xref: /titanic_51/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision 3f745f41d6d087602fbb2f748e1baabc3768f5fb)
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 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
42 
43 typedef struct prefetch_data {
44 	kmutex_t pd_mtx;
45 	kcondvar_t pd_cv;
46 	int32_t pd_bytes_fetched;
47 	int pd_flags;
48 	boolean_t pd_cancel;
49 	boolean_t pd_exited;
50 } prefetch_data_t;
51 
52 typedef struct traverse_data {
53 	spa_t *td_spa;
54 	uint64_t td_objset;
55 	blkptr_t *td_rootbp;
56 	uint64_t td_min_txg;
57 	zbookmark_phys_t *td_resume;
58 	int td_flags;
59 	prefetch_data_t *td_pfd;
60 	boolean_t td_paused;
61 	uint64_t td_hole_birth_enabled_txg;
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_phys_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_phys_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_phys_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_subtree_completed(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_phys_t *zb)
179 {
180 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_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_phys_t *zb)
213 {
214 	zbookmark_phys_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 		/*
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 		if (td->td_hole_birth_enabled_txg < td->td_min_txg)
245 			return (0);
246 	} else if (bp->blk_birth <= td->td_min_txg) {
247 		return (0);
248 	}
249 
250 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
251 		uint64_t size = BP_GET_LSIZE(bp);
252 		mutex_enter(&pd->pd_mtx);
253 		ASSERT(pd->pd_bytes_fetched >= 0);
254 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
255 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
256 		pd->pd_bytes_fetched -= size;
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 		arc_flags_t flags = ARC_FLAG_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 		arc_flags_t flags = ARC_FLAG_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 		arc_flags_t flags = ARC_FLAG_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 	if (td->td_flags & TRAVERSE_PRE) {
426 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
427 		    ZB_DNODE_BLKID);
428 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
429 		    td->td_arg);
430 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
431 			return (0);
432 		if (err != 0)
433 			return (err);
434 	}
435 
436 	for (j = 0; j < dnp->dn_nblkptr; j++) {
437 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
438 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
439 		if (err != 0)
440 			break;
441 	}
442 
443 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
444 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
445 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
446 	}
447 
448 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
449 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
450 		    ZB_DNODE_BLKID);
451 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
452 		    td->td_arg);
453 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
454 			return (0);
455 		if (err != 0)
456 			return (err);
457 	}
458 	return (err);
459 }
460 
461 /* ARGSUSED */
462 static int
463 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
464     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
465 {
466 	prefetch_data_t *pfd = arg;
467 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
468 
469 	ASSERT(pfd->pd_bytes_fetched >= 0);
470 	if (bp == NULL)
471 		return (0);
472 	if (pfd->pd_cancel)
473 		return (SET_ERROR(EINTR));
474 
475 	if (!prefetch_needed(pfd, bp))
476 		return (0);
477 
478 	mutex_enter(&pfd->pd_mtx);
479 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
480 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
481 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
482 	cv_broadcast(&pfd->pd_cv);
483 	mutex_exit(&pfd->pd_mtx);
484 
485 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
486 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
487 
488 	return (0);
489 }
490 
491 static void
492 traverse_prefetch_thread(void *arg)
493 {
494 	traverse_data_t *td_main = arg;
495 	traverse_data_t td = *td_main;
496 	zbookmark_phys_t czb;
497 
498 	td.td_func = traverse_prefetcher;
499 	td.td_arg = td_main->td_pfd;
500 	td.td_pfd = NULL;
501 
502 	SET_BOOKMARK(&czb, td.td_objset,
503 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
504 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
505 
506 	mutex_enter(&td_main->td_pfd->pd_mtx);
507 	td_main->td_pfd->pd_exited = B_TRUE;
508 	cv_broadcast(&td_main->td_pfd->pd_cv);
509 	mutex_exit(&td_main->td_pfd->pd_mtx);
510 }
511 
512 /*
513  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
514  * in syncing context).
515  */
516 static int
517 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
518     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
519     blkptr_cb_t func, void *arg)
520 {
521 	traverse_data_t td;
522 	prefetch_data_t pd = { 0 };
523 	zbookmark_phys_t czb;
524 	int err;
525 
526 	ASSERT(ds == NULL || objset == ds->ds_object);
527 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
528 
529 	/*
530 	 * The data prefetching mechanism (the prefetch thread) is incompatible
531 	 * with resuming from a bookmark.
532 	 */
533 	ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
534 
535 	td.td_spa = spa;
536 	td.td_objset = objset;
537 	td.td_rootbp = rootbp;
538 	td.td_min_txg = txg_start;
539 	td.td_resume = resume;
540 	td.td_func = func;
541 	td.td_arg = arg;
542 	td.td_pfd = &pd;
543 	td.td_flags = flags;
544 	td.td_paused = B_FALSE;
545 
546 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
547 		VERIFY(spa_feature_enabled_txg(spa,
548 		    SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
549 	} else {
550 		td.td_hole_birth_enabled_txg = 0;
551 	}
552 
553 	pd.pd_flags = flags;
554 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
555 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
556 
557 	/* See comment on ZIL traversal in dsl_scan_visitds. */
558 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
559 		arc_flags_t flags = ARC_FLAG_WAIT;
560 		objset_phys_t *osp;
561 		arc_buf_t *buf;
562 
563 		err = arc_read(NULL, td.td_spa, rootbp,
564 		    arc_getbuf_func, &buf,
565 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
566 		if (err != 0)
567 			return (err);
568 
569 		osp = buf->b_data;
570 		traverse_zil(&td, &osp->os_zil_header);
571 		(void) arc_buf_remove_ref(buf, &buf);
572 	}
573 
574 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
575 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
576 	    &td, TQ_NOQUEUE))
577 		pd.pd_exited = B_TRUE;
578 
579 	SET_BOOKMARK(&czb, td.td_objset,
580 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
581 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
582 
583 	mutex_enter(&pd.pd_mtx);
584 	pd.pd_cancel = B_TRUE;
585 	cv_broadcast(&pd.pd_cv);
586 	while (!pd.pd_exited)
587 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
588 	mutex_exit(&pd.pd_mtx);
589 
590 	mutex_destroy(&pd.pd_mtx);
591 	cv_destroy(&pd.pd_cv);
592 
593 	return (err);
594 }
595 
596 /*
597  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
598  * in syncing context).
599  */
600 int
601 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
602     blkptr_cb_t func, void *arg)
603 {
604 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
605 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, NULL, flags, func, arg));
606 }
607 
608 int
609 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
610     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
611     blkptr_cb_t func, void *arg)
612 {
613 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
614 	    blkptr, txg_start, resume, flags, func, arg));
615 }
616 
617 /*
618  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
619  */
620 int
621 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
622     blkptr_cb_t func, void *arg)
623 {
624 	int err;
625 	uint64_t obj;
626 	dsl_pool_t *dp = spa_get_dsl(spa);
627 	objset_t *mos = dp->dp_meta_objset;
628 	boolean_t hard = (flags & TRAVERSE_HARD);
629 
630 	/* visit the MOS */
631 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
632 	    txg_start, NULL, flags, func, arg);
633 	if (err != 0)
634 		return (err);
635 
636 	/* visit each dataset */
637 	for (obj = 1; err == 0;
638 	    err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
639 		dmu_object_info_t doi;
640 
641 		err = dmu_object_info(mos, obj, &doi);
642 		if (err != 0) {
643 			if (hard)
644 				continue;
645 			break;
646 		}
647 
648 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
649 			dsl_dataset_t *ds;
650 			uint64_t txg = txg_start;
651 
652 			dsl_pool_config_enter(dp, FTAG);
653 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
654 			dsl_pool_config_exit(dp, FTAG);
655 			if (err != 0) {
656 				if (hard)
657 					continue;
658 				break;
659 			}
660 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
661 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
662 			err = traverse_dataset(ds, txg, flags, func, arg);
663 			dsl_dataset_rele(ds, FTAG);
664 			if (err != 0)
665 				break;
666 		}
667 	}
668 	if (err == ESRCH)
669 		err = 0;
670 	return (err);
671 }
672