xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_traverse.c (revision df58e8b1506f241670be86a560fb6e8432043aee)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/zfs_context.h>
28 #include <sys/dmu_objset.h>
29 #include <sys/dmu_traverse.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dsl_pool.h>
33 #include <sys/dnode.h>
34 #include <sys/spa.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zio.h>
37 #include <sys/dmu_impl.h>
38 #include <sys/sa.h>
39 #include <sys/sa_impl.h>
40 #include <sys/callb.h>
41 #include <sys/zfeature.h>
42 
43 static int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
44 static int32_t send_holes_without_birth_time = 1;
45 static uint_t zfs_traverse_indirect_prefetch_limit = 32;
46 
47 typedef struct prefetch_data {
48 	kmutex_t pd_mtx;
49 	kcondvar_t pd_cv;
50 	int32_t pd_bytes_fetched;
51 	int pd_flags;
52 	boolean_t pd_cancel;
53 	boolean_t pd_exited;
54 	zbookmark_phys_t pd_resume;
55 } prefetch_data_t;
56 
57 typedef struct traverse_data {
58 	spa_t *td_spa;
59 	uint64_t td_objset;
60 	blkptr_t *td_rootbp;
61 	uint64_t td_min_txg;
62 	zbookmark_phys_t *td_resume;
63 	int td_flags;
64 	prefetch_data_t *td_pfd;
65 	boolean_t td_paused;
66 	uint64_t td_hole_birth_enabled_txg;
67 	blkptr_cb_t *td_func;
68 	void *td_arg;
69 	boolean_t td_realloc_possible;
70 } traverse_data_t;
71 
72 static int traverse_dnode(traverse_data_t *td, const blkptr_t *bp,
73     const dnode_phys_t *dnp, uint64_t objset, uint64_t object);
74 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
75     uint64_t objset, uint64_t object);
76 
77 static inline uint64_t
get_birth_time(traverse_data_t * td,const blkptr_t * bp)78 get_birth_time(traverse_data_t *td, const blkptr_t *bp)
79 {
80 	if (td->td_flags & TRAVERSE_LOGICAL)
81 		return (BP_GET_LOGICAL_BIRTH(bp));
82 	else
83 		return (BP_GET_BIRTH(bp));
84 }
85 
86 static int
traverse_zil_block(zilog_t * zilog,const blkptr_t * bp,void * arg,uint64_t claim_txg)87 traverse_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
88     uint64_t claim_txg)
89 {
90 	traverse_data_t *td = arg;
91 	zbookmark_phys_t zb;
92 
93 	if (BP_IS_HOLE(bp))
94 		return (0);
95 
96 	if (claim_txg == 0 &&
97 	    get_birth_time(td, bp) >= spa_min_claim_txg(td->td_spa))
98 		return (-1);
99 
100 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
101 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
102 
103 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
104 
105 	return (0);
106 }
107 
108 static int
traverse_zil_record(zilog_t * zilog,const lr_t * lrc,void * arg,uint64_t claim_txg)109 traverse_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
110     uint64_t claim_txg)
111 {
112 	traverse_data_t *td = arg;
113 
114 	if (lrc->lrc_txtype == TX_WRITE) {
115 		lr_write_t *lr = (lr_write_t *)lrc;
116 		blkptr_t *bp = &lr->lr_blkptr;
117 		zbookmark_phys_t zb;
118 
119 		if (BP_IS_HOLE(bp))
120 			return (0);
121 
122 		if (claim_txg == 0 || get_birth_time(td, bp) < claim_txg)
123 			return (0);
124 
125 		ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
126 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
127 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
128 
129 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
130 		    td->td_arg);
131 	}
132 	return (0);
133 }
134 
135 static void
traverse_zil(traverse_data_t * td,zil_header_t * zh)136 traverse_zil(traverse_data_t *td, zil_header_t *zh)
137 {
138 	uint64_t claim_txg = zh->zh_claim_txg;
139 
140 	/*
141 	 * We only want to visit blocks that have been claimed but not yet
142 	 * replayed; plus blocks that are already stable in read-only mode.
143 	 */
144 	if (claim_txg == 0 && spa_writeable(td->td_spa))
145 		return;
146 
147 	zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
148 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
149 	    claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
150 	zil_free(zilog);
151 }
152 
153 typedef enum resume_skip {
154 	RESUME_SKIP_ALL,
155 	RESUME_SKIP_NONE,
156 	RESUME_SKIP_CHILDREN
157 } resume_skip_t;
158 
159 /*
160  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
161  * the block indicated by zb does not need to be visited at all. Returns
162  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
163  * resume point. This indicates that this block should be visited but not its
164  * children (since they must have been visited in a previous traversal).
165  * Otherwise returns RESUME_SKIP_NONE.
166  */
167 static resume_skip_t
resume_skip_check(const traverse_data_t * td,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)168 resume_skip_check(const traverse_data_t *td, const dnode_phys_t *dnp,
169     const zbookmark_phys_t *zb)
170 {
171 	if (td->td_resume != NULL) {
172 		/*
173 		 * If we already visited this bp & everything below,
174 		 * don't bother doing it again.
175 		 */
176 		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
177 			return (RESUME_SKIP_ALL);
178 
179 		if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
180 			if (td->td_flags & TRAVERSE_POST)
181 				return (RESUME_SKIP_CHILDREN);
182 		}
183 	}
184 	return (RESUME_SKIP_NONE);
185 }
186 
187 /*
188  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
189  */
190 static boolean_t
traverse_prefetch_metadata(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)191 traverse_prefetch_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
192     const blkptr_t *bp, const zbookmark_phys_t *zb)
193 {
194 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
195 	    ARC_FLAG_PRESCIENT_PREFETCH;
196 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
197 
198 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
199 		return (B_FALSE);
200 	/*
201 	 * If this bp is before the resume point, it may have already been
202 	 * freed.
203 	 */
204 	if (resume_skip_check(td, dnp, zb) != RESUME_SKIP_NONE)
205 		return (B_FALSE);
206 	if (BP_IS_HOLE(bp) || get_birth_time(td, bp) <= td->td_min_txg)
207 		return (B_FALSE);
208 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
209 		return (B_FALSE);
210 	ASSERT(!BP_IS_REDACTED(bp));
211 
212 	if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
213 		zio_flags |= ZIO_FLAG_RAW;
214 
215 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
216 	    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
217 	return (B_TRUE);
218 }
219 
220 static boolean_t
prefetch_needed(prefetch_data_t * pfd,const blkptr_t * bp)221 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
222 {
223 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
224 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
225 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
226 		return (B_FALSE);
227 	return (B_TRUE);
228 }
229 
230 static int
traverse_visitbp(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)231 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
232     const blkptr_t *bp, const zbookmark_phys_t *zb)
233 {
234 	int err = 0;
235 	arc_buf_t *buf = NULL;
236 	prefetch_data_t *pd = td->td_pfd;
237 
238 	switch (resume_skip_check(td, dnp, zb)) {
239 	case RESUME_SKIP_ALL:
240 		return (0);
241 	case RESUME_SKIP_CHILDREN:
242 		goto post;
243 	case RESUME_SKIP_NONE:
244 		break;
245 	default:
246 		ASSERT(0);
247 	}
248 
249 	if (BP_GET_LOGICAL_BIRTH(bp) == 0) {
250 		/*
251 		 * Since this block has a birth time of 0 it must be one of
252 		 * two things: a hole created before the
253 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
254 		 * which has always been a hole in an object.
255 		 *
256 		 * If a file is written sparsely, then the unwritten parts of
257 		 * the file were "always holes" -- that is, they have been
258 		 * holes since this object was allocated.  However, we (and
259 		 * our callers) can not necessarily tell when an object was
260 		 * allocated.  Therefore, if it's possible that this object
261 		 * was freed and then its object number reused, we need to
262 		 * visit all the holes with birth==0.
263 		 *
264 		 * If it isn't possible that the object number was reused,
265 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
266 		 * all the blocks we will visit as part of this traversal,
267 		 * then this hole must have always existed, so we can skip
268 		 * it.  We visit blocks born after (exclusive) td_min_txg.
269 		 *
270 		 * Note that the meta-dnode cannot be reallocated.
271 		 */
272 		if (!send_holes_without_birth_time &&
273 		    (!td->td_realloc_possible ||
274 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
275 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
276 			return (0);
277 	} else if (get_birth_time(td, bp) <= td->td_min_txg) {
278 		return (0);
279 	}
280 
281 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
282 		uint64_t size = BP_GET_LSIZE(bp);
283 		mutex_enter(&pd->pd_mtx);
284 		ASSERT(pd->pd_bytes_fetched >= 0);
285 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
286 			cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
287 		pd->pd_bytes_fetched -= size;
288 		cv_broadcast(&pd->pd_cv);
289 		mutex_exit(&pd->pd_mtx);
290 	}
291 
292 	if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
293 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
294 		if (err != 0)
295 			goto post;
296 		return (0);
297 	}
298 
299 	if (td->td_flags & TRAVERSE_PRE) {
300 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
301 		    td->td_arg);
302 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
303 			return (0);
304 		if (err != 0)
305 			goto post;
306 	}
307 
308 	if (BP_GET_LEVEL(bp) > 0) {
309 		arc_flags_t flags = ARC_FLAG_WAIT;
310 		int32_t i, ptidx, pidx;
311 		uint32_t prefetchlimit;
312 		int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
313 		zbookmark_phys_t *czb;
314 
315 		ASSERT(!BP_IS_PROTECTED(bp));
316 
317 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
318 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
319 		if (err != 0)
320 			goto post;
321 
322 		czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
323 
324 		/*
325 		 * When performing a traversal it is beneficial to
326 		 * asynchronously read-ahead the upcoming indirect
327 		 * blocks since they will be needed shortly. However,
328 		 * since a 128k indirect (non-L0) block may contain up
329 		 * to 1024 128-byte block pointers, its preferable to not
330 		 * prefetch them all at once. Issuing a large number of
331 		 * async reads may effect performance, and the earlier
332 		 * the indirect blocks are prefetched the less likely
333 		 * they are to still be resident in the ARC when needed.
334 		 * Therefore, prefetching indirect blocks is limited to
335 		 * zfs_traverse_indirect_prefetch_limit=32 blocks by
336 		 * default.
337 		 *
338 		 * pidx: Index for which next prefetch to be issued.
339 		 * ptidx: Index at which next prefetch to be triggered.
340 		 */
341 		ptidx = 0;
342 		pidx = 1;
343 		prefetchlimit = zfs_traverse_indirect_prefetch_limit;
344 		for (i = 0; i < epb; i++) {
345 			if (prefetchlimit && i == ptidx) {
346 				ASSERT3S(ptidx, <=, pidx);
347 				for (uint32_t  prefetched = 0; pidx < epb &&
348 				    prefetched < prefetchlimit; pidx++) {
349 					SET_BOOKMARK(czb, zb->zb_objset,
350 					    zb->zb_object, zb->zb_level - 1,
351 					    zb->zb_blkid * epb + pidx);
352 					if (traverse_prefetch_metadata(td, dnp,
353 					    &((blkptr_t *)buf->b_data)[pidx],
354 					    czb) == B_TRUE) {
355 						prefetched++;
356 						if (prefetched ==
357 						    MAX(prefetchlimit / 2, 1))
358 							ptidx = pidx;
359 					}
360 				}
361 			}
362 
363 			/* recursively visitbp() blocks below this */
364 			SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
365 			    zb->zb_level - 1,
366 			    zb->zb_blkid * epb + i);
367 			err = traverse_visitbp(td, dnp,
368 			    &((blkptr_t *)buf->b_data)[i], czb);
369 			if (err != 0)
370 				break;
371 		}
372 
373 		kmem_free(czb, sizeof (zbookmark_phys_t));
374 
375 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
376 		arc_flags_t flags = ARC_FLAG_WAIT;
377 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
378 		int32_t i;
379 		int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
380 		dnode_phys_t *child_dnp;
381 
382 		/*
383 		 * dnode blocks might have their bonus buffers encrypted, so
384 		 * we must be careful to honor TRAVERSE_NO_DECRYPT
385 		 */
386 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
387 			zio_flags |= ZIO_FLAG_RAW;
388 
389 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
390 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
391 		if (err != 0)
392 			goto post;
393 
394 		child_dnp = buf->b_data;
395 
396 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
397 			prefetch_dnode_metadata(td, &child_dnp[i],
398 			    zb->zb_objset, zb->zb_blkid * epb + i);
399 		}
400 
401 		/* recursively visitbp() blocks below this */
402 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
403 			err = traverse_dnode(td, bp, &child_dnp[i],
404 			    zb->zb_objset, zb->zb_blkid * epb + i);
405 			if (err != 0)
406 				break;
407 		}
408 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
409 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
410 		arc_flags_t flags = ARC_FLAG_WAIT;
411 		objset_phys_t *osp;
412 
413 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
414 			zio_flags |= ZIO_FLAG_RAW;
415 
416 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
417 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
418 		if (err != 0)
419 			goto post;
420 
421 		osp = buf->b_data;
422 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
423 		    DMU_META_DNODE_OBJECT);
424 		/*
425 		 * See the block comment above for the goal of this variable.
426 		 * If the maxblkid of the meta-dnode is 0, then we know that
427 		 * we've never had more than DNODES_PER_BLOCK objects in the
428 		 * dataset, which means we can't have reused any object ids.
429 		 */
430 		if (osp->os_meta_dnode.dn_maxblkid == 0)
431 			td->td_realloc_possible = B_FALSE;
432 
433 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
434 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
435 				prefetch_dnode_metadata(td,
436 				    &osp->os_projectused_dnode,
437 				    zb->zb_objset, DMU_PROJECTUSED_OBJECT);
438 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
439 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
440 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
441 			    zb->zb_objset, DMU_USERUSED_OBJECT);
442 		}
443 
444 		err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
445 		    DMU_META_DNODE_OBJECT);
446 		if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
447 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
448 				err = traverse_dnode(td, bp,
449 				    &osp->os_projectused_dnode, zb->zb_objset,
450 				    DMU_PROJECTUSED_OBJECT);
451 			if (err == 0)
452 				err = traverse_dnode(td, bp,
453 				    &osp->os_groupused_dnode, zb->zb_objset,
454 				    DMU_GROUPUSED_OBJECT);
455 			if (err == 0)
456 				err = traverse_dnode(td, bp,
457 				    &osp->os_userused_dnode, zb->zb_objset,
458 				    DMU_USERUSED_OBJECT);
459 		}
460 	}
461 
462 	if (buf)
463 		arc_buf_destroy(buf, &buf);
464 
465 post:
466 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
467 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
468 
469 	if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
470 		/*
471 		 * Ignore this disk error as requested by the HARD flag,
472 		 * and continue traversal.
473 		 */
474 		err = 0;
475 	}
476 
477 	/*
478 	 * If we are stopping here, set td_resume.
479 	 */
480 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
481 		td->td_resume->zb_objset = zb->zb_objset;
482 		td->td_resume->zb_object = zb->zb_object;
483 		td->td_resume->zb_level = 0;
484 		/*
485 		 * If we have stopped on an indirect block (e.g. due to
486 		 * i/o error), we have not visited anything below it.
487 		 * Set the bookmark to the first level-0 block that we need
488 		 * to visit.  This way, the resuming code does not need to
489 		 * deal with resuming from indirect blocks.
490 		 *
491 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
492 		 * to dereference it.
493 		 */
494 		td->td_resume->zb_blkid = zb->zb_blkid;
495 		if (zb->zb_level > 0) {
496 			td->td_resume->zb_blkid <<= zb->zb_level *
497 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
498 		}
499 		td->td_paused = B_TRUE;
500 	}
501 
502 	return (err);
503 }
504 
505 static void
prefetch_dnode_metadata(traverse_data_t * td,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)506 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
507     uint64_t objset, uint64_t object)
508 {
509 	int j;
510 	zbookmark_phys_t czb;
511 
512 	for (j = 0; j < dnp->dn_nblkptr; j++) {
513 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
514 		traverse_prefetch_metadata(td, dnp, &dnp->dn_blkptr[j], &czb);
515 	}
516 
517 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
518 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
519 		traverse_prefetch_metadata(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
520 	}
521 }
522 
523 static int
traverse_dnode(traverse_data_t * td,const blkptr_t * bp,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)524 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
525     uint64_t objset, uint64_t object)
526 {
527 	int j, err = 0;
528 	zbookmark_phys_t czb;
529 
530 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
531 	    object < td->td_resume->zb_object)
532 		return (0);
533 
534 	if (td->td_flags & TRAVERSE_PRE) {
535 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
536 		    ZB_DNODE_BLKID);
537 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
538 		    td->td_arg);
539 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
540 			return (0);
541 		if (err != 0)
542 			return (err);
543 	}
544 
545 	for (j = 0; j < dnp->dn_nblkptr; j++) {
546 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
547 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
548 		if (err != 0)
549 			break;
550 	}
551 
552 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
553 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
554 		err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
555 	}
556 
557 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
558 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
559 		    ZB_DNODE_BLKID);
560 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
561 		    td->td_arg);
562 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
563 			return (0);
564 		if (err != 0)
565 			return (err);
566 	}
567 	return (err);
568 }
569 
570 static int
traverse_prefetcher(spa_t * spa,zilog_t * zilog,const blkptr_t * bp,const zbookmark_phys_t * zb,const dnode_phys_t * dnp,void * arg)571 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
572     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
573 {
574 	(void) zilog, (void) dnp;
575 	prefetch_data_t *pfd = arg;
576 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
577 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
578 	    ARC_FLAG_PRESCIENT_PREFETCH;
579 
580 	ASSERT(pfd->pd_bytes_fetched >= 0);
581 	if (zb->zb_level == ZB_DNODE_LEVEL)
582 		return (0);
583 	if (pfd->pd_cancel)
584 		return (SET_ERROR(EINTR));
585 
586 	if (!prefetch_needed(pfd, bp))
587 		return (0);
588 
589 	mutex_enter(&pfd->pd_mtx);
590 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
591 		cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
592 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
593 	cv_broadcast(&pfd->pd_cv);
594 	mutex_exit(&pfd->pd_mtx);
595 
596 	if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
597 		zio_flags |= ZIO_FLAG_RAW;
598 
599 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
600 	    zio_flags, &aflags, zb);
601 
602 	return (0);
603 }
604 
605 static void
traverse_prefetch_thread(void * arg)606 traverse_prefetch_thread(void *arg)
607 {
608 	traverse_data_t *td_main = arg;
609 	traverse_data_t td = *td_main;
610 	zbookmark_phys_t czb;
611 	fstrans_cookie_t cookie = spl_fstrans_mark();
612 
613 	td.td_func = traverse_prefetcher;
614 	td.td_arg = td_main->td_pfd;
615 	td.td_pfd = NULL;
616 	td.td_resume = &td_main->td_pfd->pd_resume;
617 
618 	SET_BOOKMARK(&czb, td.td_objset,
619 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
620 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
621 
622 	mutex_enter(&td_main->td_pfd->pd_mtx);
623 	td_main->td_pfd->pd_exited = B_TRUE;
624 	cv_broadcast(&td_main->td_pfd->pd_cv);
625 	mutex_exit(&td_main->td_pfd->pd_mtx);
626 	spl_fstrans_unmark(cookie);
627 }
628 
629 /*
630  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
631  * in syncing context).
632  */
633 static int
traverse_impl(spa_t * spa,dsl_dataset_t * ds,uint64_t objset,blkptr_t * rootbp,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)634 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
635     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
636     blkptr_cb_t func, void *arg)
637 {
638 	traverse_data_t *td;
639 	prefetch_data_t *pd;
640 	zbookmark_phys_t *czb;
641 	int err;
642 
643 	ASSERT(ds == NULL || objset == ds->ds_object);
644 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
645 
646 	td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
647 	pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
648 	czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
649 
650 	td->td_spa = spa;
651 	td->td_objset = objset;
652 	td->td_rootbp = rootbp;
653 	td->td_min_txg = txg_start;
654 	td->td_resume = resume;
655 	td->td_func = func;
656 	td->td_arg = arg;
657 	td->td_pfd = pd;
658 	td->td_flags = flags;
659 	td->td_paused = B_FALSE;
660 	td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
661 
662 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
663 		VERIFY(spa_feature_enabled_txg(spa,
664 		    SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
665 	} else {
666 		td->td_hole_birth_enabled_txg = UINT64_MAX;
667 	}
668 
669 	pd->pd_flags = flags;
670 	if (resume != NULL)
671 		pd->pd_resume = *resume;
672 	mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
673 	cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
674 
675 	SET_BOOKMARK(czb, td->td_objset,
676 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
677 
678 	/* See comment on ZIL traversal in dsl_scan_visitds. */
679 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
680 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
681 		arc_flags_t flags = ARC_FLAG_WAIT;
682 		objset_phys_t *osp;
683 		arc_buf_t *buf;
684 		ASSERT(!BP_IS_REDACTED(rootbp));
685 
686 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
687 		    BP_IS_PROTECTED(rootbp))
688 			zio_flags |= ZIO_FLAG_RAW;
689 
690 		err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
691 		    &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
692 		if (err != 0) {
693 			/*
694 			 * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
695 			 * continue to visitbp so that td_func can be called
696 			 * in pre stage, and err will reset to zero.
697 			 */
698 			if (!(td->td_flags & TRAVERSE_HARD) ||
699 			    !(td->td_flags & TRAVERSE_PRE))
700 				goto out;
701 		} else {
702 			osp = buf->b_data;
703 			traverse_zil(td, &osp->os_zil_header);
704 			arc_buf_destroy(buf, &buf);
705 		}
706 	}
707 
708 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
709 	    taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
710 	    td, TQ_NOQUEUE) == TASKQID_INVALID)
711 		pd->pd_exited = B_TRUE;
712 
713 	err = traverse_visitbp(td, NULL, rootbp, czb);
714 
715 	mutex_enter(&pd->pd_mtx);
716 	pd->pd_cancel = B_TRUE;
717 	cv_broadcast(&pd->pd_cv);
718 	while (!pd->pd_exited)
719 		cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
720 	mutex_exit(&pd->pd_mtx);
721 out:
722 	mutex_destroy(&pd->pd_mtx);
723 	cv_destroy(&pd->pd_cv);
724 
725 	kmem_free(czb, sizeof (zbookmark_phys_t));
726 	kmem_free(pd, sizeof (struct prefetch_data));
727 	kmem_free(td, sizeof (struct traverse_data));
728 
729 	return (err);
730 }
731 
732 /*
733  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
734  * in syncing context).
735  */
736 int
traverse_dataset_resume(dsl_dataset_t * ds,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)737 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
738     zbookmark_phys_t *resume,
739     int flags, blkptr_cb_t func, void *arg)
740 {
741 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
742 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
743 }
744 
745 int
traverse_dataset(dsl_dataset_t * ds,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)746 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
747     int flags, blkptr_cb_t func, void *arg)
748 {
749 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
750 }
751 
752 int
traverse_dataset_destroyed(spa_t * spa,blkptr_t * blkptr,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)753 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
754     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
755     blkptr_cb_t func, void *arg)
756 {
757 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
758 	    blkptr, txg_start, resume, flags, func, arg));
759 }
760 
761 /*
762  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
763  */
764 int
traverse_pool(spa_t * spa,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)765 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
766     blkptr_cb_t func, void *arg)
767 {
768 	int err;
769 	dsl_pool_t *dp = spa_get_dsl(spa);
770 	objset_t *mos = dp->dp_meta_objset;
771 	boolean_t hard = (flags & TRAVERSE_HARD);
772 
773 	/* visit the MOS */
774 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
775 	    txg_start, NULL, flags, func, arg);
776 	if (err != 0)
777 		return (err);
778 
779 	/* visit each dataset */
780 	for (uint64_t obj = 1; err == 0;
781 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
782 		dmu_object_info_t doi;
783 
784 		err = dmu_object_info(mos, obj, &doi);
785 		if (err != 0) {
786 			if (hard)
787 				continue;
788 			break;
789 		}
790 
791 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
792 			dsl_dataset_t *ds;
793 			uint64_t txg = txg_start;
794 
795 			dsl_pool_config_enter(dp, FTAG);
796 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
797 			dsl_pool_config_exit(dp, FTAG);
798 			if (err != 0) {
799 				if (hard)
800 					continue;
801 				break;
802 			}
803 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
804 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
805 			err = traverse_dataset(ds, txg, flags, func, arg);
806 			dsl_dataset_rele(ds, FTAG);
807 			if (err != 0)
808 				break;
809 		}
810 	}
811 	if (err == ESRCH)
812 		err = 0;
813 	return (err);
814 }
815 
816 EXPORT_SYMBOL(traverse_dataset);
817 EXPORT_SYMBOL(traverse_pool);
818 
819 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
820 	"Max number of bytes to prefetch");
821 
822 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, UINT, ZMOD_RW,
823 	"Traverse prefetch number of blocks pointed by indirect block");
824 
825 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
826 	"Ignore hole_birth txg for zfs send");
827