xref: /NextBSD/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_scan.c (revision 84d351007654069f9643c8e4b4802a7f5f08ee42)
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, 2015 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 unsigned int zfs_top_maxinflight = 32;	/* maximum I/Os per top-level */
61 unsigned int zfs_resilver_delay = 2;	/* number of ticks to delay resilver */
62 unsigned int zfs_scrub_delay = 4;	/* number of ticks to delay scrub */
63 unsigned int zfs_scan_idle = 50;	/* idle window in clock ticks */
64 
65 unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
67 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver
68 						 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 scrub prefetch */
71 
72 SYSCTL_DECL(_vfs_zfs);
73 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
74     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
75 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
76     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
77 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
78     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
79 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
80     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
81 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
82     &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg");
83 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
84     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
85 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
86     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
87 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
88     &zfs_no_scrub_io, 0, "Disable scrub I/O");
89 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
90     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
91 
92 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
93 /* max number of blocks to free in a single TXG */
94 uint64_t zfs_free_max_blocks = UINT64_MAX;
95 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
96     &zfs_free_max_blocks, 0, "Maximum number of blocks to free in one TXG");
97 
98 
99 #define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
100 	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
101 	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
102 
103 extern int zfs_txg_timeout;
104 
105 /*
106  * Enable/disable the processing of the free_bpobj object.
107  */
108 boolean_t zfs_free_bpobj_enabled = B_TRUE;
109 
110 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
111     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
112 
113 /* the order has to match pool_scan_type */
114 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
115 	NULL,
116 	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
117 	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
118 };
119 
120 int
dsl_scan_init(dsl_pool_t * dp,uint64_t txg)121 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
122 {
123 	int err;
124 	dsl_scan_t *scn;
125 	spa_t *spa = dp->dp_spa;
126 	uint64_t f;
127 
128 	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
129 	scn->scn_dp = dp;
130 
131 	/*
132 	 * It's possible that we're resuming a scan after a reboot so
133 	 * make sure that the scan_async_destroying flag is initialized
134 	 * appropriately.
135 	 */
136 	ASSERT(!scn->scn_async_destroying);
137 	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
138 	    SPA_FEATURE_ASYNC_DESTROY);
139 
140 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
141 	    "scrub_func", sizeof (uint64_t), 1, &f);
142 	if (err == 0) {
143 		/*
144 		 * There was an old-style scrub in progress.  Restart a
145 		 * new-style scrub from the beginning.
146 		 */
147 		scn->scn_restart_txg = txg;
148 		zfs_dbgmsg("old-style scrub was in progress; "
149 		    "restarting new-style scrub in txg %llu",
150 		    scn->scn_restart_txg);
151 
152 		/*
153 		 * Load the queue obj from the old location so that it
154 		 * can be freed by dsl_scan_done().
155 		 */
156 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
157 		    "scrub_queue", sizeof (uint64_t), 1,
158 		    &scn->scn_phys.scn_queue_obj);
159 	} else {
160 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
161 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
162 		    &scn->scn_phys);
163 		if (err == ENOENT)
164 			return (0);
165 		else if (err)
166 			return (err);
167 
168 		if (scn->scn_phys.scn_state == DSS_SCANNING &&
169 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
170 			/*
171 			 * A new-type scrub was in progress on an old
172 			 * pool, and the pool was accessed by old
173 			 * software.  Restart from the beginning, since
174 			 * the old software may have changed the pool in
175 			 * the meantime.
176 			 */
177 			scn->scn_restart_txg = txg;
178 			zfs_dbgmsg("new-style scrub was modified "
179 			    "by old software; restarting in txg %llu",
180 			    scn->scn_restart_txg);
181 		}
182 	}
183 
184 	spa_scan_stat_init(spa);
185 	return (0);
186 }
187 
188 void
dsl_scan_fini(dsl_pool_t * dp)189 dsl_scan_fini(dsl_pool_t *dp)
190 {
191 	if (dp->dp_scan) {
192 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
193 		dp->dp_scan = NULL;
194 	}
195 }
196 
197 /* ARGSUSED */
198 static int
dsl_scan_setup_check(void * arg,dmu_tx_t * tx)199 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
200 {
201 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
202 
203 	if (scn->scn_phys.scn_state == DSS_SCANNING)
204 		return (SET_ERROR(EBUSY));
205 
206 	return (0);
207 }
208 
209 static void
dsl_scan_setup_sync(void * arg,dmu_tx_t * tx)210 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
211 {
212 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
213 	pool_scan_func_t *funcp = arg;
214 	dmu_object_type_t ot = 0;
215 	dsl_pool_t *dp = scn->scn_dp;
216 	spa_t *spa = dp->dp_spa;
217 
218 	ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
219 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
220 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
221 	scn->scn_phys.scn_func = *funcp;
222 	scn->scn_phys.scn_state = DSS_SCANNING;
223 	scn->scn_phys.scn_min_txg = 0;
224 	scn->scn_phys.scn_max_txg = tx->tx_txg;
225 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
226 	scn->scn_phys.scn_start_time = gethrestime_sec();
227 	scn->scn_phys.scn_errors = 0;
228 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
229 	scn->scn_restart_txg = 0;
230 	scn->scn_done_txg = 0;
231 	spa_scan_stat_init(spa);
232 
233 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
234 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
235 
236 		/* rewrite all disk labels */
237 		vdev_config_dirty(spa->spa_root_vdev);
238 
239 		if (vdev_resilver_needed(spa->spa_root_vdev,
240 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
241 			spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
242 		} else {
243 			spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
244 		}
245 
246 		spa->spa_scrub_started = B_TRUE;
247 		/*
248 		 * If this is an incremental scrub, limit the DDT scrub phase
249 		 * to just the auto-ditto class (for correctness); the rest
250 		 * of the scrub should go faster using top-down pruning.
251 		 */
252 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
253 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
254 
255 	}
256 
257 	/* back to the generic stuff */
258 
259 	if (dp->dp_blkstats == NULL) {
260 		dp->dp_blkstats =
261 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
262 	}
263 	bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
264 
265 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
266 		ot = DMU_OT_ZAP_OTHER;
267 
268 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
269 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
270 
271 	dsl_scan_sync_state(scn, tx);
272 
273 	spa_history_log_internal(spa, "scan setup", tx,
274 	    "func=%u mintxg=%llu maxtxg=%llu",
275 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
276 }
277 
278 /* ARGSUSED */
279 static void
dsl_scan_done(dsl_scan_t * scn,boolean_t complete,dmu_tx_t * tx)280 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
281 {
282 	static const char *old_names[] = {
283 		"scrub_bookmark",
284 		"scrub_ddt_bookmark",
285 		"scrub_ddt_class_max",
286 		"scrub_queue",
287 		"scrub_min_txg",
288 		"scrub_max_txg",
289 		"scrub_func",
290 		"scrub_errors",
291 		NULL
292 	};
293 
294 	dsl_pool_t *dp = scn->scn_dp;
295 	spa_t *spa = dp->dp_spa;
296 	int i;
297 
298 	/* Remove any remnants of an old-style scrub. */
299 	for (i = 0; old_names[i]; i++) {
300 		(void) zap_remove(dp->dp_meta_objset,
301 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
302 	}
303 
304 	if (scn->scn_phys.scn_queue_obj != 0) {
305 		VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
306 		    scn->scn_phys.scn_queue_obj, tx));
307 		scn->scn_phys.scn_queue_obj = 0;
308 	}
309 
310 	/*
311 	 * If we were "restarted" from a stopped state, don't bother
312 	 * with anything else.
313 	 */
314 	if (scn->scn_phys.scn_state != DSS_SCANNING)
315 		return;
316 
317 	if (complete)
318 		scn->scn_phys.scn_state = DSS_FINISHED;
319 	else
320 		scn->scn_phys.scn_state = DSS_CANCELED;
321 
322 	spa_history_log_internal(spa, "scan done", tx,
323 	    "complete=%u", complete);
324 
325 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
326 		mutex_enter(&spa->spa_scrub_lock);
327 		while (spa->spa_scrub_inflight > 0) {
328 			cv_wait(&spa->spa_scrub_io_cv,
329 			    &spa->spa_scrub_lock);
330 		}
331 		mutex_exit(&spa->spa_scrub_lock);
332 		spa->spa_scrub_started = B_FALSE;
333 		spa->spa_scrub_active = B_FALSE;
334 
335 		/*
336 		 * If the scrub/resilver completed, update all DTLs to
337 		 * reflect this.  Whether it succeeded or not, vacate
338 		 * all temporary scrub DTLs.
339 		 */
340 		vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
341 		    complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
342 		if (complete) {
343 			spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
344 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
345 		}
346 		spa_errlog_rotate(spa);
347 
348 		/*
349 		 * We may have finished replacing a device.
350 		 * Let the async thread assess this and handle the detach.
351 		 */
352 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
353 	}
354 
355 	scn->scn_phys.scn_end_time = gethrestime_sec();
356 }
357 
358 /* ARGSUSED */
359 static int
dsl_scan_cancel_check(void * arg,dmu_tx_t * tx)360 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
361 {
362 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
363 
364 	if (scn->scn_phys.scn_state != DSS_SCANNING)
365 		return (SET_ERROR(ENOENT));
366 	return (0);
367 }
368 
369 /* ARGSUSED */
370 static void
dsl_scan_cancel_sync(void * arg,dmu_tx_t * tx)371 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
372 {
373 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
374 
375 	dsl_scan_done(scn, B_FALSE, tx);
376 	dsl_scan_sync_state(scn, tx);
377 }
378 
379 int
dsl_scan_cancel(dsl_pool_t * dp)380 dsl_scan_cancel(dsl_pool_t *dp)
381 {
382 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
383 	    dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
384 }
385 
386 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
387     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
388     dmu_objset_type_t ostype, dmu_tx_t *tx);
389 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
390     dmu_objset_type_t ostype,
391     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
392 
393 void
dsl_free(dsl_pool_t * dp,uint64_t txg,const blkptr_t * bp)394 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
395 {
396 	zio_free(dp->dp_spa, txg, bp);
397 }
398 
399 void
dsl_free_sync(zio_t * pio,dsl_pool_t * dp,uint64_t txg,const blkptr_t * bpp)400 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
401 {
402 	ASSERT(dsl_pool_sync_context(dp));
403 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
404 	    pio->io_flags));
405 }
406 
407 static uint64_t
dsl_scan_ds_maxtxg(dsl_dataset_t * ds)408 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
409 {
410 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
411 	if (ds->ds_is_snapshot)
412 		return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
413 	return (smt);
414 }
415 
416 static void
dsl_scan_sync_state(dsl_scan_t * scn,dmu_tx_t * tx)417 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
418 {
419 	VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
420 	    DMU_POOL_DIRECTORY_OBJECT,
421 	    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
422 	    &scn->scn_phys, tx));
423 }
424 
425 extern int zfs_vdev_async_write_active_min_dirty_percent;
426 
427 static boolean_t
dsl_scan_check_pause(dsl_scan_t * scn,const zbookmark_phys_t * zb)428 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
429 {
430 	/* we never skip user/group accounting objects */
431 	if (zb && (int64_t)zb->zb_object < 0)
432 		return (B_FALSE);
433 
434 	if (scn->scn_pausing)
435 		return (B_TRUE); /* we're already pausing */
436 
437 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
438 		return (B_FALSE); /* we're resuming */
439 
440 	/* We only know how to resume from level-0 blocks. */
441 	if (zb && zb->zb_level != 0)
442 		return (B_FALSE);
443 
444 	/*
445 	 * We pause if:
446 	 *  - we have scanned for the maximum time: an entire txg
447 	 *    timeout (default 5 sec)
448 	 *  or
449 	 *  - we have scanned for at least the minimum time (default 1 sec
450 	 *    for scrub, 3 sec for resilver), and either we have sufficient
451 	 *    dirty data that we are starting to write more quickly
452 	 *    (default 30%), or someone is explicitly waiting for this txg
453 	 *    to complete.
454 	 *  or
455 	 *  - the spa is shutting down because this pool is being exported
456 	 *    or the machine is rebooting.
457 	 */
458 	int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
459 	    zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
460 	uint64_t elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
461 	int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
462 	if (elapsed_nanosecs / NANOSEC >= zfs_txg_timeout ||
463 	    (NSEC2MSEC(elapsed_nanosecs) > mintime &&
464 	    (txg_sync_waiting(scn->scn_dp) ||
465 	    dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent)) ||
466 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
467 		if (zb) {
468 			dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
469 			    (longlong_t)zb->zb_objset,
470 			    (longlong_t)zb->zb_object,
471 			    (longlong_t)zb->zb_level,
472 			    (longlong_t)zb->zb_blkid);
473 			scn->scn_phys.scn_bookmark = *zb;
474 		}
475 		dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
476 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
477 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
478 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
479 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
480 		scn->scn_pausing = B_TRUE;
481 		return (B_TRUE);
482 	}
483 	return (B_FALSE);
484 }
485 
486 typedef struct zil_scan_arg {
487 	dsl_pool_t	*zsa_dp;
488 	zil_header_t	*zsa_zh;
489 } zil_scan_arg_t;
490 
491 /* ARGSUSED */
492 static int
dsl_scan_zil_block(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)493 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
494 {
495 	zil_scan_arg_t *zsa = arg;
496 	dsl_pool_t *dp = zsa->zsa_dp;
497 	dsl_scan_t *scn = dp->dp_scan;
498 	zil_header_t *zh = zsa->zsa_zh;
499 	zbookmark_phys_t zb;
500 
501 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
502 		return (0);
503 
504 	/*
505 	 * One block ("stubby") can be allocated a long time ago; we
506 	 * want to visit that one because it has been allocated
507 	 * (on-disk) even if it hasn't been claimed (even though for
508 	 * scrub there's nothing to do to it).
509 	 */
510 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
511 		return (0);
512 
513 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
514 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
515 
516 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
517 	return (0);
518 }
519 
520 /* ARGSUSED */
521 static int
dsl_scan_zil_record(zilog_t * zilog,lr_t * lrc,void * arg,uint64_t claim_txg)522 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
523 {
524 	if (lrc->lrc_txtype == TX_WRITE) {
525 		zil_scan_arg_t *zsa = arg;
526 		dsl_pool_t *dp = zsa->zsa_dp;
527 		dsl_scan_t *scn = dp->dp_scan;
528 		zil_header_t *zh = zsa->zsa_zh;
529 		lr_write_t *lr = (lr_write_t *)lrc;
530 		blkptr_t *bp = &lr->lr_blkptr;
531 		zbookmark_phys_t zb;
532 
533 		if (BP_IS_HOLE(bp) ||
534 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
535 			return (0);
536 
537 		/*
538 		 * birth can be < claim_txg if this record's txg is
539 		 * already txg sync'ed (but this log block contains
540 		 * other records that are not synced)
541 		 */
542 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
543 			return (0);
544 
545 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
546 		    lr->lr_foid, ZB_ZIL_LEVEL,
547 		    lr->lr_offset / BP_GET_LSIZE(bp));
548 
549 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
550 	}
551 	return (0);
552 }
553 
554 static void
dsl_scan_zil(dsl_pool_t * dp,zil_header_t * zh)555 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
556 {
557 	uint64_t claim_txg = zh->zh_claim_txg;
558 	zil_scan_arg_t zsa = { dp, zh };
559 	zilog_t *zilog;
560 
561 	/*
562 	 * We only want to visit blocks that have been claimed but not yet
563 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
564 	 */
565 	if (claim_txg == 0 && spa_writeable(dp->dp_spa))
566 		return;
567 
568 	zilog = zil_alloc(dp->dp_meta_objset, zh);
569 
570 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
571 	    claim_txg);
572 
573 	zil_free(zilog);
574 }
575 
576 /* ARGSUSED */
577 static void
dsl_scan_prefetch(dsl_scan_t * scn,arc_buf_t * buf,blkptr_t * bp,uint64_t objset,uint64_t object,uint64_t blkid)578 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
579     uint64_t objset, uint64_t object, uint64_t blkid)
580 {
581 	zbookmark_phys_t czb;
582 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
583 
584 	if (zfs_no_scrub_prefetch)
585 		return;
586 
587 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
588 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
589 		return;
590 
591 	SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
592 
593 	(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
594 	    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
595 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
596 }
597 
598 static boolean_t
dsl_scan_check_resume(dsl_scan_t * scn,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)599 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
600     const zbookmark_phys_t *zb)
601 {
602 	/*
603 	 * We never skip over user/group accounting objects (obj<0)
604 	 */
605 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
606 	    (int64_t)zb->zb_object >= 0) {
607 		/*
608 		 * If we already visited this bp & everything below (in
609 		 * a prior txg sync), don't bother doing it again.
610 		 */
611 		if (zbookmark_subtree_completed(dnp, zb,
612 		    &scn->scn_phys.scn_bookmark))
613 			return (B_TRUE);
614 
615 		/*
616 		 * If we found the block we're trying to resume from, or
617 		 * we went past it to a different object, zero it out to
618 		 * indicate that it's OK to start checking for pausing
619 		 * again.
620 		 */
621 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
622 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
623 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
624 			    (longlong_t)zb->zb_objset,
625 			    (longlong_t)zb->zb_object,
626 			    (longlong_t)zb->zb_level,
627 			    (longlong_t)zb->zb_blkid);
628 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
629 		}
630 	}
631 	return (B_FALSE);
632 }
633 
634 /*
635  * Return nonzero on i/o error.
636  * Return new buf to write out in *bufp.
637  */
638 static int
dsl_scan_recurse(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb,dmu_tx_t * tx)639 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
640     dnode_phys_t *dnp, const blkptr_t *bp,
641     const zbookmark_phys_t *zb, dmu_tx_t *tx)
642 {
643 	dsl_pool_t *dp = scn->scn_dp;
644 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
645 	int err;
646 
647 	if (BP_GET_LEVEL(bp) > 0) {
648 		arc_flags_t flags = ARC_FLAG_WAIT;
649 		int i;
650 		blkptr_t *cbp;
651 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
652 		arc_buf_t *buf;
653 
654 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
655 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
656 		if (err) {
657 			scn->scn_phys.scn_errors++;
658 			return (err);
659 		}
660 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
661 			dsl_scan_prefetch(scn, buf, cbp, zb->zb_objset,
662 			    zb->zb_object, zb->zb_blkid * epb + i);
663 		}
664 		for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
665 			zbookmark_phys_t czb;
666 
667 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
668 			    zb->zb_level - 1,
669 			    zb->zb_blkid * epb + i);
670 			dsl_scan_visitbp(cbp, &czb, dnp,
671 			    ds, scn, ostype, tx);
672 		}
673 		(void) arc_buf_remove_ref(buf, &buf);
674 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
675 		arc_flags_t flags = ARC_FLAG_WAIT;
676 		dnode_phys_t *cdnp;
677 		int i, j;
678 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
679 		arc_buf_t *buf;
680 
681 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
682 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
683 		if (err) {
684 			scn->scn_phys.scn_errors++;
685 			return (err);
686 		}
687 		for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
688 			for (j = 0; j < cdnp->dn_nblkptr; j++) {
689 				blkptr_t *cbp = &cdnp->dn_blkptr[j];
690 				dsl_scan_prefetch(scn, buf, cbp,
691 				    zb->zb_objset, zb->zb_blkid * epb + i, j);
692 			}
693 		}
694 		for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
695 			dsl_scan_visitdnode(scn, ds, ostype,
696 			    cdnp, zb->zb_blkid * epb + i, tx);
697 		}
698 
699 		(void) arc_buf_remove_ref(buf, &buf);
700 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
701 		arc_flags_t flags = ARC_FLAG_WAIT;
702 		objset_phys_t *osp;
703 		arc_buf_t *buf;
704 
705 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
706 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
707 		if (err) {
708 			scn->scn_phys.scn_errors++;
709 			return (err);
710 		}
711 
712 		osp = buf->b_data;
713 
714 		dsl_scan_visitdnode(scn, ds, osp->os_type,
715 		    &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
716 
717 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
718 			/*
719 			 * We also always visit user/group accounting
720 			 * objects, and never skip them, even if we are
721 			 * pausing.  This is necessary so that the space
722 			 * deltas from this txg get integrated.
723 			 */
724 			dsl_scan_visitdnode(scn, ds, osp->os_type,
725 			    &osp->os_groupused_dnode,
726 			    DMU_GROUPUSED_OBJECT, tx);
727 			dsl_scan_visitdnode(scn, ds, osp->os_type,
728 			    &osp->os_userused_dnode,
729 			    DMU_USERUSED_OBJECT, tx);
730 		}
731 		(void) arc_buf_remove_ref(buf, &buf);
732 	}
733 
734 	return (0);
735 }
736 
737 static void
dsl_scan_visitdnode(dsl_scan_t * scn,dsl_dataset_t * ds,dmu_objset_type_t ostype,dnode_phys_t * dnp,uint64_t object,dmu_tx_t * tx)738 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
739     dmu_objset_type_t ostype, dnode_phys_t *dnp,
740     uint64_t object, dmu_tx_t *tx)
741 {
742 	int j;
743 
744 	for (j = 0; j < dnp->dn_nblkptr; j++) {
745 		zbookmark_phys_t czb;
746 
747 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
748 		    dnp->dn_nlevels - 1, j);
749 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
750 		    &czb, dnp, ds, scn, ostype, tx);
751 	}
752 
753 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
754 		zbookmark_phys_t czb;
755 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
756 		    0, DMU_SPILL_BLKID);
757 		dsl_scan_visitbp(&dnp->dn_spill,
758 		    &czb, dnp, ds, scn, ostype, tx);
759 	}
760 }
761 
762 /*
763  * The arguments are in this order because mdb can only print the
764  * first 5; we want them to be useful.
765  */
766 static void
dsl_scan_visitbp(blkptr_t * bp,const zbookmark_phys_t * zb,dnode_phys_t * dnp,dsl_dataset_t * ds,dsl_scan_t * scn,dmu_objset_type_t ostype,dmu_tx_t * tx)767 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
768     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
769     dmu_objset_type_t ostype, dmu_tx_t *tx)
770 {
771 	dsl_pool_t *dp = scn->scn_dp;
772 	arc_buf_t *buf = NULL;
773 	blkptr_t bp_toread = *bp;
774 
775 	/* ASSERT(pbuf == NULL || arc_released(pbuf)); */
776 
777 	if (dsl_scan_check_pause(scn, zb))
778 		return;
779 
780 	if (dsl_scan_check_resume(scn, dnp, zb))
781 		return;
782 
783 	if (BP_IS_HOLE(bp))
784 		return;
785 
786 	scn->scn_visited_this_txg++;
787 
788 	dprintf_bp(bp,
789 	    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
790 	    ds, ds ? ds->ds_object : 0,
791 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
792 	    bp);
793 
794 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
795 		return;
796 
797 	if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx) != 0)
798 		return;
799 
800 	/*
801 	 * If dsl_scan_ddt() has aready visited this block, it will have
802 	 * already done any translations or scrubbing, so don't call the
803 	 * callback again.
804 	 */
805 	if (ddt_class_contains(dp->dp_spa,
806 	    scn->scn_phys.scn_ddt_class_max, bp)) {
807 		ASSERT(buf == NULL);
808 		return;
809 	}
810 
811 	/*
812 	 * If this block is from the future (after cur_max_txg), then we
813 	 * are doing this on behalf of a deleted snapshot, and we will
814 	 * revisit the future block on the next pass of this dataset.
815 	 * Don't scan it now unless we need to because something
816 	 * under it was modified.
817 	 */
818 	if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
819 		scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
820 	}
821 }
822 
823 static void
dsl_scan_visit_rootbp(dsl_scan_t * scn,dsl_dataset_t * ds,blkptr_t * bp,dmu_tx_t * tx)824 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
825     dmu_tx_t *tx)
826 {
827 	zbookmark_phys_t zb;
828 
829 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
830 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
831 	dsl_scan_visitbp(bp, &zb, NULL,
832 	    ds, scn, DMU_OST_NONE, tx);
833 
834 	dprintf_ds(ds, "finished scan%s", "");
835 }
836 
837 void
dsl_scan_ds_destroyed(dsl_dataset_t * ds,dmu_tx_t * tx)838 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
839 {
840 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
841 	dsl_scan_t *scn = dp->dp_scan;
842 	uint64_t mintxg;
843 
844 	if (scn->scn_phys.scn_state != DSS_SCANNING)
845 		return;
846 
847 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
848 		if (ds->ds_is_snapshot) {
849 			/* Note, scn_cur_{min,max}_txg stays the same. */
850 			scn->scn_phys.scn_bookmark.zb_objset =
851 			    dsl_dataset_phys(ds)->ds_next_snap_obj;
852 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
853 			    "reset zb_objset to %llu",
854 			    (u_longlong_t)ds->ds_object,
855 			    (u_longlong_t)dsl_dataset_phys(ds)->
856 			    ds_next_snap_obj);
857 			scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
858 		} else {
859 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
860 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
861 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
862 			    "reset bookmark to -1,0,0,0",
863 			    (u_longlong_t)ds->ds_object);
864 		}
865 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
866 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
867 		ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
868 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
869 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
870 		if (ds->ds_is_snapshot) {
871 			/*
872 			 * We keep the same mintxg; it could be >
873 			 * ds_creation_txg if the previous snapshot was
874 			 * deleted too.
875 			 */
876 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
877 			    scn->scn_phys.scn_queue_obj,
878 			    dsl_dataset_phys(ds)->ds_next_snap_obj,
879 			    mintxg, tx) == 0);
880 			zfs_dbgmsg("destroying ds %llu; in queue; "
881 			    "replacing with %llu",
882 			    (u_longlong_t)ds->ds_object,
883 			    (u_longlong_t)dsl_dataset_phys(ds)->
884 			    ds_next_snap_obj);
885 		} else {
886 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
887 			    (u_longlong_t)ds->ds_object);
888 		}
889 	} else {
890 		zfs_dbgmsg("destroying ds %llu; ignoring",
891 		    (u_longlong_t)ds->ds_object);
892 	}
893 
894 	/*
895 	 * dsl_scan_sync() should be called after this, and should sync
896 	 * out our changed state, but just to be safe, do it here.
897 	 */
898 	dsl_scan_sync_state(scn, tx);
899 }
900 
901 void
dsl_scan_ds_snapshotted(dsl_dataset_t * ds,dmu_tx_t * tx)902 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
903 {
904 	dsl_pool_t *dp = ds->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 	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
912 
913 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
914 		scn->scn_phys.scn_bookmark.zb_objset =
915 		    dsl_dataset_phys(ds)->ds_prev_snap_obj;
916 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
917 		    "reset zb_objset to %llu",
918 		    (u_longlong_t)ds->ds_object,
919 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
920 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
921 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
922 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
923 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
924 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
925 		    scn->scn_phys.scn_queue_obj,
926 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
927 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
928 		    "replacing with %llu",
929 		    (u_longlong_t)ds->ds_object,
930 		    (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
931 	}
932 	dsl_scan_sync_state(scn, tx);
933 }
934 
935 void
dsl_scan_ds_clone_swapped(dsl_dataset_t * ds1,dsl_dataset_t * ds2,dmu_tx_t * tx)936 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
937 {
938 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
939 	dsl_scan_t *scn = dp->dp_scan;
940 	uint64_t mintxg;
941 
942 	if (scn->scn_phys.scn_state != DSS_SCANNING)
943 		return;
944 
945 	if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
946 		scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
947 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
948 		    "reset zb_objset to %llu",
949 		    (u_longlong_t)ds1->ds_object,
950 		    (u_longlong_t)ds2->ds_object);
951 	} else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
952 		scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
953 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
954 		    "reset zb_objset to %llu",
955 		    (u_longlong_t)ds2->ds_object,
956 		    (u_longlong_t)ds1->ds_object);
957 	}
958 
959 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
960 	    ds1->ds_object, &mintxg) == 0) {
961 		int err;
962 
963 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
964 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
965 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
966 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
967 		err = zap_add_int_key(dp->dp_meta_objset,
968 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
969 		VERIFY(err == 0 || err == EEXIST);
970 		if (err == EEXIST) {
971 			/* Both were there to begin with */
972 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
973 			    scn->scn_phys.scn_queue_obj,
974 			    ds1->ds_object, mintxg, tx));
975 		}
976 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
977 		    "replacing with %llu",
978 		    (u_longlong_t)ds1->ds_object,
979 		    (u_longlong_t)ds2->ds_object);
980 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
981 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
982 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
983 		ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
984 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
985 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
986 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
987 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
988 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
989 		    "replacing with %llu",
990 		    (u_longlong_t)ds2->ds_object,
991 		    (u_longlong_t)ds1->ds_object);
992 	}
993 
994 	dsl_scan_sync_state(scn, tx);
995 }
996 
997 struct enqueue_clones_arg {
998 	dmu_tx_t *tx;
999 	uint64_t originobj;
1000 };
1001 
1002 /* ARGSUSED */
1003 static int
enqueue_clones_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)1004 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1005 {
1006 	struct enqueue_clones_arg *eca = arg;
1007 	dsl_dataset_t *ds;
1008 	int err;
1009 	dsl_scan_t *scn = dp->dp_scan;
1010 
1011 	if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != eca->originobj)
1012 		return (0);
1013 
1014 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1015 	if (err)
1016 		return (err);
1017 
1018 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != eca->originobj) {
1019 		dsl_dataset_t *prev;
1020 		err = dsl_dataset_hold_obj(dp,
1021 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1022 
1023 		dsl_dataset_rele(ds, FTAG);
1024 		if (err)
1025 			return (err);
1026 		ds = prev;
1027 	}
1028 	VERIFY(zap_add_int_key(dp->dp_meta_objset,
1029 	    scn->scn_phys.scn_queue_obj, ds->ds_object,
1030 	    dsl_dataset_phys(ds)->ds_prev_snap_txg, eca->tx) == 0);
1031 	dsl_dataset_rele(ds, FTAG);
1032 	return (0);
1033 }
1034 
1035 static void
dsl_scan_visitds(dsl_scan_t * scn,uint64_t dsobj,dmu_tx_t * tx)1036 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
1037 {
1038 	dsl_pool_t *dp = scn->scn_dp;
1039 	dsl_dataset_t *ds;
1040 	objset_t *os;
1041 
1042 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1043 
1044 	if (dmu_objset_from_ds(ds, &os))
1045 		goto out;
1046 
1047 	/*
1048 	 * Only the ZIL in the head (non-snapshot) is valid.  Even though
1049 	 * snapshots can have ZIL block pointers (which may be the same
1050 	 * BP as in the head), they must be ignored.  So we traverse the
1051 	 * ZIL here, rather than in scan_recurse(), because the regular
1052 	 * snapshot block-sharing rules don't apply to it.
1053 	 */
1054 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !ds->ds_is_snapshot)
1055 		dsl_scan_zil(dp, &os->os_zil_header);
1056 
1057 	/*
1058 	 * Iterate over the bps in this ds.
1059 	 */
1060 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
1061 	dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
1062 
1063 	char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
1064 	dsl_dataset_name(ds, dsname);
1065 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
1066 	    "pausing=%u",
1067 	    (longlong_t)dsobj, dsname,
1068 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
1069 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
1070 	    (int)scn->scn_pausing);
1071 	kmem_free(dsname, ZFS_MAXNAMELEN);
1072 
1073 	if (scn->scn_pausing)
1074 		goto out;
1075 
1076 	/*
1077 	 * We've finished this pass over this dataset.
1078 	 */
1079 
1080 	/*
1081 	 * If we did not completely visit this dataset, do another pass.
1082 	 */
1083 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
1084 		zfs_dbgmsg("incomplete pass; visiting again");
1085 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
1086 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1087 		    scn->scn_phys.scn_queue_obj, ds->ds_object,
1088 		    scn->scn_phys.scn_cur_max_txg, tx) == 0);
1089 		goto out;
1090 	}
1091 
1092 	/*
1093 	 * Add descendent datasets to work queue.
1094 	 */
1095 	if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
1096 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
1097 		    scn->scn_phys.scn_queue_obj,
1098 		    dsl_dataset_phys(ds)->ds_next_snap_obj,
1099 		    dsl_dataset_phys(ds)->ds_creation_txg, tx) == 0);
1100 	}
1101 	if (dsl_dataset_phys(ds)->ds_num_children > 1) {
1102 		boolean_t usenext = B_FALSE;
1103 		if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
1104 			uint64_t count;
1105 			/*
1106 			 * A bug in a previous version of the code could
1107 			 * cause upgrade_clones_cb() to not set
1108 			 * ds_next_snap_obj when it should, leading to a
1109 			 * missing entry.  Therefore we can only use the
1110 			 * next_clones_obj when its count is correct.
1111 			 */
1112 			int err = zap_count(dp->dp_meta_objset,
1113 			    dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
1114 			if (err == 0 &&
1115 			    count == dsl_dataset_phys(ds)->ds_num_children - 1)
1116 				usenext = B_TRUE;
1117 		}
1118 
1119 		if (usenext) {
1120 			VERIFY0(zap_join_key(dp->dp_meta_objset,
1121 			    dsl_dataset_phys(ds)->ds_next_clones_obj,
1122 			    scn->scn_phys.scn_queue_obj,
1123 			    dsl_dataset_phys(ds)->ds_creation_txg, tx));
1124 		} else {
1125 			struct enqueue_clones_arg eca;
1126 			eca.tx = tx;
1127 			eca.originobj = ds->ds_object;
1128 
1129 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1130 			    enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
1131 		}
1132 	}
1133 
1134 out:
1135 	dsl_dataset_rele(ds, FTAG);
1136 }
1137 
1138 /* ARGSUSED */
1139 static int
enqueue_cb(dsl_pool_t * dp,dsl_dataset_t * hds,void * arg)1140 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
1141 {
1142 	dmu_tx_t *tx = arg;
1143 	dsl_dataset_t *ds;
1144 	int err;
1145 	dsl_scan_t *scn = dp->dp_scan;
1146 
1147 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
1148 	if (err)
1149 		return (err);
1150 
1151 	while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
1152 		dsl_dataset_t *prev;
1153 		err = dsl_dataset_hold_obj(dp,
1154 		    dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
1155 		if (err) {
1156 			dsl_dataset_rele(ds, FTAG);
1157 			return (err);
1158 		}
1159 
1160 		/*
1161 		 * If this is a clone, we don't need to worry about it for now.
1162 		 */
1163 		if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
1164 			dsl_dataset_rele(ds, FTAG);
1165 			dsl_dataset_rele(prev, FTAG);
1166 			return (0);
1167 		}
1168 		dsl_dataset_rele(ds, FTAG);
1169 		ds = prev;
1170 	}
1171 
1172 	VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1173 	    ds->ds_object, dsl_dataset_phys(ds)->ds_prev_snap_txg, tx) == 0);
1174 	dsl_dataset_rele(ds, FTAG);
1175 	return (0);
1176 }
1177 
1178 /*
1179  * Scrub/dedup interaction.
1180  *
1181  * If there are N references to a deduped block, we don't want to scrub it
1182  * N times -- ideally, we should scrub it exactly once.
1183  *
1184  * We leverage the fact that the dde's replication class (enum ddt_class)
1185  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
1186  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
1187  *
1188  * To prevent excess scrubbing, the scrub begins by walking the DDT
1189  * to find all blocks with refcnt > 1, and scrubs each of these once.
1190  * Since there are two replication classes which contain blocks with
1191  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
1192  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
1193  *
1194  * There would be nothing more to say if a block's refcnt couldn't change
1195  * during a scrub, but of course it can so we must account for changes
1196  * in a block's replication class.
1197  *
1198  * Here's an example of what can occur:
1199  *
1200  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
1201  * when visited during the top-down scrub phase, it will be scrubbed twice.
1202  * This negates our scrub optimization, but is otherwise harmless.
1203  *
1204  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
1205  * on each visit during the top-down scrub phase, it will never be scrubbed.
1206  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
1207  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
1208  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
1209  * while a scrub is in progress, it scrubs the block right then.
1210  */
1211 static void
dsl_scan_ddt(dsl_scan_t * scn,dmu_tx_t * tx)1212 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
1213 {
1214 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
1215 	ddt_entry_t dde = { 0 };
1216 	int error;
1217 	uint64_t n = 0;
1218 
1219 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
1220 		ddt_t *ddt;
1221 
1222 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
1223 			break;
1224 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
1225 		    (longlong_t)ddb->ddb_class,
1226 		    (longlong_t)ddb->ddb_type,
1227 		    (longlong_t)ddb->ddb_checksum,
1228 		    (longlong_t)ddb->ddb_cursor);
1229 
1230 		/* There should be no pending changes to the dedup table */
1231 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
1232 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
1233 
1234 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
1235 		n++;
1236 
1237 		if (dsl_scan_check_pause(scn, NULL))
1238 			break;
1239 	}
1240 
1241 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
1242 	    (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
1243 	    (int)scn->scn_pausing);
1244 
1245 	ASSERT(error == 0 || error == ENOENT);
1246 	ASSERT(error != ENOENT ||
1247 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
1248 }
1249 
1250 /* ARGSUSED */
1251 void
dsl_scan_ddt_entry(dsl_scan_t * scn,enum zio_checksum checksum,ddt_entry_t * dde,dmu_tx_t * tx)1252 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
1253     ddt_entry_t *dde, dmu_tx_t *tx)
1254 {
1255 	const ddt_key_t *ddk = &dde->dde_key;
1256 	ddt_phys_t *ddp = dde->dde_phys;
1257 	blkptr_t bp;
1258 	zbookmark_phys_t zb = { 0 };
1259 
1260 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1261 		return;
1262 
1263 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
1264 		if (ddp->ddp_phys_birth == 0 ||
1265 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
1266 			continue;
1267 		ddt_bp_create(checksum, ddk, ddp, &bp);
1268 
1269 		scn->scn_visited_this_txg++;
1270 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
1271 	}
1272 }
1273 
1274 static void
dsl_scan_visit(dsl_scan_t * scn,dmu_tx_t * tx)1275 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
1276 {
1277 	dsl_pool_t *dp = scn->scn_dp;
1278 	zap_cursor_t zc;
1279 	zap_attribute_t za;
1280 
1281 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1282 	    scn->scn_phys.scn_ddt_class_max) {
1283 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1284 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1285 		dsl_scan_ddt(scn, tx);
1286 		if (scn->scn_pausing)
1287 			return;
1288 	}
1289 
1290 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
1291 		/* First do the MOS & ORIGIN */
1292 
1293 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
1294 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
1295 		dsl_scan_visit_rootbp(scn, NULL,
1296 		    &dp->dp_meta_rootbp, tx);
1297 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
1298 		if (scn->scn_pausing)
1299 			return;
1300 
1301 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
1302 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1303 			    enqueue_cb, tx, DS_FIND_CHILDREN));
1304 		} else {
1305 			dsl_scan_visitds(scn,
1306 			    dp->dp_origin_snap->ds_object, tx);
1307 		}
1308 		ASSERT(!scn->scn_pausing);
1309 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
1310 	    ZB_DESTROYED_OBJSET) {
1311 		/*
1312 		 * If we were paused, continue from here.  Note if the
1313 		 * ds we were paused on was deleted, the zb_objset may
1314 		 * be -1, so we will skip this and find a new objset
1315 		 * below.
1316 		 */
1317 		dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
1318 		if (scn->scn_pausing)
1319 			return;
1320 	}
1321 
1322 	/*
1323 	 * In case we were paused right at the end of the ds, zero the
1324 	 * bookmark so we don't think that we're still trying to resume.
1325 	 */
1326 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
1327 
1328 	/* keep pulling things out of the zap-object-as-queue */
1329 	while (zap_cursor_init(&zc, dp->dp_meta_objset,
1330 	    scn->scn_phys.scn_queue_obj),
1331 	    zap_cursor_retrieve(&zc, &za) == 0) {
1332 		dsl_dataset_t *ds;
1333 		uint64_t dsobj;
1334 
1335 		dsobj = strtonum(za.za_name, NULL);
1336 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1337 		    scn->scn_phys.scn_queue_obj, dsobj, tx));
1338 
1339 		/* Set up min/max txg */
1340 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
1341 		if (za.za_first_integer != 0) {
1342 			scn->scn_phys.scn_cur_min_txg =
1343 			    MAX(scn->scn_phys.scn_min_txg,
1344 			    za.za_first_integer);
1345 		} else {
1346 			scn->scn_phys.scn_cur_min_txg =
1347 			    MAX(scn->scn_phys.scn_min_txg,
1348 			    dsl_dataset_phys(ds)->ds_prev_snap_txg);
1349 		}
1350 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
1351 		dsl_dataset_rele(ds, FTAG);
1352 
1353 		dsl_scan_visitds(scn, dsobj, tx);
1354 		zap_cursor_fini(&zc);
1355 		if (scn->scn_pausing)
1356 			return;
1357 	}
1358 	zap_cursor_fini(&zc);
1359 }
1360 
1361 static boolean_t
dsl_scan_free_should_pause(dsl_scan_t * scn)1362 dsl_scan_free_should_pause(dsl_scan_t *scn)
1363 {
1364 	uint64_t elapsed_nanosecs;
1365 
1366 	if (zfs_recover)
1367 		return (B_FALSE);
1368 
1369 	if (scn->scn_visited_this_txg >= zfs_free_max_blocks)
1370 		return (B_TRUE);
1371 
1372 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1373 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
1374 	    (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1375 	    txg_sync_waiting(scn->scn_dp)) ||
1376 	    spa_shutting_down(scn->scn_dp->dp_spa));
1377 }
1378 
1379 static int
dsl_scan_free_block_cb(void * arg,const blkptr_t * bp,dmu_tx_t * tx)1380 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1381 {
1382 	dsl_scan_t *scn = arg;
1383 
1384 	if (!scn->scn_is_bptree ||
1385 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1386 		if (dsl_scan_free_should_pause(scn))
1387 			return (SET_ERROR(ERESTART));
1388 	}
1389 
1390 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1391 	    dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
1392 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1393 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1394 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1395 	scn->scn_visited_this_txg++;
1396 	return (0);
1397 }
1398 
1399 boolean_t
dsl_scan_active(dsl_scan_t * scn)1400 dsl_scan_active(dsl_scan_t *scn)
1401 {
1402 	spa_t *spa = scn->scn_dp->dp_spa;
1403 	uint64_t used = 0, comp, uncomp;
1404 
1405 	if (spa->spa_load_state != SPA_LOAD_NONE)
1406 		return (B_FALSE);
1407 	if (spa_shutting_down(spa))
1408 		return (B_FALSE);
1409 	if (scn->scn_phys.scn_state == DSS_SCANNING ||
1410 	    (scn->scn_async_destroying && !scn->scn_async_stalled))
1411 		return (B_TRUE);
1412 
1413 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1414 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1415 		    &used, &comp, &uncomp);
1416 	}
1417 	return (used != 0);
1418 }
1419 
1420 void
dsl_scan_sync(dsl_pool_t * dp,dmu_tx_t * tx)1421 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
1422 {
1423 	dsl_scan_t *scn = dp->dp_scan;
1424 	spa_t *spa = dp->dp_spa;
1425 	int err = 0;
1426 
1427 	/*
1428 	 * Check for scn_restart_txg before checking spa_load_state, so
1429 	 * that we can restart an old-style scan while the pool is being
1430 	 * imported (see dsl_scan_init).
1431 	 */
1432 	if (scn->scn_restart_txg != 0 &&
1433 	    scn->scn_restart_txg <= tx->tx_txg) {
1434 		pool_scan_func_t func = POOL_SCAN_SCRUB;
1435 		dsl_scan_done(scn, B_FALSE, tx);
1436 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
1437 			func = POOL_SCAN_RESILVER;
1438 		zfs_dbgmsg("restarting scan func=%u txg=%llu",
1439 		    func, tx->tx_txg);
1440 		dsl_scan_setup_sync(&func, tx);
1441 	}
1442 
1443 	/*
1444 	 * If the scan is inactive due to a stalled async destroy, try again.
1445 	 */
1446 	if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1447 	    spa_sync_pass(dp->dp_spa) > 1)
1448 		return;
1449 
1450 	scn->scn_visited_this_txg = 0;
1451 	scn->scn_pausing = B_FALSE;
1452 	scn->scn_sync_start_time = gethrtime();
1453 	spa->spa_scrub_active = B_TRUE;
1454 
1455 	/*
1456 	 * First process the async destroys.  If we pause, don't do
1457 	 * any scrubbing or resilvering.  This ensures that there are no
1458 	 * async destroys while we are scanning, so the scan code doesn't
1459 	 * have to worry about traversing it.  It is also faster to free the
1460 	 * blocks than to scrub them.
1461 	 */
1462 	if (zfs_free_bpobj_enabled &&
1463 	    spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1464 		scn->scn_is_bptree = B_FALSE;
1465 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1466 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1467 		err = bpobj_iterate(&dp->dp_free_bpobj,
1468 		    dsl_scan_free_block_cb, scn, tx);
1469 		VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1470 
1471 		if (err != 0 && err != ERESTART)
1472 			zfs_panic_recover("error %u from bpobj_iterate()", err);
1473 	}
1474 
1475 	if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
1476 		ASSERT(scn->scn_async_destroying);
1477 		scn->scn_is_bptree = B_TRUE;
1478 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1479 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1480 		err = bptree_iterate(dp->dp_meta_objset,
1481 		    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
1482 		VERIFY0(zio_wait(scn->scn_zio_root));
1483 
1484 		if (err == EIO || err == ECKSUM) {
1485 			err = 0;
1486 		} else if (err != 0 && err != ERESTART) {
1487 			zfs_panic_recover("error %u from "
1488 			    "traverse_dataset_destroyed()", err);
1489 		}
1490 
1491 		if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
1492 			/* finished; deactivate async destroy feature */
1493 			spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
1494 			ASSERT(!spa_feature_is_active(spa,
1495 			    SPA_FEATURE_ASYNC_DESTROY));
1496 			VERIFY0(zap_remove(dp->dp_meta_objset,
1497 			    DMU_POOL_DIRECTORY_OBJECT,
1498 			    DMU_POOL_BPTREE_OBJ, tx));
1499 			VERIFY0(bptree_free(dp->dp_meta_objset,
1500 			    dp->dp_bptree_obj, tx));
1501 			dp->dp_bptree_obj = 0;
1502 			scn->scn_async_destroying = B_FALSE;
1503 			scn->scn_async_stalled = B_FALSE;
1504 		} else {
1505 			/*
1506 			 * If we didn't make progress, mark the async
1507 			 * destroy as stalled, so that we will not initiate
1508 			 * a spa_sync() on its behalf.  Note that we only
1509 			 * check this if we are not finished, because if the
1510 			 * bptree had no blocks for us to visit, we can
1511 			 * finish without "making progress".
1512 			 */
1513 			scn->scn_async_stalled =
1514 			    (scn->scn_visited_this_txg == 0);
1515 		}
1516 	}
1517 	if (scn->scn_visited_this_txg) {
1518 		zfs_dbgmsg("freed %llu blocks in %llums from "
1519 		    "free_bpobj/bptree txg %llu; err=%d",
1520 		    (longlong_t)scn->scn_visited_this_txg,
1521 		    (longlong_t)
1522 		    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
1523 		    (longlong_t)tx->tx_txg, err);
1524 		scn->scn_visited_this_txg = 0;
1525 
1526 		/*
1527 		 * Write out changes to the DDT that may be required as a
1528 		 * result of the blocks freed.  This ensures that the DDT
1529 		 * is clean when a scrub/resilver runs.
1530 		 */
1531 		ddt_sync(spa, tx->tx_txg);
1532 	}
1533 	if (err != 0)
1534 		return;
1535 	if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
1536 	    (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
1537 	    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
1538 	    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
1539 		/*
1540 		 * We have finished background destroying, but there is still
1541 		 * some space left in the dp_free_dir. Transfer this leaked
1542 		 * space to the dp_leak_dir.
1543 		 */
1544 		if (dp->dp_leak_dir == NULL) {
1545 			rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
1546 			(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
1547 			    LEAK_DIR_NAME, tx);
1548 			VERIFY0(dsl_pool_open_special_dir(dp,
1549 			    LEAK_DIR_NAME, &dp->dp_leak_dir));
1550 			rrw_exit(&dp->dp_config_rwlock, FTAG);
1551 		}
1552 		dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
1553 		    dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1554 		    dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1555 		    dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1556 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
1557 		    -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
1558 		    -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
1559 		    -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
1560 	}
1561 	if (!scn->scn_async_destroying) {
1562 		/* finished; verify that space accounting went to zero */
1563 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
1564 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
1565 		ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
1566 	}
1567 
1568 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1569 		return;
1570 
1571 	if (scn->scn_done_txg == tx->tx_txg) {
1572 		ASSERT(!scn->scn_pausing);
1573 		/* finished with scan. */
1574 		zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1575 		dsl_scan_done(scn, B_TRUE, tx);
1576 		ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1577 		dsl_scan_sync_state(scn, tx);
1578 		return;
1579 	}
1580 
1581 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
1582 	    scn->scn_phys.scn_ddt_class_max) {
1583 		zfs_dbgmsg("doing scan sync txg %llu; "
1584 		    "ddt bm=%llu/%llu/%llu/%llx",
1585 		    (longlong_t)tx->tx_txg,
1586 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
1587 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
1588 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
1589 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
1590 		ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
1591 		ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
1592 		ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
1593 		ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
1594 	} else {
1595 		zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
1596 		    (longlong_t)tx->tx_txg,
1597 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
1598 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
1599 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
1600 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
1601 	}
1602 
1603 	scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1604 	    NULL, ZIO_FLAG_CANFAIL);
1605 	dsl_pool_config_enter(dp, FTAG);
1606 	dsl_scan_visit(scn, tx);
1607 	dsl_pool_config_exit(dp, FTAG);
1608 	(void) zio_wait(scn->scn_zio_root);
1609 	scn->scn_zio_root = NULL;
1610 
1611 	zfs_dbgmsg("visited %llu blocks in %llums",
1612 	    (longlong_t)scn->scn_visited_this_txg,
1613 	    (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
1614 
1615 	if (!scn->scn_pausing) {
1616 		scn->scn_done_txg = tx->tx_txg + 1;
1617 		zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1618 		    tx->tx_txg, scn->scn_done_txg);
1619 	}
1620 
1621 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
1622 		mutex_enter(&spa->spa_scrub_lock);
1623 		while (spa->spa_scrub_inflight > 0) {
1624 			cv_wait(&spa->spa_scrub_io_cv,
1625 			    &spa->spa_scrub_lock);
1626 		}
1627 		mutex_exit(&spa->spa_scrub_lock);
1628 	}
1629 
1630 	dsl_scan_sync_state(scn, tx);
1631 }
1632 
1633 /*
1634  * This will start a new scan, or restart an existing one.
1635  */
1636 void
dsl_resilver_restart(dsl_pool_t * dp,uint64_t txg)1637 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
1638 {
1639 	if (txg == 0) {
1640 		dmu_tx_t *tx;
1641 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
1642 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
1643 
1644 		txg = dmu_tx_get_txg(tx);
1645 		dp->dp_scan->scn_restart_txg = txg;
1646 		dmu_tx_commit(tx);
1647 	} else {
1648 		dp->dp_scan->scn_restart_txg = txg;
1649 	}
1650 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
1651 }
1652 
1653 boolean_t
dsl_scan_resilvering(dsl_pool_t * dp)1654 dsl_scan_resilvering(dsl_pool_t *dp)
1655 {
1656 	return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
1657 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
1658 }
1659 
1660 /*
1661  * scrub consumers
1662  */
1663 
1664 static void
count_block(zfs_all_blkstats_t * zab,const blkptr_t * bp)1665 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
1666 {
1667 	int i;
1668 
1669 	/*
1670 	 * If we resume after a reboot, zab will be NULL; don't record
1671 	 * incomplete stats in that case.
1672 	 */
1673 	if (zab == NULL)
1674 		return;
1675 
1676 	for (i = 0; i < 4; i++) {
1677 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
1678 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1679 		if (t & DMU_OT_NEWTYPE)
1680 			t = DMU_OT_OTHER;
1681 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
1682 		int equal;
1683 
1684 		zb->zb_count++;
1685 		zb->zb_asize += BP_GET_ASIZE(bp);
1686 		zb->zb_lsize += BP_GET_LSIZE(bp);
1687 		zb->zb_psize += BP_GET_PSIZE(bp);
1688 		zb->zb_gangs += BP_COUNT_GANG(bp);
1689 
1690 		switch (BP_GET_NDVAS(bp)) {
1691 		case 2:
1692 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1693 			    DVA_GET_VDEV(&bp->blk_dva[1]))
1694 				zb->zb_ditto_2_of_2_samevdev++;
1695 			break;
1696 		case 3:
1697 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1698 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
1699 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
1700 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
1701 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
1702 			    DVA_GET_VDEV(&bp->blk_dva[2]));
1703 			if (equal == 1)
1704 				zb->zb_ditto_2_of_3_samevdev++;
1705 			else if (equal == 3)
1706 				zb->zb_ditto_3_of_3_samevdev++;
1707 			break;
1708 		}
1709 	}
1710 }
1711 
1712 static void
dsl_scan_scrub_done(zio_t * zio)1713 dsl_scan_scrub_done(zio_t *zio)
1714 {
1715 	spa_t *spa = zio->io_spa;
1716 
1717 	zio_data_buf_free(zio->io_data, zio->io_size);
1718 
1719 	mutex_enter(&spa->spa_scrub_lock);
1720 	spa->spa_scrub_inflight--;
1721 	cv_broadcast(&spa->spa_scrub_io_cv);
1722 
1723 	if (zio->io_error && (zio->io_error != ECKSUM ||
1724 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
1725 		spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
1726 	}
1727 	mutex_exit(&spa->spa_scrub_lock);
1728 }
1729 
1730 static int
dsl_scan_scrub_cb(dsl_pool_t * dp,const blkptr_t * bp,const zbookmark_phys_t * zb)1731 dsl_scan_scrub_cb(dsl_pool_t *dp,
1732     const blkptr_t *bp, const zbookmark_phys_t *zb)
1733 {
1734 	dsl_scan_t *scn = dp->dp_scan;
1735 	size_t size = BP_GET_PSIZE(bp);
1736 	spa_t *spa = dp->dp_spa;
1737 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
1738 	boolean_t needs_io;
1739 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
1740 	unsigned int scan_delay = 0;
1741 
1742 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
1743 	    phys_birth >= scn->scn_phys.scn_max_txg)
1744 		return (0);
1745 
1746 	count_block(dp->dp_blkstats, bp);
1747 
1748 	if (BP_IS_EMBEDDED(bp))
1749 		return (0);
1750 
1751 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
1752 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
1753 		zio_flags |= ZIO_FLAG_SCRUB;
1754 		needs_io = B_TRUE;
1755 		scan_delay = zfs_scrub_delay;
1756 	} else {
1757 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
1758 		zio_flags |= ZIO_FLAG_RESILVER;
1759 		needs_io = B_FALSE;
1760 		scan_delay = zfs_resilver_delay;
1761 	}
1762 
1763 	/* If it's an intent log block, failure is expected. */
1764 	if (zb->zb_level == ZB_ZIL_LEVEL)
1765 		zio_flags |= ZIO_FLAG_SPECULATIVE;
1766 
1767 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
1768 		vdev_t *vd = vdev_lookup_top(spa,
1769 		    DVA_GET_VDEV(&bp->blk_dva[d]));
1770 
1771 		/*
1772 		 * Keep track of how much data we've examined so that
1773 		 * zpool(1M) status can make useful progress reports.
1774 		 */
1775 		scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
1776 		spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
1777 
1778 		/* if it's a resilver, this may not be in the target range */
1779 		if (!needs_io) {
1780 			if (DVA_GET_GANG(&bp->blk_dva[d])) {
1781 				/*
1782 				 * Gang members may be spread across multiple
1783 				 * vdevs, so the best estimate we have is the
1784 				 * scrub range, which has already been checked.
1785 				 * XXX -- it would be better to change our
1786 				 * allocation policy to ensure that all
1787 				 * gang members reside on the same vdev.
1788 				 */
1789 				needs_io = B_TRUE;
1790 			} else {
1791 				needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
1792 				    phys_birth, 1);
1793 			}
1794 		}
1795 	}
1796 
1797 	if (needs_io && !zfs_no_scrub_io) {
1798 		vdev_t *rvd = spa->spa_root_vdev;
1799 		uint64_t maxinflight = rvd->vdev_children *
1800 		    MAX(zfs_top_maxinflight, 1);
1801 		void *data = zio_data_buf_alloc(size);
1802 
1803 		mutex_enter(&spa->spa_scrub_lock);
1804 		while (spa->spa_scrub_inflight >= maxinflight)
1805 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1806 		spa->spa_scrub_inflight++;
1807 		mutex_exit(&spa->spa_scrub_lock);
1808 
1809 		/*
1810 		 * If we're seeing recent (zfs_scan_idle) "important" I/Os
1811 		 * then throttle our workload to limit the impact of a scan.
1812 		 */
1813 		if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
1814 			delay(MAX((int)scan_delay, 0));
1815 
1816 		zio_nowait(zio_read(NULL, spa, bp, data, size,
1817 		    dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
1818 		    zio_flags, zb));
1819 	}
1820 
1821 	/* do not relocate this block */
1822 	return (0);
1823 }
1824 
1825 int
dsl_scan(dsl_pool_t * dp,pool_scan_func_t func)1826 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
1827 {
1828 	spa_t *spa = dp->dp_spa;
1829 
1830 	/*
1831 	 * Purge all vdev caches and probe all devices.  We do this here
1832 	 * rather than in sync context because this requires a writer lock
1833 	 * on the spa_config lock, which we can't do from sync context.  The
1834 	 * spa_scrub_reopen flag indicates that vdev_open() should not
1835 	 * attempt to start another scrub.
1836 	 */
1837 	spa_vdev_state_enter(spa, SCL_NONE);
1838 	spa->spa_scrub_reopen = B_TRUE;
1839 	vdev_reopen(spa->spa_root_vdev);
1840 	spa->spa_scrub_reopen = B_FALSE;
1841 	(void) spa_vdev_state_exit(spa, NULL, 0);
1842 
1843 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
1844 	    dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_NONE));
1845 }
1846