xref: /illumos-gate/usr/src/uts/common/fs/zfs/dmu_traverse.c (revision ed093b41a93e8563e6e1e5dae0768dda2a7bcc27)
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, 2018 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/spa_impl.h>
35 #include <sys/zio.h>
36 #include <sys/dmu_impl.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/callb.h>
40 #include <sys/zfeature.h>
41 
42 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
43 boolean_t send_holes_without_birth_time = B_TRUE;
44 
45 typedef struct prefetch_data {
46 	kmutex_t pd_mtx;
47 	kcondvar_t pd_cv;
48 	int32_t pd_bytes_fetched;
49 	int pd_flags;
50 	boolean_t pd_cancel;
51 	boolean_t pd_exited;
52 	zbookmark_phys_t pd_resume;
53 } prefetch_data_t;
54 
55 typedef struct traverse_data {
56 	spa_t *td_spa;
57 	uint64_t td_objset;
58 	blkptr_t *td_rootbp;
59 	uint64_t td_min_txg;
60 	zbookmark_phys_t *td_resume;
61 	int td_flags;
62 	prefetch_data_t *td_pfd;
63 	boolean_t td_paused;
64 	uint64_t td_hole_birth_enabled_txg;
65 	blkptr_cb_t *td_func;
66 	void *td_arg;
67 	boolean_t td_realloc_possible;
68 } traverse_data_t;
69 
70 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
71     uint64_t objset, uint64_t object);
72 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
73     uint64_t objset, uint64_t object);
74 
75 static int
76 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
77 {
78 	traverse_data_t *td = arg;
79 	zbookmark_phys_t zb;
80 
81 	if (BP_IS_HOLE(bp))
82 		return (0);
83 
84 	if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa))
85 		return (-1);
86 
87 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
88 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
89 
90 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
91 
92 	return (0);
93 }
94 
95 static int
96 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
97 {
98 	traverse_data_t *td = arg;
99 
100 	if (lrc->lrc_txtype == TX_WRITE) {
101 		lr_write_t *lr = (lr_write_t *)lrc;
102 		blkptr_t *bp = &lr->lr_blkptr;
103 		zbookmark_phys_t zb;
104 
105 		if (BP_IS_HOLE(bp))
106 			return (0);
107 
108 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
109 			return (0);
110 
111 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
112 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
113 
114 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
115 		    td->td_arg);
116 	}
117 	return (0);
118 }
119 
120 static void
121 traverse_zil(traverse_data_t *td, zil_header_t *zh)
122 {
123 	uint64_t claim_txg = zh->zh_claim_txg;
124 
125 	/*
126 	 * We only want to visit blocks that have been claimed but not yet
127 	 * replayed; plus blocks that are already stable in read-only mode.
128 	 */
129 	if (claim_txg == 0 && spa_writeable(td->td_spa))
130 		return;
131 
132 	zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
133 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
134 	    claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
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_subtree_completed(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 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
182 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
183 
184 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
185 		return;
186 	/*
187 	 * If we are in the process of resuming, don't prefetch, because
188 	 * some children will not be needed (and in fact may have already
189 	 * been freed).
190 	 */
191 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
192 		return;
193 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
194 		return;
195 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
196 		return;
197 
198 	if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
199 		zio_flags |= ZIO_FLAG_RAW;
200 
201 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
202 	    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
203 }
204 
205 static boolean_t
206 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
207 {
208 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
209 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
210 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
211 		return (B_FALSE);
212 	return (B_TRUE);
213 }
214 
215 static int
216 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
217     const blkptr_t *bp, const zbookmark_phys_t *zb)
218 {
219 	zbookmark_phys_t czb;
220 	int err = 0;
221 	arc_buf_t *buf = NULL;
222 	prefetch_data_t *pd = td->td_pfd;
223 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
224 
225 	switch (resume_skip_check(td, dnp, zb)) {
226 	case RESUME_SKIP_ALL:
227 		return (0);
228 	case RESUME_SKIP_CHILDREN:
229 		goto post;
230 	case RESUME_SKIP_NONE:
231 		break;
232 	default:
233 		ASSERT(0);
234 	}
235 
236 	if (bp->blk_birth == 0) {
237 		/*
238 		 * Since this block has a birth time of 0 it must be one of
239 		 * two things: a hole created before the
240 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
241 		 * which has always been a hole in an object.
242 		 *
243 		 * If a file is written sparsely, then the unwritten parts of
244 		 * the file were "always holes" -- that is, they have been
245 		 * holes since this object was allocated.  However, we (and
246 		 * our callers) can not necessarily tell when an object was
247 		 * allocated.  Therefore, if it's possible that this object
248 		 * was freed and then its object number reused, we need to
249 		 * visit all the holes with birth==0.
250 		 *
251 		 * If it isn't possible that the object number was reused,
252 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
253 		 * all the blocks we will visit as part of this traversal,
254 		 * then this hole must have always existed, so we can skip
255 		 * it.  We visit blocks born after (exclusive) td_min_txg.
256 		 *
257 		 * Note that the meta-dnode cannot be reallocated.
258 		 */
259 		if (!send_holes_without_birth_time &&
260 		    (!td->td_realloc_possible ||
261 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
262 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
263 			return (0);
264 	} else if (bp->blk_birth <= td->td_min_txg) {
265 		return (0);
266 	}
267 
268 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
269 		uint64_t size = BP_GET_LSIZE(bp);
270 		mutex_enter(&pd->pd_mtx);
271 		ASSERT(pd->pd_bytes_fetched >= 0);
272 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
273 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
274 		pd->pd_bytes_fetched -= size;
275 		cv_broadcast(&pd->pd_cv);
276 		mutex_exit(&pd->pd_mtx);
277 	}
278 
279 	if (BP_IS_HOLE(bp)) {
280 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
281 		if (err != 0)
282 			goto post;
283 		return (0);
284 	}
285 
286 	if (td->td_flags & TRAVERSE_PRE) {
287 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
288 		    td->td_arg);
289 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
290 			return (0);
291 		if (err != 0)
292 			goto post;
293 	}
294 
295 	if (BP_GET_LEVEL(bp) > 0) {
296 		arc_flags_t flags = ARC_FLAG_WAIT;
297 		int i;
298 		blkptr_t *cbp;
299 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
300 
301 		ASSERT(!BP_IS_PROTECTED(bp));
302 
303 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
304 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
305 		if (err != 0)
306 			goto post;
307 		cbp = buf->b_data;
308 
309 		for (i = 0; i < epb; i++) {
310 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
311 			    zb->zb_level - 1,
312 			    zb->zb_blkid * epb + i);
313 			traverse_prefetch_metadata(td, &cbp[i], &czb);
314 		}
315 
316 		/* recursively visitbp() blocks below this */
317 		for (i = 0; i < epb; i++) {
318 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
319 			    zb->zb_level - 1,
320 			    zb->zb_blkid * epb + i);
321 			err = traverse_visitbp(td, dnp, &cbp[i], &czb);
322 			if (err != 0)
323 				break;
324 		}
325 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
326 		arc_flags_t flags = ARC_FLAG_WAIT;
327 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
328 		int i;
329 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
330 
331 		/*
332 		 * dnode blocks might have their bonus buffers encrypted, so
333 		 * we must be careful to honor TRAVERSE_NO_DECRYPT
334 		 */
335 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
336 			zio_flags |= ZIO_FLAG_RAW;
337 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
338 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
339 		if (err != 0)
340 			goto post;
341 		dnode_phys_t *child_dnp = buf->b_data;
342 
343 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
344 			prefetch_dnode_metadata(td, &child_dnp[i],
345 			    zb->zb_objset, zb->zb_blkid * epb + i);
346 		}
347 
348 		/* recursively visitbp() blocks below this */
349 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
350 			err = traverse_dnode(td, &child_dnp[i],
351 			    zb->zb_objset, zb->zb_blkid * epb + i);
352 			if (err != 0)
353 				break;
354 		}
355 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
356 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
357 		arc_flags_t flags = ARC_FLAG_WAIT;
358 
359 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
360 			zio_flags |= ZIO_FLAG_RAW;
361 
362 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
363 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
364 		if (err != 0)
365 			goto post;
366 
367 		objset_phys_t *osp = buf->b_data;
368 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
369 		    DMU_META_DNODE_OBJECT);
370 		/*
371 		 * See the block comment above for the goal of this variable.
372 		 * If the maxblkid of the meta-dnode is 0, then we know that
373 		 * we've never had more than DNODES_PER_BLOCK objects in the
374 		 * dataset, which means we can't have reused any object ids.
375 		 */
376 		if (osp->os_meta_dnode.dn_maxblkid == 0)
377 			td->td_realloc_possible = B_FALSE;
378 
379 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
380 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
381 				prefetch_dnode_metadata(td,
382 				    &osp->os_projectused_dnode,
383 				    zb->zb_objset, DMU_PROJECTUSED_OBJECT);
384 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
385 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
386 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
387 			    zb->zb_objset, DMU_USERUSED_OBJECT);
388 		}
389 
390 		err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
391 		    DMU_META_DNODE_OBJECT);
392 		if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
393 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
394 				err = traverse_dnode(td,
395 				    &osp->os_projectused_dnode, zb->zb_objset,
396 				    DMU_PROJECTUSED_OBJECT);
397 			if (err == 0)
398 				err = traverse_dnode(td,
399 				    &osp->os_groupused_dnode, zb->zb_objset,
400 				    DMU_GROUPUSED_OBJECT);
401 			if (err == 0)
402 				err = traverse_dnode(td,
403 				    &osp->os_userused_dnode, zb->zb_objset,
404 				    DMU_USERUSED_OBJECT);
405 		}
406 	}
407 
408 	if (buf)
409 		arc_buf_destroy(buf, &buf);
410 
411 post:
412 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
413 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
414 
415 	if (hard && (err == EIO || err == ECKSUM)) {
416 		/*
417 		 * Ignore this disk error as requested by the HARD flag,
418 		 * and continue traversal.
419 		 */
420 		err = 0;
421 	}
422 
423 	/*
424 	 * If we are stopping here, set td_resume.
425 	 */
426 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
427 		td->td_resume->zb_objset = zb->zb_objset;
428 		td->td_resume->zb_object = zb->zb_object;
429 		td->td_resume->zb_level = 0;
430 		/*
431 		 * If we have stopped on an indirect block (e.g. due to
432 		 * i/o error), we have not visited anything below it.
433 		 * Set the bookmark to the first level-0 block that we need
434 		 * to visit.  This way, the resuming code does not need to
435 		 * deal with resuming from indirect blocks.
436 		 *
437 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
438 		 * to dereference it.
439 		 */
440 		td->td_resume->zb_blkid = zb->zb_blkid;
441 		if (zb->zb_level > 0) {
442 			td->td_resume->zb_blkid <<= zb->zb_level *
443 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
444 		}
445 		td->td_paused = B_TRUE;
446 	}
447 
448 	return (err);
449 }
450 
451 static void
452 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
453     uint64_t objset, uint64_t object)
454 {
455 	int j;
456 	zbookmark_phys_t czb;
457 
458 	for (j = 0; j < dnp->dn_nblkptr; j++) {
459 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
460 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
461 	}
462 
463 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
464 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
465 		traverse_prefetch_metadata(td, DN_SPILL_BLKPTR(dnp), &czb);
466 	}
467 }
468 
469 static int
470 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
471     uint64_t objset, uint64_t object)
472 {
473 	int j, err = 0;
474 	zbookmark_phys_t czb;
475 
476 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
477 	    object < td->td_resume->zb_object)
478 		return (0);
479 
480 	if (td->td_flags & TRAVERSE_PRE) {
481 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
482 		    ZB_DNODE_BLKID);
483 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
484 		    td->td_arg);
485 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
486 			return (0);
487 		if (err != 0)
488 			return (err);
489 	}
490 
491 	for (j = 0; j < dnp->dn_nblkptr; j++) {
492 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
493 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
494 		if (err != 0)
495 			break;
496 	}
497 
498 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
499 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
500 		err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
501 	}
502 
503 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
504 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
505 		    ZB_DNODE_BLKID);
506 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
507 		    td->td_arg);
508 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
509 			return (0);
510 		if (err != 0)
511 			return (err);
512 	}
513 	return (err);
514 }
515 
516 /* ARGSUSED */
517 static int
518 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
519     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
520 {
521 	prefetch_data_t *pfd = arg;
522 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
523 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
524 	    ARC_FLAG_PRESCIENT_PREFETCH;
525 
526 	ASSERT(pfd->pd_bytes_fetched >= 0);
527 	if (bp == NULL)
528 		return (0);
529 	if (pfd->pd_cancel)
530 		return (SET_ERROR(EINTR));
531 
532 	if (!prefetch_needed(pfd, bp))
533 		return (0);
534 
535 	mutex_enter(&pfd->pd_mtx);
536 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
537 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
538 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
539 	cv_broadcast(&pfd->pd_cv);
540 	mutex_exit(&pfd->pd_mtx);
541 
542 	if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
543 		zio_flags |= ZIO_FLAG_RAW;
544 
545 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
546 	    zio_flags, &aflags, zb);
547 
548 	return (0);
549 }
550 
551 static void
552 traverse_prefetch_thread(void *arg)
553 {
554 	traverse_data_t *td_main = arg;
555 	traverse_data_t td = *td_main;
556 	zbookmark_phys_t czb;
557 
558 	td.td_func = traverse_prefetcher;
559 	td.td_arg = td_main->td_pfd;
560 	td.td_pfd = NULL;
561 	td.td_resume = &td_main->td_pfd->pd_resume;
562 
563 	SET_BOOKMARK(&czb, td.td_objset,
564 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
565 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
566 
567 	mutex_enter(&td_main->td_pfd->pd_mtx);
568 	td_main->td_pfd->pd_exited = B_TRUE;
569 	cv_broadcast(&td_main->td_pfd->pd_cv);
570 	mutex_exit(&td_main->td_pfd->pd_mtx);
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 static int
578 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
579     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
580     blkptr_cb_t func, void *arg)
581 {
582 	traverse_data_t td;
583 	prefetch_data_t pd = { 0 };
584 	zbookmark_phys_t czb;
585 	int err;
586 
587 	ASSERT(ds == NULL || objset == ds->ds_object);
588 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
589 
590 	td.td_spa = spa;
591 	td.td_objset = objset;
592 	td.td_rootbp = rootbp;
593 	td.td_min_txg = txg_start;
594 	td.td_resume = resume;
595 	td.td_func = func;
596 	td.td_arg = arg;
597 	td.td_pfd = &pd;
598 	td.td_flags = flags;
599 	td.td_paused = B_FALSE;
600 	td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
601 
602 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
603 		VERIFY(spa_feature_enabled_txg(spa,
604 		    SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
605 	} else {
606 		td.td_hole_birth_enabled_txg = UINT64_MAX;
607 	}
608 
609 	pd.pd_flags = flags;
610 	if (resume != NULL)
611 		pd.pd_resume = *resume;
612 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
613 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
614 
615 	SET_BOOKMARK(&czb, td.td_objset,
616 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
617 
618 	/* See comment on ZIL traversal in dsl_scan_visitds. */
619 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
620 		enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
621 		arc_flags_t flags = ARC_FLAG_WAIT;
622 		objset_phys_t *osp;
623 		arc_buf_t *buf;
624 
625 		if ((td.td_flags & TRAVERSE_NO_DECRYPT) &&
626 		    BP_IS_PROTECTED(rootbp))
627 			zio_flags |= ZIO_FLAG_RAW;
628 
629 		err = arc_read(NULL, td.td_spa, rootbp, arc_getbuf_func,
630 		    &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, &czb);
631 		if (err != 0)
632 			return (err);
633 
634 		osp = buf->b_data;
635 		traverse_zil(&td, &osp->os_zil_header);
636 		arc_buf_destroy(buf, &buf);
637 	}
638 
639 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
640 	    taskq_dispatch(system_taskq, traverse_prefetch_thread,
641 	    &td, TQ_NOQUEUE) == TASKQID_INVALID)
642 		pd.pd_exited = B_TRUE;
643 
644 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
645 
646 	mutex_enter(&pd.pd_mtx);
647 	pd.pd_cancel = B_TRUE;
648 	cv_broadcast(&pd.pd_cv);
649 	while (!pd.pd_exited)
650 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
651 	mutex_exit(&pd.pd_mtx);
652 
653 	mutex_destroy(&pd.pd_mtx);
654 	cv_destroy(&pd.pd_cv);
655 
656 	return (err);
657 }
658 
659 /*
660  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
661  * in syncing context).
662  */
663 int
664 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
665     zbookmark_phys_t *resume,
666     int flags, blkptr_cb_t func, void *arg)
667 {
668 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
669 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
670 }
671 
672 int
673 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
674     int flags, blkptr_cb_t func, void *arg)
675 {
676 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
677 }
678 
679 int
680 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
681     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
682     blkptr_cb_t func, void *arg)
683 {
684 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
685 	    blkptr, txg_start, resume, flags, func, arg));
686 }
687 
688 /*
689  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
690  */
691 int
692 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
693     blkptr_cb_t func, void *arg)
694 {
695 	int err;
696 	dsl_pool_t *dp = spa_get_dsl(spa);
697 	objset_t *mos = dp->dp_meta_objset;
698 	boolean_t hard = (flags & TRAVERSE_HARD);
699 
700 	/* visit the MOS */
701 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
702 	    txg_start, NULL, flags, func, arg);
703 	if (err != 0)
704 		return (err);
705 
706 	/* visit each dataset */
707 	for (uint64_t obj = 1; err == 0;
708 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
709 		dmu_object_info_t doi;
710 
711 		err = dmu_object_info(mos, obj, &doi);
712 		if (err != 0) {
713 			if (hard)
714 				continue;
715 			break;
716 		}
717 
718 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
719 			dsl_dataset_t *ds;
720 			uint64_t txg = txg_start;
721 
722 			dsl_pool_config_enter(dp, FTAG);
723 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
724 			dsl_pool_config_exit(dp, FTAG);
725 			if (err != 0) {
726 				if (hard)
727 					continue;
728 				break;
729 			}
730 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
731 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
732 			err = traverse_dataset(ds, txg, flags, func, arg);
733 			dsl_dataset_rele(ds, FTAG);
734 			if (err != 0)
735 				break;
736 		}
737 	}
738 	if (err == ESRCH)
739 		err = 0;
740 	return (err);
741 }
742