all DTLs to
304 		 * reflect this.  Whether it succeeded or not, vacate
305 		 * all temporary scrub DTLs.
306 		 */
307 		vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
308 		    complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
309 		if (complete) {
310 			spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
311 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
312 		}
313 		spa_errlog_rotate(spa);
314 
315 		/*
316 		 * We may have finished replacing a device.
317 		 * Let the async thread assess this and handle the detach.
318 		 */
319 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
320 	}
321 
322 	scn->scn_phys.scn_end_time = gethrestime_sec();
323 }
324 
325 /* ARGSUSED */
326 static int
327 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
328 {
329 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
330 
331 	if (scn->scn_phys.scn_state != DSS_SCANNING)
332 		return (SET_ERROR(ENOENT));
333 	return (0);
334 }
335 
336 /* ARGSUSED */
337 static void
338 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
339 {
340 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
341 
342 	dsl_scan_done(scn, B_FALSE, tx);
343 	dsl_scan_sync_state(scn, tx);
344 }
345 
346 int
347 dsl_scan_cancel(dsl_pool_t *dp)
348 {
349 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
350 	    dsl_scan_cancel_sync, NULL, 3));
351 }
352 
353 static void dsl_scan_visitbp(blkptr_t *bp,
354     const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
355     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
356     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, arc_buf_t *buf, 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 (dsl_dataset_is_snapshot(ds))
379 		return (MIN(smt, ds->ds_phys->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 static boolean_t
393 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb)
394 {
395 	uint64_t elapsed_nanosecs;
396 	int mintime;
397 
398 	/* we never skip user/group accounting objects */
399 	if (zb && (int64_t)zb->zb_object < 0)
400 		return (B_FALSE);
401 
402 	if (scn->scn_pausing)
403 		return (B_TRUE); /* we're already pausing */
404 
405 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
406 		return (B_FALSE); /* we're resuming */
407 
408 	/* We only know how to resume from level-0 blocks. */
409 	if (zb && zb->zb_level != 0)
410 		return (B_FALSE);
411 
412 	mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
413 	    zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
414 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
415 	if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
416 	    (NSEC2MSEC(elapsed_nanosecs) > mintime &&
417 	    txg_sync_waiting(scn->scn_dp)) ||
418 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
419 		if (zb) {
420 			dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
421 			    (longlong_t)zb->zb_objset,
422 			    (longlong_t)zb->zb_object,
423 			    (longlong_t)zb->zb_level,
424 			    (longlong_t)zb->zb_blkid);
425 			scn->scn_phys.scn_bookmark = *zb;
426 		}
427 		dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
428 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
429 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
430 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
431 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
432 		scn->scn_pausing = B_TRUE;
433 		return (B_TRUE);
434 	}
435 	return (B_FALSE);
436 }
437 
438 typedef struct zil_scan_arg {
439 	dsl_pool_t	*zsa_dp;
440 	zil_header_t	*zsa_zh;
441 } zil_scan_arg_t;
442 
443 /* ARGSUSED */
444 static int
445 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
446 {
447 	zil_scan_arg_t *zsa = arg;
448 	dsl_pool_t *dp = zsa->zsa_dp;
449 	dsl_scan_t *scn = dp->dp_scan;
450 	zil_header_t *zh = zsa->zsa_zh;
451 	zbookmark_t zb;
452 
453 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
454 		return (0);
455 
456 	/*
457 	 * One block ("stubby") can be allocated a long time ago; we
458 	 * want to visit that one because it has been allocated
459 	 * (on-disk) even if it hasn't been claimed (even though for
460 	 * scrub there's nothing to do to it).
461 	 */
462 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
463 		return (0);
464 
465 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
466 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
467 
468 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
469 	return (0);
470 }
471 
472 /* ARGSUSED */
473 static int
474 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
475 {
476 	if (lrc->lrc_txtype == TX_WRITE) {
477 		zil_scan_arg_t *zsa = arg;
478 		dsl_pool_t *dp = zsa->zsa_dp;
479 		dsl_scan_t *scn = dp->dp_scan;
480 		zil_header_t *zh = zsa->zsa_zh;
481 		lr_write_t *lr = (lr_write_t *)lrc;
482 		blkptr_t *bp = &lr->lr_blkptr;
483 		zbookmark_t zb;
484 
485 		if (BP_IS_HOLE(bp) ||
486 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
487 			return (0);
488 
489 		/*
490 		 * birth can be < claim_txg if this record's txg is
491 		 * already txg sync'ed (but this log block contains
492 		 * other records that are not synced)
493 		 */
494 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
495 			return (0);
496 
497 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
498 		    lr->lr_foid, ZB_ZIL_LEVEL,
499 		    lr->lr_offset / BP_GET_LSIZE(bp));
500 
501 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
502 	}
503 	return (0);
504 }
505 
506 static void
507 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
508 {
509 	uint64_t claim_txg = zh->zh_claim_txg;
510 	zil_scan_arg_t zsa = { dp, zh };
511 	zilog_t *zilog;
512 
513 	/*
514 	 * We only want to visit blocks that have been claimed but not yet
515 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
516 	 */
517 	if (claim_txg == 0 && spa_writeable(dp->dp_spa))
518 		return;
519 
520 	zilog = zil_alloc(dp->dp_meta_objset, zh);
521 
522 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
523 	    claim_txg);
524 
525 	zil_free(zilog);
526 }
527 
528 /* ARGSUSED */
529 static void
530 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
531     uint64_t objset, uint64_t object, uint64_t blkid)
532 {
533 	zbookmark_t czb;
534 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
535 
536 	if (zfs_no_scrub_prefetch)
537 		return;
538 
539 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
540 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
541 		return;
542 
543 	SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
544 
545 	(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
546 	    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
547 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
548 }
549 
550 static boolean_t
551 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
552     const zbookmark_t *zb)
553 {
554 	/*
555 	 * We never skip over user/group accounting objects (obj<0)
556 	 */
557 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
558 	    (int64_t)zb->zb_object >= 0) {
559 		/*
560 		 * If we already visited this bp & everything below (in
561 		 * a prior txg sync), don't bother doing it again.
562 		 */
563 		if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
564 			return (B_TRUE);
565 
566 		/*
567 		 * If we found the block we're trying to resume from, or
568 		 * we went past it to a different object, zero it out to
569 		 * indicate that it's OK to start checking for pausing
570 		 * again.
571 		 */
572 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
573 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
574 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
575 			    (longlong_t)zb->zb_objset,
576 			    (longlong_t)zb->zb_object,
577 			    (longlong_t)zb->zb_level,
578 			    (longlong_t)zb->zb_blkid);
579 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
580 		}
581 	}
582 	return (B_FALSE);
583 }
584 
585 /*
586  * Return nonzero on i/o error.
587  * Return new buf to write out in *bufp.
588  */
589 static int
590 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
591     dnode_phys_t *dnp, const blkptr_t *bp,
592     const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
593 {
594 	dsl_pool_t *dp = scn->scn_dp;
595 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
596 	int err;
597 
598 	if (BP_GET_LEVEL(bp) > 0) {
599 		uint32_t flags = ARC_WAIT;
600 		int i;
601 		blkptr_t *cbp;
602 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
603 
604 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
605 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
606 		if (err) {
607 			scn->scn_phys.scn_errors++;
608 			return (err);
609 		}
610 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
611 			dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
612 			    zb->zb_object, zb->zb_blkid * epb + i);
613 		}
614 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
615 			zbookmark_t czb;
616 
617 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
618 			    zb->zb_level - 1,
619 			    zb->zb_blkid * epb + i);
620 			dsl_scan_visitbp(cbp, &czb, dnp,
621 			    *bufp, ds, scn, ostype, tx);
622 		}
623 	} else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
624 		uint32_t flags = ARC_WAIT;
625 
626 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
627 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
628 		if (err) {
629 			scn->scn_phys.scn_errors++;
630 			return (err);
631 		}
632 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
633 		uint32_t flags = ARC_WAIT;
634 		dnode_phys_t *cdnp;
635 		int i, j;
636 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
637 
638 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
639 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
640 		if (err) {
641 			scn->scn_phys.scn_errors++;
642 			return (err);
643 		}
644 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
645 			for (j = 0; j < cdnp->dn_nblkptr; j++) {
646 				blkptr_t *cbp = &cdnp->dn_blkptr[j];
647 				dsl_scan_prefetch(scn, *bufp, cbp,
648 				    zb->zb_objset, zb->zb_blkid * epb + i, j);
649 			}
650 		}
651 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
652 			dsl_scan_visitdnode(scn, ds, ostype,
653 			    cdnp, *bufp, zb->zb_blkid * epb + i, tx);
654 		}
655 
656 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
657 		uint32_t flags = ARC_WAIT;
658 		objset_phys_t *osp;
659 
660 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
661 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
662 		if (err) {
663 			scn->scn_phys.scn_errors++;
664 			return (err);
665 		}
666 
667 		osp = (*bufp)->b_data;
668 
669 		dsl_scan_visitdnode(scn, ds, osp->os_type,
670 		    &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
671 
672 		if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
673 			/*
674 			 * We also always visit user/group accounting
675 			 * objects, and never skip them, even if we are
676 			 * pausing.  This is necessary so that the space
677 			 * deltas from this txg get integrated.
678 			 */
679 			dsl_scan_visitdnode(scn, ds, osp->os_type,
680 			    &osp->os_groupused_dnode, *bufp,
681 			    DMU_GROUPUSED_OBJECT, tx);
682 			dsl_scan_visitdnode(scn, ds, osp->os_type,
683 			    &osp->os_userused_dnode, *bufp,
684 			    DMU_USERUSED_OBJECT, tx);
685 		}
686 	}
687 
688 	return (0);
689 }
690 
691 static void
692 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
693     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
694     uint64_t object, dmu_tx_t *tx)
695 {
696 	int j;
697 
698 	for (j = 0; j < dnp->dn_nblkptr; j++) {
699 		zbookmark_t czb;
700 
701 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
702 		    dnp->dn_nlevels - 1, j);
703 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
704 		    &czb, dnp, buf, ds, scn, ostype, tx);
705 	}
706 
707 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
708 		zbookmark_t czb;
709 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
710 		    0, DMU_SPILL_BLKID);
711 		dsl_scan_visitbp(&dnp->dn_spill,
712 		    &czb, dnp, buf, ds, scn, ostype, tx);
713 	}
714 }
715 
716 /*
717  * The arguments are in this order because mdb can only print the
718  * first 5; we want them to be useful.
719  */
720 static void
721 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb,
722     dnode_phys_t *dnp, arc_buf_t *pbuf,
723     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
724     dmu_tx_t *tx)
725 {
726 	dsl_pool_t *dp = scn->scn_dp;
727 	arc_buf_t *buf = NULL;
728 	blkptr_t bp_toread = *bp;
729 
730 	/* ASSERT(pbuf == NULL || arc_released(pbuf)); */
731 
732 	if (dsl_scan_check_pause(scn, zb))
733 		return;
734 
735 	if (dsl_scan_check_resume(scn, dnp, zb))
736 		return;
737 
738 	if (BP_IS_HOLE(bp))
739 		return;
740 
741 	scn->scn_visited_this_txg++;
742 
743 	dprintf_bp(bp,
744 	    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
745 	    ds, ds ? ds->ds_object : 0,
746 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
747 	    pbuf, bp);
748 
749 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
750 		return;
751 
752 	if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
753 	    &buf) != 0)
754 		return;
755 
756 	/*
757 	 * If dsl_scan_ddt() has aready visited this block, it will have
758 	 * already done any translations or scrubbing, so don't call the
759 	 * callback again.
760 	 */
761 	if (ddt_class_contains(dp->dp_spa,
762 	    scn->scn_phys.scn_ddt_class_max, bp)) {
763 		ASSERT(buf == NULL);
764 		return;
765 	}
766 
767 	/*
768 	 * If this block is from the future (after cur_max_txg), then we
769 	 * are doing this on behalf of a deleted snapshot, and we will
770 	 * revisit the future block on the next pass of this dataset.
771 	 * Don't scan it now unless we need to because something
772 	 * under it was modified.
773 	 */
774 	if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
775 		scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
776 	}
777 	if (buf)
778 		(void) arc_buf_remove_ref(buf, &buf);
779 }
780 
781 static void
782 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
783     dmu_tx_t *tx)
784 {
785 	zbookmark_t zb;
786 
787 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
788 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
789 	dsl_scan_visitbp(bp, &zb, NULL, NULL,
790 	    ds, scn, DMU_OST_NONE, tx);
791 
792 	dprintf_ds(ds, "finished scan%s", "");
793 }
794 
795 void
796 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
797 {
798 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
799 	dsl_scan_t *scn = dp->dp_scan;
800 	uint64_t mintxg;
801 
802 	if (scn->scn_phys.scn_state != DSS_SCANNING)
803 		return;
804 
805 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
806 		if (dsl_dataset_is_snapshot(ds)) {
807 			/* Note, scn_cur_{min,max}_txg stays the same. */
808 			scn->scn_phys.scn_bookmark.zb_objset =
809 			    ds->ds_phys->ds_next_snap_obj;
810 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
811 			    "reset zb_objset to %llu",
812 			    (u_longlong_t)ds->ds_object,
813 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
814 			scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
815 		} else {
816 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
817 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
818 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
819 			    "reset bookmark to -1,0,0,0",
820 			    (u_longlong_t)ds->ds_object);
821 		}
822 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
823 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
824 		ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
825 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
826 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
827 		if (dsl_dataset_is_snapshot(ds)) {
828 			/*
829 			 * We keep the same mintxg; it could be >
830 			 * ds_creation_txg if the previous snapshot was
831 			 * deleted too.
832 			 */
833 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
834 			    scn->scn_phys.scn_queue_obj,
835 			    ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
836 			zfs_dbgmsg("destroying ds %llu; in queue; "
837 			    "replacing with %llu",
838 			    (u_longlong_t)ds->ds_object,
839 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
840 		} else {
841 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
842 			    (u_longlong_t)ds->ds_object);
843 		}
844 	} else {
845 		zfs_dbgmsg("destroying ds %llu; ignoring",
846 		    (u_longlong_t)ds->ds_object);
847 	}
848 
849 	/*
850 	 * dsl_scan_sync() should be called after this, and should sync
851 	 * out our changed state, but just to be safe, do it here.
852 	 */
853 	dsl_scan_sync_state(scn, tx);
854 }
855 
856 void
857 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
858 {
859 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
860 	dsl_scan_t *scn = dp->dp_scan;
861 	uint64_t mintxg;
862 
863 	if (scn->scn_phys.scn_state != DSS_SCANNING)
864 		return;
865 
866 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
867 
868 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
869 		scn->scn_phys.scn_bookmark.zb_objset =
870 		    ds->ds_phys->ds_prev_snap_obj;
871 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
872 		    "reset zb_objset to %llu",
873 		    (u_longlong_t)ds->ds_object,
874 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
875 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
876 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
877 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
878 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
879 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
880 		    scn->scn_phys.scn_queue_obj,
881 		    ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
882 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
883 		    "replacing with %llu",
884 		    (u_longlong_t)ds->ds_object,
885 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
886 	}
887 	dsl_scan_sync_state(scn, tx);
888 }
889 
890 void
891 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
892 {
893 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
894 	dsl_scan_t *scn = dp->dp_scan;
895 	uint64_t mintxg;
896 
897 	if (scn->scn_phys.scn_state != DSS_SCANNING)
898 		return;
899 
900 	if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
901 		scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
902 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
903 		    "reset zb_objset to %llu",
904 		    (u_longlong_t)ds1->ds_object,
905 		    (u_longlong_t)ds2->ds_object);
906 	} else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
907 		scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
908 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
909 		    "reset zb_objset to %llu",
910 		    (u_longlong_t)ds2->ds_object,
911 		    (u_longlong_t)ds1->ds_object);
912 	}
913 
914 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
915 	    ds1->ds_object, &mintxg) == 0) {
916 		int err;
917 
918 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
919 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
920 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
921 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
922 		err = zap_add_int_key(dp->dp_meta_objset,
923 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
924 		VERIFY(err == 0 || err == EEXIST);
925 		if (err == EEXIST) {
926 			/* Both were there to begin with */
927 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
928 			    scn->scn_phys.scn_queue_obj,
929 			    ds1->ds_object, mintxg, tx));
930 		}
931 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
932 		    "replacing with %llu",
933 		    (u_longlong_t)ds1->ds_object,
934 		    (u_longlong_t)ds2->ds_object);
935 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
936 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
937 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
938 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
939 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
940 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
941 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
942 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
943 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
944 		    "replacing with %llu",
945 		    (u_longlong_t)ds2->ds_object,
946 		    (u_longlong_t)ds1->ds_object);
947 	}
948 
949 	dsl_scan_sync_state(scn, tx);
950 }
951 
952 struct enqueue_clones_arg {
953 	dmu_tx_t *tx;
954 	uint64_t originobj;
955 };
956 
957 /* ARGSUSED */
958 static int
959 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
960 {
961 	struct enqueue_clones_arg *eca = arg;
962 	dsl_dataset_t *ds;
963 	int err;
964 	dsl_scan_t *scn = dp->dp_scan;
965 
966 	if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
967 		return (0);
968 
969 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
970 	if (err)
971 		return (err);
972 
973 	while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
974 		dsl_dataset_t *prev;
975 		err = dsl_dataset_hold_obj(dp,
976 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
977 
978 		dsl_dataset_rele(ds, FTAG);
979 		if (err)
980 			return (err);
981 		ds = prev;
982 	}
983 	VERIFY(zap_add_int_key(dp->dp_meta_objset,
984 	    scn->scn_phys.scn_queue_obj, ds->ds_object,
985 	    ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
986 	dsl_dataset_rele(ds, FTAG);
987 	return (0);
988 }
989 
990 static void
991 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
992 {
993 	dsl_pool_t *dp = scn->scn_dp;
994 	dsl_dataset_t *ds;
995 	objset_t *os;
996 
997 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
998 
999 	if (dmu_objset_from_ds(ds, &os))
1000 		goto out;
1001 
1002 	/*
1003 	 * Only the ZIL in the head (non-snapshot) is valid.  Even though
1004 	 * snapshots can have ZIL block pointers (which may be the same
1005 	 * BP as in the head), they must be ignored.  So we traverse the
1006 	 * ZIL here, rather than in scan_recurse(), because the regular
1007 	 * snapshot block-sharing rules don't apply to it.
1008 	 */
1009 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
1010 		dsl_scan_zil(dp, &os->os_zil_header);
1011 
1012 	/*
1013 	 * Iterate over the bps in this ds.
1014 	 */
1015 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1016 	dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
1017 
1018 	char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1019 	dsl_dataset_name(ds, dsname);
1020 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1021 	    "pausing=%u",
1022 	    (longlong_t)dsobj, dsname,
1023 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
1024 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
1025 	    (int)scn->scn_pausing);
1026 	kmem_free(dsname, ZFS_MAXNAMELEN);
1027 
1028 	if (scn->scn_pausing)
1029 		goto out;
1030 
1031 	/*
1032 	 * We've finished this pass over this dataset.
1033 	 */
1034 
1035 	/*
1036 	 * If we did not completely visit this dataset, do another pass.
1037 	 */
1038 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1039 		zfs_dbgmsg("incomplete pass; visiting again");
1040 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1041 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1042 		    scn->scn_phys.scn_queue_obj, ds->ds_object,
1043 		    scn->scn_phys.scn_cur_max_txg, tx) == 0);
1044 		goto out;
1045 	}
1046 
1047 	/*
1048 	 * Add descendent datasets to work queue.
1049 	 */
1050 	if (ds->ds_phys->ds_next_snap_obj != 0) {
1051 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1052 		    scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
1053 		    ds->ds_phys->ds_creation_txg, tx) == 0);
1054 	}
1055 	if (ds->ds_phys->ds_num_children > 1) {
1056 		boolean_t usenext = B_FALSE;
1057 		if (ds->ds_phys->ds_next_clones_obj != 0) {
1058 			uint64_t count;
1059 			/*
1060 			 * A bug in a previous version of the code could
1061 			 * cause upgrade_clones_cb() to not set
1062 			 * ds_next_snap_obj when it should, leading to a
1063 			 * missing entry.  Therefore we can only use the
1064 			 * next_clones_obj when its count is correct.
1065 			 */
1066 			int err = zap_count(dp->dp_meta_objset,
1067 			    ds->ds_phys->ds_next_clones_obj, &count);
1068 			if (err == 0 &&
1069 			    count == ds->ds_phys->ds_num_children - 1)
1070 				usenext = B_TRUE;
1071 		}
1072 
1073 		if (usenext) {
1074 			VERIFY0(zap_join_key(dp->dp_meta_objset,
1075 			    ds->ds_phys->ds_next_clones_obj,
1076 			    scn->scn_phys.scn_queue_obj,
1077 			    ds->ds_phys->ds_creation_txg, tx));
1078 		} else {
1079 			struct enqueue_clones_arg eca;
1080 			eca.tx = tx;
1081 			eca.originobj = ds->ds_object;
1082 
1083 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1084 			    enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1085 		}
1086 	}
1087 
1088 out:
1089 	dsl_dataset_rele(ds, FTAG);
1090 }
1091 
1092 /* ARGSUSED */
1093 static int
1094 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1095 {
1096 	dmu_tx_t *tx = arg;
1097 	dsl_dataset_t *ds;
1098 	int err;
1099 	dsl_scan_t *scn = dp->dp_scan;
1100 
1101 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1102 	if (err)
1103 		return (err);
1104 
1105 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
1106 		dsl_dataset_t *prev;
1107 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
1108 		    FTAG, &prev);
1109 		if (err) {
1110 			dsl_dataset_rele(ds, FTAG);
1111 			return (err);
1112 		}
1113 
1114 		/*
1115 		 * If this is a clone, we don't need to worry about it for now.
1116 		 */
1117 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
1118 			dsl_dataset_rele(ds, FTAG);
1119 			dsl_dataset_rele(prev, FTAG);
1120 			return (0);
1121 		}
1122 		dsl_dataset_rele(ds, FTAG);
1123 		ds = prev;
1124 	}
1125 
1126 	VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1127 	    ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
1128 	dsl_dataset_rele(ds, FTAG);
1129 	return (0);
1130 }
1131 
1132 /*
1133  * Scrub/dedup interaction.
1134  *
1135  * If there are N references to a deduped block, we don't want to scrub it
1136  * N times -- ideally, we should scrub it exactly once.
1137  *
1138  * We leverage the fact that the dde's replication class (enum ddt_class)
1139  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1140  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1141  *
1142  * To prevent excess scrubbing, the scrub begins by walking the DDT
1143  * to find all blocks with refcnt > 1, and scrubs each of these once.
1144  * Since there are two replication classes which contain blocks with
1145  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1146  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1147  *
1148  * There would be nothing more to say if a block's refcnt couldn't change
1149  * during a scrub, but of course it can so we must account for changes
1150  * in a block's replication class.
1151  *
1152  * Here's an example of what can occur:
1153  *
1154  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1155  * when visited during the top-down scrub phase, it will be scrubbed twice.
1156  * This negates our scrub optimization, but is otherwise harmless.
1157  *
1158  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1159  * on each visit during the top-down scrub phase, it will never be scrubbed.
1160  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1161  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1162  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1163  * while a scrub is in progress, it scrubs the block right then.
1164  */
1165 static void
1166 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1167 {
1168 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1169 	ddt_entry_t dde = { 0 };
1170 	int error;
1171 	uint64_t n = 0;
1172 
1173 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1174 		ddt_t *ddt;
1175 
1176 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1177 			break;
1178 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1179 		    (longlong_t)ddb->ddb_class,
1180 		    (longlong_t)ddb->ddb_type,
1181 		    (longlong_t)ddb->ddb_checksum,
1182 		    (longlong_t)ddb->ddb_cursor);
1183 
1184 		/* There should be no pending changes to the dedup table */
1185 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1186 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1187 
1188 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1189 		n++;
1190 
1191 		if (dsl_scan_check_pause(scn, NULL))
1192 			break;
1193 	}
1194 
1195 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1196 	    (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1197 	    (int)scn->scn_pausing);
1198 
1199 	ASSERT(error == 0 || error == ENOENT);
1200 	ASSERT(error != ENOENT ||
1201 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1202 }
1203 
1204 /* ARGSUSED */
1205 void
1206 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1207     ddt_entry_t *dde, dmu_tx_t *tx)
1208 {
1209 	const ddt_key_t *ddk = &dde->dde_key;
1210 	ddt_phys_t *ddp = dde->dde_phys;
1211 	blkptr_t bp;
1212 	zbookmark_t zb = { 0 };
1213 
1214 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1215 		return;
1216 
1217 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1218 		if (ddp->ddp_phys_birth == 0 ||
1219 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1220 			continue;
1221 		ddt_bp_create(checksum, ddk, ddp, &bp);
1222 
1223 		scn->scn_visited_this_txg++;
1224 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1225 	}
1226 }
1227 
1228 static void
1229 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1230 {
1231 	dsl_pool_t *dp = scn->scn_dp;
1232 	zap_cursor_t zc;
1233 	zap_attribute_t za;
1234 
1235 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1236 	    scn->scn_phys.scn_ddt_class_max) {
1237 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1238 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1239 		dsl_scan_ddt(scn, tx);
1240 		if (scn->scn_pausing)
1241 			return;
1242 	}
1243 
1244 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1245 		/* First do the MOS & ORIGIN */
1246 
1247 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1248 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1249 		dsl_scan_visit_rootbp(scn, NULL,
1250 		    &dp->dp_meta_rootbp, tx);
1251 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1252 		if (scn->scn_pausing)
1253 			return;
1254 
1255 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1256 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1257 			    enqueue_cb, tx, DS_FIND_CHILDREN));
1258 		} else {
1259 			dsl_scan_visitds(scn,
1260 			    dp->dp_origin_snap->ds_object, tx);
1261 		}
1262 		ASSERT(!scn->scn_pausing);
1263 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
1264 	    ZB_DESTROYED_OBJSET) {
1265 		/*
1266 		 * If we were paused, continue from here.  Note if the
1267 		 * ds we were paused on was deleted, the zb_objset may
1268 		 * be -1, so we will skip this and find a new objset
1269 		 * below.
1270 		 */
1271 		dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1272 		if (scn->scn_pausing)
1273 			return;
1274 	}
1275 
1276 	/*
1277 	 * In case we were paused right at the end of the ds, zero the
1278 	 * bookmark so we don't think that we're still trying to resume.
1279 	 */
1280 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t));
1281 
1282 	/* keep pulling things out of the zap-object-as-queue */
1283 	while (zap_cursor_init(&zc, dp->dp_meta_objset,
1284 	    scn->scn_phys.scn_queue_obj),
1285 	    zap_cursor_retrieve(&zc, &za) == 0) {
1286 		dsl_dataset_t *ds;
1287 		uint64_t dsobj;
1288 
1289 		dsobj = strtonum(za.za_name, NULL);
1290 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1291 		    scn->scn_phys.scn_queue_obj, dsobj, tx));
1292 
1293 		/* Set up min/max txg */
1294 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1295 		if (za.za_first_integer != 0) {
1296 			scn->scn_phys.scn_cur_min_txg =
1297 			    MAX(scn->scn_phys.scn_min_txg,
1298 			    za.za_first_integer);
1299 		} else {
1300 			scn->scn_phys.scn_cur_min_txg =
1301 			    MAX(scn->scn_phys.scn_min_txg,
1302 			    ds->ds_phys->ds_prev_snap_txg);
1303 		}
1304 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1305 		dsl_dataset_rele(ds, FTAG);
1306 
1307 		dsl_scan_visitds(scn, dsobj, tx);
1308 		zap_cursor_fini(&zc);
1309 		if (scn->scn_pausing)
1310 			return;
1311 	}
1312 	zap_cursor_fini(&zc);
1313 }
1314 
1315 static boolean_t
1316 dsl_scan_free_should_pause(dsl_scan_t *scn)
1317 {
1318 	uint64_t elapsed_nanosecs;
1319 
1320 	if (zfs_recover)
1321 		return (B_FALSE);
1322 
1323 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1324 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1325 	    (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1326 	    txg_sync_waiting(scn->scn_dp)) ||
1327 	    spa_shutting_down(scn->scn_dp->dp_spa));
1328 }
1329 
1330 static int
1331 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1332 {
1333 	dsl_scan_t *scn = arg;
1334 
1335 	if (!scn->scn_is_bptree ||
1336 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1337 		if (dsl_scan_free_should_pause(scn))
1338 			return (SET_ERROR(ERESTART));
1339 	}
1340 
1341 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1342 	    dmu_tx_get_txg(tx), bp, 0));
1343 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1344 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1345 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1346 	scn->scn_visited_this_txg++;
1347 	return (0);
1348 }
1349 
1350 boolean_t
1351 dsl_scan_active(dsl_scan_t *scn)
1352 {
1353 	spa_t *spa = scn->scn_dp->dp_spa;
1354 	uint64_t used = 0, comp, uncomp;
1355 
1356 	if (spa->spa_load_state != SPA_LOAD_NONE)
1357 		return (B_FALSE);
1358 	if (spa_shutting_down(spa))
1359 		return (B_FALSE);
1360 	if (scn->scn_phys.scn_state == DSS_SCANNING ||
1361 	    scn->scn_async_destroying)
1362 		return (B_TRUE);
1363 
1364 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1365 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1366 		    &used, &comp, &uncomp);
1367 	}
1368 	return (used != 0);
1369 }
1370 
1371 void
1372 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1373 {
1374 	dsl_scan_t *scn = dp->dp_scan;
1375 	spa_t *spa = dp->dp_spa;
1376 	int err;
1377 
1378 	/*
1379 	 * Check for scn_restart_txg before checking spa_load_state, so
1380 	 * that we can restart an old-style scan while the pool is being
1381 	 * imported (see dsl_scan_init).
1382 	 */
1383 	if (scn->scn_restart_txg != 0 &&
1384 	    scn->scn_restart_txg <= tx->tx_txg) {
1385 		pool_scan_func_t func = POOL_SCAN_SCRUB;
1386 		dsl_scan_done(scn, B_FALSE, tx);
1387 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1388 			func = POOL_SCAN_RESILVER;
1389 		zfs_dbgmsg("restarting scan func=%u txg=%llu",
1390 		    func, tx->tx_txg);
1391 		dsl_scan_setup_sync(&func, tx);
1392 	}
1393 
1394 	if (!dsl_scan_active(scn) ||
1395 	    spa_sync_pass(dp->dp_spa) > 1)
1396 		return;
1397 
1398 	scn->scn_visited_this_txg = 0;
1399 	scn->scn_pausing = B_FALSE;
1400 	scn->scn_sync_start_time = gethrtime();
1401 	spa->spa_scrub_active = B_TRUE;
1402 
1403 	/*
1404 	 * First process the free list.  If we pause the free, don't do
1405 	 * any scanning.  This ensures that there is no free list when
1406 	 * we are scanning, so the scan code doesn't have to worry about
1407 	 * traversing it.
1408 	 */
1409 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1410 		scn->scn_is_bptree = B_FALSE;
1411 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1412 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1413 		err = bpobj_iterate(&dp->dp_free_bpobj,
1414 		    dsl_scan_free_block_cb, scn, tx);
1415 		VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1416 
1417 		if (err == 0 && spa_feature_is_active(spa,
1418 		    SPA_FEATURE_ASYNC_DESTROY)) {
1419 			ASSERT(scn->scn_async_destroying);
1420 			scn->scn_is_bptree = B_TRUE;
1421 			scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1422 			    NULL, ZIO_FLAG_MUSTSUCCEED);
1423 			err = bptree_iterate(dp->dp_meta_objset,
1424 			    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb,
1425 			    scn, tx);
1426 			VERIFY0(zio_wait(scn->scn_zio_root));
1427 
1428 			if (err == 0) {
1429 				/* finished; deactivate async destroy feature */
1430 				spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY,
1431 				    tx);
1432 				ASSERT(!spa_feature_is_active(spa,
1433 				    SPA_FEATURE_ASYNC_DESTROY));
1434 				VERIFY0(zap_remove(dp->dp_meta_objset,
1435 				    DMU_POOL_DIRECTORY_OBJECT,
1436 				    DMU_POOL_BPTREE_OBJ, tx));
1437 				VERIFY0(bptree_free(dp->dp_meta_objset,
1438 				    dp->dp_bptree_obj, tx));
1439 				dp->dp_bptree_obj = 0;
1440 				scn->scn_async_destroying = B_FALSE;
1441 			}
1442 		}
1443 		if (scn->scn_visited_this_txg) {
1444 			zfs_dbgmsg("freed %llu blocks in %llums from "
1445 			    "free_bpobj/bptree txg %llu",
1446 			    (longlong_t)scn->scn_visited_this_txg,
1447 			    (longlong_t)
1448 			    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1449 			    (longlong_t)tx->tx_txg);
1450 			scn->scn_visited_this_txg = 0;
1451 			/*
1452 			 * Re-sync the ddt so that we can further modify
1453 			 * it when doing bprewrite.
1454 			 */
1455 			ddt_sync(spa, tx->tx_txg);
1456 		}
1457 		if (err == ERESTART)
1458 			return;
1459 		/* finished; verify that space accounting went to zero */
1460 		ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
1461 		ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
1462 		ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1463 	}
1464 
1465 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1466 		return;
1467 
1468 	if (scn->scn_done_txg == tx->tx_txg) {
1469 		ASSERT(!scn->scn_pausing);
1470 		/* finished with scan. */
1471 		zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1472 		dsl_scan_done(scn, B_TRUE, tx);
1473 		ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1474 		dsl_scan_sync_state(scn, tx);
1475 		return;
1476 	}
1477 
1478 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1479 	    scn->scn_phys.scn_ddt_class_max) {
1480 		zfs_dbgmsg("doing scan sync txg %llu; "
1481 		    "ddt bm=%llu/%llu/%llu/%llx",
1482 		    (longlong_t)tx->tx_txg,
1483 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1484 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1485 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1486 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1487 		ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1488 		ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1489 		ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1490 		ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1491 	} else {
1492 		zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1493 		    (longlong_t)tx->tx_txg,
1494 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1495 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1496 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1497 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1498 	}
1499 
1500 	scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1501 	    NULL, ZIO_FLAG_CANFAIL);
1502 	dsl_pool_config_enter(dp, FTAG);
1503 	dsl_scan_visit(scn, tx);
1504 	dsl_pool_config_exit(dp, FTAG);
1505 	(void) zio_wait(scn->scn_zio_root);
1506 	scn->scn_zio_root = NULL;
1507 
1508 	zfs_dbgmsg("visited %llu blocks in %llums",
1509 	    (longlong_t)scn->scn_visited_this_txg,
1510 	    (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1511 
1512 	if (!scn->scn_pausing) {
1513 		scn->scn_done_txg = tx->tx_txg + 1;
1514 		zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1515 		    tx->tx_txg, scn->scn_done_txg);
1516 	}
1517 
1518 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1519 		mutex_enter(&spa->spa_scrub_lock);
1520 		while (spa->spa_scrub_inflight > 0) {
1521 			cv_wait(&spa->spa_scrub_io_cv,
1522 			    &spa->spa_scrub_lock);
1523 		}
1524 		mutex_exit(&spa->spa_scrub_lock);
1525 	}
1526 
1527 	dsl_scan_sync_state(scn, tx);
1528 }
1529 
1530 /*
1531  * This will start a new scan, or restart an existing one.
1532  */
1533 void
1534 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1535 {
1536 	if (txg == 0) {
1537 		dmu_tx_t *tx;
1538 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1539 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1540 
1541 		txg = dmu_tx_get_txg(tx);
1542 		dp->dp_scan->scn_restart_txg = txg;
1543 		dmu_tx_commit(tx);
1544 	} else {
1545 		dp->dp_scan->scn_restart_txg = txg;
1546 	}
1547 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1548 }
1549 
1550 boolean_t
1551 dsl_scan_resilvering(dsl_pool_t *dp)
1552 {
1553 	return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1554 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1555 }
1556 
1557 /*
1558  * scrub consumers
1559  */
1560 
1561 static void
1562 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1563 {
1564 	int i;
1565 
1566 	/*
1567 	 * If we resume after a reboot, zab will be NULL; don't record
1568 	 * incomplete stats in that case.
1569 	 */
1570 	if (zab == NULL)
1571 		return;
1572 
1573 	for (i = 0; i < 4; i++) {
1574 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1575 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1576 		if (t & DMU_OT_NEWTYPE)
1577 			t = DMU_OT_OTHER;
1578 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
1579 		int equal;
1580 
1581 		zb->zb_count++;
1582 		zb->zb_asize += BP_GET_ASIZE(bp);
1583 		zb->zb_lsize += BP_GET_LSIZE(bp);
1584 		zb->zb_psize += BP_GET_PSIZE(bp);
1585 		zb->zb_gangs += BP_COUNT_GANG(bp);
1586 
1587 		switch (BP_GET_NDVAS(bp)) {
1588 		case 2:
1589 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1590 			    DVA_GET_VDEV(&bp->blk_dva[1]))
1591 				zb->zb_ditto_2_of_2_samevdev++;
1592 			break;
1593 		case 3:
1594 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1595 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
1596 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1597 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
1598 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1599 			    DVA_GET_VDEV(&bp->blk_dva[2]));
1600 			if (equal == 1)
1601 				zb->zb_ditto_2_of_3_samevdev++;
1602 			else if (equal == 3)
1603 				zb->zb_ditto_3_of_3_samevdev++;
1604 			break;
1605 		}
1606 	}
1607 }
1608 
1609 static void
1610 dsl_scan_scrub_done(zio_t *zio)
1611 {
1612 	spa_t *spa = zio->io_spa;
1613 
1614 	zio_data_buf_free(zio->io_data, zio->io_size);
1615 
1616 	mutex_enter(&spa->spa_scrub_lock);
1617 	spa->spa_scrub_inflight--;
1618 	cv_broadcast(&spa->spa_scrub_io_cv);
1619 
1620 	if (zio->io_error && (zio->io_error != ECKSUM ||
1621 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1622 		spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1623 	}
1624 	mutex_exit(&spa->spa_scrub_lock);
1625 }
1626 
1627 static int
1628 dsl_scan_scrub_cb(dsl_pool_t *dp,
1629     const blkptr_t *bp, const zbookmark_t *zb)
1630 {
1631 	dsl_scan_t *scn = dp->dp_scan;
1632 	size_t size = BP_GET_PSIZE(bp);
1633 	spa_t *spa = dp->dp_spa;
1634 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1635 	boolean_t needs_io;
1636 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1637 	int scan_delay = 0;
1638 
1639 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
1640 	    phys_birth >= scn->scn_phys.scn_max_txg)
1641 		return (0);
1642 
1643 	count_block(dp->dp_blkstats, bp);
1644 
1645 	if (BP_IS_EMBEDDED(bp))
1646 		return (0);
1647 
1648 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1649 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1650 		zio_flags |= ZIO_FLAG_SCRUB;
1651 		needs_io = B_TRUE;
1652 		scan_delay = zfs_scrub_delay;
1653 	} else {
1654 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1655 		zio_flags |= ZIO_FLAG_RESILVER;
1656 		needs_io = B_FALSE;
1657 		scan_delay = zfs_resilver_delay;
1658 	}
1659 
1660 	/* If it's an intent log block, failure is expected. */
1661 	if (zb->zb_level == ZB_ZIL_LEVEL)
1662 		zio_flags |= ZIO_FLAG_SPECULATIVE;
1663 
1664 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1665 		vdev_t *vd = vdev_lookup_top(spa,
1666 		    DVA_GET_VDEV(&bp->blk_dva[d]));
1667 
1668 		/*
1669 		 * Keep track of how much data we've examined so that
1670 		 * zpool(1M) status can make useful progress reports.
1671 		 */
1672 		scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1673 		spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1674 
1675 		/* if it's a resilver, this may not be in the target range */
1676 		if (!needs_io) {
1677 			if (DVA_GET_GANG(&bp->blk_dva[d])) {
1678 				/*
1679 				 * Gang members may be spread across multiple
1680 				 * vdevs, so the best estimate we have is the
1681 				 * scrub range, which has already been checked.
1682 				 * XXX -- it would be better to change our
1683 				 * allocation policy to ensure that all
1684 				 * gang members reside on the same vdev.
1685 				 */
1686 				needs_io = B_TRUE;
1687 			} else {
1688 				needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1689 				    phys_birth, 1);
1690 			}
1691 		}
1692 	}
1693 
1694 	if (needs_io && !zfs_no_scrub_io) {
1695 		vdev_t *rvd = spa->spa_root_vdev;
1696 		uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1697 		void *data = zio_data_buf_alloc(size);
1698 
1699 		mutex_enter(&spa->spa_scrub_lock);
1700 		while (spa->spa_scrub_inflight >= maxinflight)
1701 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1702 		spa->spa_scrub_inflight++;
1703 		mutex_exit(&spa->spa_scrub_lock);
1704 
1705 		/*
1706 		 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1707 		 * then throttle our workload to limit the impact of a scan.
1708 		 */
1709 		if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1710 			delay(scan_delay);
1711 
1712 		zio_nowait(zio_read(NULL, spa, bp, data, size,
1713 		    dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1714 		    zio_flags, zb));
1715 	}
1716 
1717 	/* do not relocate this block */
1718 	return (0);
1719 }
1720 
1721 int
1722 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1723 {
1724 	spa_t *spa = dp->dp_spa;
1725 
1726 	/*
1727 	 * Purge all vdev caches and probe all devices.  We do this here
1728 	 * rather than in sync context because this requires a writer lock
1729 	 * on the spa_config lock, which we can't do from sync context.  The
1730 	 * spa_scrub_reopen flag indicates that vdev_open() should not
1731 	 * attempt to start another scrub.
1732 	 */
1733 	spa_vdev_state_enter(spa, SCL_NONE);
1734 	spa->spa_scrub_reopen = B_TRUE;
1735 	vdev_reopen(spa->spa_root_vdev);
1736 	spa->spa_scrub_reopen = B_FALSE;
1737 	(void) spa_vdev_state_exit(spa, NULL, 0);
1738 
1739 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1740 	    dsl_scan_setup_sync, &func, 0));
1741 }
1742