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