xref: /titanic_44/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision d1b18d1a1255ac607d5e072515d727cdfe52f878)
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) 2013 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 	blkptr_cb_t *td_func;
62 	void *td_arg;
63 } traverse_data_t;
64 
65 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
66     uint64_t objset, uint64_t object);
67 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
68     uint64_t objset, uint64_t object);
69 
70 static int
71 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
72 {
73 	traverse_data_t *td = arg;
74 	zbookmark_t zb;
75 
76 	if (BP_IS_HOLE(bp))
77 		return (0);
78 
79 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
80 		return (0);
81 
82 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
83 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
84 
85 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
86 
87 	return (0);
88 }
89 
90 static int
91 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
92 {
93 	traverse_data_t *td = arg;
94 
95 	if (lrc->lrc_txtype == TX_WRITE) {
96 		lr_write_t *lr = (lr_write_t *)lrc;
97 		blkptr_t *bp = &lr->lr_blkptr;
98 		zbookmark_t zb;
99 
100 		if (BP_IS_HOLE(bp))
101 			return (0);
102 
103 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
104 			return (0);
105 
106 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
107 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
108 
109 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
110 		    td->td_arg);
111 	}
112 	return (0);
113 }
114 
115 static void
116 traverse_zil(traverse_data_t *td, zil_header_t *zh)
117 {
118 	uint64_t claim_txg = zh->zh_claim_txg;
119 	zilog_t *zilog;
120 
121 	/*
122 	 * We only want to visit blocks that have been claimed but not yet
123 	 * replayed; plus, in read-only mode, blocks that are already stable.
124 	 */
125 	if (claim_txg == 0 && spa_writeable(td->td_spa))
126 		return;
127 
128 	zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
129 
130 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
131 	    claim_txg);
132 
133 	zil_free(zilog);
134 }
135 
136 typedef enum resume_skip {
137 	RESUME_SKIP_ALL,
138 	RESUME_SKIP_NONE,
139 	RESUME_SKIP_CHILDREN
140 } resume_skip_t;
141 
142 /*
143  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
144  * the block indicated by zb does not need to be visited at all. Returns
145  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
146  * resume point. This indicates that this block should be visited but not its
147  * children (since they must have been visited in a previous traversal).
148  * Otherwise returns RESUME_SKIP_NONE.
149  */
150 static resume_skip_t
151 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
152     const zbookmark_t *zb)
153 {
154 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
155 		/*
156 		 * If we already visited this bp & everything below,
157 		 * don't bother doing it again.
158 		 */
159 		if (zbookmark_is_before(dnp, zb, td->td_resume))
160 			return (RESUME_SKIP_ALL);
161 
162 		/*
163 		 * If we found the block we're trying to resume from, zero
164 		 * the bookmark out to indicate that we have resumed.
165 		 */
166 		ASSERT3U(zb->zb_object, <=, td->td_resume->zb_object);
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_pause(traverse_data_t *td, const zbookmark_t *zb)
178 {
179 	ASSERT(td->td_resume != NULL);
180 	ASSERT0(zb->zb_level);
181 	bcopy(zb, td->td_resume, sizeof (*td->td_resume));
182 }
183 
184 static void
185 traverse_prefetch_metadata(traverse_data_t *td,
186     const blkptr_t *bp, const zbookmark_t *zb)
187 {
188 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
189 
190 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
191 		return;
192 	/*
193 	 * If we are in the process of resuming, don't prefetch, because
194 	 * some children will not be needed (and in fact may have already
195 	 * been freed).
196 	 */
197 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
198 		return;
199 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
200 		return;
201 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
202 		return;
203 
204 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
205 	    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
206 }
207 
208 static int
209 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
210     const blkptr_t *bp, const zbookmark_t *zb)
211 {
212 	zbookmark_t czb;
213 	int err = 0, lasterr = 0;
214 	arc_buf_t *buf = NULL;
215 	prefetch_data_t *pd = td->td_pfd;
216 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
217 	boolean_t pause = B_FALSE;
218 
219 	switch (resume_skip_check(td, dnp, zb)) {
220 	case RESUME_SKIP_ALL:
221 		return (0);
222 	case RESUME_SKIP_CHILDREN:
223 		goto post;
224 	case RESUME_SKIP_NONE:
225 		break;
226 	default:
227 		ASSERT(0);
228 	}
229 
230 	if (bp->blk_birth == 0) {
231 		if (spa_feature_is_active(td->td_spa, SPA_FEATURE_HOLE_BIRTH)) {
232 			/*
233 			 * Since this block has a birth time of 0 it must be a
234 			 * hole created before the SPA_FEATURE_HOLE_BIRTH
235 			 * feature was enabled.  If SPA_FEATURE_HOLE_BIRTH
236 			 * was enabled before the min_txg for this traveral we
237 			 * know the hole must have been created before the
238 			 * min_txg for this traveral, so we can skip it. If
239 			 * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg
240 			 * for this traveral we cannot tell if the hole was
241 			 * created before or after the min_txg for this
242 			 * traversal, so we cannot skip it.
243 			 */
244 			uint64_t hole_birth_enabled_txg;
245 			VERIFY(spa_feature_enabled_txg(td->td_spa,
246 			    SPA_FEATURE_HOLE_BIRTH, &hole_birth_enabled_txg));
247 			if (hole_birth_enabled_txg < td->td_min_txg)
248 				return (0);
249 		}
250 	} else if (bp->blk_birth <= td->td_min_txg) {
251 		return (0);
252 	}
253 
254 	if (BP_IS_HOLE(bp)) {
255 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
256 		return (err);
257 	}
258 
259 	if (pd && !pd->pd_exited &&
260 	    ((pd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
261 	    BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) {
262 		mutex_enter(&pd->pd_mtx);
263 		ASSERT(pd->pd_blks_fetched >= 0);
264 		while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
265 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
266 		pd->pd_blks_fetched--;
267 		cv_broadcast(&pd->pd_cv);
268 		mutex_exit(&pd->pd_mtx);
269 	}
270 
271 	if (td->td_flags & TRAVERSE_PRE) {
272 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
273 		    td->td_arg);
274 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
275 			return (0);
276 		if (err == ERESTART)
277 			pause = B_TRUE; /* handle pausing at a common point */
278 		if (err != 0)
279 			goto post;
280 	}
281 
282 	if (BP_GET_LEVEL(bp) > 0) {
283 		uint32_t flags = ARC_WAIT;
284 		int i;
285 		blkptr_t *cbp;
286 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
287 
288 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
289 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
290 		if (err != 0)
291 			return (err);
292 		cbp = buf->b_data;
293 
294 		for (i = 0; i < epb; i++) {
295 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
296 			    zb->zb_level - 1,
297 			    zb->zb_blkid * epb + i);
298 			traverse_prefetch_metadata(td, &cbp[i], &czb);
299 		}
300 
301 		/* recursively visitbp() blocks below this */
302 		for (i = 0; i < epb; i++) {
303 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
304 			    zb->zb_level - 1,
305 			    zb->zb_blkid * epb + i);
306 			err = traverse_visitbp(td, dnp, &cbp[i], &czb);
307 			if (err != 0) {
308 				if (!hard)
309 					break;
310 				lasterr = err;
311 			}
312 		}
313 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
314 		uint32_t flags = ARC_WAIT;
315 		int i;
316 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
317 
318 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
319 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
320 		if (err != 0)
321 			return (err);
322 		dnp = buf->b_data;
323 
324 		for (i = 0; i < epb; i++) {
325 			prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
326 			    zb->zb_blkid * epb + i);
327 		}
328 
329 		/* recursively visitbp() blocks below this */
330 		for (i = 0; i < epb; i++) {
331 			err = traverse_dnode(td, &dnp[i], zb->zb_objset,
332 			    zb->zb_blkid * epb + i);
333 			if (err != 0) {
334 				if (!hard)
335 					break;
336 				lasterr = err;
337 			}
338 		}
339 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
340 		uint32_t flags = ARC_WAIT;
341 		objset_phys_t *osp;
342 		dnode_phys_t *dnp;
343 
344 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
345 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
346 		if (err != 0)
347 			return (err);
348 
349 		osp = buf->b_data;
350 		dnp = &osp->os_meta_dnode;
351 		prefetch_dnode_metadata(td, dnp, zb->zb_objset,
352 		    DMU_META_DNODE_OBJECT);
353 		if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
354 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
355 			    zb->zb_objset, DMU_USERUSED_OBJECT);
356 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
357 			    zb->zb_objset, DMU_USERUSED_OBJECT);
358 		}
359 
360 		err = traverse_dnode(td, dnp, zb->zb_objset,
361 		    DMU_META_DNODE_OBJECT);
362 		if (err && hard) {
363 			lasterr = err;
364 			err = 0;
365 		}
366 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
367 			dnp = &osp->os_userused_dnode;
368 			err = traverse_dnode(td, dnp, zb->zb_objset,
369 			    DMU_USERUSED_OBJECT);
370 		}
371 		if (err && hard) {
372 			lasterr = err;
373 			err = 0;
374 		}
375 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
376 			dnp = &osp->os_groupused_dnode;
377 			err = traverse_dnode(td, dnp, zb->zb_objset,
378 			    DMU_GROUPUSED_OBJECT);
379 		}
380 	}
381 
382 	if (buf)
383 		(void) arc_buf_remove_ref(buf, &buf);
384 
385 post:
386 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
387 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
388 		if (err == ERESTART)
389 			pause = B_TRUE;
390 	}
391 
392 	if (pause && td->td_resume != NULL) {
393 		ASSERT3U(err, ==, ERESTART);
394 		ASSERT(!hard);
395 		traverse_pause(td, zb);
396 	}
397 
398 	return (err != 0 ? err : lasterr);
399 }
400 
401 static void
402 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
403     uint64_t objset, uint64_t object)
404 {
405 	int j;
406 	zbookmark_t czb;
407 
408 	for (j = 0; j < dnp->dn_nblkptr; j++) {
409 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
410 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
411 	}
412 
413 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
414 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
415 		traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
416 	}
417 }
418 
419 static int
420 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
421     uint64_t objset, uint64_t object)
422 {
423 	int j, err = 0, lasterr = 0;
424 	zbookmark_t czb;
425 	boolean_t hard = (td->td_flags & TRAVERSE_HARD);
426 
427 	for (j = 0; j < dnp->dn_nblkptr; j++) {
428 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
429 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
430 		if (err != 0) {
431 			if (!hard)
432 				break;
433 			lasterr = err;
434 		}
435 	}
436 
437 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
438 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
439 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
440 		if (err != 0) {
441 			if (!hard)
442 				return (err);
443 			lasterr = err;
444 		}
445 	}
446 	return (err != 0 ? err : lasterr);
447 }
448 
449 /* ARGSUSED */
450 static int
451 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
452     const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
453 {
454 	prefetch_data_t *pfd = arg;
455 	uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
456 
457 	ASSERT(pfd->pd_blks_fetched >= 0);
458 	if (pfd->pd_cancel)
459 		return (SET_ERROR(EINTR));
460 
461 	if (BP_IS_HOLE(bp) ||
462 	    !((pfd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
463 	    BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0) ||
464 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
465 		return (0);
466 
467 	mutex_enter(&pfd->pd_mtx);
468 	while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
469 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
470 	pfd->pd_blks_fetched++;
471 	cv_broadcast(&pfd->pd_cv);
472 	mutex_exit(&pfd->pd_mtx);
473 
474 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
475 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
476 
477 	return (0);
478 }
479 
480 static void
481 traverse_prefetch_thread(void *arg)
482 {
483 	traverse_data_t *td_main = arg;
484 	traverse_data_t td = *td_main;
485 	zbookmark_t czb;
486 
487 	td.td_func = traverse_prefetcher;
488 	td.td_arg = td_main->td_pfd;
489 	td.td_pfd = NULL;
490 
491 	SET_BOOKMARK(&czb, td.td_objset,
492 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
493 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
494 
495 	mutex_enter(&td_main->td_pfd->pd_mtx);
496 	td_main->td_pfd->pd_exited = B_TRUE;
497 	cv_broadcast(&td_main->td_pfd->pd_cv);
498 	mutex_exit(&td_main->td_pfd->pd_mtx);
499 }
500 
501 /*
502  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
503  * in syncing context).
504  */
505 static int
506 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
507     uint64_t txg_start, zbookmark_t *resume, int flags,
508     blkptr_cb_t func, void *arg)
509 {
510 	traverse_data_t td;
511 	prefetch_data_t pd = { 0 };
512 	zbookmark_t czb;
513 	int err;
514 
515 	ASSERT(ds == NULL || objset == ds->ds_object);
516 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
517 
518 	/*
519 	 * The data prefetching mechanism (the prefetch thread) is incompatible
520 	 * with resuming from a bookmark.
521 	 */
522 	ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
523 
524 	td.td_spa = spa;
525 	td.td_objset = objset;
526 	td.td_rootbp = rootbp;
527 	td.td_min_txg = txg_start;
528 	td.td_resume = resume;
529 	td.td_func = func;
530 	td.td_arg = arg;
531 	td.td_pfd = &pd;
532 	td.td_flags = flags;
533 
534 	pd.pd_blks_max = zfs_pd_blks_max;
535 	pd.pd_flags = flags;
536 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
537 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
538 
539 	/* See comment on ZIL traversal in dsl_scan_visitds. */
540 	if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
541 		uint32_t flags = ARC_WAIT;
542 		objset_phys_t *osp;
543 		arc_buf_t *buf;
544 
545 		err = arc_read(NULL, td.td_spa, rootbp,
546 		    arc_getbuf_func, &buf,
547 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
548 		if (err != 0)
549 			return (err);
550 
551 		osp = buf->b_data;
552 		traverse_zil(&td, &osp->os_zil_header);
553 		(void) arc_buf_remove_ref(buf, &buf);
554 	}
555 
556 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
557 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
558 	    &td, TQ_NOQUEUE))
559 		pd.pd_exited = B_TRUE;
560 
561 	SET_BOOKMARK(&czb, td.td_objset,
562 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
563 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
564 
565 	mutex_enter(&pd.pd_mtx);
566 	pd.pd_cancel = B_TRUE;
567 	cv_broadcast(&pd.pd_cv);
568 	while (!pd.pd_exited)
569 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
570 	mutex_exit(&pd.pd_mtx);
571 
572 	mutex_destroy(&pd.pd_mtx);
573 	cv_destroy(&pd.pd_cv);
574 
575 	return (err);
576 }
577 
578 /*
579  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
580  * in syncing context).
581  */
582 int
583 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
584     blkptr_cb_t func, void *arg)
585 {
586 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
587 	    &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
588 }
589 
590 int
591 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
592     uint64_t txg_start, zbookmark_t *resume, int flags,
593     blkptr_cb_t func, void *arg)
594 {
595 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
596 	    blkptr, txg_start, resume, flags, func, arg));
597 }
598 
599 /*
600  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
601  */
602 int
603 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
604     blkptr_cb_t func, void *arg)
605 {
606 	int err, lasterr = 0;
607 	uint64_t obj;
608 	dsl_pool_t *dp = spa_get_dsl(spa);
609 	objset_t *mos = dp->dp_meta_objset;
610 	boolean_t hard = (flags & TRAVERSE_HARD);
611 
612 	/* visit the MOS */
613 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
614 	    txg_start, NULL, flags, func, arg);
615 	if (err != 0)
616 		return (err);
617 
618 	/* visit each dataset */
619 	for (obj = 1; err == 0 || (err != ESRCH && hard);
620 	    err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
621 		dmu_object_info_t doi;
622 
623 		err = dmu_object_info(mos, obj, &doi);
624 		if (err != 0) {
625 			if (!hard)
626 				return (err);
627 			lasterr = err;
628 			continue;
629 		}
630 
631 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
632 			dsl_dataset_t *ds;
633 			uint64_t txg = txg_start;
634 
635 			dsl_pool_config_enter(dp, FTAG);
636 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
637 			dsl_pool_config_exit(dp, FTAG);
638 			if (err != 0) {
639 				if (!hard)
640 					return (err);
641 				lasterr = err;
642 				continue;
643 			}
644 			if (ds->ds_phys->ds_prev_snap_txg > txg)
645 				txg = ds->ds_phys->ds_prev_snap_txg;
646 			err = traverse_dataset(ds, txg, flags, func, arg);
647 			dsl_dataset_rele(ds, FTAG);
648 			if (err != 0) {
649 				if (!hard)
650 					return (err);
651 				lasterr = err;
652 			}
653 		}
654 	}
655 	if (err == ESRCH)
656 		err = 0;
657 	return (err != 0 ? err : lasterr);
658 }
659