xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_traverse.c (revision 783d3ff6d7fae619db8a7990b8a6387de0c677b5)
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
traverse_zil_block(zilog_t * zilog,const blkptr_t * bp,void * arg,uint64_t claim_txg)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 &&
87 	    BP_GET_LOGICAL_BIRTH(bp) >= spa_min_claim_txg(td->td_spa))
88 		return (-1);
89 
90 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
91 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
92 
93 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
94 
95 	return (0);
96 }
97 
98 static int
traverse_zil_record(zilog_t * zilog,const lr_t * lrc,void * arg,uint64_t claim_txg)99 traverse_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
100     uint64_t claim_txg)
101 {
102 	traverse_data_t *td = arg;
103 
104 	if (lrc->lrc_txtype == TX_WRITE) {
105 		lr_write_t *lr = (lr_write_t *)lrc;
106 		blkptr_t *bp = &lr->lr_blkptr;
107 		zbookmark_phys_t zb;
108 
109 		if (BP_IS_HOLE(bp))
110 			return (0);
111 
112 		if (claim_txg == 0 || BP_GET_LOGICAL_BIRTH(bp) < claim_txg)
113 			return (0);
114 
115 		ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
116 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
117 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
118 
119 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
120 		    td->td_arg);
121 	}
122 	return (0);
123 }
124 
125 static void
traverse_zil(traverse_data_t * td,zil_header_t * zh)126 traverse_zil(traverse_data_t *td, zil_header_t *zh)
127 {
128 	uint64_t claim_txg = zh->zh_claim_txg;
129 
130 	/*
131 	 * We only want to visit blocks that have been claimed but not yet
132 	 * replayed; plus blocks that are already stable in read-only mode.
133 	 */
134 	if (claim_txg == 0 && spa_writeable(td->td_spa))
135 		return;
136 
137 	zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
138 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
139 	    claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
140 	zil_free(zilog);
141 }
142 
143 typedef enum resume_skip {
144 	RESUME_SKIP_ALL,
145 	RESUME_SKIP_NONE,
146 	RESUME_SKIP_CHILDREN
147 } resume_skip_t;
148 
149 /*
150  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
151  * the block indicated by zb does not need to be visited at all. Returns
152  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
153  * resume point. This indicates that this block should be visited but not its
154  * children (since they must have been visited in a previous traversal).
155  * Otherwise returns RESUME_SKIP_NONE.
156  */
157 static resume_skip_t
resume_skip_check(const traverse_data_t * td,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)158 resume_skip_check(const traverse_data_t *td, const dnode_phys_t *dnp,
159     const zbookmark_phys_t *zb)
160 {
161 	if (td->td_resume != NULL) {
162 		/*
163 		 * If we already visited this bp & everything below,
164 		 * don't bother doing it again.
165 		 */
166 		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
167 			return (RESUME_SKIP_ALL);
168 
169 		if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
170 			if (td->td_flags & TRAVERSE_POST)
171 				return (RESUME_SKIP_CHILDREN);
172 		}
173 	}
174 	return (RESUME_SKIP_NONE);
175 }
176 
177 /*
178  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
179  */
180 static boolean_t
traverse_prefetch_metadata(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)181 traverse_prefetch_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
182     const blkptr_t *bp, const zbookmark_phys_t *zb)
183 {
184 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
185 	    ARC_FLAG_PRESCIENT_PREFETCH;
186 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
187 
188 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
189 		return (B_FALSE);
190 	/*
191 	 * If this bp is before the resume point, it may have already been
192 	 * freed.
193 	 */
194 	if (resume_skip_check(td, dnp, zb) != RESUME_SKIP_NONE)
195 		return (B_FALSE);
196 	if (BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) <= td->td_min_txg)
197 		return (B_FALSE);
198 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
199 		return (B_FALSE);
200 	ASSERT(!BP_IS_REDACTED(bp));
201 
202 	if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
203 		zio_flags |= ZIO_FLAG_RAW;
204 
205 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
206 	    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
207 	return (B_TRUE);
208 }
209 
210 static boolean_t
prefetch_needed(prefetch_data_t * pfd,const blkptr_t * bp)211 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
212 {
213 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
214 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
215 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
216 		return (B_FALSE);
217 	return (B_TRUE);
218 }
219 
220 static int
traverse_visitbp(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)221 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
222     const blkptr_t *bp, const zbookmark_phys_t *zb)
223 {
224 	int err = 0;
225 	arc_buf_t *buf = NULL;
226 	prefetch_data_t *pd = td->td_pfd;
227 
228 	switch (resume_skip_check(td, dnp, zb)) {
229 	case RESUME_SKIP_ALL:
230 		return (0);
231 	case RESUME_SKIP_CHILDREN:
232 		goto post;
233 	case RESUME_SKIP_NONE:
234 		break;
235 	default:
236 		ASSERT(0);
237 	}
238 
239 	if (BP_GET_LOGICAL_BIRTH(bp) == 0) {
240 		/*
241 		 * Since this block has a birth time of 0 it must be one of
242 		 * two things: a hole created before the
243 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
244 		 * which has always been a hole in an object.
245 		 *
246 		 * If a file is written sparsely, then the unwritten parts of
247 		 * the file were "always holes" -- that is, they have been
248 		 * holes since this object was allocated.  However, we (and
249 		 * our callers) can not necessarily tell when an object was
250 		 * allocated.  Therefore, if it's possible that this object
251 		 * was freed and then its object number reused, we need to
252 		 * visit all the holes with birth==0.
253 		 *
254 		 * If it isn't possible that the object number was reused,
255 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
256 		 * all the blocks we will visit as part of this traversal,
257 		 * then this hole must have always existed, so we can skip
258 		 * it.  We visit blocks born after (exclusive) td_min_txg.
259 		 *
260 		 * Note that the meta-dnode cannot be reallocated.
261 		 */
262 		if (!send_holes_without_birth_time &&
263 		    (!td->td_realloc_possible ||
264 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
265 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
266 			return (0);
267 	} else if (BP_GET_LOGICAL_BIRTH(bp) <= td->td_min_txg) {
268 		return (0);
269 	}
270 
271 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
272 		uint64_t size = BP_GET_LSIZE(bp);
273 		mutex_enter(&pd->pd_mtx);
274 		ASSERT(pd->pd_bytes_fetched >= 0);
275 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
276 			cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
277 		pd->pd_bytes_fetched -= size;
278 		cv_broadcast(&pd->pd_cv);
279 		mutex_exit(&pd->pd_mtx);
280 	}
281 
282 	if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
283 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
284 		if (err != 0)
285 			goto post;
286 		return (0);
287 	}
288 
289 	if (td->td_flags & TRAVERSE_PRE) {
290 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
291 		    td->td_arg);
292 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
293 			return (0);
294 		if (err != 0)
295 			goto post;
296 	}
297 
298 	if (BP_GET_LEVEL(bp) > 0) {
299 		uint32_t flags = ARC_FLAG_WAIT;
300 		int32_t i, ptidx, pidx;
301 		uint32_t prefetchlimit;
302 		int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
303 		zbookmark_phys_t *czb;
304 
305 		ASSERT(!BP_IS_PROTECTED(bp));
306 
307 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
308 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
309 		if (err != 0)
310 			goto post;
311 
312 		czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
313 
314 		/*
315 		 * When performing a traversal it is beneficial to
316 		 * asynchronously read-ahead the upcoming indirect
317 		 * blocks since they will be needed shortly. However,
318 		 * since a 128k indirect (non-L0) block may contain up
319 		 * to 1024 128-byte block pointers, its preferable to not
320 		 * prefetch them all at once. Issuing a large number of
321 		 * async reads may effect performance, and the earlier
322 		 * the indirect blocks are prefetched the less likely
323 		 * they are to still be resident in the ARC when needed.
324 		 * Therefore, prefetching indirect blocks is limited to
325 		 * zfs_traverse_indirect_prefetch_limit=32 blocks by
326 		 * default.
327 		 *
328 		 * pidx: Index for which next prefetch to be issued.
329 		 * ptidx: Index at which next prefetch to be triggered.
330 		 */
331 		ptidx = 0;
332 		pidx = 1;
333 		prefetchlimit = zfs_traverse_indirect_prefetch_limit;
334 		for (i = 0; i < epb; i++) {
335 			if (prefetchlimit && i == ptidx) {
336 				ASSERT3S(ptidx, <=, pidx);
337 				for (uint32_t  prefetched = 0; pidx < epb &&
338 				    prefetched < prefetchlimit; pidx++) {
339 					SET_BOOKMARK(czb, zb->zb_objset,
340 					    zb->zb_object, zb->zb_level - 1,
341 					    zb->zb_blkid * epb + pidx);
342 					if (traverse_prefetch_metadata(td, dnp,
343 					    &((blkptr_t *)buf->b_data)[pidx],
344 					    czb) == B_TRUE) {
345 						prefetched++;
346 						if (prefetched ==
347 						    MAX(prefetchlimit / 2, 1))
348 							ptidx = pidx;
349 					}
350 				}
351 			}
352 
353 			/* recursively visitbp() blocks below this */
354 			SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
355 			    zb->zb_level - 1,
356 			    zb->zb_blkid * epb + i);
357 			err = traverse_visitbp(td, dnp,
358 			    &((blkptr_t *)buf->b_data)[i], czb);
359 			if (err != 0)
360 				break;
361 		}
362 
363 		kmem_free(czb, sizeof (zbookmark_phys_t));
364 
365 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
366 		uint32_t flags = ARC_FLAG_WAIT;
367 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
368 		int32_t i;
369 		int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
370 		dnode_phys_t *child_dnp;
371 
372 		/*
373 		 * dnode blocks might have their bonus buffers encrypted, so
374 		 * we must be careful to honor TRAVERSE_NO_DECRYPT
375 		 */
376 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
377 			zio_flags |= ZIO_FLAG_RAW;
378 
379 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
380 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
381 		if (err != 0)
382 			goto post;
383 
384 		child_dnp = buf->b_data;
385 
386 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
387 			prefetch_dnode_metadata(td, &child_dnp[i],
388 			    zb->zb_objset, zb->zb_blkid * epb + i);
389 		}
390 
391 		/* recursively visitbp() blocks below this */
392 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
393 			err = traverse_dnode(td, bp, &child_dnp[i],
394 			    zb->zb_objset, zb->zb_blkid * epb + i);
395 			if (err != 0)
396 				break;
397 		}
398 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
399 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
400 		arc_flags_t flags = ARC_FLAG_WAIT;
401 		objset_phys_t *osp;
402 
403 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
404 			zio_flags |= ZIO_FLAG_RAW;
405 
406 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
407 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
408 		if (err != 0)
409 			goto post;
410 
411 		osp = buf->b_data;
412 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
413 		    DMU_META_DNODE_OBJECT);
414 		/*
415 		 * See the block comment above for the goal of this variable.
416 		 * If the maxblkid of the meta-dnode is 0, then we know that
417 		 * we've never had more than DNODES_PER_BLOCK objects in the
418 		 * dataset, which means we can't have reused any object ids.
419 		 */
420 		if (osp->os_meta_dnode.dn_maxblkid == 0)
421 			td->td_realloc_possible = B_FALSE;
422 
423 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
424 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
425 				prefetch_dnode_metadata(td,
426 				    &osp->os_projectused_dnode,
427 				    zb->zb_objset, DMU_PROJECTUSED_OBJECT);
428 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
429 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
430 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
431 			    zb->zb_objset, DMU_USERUSED_OBJECT);
432 		}
433 
434 		err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
435 		    DMU_META_DNODE_OBJECT);
436 		if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
437 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
438 				err = traverse_dnode(td, bp,
439 				    &osp->os_projectused_dnode, zb->zb_objset,
440 				    DMU_PROJECTUSED_OBJECT);
441 			if (err == 0)
442 				err = traverse_dnode(td, bp,
443 				    &osp->os_groupused_dnode, zb->zb_objset,
444 				    DMU_GROUPUSED_OBJECT);
445 			if (err == 0)
446 				err = traverse_dnode(td, bp,
447 				    &osp->os_userused_dnode, zb->zb_objset,
448 				    DMU_USERUSED_OBJECT);
449 		}
450 	}
451 
452 	if (buf)
453 		arc_buf_destroy(buf, &buf);
454 
455 post:
456 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
457 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
458 
459 	if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
460 		/*
461 		 * Ignore this disk error as requested by the HARD flag,
462 		 * and continue traversal.
463 		 */
464 		err = 0;
465 	}
466 
467 	/*
468 	 * If we are stopping here, set td_resume.
469 	 */
470 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
471 		td->td_resume->zb_objset = zb->zb_objset;
472 		td->td_resume->zb_object = zb->zb_object;
473 		td->td_resume->zb_level = 0;
474 		/*
475 		 * If we have stopped on an indirect block (e.g. due to
476 		 * i/o error), we have not visited anything below it.
477 		 * Set the bookmark to the first level-0 block that we need
478 		 * to visit.  This way, the resuming code does not need to
479 		 * deal with resuming from indirect blocks.
480 		 *
481 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
482 		 * to dereference it.
483 		 */
484 		td->td_resume->zb_blkid = zb->zb_blkid;
485 		if (zb->zb_level > 0) {
486 			td->td_resume->zb_blkid <<= zb->zb_level *
487 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
488 		}
489 		td->td_paused = B_TRUE;
490 	}
491 
492 	return (err);
493 }
494 
495 static void
prefetch_dnode_metadata(traverse_data_t * td,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)496 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
497     uint64_t objset, uint64_t object)
498 {
499 	int j;
500 	zbookmark_phys_t czb;
501 
502 	for (j = 0; j < dnp->dn_nblkptr; j++) {
503 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
504 		traverse_prefetch_metadata(td, dnp, &dnp->dn_blkptr[j], &czb);
505 	}
506 
507 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
508 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
509 		traverse_prefetch_metadata(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
510 	}
511 }
512 
513 static int
traverse_dnode(traverse_data_t * td,const blkptr_t * bp,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)514 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
515     uint64_t objset, uint64_t object)
516 {
517 	int j, err = 0;
518 	zbookmark_phys_t czb;
519 
520 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
521 	    object < td->td_resume->zb_object)
522 		return (0);
523 
524 	if (td->td_flags & TRAVERSE_PRE) {
525 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
526 		    ZB_DNODE_BLKID);
527 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
528 		    td->td_arg);
529 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
530 			return (0);
531 		if (err != 0)
532 			return (err);
533 	}
534 
535 	for (j = 0; j < dnp->dn_nblkptr; j++) {
536 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
537 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
538 		if (err != 0)
539 			break;
540 	}
541 
542 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
543 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
544 		err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
545 	}
546 
547 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
548 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
549 		    ZB_DNODE_BLKID);
550 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
551 		    td->td_arg);
552 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
553 			return (0);
554 		if (err != 0)
555 			return (err);
556 	}
557 	return (err);
558 }
559 
560 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)561 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
562     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
563 {
564 	(void) zilog, (void) dnp;
565 	prefetch_data_t *pfd = arg;
566 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
567 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
568 	    ARC_FLAG_PRESCIENT_PREFETCH;
569 
570 	ASSERT(pfd->pd_bytes_fetched >= 0);
571 	if (zb->zb_level == ZB_DNODE_LEVEL)
572 		return (0);
573 	if (pfd->pd_cancel)
574 		return (SET_ERROR(EINTR));
575 
576 	if (!prefetch_needed(pfd, bp))
577 		return (0);
578 
579 	mutex_enter(&pfd->pd_mtx);
580 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
581 		cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
582 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
583 	cv_broadcast(&pfd->pd_cv);
584 	mutex_exit(&pfd->pd_mtx);
585 
586 	if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
587 		zio_flags |= ZIO_FLAG_RAW;
588 
589 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
590 	    zio_flags, &aflags, zb);
591 
592 	return (0);
593 }
594 
595 static void
traverse_prefetch_thread(void * arg)596 traverse_prefetch_thread(void *arg)
597 {
598 	traverse_data_t *td_main = arg;
599 	traverse_data_t td = *td_main;
600 	zbookmark_phys_t czb;
601 	fstrans_cookie_t cookie = spl_fstrans_mark();
602 
603 	td.td_func = traverse_prefetcher;
604 	td.td_arg = td_main->td_pfd;
605 	td.td_pfd = NULL;
606 	td.td_resume = &td_main->td_pfd->pd_resume;
607 
608 	SET_BOOKMARK(&czb, td.td_objset,
609 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
610 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
611 
612 	mutex_enter(&td_main->td_pfd->pd_mtx);
613 	td_main->td_pfd->pd_exited = B_TRUE;
614 	cv_broadcast(&td_main->td_pfd->pd_cv);
615 	mutex_exit(&td_main->td_pfd->pd_mtx);
616 	spl_fstrans_unmark(cookie);
617 }
618 
619 /*
620  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
621  * in syncing context).
622  */
623 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)624 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
625     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
626     blkptr_cb_t func, void *arg)
627 {
628 	traverse_data_t *td;
629 	prefetch_data_t *pd;
630 	zbookmark_phys_t *czb;
631 	int err;
632 
633 	ASSERT(ds == NULL || objset == ds->ds_object);
634 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
635 
636 	td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
637 	pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
638 	czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
639 
640 	td->td_spa = spa;
641 	td->td_objset = objset;
642 	td->td_rootbp = rootbp;
643 	td->td_min_txg = txg_start;
644 	td->td_resume = resume;
645 	td->td_func = func;
646 	td->td_arg = arg;
647 	td->td_pfd = pd;
648 	td->td_flags = flags;
649 	td->td_paused = B_FALSE;
650 	td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
651 
652 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
653 		VERIFY(spa_feature_enabled_txg(spa,
654 		    SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
655 	} else {
656 		td->td_hole_birth_enabled_txg = UINT64_MAX;
657 	}
658 
659 	pd->pd_flags = flags;
660 	if (resume != NULL)
661 		pd->pd_resume = *resume;
662 	mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
663 	cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
664 
665 	SET_BOOKMARK(czb, td->td_objset,
666 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
667 
668 	/* See comment on ZIL traversal in dsl_scan_visitds. */
669 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
670 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
671 		uint32_t flags = ARC_FLAG_WAIT;
672 		objset_phys_t *osp;
673 		arc_buf_t *buf;
674 		ASSERT(!BP_IS_REDACTED(rootbp));
675 
676 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
677 		    BP_IS_PROTECTED(rootbp))
678 			zio_flags |= ZIO_FLAG_RAW;
679 
680 		err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
681 		    &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
682 		if (err != 0) {
683 			/*
684 			 * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
685 			 * continue to visitbp so that td_func can be called
686 			 * in pre stage, and err will reset to zero.
687 			 */
688 			if (!(td->td_flags & TRAVERSE_HARD) ||
689 			    !(td->td_flags & TRAVERSE_PRE))
690 				goto out;
691 		} else {
692 			osp = buf->b_data;
693 			traverse_zil(td, &osp->os_zil_header);
694 			arc_buf_destroy(buf, &buf);
695 		}
696 	}
697 
698 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
699 	    taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
700 	    td, TQ_NOQUEUE) == TASKQID_INVALID)
701 		pd->pd_exited = B_TRUE;
702 
703 	err = traverse_visitbp(td, NULL, rootbp, czb);
704 
705 	mutex_enter(&pd->pd_mtx);
706 	pd->pd_cancel = B_TRUE;
707 	cv_broadcast(&pd->pd_cv);
708 	while (!pd->pd_exited)
709 		cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
710 	mutex_exit(&pd->pd_mtx);
711 out:
712 	mutex_destroy(&pd->pd_mtx);
713 	cv_destroy(&pd->pd_cv);
714 
715 	kmem_free(czb, sizeof (zbookmark_phys_t));
716 	kmem_free(pd, sizeof (struct prefetch_data));
717 	kmem_free(td, sizeof (struct traverse_data));
718 
719 	return (err);
720 }
721 
722 /*
723  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
724  * in syncing context).
725  */
726 int
traverse_dataset_resume(dsl_dataset_t * ds,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)727 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
728     zbookmark_phys_t *resume,
729     int flags, blkptr_cb_t func, void *arg)
730 {
731 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
732 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
733 }
734 
735 int
traverse_dataset(dsl_dataset_t * ds,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)736 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
737     int flags, blkptr_cb_t func, void *arg)
738 {
739 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
740 }
741 
742 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)743 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
744     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
745     blkptr_cb_t func, void *arg)
746 {
747 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
748 	    blkptr, txg_start, resume, flags, func, arg));
749 }
750 
751 /*
752  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
753  */
754 int
traverse_pool(spa_t * spa,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)755 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
756     blkptr_cb_t func, void *arg)
757 {
758 	int err;
759 	dsl_pool_t *dp = spa_get_dsl(spa);
760 	objset_t *mos = dp->dp_meta_objset;
761 	boolean_t hard = (flags & TRAVERSE_HARD);
762 
763 	/* visit the MOS */
764 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
765 	    txg_start, NULL, flags, func, arg);
766 	if (err != 0)
767 		return (err);
768 
769 	/* visit each dataset */
770 	for (uint64_t obj = 1; err == 0;
771 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
772 		dmu_object_info_t doi;
773 
774 		err = dmu_object_info(mos, obj, &doi);
775 		if (err != 0) {
776 			if (hard)
777 				continue;
778 			break;
779 		}
780 
781 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
782 			dsl_dataset_t *ds;
783 			uint64_t txg = txg_start;
784 
785 			dsl_pool_config_enter(dp, FTAG);
786 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
787 			dsl_pool_config_exit(dp, FTAG);
788 			if (err != 0) {
789 				if (hard)
790 					continue;
791 				break;
792 			}
793 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
794 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
795 			err = traverse_dataset(ds, txg, flags, func, arg);
796 			dsl_dataset_rele(ds, FTAG);
797 			if (err != 0)
798 				break;
799 		}
800 	}
801 	if (err == ESRCH)
802 		err = 0;
803 	return (err);
804 }
805 
806 EXPORT_SYMBOL(traverse_dataset);
807 EXPORT_SYMBOL(traverse_pool);
808 
809 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
810 	"Max number of bytes to prefetch");
811 
812 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, UINT, ZMOD_RW,
813 	"Traverse prefetch number of blocks pointed by indirect block");
814 
815 #if defined(_KERNEL)
816 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644);
817 MODULE_PARM_DESC(ignore_hole_birth,
818 	"Alias for send_holes_without_birth_time");
819 #endif
820 
821 /* CSTYLED */
822 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
823 	"Ignore hole_birth txg for zfs send");
824