xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision 06315b795c0d54f0228e0b8af497a28752dd92da)
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) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dmu_objset.h>
28 #include <sys/dmu_traverse.h>
29 #include <sys/dsl_dataset.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dnode.h>
33 #include <sys/spa.h>
34 #include <sys/zio.h>
35 #include <sys/dmu_impl.h>
36 #include <sys/sa.h>
37 #include <sys/sa_impl.h>
38 #include <sys/callb.h>
39 #include <sys/zfeature.h>
40 
41 int zfs_pd_blks_max = 100;
42 
43 typedef struct prefetch_data {
44 	kmutex_t pd_mtx;
45 	kcondvar_t pd_cv;
46 	int pd_blks_max;
47 	int pd_blks_fetched;
48 	int pd_flags;
49 	boolean_t pd_cancel;
50 	boolean_t pd_exited;
51 } prefetch_data_t;
52 
53 typedef struct traverse_data {
54 	spa_t *td_spa;
55 	uint64_t td_objset;
56 	blkptr_t *td_rootbp;
57 	uint64_t td_min_txg;
58 	zbookmark_t *td_resume;
59 	int td_flags;
60 	prefetch_data_t *td_pfd;
61 	boolean_t td_paused;
62 	blkptr_cb_t *td_func;
63 	void *td_arg;
64 } traverse_data_t;
65 
66 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
67     uint64_t objset, uint64_t object);
68 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
69     uint64_t objset, uint64_t object);
70 
71 static int
72 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
73 {
74 	traverse_data_t *td = arg;
75 	zbookmark_t zb;
76 
77 	if (BP_IS_HOLE(bp))
78 		return (0);
79 
80 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
81 		return (0);
82 
83 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
84 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
85 
86 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
87 
88 	return (0);
89 }
90 
91 static int
92 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
93 {
94 	traverse_data_t *td = arg;
95 
96 	if (lrc->lrc_txtype == TX_WRITE) {
97 		lr_write_t *lr = (lr_write_t *)lrc;
98 		blkptr_t *bp = &lr->lr_blkptr;
99 		zbookmark_t zb;
100 
101 		if (BP_IS_HOLE(bp))
102 			return (0);
103 
104 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
105 			return (0);
106 
107 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
108 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
109 
110 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
111 		    td->td_arg);
112 	}
113 	return (0);
114 }
115 
116 static void
117 traverse_zil(traverse_data_t *td, zil_header_t *zh)
118 {
119 	uint64_t claim_txg = zh->zh_claim_txg;
120 	zilog_t *zilog;
121 
122 	/*
123 	 * We only want to visit blocks that have been claimed but not yet
124 	 * replayed; plus, in read-only mode, blocks that are already stable.
125 	 */
126 	if (claim_txg == 0 && spa_writeable(td->td_spa))
127 		return;
128 
129 	zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
130 
131 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
132 	    claim_txg);
133 
134 	zil_free(zilog);
135 }
136 
137 typedef enum resume_skip {
138 	RESUME_SKIP_ALL,
139 	RESUME_SKIP_NONE,
140 	RESUME_SKIP_CHILDREN
141 } resume_skip_t;
142 
143 /*
144  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
145  * the block indicated by zb does not need to be visited at all. Returns
146  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
147  * resume point. This indicates that this block should be visited but not its
148  * children (since they must have been visited in a previous traversal).
149  * Otherwise returns RESUME_SKIP_NONE.
150  */
151 static resume_skip_t
152 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
153     const zbookmark_t *zb)
154 {
155 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
156 		/*
157 		 * If we already visited this bp & everything below,
158 		 * don't bother doing it again.
159 		 */
160 		if (zbookmark_is_before(dnp, zb, td->td_resume))
161 			return (RESUME_SKIP_ALL);
162 
163 		/*
164 		 * If we found the block we're trying to resume from, zero
165 		 * the bookmark out to indicate that we have resumed.
166 		 */
167 		if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
168 			bzero(td->td_resume, sizeof (*zb));
169 			if (td->td_flags & TRAVERSE_POST)
170 				return (RESUME_SKIP_CHILDREN);
171 		}
172 	}
173 	return (RESUME_SKIP_NONE);
174 }
175 
176 static void
177 traverse_prefetch_metadata(traverse_data_t *td,
178     const blkptr_t *bp, const zbookmark_t *zb)
179 {
180 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
181 
182 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
183 		return;
184 	/*
185 	 * If we are in the process of resuming, don't prefetch, because
186 	 * some children will not be needed (and in fact may have already
187 	 * been freed).
188 	 */
189 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
190 		return;
191 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
192 		return;
193 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
194 		return;
195 
196 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
197 	    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
198 }
199 
200 static boolean_t
201 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
202 {
203 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
204 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
205 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
206 		return (B_FALSE);
207 	return (B_TRUE);
208 }
209 
210 static int
211 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
212     const blkptr_t *bp, const zbookmark_t *zb)
213 {
214 	zbookmark_t czb;
215 	int err = 0;
216 	arc_buf_t *buf = NULL;
217 	prefetch_data_t *pd = td->td_pfd;
218 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
219 
220 	switch (resume_skip_check(td, dnp, zb)) {
221 	case RESUME_SKIP_ALL:
222 		return (0);
223 	case RESUME_SKIP_CHILDREN:
224 		goto post;
225 	case RESUME_SKIP_NONE:
226 		break;
227 	default:
228 		ASSERT(0);
229 	}
230 
231 	if (bp->blk_birth == 0) {
232 		if (spa_feature_is_active(td->td_spa, SPA_FEATURE_HOLE_BIRTH)) {
233 			/*
234 			 * Since this block has a birth time of 0 it must be a
235 			 * hole created before the SPA_FEATURE_HOLE_BIRTH
236 			 * feature was enabled.  If SPA_FEATURE_HOLE_BIRTH
237 			 * was enabled before the min_txg for this traveral we
238 			 * know the hole must have been created before the
239 			 * min_txg for this traveral, so we can skip it. If
240 			 * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg
241 			 * for this traveral we cannot tell if the hole was
242 			 * created before or after the min_txg for this
243 			 * traversal, so we cannot skip it.
244 			 */
245 			uint64_t hole_birth_enabled_txg;
246 			VERIFY(spa_feature_enabled_txg(td->td_spa,
247 			    SPA_FEATURE_HOLE_BIRTH, &hole_birth_enabled_txg));
248 			if (hole_birth_enabled_txg < td->td_min_txg)
249 				return (0);
250 		}
251 	} else if (bp->blk_birth <= td->td_min_txg) {
252 		return (0);
253 	}
254 
255 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
256 		mutex_enter(&pd->pd_mtx);
257 		ASSERT(pd->pd_blks_fetched >= 0);
258 		while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
259 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
260 		pd->pd_blks_fetched--;
261 		cv_broadcast(&pd->pd_cv);
262 		mutex_exit(&pd->pd_mtx);
263 	}
264 
265 	if (BP_IS_HOLE(bp)) {
266 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
267 		if (err != 0)
268 			goto post;
269 		return (0);
270 	}
271 
272 	if (td->td_flags & TRAVERSE_PRE) {
273 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
274 		    td->td_arg);
275 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
276 			return (0);
277 		if (err != 0)
278 			goto post;
279 	}
280 
281 	if (BP_GET_LEVEL(bp) > 0) {
282 		uint32_t flags = ARC_WAIT;
283 		int i;
284 		blkptr_t *cbp;
285 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
286 
287 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
288 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
289 		if (err != 0)
290 			goto post;
291 		cbp = buf->b_data;
292 
293 		for (i = 0; i < epb; i++) {
294 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
295 			    zb->zb_level - 1,
296 			    zb->zb_blkid * epb + i);
297 			traverse_prefetch_metadata(td, &cbp[i], &czb);
298 		}
299 
300 		/* recursively visitbp() blocks below this */
301 		for (i = 0; i < epb; i++) {
302 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
303 			    zb->zb_level - 1,
304 			    zb->zb_blkid * epb + i);
305 			err = traverse_visitbp(td, dnp, &cbp[i], &czb);
306 			if (err != 0)
307 				break;
308 		}
309 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
310 		uint32_t flags = ARC_WAIT;
311 		int i;
312 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
313 
314 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
315 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
316 		if (err != 0)
317 			goto post;
318 		dnp = buf->b_data;
319 
320 		for (i = 0; i < epb; i++) {
321 			prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
322 			    zb->zb_blkid * epb + i);
323 		}
324 
325 		/* recursively visitbp() blocks below this */
326 		for (i = 0; i < epb; i++) {
327 			err = traverse_dnode(td, &dnp[i], zb->zb_objset,
328 			    zb->zb_blkid * epb + i);
329 			if (err != 0)
330 				break;
331 		}
332 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
333 		uint32_t flags = ARC_WAIT;
334 		objset_phys_t *osp;
335 		dnode_phys_t *dnp;
336 
337 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
338 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
339 		if (err != 0)
340 			goto post;
341 
342 		osp = buf->b_data;
343 		dnp = &osp->os_meta_dnode;
344 		prefetch_dnode_metadata(td, dnp, zb->zb_objset,
345 		    DMU_META_DNODE_OBJECT);
346 		if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
347 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
348 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
349 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
350 			    zb->zb_objset, DMU_USERUSED_OBJECT);
351 		}
352 
353 		err = traverse_dnode(td, dnp, zb->zb_objset,
354 		    DMU_META_DNODE_OBJECT);
355 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
356 			dnp = &osp->os_groupused_dnode;
357 			err = traverse_dnode(td, dnp, zb->zb_objset,
358 			    DMU_GROUPUSED_OBJECT);
359 		}
360 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
361 			dnp = &osp->os_userused_dnode;
362 			err = traverse_dnode(td, dnp, zb->zb_objset,
363 			    DMU_USERUSED_OBJECT);
364 		}
365 	}
366 
367 	if (buf)
368 		(void) arc_buf_remove_ref(buf, &buf);
369 
370 post:
371 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
372 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
373 
374 	if (hard && (err == EIO || err == ECKSUM)) {
375 		/*
376 		 * Ignore this disk error as requested by the HARD flag,
377 		 * and continue traversal.
378 		 */
379 		err = 0;
380 	}
381 
382 	/*
383 	 * If we are stopping here, set td_resume.
384 	 */
385 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
386 		td->td_resume->zb_objset = zb->zb_objset;
387 		td->td_resume->zb_object = zb->zb_object;
388 		td->td_resume->zb_level = 0;
389 		/*
390 		 * If we have stopped on an indirect block (e.g. due to
391 		 * i/o error), we have not visited anything below it.
392 		 * Set the bookmark to the first level-0 block that we need
393 		 * to visit.  This way, the resuming code does not need to
394 		 * deal with resuming from indirect blocks.
395 		 */
396 		td->td_resume->zb_blkid = zb->zb_blkid <<
397 		    (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
398 		td->td_paused = B_TRUE;
399 	}
400 
401 	return (err);
402 }
403 
404 static void
405 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
406     uint64_t objset, uint64_t object)
407 {
408 	int j;
409 	zbookmark_t czb;
410 
411 	for (j = 0; j < dnp->dn_nblkptr; j++) {
412 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
413 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
414 	}
415 
416 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
417 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
418 		traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
419 	}
420 }
421 
422 static int
423 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
424     uint64_t objset, uint64_t object)
425 {
426 	int j, err = 0;
427 	zbookmark_t czb;
428 
429 	for (j = 0; j < dnp->dn_nblkptr; j++) {
430 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
431 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
432 		if (err != 0)
433 			break;
434 	}
435 
436 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
437 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
438 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
439 	}
440 	return (err);
441 }
442 
443 /* ARGSUSED */
444 static int
445 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
446     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
447 {
448 	prefetch_data_t *pfd = arg;
449 	uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
450 
451 	ASSERT(pfd->pd_blks_fetched >= 0);
452 	if (pfd->pd_cancel)
453 		return (SET_ERROR(EINTR));
454 
455 	if (!prefetch_needed(pfd, bp))
456 		return (0);
457 
458 	mutex_enter(&pfd->pd_mtx);
459 	while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
460 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
461 	pfd->pd_blks_fetched++;
462 	cv_broadcast(&pfd->pd_cv);
463 	mutex_exit(&pfd->pd_mtx);
464 
465 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
466 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
467 
468 	return (0);
469 }
470 
471 static void
472 traverse_prefetch_thread(void *arg)
473 {
474 	traverse_data_t *td_main = arg;
475 	traverse_data_t td = *td_main;
476 	zbookmark_t czb;
477 
478 	td.td_func = traverse_prefetcher;
479 	td.td_arg = td_main->td_pfd;
480 	td.td_pfd = NULL;
481 
482 	SET_BOOKMARK(&czb, td.td_objset,
483 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
484 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
485 
486 	mutex_enter(&td_main->td_pfd->pd_mtx);
487 	td_main->td_pfd->pd_exited = B_TRUE;
488 	cv_broadcast(&td_main->td_pfd->pd_cv);
489 	mutex_exit(&td_main->td_pfd->pd_mtx);
490 }
491 
492 /*
493  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
494  * in syncing context).
495  */
496 static int
497 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
498     uint64_t txg_start, zbookmark_t *resume, int flags,
499     blkptr_cb_t func, void *arg)
500 {
501 	traverse_data_t td;
502 	prefetch_data_t pd = { 0 };
503 	zbookmark_t czb;
504 	int err;
505 
506 	ASSERT(ds == NULL || objset == ds->ds_object);
507 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
508 
509 	/*
510 	 * The data prefetching mechanism (the prefetch thread) is incompatible
511 	 * with resuming from a bookmark.
512 	 */
513 	ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
514 
515 	td.td_spa = spa;
516 	td.td_objset = objset;
517 	td.td_rootbp = rootbp;
518 	td.td_min_txg = txg_start;
519 	td.td_resume = resume;
520 	td.td_func = func;
521 	td.td_arg = arg;
522 	td.td_pfd = &pd;
523 	td.td_flags = flags;
524 	td.td_paused = B_FALSE;
525 
526 	pd.pd_blks_max = zfs_pd_blks_max;
527 	pd.pd_flags = flags;
528 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
529 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
530 
531 	/* See comment on ZIL traversal in dsl_scan_visitds. */
532 	if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
533 		uint32_t flags = ARC_WAIT;
534 		objset_phys_t *osp;
535 		arc_buf_t *buf;
536 
537 		err = arc_read(NULL, td.td_spa, rootbp,
538 		    arc_getbuf_func, &buf,
539 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
540 		if (err != 0)
541 			return (err);
542 
543 		osp = buf->b_data;
544 		traverse_zil(&td, &osp->os_zil_header);
545 		(void) arc_buf_remove_ref(buf, &buf);
546 	}
547 
548 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
549 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
550 	    &td, TQ_NOQUEUE))
551 		pd.pd_exited = B_TRUE;
552 
553 	SET_BOOKMARK(&czb, td.td_objset,
554 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
555 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
556 
557 	mutex_enter(&pd.pd_mtx);
558 	pd.pd_cancel = B_TRUE;
559 	cv_broadcast(&pd.pd_cv);
560 	while (!pd.pd_exited)
561 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
562 	mutex_exit(&pd.pd_mtx);
563 
564 	mutex_destroy(&pd.pd_mtx);
565 	cv_destroy(&pd.pd_cv);
566 
567 	return (err);
568 }
569 
570 /*
571  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
572  * in syncing context).
573  */
574 int
575 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
576     blkptr_cb_t func, void *arg)
577 {
578 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
579 	    &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
580 }
581 
582 int
583 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
584     uint64_t txg_start, zbookmark_t *resume, int flags,
585     blkptr_cb_t func, void *arg)
586 {
587 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
588 	    blkptr, txg_start, resume, flags, func, arg));
589 }
590 
591 /*
592  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
593  */
594 int
595 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
596     blkptr_cb_t func, void *arg)
597 {
598 	int err;
599 	uint64_t obj;
600 	dsl_pool_t *dp = spa_get_dsl(spa);
601 	objset_t *mos = dp->dp_meta_objset;
602 	boolean_t hard = (flags & TRAVERSE_HARD);
603 
604 	/* visit the MOS */
605 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
606 	    txg_start, NULL, flags, func, arg);
607 	if (err != 0)
608 		return (err);
609 
610 	/* visit each dataset */
611 	for (obj = 1; err == 0;
612 	    err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
613 		dmu_object_info_t doi;
614 
615 		err = dmu_object_info(mos, obj, &doi);
616 		if (err != 0) {
617 			if (hard)
618 				continue;
619 			break;
620 		}
621 
622 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
623 			dsl_dataset_t *ds;
624 			uint64_t txg = txg_start;
625 
626 			dsl_pool_config_enter(dp, FTAG);
627 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
628 			dsl_pool_config_exit(dp, FTAG);
629 			if (err != 0) {
630 				if (hard)
631 					continue;
632 				break;
633 			}
634 			if (ds->ds_phys->ds_prev_snap_txg > txg)
635 				txg = ds->ds_phys->ds_prev_snap_txg;
636 			err = traverse_dataset(ds, txg, flags, func, arg);
637 			dsl_dataset_rele(ds, FTAG);
638 			if (err != 0)
639 				break;
640 		}
641 	}
642 	if (err == ESRCH)
643 		err = 0;
644 	return (err);
645 }
646