]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
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 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
d2734cce | 23 | * Copyright (c) 2012, 2018 by Delphix. All rights reserved. |
34dc7c2f BB |
24 | */ |
25 | ||
34dc7c2f BB |
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> | |
77d8a0f1 | 34 | #include <sys/spa_impl.h> |
34dc7c2f BB |
35 | #include <sys/zio.h> |
36 | #include <sys/dmu_impl.h> | |
428870ff BB |
37 | #include <sys/sa.h> |
38 | #include <sys/sa_impl.h> | |
b128c09f | 39 | #include <sys/callb.h> |
b0bc7a84 | 40 | #include <sys/zfeature.h> |
b128c09f | 41 | |
b738bc5a | 42 | int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */ |
690fe647 | 43 | int32_t send_holes_without_birth_time = 1; |
572e2857 BB |
44 | |
45 | typedef struct prefetch_data { | |
b128c09f BB |
46 | kmutex_t pd_mtx; |
47 | kcondvar_t pd_cv; | |
b738bc5a | 48 | int32_t pd_bytes_fetched; |
b128c09f BB |
49 | int pd_flags; |
50 | boolean_t pd_cancel; | |
51 | boolean_t pd_exited; | |
47dfff3b | 52 | zbookmark_phys_t pd_resume; |
572e2857 | 53 | } prefetch_data_t; |
b128c09f | 54 | |
572e2857 | 55 | typedef struct traverse_data { |
b128c09f BB |
56 | spa_t *td_spa; |
57 | uint64_t td_objset; | |
58 | blkptr_t *td_rootbp; | |
59 | uint64_t td_min_txg; | |
5dbd68a3 | 60 | zbookmark_phys_t *td_resume; |
b128c09f | 61 | int td_flags; |
572e2857 | 62 | prefetch_data_t *td_pfd; |
fbeddd60 | 63 | boolean_t td_paused; |
e5fd1dd6 | 64 | uint64_t td_hole_birth_enabled_txg; |
b128c09f BB |
65 | blkptr_cb_t *td_func; |
66 | void *td_arg; | |
c352ec27 | 67 | boolean_t td_realloc_possible; |
572e2857 | 68 | } traverse_data_t; |
34dc7c2f | 69 | |
572e2857 | 70 | static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 71 | uint64_t objset, uint64_t object); |
96b89346 | 72 | static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, |
294f6806 | 73 | uint64_t objset, uint64_t object); |
9babb374 | 74 | |
428870ff | 75 | static int |
34dc7c2f BB |
76 | traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) |
77 | { | |
572e2857 | 78 | traverse_data_t *td = arg; |
5dbd68a3 | 79 | zbookmark_phys_t zb; |
34dc7c2f | 80 | |
b0bc7a84 | 81 | if (BP_IS_HOLE(bp)) |
428870ff | 82 | return (0); |
34dc7c2f | 83 | |
d2734cce SD |
84 | if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa)) |
85 | return (-1); | |
428870ff BB |
86 | |
87 | SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
88 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
89 | ||
294f6806 | 90 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); |
b128c09f | 91 | |
428870ff | 92 | return (0); |
34dc7c2f BB |
93 | } |
94 | ||
428870ff | 95 | static int |
34dc7c2f BB |
96 | traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) |
97 | { | |
572e2857 | 98 | traverse_data_t *td = arg; |
34dc7c2f BB |
99 | |
100 | if (lrc->lrc_txtype == TX_WRITE) { | |
101 | lr_write_t *lr = (lr_write_t *)lrc; | |
102 | blkptr_t *bp = &lr->lr_blkptr; | |
5dbd68a3 | 103 | zbookmark_phys_t zb; |
34dc7c2f | 104 | |
b0bc7a84 | 105 | if (BP_IS_HOLE(bp)) |
428870ff | 106 | return (0); |
34dc7c2f | 107 | |
b128c09f | 108 | if (claim_txg == 0 || bp->blk_birth < claim_txg) |
428870ff | 109 | return (0); |
b128c09f | 110 | |
572e2857 BB |
111 | SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, |
112 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
428870ff | 113 | |
294f6806 | 114 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, |
428870ff | 115 | td->td_arg); |
34dc7c2f | 116 | } |
428870ff | 117 | return (0); |
34dc7c2f BB |
118 | } |
119 | ||
120 | static void | |
572e2857 | 121 | traverse_zil(traverse_data_t *td, zil_header_t *zh) |
34dc7c2f | 122 | { |
34dc7c2f | 123 | uint64_t claim_txg = zh->zh_claim_txg; |
34dc7c2f | 124 | |
34dc7c2f BB |
125 | /* |
126 | * We only want to visit blocks that have been claimed but not yet | |
d2734cce | 127 | * replayed; plus blocks that are already stable in read-only mode. |
34dc7c2f | 128 | */ |
fb5f0bc8 | 129 | if (claim_txg == 0 && spa_writeable(td->td_spa)) |
34dc7c2f BB |
130 | return; |
131 | ||
d2734cce | 132 | zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh); |
b128c09f | 133 | (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td, |
b5256303 | 134 | claim_txg, !(td->td_flags & TRAVERSE_NO_DECRYPT)); |
34dc7c2f BB |
135 | zil_free(zilog); |
136 | } | |
137 | ||
9ae529ec CS |
138 | typedef enum resume_skip { |
139 | RESUME_SKIP_ALL, | |
140 | RESUME_SKIP_NONE, | |
141 | RESUME_SKIP_CHILDREN | |
142 | } resume_skip_t; | |
143 | ||
144 | /* | |
145 | * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and | |
146 | * the block indicated by zb does not need to be visited at all. Returns | |
147 | * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the | |
148 | * resume point. This indicates that this block should be visited but not its | |
149 | * children (since they must have been visited in a previous traversal). | |
150 | * Otherwise returns RESUME_SKIP_NONE. | |
151 | */ | |
152 | static resume_skip_t | |
153 | resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, | |
5dbd68a3 | 154 | const zbookmark_phys_t *zb) |
9ae529ec CS |
155 | { |
156 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { | |
157 | /* | |
158 | * If we already visited this bp & everything below, | |
159 | * don't bother doing it again. | |
160 | */ | |
fcff0f35 | 161 | if (zbookmark_subtree_completed(dnp, zb, td->td_resume)) |
9ae529ec CS |
162 | return (RESUME_SKIP_ALL); |
163 | ||
164 | /* | |
165 | * If we found the block we're trying to resume from, zero | |
166 | * the bookmark out to indicate that we have resumed. | |
167 | */ | |
9ae529ec CS |
168 | if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) { |
169 | bzero(td->td_resume, sizeof (*zb)); | |
170 | if (td->td_flags & TRAVERSE_POST) | |
171 | return (RESUME_SKIP_CHILDREN); | |
172 | } | |
173 | } | |
174 | return (RESUME_SKIP_NONE); | |
175 | } | |
176 | ||
96b89346 MA |
177 | static void |
178 | traverse_prefetch_metadata(traverse_data_t *td, | |
5dbd68a3 | 179 | const blkptr_t *bp, const zbookmark_phys_t *zb) |
96b89346 | 180 | { |
2a432414 | 181 | arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; |
440a3eb9 | 182 | int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; |
96b89346 MA |
183 | |
184 | if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) | |
185 | return; | |
186 | /* | |
187 | * If we are in the process of resuming, don't prefetch, because | |
188 | * some children will not be needed (and in fact may have already | |
189 | * been freed). | |
190 | */ | |
191 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) | |
192 | return; | |
193 | if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) | |
194 | return; | |
195 | if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) | |
196 | return; | |
197 | ||
b5256303 TC |
198 | if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) |
199 | zio_flags |= ZIO_FLAG_RAW; | |
200 | ||
294f6806 | 201 | (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, |
b5256303 | 202 | ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); |
96b89346 MA |
203 | } |
204 | ||
1fa8f795 MA |
205 | static boolean_t |
206 | prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp) | |
207 | { | |
208 | ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA); | |
209 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || | |
210 | BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) | |
211 | return (B_FALSE); | |
212 | return (B_TRUE); | |
213 | } | |
214 | ||
c7f8f831 BB |
215 | static int |
216 | traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, | |
5dbd68a3 | 217 | const blkptr_t *bp, const zbookmark_phys_t *zb) |
6656bf56 | 218 | { |
fbeddd60 | 219 | int err = 0; |
c7f8f831 | 220 | arc_buf_t *buf = NULL; |
ee2f17aa | 221 | prefetch_data_t *pd = td->td_pfd; |
9ae529ec CS |
222 | |
223 | switch (resume_skip_check(td, dnp, zb)) { | |
224 | case RESUME_SKIP_ALL: | |
225 | return (0); | |
226 | case RESUME_SKIP_CHILDREN: | |
227 | goto post; | |
228 | case RESUME_SKIP_NONE: | |
229 | break; | |
230 | default: | |
231 | ASSERT(0); | |
232 | } | |
6656bf56 | 233 | |
b0bc7a84 | 234 | if (bp->blk_birth == 0) { |
e5fd1dd6 | 235 | /* |
c352ec27 PD |
236 | * Since this block has a birth time of 0 it must be one of |
237 | * two things: a hole created before the | |
238 | * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole | |
239 | * which has always been a hole in an object. | |
240 | * | |
241 | * If a file is written sparsely, then the unwritten parts of | |
242 | * the file were "always holes" -- that is, they have been | |
243 | * holes since this object was allocated. However, we (and | |
244 | * our callers) can not necessarily tell when an object was | |
245 | * allocated. Therefore, if it's possible that this object | |
246 | * was freed and then its object number reused, we need to | |
247 | * visit all the holes with birth==0. | |
248 | * | |
249 | * If it isn't possible that the object number was reused, | |
250 | * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote | |
251 | * all the blocks we will visit as part of this traversal, | |
252 | * then this hole must have always existed, so we can skip | |
253 | * it. We visit blocks born after (exclusive) td_min_txg. | |
254 | * | |
255 | * Note that the meta-dnode cannot be reallocated. | |
e5fd1dd6 | 256 | */ |
690fe647 BB |
257 | if (!send_holes_without_birth_time && |
258 | (!td->td_realloc_possible || | |
259 | zb->zb_object == DMU_META_DNODE_OBJECT) && | |
260 | td->td_hole_birth_enabled_txg <= td->td_min_txg) | |
e5fd1dd6 | 261 | return (0); |
b0bc7a84 MG |
262 | } else if (bp->blk_birth <= td->td_min_txg) { |
263 | return (0); | |
264 | } | |
265 | ||
ee2f17aa | 266 | if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) { |
b738bc5a | 267 | uint64_t size = BP_GET_LSIZE(bp); |
ee2f17aa | 268 | mutex_enter(&pd->pd_mtx); |
b738bc5a GW |
269 | ASSERT(pd->pd_bytes_fetched >= 0); |
270 | while (pd->pd_bytes_fetched < size && !pd->pd_exited) | |
8e70975f | 271 | cv_wait_sig(&pd->pd_cv, &pd->pd_mtx); |
b738bc5a | 272 | pd->pd_bytes_fetched -= size; |
ee2f17aa CD |
273 | cv_broadcast(&pd->pd_cv); |
274 | mutex_exit(&pd->pd_mtx); | |
34dc7c2f BB |
275 | } |
276 | ||
1fa8f795 MA |
277 | if (BP_IS_HOLE(bp)) { |
278 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); | |
279 | if (err != 0) | |
280 | goto post; | |
281 | return (0); | |
282 | } | |
283 | ||
c7f8f831 | 284 | if (td->td_flags & TRAVERSE_PRE) { |
294f6806 | 285 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, |
c7f8f831 BB |
286 | td->td_arg); |
287 | if (err == TRAVERSE_VISIT_NO_CHILDREN) | |
572e2857 | 288 | return (0); |
9ae529ec CS |
289 | if (err != 0) |
290 | goto post; | |
34dc7c2f BB |
291 | } |
292 | ||
c7f8f831 | 293 | if (BP_GET_LEVEL(bp) > 0) { |
2a432414 | 294 | uint32_t flags = ARC_FLAG_WAIT; |
a1687880 BB |
295 | int32_t i; |
296 | int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; | |
5dbd68a3 | 297 | zbookmark_phys_t *czb; |
b128c09f | 298 | |
b5256303 TC |
299 | ASSERT(!BP_IS_PROTECTED(bp)); |
300 | ||
294f6806 | 301 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 302 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 303 | if (err != 0) |
fbeddd60 | 304 | goto post; |
a1687880 | 305 | |
79c76d5b | 306 | czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); |
96b89346 MA |
307 | |
308 | for (i = 0; i < epb; i++) { | |
a1687880 | 309 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
96b89346 MA |
310 | zb->zb_level - 1, |
311 | zb->zb_blkid * epb + i); | |
a1687880 BB |
312 | traverse_prefetch_metadata(td, |
313 | &((blkptr_t *)buf->b_data)[i], czb); | |
96b89346 | 314 | } |
b128c09f BB |
315 | |
316 | /* recursively visitbp() blocks below this */ | |
96b89346 | 317 | for (i = 0; i < epb; i++) { |
a1687880 | 318 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
c7f8f831 BB |
319 | zb->zb_level - 1, |
320 | zb->zb_blkid * epb + i); | |
a1687880 BB |
321 | err = traverse_visitbp(td, dnp, |
322 | &((blkptr_t *)buf->b_data)[i], czb); | |
fbeddd60 MA |
323 | if (err != 0) |
324 | break; | |
b128c09f | 325 | } |
a1687880 | 326 | |
5dbd68a3 | 327 | kmem_free(czb, sizeof (zbookmark_phys_t)); |
a1687880 | 328 | |
c7f8f831 | 329 | } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { |
2a432414 | 330 | uint32_t flags = ARC_FLAG_WAIT; |
b5256303 | 331 | uint32_t zio_flags = ZIO_FLAG_CANFAIL; |
a1687880 BB |
332 | int32_t i; |
333 | int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; | |
47dfff3b | 334 | dnode_phys_t *child_dnp; |
b128c09f | 335 | |
b5256303 TC |
336 | /* |
337 | * dnode blocks might have their bonus buffers encrypted, so | |
338 | * we must be careful to honor TRAVERSE_NO_DECRYPT | |
339 | */ | |
340 | if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) | |
341 | zio_flags |= ZIO_FLAG_RAW; | |
342 | ||
294f6806 | 343 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
b5256303 | 344 | ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); |
13fe0198 | 345 | if (err != 0) |
fbeddd60 | 346 | goto post; |
b5256303 | 347 | |
47dfff3b | 348 | child_dnp = buf->b_data; |
96b89346 | 349 | |
47dfff3b MA |
350 | for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) { |
351 | prefetch_dnode_metadata(td, &child_dnp[i], | |
352 | zb->zb_objset, zb->zb_blkid * epb + i); | |
96b89346 | 353 | } |
b128c09f BB |
354 | |
355 | /* recursively visitbp() blocks below this */ | |
47dfff3b MA |
356 | for (i = 0; i < epb; i += child_dnp[i].dn_extra_slots + 1) { |
357 | err = traverse_dnode(td, &child_dnp[i], | |
358 | zb->zb_objset, zb->zb_blkid * epb + i); | |
fbeddd60 MA |
359 | if (err != 0) |
360 | break; | |
34dc7c2f | 361 | } |
c7f8f831 | 362 | } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { |
b5256303 | 363 | uint32_t zio_flags = ZIO_FLAG_CANFAIL; |
2a432414 | 364 | arc_flags_t flags = ARC_FLAG_WAIT; |
c7f8f831 | 365 | objset_phys_t *osp; |
c7f8f831 | 366 | |
b5256303 TC |
367 | if ((td->td_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) |
368 | zio_flags |= ZIO_FLAG_RAW; | |
369 | ||
294f6806 | 370 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
b5256303 | 371 | ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); |
13fe0198 | 372 | if (err != 0) |
fbeddd60 | 373 | goto post; |
c7f8f831 BB |
374 | |
375 | osp = buf->b_data; | |
47dfff3b | 376 | prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset, |
96b89346 | 377 | DMU_META_DNODE_OBJECT); |
c352ec27 PD |
378 | /* |
379 | * See the block comment above for the goal of this variable. | |
380 | * If the maxblkid of the meta-dnode is 0, then we know that | |
381 | * we've never had more than DNODES_PER_BLOCK objects in the | |
382 | * dataset, which means we can't have reused any object ids. | |
383 | */ | |
384 | if (osp->os_meta_dnode.dn_maxblkid == 0) | |
385 | td->td_realloc_possible = B_FALSE; | |
386 | ||
9c5167d1 NF |
387 | if (OBJSET_BUF_HAS_USERUSED(buf)) { |
388 | if (OBJSET_BUF_HAS_PROJECTUSED(buf)) | |
389 | prefetch_dnode_metadata(td, | |
390 | &osp->os_projectused_dnode, | |
391 | zb->zb_objset, DMU_PROJECTUSED_OBJECT); | |
47dfff3b MA |
392 | prefetch_dnode_metadata(td, &osp->os_groupused_dnode, |
393 | zb->zb_objset, DMU_GROUPUSED_OBJECT); | |
394 | prefetch_dnode_metadata(td, &osp->os_userused_dnode, | |
395 | zb->zb_objset, DMU_USERUSED_OBJECT); | |
96b89346 MA |
396 | } |
397 | ||
47dfff3b | 398 | err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset, |
c7f8f831 | 399 | DMU_META_DNODE_OBJECT); |
9c5167d1 NF |
400 | if (err == 0 && OBJSET_BUF_HAS_USERUSED(buf)) { |
401 | if (OBJSET_BUF_HAS_PROJECTUSED(buf)) | |
402 | err = traverse_dnode(td, | |
403 | &osp->os_projectused_dnode, zb->zb_objset, | |
404 | DMU_PROJECTUSED_OBJECT); | |
405 | if (err == 0) | |
406 | err = traverse_dnode(td, | |
407 | &osp->os_groupused_dnode, zb->zb_objset, | |
408 | DMU_GROUPUSED_OBJECT); | |
409 | if (err == 0) | |
410 | err = traverse_dnode(td, | |
411 | &osp->os_userused_dnode, zb->zb_objset, | |
412 | DMU_USERUSED_OBJECT); | |
34dc7c2f | 413 | } |
34dc7c2f BB |
414 | } |
415 | ||
c7f8f831 | 416 | if (buf) |
d3c2ae1c | 417 | arc_buf_destroy(buf, &buf); |
34dc7c2f | 418 | |
9ae529ec | 419 | post: |
fbeddd60 | 420 | if (err == 0 && (td->td_flags & TRAVERSE_POST)) |
294f6806 | 421 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); |
fbeddd60 MA |
422 | |
423 | if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) { | |
424 | /* | |
425 | * Ignore this disk error as requested by the HARD flag, | |
426 | * and continue traversal. | |
427 | */ | |
428 | err = 0; | |
9ae529ec CS |
429 | } |
430 | ||
fbeddd60 MA |
431 | /* |
432 | * If we are stopping here, set td_resume. | |
433 | */ | |
434 | if (td->td_resume != NULL && err != 0 && !td->td_paused) { | |
435 | td->td_resume->zb_objset = zb->zb_objset; | |
436 | td->td_resume->zb_object = zb->zb_object; | |
437 | td->td_resume->zb_level = 0; | |
438 | /* | |
439 | * If we have stopped on an indirect block (e.g. due to | |
440 | * i/o error), we have not visited anything below it. | |
441 | * Set the bookmark to the first level-0 block that we need | |
442 | * to visit. This way, the resuming code does not need to | |
443 | * deal with resuming from indirect blocks. | |
47dfff3b MA |
444 | * |
445 | * Note, if zb_level <= 0, dnp may be NULL, so we don't want | |
446 | * to dereference it. | |
fbeddd60 | 447 | */ |
47dfff3b MA |
448 | td->td_resume->zb_blkid = zb->zb_blkid; |
449 | if (zb->zb_level > 0) { | |
450 | td->td_resume->zb_blkid <<= zb->zb_level * | |
451 | (dnp->dn_indblkshift - SPA_BLKPTRSHIFT); | |
452 | } | |
fbeddd60 | 453 | td->td_paused = B_TRUE; |
428870ff | 454 | } |
34dc7c2f | 455 | |
fbeddd60 | 456 | return (err); |
34dc7c2f BB |
457 | } |
458 | ||
96b89346 MA |
459 | static void |
460 | prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 461 | uint64_t objset, uint64_t object) |
96b89346 MA |
462 | { |
463 | int j; | |
5dbd68a3 | 464 | zbookmark_phys_t czb; |
96b89346 MA |
465 | |
466 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
467 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 468 | traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); |
96b89346 MA |
469 | } |
470 | ||
471 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
472 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); | |
50c957f7 | 473 | traverse_prefetch_metadata(td, DN_SPILL_BLKPTR(dnp), &czb); |
96b89346 MA |
474 | } |
475 | } | |
476 | ||
9babb374 | 477 | static int |
572e2857 | 478 | traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 479 | uint64_t objset, uint64_t object) |
9babb374 | 480 | { |
fbeddd60 | 481 | int j, err = 0; |
5dbd68a3 | 482 | zbookmark_phys_t czb; |
9babb374 | 483 | |
47dfff3b MA |
484 | if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL && |
485 | object < td->td_resume->zb_object) | |
486 | return (0); | |
487 | ||
fcff0f35 PD |
488 | if (td->td_flags & TRAVERSE_PRE) { |
489 | SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL, | |
490 | ZB_DNODE_BLKID); | |
491 | err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp, | |
492 | td->td_arg); | |
493 | if (err == TRAVERSE_VISIT_NO_CHILDREN) | |
494 | return (0); | |
495 | if (err != 0) | |
496 | return (err); | |
497 | } | |
498 | ||
9babb374 BB |
499 | for (j = 0; j < dnp->dn_nblkptr; j++) { |
500 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 501 | err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); |
fbeddd60 MA |
502 | if (err != 0) |
503 | break; | |
9babb374 | 504 | } |
428870ff | 505 | |
fcff0f35 | 506 | if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { |
96b89346 | 507 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); |
50c957f7 | 508 | err = traverse_visitbp(td, dnp, DN_SPILL_BLKPTR(dnp), &czb); |
428870ff | 509 | } |
fcff0f35 PD |
510 | |
511 | if (err == 0 && (td->td_flags & TRAVERSE_POST)) { | |
512 | SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL, | |
513 | ZB_DNODE_BLKID); | |
514 | err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp, | |
515 | td->td_arg); | |
516 | if (err == TRAVERSE_VISIT_NO_CHILDREN) | |
517 | return (0); | |
518 | if (err != 0) | |
519 | return (err); | |
520 | } | |
fbeddd60 | 521 | return (err); |
9babb374 BB |
522 | } |
523 | ||
b128c09f BB |
524 | /* ARGSUSED */ |
525 | static int | |
428870ff | 526 | traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, |
5dbd68a3 | 527 | const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) |
34dc7c2f | 528 | { |
572e2857 | 529 | prefetch_data_t *pfd = arg; |
b5256303 | 530 | int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; |
d4a72f23 TC |
531 | arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH | |
532 | ARC_FLAG_PRESCIENT_PREFETCH; | |
34dc7c2f | 533 | |
b738bc5a | 534 | ASSERT(pfd->pd_bytes_fetched >= 0); |
fcff0f35 PD |
535 | if (bp == NULL) |
536 | return (0); | |
b128c09f | 537 | if (pfd->pd_cancel) |
2e528b49 | 538 | return (SET_ERROR(EINTR)); |
34dc7c2f | 539 | |
1fa8f795 | 540 | if (!prefetch_needed(pfd, bp)) |
34dc7c2f BB |
541 | return (0); |
542 | ||
b128c09f | 543 | mutex_enter(&pfd->pd_mtx); |
b738bc5a | 544 | while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max) |
8e70975f | 545 | cv_wait_sig(&pfd->pd_cv, &pfd->pd_mtx); |
b738bc5a | 546 | pfd->pd_bytes_fetched += BP_GET_LSIZE(bp); |
b128c09f BB |
547 | cv_broadcast(&pfd->pd_cv); |
548 | mutex_exit(&pfd->pd_mtx); | |
34dc7c2f | 549 | |
b5256303 TC |
550 | if ((pfd->pd_flags & TRAVERSE_NO_DECRYPT) && BP_IS_PROTECTED(bp)) |
551 | zio_flags |= ZIO_FLAG_RAW; | |
552 | ||
294f6806 | 553 | (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, |
b5256303 | 554 | zio_flags, &aflags, zb); |
34dc7c2f | 555 | |
b128c09f | 556 | return (0); |
34dc7c2f BB |
557 | } |
558 | ||
34dc7c2f | 559 | static void |
b128c09f | 560 | traverse_prefetch_thread(void *arg) |
34dc7c2f | 561 | { |
572e2857 BB |
562 | traverse_data_t *td_main = arg; |
563 | traverse_data_t td = *td_main; | |
5dbd68a3 | 564 | zbookmark_phys_t czb; |
dca8c34d | 565 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
34dc7c2f | 566 | |
b128c09f BB |
567 | td.td_func = traverse_prefetcher; |
568 | td.td_arg = td_main->td_pfd; | |
569 | td.td_pfd = NULL; | |
47dfff3b | 570 | td.td_resume = &td_main->td_pfd->pd_resume; |
34dc7c2f | 571 | |
428870ff BB |
572 | SET_BOOKMARK(&czb, td.td_objset, |
573 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
294f6806 | 574 | (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); |
34dc7c2f | 575 | |
b128c09f BB |
576 | mutex_enter(&td_main->td_pfd->pd_mtx); |
577 | td_main->td_pfd->pd_exited = B_TRUE; | |
578 | cv_broadcast(&td_main->td_pfd->pd_cv); | |
579 | mutex_exit(&td_main->td_pfd->pd_mtx); | |
dca8c34d | 580 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
581 | } |
582 | ||
b128c09f BB |
583 | /* |
584 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
585 | * in syncing context). | |
586 | */ | |
587 | static int | |
9ae529ec | 588 | traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, |
5dbd68a3 | 589 | uint64_t txg_start, zbookmark_phys_t *resume, int flags, |
9ae529ec | 590 | blkptr_cb_t func, void *arg) |
34dc7c2f | 591 | { |
47050a88 BB |
592 | traverse_data_t *td; |
593 | prefetch_data_t *pd; | |
5dbd68a3 | 594 | zbookmark_phys_t *czb; |
b128c09f | 595 | int err; |
34dc7c2f | 596 | |
9ae529ec CS |
597 | ASSERT(ds == NULL || objset == ds->ds_object); |
598 | ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); | |
599 | ||
79c76d5b BB |
600 | td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP); |
601 | pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP); | |
602 | czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); | |
47050a88 BB |
603 | |
604 | td->td_spa = spa; | |
9ae529ec | 605 | td->td_objset = objset; |
47050a88 BB |
606 | td->td_rootbp = rootbp; |
607 | td->td_min_txg = txg_start; | |
9ae529ec | 608 | td->td_resume = resume; |
47050a88 BB |
609 | td->td_func = func; |
610 | td->td_arg = arg; | |
611 | td->td_pfd = pd; | |
612 | td->td_flags = flags; | |
fbeddd60 | 613 | td->td_paused = B_FALSE; |
c352ec27 | 614 | td->td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE); |
b128c09f | 615 | |
e5fd1dd6 MA |
616 | if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { |
617 | VERIFY(spa_feature_enabled_txg(spa, | |
618 | SPA_FEATURE_HOLE_BIRTH, &td->td_hole_birth_enabled_txg)); | |
619 | } else { | |
c352ec27 | 620 | td->td_hole_birth_enabled_txg = UINT64_MAX; |
e5fd1dd6 MA |
621 | } |
622 | ||
47050a88 | 623 | pd->pd_flags = flags; |
47dfff3b MA |
624 | if (resume != NULL) |
625 | pd->pd_resume = *resume; | |
47050a88 BB |
626 | mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); |
627 | cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); | |
b128c09f | 628 | |
1421c891 PS |
629 | SET_BOOKMARK(czb, td->td_objset, |
630 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
631 | ||
572e2857 | 632 | /* See comment on ZIL traversal in dsl_scan_visitds. */ |
0c66c32d | 633 | if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) { |
b5256303 | 634 | enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; |
2a432414 | 635 | uint32_t flags = ARC_FLAG_WAIT; |
13fe0198 MA |
636 | objset_phys_t *osp; |
637 | arc_buf_t *buf; | |
572e2857 | 638 | |
b5256303 TC |
639 | if ((td->td_flags & TRAVERSE_NO_DECRYPT) && |
640 | BP_IS_PROTECTED(rootbp)) | |
641 | zio_flags |= ZIO_FLAG_RAW; | |
642 | ||
643 | err = arc_read(NULL, td->td_spa, rootbp, arc_getbuf_func, | |
644 | &buf, ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, czb); | |
d3190c5f CC |
645 | if (err != 0) { |
646 | /* | |
647 | * If both TRAVERSE_HARD and TRAVERSE_PRE are set, | |
648 | * continue to visitbp so that td_func can be called | |
649 | * in pre stage, and err will reset to zero. | |
650 | */ | |
651 | if (!(td->td_flags & TRAVERSE_HARD) || | |
652 | !(td->td_flags & TRAVERSE_PRE)) | |
4338c5c0 | 653 | goto out; |
d3190c5f CC |
654 | } else { |
655 | osp = buf->b_data; | |
656 | traverse_zil(td, &osp->os_zil_header); | |
657 | arc_buf_destroy(buf, &buf); | |
658 | } | |
572e2857 BB |
659 | } |
660 | ||
96b89346 | 661 | if (!(flags & TRAVERSE_PREFETCH_DATA) || |
77d8a0f1 | 662 | taskq_dispatch(spa->spa_prefetch_taskq, traverse_prefetch_thread, |
48d3eb40 | 663 | td, TQ_NOQUEUE) == TASKQID_INVALID) |
47050a88 | 664 | pd->pd_exited = B_TRUE; |
b128c09f | 665 | |
294f6806 | 666 | err = traverse_visitbp(td, NULL, rootbp, czb); |
47050a88 BB |
667 | |
668 | mutex_enter(&pd->pd_mtx); | |
669 | pd->pd_cancel = B_TRUE; | |
670 | cv_broadcast(&pd->pd_cv); | |
671 | while (!pd->pd_exited) | |
8e70975f | 672 | cv_wait_sig(&pd->pd_cv, &pd->pd_mtx); |
47050a88 | 673 | mutex_exit(&pd->pd_mtx); |
4338c5c0 | 674 | out: |
47050a88 BB |
675 | mutex_destroy(&pd->pd_mtx); |
676 | cv_destroy(&pd->pd_cv); | |
b128c09f | 677 | |
5dbd68a3 | 678 | kmem_free(czb, sizeof (zbookmark_phys_t)); |
a1687880 BB |
679 | kmem_free(pd, sizeof (struct prefetch_data)); |
680 | kmem_free(td, sizeof (struct traverse_data)); | |
34dc7c2f | 681 | |
b128c09f | 682 | return (err); |
34dc7c2f BB |
683 | } |
684 | ||
b128c09f BB |
685 | /* |
686 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
687 | * in syncing context). | |
688 | */ | |
689 | int | |
47dfff3b MA |
690 | traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start, |
691 | zbookmark_phys_t *resume, | |
692 | int flags, blkptr_cb_t func, void *arg) | |
34dc7c2f | 693 | { |
9ae529ec | 694 | return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, |
47dfff3b MA |
695 | &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg)); |
696 | } | |
697 | ||
698 | int | |
699 | traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, | |
700 | int flags, blkptr_cb_t func, void *arg) | |
701 | { | |
702 | return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg)); | |
9ae529ec CS |
703 | } |
704 | ||
705 | int | |
706 | traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, | |
5dbd68a3 | 707 | uint64_t txg_start, zbookmark_phys_t *resume, int flags, |
9ae529ec CS |
708 | blkptr_cb_t func, void *arg) |
709 | { | |
710 | return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, | |
711 | blkptr, txg_start, resume, flags, func, arg)); | |
34dc7c2f BB |
712 | } |
713 | ||
b128c09f BB |
714 | /* |
715 | * NB: pool must not be changing on-disk (eg, from zdb or sync context). | |
716 | */ | |
717 | int | |
428870ff BB |
718 | traverse_pool(spa_t *spa, uint64_t txg_start, int flags, |
719 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 720 | { |
fbeddd60 | 721 | int err; |
b128c09f BB |
722 | dsl_pool_t *dp = spa_get_dsl(spa); |
723 | objset_t *mos = dp->dp_meta_objset; | |
428870ff | 724 | boolean_t hard = (flags & TRAVERSE_HARD); |
b128c09f BB |
725 | |
726 | /* visit the MOS */ | |
9ae529ec CS |
727 | err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), |
728 | txg_start, NULL, flags, func, arg); | |
13fe0198 | 729 | if (err != 0) |
b128c09f BB |
730 | return (err); |
731 | ||
732 | /* visit each dataset */ | |
1c27024e | 733 | for (uint64_t obj = 1; err == 0; |
47dfff3b | 734 | err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) { |
b128c09f BB |
735 | dmu_object_info_t doi; |
736 | ||
737 | err = dmu_object_info(mos, obj, &doi); | |
13fe0198 | 738 | if (err != 0) { |
fbeddd60 MA |
739 | if (hard) |
740 | continue; | |
741 | break; | |
428870ff | 742 | } |
b128c09f | 743 | |
fa86b5db | 744 | if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) { |
b128c09f | 745 | dsl_dataset_t *ds; |
428870ff BB |
746 | uint64_t txg = txg_start; |
747 | ||
13fe0198 | 748 | dsl_pool_config_enter(dp, FTAG); |
b128c09f | 749 | err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds); |
13fe0198 MA |
750 | dsl_pool_config_exit(dp, FTAG); |
751 | if (err != 0) { | |
fbeddd60 MA |
752 | if (hard) |
753 | continue; | |
754 | break; | |
428870ff | 755 | } |
d683ddbb JG |
756 | if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg) |
757 | txg = dsl_dataset_phys(ds)->ds_prev_snap_txg; | |
428870ff | 758 | err = traverse_dataset(ds, txg, flags, func, arg); |
b128c09f | 759 | dsl_dataset_rele(ds, FTAG); |
fbeddd60 MA |
760 | if (err != 0) |
761 | break; | |
b128c09f | 762 | } |
34dc7c2f | 763 | } |
b128c09f BB |
764 | if (err == ESRCH) |
765 | err = 0; | |
fbeddd60 | 766 | return (err); |
34dc7c2f | 767 | } |
c28b2279 | 768 | |
93ce2b4c | 769 | #if defined(_KERNEL) |
c28b2279 BB |
770 | EXPORT_SYMBOL(traverse_dataset); |
771 | EXPORT_SYMBOL(traverse_pool); | |
c409e464 | 772 | |
b738bc5a GW |
773 | module_param(zfs_pd_bytes_max, int, 0644); |
774 | MODULE_PARM_DESC(zfs_pd_bytes_max, "Max number of bytes to prefetch"); | |
6d836e6f | 775 | |
690fe647 BB |
776 | module_param_named(ignore_hole_birth, send_holes_without_birth_time, int, 0644); |
777 | MODULE_PARM_DESC(ignore_hole_birth, "Alias for send_holes_without_birth_time"); | |
778 | ||
779 | module_param_named(send_holes_without_birth_time, | |
780 | send_holes_without_birth_time, int, 0644); | |
781 | MODULE_PARM_DESC(send_holes_without_birth_time, | |
782 | "Ignore hole_birth txg for zfs send"); | |
c28b2279 | 783 | #endif |