]>
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. |
1fa8f795 | 23 | * Copyright (c) 2012, 2014 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> | |
34 | #include <sys/zio.h> | |
35 | #include <sys/dmu_impl.h> | |
428870ff BB |
36 | #include <sys/sa.h> |
37 | #include <sys/sa_impl.h> | |
b128c09f | 38 | #include <sys/callb.h> |
b0bc7a84 | 39 | #include <sys/zfeature.h> |
b128c09f | 40 | |
b738bc5a | 41 | int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */ |
572e2857 BB |
42 | |
43 | typedef struct prefetch_data { | |
b128c09f BB |
44 | kmutex_t pd_mtx; |
45 | kcondvar_t pd_cv; | |
b738bc5a | 46 | int32_t pd_bytes_fetched; |
b128c09f BB |
47 | int pd_flags; |
48 | boolean_t pd_cancel; | |
49 | boolean_t pd_exited; | |
572e2857 | 50 | } prefetch_data_t; |
b128c09f | 51 | |
572e2857 | 52 | typedef struct traverse_data { |
b128c09f BB |
53 | spa_t *td_spa; |
54 | uint64_t td_objset; | |
55 | blkptr_t *td_rootbp; | |
56 | uint64_t td_min_txg; | |
5dbd68a3 | 57 | zbookmark_phys_t *td_resume; |
b128c09f | 58 | int td_flags; |
572e2857 | 59 | prefetch_data_t *td_pfd; |
fbeddd60 | 60 | boolean_t td_paused; |
b128c09f BB |
61 | blkptr_cb_t *td_func; |
62 | void *td_arg; | |
572e2857 | 63 | } traverse_data_t; |
34dc7c2f | 64 | |
572e2857 | 65 | static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 66 | uint64_t objset, uint64_t object); |
96b89346 | 67 | static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, |
294f6806 | 68 | uint64_t objset, uint64_t object); |
9babb374 | 69 | |
428870ff | 70 | static int |
34dc7c2f BB |
71 | traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) |
72 | { | |
572e2857 | 73 | traverse_data_t *td = arg; |
5dbd68a3 | 74 | zbookmark_phys_t zb; |
34dc7c2f | 75 | |
b0bc7a84 | 76 | if (BP_IS_HOLE(bp)) |
428870ff | 77 | return (0); |
34dc7c2f | 78 | |
b128c09f | 79 | if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa)) |
428870ff BB |
80 | return (0); |
81 | ||
82 | SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
83 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
84 | ||
294f6806 | 85 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); |
b128c09f | 86 | |
428870ff | 87 | return (0); |
34dc7c2f BB |
88 | } |
89 | ||
428870ff | 90 | static int |
34dc7c2f BB |
91 | traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) |
92 | { | |
572e2857 | 93 | traverse_data_t *td = arg; |
34dc7c2f BB |
94 | |
95 | if (lrc->lrc_txtype == TX_WRITE) { | |
96 | lr_write_t *lr = (lr_write_t *)lrc; | |
97 | blkptr_t *bp = &lr->lr_blkptr; | |
5dbd68a3 | 98 | zbookmark_phys_t zb; |
34dc7c2f | 99 | |
b0bc7a84 | 100 | if (BP_IS_HOLE(bp)) |
428870ff | 101 | return (0); |
34dc7c2f | 102 | |
b128c09f | 103 | if (claim_txg == 0 || bp->blk_birth < claim_txg) |
428870ff | 104 | return (0); |
b128c09f | 105 | |
572e2857 BB |
106 | SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, |
107 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
428870ff | 108 | |
294f6806 | 109 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, |
428870ff | 110 | td->td_arg); |
34dc7c2f | 111 | } |
428870ff | 112 | return (0); |
34dc7c2f BB |
113 | } |
114 | ||
115 | static void | |
572e2857 | 116 | traverse_zil(traverse_data_t *td, zil_header_t *zh) |
34dc7c2f | 117 | { |
34dc7c2f BB |
118 | uint64_t claim_txg = zh->zh_claim_txg; |
119 | zilog_t *zilog; | |
120 | ||
34dc7c2f BB |
121 | /* |
122 | * We only want to visit blocks that have been claimed but not yet | |
428870ff | 123 | * replayed; plus, in read-only mode, blocks that are already stable. |
34dc7c2f | 124 | */ |
fb5f0bc8 | 125 | if (claim_txg == 0 && spa_writeable(td->td_spa)) |
34dc7c2f BB |
126 | return; |
127 | ||
b128c09f | 128 | zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh); |
34dc7c2f | 129 | |
b128c09f | 130 | (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td, |
34dc7c2f BB |
131 | claim_txg); |
132 | ||
133 | zil_free(zilog); | |
134 | } | |
135 | ||
9ae529ec CS |
136 | typedef enum resume_skip { |
137 | RESUME_SKIP_ALL, | |
138 | RESUME_SKIP_NONE, | |
139 | RESUME_SKIP_CHILDREN | |
140 | } resume_skip_t; | |
141 | ||
142 | /* | |
143 | * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and | |
144 | * the block indicated by zb does not need to be visited at all. Returns | |
145 | * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the | |
146 | * resume point. This indicates that this block should be visited but not its | |
147 | * children (since they must have been visited in a previous traversal). | |
148 | * Otherwise returns RESUME_SKIP_NONE. | |
149 | */ | |
150 | static resume_skip_t | |
151 | resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, | |
5dbd68a3 | 152 | const zbookmark_phys_t *zb) |
9ae529ec CS |
153 | { |
154 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { | |
155 | /* | |
156 | * If we already visited this bp & everything below, | |
157 | * don't bother doing it again. | |
158 | */ | |
159 | if (zbookmark_is_before(dnp, zb, td->td_resume)) | |
160 | return (RESUME_SKIP_ALL); | |
161 | ||
162 | /* | |
163 | * If we found the block we're trying to resume from, zero | |
164 | * the bookmark out to indicate that we have resumed. | |
165 | */ | |
9ae529ec CS |
166 | if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) { |
167 | bzero(td->td_resume, sizeof (*zb)); | |
168 | if (td->td_flags & TRAVERSE_POST) | |
169 | return (RESUME_SKIP_CHILDREN); | |
170 | } | |
171 | } | |
172 | return (RESUME_SKIP_NONE); | |
173 | } | |
174 | ||
96b89346 MA |
175 | static void |
176 | traverse_prefetch_metadata(traverse_data_t *td, | |
5dbd68a3 | 177 | const blkptr_t *bp, const zbookmark_phys_t *zb) |
96b89346 MA |
178 | { |
179 | uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; | |
180 | ||
181 | if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) | |
182 | return; | |
183 | /* | |
184 | * If we are in the process of resuming, don't prefetch, because | |
185 | * some children will not be needed (and in fact may have already | |
186 | * been freed). | |
187 | */ | |
188 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) | |
189 | return; | |
190 | if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) | |
191 | return; | |
192 | if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) | |
193 | return; | |
194 | ||
294f6806 GW |
195 | (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, |
196 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); | |
96b89346 MA |
197 | } |
198 | ||
1fa8f795 MA |
199 | static boolean_t |
200 | prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp) | |
201 | { | |
202 | ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA); | |
203 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || | |
204 | BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) | |
205 | return (B_FALSE); | |
206 | return (B_TRUE); | |
207 | } | |
208 | ||
c7f8f831 BB |
209 | static int |
210 | traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, | |
5dbd68a3 | 211 | const blkptr_t *bp, const zbookmark_phys_t *zb) |
6656bf56 | 212 | { |
fbeddd60 | 213 | int err = 0; |
c7f8f831 | 214 | arc_buf_t *buf = NULL; |
ee2f17aa | 215 | prefetch_data_t *pd = td->td_pfd; |
9ae529ec CS |
216 | |
217 | switch (resume_skip_check(td, dnp, zb)) { | |
218 | case RESUME_SKIP_ALL: | |
219 | return (0); | |
220 | case RESUME_SKIP_CHILDREN: | |
221 | goto post; | |
222 | case RESUME_SKIP_NONE: | |
223 | break; | |
224 | default: | |
225 | ASSERT(0); | |
226 | } | |
6656bf56 | 227 | |
b0bc7a84 MG |
228 | if (bp->blk_birth == 0) { |
229 | if (spa_feature_is_active(td->td_spa, SPA_FEATURE_HOLE_BIRTH)) { | |
230 | /* | |
231 | * Since this block has a birth time of 0 it must be a | |
232 | * hole created before the SPA_FEATURE_HOLE_BIRTH | |
233 | * feature was enabled. If SPA_FEATURE_HOLE_BIRTH | |
234 | * was enabled before the min_txg for this traveral we | |
235 | * know the hole must have been created before the | |
236 | * min_txg for this traveral, so we can skip it. If | |
237 | * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg | |
238 | * for this traveral we cannot tell if the hole was | |
239 | * created before or after the min_txg for this | |
240 | * traversal, so we cannot skip it. | |
241 | */ | |
242 | uint64_t hole_birth_enabled_txg; | |
243 | VERIFY(spa_feature_enabled_txg(td->td_spa, | |
244 | SPA_FEATURE_HOLE_BIRTH, &hole_birth_enabled_txg)); | |
245 | if (hole_birth_enabled_txg < td->td_min_txg) | |
246 | return (0); | |
247 | } | |
248 | } else if (bp->blk_birth <= td->td_min_txg) { | |
249 | return (0); | |
250 | } | |
251 | ||
ee2f17aa | 252 | if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) { |
b738bc5a | 253 | uint64_t size = BP_GET_LSIZE(bp); |
ee2f17aa | 254 | mutex_enter(&pd->pd_mtx); |
b738bc5a GW |
255 | ASSERT(pd->pd_bytes_fetched >= 0); |
256 | while (pd->pd_bytes_fetched < size && !pd->pd_exited) | |
ee2f17aa | 257 | cv_wait(&pd->pd_cv, &pd->pd_mtx); |
b738bc5a | 258 | pd->pd_bytes_fetched -= size; |
ee2f17aa CD |
259 | cv_broadcast(&pd->pd_cv); |
260 | mutex_exit(&pd->pd_mtx); | |
34dc7c2f BB |
261 | } |
262 | ||
1fa8f795 MA |
263 | if (BP_IS_HOLE(bp)) { |
264 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); | |
265 | if (err != 0) | |
266 | goto post; | |
267 | return (0); | |
268 | } | |
269 | ||
c7f8f831 | 270 | if (td->td_flags & TRAVERSE_PRE) { |
294f6806 | 271 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, |
c7f8f831 BB |
272 | td->td_arg); |
273 | if (err == TRAVERSE_VISIT_NO_CHILDREN) | |
572e2857 | 274 | return (0); |
9ae529ec CS |
275 | if (err != 0) |
276 | goto post; | |
34dc7c2f BB |
277 | } |
278 | ||
c7f8f831 BB |
279 | if (BP_GET_LEVEL(bp) > 0) { |
280 | uint32_t flags = ARC_WAIT; | |
a1687880 BB |
281 | int32_t i; |
282 | int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; | |
5dbd68a3 | 283 | zbookmark_phys_t *czb; |
b128c09f | 284 | |
294f6806 | 285 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 286 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 287 | if (err != 0) |
fbeddd60 | 288 | goto post; |
a1687880 | 289 | |
79c76d5b | 290 | czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); |
96b89346 MA |
291 | |
292 | for (i = 0; i < epb; i++) { | |
a1687880 | 293 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
96b89346 MA |
294 | zb->zb_level - 1, |
295 | zb->zb_blkid * epb + i); | |
a1687880 BB |
296 | traverse_prefetch_metadata(td, |
297 | &((blkptr_t *)buf->b_data)[i], czb); | |
96b89346 | 298 | } |
b128c09f BB |
299 | |
300 | /* recursively visitbp() blocks below this */ | |
96b89346 | 301 | for (i = 0; i < epb; i++) { |
a1687880 | 302 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
c7f8f831 BB |
303 | zb->zb_level - 1, |
304 | zb->zb_blkid * epb + i); | |
a1687880 BB |
305 | err = traverse_visitbp(td, dnp, |
306 | &((blkptr_t *)buf->b_data)[i], czb); | |
fbeddd60 MA |
307 | if (err != 0) |
308 | break; | |
b128c09f | 309 | } |
a1687880 | 310 | |
5dbd68a3 | 311 | kmem_free(czb, sizeof (zbookmark_phys_t)); |
a1687880 | 312 | |
c7f8f831 BB |
313 | } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { |
314 | uint32_t flags = ARC_WAIT; | |
a1687880 BB |
315 | int32_t i; |
316 | int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; | |
ecfb0b5f | 317 | dnode_phys_t *cdnp; |
b128c09f | 318 | |
294f6806 | 319 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 320 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 321 | if (err != 0) |
fbeddd60 | 322 | goto post; |
ecfb0b5f | 323 | cdnp = buf->b_data; |
96b89346 MA |
324 | |
325 | for (i = 0; i < epb; i++) { | |
ecfb0b5f | 326 | prefetch_dnode_metadata(td, &cdnp[i], zb->zb_objset, |
96b89346 MA |
327 | zb->zb_blkid * epb + i); |
328 | } | |
b128c09f BB |
329 | |
330 | /* recursively visitbp() blocks below this */ | |
96b89346 | 331 | for (i = 0; i < epb; i++) { |
ecfb0b5f | 332 | err = traverse_dnode(td, &cdnp[i], zb->zb_objset, |
c7f8f831 | 333 | zb->zb_blkid * epb + i); |
fbeddd60 MA |
334 | if (err != 0) |
335 | break; | |
34dc7c2f | 336 | } |
c7f8f831 BB |
337 | } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { |
338 | uint32_t flags = ARC_WAIT; | |
339 | objset_phys_t *osp; | |
ecfb0b5f | 340 | dnode_phys_t *mdnp, *gdnp, *udnp; |
c7f8f831 | 341 | |
294f6806 | 342 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 343 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 344 | if (err != 0) |
fbeddd60 | 345 | goto post; |
c7f8f831 BB |
346 | |
347 | osp = buf->b_data; | |
ecfb0b5f | 348 | mdnp = &osp->os_meta_dnode; |
349 | gdnp = &osp->os_groupused_dnode; | |
350 | udnp = &osp->os_userused_dnode; | |
351 | ||
352 | prefetch_dnode_metadata(td, mdnp, zb->zb_objset, | |
96b89346 MA |
353 | DMU_META_DNODE_OBJECT); |
354 | if (arc_buf_size(buf) >= sizeof (objset_phys_t)) { | |
ecfb0b5f | 355 | prefetch_dnode_metadata(td, gdnp, zb->zb_objset, |
356 | DMU_GROUPUSED_OBJECT); | |
357 | prefetch_dnode_metadata(td, udnp, zb->zb_objset, | |
358 | DMU_USERUSED_OBJECT); | |
96b89346 MA |
359 | } |
360 | ||
ecfb0b5f | 361 | err = traverse_dnode(td, mdnp, zb->zb_objset, |
c7f8f831 | 362 | DMU_META_DNODE_OBJECT); |
c7f8f831 | 363 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
ecfb0b5f | 364 | err = traverse_dnode(td, gdnp, zb->zb_objset, |
8b464649 | 365 | DMU_GROUPUSED_OBJECT); |
9babb374 | 366 | } |
c7f8f831 | 367 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
ecfb0b5f | 368 | err = traverse_dnode(td, udnp, zb->zb_objset, |
8b464649 | 369 | DMU_USERUSED_OBJECT); |
34dc7c2f | 370 | } |
34dc7c2f BB |
371 | } |
372 | ||
c7f8f831 BB |
373 | if (buf) |
374 | (void) arc_buf_remove_ref(buf, &buf); | |
34dc7c2f | 375 | |
9ae529ec | 376 | post: |
fbeddd60 | 377 | if (err == 0 && (td->td_flags & TRAVERSE_POST)) |
294f6806 | 378 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); |
fbeddd60 MA |
379 | |
380 | if ((td->td_flags & TRAVERSE_HARD) && (err == EIO || err == ECKSUM)) { | |
381 | /* | |
382 | * Ignore this disk error as requested by the HARD flag, | |
383 | * and continue traversal. | |
384 | */ | |
385 | err = 0; | |
9ae529ec CS |
386 | } |
387 | ||
fbeddd60 MA |
388 | /* |
389 | * If we are stopping here, set td_resume. | |
390 | */ | |
391 | if (td->td_resume != NULL && err != 0 && !td->td_paused) { | |
392 | td->td_resume->zb_objset = zb->zb_objset; | |
393 | td->td_resume->zb_object = zb->zb_object; | |
394 | td->td_resume->zb_level = 0; | |
395 | /* | |
396 | * If we have stopped on an indirect block (e.g. due to | |
397 | * i/o error), we have not visited anything below it. | |
398 | * Set the bookmark to the first level-0 block that we need | |
399 | * to visit. This way, the resuming code does not need to | |
400 | * deal with resuming from indirect blocks. | |
401 | */ | |
402 | td->td_resume->zb_blkid = zb->zb_blkid << | |
403 | (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); | |
404 | td->td_paused = B_TRUE; | |
428870ff | 405 | } |
34dc7c2f | 406 | |
fbeddd60 | 407 | return (err); |
34dc7c2f BB |
408 | } |
409 | ||
96b89346 MA |
410 | static void |
411 | prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 412 | uint64_t objset, uint64_t object) |
96b89346 MA |
413 | { |
414 | int j; | |
5dbd68a3 | 415 | zbookmark_phys_t czb; |
96b89346 MA |
416 | |
417 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
418 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 419 | traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); |
96b89346 MA |
420 | } |
421 | ||
422 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
423 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); | |
294f6806 | 424 | traverse_prefetch_metadata(td, &dnp->dn_spill, &czb); |
96b89346 MA |
425 | } |
426 | } | |
427 | ||
9babb374 | 428 | static int |
572e2857 | 429 | traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 430 | uint64_t objset, uint64_t object) |
9babb374 | 431 | { |
fbeddd60 | 432 | int j, err = 0; |
5dbd68a3 | 433 | zbookmark_phys_t czb; |
9babb374 BB |
434 | |
435 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
436 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 437 | err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); |
fbeddd60 MA |
438 | if (err != 0) |
439 | break; | |
9babb374 | 440 | } |
428870ff | 441 | |
33b4de51 | 442 | if (err == 0 && dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { |
96b89346 | 443 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); |
294f6806 | 444 | err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb); |
428870ff | 445 | } |
fbeddd60 | 446 | return (err); |
9babb374 BB |
447 | } |
448 | ||
b128c09f BB |
449 | /* ARGSUSED */ |
450 | static int | |
428870ff | 451 | traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, |
5dbd68a3 | 452 | const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) |
34dc7c2f | 453 | { |
572e2857 | 454 | prefetch_data_t *pfd = arg; |
b128c09f | 455 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
34dc7c2f | 456 | |
b738bc5a | 457 | ASSERT(pfd->pd_bytes_fetched >= 0); |
b128c09f | 458 | if (pfd->pd_cancel) |
2e528b49 | 459 | return (SET_ERROR(EINTR)); |
34dc7c2f | 460 | |
1fa8f795 | 461 | if (!prefetch_needed(pfd, bp)) |
34dc7c2f BB |
462 | return (0); |
463 | ||
b128c09f | 464 | mutex_enter(&pfd->pd_mtx); |
b738bc5a | 465 | while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max) |
b128c09f | 466 | cv_wait(&pfd->pd_cv, &pfd->pd_mtx); |
b738bc5a | 467 | pfd->pd_bytes_fetched += BP_GET_LSIZE(bp); |
b128c09f BB |
468 | cv_broadcast(&pfd->pd_cv); |
469 | mutex_exit(&pfd->pd_mtx); | |
34dc7c2f | 470 | |
294f6806 GW |
471 | (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, |
472 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb); | |
34dc7c2f | 473 | |
b128c09f | 474 | return (0); |
34dc7c2f BB |
475 | } |
476 | ||
34dc7c2f | 477 | static void |
b128c09f | 478 | traverse_prefetch_thread(void *arg) |
34dc7c2f | 479 | { |
572e2857 BB |
480 | traverse_data_t *td_main = arg; |
481 | traverse_data_t td = *td_main; | |
5dbd68a3 | 482 | zbookmark_phys_t czb; |
34dc7c2f | 483 | |
b128c09f BB |
484 | td.td_func = traverse_prefetcher; |
485 | td.td_arg = td_main->td_pfd; | |
486 | td.td_pfd = NULL; | |
34dc7c2f | 487 | |
428870ff BB |
488 | SET_BOOKMARK(&czb, td.td_objset, |
489 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
294f6806 | 490 | (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); |
34dc7c2f | 491 | |
b128c09f BB |
492 | mutex_enter(&td_main->td_pfd->pd_mtx); |
493 | td_main->td_pfd->pd_exited = B_TRUE; | |
494 | cv_broadcast(&td_main->td_pfd->pd_cv); | |
495 | mutex_exit(&td_main->td_pfd->pd_mtx); | |
34dc7c2f BB |
496 | } |
497 | ||
b128c09f BB |
498 | /* |
499 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
500 | * in syncing context). | |
501 | */ | |
502 | static int | |
9ae529ec | 503 | traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, |
5dbd68a3 | 504 | uint64_t txg_start, zbookmark_phys_t *resume, int flags, |
9ae529ec | 505 | blkptr_cb_t func, void *arg) |
34dc7c2f | 506 | { |
47050a88 BB |
507 | traverse_data_t *td; |
508 | prefetch_data_t *pd; | |
5dbd68a3 | 509 | zbookmark_phys_t *czb; |
b128c09f | 510 | int err; |
34dc7c2f | 511 | |
9ae529ec CS |
512 | ASSERT(ds == NULL || objset == ds->ds_object); |
513 | ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); | |
514 | ||
96b89346 MA |
515 | /* |
516 | * The data prefetching mechanism (the prefetch thread) is incompatible | |
517 | * with resuming from a bookmark. | |
518 | */ | |
519 | ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA)); | |
520 | ||
79c76d5b BB |
521 | td = kmem_alloc(sizeof (traverse_data_t), KM_SLEEP); |
522 | pd = kmem_zalloc(sizeof (prefetch_data_t), KM_SLEEP); | |
523 | czb = kmem_alloc(sizeof (zbookmark_phys_t), KM_SLEEP); | |
47050a88 BB |
524 | |
525 | td->td_spa = spa; | |
9ae529ec | 526 | td->td_objset = objset; |
47050a88 BB |
527 | td->td_rootbp = rootbp; |
528 | td->td_min_txg = txg_start; | |
9ae529ec | 529 | td->td_resume = resume; |
47050a88 BB |
530 | td->td_func = func; |
531 | td->td_arg = arg; | |
532 | td->td_pfd = pd; | |
533 | td->td_flags = flags; | |
fbeddd60 | 534 | td->td_paused = B_FALSE; |
b128c09f | 535 | |
47050a88 BB |
536 | pd->pd_flags = flags; |
537 | mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); | |
538 | cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); | |
b128c09f | 539 | |
1421c891 PS |
540 | SET_BOOKMARK(czb, td->td_objset, |
541 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
542 | ||
572e2857 | 543 | /* See comment on ZIL traversal in dsl_scan_visitds. */ |
13fe0198 MA |
544 | if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) { |
545 | uint32_t flags = ARC_WAIT; | |
546 | objset_phys_t *osp; | |
547 | arc_buf_t *buf; | |
572e2857 | 548 | |
13fe0198 MA |
549 | err = arc_read(NULL, td->td_spa, rootbp, |
550 | arc_getbuf_func, &buf, | |
1421c891 | 551 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, czb); |
13fe0198 | 552 | if (err != 0) |
572e2857 BB |
553 | return (err); |
554 | ||
13fe0198 MA |
555 | osp = buf->b_data; |
556 | traverse_zil(td, &osp->os_zil_header); | |
557 | (void) arc_buf_remove_ref(buf, &buf); | |
572e2857 BB |
558 | } |
559 | ||
96b89346 | 560 | if (!(flags & TRAVERSE_PREFETCH_DATA) || |
b128c09f | 561 | 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread, |
47050a88 BB |
562 | td, TQ_NOQUEUE)) |
563 | pd->pd_exited = B_TRUE; | |
b128c09f | 564 | |
294f6806 | 565 | err = traverse_visitbp(td, NULL, rootbp, czb); |
47050a88 BB |
566 | |
567 | mutex_enter(&pd->pd_mtx); | |
568 | pd->pd_cancel = B_TRUE; | |
569 | cv_broadcast(&pd->pd_cv); | |
570 | while (!pd->pd_exited) | |
571 | cv_wait(&pd->pd_cv, &pd->pd_mtx); | |
572 | mutex_exit(&pd->pd_mtx); | |
b128c09f | 573 | |
47050a88 BB |
574 | mutex_destroy(&pd->pd_mtx); |
575 | cv_destroy(&pd->pd_cv); | |
b128c09f | 576 | |
5dbd68a3 | 577 | kmem_free(czb, sizeof (zbookmark_phys_t)); |
a1687880 BB |
578 | kmem_free(pd, sizeof (struct prefetch_data)); |
579 | kmem_free(td, sizeof (struct traverse_data)); | |
34dc7c2f | 580 | |
b128c09f | 581 | return (err); |
34dc7c2f BB |
582 | } |
583 | ||
b128c09f BB |
584 | /* |
585 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
586 | * in syncing context). | |
587 | */ | |
588 | int | |
589 | traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags, | |
590 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 591 | { |
9ae529ec CS |
592 | return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, |
593 | &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg)); | |
594 | } | |
595 | ||
596 | int | |
597 | traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, | |
5dbd68a3 | 598 | uint64_t txg_start, zbookmark_phys_t *resume, int flags, |
9ae529ec CS |
599 | blkptr_cb_t func, void *arg) |
600 | { | |
601 | return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, | |
602 | blkptr, txg_start, resume, flags, func, arg)); | |
34dc7c2f BB |
603 | } |
604 | ||
b128c09f BB |
605 | /* |
606 | * NB: pool must not be changing on-disk (eg, from zdb or sync context). | |
607 | */ | |
608 | int | |
428870ff BB |
609 | traverse_pool(spa_t *spa, uint64_t txg_start, int flags, |
610 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 611 | { |
fbeddd60 | 612 | int err; |
b128c09f BB |
613 | uint64_t obj; |
614 | dsl_pool_t *dp = spa_get_dsl(spa); | |
615 | objset_t *mos = dp->dp_meta_objset; | |
428870ff | 616 | boolean_t hard = (flags & TRAVERSE_HARD); |
b128c09f BB |
617 | |
618 | /* visit the MOS */ | |
9ae529ec CS |
619 | err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), |
620 | txg_start, NULL, flags, func, arg); | |
13fe0198 | 621 | if (err != 0) |
b128c09f BB |
622 | return (err); |
623 | ||
624 | /* visit each dataset */ | |
fbeddd60 | 625 | for (obj = 1; err == 0; |
428870ff | 626 | err = dmu_object_next(mos, &obj, FALSE, txg_start)) { |
b128c09f BB |
627 | dmu_object_info_t doi; |
628 | ||
629 | err = dmu_object_info(mos, obj, &doi); | |
13fe0198 | 630 | if (err != 0) { |
fbeddd60 MA |
631 | if (hard) |
632 | continue; | |
633 | break; | |
428870ff | 634 | } |
b128c09f | 635 | |
fa86b5db | 636 | if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) { |
b128c09f | 637 | dsl_dataset_t *ds; |
428870ff BB |
638 | uint64_t txg = txg_start; |
639 | ||
13fe0198 | 640 | dsl_pool_config_enter(dp, FTAG); |
b128c09f | 641 | err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds); |
13fe0198 MA |
642 | dsl_pool_config_exit(dp, FTAG); |
643 | if (err != 0) { | |
fbeddd60 MA |
644 | if (hard) |
645 | continue; | |
646 | break; | |
428870ff BB |
647 | } |
648 | if (ds->ds_phys->ds_prev_snap_txg > txg) | |
649 | txg = ds->ds_phys->ds_prev_snap_txg; | |
650 | err = traverse_dataset(ds, txg, flags, func, arg); | |
b128c09f | 651 | dsl_dataset_rele(ds, FTAG); |
fbeddd60 MA |
652 | if (err != 0) |
653 | break; | |
b128c09f | 654 | } |
34dc7c2f | 655 | } |
b128c09f BB |
656 | if (err == ESRCH) |
657 | err = 0; | |
fbeddd60 | 658 | return (err); |
34dc7c2f | 659 | } |
c28b2279 BB |
660 | |
661 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
662 | EXPORT_SYMBOL(traverse_dataset); | |
663 | EXPORT_SYMBOL(traverse_pool); | |
c409e464 | 664 | |
b738bc5a GW |
665 | module_param(zfs_pd_bytes_max, int, 0644); |
666 | MODULE_PARM_DESC(zfs_pd_bytes_max, "Max number of bytes to prefetch"); | |
c28b2279 | 667 | #endif |