xref: /freebsd/sys/contrib/openzfs/module/zfs/dmu_traverse.c (revision 61145dc2b94f12f6a47344fb9aac702321880e43)
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 int
traverse_zil_block(zilog_t * zilog,const blkptr_t * bp,void * arg,uint64_t claim_txg)78 traverse_zil_block(zilog_t *zilog, const blkptr_t *bp, void *arg,
79     uint64_t claim_txg)
80 {
81 	traverse_data_t *td = arg;
82 	zbookmark_phys_t zb;
83 
84 	if (BP_IS_HOLE(bp))
85 		return (0);
86 
87 	if (claim_txg == 0 &&
88 	    BP_GET_LOGICAL_BIRTH(bp) >= spa_min_claim_txg(td->td_spa))
89 		return (-1);
90 
91 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
92 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
93 
94 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
95 
96 	return (0);
97 }
98 
99 static int
traverse_zil_record(zilog_t * zilog,const lr_t * lrc,void * arg,uint64_t claim_txg)100 traverse_zil_record(zilog_t *zilog, const lr_t *lrc, void *arg,
101     uint64_t claim_txg)
102 {
103 	traverse_data_t *td = arg;
104 
105 	if (lrc->lrc_txtype == TX_WRITE) {
106 		lr_write_t *lr = (lr_write_t *)lrc;
107 		blkptr_t *bp = &lr->lr_blkptr;
108 		zbookmark_phys_t zb;
109 
110 		if (BP_IS_HOLE(bp))
111 			return (0);
112 
113 		if (claim_txg == 0 || BP_GET_LOGICAL_BIRTH(bp) < claim_txg)
114 			return (0);
115 
116 		ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
117 		SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
118 		    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
119 
120 		(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
121 		    td->td_arg);
122 	}
123 	return (0);
124 }
125 
126 static void
traverse_zil(traverse_data_t * td,zil_header_t * zh)127 traverse_zil(traverse_data_t *td, zil_header_t *zh)
128 {
129 	uint64_t claim_txg = zh->zh_claim_txg;
130 
131 	/*
132 	 * We only want to visit blocks that have been claimed but not yet
133 	 * replayed; plus blocks that are already stable in read-only mode.
134 	 */
135 	if (claim_txg == 0 && spa_writeable(td->td_spa))
136 		return;
137 
138 	zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
139 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
140 	    claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT));
141 	zil_free(zilog);
142 }
143 
144 typedef enum resume_skip {
145 	RESUME_SKIP_ALL,
146 	RESUME_SKIP_NONE,
147 	RESUME_SKIP_CHILDREN
148 } resume_skip_t;
149 
150 /*
151  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
152  * the block indicated by zb does not need to be visited at all. Returns
153  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
154  * resume point. This indicates that this block should be visited but not its
155  * children (since they must have been visited in a previous traversal).
156  * Otherwise returns RESUME_SKIP_NONE.
157  */
158 static resume_skip_t
resume_skip_check(const traverse_data_t * td,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)159 resume_skip_check(const traverse_data_t *td, const dnode_phys_t *dnp,
160     const zbookmark_phys_t *zb)
161 {
162 	if (td->td_resume != NULL) {
163 		/*
164 		 * If we already visited this bp & everything below,
165 		 * don't bother doing it again.
166 		 */
167 		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
168 			return (RESUME_SKIP_ALL);
169 
170 		if (memcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
171 			if (td->td_flags & TRAVERSE_POST)
172 				return (RESUME_SKIP_CHILDREN);
173 		}
174 	}
175 	return (RESUME_SKIP_NONE);
176 }
177 
178 /*
179  * Returns B_TRUE, if prefetch read is issued, otherwise B_FALSE.
180  */
181 static boolean_t
traverse_prefetch_metadata(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)182 traverse_prefetch_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
183     const blkptr_t *bp, const zbookmark_phys_t *zb)
184 {
185 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
186 	    ARC_FLAG_PRESCIENT_PREFETCH;
187 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
188 
189 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
190 		return (B_FALSE);
191 	/*
192 	 * If this bp is before the resume point, it may have already been
193 	 * freed.
194 	 */
195 	if (resume_skip_check(td, dnp, zb) != RESUME_SKIP_NONE)
196 		return (B_FALSE);
197 	if (BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) <= td->td_min_txg)
198 		return (B_FALSE);
199 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
200 		return (B_FALSE);
201 	ASSERT(!BP_IS_REDACTED(bp));
202 
203 	if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
204 		zio_flags |= ZIO_FLAG_RAW;
205 
206 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
207 	    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
208 	return (B_TRUE);
209 }
210 
211 static boolean_t
prefetch_needed(prefetch_data_t * pfd,const blkptr_t * bp)212 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
213 {
214 	ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
215 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
216 	    BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG || BP_IS_REDACTED(bp))
217 		return (B_FALSE);
218 	return (B_TRUE);
219 }
220 
221 static int
traverse_visitbp(traverse_data_t * td,const dnode_phys_t * dnp,const blkptr_t * bp,const zbookmark_phys_t * zb)222 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
223     const blkptr_t *bp, const zbookmark_phys_t *zb)
224 {
225 	int err = 0;
226 	arc_buf_t *buf = NULL;
227 	prefetch_data_t *pd = td->td_pfd;
228 
229 	switch (resume_skip_check(td, dnp, zb)) {
230 	case RESUME_SKIP_ALL:
231 		return (0);
232 	case RESUME_SKIP_CHILDREN:
233 		goto post;
234 	case RESUME_SKIP_NONE:
235 		break;
236 	default:
237 		ASSERT(0);
238 	}
239 
240 	if (BP_GET_LOGICAL_BIRTH(bp) == 0) {
241 		/*
242 		 * Since this block has a birth time of 0 it must be one of
243 		 * two things: a hole created before the
244 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
245 		 * which has always been a hole in an object.
246 		 *
247 		 * If a file is written sparsely, then the unwritten parts of
248 		 * the file were "always holes" -- that is, they have been
249 		 * holes since this object was allocated.  However, we (and
250 		 * our callers) can not necessarily tell when an object was
251 		 * allocated.  Therefore, if it's possible that this object
252 		 * was freed and then its object number reused, we need to
253 		 * visit all the holes with birth==0.
254 		 *
255 		 * If it isn't possible that the object number was reused,
256 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
257 		 * all the blocks we will visit as part of this traversal,
258 		 * then this hole must have always existed, so we can skip
259 		 * it.  We visit blocks born after (exclusive) td_min_txg.
260 		 *
261 		 * Note that the meta-dnode cannot be reallocated.
262 		 */
263 		if (!send_holes_without_birth_time &&
264 		    (!td->td_realloc_possible ||
265 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
266 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
267 			return (0);
268 	} else if (BP_GET_LOGICAL_BIRTH(bp) <= td->td_min_txg) {
269 		return (0);
270 	}
271 
272 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
273 		uint64_t size = BP_GET_LSIZE(bp);
274 		mutex_enter(&pd->pd_mtx);
275 		ASSERT(pd->pd_bytes_fetched >= 0);
276 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
277 			cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
278 		pd->pd_bytes_fetched -= size;
279 		cv_broadcast(&pd->pd_cv);
280 		mutex_exit(&pd->pd_mtx);
281 	}
282 
283 	if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) {
284 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
285 		if (err != 0)
286 			goto post;
287 		return (0);
288 	}
289 
290 	if (td->td_flags & TRAVERSE_PRE) {
291 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
292 		    td->td_arg);
293 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
294 			return (0);
295 		if (err != 0)
296 			goto post;
297 	}
298 
299 	if (BP_GET_LEVEL(bp) > 0) {
300 		uint32_t flags = ARC_FLAG_WAIT;
301 		int32_t i, ptidx, pidx;
302 		uint32_t prefetchlimit;
303 		int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
304 		zbookmark_phys_t *czb;
305 
306 		ASSERT(!BP_IS_PROTECTED(bp));
307 
308 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
309 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
310 		if (err != 0)
311 			goto post;
312 
313 		czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
314 
315 		/*
316 		 * When performing a traversal it is beneficial to
317 		 * asynchronously read-ahead the upcoming indirect
318 		 * blocks since they will be needed shortly. However,
319 		 * since a 128k indirect (non-L0) block may contain up
320 		 * to 1024 128-byte block pointers, its preferable to not
321 		 * prefetch them all at once. Issuing a large number of
322 		 * async reads may effect performance, and the earlier
323 		 * the indirect blocks are prefetched the less likely
324 		 * they are to still be resident in the ARC when needed.
325 		 * Therefore, prefetching indirect blocks is limited to
326 		 * zfs_traverse_indirect_prefetch_limit=32 blocks by
327 		 * default.
328 		 *
329 		 * pidx: Index for which next prefetch to be issued.
330 		 * ptidx: Index at which next prefetch to be triggered.
331 		 */
332 		ptidx = 0;
333 		pidx = 1;
334 		prefetchlimit = zfs_traverse_indirect_prefetch_limit;
335 		for (i = 0; i < epb; i++) {
336 			if (prefetchlimit && i == ptidx) {
337 				ASSERT3S(ptidx, <=, pidx);
338 				for (uint32_t  prefetched = 0; pidx < epb &&
339 				    prefetched < prefetchlimit; pidx++) {
340 					SET_BOOKMARK(czb, zb->zb_objset,
341 					    zb->zb_object, zb->zb_level - 1,
342 					    zb->zb_blkid * epb + pidx);
343 					if (traverse_prefetch_metadata(td, dnp,
344 					    &((blkptr_t *)buf->b_data)[pidx],
345 					    czb) == B_TRUE) {
346 						prefetched++;
347 						if (prefetched ==
348 						    MAX(prefetchlimit / 2, 1))
349 							ptidx = pidx;
350 					}
351 				}
352 			}
353 
354 			/* recursively visitbp() blocks below this */
355 			SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object,
356 			    zb->zb_level - 1,
357 			    zb->zb_blkid * epb + i);
358 			err = traverse_visitbp(td, dnp,
359 			    &((blkptr_t *)buf->b_data)[i], czb);
360 			if (err != 0)
361 				break;
362 		}
363 
364 		kmem_free(czb, sizeof (zbookmark_phys_t));
365 
366 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
367 		uint32_t flags = ARC_FLAG_WAIT;
368 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
369 		int32_t i;
370 		int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
371 		dnode_phys_t *child_dnp;
372 
373 		/*
374 		 * dnode blocks might have their bonus buffers encrypted, so
375 		 * we must be careful to honor TRAVERSE_NO_DECRYPT
376 		 */
377 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
378 			zio_flags |= ZIO_FLAG_RAW;
379 
380 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
381 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
382 		if (err != 0)
383 			goto post;
384 
385 		child_dnp = buf->b_data;
386 
387 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
388 			prefetch_dnode_metadata(td, &child_dnp[i],
389 			    zb->zb_objset, zb->zb_blkid * epb + i);
390 		}
391 
392 		/* recursively visitbp() blocks below this */
393 		for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) {
394 			err = traverse_dnode(td, bp, &child_dnp[i],
395 			    zb->zb_objset, zb->zb_blkid * epb + i);
396 			if (err != 0)
397 				break;
398 		}
399 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
400 		uint32_t zio_flags = ZIO_FLAG_CANFAIL;
401 		arc_flags_t flags = ARC_FLAG_WAIT;
402 		objset_phys_t *osp;
403 
404 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
405 			zio_flags |= ZIO_FLAG_RAW;
406 
407 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
408 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
409 		if (err != 0)
410 			goto post;
411 
412 		osp = buf->b_data;
413 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
414 		    DMU_META_DNODE_OBJECT);
415 		/*
416 		 * See the block comment above for the goal of this variable.
417 		 * If the maxblkid of the meta-dnode is 0, then we know that
418 		 * we've never had more than DNODES_PER_BLOCK objects in the
419 		 * dataset, which means we can't have reused any object ids.
420 		 */
421 		if (osp->os_meta_dnode.dn_maxblkid == 0)
422 			td->td_realloc_possible = B_FALSE;
423 
424 		if (OBJSET_BUF_HAS_USERUSED(buf)) {
425 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
426 				prefetch_dnode_metadata(td,
427 				    &osp->os_projectused_dnode,
428 				    zb->zb_objset, DMU_PROJECTUSED_OBJECT);
429 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
430 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
431 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
432 			    zb->zb_objset, DMU_USERUSED_OBJECT);
433 		}
434 
435 		err = traverse_dnode(td, bp, &osp->os_meta_dnode, zb->zb_objset,
436 		    DMU_META_DNODE_OBJECT);
437 		if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) {
438 			if (OBJSET_BUF_HAS_PROJECTUSED(buf))
439 				err = traverse_dnode(td, bp,
440 				    &osp->os_projectused_dnode, zb->zb_objset,
441 				    DMU_PROJECTUSED_OBJECT);
442 			if (err == 0)
443 				err = traverse_dnode(td, bp,
444 				    &osp->os_groupused_dnode, zb->zb_objset,
445 				    DMU_GROUPUSED_OBJECT);
446 			if (err == 0)
447 				err = traverse_dnode(td, bp,
448 				    &osp->os_userused_dnode, zb->zb_objset,
449 				    DMU_USERUSED_OBJECT);
450 		}
451 	}
452 
453 	if (buf)
454 		arc_buf_destroy(buf, &buf);
455 
456 post:
457 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
458 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
459 
460 	if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) {
461 		/*
462 		 * Ignore this disk error as requested by the HARD flag,
463 		 * and continue traversal.
464 		 */
465 		err = 0;
466 	}
467 
468 	/*
469 	 * If we are stopping here, set td_resume.
470 	 */
471 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
472 		td->td_resume->zb_objset = zb->zb_objset;
473 		td->td_resume->zb_object = zb->zb_object;
474 		td->td_resume->zb_level = 0;
475 		/*
476 		 * If we have stopped on an indirect block (e.g. due to
477 		 * i/o error), we have not visited anything below it.
478 		 * Set the bookmark to the first level-0 block that we need
479 		 * to visit.  This way, the resuming code does not need to
480 		 * deal with resuming from indirect blocks.
481 		 *
482 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
483 		 * to dereference it.
484 		 */
485 		td->td_resume->zb_blkid = zb->zb_blkid;
486 		if (zb->zb_level > 0) {
487 			td->td_resume->zb_blkid <<= zb->zb_level *
488 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
489 		}
490 		td->td_paused = B_TRUE;
491 	}
492 
493 	return (err);
494 }
495 
496 static void
prefetch_dnode_metadata(traverse_data_t * td,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)497 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
498     uint64_t objset, uint64_t object)
499 {
500 	int j;
501 	zbookmark_phys_t czb;
502 
503 	for (j = 0; j < dnp->dn_nblkptr; j++) {
504 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
505 		traverse_prefetch_metadata(td, dnp, &dnp->dn_blkptr[j], &czb);
506 	}
507 
508 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
509 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
510 		traverse_prefetch_metadata(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
511 	}
512 }
513 
514 static int
traverse_dnode(traverse_data_t * td,const blkptr_t * bp,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)515 traverse_dnode(traverse_data_t *td, const blkptr_t *bp, const dnode_phys_t *dnp,
516     uint64_t objset, uint64_t object)
517 {
518 	int j, err = 0;
519 	zbookmark_phys_t czb;
520 
521 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
522 	    object < td->td_resume->zb_object)
523 		return (0);
524 
525 	if (td->td_flags & TRAVERSE_PRE) {
526 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
527 		    ZB_DNODE_BLKID);
528 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
529 		    td->td_arg);
530 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
531 			return (0);
532 		if (err != 0)
533 			return (err);
534 	}
535 
536 	for (j = 0; j < dnp->dn_nblkptr; j++) {
537 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
538 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
539 		if (err != 0)
540 			break;
541 	}
542 
543 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
544 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
545 		err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb);
546 	}
547 
548 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
549 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
550 		    ZB_DNODE_BLKID);
551 		err = td->td_func(td->td_spa, NULL, bp, &czb, dnp,
552 		    td->td_arg);
553 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
554 			return (0);
555 		if (err != 0)
556 			return (err);
557 	}
558 	return (err);
559 }
560 
561 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)562 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
563     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
564 {
565 	(void) zilog, (void) dnp;
566 	prefetch_data_t *pfd = arg;
567 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
568 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH |
569 	    ARC_FLAG_PRESCIENT_PREFETCH;
570 
571 	ASSERT(pfd->pd_bytes_fetched >= 0);
572 	if (zb->zb_level == ZB_DNODE_LEVEL)
573 		return (0);
574 	if (pfd->pd_cancel)
575 		return (SET_ERROR(EINTR));
576 
577 	if (!prefetch_needed(pfd, bp))
578 		return (0);
579 
580 	mutex_enter(&pfd->pd_mtx);
581 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
582 		cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx);
583 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
584 	cv_broadcast(&pfd->pd_cv);
585 	mutex_exit(&pfd->pd_mtx);
586 
587 	if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp))
588 		zio_flags |= ZIO_FLAG_RAW;
589 
590 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
591 	    zio_flags, &aflags, zb);
592 
593 	return (0);
594 }
595 
596 static void
traverse_prefetch_thread(void * arg)597 traverse_prefetch_thread(void *arg)
598 {
599 	traverse_data_t *td_main = arg;
600 	traverse_data_t td = *td_main;
601 	zbookmark_phys_t czb;
602 	fstrans_cookie_t cookie = spl_fstrans_mark();
603 
604 	td.td_func = traverse_prefetcher;
605 	td.td_arg = td_main->td_pfd;
606 	td.td_pfd = NULL;
607 	td.td_resume = &td_main->td_pfd->pd_resume;
608 
609 	SET_BOOKMARK(&czb, td.td_objset,
610 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
611 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
612 
613 	mutex_enter(&td_main->td_pfd->pd_mtx);
614 	td_main->td_pfd->pd_exited = B_TRUE;
615 	cv_broadcast(&td_main->td_pfd->pd_cv);
616 	mutex_exit(&td_main->td_pfd->pd_mtx);
617 	spl_fstrans_unmark(cookie);
618 }
619 
620 /*
621  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
622  * in syncing context).
623  */
624 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)625 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
626     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
627     blkptr_cb_t func, void *arg)
628 {
629 	traverse_data_t *td;
630 	prefetch_data_t *pd;
631 	zbookmark_phys_t *czb;
632 	int err;
633 
634 	ASSERT(ds == NULL || objset == ds->ds_object);
635 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
636 
637 	td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP);
638 	pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP);
639 	czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP);
640 
641 	td->td_spa = spa;
642 	td->td_objset = objset;
643 	td->td_rootbp = rootbp;
644 	td->td_min_txg = txg_start;
645 	td->td_resume = resume;
646 	td->td_func = func;
647 	td->td_arg = arg;
648 	td->td_pfd = pd;
649 	td->td_flags = flags;
650 	td->td_paused = B_FALSE;
651 	td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
652 
653 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
654 		VERIFY(spa_feature_enabled_txg(spa,
655 		    SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg));
656 	} else {
657 		td->td_hole_birth_enabled_txg = UINT64_MAX;
658 	}
659 
660 	pd->pd_flags = flags;
661 	if (resume != NULL)
662 		pd->pd_resume = *resume;
663 	mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
664 	cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
665 
666 	SET_BOOKMARK(czb, td->td_objset,
667 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
668 
669 	/* See comment on ZIL traversal in dsl_scan_visitds. */
670 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
671 		zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
672 		uint32_t flags = ARC_FLAG_WAIT;
673 		objset_phys_t *osp;
674 		arc_buf_t *buf;
675 		ASSERT(!BP_IS_REDACTED(rootbp));
676 
677 		if ((td->td_flags & TRAVERSE_NO_DECRYPT) &&
678 		    BP_IS_PROTECTED(rootbp))
679 			zio_flags |= ZIO_FLAG_RAW;
680 
681 		err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func,
682 		    &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb);
683 		if (err != 0) {
684 			/*
685 			 * If both TRAVERSE_HARD and TRAVERSE_PRE are set,
686 			 * continue to visitbp so that td_func can be called
687 			 * in pre stage, and err will reset to zero.
688 			 */
689 			if (!(td->td_flags & TRAVERSE_HARD) ||
690 			    !(td->td_flags & TRAVERSE_PRE))
691 				goto out;
692 		} else {
693 			osp = buf->b_data;
694 			traverse_zil(td, &osp->os_zil_header);
695 			arc_buf_destroy(buf, &buf);
696 		}
697 	}
698 
699 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
700 	    taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread,
701 	    td, TQ_NOQUEUE) == TASKQID_INVALID)
702 		pd->pd_exited = B_TRUE;
703 
704 	err = traverse_visitbp(td, NULL, rootbp, czb);
705 
706 	mutex_enter(&pd->pd_mtx);
707 	pd->pd_cancel = B_TRUE;
708 	cv_broadcast(&pd->pd_cv);
709 	while (!pd->pd_exited)
710 		cv_wait_sig(&pd->pd_cv, &pd->pd_mtx);
711 	mutex_exit(&pd->pd_mtx);
712 out:
713 	mutex_destroy(&pd->pd_mtx);
714 	cv_destroy(&pd->pd_cv);
715 
716 	kmem_free(czb, sizeof (zbookmark_phys_t));
717 	kmem_free(pd, sizeof (struct prefetch_data));
718 	kmem_free(td, sizeof (struct traverse_data));
719 
720 	return (err);
721 }
722 
723 /*
724  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
725  * in syncing context).
726  */
727 int
traverse_dataset_resume(dsl_dataset_t * ds,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)728 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
729     zbookmark_phys_t *resume,
730     int flags, blkptr_cb_t func, void *arg)
731 {
732 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
733 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
734 }
735 
736 int
traverse_dataset(dsl_dataset_t * ds,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)737 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
738     int flags, blkptr_cb_t func, void *arg)
739 {
740 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
741 }
742 
743 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)744 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
745     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
746     blkptr_cb_t func, void *arg)
747 {
748 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
749 	    blkptr, txg_start, resume, flags, func, arg));
750 }
751 
752 /*
753  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
754  */
755 int
traverse_pool(spa_t * spa,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)756 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
757     blkptr_cb_t func, void *arg)
758 {
759 	int err;
760 	dsl_pool_t *dp = spa_get_dsl(spa);
761 	objset_t *mos = dp->dp_meta_objset;
762 	boolean_t hard = (flags & TRAVERSE_HARD);
763 
764 	/* visit the MOS */
765 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
766 	    txg_start, NULL, flags, func, arg);
767 	if (err != 0)
768 		return (err);
769 
770 	/* visit each dataset */
771 	for (uint64_t obj = 1; err == 0;
772 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
773 		dmu_object_info_t doi;
774 
775 		err = dmu_object_info(mos, obj, &doi);
776 		if (err != 0) {
777 			if (hard)
778 				continue;
779 			break;
780 		}
781 
782 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
783 			dsl_dataset_t *ds;
784 			uint64_t txg = txg_start;
785 
786 			dsl_pool_config_enter(dp, FTAG);
787 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
788 			dsl_pool_config_exit(dp, FTAG);
789 			if (err != 0) {
790 				if (hard)
791 					continue;
792 				break;
793 			}
794 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
795 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
796 			err = traverse_dataset(ds, txg, flags, func, arg);
797 			dsl_dataset_rele(ds, FTAG);
798 			if (err != 0)
799 				break;
800 		}
801 	}
802 	if (err == ESRCH)
803 		err = 0;
804 	return (err);
805 }
806 
807 EXPORT_SYMBOL(traverse_dataset);
808 EXPORT_SYMBOL(traverse_pool);
809 
810 ZFS_MODULE_PARAM(zfs, zfs_, pd_bytes_max, INT, ZMOD_RW,
811 	"Max number of bytes to prefetch");
812 
813 ZFS_MODULE_PARAM(zfs, zfs_, traverse_indirect_prefetch_limit, UINT, ZMOD_RW,
814 	"Traverse prefetch number of blocks pointed by indirect block");
815 
816 #if defined(_KERNEL)
817 module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644);
818 MODULE_PARM_DESC(ignore_hole_birth,
819 	"Alias for send_holes_without_birth_time");
820 #endif
821 
822 ZFS_MODULE_PARAM(zfs, , send_holes_without_birth_time, INT, ZMOD_RW,
823 	"Ignore hole_birth txg for zfs send");
824