xref: /titanic_41/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision 7c667cfc4d78bd7223e69f45e4cc4db8188d1537)
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_phys_t *td_resume;
59 	int td_flags;
60 	prefetch_data_t *td_pfd;
61 	boolean_t td_paused;
62 	uint64_t td_hole_birth_enabled_txg;
63 	blkptr_cb_t *td_func;
64 	void *td_arg;
65 } traverse_data_t;
66 
67 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
68     uint64_t objset, uint64_t object);
69 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
70     uint64_t objset, uint64_t object);
71 
72 static int
73 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
74 {
75 	traverse_data_t *td = arg;
76 	zbookmark_phys_t zb;
77 
78 	if (BP_IS_HOLE(bp))
79 		return (0);
80 
81 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
82 		return (0);
83 
84 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
85 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
86 
87 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
88 
89 	return (0);
90 }
91 
92 static int
93 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
94 {
95 	traverse_data_t *td = arg;
96 
97 	if (lrc->lrc_txtype == TX_WRITE) {
98 		lr_write_t *lr = (lr_write_t *)lrc;
99 		blkptr_t *bp = &lr->lr_blkptr;
100 		zbookmark_phys_t zb;
101 
102 		if (BP_IS_HOLE(bp))
103 			return (0);
104 
105 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
106 			return (0);
107 
108 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
109 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
110 
111 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
112 		    td->td_arg);
113 	}
114 	return (0);
115 }
116 
117 static void
118 traverse_zil(traverse_data_t *td, zil_header_t *zh)
119 {
120 	uint64_t claim_txg = zh->zh_claim_txg;
121 	zilog_t *zilog;
122 
123 	/*
124 	 * We only want to visit blocks that have been claimed but not yet
125 	 * replayed; plus, in read-only mode, blocks that are already stable.
126 	 */
127 	if (claim_txg == 0 && spa_writeable(td->td_spa))
128 		return;
129 
130 	zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
131 
132 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
133 	    claim_txg);
134 
135 	zil_free(zilog);
136 }
137 
138 typedef enum resume_skip {
139 	RESUME_SKIP_ALL,
140 	RESUME_SKIP_NONE,
141 	RESUME_SKIP_CHILDREN
142 } resume_skip_t;
143 
144 /*
145  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
146  * the block indicated by zb does not need to be visited at all. Returns
147  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
148  * resume point. This indicates that this block should be visited but not its
149  * children (since they must have been visited in a previous traversal).
150  * Otherwise returns RESUME_SKIP_NONE.
151  */
152 static resume_skip_t
153 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
154     const zbookmark_phys_t *zb)
155 {
156 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
157 		/*
158 		 * If we already visited this bp & everything below,
159 		 * don't bother doing it again.
160 		 */
161 		if (zbookmark_is_before(dnp, zb, td->td_resume))
162 			return (RESUME_SKIP_ALL);
163 
164 		/*
165 		 * If we found the block we're trying to resume from, zero
166 		 * the bookmark out to indicate that we have resumed.
167 		 */
168 		if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
169 			bzero(td->td_resume, sizeof (*zb));
170 			if (td->td_flags & TRAVERSE_POST)
171 				return (RESUME_SKIP_CHILDREN);
172 		}
173 	}
174 	return (RESUME_SKIP_NONE);
175 }
176 
177 static void
178 traverse_prefetch_metadata(traverse_data_t *td,
179     const blkptr_t *bp, const zbookmark_phys_t *zb)
180 {
181 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
182 
183 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
184 		return;
185 	/*
186 	 * If we are in the process of resuming, don't prefetch, because
187 	 * some children will not be needed (and in fact may have already
188 	 * been freed).
189 	 */
190 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
191 		return;
192 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
193 		return;
194 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
195 		return;
196 
197 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
198 	    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
199 }
200 
201 static boolean_t
202 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
203 {
204 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
205 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
206 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
207 		return (B_FALSE);
208 	return (B_TRUE);
209 }
210 
211 static int
212 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
213     const blkptr_t *bp, const zbookmark_phys_t *zb)
214 {
215 	zbookmark_phys_t czb;
216 	int err = 0;
217 	arc_buf_t *buf = NULL;
218 	prefetch_data_t *pd = td->td_pfd;
219 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
220 
221 	switch (resume_skip_check(td, dnp, zb)) {
222 	case RESUME_SKIP_ALL:
223 		return (0);
224 	case RESUME_SKIP_CHILDREN:
225 		goto post;
226 	case RESUME_SKIP_NONE:
227 		break;
228 	default:
229 		ASSERT(0);
230 	}
231 
232 	if (bp->blk_birth == 0) {
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 		if (td->td_hole_birth_enabled_txg < td->td_min_txg)
246 			return (0);
247 	} else if (bp->blk_birth <= td->td_min_txg) {
248 		return (0);
249 	}
250 
251 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
252 		mutex_enter(&pd->pd_mtx);
253 		ASSERT(pd->pd_blks_fetched >= 0);
254 		while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
255 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
256 		pd->pd_blks_fetched--;
257 		cv_broadcast(&pd->pd_cv);
258 		mutex_exit(&pd->pd_mtx);
259 	}
260 
261 	if (BP_IS_HOLE(bp)) {
262 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
263 		if (err != 0)
264 			goto post;
265 		return (0);
266 	}
267 
268 	if (td->td_flags & TRAVERSE_PRE) {
269 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
270 		    td->td_arg);
271 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
272 			return (0);
273 		if (err != 0)
274 			goto post;
275 	}
276 
277 	if (BP_GET_LEVEL(bp) > 0) {
278 		uint32_t flags = ARC_WAIT;
279 		int i;
280 		blkptr_t *cbp;
281 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
282 
283 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
284 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
285 		if (err != 0)
286 			goto post;
287 		cbp = buf->b_data;
288 
289 		for (i = 0; i < epb; i++) {
290 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
291 			    zb->zb_level - 1,
292 			    zb->zb_blkid * epb + i);
293 			traverse_prefetch_metadata(td, &cbp[i], &czb);
294 		}
295 
296 		/* recursively visitbp() blocks below this */
297 		for (i = 0; i < epb; i++) {
298 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
299 			    zb->zb_level - 1,
300 			    zb->zb_blkid * epb + i);
301 			err = traverse_visitbp(td, dnp, &cbp[i], &czb);
302 			if (err != 0)
303 				break;
304 		}
305 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
306 		uint32_t flags = ARC_WAIT;
307 		int i;
308 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
309 
310 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
311 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
312 		if (err != 0)
313 			goto post;
314 		dnp = buf->b_data;
315 
316 		for (i = 0; i < epb; i++) {
317 			prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
318 			    zb->zb_blkid * epb + i);
319 		}
320 
321 		/* recursively visitbp() blocks below this */
322 		for (i = 0; i < epb; i++) {
323 			err = traverse_dnode(td, &dnp[i], zb->zb_objset,
324 			    zb->zb_blkid * epb + i);
325 			if (err != 0)
326 				break;
327 		}
328 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
329 		uint32_t flags = ARC_WAIT;
330 		objset_phys_t *osp;
331 		dnode_phys_t *dnp;
332 
333 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
334 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
335 		if (err != 0)
336 			goto post;
337 
338 		osp = buf->b_data;
339 		dnp = &osp->os_meta_dnode;
340 		prefetch_dnode_metadata(td, dnp, zb->zb_objset,
341 		    DMU_META_DNODE_OBJECT);
342 		if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
343 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
344 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
345 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
346 			    zb->zb_objset, DMU_USERUSED_OBJECT);
347 		}
348 
349 		err = traverse_dnode(td, dnp, zb->zb_objset,
350 		    DMU_META_DNODE_OBJECT);
351 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
352 			dnp = &osp->os_groupused_dnode;
353 			err = traverse_dnode(td, dnp, zb->zb_objset,
354 			    DMU_GROUPUSED_OBJECT);
355 		}
356 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
357 			dnp = &osp->os_userused_dnode;
358 			err = traverse_dnode(td, dnp, zb->zb_objset,
359 			    DMU_USERUSED_OBJECT);
360 		}
361 	}
362 
363 	if (buf)
364 		(void) arc_buf_remove_ref(buf, &buf);
365 
366 post:
367 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
368 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
369 
370 	if (hard && (err == EIO || err == ECKSUM)) {
371 		/*
372 		 * Ignore this disk error as requested by the HARD flag,
373 		 * and continue traversal.
374 		 */
375 		err = 0;
376 	}
377 
378 	/*
379 	 * If we are stopping here, set td_resume.
380 	 */
381 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
382 		td->td_resume->zb_objset = zb->zb_objset;
383 		td->td_resume->zb_object = zb->zb_object;
384 		td->td_resume->zb_level = 0;
385 		/*
386 		 * If we have stopped on an indirect block (e.g. due to
387 		 * i/o error), we have not visited anything below it.
388 		 * Set the bookmark to the first level-0 block that we need
389 		 * to visit.  This way, the resuming code does not need to
390 		 * deal with resuming from indirect blocks.
391 		 */
392 		td->td_resume->zb_blkid = zb->zb_blkid <<
393 		    (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
394 		td->td_paused = B_TRUE;
395 	}
396 
397 	return (err);
398 }
399 
400 static void
401 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
402     uint64_t objset, uint64_t object)
403 {
404 	int j;
405 	zbookmark_phys_t czb;
406 
407 	for (j = 0; j < dnp->dn_nblkptr; j++) {
408 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
409 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
410 	}
411 
412 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
413 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
414 		traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
415 	}
416 }
417 
418 static int
419 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
420     uint64_t objset, uint64_t object)
421 {
422 	int j, err = 0;
423 	zbookmark_phys_t czb;
424 
425 	for (j = 0; j < dnp->dn_nblkptr; j++) {
426 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
427 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
428 		if (err != 0)
429 			break;
430 	}
431 
432 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
433 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
434 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
435 	}
436 	return (err);
437 }
438 
439 /* ARGSUSED */
440 static int
441 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
442     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
443 {
444 	prefetch_data_t *pfd = arg;
445 	uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
446 
447 	ASSERT(pfd->pd_blks_fetched >= 0);
448 	if (pfd->pd_cancel)
449 		return (SET_ERROR(EINTR));
450 
451 	if (!prefetch_needed(pfd, bp))
452 		return (0);
453 
454 	mutex_enter(&pfd->pd_mtx);
455 	while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
456 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
457 	pfd->pd_blks_fetched++;
458 	cv_broadcast(&pfd->pd_cv);
459 	mutex_exit(&pfd->pd_mtx);
460 
461 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
462 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
463 
464 	return (0);
465 }
466 
467 static void
468 traverse_prefetch_thread(void *arg)
469 {
470 	traverse_data_t *td_main = arg;
471 	traverse_data_t td = *td_main;
472 	zbookmark_phys_t czb;
473 
474 	td.td_func = traverse_prefetcher;
475 	td.td_arg = td_main->td_pfd;
476 	td.td_pfd = NULL;
477 
478 	SET_BOOKMARK(&czb, td.td_objset,
479 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
480 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
481 
482 	mutex_enter(&td_main->td_pfd->pd_mtx);
483 	td_main->td_pfd->pd_exited = B_TRUE;
484 	cv_broadcast(&td_main->td_pfd->pd_cv);
485 	mutex_exit(&td_main->td_pfd->pd_mtx);
486 }
487 
488 /*
489  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
490  * in syncing context).
491  */
492 static int
493 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
494     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
495     blkptr_cb_t func, void *arg)
496 {
497 	traverse_data_t td;
498 	prefetch_data_t pd = { 0 };
499 	zbookmark_phys_t czb;
500 	int err;
501 
502 	ASSERT(ds == NULL || objset == ds->ds_object);
503 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
504 
505 	/*
506 	 * The data prefetching mechanism (the prefetch thread) is incompatible
507 	 * with resuming from a bookmark.
508 	 */
509 	ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
510 
511 	td.td_spa = spa;
512 	td.td_objset = objset;
513 	td.td_rootbp = rootbp;
514 	td.td_min_txg = txg_start;
515 	td.td_resume = resume;
516 	td.td_func = func;
517 	td.td_arg = arg;
518 	td.td_pfd = &pd;
519 	td.td_flags = flags;
520 	td.td_paused = B_FALSE;
521 
522 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
523 		VERIFY(spa_feature_enabled_txg(spa,
524 		    SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
525 	} else {
526 		td.td_hole_birth_enabled_txg = 0;
527 	}
528 
529 	pd.pd_blks_max = zfs_pd_blks_max;
530 	pd.pd_flags = flags;
531 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
532 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
533 
534 	/* See comment on ZIL traversal in dsl_scan_visitds. */
535 	if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
536 		uint32_t flags = ARC_WAIT;
537 		objset_phys_t *osp;
538 		arc_buf_t *buf;
539 
540 		err = arc_read(NULL, td.td_spa, rootbp,
541 		    arc_getbuf_func, &buf,
542 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
543 		if (err != 0)
544 			return (err);
545 
546 		osp = buf->b_data;
547 		traverse_zil(&td, &osp->os_zil_header);
548 		(void) arc_buf_remove_ref(buf, &buf);
549 	}
550 
551 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
552 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
553 	    &td, TQ_NOQUEUE))
554 		pd.pd_exited = B_TRUE;
555 
556 	SET_BOOKMARK(&czb, td.td_objset,
557 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
558 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
559 
560 	mutex_enter(&pd.pd_mtx);
561 	pd.pd_cancel = B_TRUE;
562 	cv_broadcast(&pd.pd_cv);
563 	while (!pd.pd_exited)
564 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
565 	mutex_exit(&pd.pd_mtx);
566 
567 	mutex_destroy(&pd.pd_mtx);
568 	cv_destroy(&pd.pd_cv);
569 
570 	return (err);
571 }
572 
573 /*
574  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
575  * in syncing context).
576  */
577 int
578 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
579     blkptr_cb_t func, void *arg)
580 {
581 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
582 	    &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
583 }
584 
585 int
586 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
587     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
588     blkptr_cb_t func, void *arg)
589 {
590 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
591 	    blkptr, txg_start, resume, flags, func, arg));
592 }
593 
594 /*
595  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
596  */
597 int
598 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
599     blkptr_cb_t func, void *arg)
600 {
601 	int err;
602 	uint64_t obj;
603 	dsl_pool_t *dp = spa_get_dsl(spa);
604 	objset_t *mos = dp->dp_meta_objset;
605 	boolean_t hard = (flags & TRAVERSE_HARD);
606 
607 	/* visit the MOS */
608 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
609 	    txg_start, NULL, flags, func, arg);
610 	if (err != 0)
611 		return (err);
612 
613 	/* visit each dataset */
614 	for (obj = 1; err == 0;
615 	    err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
616 		dmu_object_info_t doi;
617 
618 		err = dmu_object_info(mos, obj, &doi);
619 		if (err != 0) {
620 			if (hard)
621 				continue;
622 			break;
623 		}
624 
625 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
626 			dsl_dataset_t *ds;
627 			uint64_t txg = txg_start;
628 
629 			dsl_pool_config_enter(dp, FTAG);
630 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
631 			dsl_pool_config_exit(dp, FTAG);
632 			if (err != 0) {
633 				if (hard)
634 					continue;
635 				break;
636 			}
637 			if (ds->ds_phys->ds_prev_snap_txg > txg)
638 				txg = ds->ds_phys->ds_prev_snap_txg;
639 			err = traverse_dataset(ds, txg, flags, func, arg);
640 			dsl_dataset_rele(ds, FTAG);
641 			if (err != 0)
642 				break;
643 		}
644 	}
645 	if (err == ESRCH)
646 		err = 0;
647 	return (err);
648 }
649