xref: /netbsd-src/external/cddl/osnet/dist/uts/common/fs/zfs/dmu_traverse.c (revision 3227e6cf668bd374971740bd6660f43cee4417ac)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
24  * Copyright (c) 2015 Chunwei Chen. 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/zio.h>
36 #include <sys/dmu_impl.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/callb.h>
40 #include <sys/zfeature.h>
41 
42 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
43 boolean_t send_holes_without_birth_time = B_TRUE;
44 
45 #ifdef _KERNEL
46 SYSCTL_DECL(_vfs_zfs);
47 SYSCTL_UINT(_vfs_zfs, OID_AUTO, send_holes_without_birth_time, CTLFLAG_RWTUN,
48     &send_holes_without_birth_time, 0, "Send holes without birth time");
49 #endif
50 
51 typedef struct prefetch_data {
52 	kmutex_t pd_mtx;
53 	kcondvar_t pd_cv;
54 	int32_t pd_bytes_fetched;
55 	int pd_flags;
56 	boolean_t pd_cancel;
57 	boolean_t pd_exited;
58 	zbookmark_phys_t pd_resume;
59 } prefetch_data_t;
60 
61 typedef struct traverse_data {
62 	spa_t *td_spa;
63 	uint64_t td_objset;
64 	blkptr_t *td_rootbp;
65 	uint64_t td_min_txg;
66 	zbookmark_phys_t *td_resume;
67 	int td_flags;
68 	prefetch_data_t *td_pfd;
69 	boolean_t td_paused;
70 	uint64_t td_hole_birth_enabled_txg;
71 	blkptr_cb_t *td_func;
72 	void *td_arg;
73 	boolean_t td_realloc_possible;
74 } traverse_data_t;
75 
76 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
77     uint64_t objset, uint64_t object);
78 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
79     uint64_t objset, uint64_t object);
80 
81 static int
traverse_zil_block(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)82 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
83 {
84 	traverse_data_t *td = arg;
85 	zbookmark_phys_t zb;
86 
87 	if (BP_IS_HOLE(bp))
88 		return (0);
89 
90 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
91 		return (0);
92 
93 	SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
94 	    bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
95 
96 	(void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
97 
98 	return (0);
99 }
100 
101 static int
traverse_zil_record(zilog_t * zilog,lr_t * lrc,void * arg,uint64_t claim_txg)102 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
103 {
104 	traverse_data_t *td = arg;
105 
106 	if (lrc->lrc_txtype == TX_WRITE) {
107 		lr_write_t *lr = (lr_write_t *)lrc;
108 		blkptr_t *bp = &lr->lr_blkptr;
109 		zbookmark_phys_t zb;
110 
111 		if (BP_IS_HOLE(bp))
112 			return (0);
113 
114 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
115 			return (0);
116 
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 	zilog_t *zilog;
131 
132 	/*
133 	 * We only want to visit blocks that have been claimed but not yet
134 	 * replayed; plus, in read-only mode, blocks that are already stable.
135 	 */
136 	if (claim_txg == 0 && spa_writeable(td->td_spa))
137 		return;
138 
139 	zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
140 
141 	(void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
142 	    claim_txg);
143 
144 	zil_free(zilog);
145 }
146 
147 typedef enum resume_skip {
148 	RESUME_SKIP_ALL,
149 	RESUME_SKIP_NONE,
150 	RESUME_SKIP_CHILDREN
151 } resume_skip_t;
152 
153 /*
154  * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
155  * the block indicated by zb does not need to be visited at all. Returns
156  * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
157  * resume point. This indicates that this block should be visited but not its
158  * children (since they must have been visited in a previous traversal).
159  * Otherwise returns RESUME_SKIP_NONE.
160  */
161 static resume_skip_t
resume_skip_check(traverse_data_t * td,const dnode_phys_t * dnp,const zbookmark_phys_t * zb)162 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
163     const zbookmark_phys_t *zb)
164 {
165 	if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
166 		/*
167 		 * If we already visited this bp & everything below,
168 		 * don't bother doing it again.
169 		 */
170 		if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
171 			return (RESUME_SKIP_ALL);
172 
173 		/*
174 		 * If we found the block we're trying to resume from, zero
175 		 * the bookmark out to indicate that we have resumed.
176 		 */
177 		if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
178 			bzero(td->td_resume, sizeof (*zb));
179 			if (td->td_flags & TRAVERSE_POST)
180 				return (RESUME_SKIP_CHILDREN);
181 		}
182 	}
183 	return (RESUME_SKIP_NONE);
184 }
185 
186 static void
traverse_prefetch_metadata(traverse_data_t * td,const blkptr_t * bp,const zbookmark_phys_t * zb)187 traverse_prefetch_metadata(traverse_data_t *td,
188     const blkptr_t *bp, const zbookmark_phys_t *zb)
189 {
190 	arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
191 
192 	if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
193 		return;
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;
201 	if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
202 		return;
203 	if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
204 		return;
205 
206 	(void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
207 	    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
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)
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 	zbookmark_phys_t czb;
225 	int err = 0;
226 	arc_buf_t *buf = NULL;
227 	prefetch_data_t *pd = td->td_pfd;
228 	boolean_t hard = td->td_flags & TRAVERSE_HARD;
229 
230 	switch (resume_skip_check(td, dnp, zb)) {
231 	case RESUME_SKIP_ALL:
232 		return (0);
233 	case RESUME_SKIP_CHILDREN:
234 		goto post;
235 	case RESUME_SKIP_NONE:
236 		break;
237 	default:
238 		ASSERT(0);
239 	}
240 
241 	if (bp->blk_birth == 0) {
242 		/*
243 		 * Since this block has a birth time of 0 it must be one of
244 		 * two things: a hole created before the
245 		 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
246 		 * which has always been a hole in an object.
247 		 *
248 		 * If a file is written sparsely, then the unwritten parts of
249 		 * the file were "always holes" -- that is, they have been
250 		 * holes since this object was allocated.  However, we (and
251 		 * our callers) can not necessarily tell when an object was
252 		 * allocated.  Therefore, if it's possible that this object
253 		 * was freed and then its object number reused, we need to
254 		 * visit all the holes with birth==0.
255 		 *
256 		 * If it isn't possible that the object number was reused,
257 		 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
258 		 * all the blocks we will visit as part of this traversal,
259 		 * then this hole must have always existed, so we can skip
260 		 * it.  We visit blocks born after (exclusive) td_min_txg.
261 		 *
262 		 * Note that the meta-dnode cannot be reallocated.
263 		 */
264 		if (!send_holes_without_birth_time &&
265 		    (!td->td_realloc_possible ||
266 		    zb->zb_object == DMU_META_DNODE_OBJECT) &&
267 		    td->td_hole_birth_enabled_txg <= td->td_min_txg)
268 			return (0);
269 	} else if (bp->blk_birth <= td->td_min_txg) {
270 		return (0);
271 	}
272 
273 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
274 		uint64_t size = BP_GET_LSIZE(bp);
275 		mutex_enter(&pd->pd_mtx);
276 		ASSERT(pd->pd_bytes_fetched >= 0);
277 		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
278 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
279 		pd->pd_bytes_fetched -= size;
280 		cv_broadcast(&pd->pd_cv);
281 		mutex_exit(&pd->pd_mtx);
282 	}
283 
284 	if (BP_IS_HOLE(bp)) {
285 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
286 		if (err != 0)
287 			goto post;
288 		return (0);
289 	}
290 
291 	if (td->td_flags & TRAVERSE_PRE) {
292 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
293 		    td->td_arg);
294 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
295 			return (0);
296 		if (err != 0)
297 			goto post;
298 	}
299 
300 	if (BP_GET_LEVEL(bp) > 0) {
301 		arc_flags_t flags = ARC_FLAG_WAIT;
302 		int i;
303 		blkptr_t *cbp;
304 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
305 
306 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
307 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
308 		if (err != 0)
309 			goto post;
310 		cbp = buf->b_data;
311 
312 		for (i = 0; i < epb; i++) {
313 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
314 			    zb->zb_level - 1,
315 			    zb->zb_blkid * epb + i);
316 			traverse_prefetch_metadata(td, &cbp[i], &czb);
317 		}
318 
319 		/* recursively visitbp() blocks below this */
320 		for (i = 0; i < epb; i++) {
321 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
322 			    zb->zb_level - 1,
323 			    zb->zb_blkid * epb + i);
324 			err = traverse_visitbp(td, dnp, &cbp[i], &czb);
325 			if (err != 0)
326 				break;
327 		}
328 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
329 		arc_flags_t flags = ARC_FLAG_WAIT;
330 		int i;
331 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
332 
333 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
334 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
335 		if (err != 0)
336 			goto post;
337 		dnode_phys_t *child_dnp = buf->b_data;
338 
339 		for (i = 0; i < epb; i++) {
340 			prefetch_dnode_metadata(td, &child_dnp[i],
341 			    zb->zb_objset, zb->zb_blkid * epb + i);
342 		}
343 
344 		/* recursively visitbp() blocks below this */
345 		for (i = 0; i < epb; i++) {
346 			err = traverse_dnode(td, &child_dnp[i],
347 			    zb->zb_objset, zb->zb_blkid * epb + i);
348 			if (err != 0)
349 				break;
350 		}
351 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
352 		arc_flags_t flags = ARC_FLAG_WAIT;
353 
354 		err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
355 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
356 		if (err != 0)
357 			goto post;
358 
359 		objset_phys_t *osp = buf->b_data;
360 		prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
361 		    DMU_META_DNODE_OBJECT);
362 		/*
363 		 * See the block comment above for the goal of this variable.
364 		 * If the maxblkid of the meta-dnode is 0, then we know that
365 		 * we've never had more than DNODES_PER_BLOCK objects in the
366 		 * dataset, which means we can't have reused any object ids.
367 		 */
368 		if (osp->os_meta_dnode.dn_maxblkid == 0)
369 			td->td_realloc_possible = B_FALSE;
370 
371 		if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
372 			prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
373 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
374 			prefetch_dnode_metadata(td, &osp->os_userused_dnode,
375 			    zb->zb_objset, DMU_USERUSED_OBJECT);
376 		}
377 
378 		err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
379 		    DMU_META_DNODE_OBJECT);
380 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
381 			err = traverse_dnode(td, &osp->os_groupused_dnode,
382 			    zb->zb_objset, DMU_GROUPUSED_OBJECT);
383 		}
384 		if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
385 			err = traverse_dnode(td, &osp->os_userused_dnode,
386 			    zb->zb_objset, DMU_USERUSED_OBJECT);
387 		}
388 	}
389 
390 	if (buf)
391 		arc_buf_destroy(buf, &buf);
392 
393 post:
394 	if (err == 0 && (td->td_flags & TRAVERSE_POST))
395 		err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
396 
397 	if (hard && (err == EIO || err == ECKSUM)) {
398 		/*
399 		 * Ignore this disk error as requested by the HARD flag,
400 		 * and continue traversal.
401 		 */
402 		err = 0;
403 	}
404 
405 	/*
406 	 * If we are stopping here, set td_resume.
407 	 */
408 	if (td->td_resume != NULL && err != 0 && !td->td_paused) {
409 		td->td_resume->zb_objset = zb->zb_objset;
410 		td->td_resume->zb_object = zb->zb_object;
411 		td->td_resume->zb_level = 0;
412 		/*
413 		 * If we have stopped on an indirect block (e.g. due to
414 		 * i/o error), we have not visited anything below it.
415 		 * Set the bookmark to the first level-0 block that we need
416 		 * to visit.  This way, the resuming code does not need to
417 		 * deal with resuming from indirect blocks.
418 		 *
419 		 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
420 		 * to dereference it.
421 		 */
422 		td->td_resume->zb_blkid = zb->zb_blkid;
423 		if (zb->zb_level > 0) {
424 			td->td_resume->zb_blkid <<= zb->zb_level *
425 			    (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
426 		}
427 		td->td_paused = B_TRUE;
428 	}
429 
430 	return (err);
431 }
432 
433 static void
prefetch_dnode_metadata(traverse_data_t * td,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)434 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
435     uint64_t objset, uint64_t object)
436 {
437 	int j;
438 	zbookmark_phys_t czb;
439 
440 	for (j = 0; j < dnp->dn_nblkptr; j++) {
441 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
442 		traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
443 	}
444 
445 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
446 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
447 		traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
448 	}
449 }
450 
451 static int
traverse_dnode(traverse_data_t * td,const dnode_phys_t * dnp,uint64_t objset,uint64_t object)452 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
453     uint64_t objset, uint64_t object)
454 {
455 	int j, err = 0;
456 	zbookmark_phys_t czb;
457 
458 	if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
459 	    object < td->td_resume->zb_object)
460 		return (0);
461 
462 	if (td->td_flags & TRAVERSE_PRE) {
463 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
464 		    ZB_DNODE_BLKID);
465 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
466 		    td->td_arg);
467 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
468 			return (0);
469 		if (err != 0)
470 			return (err);
471 	}
472 
473 	for (j = 0; j < dnp->dn_nblkptr; j++) {
474 		SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
475 		err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
476 		if (err != 0)
477 			break;
478 	}
479 
480 	if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
481 		SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
482 		err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
483 	}
484 
485 	if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
486 		SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
487 		    ZB_DNODE_BLKID);
488 		err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
489 		    td->td_arg);
490 		if (err == TRAVERSE_VISIT_NO_CHILDREN)
491 			return (0);
492 		if (err != 0)
493 			return (err);
494 	}
495 	return (err);
496 }
497 
498 /* ARGSUSED */
499 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)500 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
501     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
502 {
503 	prefetch_data_t *pfd = arg;
504 	arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
505 
506 	ASSERT(pfd->pd_bytes_fetched >= 0);
507 	if (bp == NULL)
508 		return (0);
509 	if (pfd->pd_cancel)
510 		return (SET_ERROR(EINTR));
511 
512 	if (!prefetch_needed(pfd, bp))
513 		return (0);
514 
515 	mutex_enter(&pfd->pd_mtx);
516 	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
517 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
518 	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
519 	cv_broadcast(&pfd->pd_cv);
520 	mutex_exit(&pfd->pd_mtx);
521 
522 	(void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
523 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
524 
525 	return (0);
526 }
527 
528 static void
traverse_prefetch_thread(void * arg)529 traverse_prefetch_thread(void *arg)
530 {
531 	traverse_data_t *td_main = arg;
532 	traverse_data_t td = *td_main;
533 	zbookmark_phys_t czb;
534 
535 	td.td_func = traverse_prefetcher;
536 	td.td_arg = td_main->td_pfd;
537 	td.td_pfd = NULL;
538 	td.td_resume = &td_main->td_pfd->pd_resume;
539 
540 	SET_BOOKMARK(&czb, td.td_objset,
541 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
542 	(void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
543 
544 	mutex_enter(&td_main->td_pfd->pd_mtx);
545 	td_main->td_pfd->pd_exited = B_TRUE;
546 	cv_broadcast(&td_main->td_pfd->pd_cv);
547 	mutex_exit(&td_main->td_pfd->pd_mtx);
548 }
549 
550 /*
551  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
552  * in syncing context).
553  */
554 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)555 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
556     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
557     blkptr_cb_t func, void *arg)
558 {
559 	traverse_data_t td;
560 	prefetch_data_t pd = { 0 };
561 	zbookmark_phys_t czb;
562 	int err;
563 
564 	ASSERT(ds == NULL || objset == ds->ds_object);
565 	ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
566 
567 	td.td_spa = spa;
568 	td.td_objset = objset;
569 	td.td_rootbp = rootbp;
570 	td.td_min_txg = txg_start;
571 	td.td_resume = resume;
572 	td.td_func = func;
573 	td.td_arg = arg;
574 	td.td_pfd = &pd;
575 	td.td_flags = flags;
576 	td.td_paused = B_FALSE;
577 	td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
578 
579 	if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
580 		VERIFY(spa_feature_enabled_txg(spa,
581 		    SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
582 	} else {
583 		td.td_hole_birth_enabled_txg = UINT64_MAX;
584 	}
585 
586 	pd.pd_flags = flags;
587 	if (resume != NULL)
588 		pd.pd_resume = *resume;
589 	mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
590 	cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
591 
592 	/* See comment on ZIL traversal in dsl_scan_visitds. */
593 	if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
594 		arc_flags_t flags = ARC_FLAG_WAIT;
595 		objset_phys_t *osp;
596 		arc_buf_t *buf;
597 
598 		err = arc_read(NULL, td.td_spa, rootbp,
599 		    arc_getbuf_func, &buf,
600 		    ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
601 		if (err != 0)
602 			return (err);
603 
604 		osp = buf->b_data;
605 		traverse_zil(&td, &osp->os_zil_header);
606 		arc_buf_destroy(buf, &buf);
607 	}
608 
609 	if (!(flags & TRAVERSE_PREFETCH_DATA) ||
610 	    0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
611 	    &td, TQ_NOQUEUE))
612 		pd.pd_exited = B_TRUE;
613 
614 	SET_BOOKMARK(&czb, td.td_objset,
615 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
616 	err = traverse_visitbp(&td, NULL, rootbp, &czb);
617 
618 	mutex_enter(&pd.pd_mtx);
619 	pd.pd_cancel = B_TRUE;
620 	cv_broadcast(&pd.pd_cv);
621 	while (!pd.pd_exited)
622 		cv_wait(&pd.pd_cv, &pd.pd_mtx);
623 	mutex_exit(&pd.pd_mtx);
624 
625 	mutex_destroy(&pd.pd_mtx);
626 	cv_destroy(&pd.pd_cv);
627 
628 	return (err);
629 }
630 
631 /*
632  * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
633  * in syncing context).
634  */
635 int
traverse_dataset_resume(dsl_dataset_t * ds,uint64_t txg_start,zbookmark_phys_t * resume,int flags,blkptr_cb_t func,void * arg)636 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
637     zbookmark_phys_t *resume,
638     int flags, blkptr_cb_t func, void *arg)
639 {
640 	return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
641 	    &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
642 }
643 
644 int
traverse_dataset(dsl_dataset_t * ds,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)645 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
646     int flags, blkptr_cb_t func, void *arg)
647 {
648 	return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
649 }
650 
651 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)652 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
653     uint64_t txg_start, zbookmark_phys_t *resume, int flags,
654     blkptr_cb_t func, void *arg)
655 {
656 	return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
657 	    blkptr, txg_start, resume, flags, func, arg));
658 }
659 
660 /*
661  * NB: pool must not be changing on-disk (eg, from zdb or sync context).
662  */
663 int
traverse_pool(spa_t * spa,uint64_t txg_start,int flags,blkptr_cb_t func,void * arg)664 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
665     blkptr_cb_t func, void *arg)
666 {
667 	int err;
668 	dsl_pool_t *dp = spa_get_dsl(spa);
669 	objset_t *mos = dp->dp_meta_objset;
670 	boolean_t hard = (flags & TRAVERSE_HARD);
671 
672 	/* visit the MOS */
673 	err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
674 	    txg_start, NULL, flags, func, arg);
675 	if (err != 0)
676 		return (err);
677 
678 	/* visit each dataset */
679 	for (uint64_t obj = 1; err == 0;
680 	    err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
681 		dmu_object_info_t doi;
682 
683 		err = dmu_object_info(mos, obj, &doi);
684 		if (err != 0) {
685 			if (hard)
686 				continue;
687 			break;
688 		}
689 
690 		if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
691 			dsl_dataset_t *ds;
692 			uint64_t txg = txg_start;
693 
694 			dsl_pool_config_enter(dp, FTAG);
695 			err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
696 			dsl_pool_config_exit(dp, FTAG);
697 			if (err != 0) {
698 				if (hard)
699 					continue;
700 				break;
701 			}
702 			if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
703 				txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
704 			err = traverse_dataset(ds, txg, flags, func, arg);
705 			dsl_dataset_rele(ds, FTAG);
706 			if (err != 0)
707 				break;
708 		}
709 	}
710 	if (err == ESRCH)
711 		err = 0;
712 	return (err);
713 }
714