xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_traverse.c (revision 097e192a9c203d98c0b097d1517f87342896f676)
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 https://opensource.org/licenses/CDDL-1.0.
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 static int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
43 static int32_t send_holes_without_birth_time = 1;
44 static uint_t zfs_traverse_indirect_prefetch_limit = 32;
45 
46 typedef struct prefetch_data {
47 	kmutex_t pd_mtx;
48 	kcondvar_t pd_cv;
49 	int32_t pd_bytes_fetched;
50 	int pd_flags;
51 	boolean_t pd_cancel;
52 	boolean_t pd_exited;
53 	zbookmark_phys_t pd_resume;
54 } prefetch_data_t;
55 
56 typedef struct traverse_data {
57 	spa_t *td_spa;
58 	uint64_t td_objset;
59 	blkptr_t *td_rootbp;
60 	uint64_t td_min_txg;
61 	zbookmark_phys_t *td_resume;
62 	int td_flags;
63 	prefetch_data_t *td_pfd;
64 	boolean_t td_paused;
65 	uint64_t td_hole_birth_enabled_txg;
66 	blkptr_cb_t *td_func;
67 	void *td_arg;
68 	boolean_t td_realloc_possible;
69 } traverse_data_t;
70 
71 static int traverse_dnode(traverse_data_t *td, const blkptr_t *bp,
72     const dnode_phys_t *dnp, uint64_t objset, uint64_t object);
73 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
74     uint64_t objset, uint64_t object);
75 
76 static int
77 traverse_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
78     uint64_t claim_txg)
79 {
80 	traverse_data_t *td = arg;
81 	zbookmark_phys_t zb;
82 
83 	if (BP_IS_HOLE(bp))
84 		return (0);
85 
86 	if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa))
87 		return (-1);
88 
89 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
90 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
91 
92 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
93 
94 	return (0);
95 }
96 
97 static int
98 traverse_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
99     uint64_t claim_txg)
100 {
101 	traverse_data_t *td = arg;
102 
103 	if (lrc->lrc_txtype == TX_WRITE) {
104 		lr_write_t *lr = (lr_write_t *)lrc;
105 		blkptr_t *bp = &lr->lr_blkptr;
106 		zbookmark_phys_t zb;
107 
108 		if (BP_IS_HOLE(bp))
109 			return (0);
110 
111 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
112 			return (0);
113 
114 		ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
115 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
116 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
117 
118 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
119 		    td->td_arg);
120 	}
121 	return (0);
122 }
123 
124 static void
125 traverse_zil(traverse_data_t *td, zil_header_t *zh)
126 {
127 	uint64_t claim_txg = zh->zh_claim_txg;
128 
129 	/*
130 	 * We only want to visit blocks that have been claimed but not yet
131 	 * replayed; plus blocks that are already stable in read-only mode.
132 	 */
133 	if (claim_txg == 0 && spa_writeable(td->td_spa))
134 		return;
135 
136 	zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
137 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
138 	    claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
139 	zil_free(zilog);
140 }
141 
142 typedef enum resume_skip {
143 	RESUME_SKIP_ALL,
144 	RESUME_SKIP_NONE,
145 	RESUME_SKIP_CHILDREN
146 } resume_skip_t;
147 
148 /*
149  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
150  * the block indicated by zb does not need to be visited at all. Returns
151  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
152  * resume point. This indicates that this block should be visited but not its
153  * children (since they must have been visited in a previous traversal).
154  * Otherwise returns RESUME_SKIP_NONE.
155  */
156 static resume_skip_t
157 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
158     const zbookmark_phys_t *zb)
159 {
160 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
161 		/*
162 		 * If we already visited this bp & everything below,
163 		 * don't bother doing it again.
164 		 */
165 		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
166 			return (RESUME_SKIP_ALL);
167 
168 		/*
169 		 * If we found the block we're trying to resume from, zero
170 		 * the bookmark out to indicate that we have resumed.
171 		 */
172 		if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
173 			memset(td->td_resume, 0, sizeof (*zb));
174 			if (td->td_flags & TRAVERSE_POST)
175 				return (RESUME_SKIP_CHILDREN);
176 		}
177 	}
178 	return (RESUME_SKIP_NONE);
179 }
180 
181 /*
182  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
183  */
184 static boolean_t
185 traverse_prefetch_metadata(traverse_data_t *td,
186     const blkptr_t *bp, const zbookmark_phys_t *zb)
187 {
188 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
189 	    ARC_FLAG_PRESCIENT_PREFETCH;
190 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
191 
192 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
193 		return (B_FALSE);
194 	/*
195 	 * If we are in the process of resuming, don't prefetch, because
196 	 * some children will not be needed (and in fact may have already
197 	 * been freed).
198 	 */
199 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
200 		return (B_FALSE);
201 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
202 		return (B_FALSE);
203 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
204 		return (B_FALSE);
205 	ASSERT(!BP_IS_REDACTED(bp));
206 
207 	if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
208 		zio_flags |= ZIO_FLAG_RAW;
209 
210 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
211 	    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
212 	return (B_TRUE);
213 }
214 
215 static boolean_t
216 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
217 {
218 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
219 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
220 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
221 		return (B_FALSE);
222 	return (B_TRUE);
223 }
224 
225 static int
226 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
227     const blkptr_t *bp, const zbookmark_phys_t *zb)
228 {
229 	int err = 0;
230 	arc_buf_t *buf = NULL;
231 	prefetch_data_t *pd = td->td_pfd;
232 
233 	switch (resume_skip_check(td, dnp, zb)) {
234 	case RESUME_SKIP_ALL:
235 		return (0);
236 	case RESUME_SKIP_CHILDREN:
237 		goto post;
238 	case RESUME_SKIP_NONE:
239 		break;
240 	default:
241 		ASSERT(0);
242 	}
243 
244 	if (bp->blk_birth == 0) {
245 		/*
246 		 * Since this block has a birth time of 0 it must be one of
247 		 * two things: a hole created before the
248 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
249 		 * which has always been a hole in an object.
250 		 *
251 		 * If a file is written sparsely, then the unwritten parts of
252 		 * the file were "always holes" -- that is, they have been
253 		 * holes since this object was allocated.  However, we (and
254 		 * our callers) can not necessarily tell when an object was
255 		 * allocated.  Therefore, if it's possible that this object
256 		 * was freed and then its object number reused, we need to
257 		 * visit all the holes with birth==0.
258 		 *
259 		 * If it isn't possible that the object number was reused,
260 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
261 		 * all the blocks we will visit as part of this traversal,
262 		 * then this hole must have always existed, so we can skip
263 		 * it.  We visit blocks born after (exclusive) td_min_txg.
264 		 *
265 		 * Note that the meta-dnode cannot be reallocated.
266 		 */
267 		if (!send_holes_without_birth_time &&
268 		    (!td->td_realloc_possible ||
269 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
270 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
271 			return (0);
272 	} else if (bp->blk_birth <= td->td_min_txg) {
273 		return (0);
274 	}
275 
276 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
277 		uint64_t size = BP_GET_LSIZE(bp);
278 		mutex_enter(&pd->pd_mtx);
279 		ASSERT(pd->pd_bytes_fetched >= 0);
280 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
281 			cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
282 		pd->pd_bytes_fetched -= size;
283 		cv_broadcast(&pd->pd_cv);
284 		mutex_exit(&pd->pd_mtx);
285 	}
286 
287 	if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
288 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
289 		if (err != 0)
290 			goto post;
291 		return (0);
292 	}
293 
294 	if (td->td_flags & TRAVERSE_PRE) {
295 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
296 		    td->td_arg);
297 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
298 			return (0);
299 		if (err != 0)
300 			goto post;
301 	}
302 
303 	if (BP_GET_LEVEL(bp) > 0) {
304 		uint32_t flags = ARC_FLAG_WAIT;
305 		int32_t i, ptidx, pidx;
306 		uint32_t prefetchlimit;
307 		int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
308 		zbookmark_phys_t *czb;
309 
310 		ASSERT(!BP_IS_PROTECTED(bp));
311 
312 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
313 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
314 		if (err != 0)
315 			goto post;
316 
317 		czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
318 
319 		/*
320 		 * When performing a traversal it is beneficial to
321 		 * asynchronously read-ahead the upcoming indirect
322 		 * blocks since they will be needed shortly. However,
323 		 * since a 128k indirect (non-L0) block may contain up
324 		 * to 1024 128-byte block pointers, its preferable to not
325 		 * prefetch them all at once. Issuing a large number of
326 		 * async reads may effect performance, and the earlier
327 		 * the indirect blocks are prefetched the less likely
328 		 * they are to still be resident in the ARC when needed.
329 		 * Therefore, prefetching indirect blocks is limited to
330 		 * zfs_traverse_indirect_prefetch_limit=32 blocks by
331 		 * default.
332 		 *
333 		 * pidx: Index for which next prefetch to be issued.
334 		 * ptidx: Index at which next prefetch to be triggered.
335 		 */
336 		ptidx = 0;
337 		pidx = 1;
338 		prefetchlimit = zfs_traverse_indirect_prefetch_limit;
339 		for (i = 0; i < epb; i++) {
340 			if (prefetchlimit && i == ptidx) {
341 				ASSERT3S(ptidx, <=, pidx);
342 				for (uint32_t  prefetched = 0; pidx < epb &&
343 				    prefetched < prefetchlimit; pidx++) {
344 					SET_BOOKMARK(czb, zb->zb_objset,
345 					    zb->zb_object, zb->zb_level - 1,
346 					    zb->zb_blkid * epb + pidx);
347 					if (traverse_prefetch_metadata(td,
348 					    &((blkptr_t *)buf->b_data)[pidx],
349 					    czb) == B_TRUE) {
350 						prefetched++;
351 						if (prefetched ==
352 						    MAX(prefetchlimit / 2, 1))
353 							ptidx = pidx;
354 					}
355 				}
356 			}
357 
358 			/* recursively visitbp() blocks below this */
359 			SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
360 			    zb->zb_level - 1,
361 			    zb->zb_blkid * epb + i);
362 			err = traverse_visitbp(td, dnp,
363 			    &((blkptr_t *)buf->b_data)[i], czb);
364 			if (err != 0)
365 				break;
366 		}
367 
368 		kmem_free(czb, sizeof (zbookmark_phys_t));
369 
370 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
371 		uint32_t flags = ARC_FLAG_WAIT;
372 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
373 		int32_t i;
374 		int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
375 		dnode_phys_t *child_dnp;
376 
377 		/*
378 		 * dnode blocks might have their bonus buffers encrypted, so
379 		 * we must be careful to honor TRAVERSE_NO_DECRYPT
380 		 */
381 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
382 			zio_flags |= ZIO_FLAG_RAW;
383 
384 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
385 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
386 		if (err != 0)
387 			goto post;
388 
389 		child_dnp = buf->b_data;
390 
391 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
392 			prefetch_dnode_metadata(td, &child_dnp[i],
393 			    zb->zb_objset, zb->zb_blkid * epb + i);
394 		}
395 
396 		/* recursively visitbp() blocks below this */
397 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
398 			err = traverse_dnode(td, bp, &child_dnp[i],
399 			    zb->zb_objset, zb->zb_blkid * epb + i);
400 			if (err != 0)
401 				break;
402 		}
403 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
404 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
405 		arc_flags_t flags = ARC_FLAG_WAIT;
406 		objset_phys_t *osp;
407 
408 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
409 			zio_flags |= ZIO_FLAG_RAW;
410 
411 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
412 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
413 		if (err != 0)
414 			goto post;
415 
416 		osp = buf->b_data;
417 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
418 		    DMU_META_DNODE_OBJECT);
419 		/*
420 		 * See the block comment above for the goal of this variable.
421 		 * If the maxblkid of the meta-dnode is 0, then we know that
422 		 * we've never had more than DNODES_PER_BLOCK objects in the
423 		 * dataset, which means we can't have reused any object ids.
424 		 */
425 		if (osp->os_meta_dnode.dn_maxblkid == 0)
426 			td->td_realloc_possible = B_FALSE;
427 
428 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
429 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
430 				prefetch_dnode_metadata(td,
431 				    &osp->os_projectused_dnode,
432 				    zb->zb_objset, DMU_PROJECTUSED_OBJECT);
433 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
434 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
435 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
436 			    zb->zb_objset, DMU_USERUSED_OBJECT);
437 		}
438 
439 		err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
440 		    DMU_META_DNODE_OBJECT);
441 		if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
442 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
443 				err = traverse_dnode(td, bp,
444 				    &osp->os_projectused_dnode, zb->zb_objset,
445 				    DMU_PROJECTUSED_OBJECT);
446 			if (err == 0)
447 				err = traverse_dnode(td, bp,
448 				    &osp->os_groupused_dnode, zb->zb_objset,
449 				    DMU_GROUPUSED_OBJECT);
450 			if (err == 0)
451 				err = traverse_dnode(td, bp,
452 				    &osp->os_userused_dnode, zb->zb_objset,
453 				    DMU_USERUSED_OBJECT);
454 		}
455 	}
456 
457 	if (buf)
458 		arc_buf_destroy(buf, &buf);
459 
460 post:
461 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
462 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
463 
464 	if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
465 		/*
466 		 * Ignore this disk error as requested by the HARD flag,
467 		 * and continue traversal.
468 		 */
469 		err = 0;
470 	}
471 
472 	/*
473 	 * If we are stopping here, set td_resume.
474 	 */
475 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
476 		td->td_resume->zb_objset = zb->zb_objset;
477 		td->td_resume->zb_object = zb->zb_object;
478 		td->td_resume->zb_level = 0;
479 		/*
480 		 * If we have stopped on an indirect block (e.g. due to
481 		 * i/o error), we have not visited anything below it.
482 		 * Set the bookmark to the first level-0 block that we need
483 		 * to visit.  This way, the resuming code does not need to
484 		 * deal with resuming from indirect blocks.
485 		 *
486 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
487 		 * to dereference it.
488 		 */
489 		td->td_resume->zb_blkid = zb->zb_blkid;
490 		if (zb->zb_level > 0) {
491 			td->td_resume->zb_blkid <<= zb->zb_level *
492 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
493 		}
494 		td->td_paused = B_TRUE;
495 	}
496 
497 	return (err);
498 }
499 
500 static void
501 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
502     uint64_t objset, uint64_t object)
503 {
504 	int j;
505 	zbookmark_phys_t czb;
506 
507 	for (j = 0; j < dnp->dn_nblkptr; j++) {
508 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
509 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
510 	}
511 
512 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
513 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
514 		traverse_prefetch_metadata(td, DN_SPILL_BLKPTR(dnp), &czb);
515 	}
516 }
517 
518 static int
519 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
520     uint64_t objset, uint64_t object)
521 {
522 	int j, err = 0;
523 	zbookmark_phys_t czb;
524 
525 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
526 	    object < td->td_resume->zb_object)
527 		return (0);
528 
529 	if (td->td_flags & TRAVERSE_PRE) {
530 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
531 		    ZB_DNODE_BLKID);
532 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
533 		    td->td_arg);
534 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
535 			return (0);
536 		if (err != 0)
537 			return (err);
538 	}
539 
540 	for (j = 0; j < dnp->dn_nblkptr; j++) {
541 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
542 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
543 		if (err != 0)
544 			break;
545 	}
546 
547 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
548 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
549 		err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
550 	}
551 
552 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
553 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
554 		    ZB_DNODE_BLKID);
555 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
556 		    td->td_arg);
557 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
558 			return (0);
559 		if (err != 0)
560 			return (err);
561 	}
562 	return (err);
563 }
564 
565 static int
566 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
567     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
568 {
569 	(void) zilog, (void) dnp;
570 	prefetch_data_t *pfd = arg;
571 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
572 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
573 	    ARC_FLAG_PRESCIENT_PREFETCH;
574 
575 	ASSERT(pfd->pd_bytes_fetched >= 0);
576 	if (zb->zb_level == ZB_DNODE_LEVEL)
577 		return (0);
578 	if (pfd->pd_cancel)
579 		return (SET_ERROR(EINTR));
580 
581 	if (!prefetch_needed(pfd, bp))
582 		return (0);
583 
584 	mutex_enter(&pfd->pd_mtx);
585 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
586 		cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
587 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
588 	cv_broadcast(&pfd->pd_cv);
589 	mutex_exit(&pfd->pd_mtx);
590 
591 	if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
592 		zio_flags |= ZIO_FLAG_RAW;
593 
594 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
595 	    zio_flags, &aflags, zb);
596 
597 	return (0);
598 }
599 
600 static void
601 traverse_prefetch_thread(void *arg)
602 {
603 	traverse_data_t *td_main = arg;
604 	traverse_data_t td = *td_main;
605 	zbookmark_phys_t czb;
606 	fstrans_cookie_t cookie = spl_fstrans_mark();
607 
608 	td.td_func = traverse_prefetcher;
609 	td.td_arg = td_main->td_pfd;
610 	td.td_pfd = NULL;
611 	td.td_resume = &td_main->td_pfd->pd_resume;
612 
613 	SET_BOOKMARK(&czb, td.td_objset,
614 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
615 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
616 
617 	mutex_enter(&td_main->td_pfd->pd_mtx);
618 	td_main->td_pfd->pd_exited = B_TRUE;
619 	cv_broadcast(&td_main->td_pfd->pd_cv);
620 	mutex_exit(&td_main->td_pfd->pd_mtx);
621 	spl_fstrans_unmark(cookie);
622 }
623 
624 /*
625  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
626  * in syncing context).
627  */
628 static int
629 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
630     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
631     blkptr_cb_t func, void *arg)
632 {
633 	traverse_data_t *td;
634 	prefetch_data_t *pd;
635 	zbookmark_phys_t *czb;
636 	int err;
637 
638 	ASSERT(ds == NULL || objset == ds->ds_object);
639 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
640 
641 	td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
642 	pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
643 	czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
644 
645 	td->td_spa = spa;
646 	td->td_objset = objset;
647 	td->td_rootbp = rootbp;
648 	td->td_min_txg = txg_start;
649 	td->td_resume = resume;
650 	td->td_func = func;
651 	td->td_arg = arg;
652 	td->td_pfd = pd;
653 	td->td_flags = flags;
654 	td->td_paused = B_FALSE;
655 	td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
656 
657 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
658 		VERIFY(spa_feature_enabled_txg(spa,
659 		    SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
660 	} else {
661 		td->td_hole_birth_enabled_txg = UINT64_MAX;
662 	}
663 
664 	pd->pd_flags = flags;
665 	if (resume != NULL)
666 		pd->pd_resume = *resume;
667 	mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
668 	cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
669 
670 	SET_BOOKMARK(czb, td->td_objset,
671 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
672 
673 	/* See comment on ZIL traversal in dsl_scan_visitds. */
674 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
675 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
676 		uint32_t flags = ARC_FLAG_WAIT;
677 		objset_phys_t *osp;
678 		arc_buf_t *buf;
679 		ASSERT(!BP_IS_REDACTED(rootbp));
680 
681 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
682 		    BP_IS_PROTECTED(rootbp))
683 			zio_flags |= ZIO_FLAG_RAW;
684 
685 		err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
686 		    &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
687 		if (err != 0) {
688 			/*
689 			 * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
690 			 * continue to visitbp so that td_func can be called
691 			 * in pre stage, and err will reset to zero.
692 			 */
693 			if (!(td->td_flags & TRAVERSE_HARD) ||
694 			    !(td->td_flags & TRAVERSE_PRE))
695 				goto out;
696 		} else {
697 			osp = buf->b_data;
698 			traverse_zil(td, &osp->os_zil_header);
699 			arc_buf_destroy(buf, &buf);
700 		}
701 	}
702 
703 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
704 	    taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
705 	    td, TQ_NOQUEUE) == TASKQID_INVALID)
706 		pd->pd_exited = B_TRUE;
707 
708 	err = traverse_visitbp(td, NULL, rootbp, czb);
709 
710 	mutex_enter(&pd->pd_mtx);
711 	pd->pd_cancel = B_TRUE;
712 	cv_broadcast(&pd->pd_cv);
713 	while (!pd->pd_exited)
714 		cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
715 	mutex_exit(&pd->pd_mtx);
716 out:
717 	mutex_destroy(&pd->pd_mtx);
718 	cv_destroy(&pd->pd_cv);
719 
720 	kmem_free(czb, sizeof (zbookmark_phys_t));
721 	kmem_free(pd, sizeof (struct prefetch_data));
722 	kmem_free(td, sizeof (struct traverse_data));
723 
724 	return (err);
725 }
726 
727 /*
728  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
729  * in syncing context).
730  */
731 int
732 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
733     zbookmark_phys_t *resume,
734     int flags, blkptr_cb_t func, void *arg)
735 {
736 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
737 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
738 }
739 
740 int
741 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
742     int flags, blkptr_cb_t func, void *arg)
743 {
744 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
745 }
746 
747 int
748 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
749     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
750     blkptr_cb_t func, void *arg)
751 {
752 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
753 	    blkptr, txg_start, resume, flags, func, arg));
754 }
755 
756 /*
757  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
758  */
759 int
760 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
761     blkptr_cb_t func, void *arg)
762 {
763 	int err;
764 	dsl_pool_t *dp = spa_get_dsl(spa);
765 	objset_t *mos = dp->dp_meta_objset;
766 	boolean_t hard = (flags & TRAVERSE_HARD);
767 
768 	/* visit the MOS */
769 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
770 	    txg_start, NULL, flags, func, arg);
771 	if (err != 0)
772 		return (err);
773 
774 	/* visit each dataset */
775 	for (uint64_t obj = 1; err == 0;
776 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
777 		dmu_object_info_t doi;
778 
779 		err = dmu_object_info(mos, obj, &doi);
780 		if (err != 0) {
781 			if (hard)
782 				continue;
783 			break;
784 		}
785 
786 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
787 			dsl_dataset_t *ds;
788 			uint64_t txg = txg_start;
789 
790 			dsl_pool_config_enter(dp, FTAG);
791 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
792 			dsl_pool_config_exit(dp, FTAG);
793 			if (err != 0) {
794 				if (hard)
795 					continue;
796 				break;
797 			}
798 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
799 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
800 			err = traverse_dataset(ds, txg, flags, func, arg);
801 			dsl_dataset_rele(ds, FTAG);
802 			if (err != 0)
803 				break;
804 		}
805 	}
806 	if (err == ESRCH)
807 		err = 0;
808 	return (err);
809 }
810 
811 EXPORT_SYMBOL(traverse_dataset);
812 EXPORT_SYMBOL(traverse_pool);
813 
814 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
815 	"Max number of bytes to prefetch");
816 
817 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, UINT, ZMOD_RW,
818 	"Traverse prefetch number of blocks pointed by indirect block");
819 
820 #if defined(_KERNEL)
821 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644);
822 MODULE_PARM_DESC(ignore_hole_birth,
823 	"Alias for send_holes_without_birth_time");
824 #endif
825 
826 /* CSTYLED */
827 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
828 	"Ignore hole_birth txg for zfs send");
829