]>
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. |
2e528b49 | 23 | * Copyright (c) 2013 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 BB |
38 | #include <sys/callb.h> |
39 | ||
572e2857 BB |
40 | int zfs_pd_blks_max = 100; |
41 | ||
42 | typedef struct prefetch_data { | |
b128c09f BB |
43 | kmutex_t pd_mtx; |
44 | kcondvar_t pd_cv; | |
45 | int pd_blks_max; | |
46 | int pd_blks_fetched; | |
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; | |
9ae529ec | 57 | zbookmark_t *td_resume; |
b128c09f | 58 | int td_flags; |
572e2857 | 59 | prefetch_data_t *td_pfd; |
b128c09f BB |
60 | blkptr_cb_t *td_func; |
61 | void *td_arg; | |
572e2857 | 62 | } traverse_data_t; |
34dc7c2f | 63 | |
a1687880 BB |
64 | #define TD_HARD(td) (td->td_flags & TRAVERSE_HARD) |
65 | ||
572e2857 | 66 | static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 67 | uint64_t objset, uint64_t object); |
96b89346 | 68 | static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, |
294f6806 | 69 | uint64_t objset, uint64_t object); |
9babb374 | 70 | |
428870ff | 71 | static int |
34dc7c2f BB |
72 | traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) |
73 | { | |
572e2857 | 74 | traverse_data_t *td = arg; |
b128c09f | 75 | zbookmark_t zb; |
34dc7c2f | 76 | |
b128c09f | 77 | if (bp->blk_birth == 0) |
428870ff | 78 | return (0); |
34dc7c2f | 79 | |
b128c09f | 80 | if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa)) |
428870ff BB |
81 | return (0); |
82 | ||
83 | SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
84 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
85 | ||
294f6806 | 86 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); |
b128c09f | 87 | |
428870ff | 88 | return (0); |
34dc7c2f BB |
89 | } |
90 | ||
428870ff | 91 | static int |
34dc7c2f BB |
92 | traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) |
93 | { | |
572e2857 | 94 | traverse_data_t *td = arg; |
34dc7c2f BB |
95 | |
96 | if (lrc->lrc_txtype == TX_WRITE) { | |
97 | lr_write_t *lr = (lr_write_t *)lrc; | |
98 | blkptr_t *bp = &lr->lr_blkptr; | |
b128c09f | 99 | zbookmark_t zb; |
34dc7c2f | 100 | |
b128c09f | 101 | if (bp->blk_birth == 0) |
428870ff | 102 | return (0); |
34dc7c2f | 103 | |
b128c09f | 104 | if (claim_txg == 0 || bp->blk_birth < claim_txg) |
428870ff | 105 | return (0); |
b128c09f | 106 | |
572e2857 BB |
107 | SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, |
108 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
428870ff | 109 | |
294f6806 | 110 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, |
428870ff | 111 | td->td_arg); |
34dc7c2f | 112 | } |
428870ff | 113 | return (0); |
34dc7c2f BB |
114 | } |
115 | ||
116 | static void | |
572e2857 | 117 | traverse_zil(traverse_data_t *td, zil_header_t *zh) |
34dc7c2f | 118 | { |
34dc7c2f BB |
119 | uint64_t claim_txg = zh->zh_claim_txg; |
120 | zilog_t *zilog; | |
121 | ||
34dc7c2f BB |
122 | /* |
123 | * We only want to visit blocks that have been claimed but not yet | |
428870ff | 124 | * replayed; plus, in read-only mode, blocks that are already stable. |
34dc7c2f | 125 | */ |
fb5f0bc8 | 126 | if (claim_txg == 0 && spa_writeable(td->td_spa)) |
34dc7c2f BB |
127 | return; |
128 | ||
b128c09f | 129 | zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh); |
34dc7c2f | 130 | |
b128c09f | 131 | (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td, |
34dc7c2f BB |
132 | claim_txg); |
133 | ||
134 | zil_free(zilog); | |
135 | } | |
136 | ||
9ae529ec CS |
137 | typedef enum resume_skip { |
138 | RESUME_SKIP_ALL, | |
139 | RESUME_SKIP_NONE, | |
140 | RESUME_SKIP_CHILDREN | |
141 | } resume_skip_t; | |
142 | ||
143 | /* | |
144 | * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and | |
145 | * the block indicated by zb does not need to be visited at all. Returns | |
146 | * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the | |
147 | * resume point. This indicates that this block should be visited but not its | |
148 | * children (since they must have been visited in a previous traversal). | |
149 | * Otherwise returns RESUME_SKIP_NONE. | |
150 | */ | |
151 | static resume_skip_t | |
152 | resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, | |
153 | const zbookmark_t *zb) | |
154 | { | |
155 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { | |
156 | /* | |
157 | * If we already visited this bp & everything below, | |
158 | * don't bother doing it again. | |
159 | */ | |
160 | if (zbookmark_is_before(dnp, zb, td->td_resume)) | |
161 | return (RESUME_SKIP_ALL); | |
162 | ||
163 | /* | |
164 | * If we found the block we're trying to resume from, zero | |
165 | * the bookmark out to indicate that we have resumed. | |
166 | */ | |
167 | ASSERT3U(zb->zb_object, <=, td->td_resume->zb_object); | |
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 | ||
177 | static void | |
178 | traverse_pause(traverse_data_t *td, const zbookmark_t *zb) | |
179 | { | |
180 | ASSERT(td->td_resume != NULL); | |
c99c9001 | 181 | ASSERT0(zb->zb_level); |
9ae529ec CS |
182 | bcopy(zb, td->td_resume, sizeof (*td->td_resume)); |
183 | } | |
184 | ||
96b89346 MA |
185 | static void |
186 | traverse_prefetch_metadata(traverse_data_t *td, | |
294f6806 | 187 | const blkptr_t *bp, const zbookmark_t *zb) |
96b89346 MA |
188 | { |
189 | uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; | |
190 | ||
191 | if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) | |
192 | return; | |
193 | /* | |
194 | * If we are in the process of resuming, don't prefetch, because | |
195 | * some children will not be needed (and in fact may have already | |
196 | * been freed). | |
197 | */ | |
198 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) | |
199 | return; | |
200 | if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) | |
201 | return; | |
202 | if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) | |
203 | return; | |
204 | ||
294f6806 GW |
205 | (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, |
206 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); | |
96b89346 MA |
207 | } |
208 | ||
c7f8f831 BB |
209 | static int |
210 | traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 211 | const blkptr_t *bp, const zbookmark_t *zb) |
6656bf56 | 212 | { |
c7f8f831 BB |
213 | int err = 0, lasterr = 0; |
214 | arc_buf_t *buf = NULL; | |
9ae529ec CS |
215 | boolean_t pause = B_FALSE; |
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 | |
9ae529ec | 228 | if (BP_IS_HOLE(bp)) { |
294f6806 | 229 | err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg); |
c7f8f831 | 230 | return (err); |
34dc7c2f BB |
231 | } |
232 | ||
c7f8f831 | 233 | if (bp->blk_birth <= td->td_min_txg) |
b128c09f | 234 | return (0); |
34dc7c2f | 235 | |
a1687880 BB |
236 | if (td->td_pfd && !td->td_pfd->pd_exited && |
237 | ((td->td_pfd->pd_flags & TRAVERSE_PREFETCH_DATA) || | |
c7f8f831 | 238 | BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) { |
a1687880 BB |
239 | mutex_enter(&td->td_pfd->pd_mtx); |
240 | ASSERT(td->td_pfd->pd_blks_fetched >= 0); | |
241 | while (td->td_pfd->pd_blks_fetched == 0 && | |
242 | !td->td_pfd->pd_exited) | |
243 | cv_wait(&td->td_pfd->pd_cv, &td->td_pfd->pd_mtx); | |
244 | td->td_pfd->pd_blks_fetched--; | |
245 | cv_broadcast(&td->td_pfd->pd_cv); | |
246 | mutex_exit(&td->td_pfd->pd_mtx); | |
34dc7c2f BB |
247 | } |
248 | ||
c7f8f831 | 249 | if (td->td_flags & TRAVERSE_PRE) { |
294f6806 | 250 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, |
c7f8f831 BB |
251 | td->td_arg); |
252 | if (err == TRAVERSE_VISIT_NO_CHILDREN) | |
572e2857 | 253 | return (0); |
9ae529ec CS |
254 | if (err == ERESTART) |
255 | pause = B_TRUE; /* handle pausing at a common point */ | |
256 | if (err != 0) | |
257 | goto post; | |
34dc7c2f BB |
258 | } |
259 | ||
c7f8f831 BB |
260 | if (BP_GET_LEVEL(bp) > 0) { |
261 | uint32_t flags = ARC_WAIT; | |
a1687880 BB |
262 | int32_t i; |
263 | int32_t epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; | |
264 | zbookmark_t *czb; | |
b128c09f | 265 | |
294f6806 | 266 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 267 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 268 | if (err != 0) |
c7f8f831 | 269 | return (err); |
a1687880 BB |
270 | |
271 | czb = kmem_alloc(sizeof (zbookmark_t), KM_PUSHPAGE); | |
96b89346 MA |
272 | |
273 | for (i = 0; i < epb; i++) { | |
a1687880 | 274 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
96b89346 MA |
275 | zb->zb_level - 1, |
276 | zb->zb_blkid * epb + i); | |
a1687880 BB |
277 | traverse_prefetch_metadata(td, |
278 | &((blkptr_t *)buf->b_data)[i], czb); | |
96b89346 | 279 | } |
b128c09f BB |
280 | |
281 | /* recursively visitbp() blocks below this */ | |
96b89346 | 282 | for (i = 0; i < epb; i++) { |
a1687880 | 283 | SET_BOOKMARK(czb, zb->zb_objset, zb->zb_object, |
c7f8f831 BB |
284 | zb->zb_level - 1, |
285 | zb->zb_blkid * epb + i); | |
a1687880 BB |
286 | err = traverse_visitbp(td, dnp, |
287 | &((blkptr_t *)buf->b_data)[i], czb); | |
13fe0198 | 288 | if (err != 0) { |
a1687880 | 289 | if (!TD_HARD(td)) |
428870ff | 290 | break; |
c7f8f831 | 291 | lasterr = err; |
428870ff | 292 | } |
b128c09f | 293 | } |
a1687880 BB |
294 | |
295 | kmem_free(czb, sizeof (zbookmark_t)); | |
296 | ||
c7f8f831 BB |
297 | } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { |
298 | uint32_t flags = ARC_WAIT; | |
a1687880 BB |
299 | int32_t i; |
300 | int32_t epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; | |
b128c09f | 301 | |
294f6806 | 302 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 303 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 304 | if (err != 0) |
c7f8f831 | 305 | return (err); |
96b89346 MA |
306 | dnp = buf->b_data; |
307 | ||
308 | for (i = 0; i < epb; i++) { | |
294f6806 | 309 | prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset, |
96b89346 MA |
310 | zb->zb_blkid * epb + i); |
311 | } | |
b128c09f BB |
312 | |
313 | /* recursively visitbp() blocks below this */ | |
96b89346 | 314 | for (i = 0; i < epb; i++) { |
294f6806 | 315 | err = traverse_dnode(td, &dnp[i], zb->zb_objset, |
c7f8f831 | 316 | zb->zb_blkid * epb + i); |
13fe0198 | 317 | if (err != 0) { |
a1687880 | 318 | if (!TD_HARD(td)) |
428870ff | 319 | break; |
c7f8f831 | 320 | lasterr = err; |
428870ff | 321 | } |
34dc7c2f | 322 | } |
c7f8f831 BB |
323 | } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { |
324 | uint32_t flags = ARC_WAIT; | |
325 | objset_phys_t *osp; | |
326 | dnode_phys_t *dnp; | |
327 | ||
294f6806 | 328 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 329 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 330 | if (err != 0) |
c7f8f831 BB |
331 | return (err); |
332 | ||
333 | osp = buf->b_data; | |
334 | dnp = &osp->os_meta_dnode; | |
294f6806 | 335 | prefetch_dnode_metadata(td, dnp, zb->zb_objset, |
96b89346 MA |
336 | DMU_META_DNODE_OBJECT); |
337 | if (arc_buf_size(buf) >= sizeof (objset_phys_t)) { | |
96b89346 | 338 | prefetch_dnode_metadata(td, &osp->os_groupused_dnode, |
8b464649 MA |
339 | zb->zb_objset, DMU_GROUPUSED_OBJECT); |
340 | prefetch_dnode_metadata(td, &osp->os_userused_dnode, | |
294f6806 | 341 | zb->zb_objset, DMU_USERUSED_OBJECT); |
96b89346 MA |
342 | } |
343 | ||
294f6806 | 344 | err = traverse_dnode(td, dnp, zb->zb_objset, |
c7f8f831 | 345 | DMU_META_DNODE_OBJECT); |
a1687880 | 346 | if (err && TD_HARD(td)) { |
c7f8f831 BB |
347 | lasterr = err; |
348 | err = 0; | |
428870ff | 349 | } |
c7f8f831 | 350 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
8b464649 | 351 | dnp = &osp->os_groupused_dnode; |
294f6806 | 352 | err = traverse_dnode(td, dnp, zb->zb_objset, |
8b464649 | 353 | DMU_GROUPUSED_OBJECT); |
9babb374 | 354 | } |
a1687880 | 355 | if (err && TD_HARD(td)) { |
c7f8f831 BB |
356 | lasterr = err; |
357 | err = 0; | |
428870ff | 358 | } |
c7f8f831 | 359 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
8b464649 | 360 | dnp = &osp->os_userused_dnode; |
294f6806 | 361 | err = traverse_dnode(td, dnp, zb->zb_objset, |
8b464649 | 362 | DMU_USERUSED_OBJECT); |
34dc7c2f | 363 | } |
34dc7c2f BB |
364 | } |
365 | ||
c7f8f831 BB |
366 | if (buf) |
367 | (void) arc_buf_remove_ref(buf, &buf); | |
34dc7c2f | 368 | |
9ae529ec | 369 | post: |
78e2739d | 370 | if (err == 0 && (td->td_flags & TRAVERSE_POST)) { |
294f6806 | 371 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); |
9ae529ec CS |
372 | if (err == ERESTART) |
373 | pause = B_TRUE; | |
374 | } | |
375 | ||
376 | if (pause && td->td_resume != NULL) { | |
377 | ASSERT3U(err, ==, ERESTART); | |
a1687880 | 378 | ASSERT(!TD_HARD(td)); |
9ae529ec | 379 | traverse_pause(td, zb); |
428870ff | 380 | } |
34dc7c2f | 381 | |
c7f8f831 | 382 | return (err != 0 ? err : lasterr); |
34dc7c2f BB |
383 | } |
384 | ||
96b89346 MA |
385 | static void |
386 | prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 387 | uint64_t objset, uint64_t object) |
96b89346 MA |
388 | { |
389 | int j; | |
390 | zbookmark_t czb; | |
391 | ||
392 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
393 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 394 | traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); |
96b89346 MA |
395 | } |
396 | ||
397 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
398 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); | |
294f6806 | 399 | traverse_prefetch_metadata(td, &dnp->dn_spill, &czb); |
96b89346 MA |
400 | } |
401 | } | |
402 | ||
9babb374 | 403 | static int |
572e2857 | 404 | traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 405 | uint64_t objset, uint64_t object) |
9babb374 | 406 | { |
428870ff | 407 | int j, err = 0, lasterr = 0; |
9babb374 BB |
408 | zbookmark_t czb; |
409 | ||
410 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
411 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 412 | err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); |
13fe0198 | 413 | if (err != 0) { |
a1687880 | 414 | if (!TD_HARD(td)) |
428870ff BB |
415 | break; |
416 | lasterr = err; | |
417 | } | |
9babb374 | 418 | } |
428870ff BB |
419 | |
420 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
96b89346 | 421 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); |
294f6806 | 422 | err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb); |
13fe0198 | 423 | if (err != 0) { |
a1687880 | 424 | if (!TD_HARD(td)) |
428870ff BB |
425 | return (err); |
426 | lasterr = err; | |
427 | } | |
428 | } | |
429 | return (err != 0 ? err : lasterr); | |
9babb374 BB |
430 | } |
431 | ||
b128c09f BB |
432 | /* ARGSUSED */ |
433 | static int | |
428870ff | 434 | traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, |
294f6806 | 435 | const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg) |
34dc7c2f | 436 | { |
572e2857 | 437 | prefetch_data_t *pfd = arg; |
b128c09f | 438 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
34dc7c2f | 439 | |
b128c09f BB |
440 | ASSERT(pfd->pd_blks_fetched >= 0); |
441 | if (pfd->pd_cancel) | |
2e528b49 | 442 | return (SET_ERROR(EINTR)); |
34dc7c2f | 443 | |
b128c09f | 444 | if (bp == NULL || !((pfd->pd_flags & TRAVERSE_PREFETCH_DATA) || |
428870ff BB |
445 | BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0) || |
446 | BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) | |
34dc7c2f BB |
447 | return (0); |
448 | ||
b128c09f BB |
449 | mutex_enter(&pfd->pd_mtx); |
450 | while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max) | |
451 | cv_wait(&pfd->pd_cv, &pfd->pd_mtx); | |
452 | pfd->pd_blks_fetched++; | |
453 | cv_broadcast(&pfd->pd_cv); | |
454 | mutex_exit(&pfd->pd_mtx); | |
34dc7c2f | 455 | |
294f6806 GW |
456 | (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, |
457 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb); | |
34dc7c2f | 458 | |
b128c09f | 459 | return (0); |
34dc7c2f BB |
460 | } |
461 | ||
34dc7c2f | 462 | static void |
b128c09f | 463 | traverse_prefetch_thread(void *arg) |
34dc7c2f | 464 | { |
572e2857 BB |
465 | traverse_data_t *td_main = arg; |
466 | traverse_data_t td = *td_main; | |
b128c09f | 467 | zbookmark_t czb; |
34dc7c2f | 468 | |
b128c09f BB |
469 | td.td_func = traverse_prefetcher; |
470 | td.td_arg = td_main->td_pfd; | |
471 | td.td_pfd = NULL; | |
34dc7c2f | 472 | |
428870ff BB |
473 | SET_BOOKMARK(&czb, td.td_objset, |
474 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
294f6806 | 475 | (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); |
34dc7c2f | 476 | |
b128c09f BB |
477 | mutex_enter(&td_main->td_pfd->pd_mtx); |
478 | td_main->td_pfd->pd_exited = B_TRUE; | |
479 | cv_broadcast(&td_main->td_pfd->pd_cv); | |
480 | mutex_exit(&td_main->td_pfd->pd_mtx); | |
34dc7c2f BB |
481 | } |
482 | ||
b128c09f BB |
483 | /* |
484 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
485 | * in syncing context). | |
486 | */ | |
487 | static int | |
9ae529ec CS |
488 | traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, |
489 | uint64_t txg_start, zbookmark_t *resume, int flags, | |
490 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 491 | { |
47050a88 BB |
492 | traverse_data_t *td; |
493 | prefetch_data_t *pd; | |
494 | zbookmark_t *czb; | |
b128c09f | 495 | int err; |
34dc7c2f | 496 | |
9ae529ec CS |
497 | ASSERT(ds == NULL || objset == ds->ds_object); |
498 | ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); | |
499 | ||
96b89346 MA |
500 | /* |
501 | * The data prefetching mechanism (the prefetch thread) is incompatible | |
502 | * with resuming from a bookmark. | |
503 | */ | |
504 | ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA)); | |
505 | ||
a1687880 BB |
506 | td = kmem_alloc(sizeof (traverse_data_t), KM_PUSHPAGE); |
507 | pd = kmem_zalloc(sizeof (prefetch_data_t), KM_PUSHPAGE); | |
508 | czb = kmem_alloc(sizeof (zbookmark_t), KM_PUSHPAGE); | |
47050a88 BB |
509 | |
510 | td->td_spa = spa; | |
9ae529ec | 511 | td->td_objset = objset; |
47050a88 BB |
512 | td->td_rootbp = rootbp; |
513 | td->td_min_txg = txg_start; | |
9ae529ec | 514 | td->td_resume = resume; |
47050a88 BB |
515 | td->td_func = func; |
516 | td->td_arg = arg; | |
517 | td->td_pfd = pd; | |
518 | td->td_flags = flags; | |
b128c09f | 519 | |
47050a88 BB |
520 | pd->pd_blks_max = zfs_pd_blks_max; |
521 | pd->pd_flags = flags; | |
522 | mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); | |
523 | cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); | |
b128c09f | 524 | |
1421c891 PS |
525 | SET_BOOKMARK(czb, td->td_objset, |
526 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
527 | ||
572e2857 | 528 | /* See comment on ZIL traversal in dsl_scan_visitds. */ |
13fe0198 MA |
529 | if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) { |
530 | uint32_t flags = ARC_WAIT; | |
531 | objset_phys_t *osp; | |
532 | arc_buf_t *buf; | |
572e2857 | 533 | |
13fe0198 MA |
534 | err = arc_read(NULL, td->td_spa, rootbp, |
535 | arc_getbuf_func, &buf, | |
1421c891 | 536 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, czb); |
13fe0198 | 537 | if (err != 0) |
572e2857 BB |
538 | return (err); |
539 | ||
13fe0198 MA |
540 | osp = buf->b_data; |
541 | traverse_zil(td, &osp->os_zil_header); | |
542 | (void) arc_buf_remove_ref(buf, &buf); | |
572e2857 BB |
543 | } |
544 | ||
96b89346 | 545 | if (!(flags & TRAVERSE_PREFETCH_DATA) || |
b128c09f | 546 | 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread, |
47050a88 BB |
547 | td, TQ_NOQUEUE)) |
548 | pd->pd_exited = B_TRUE; | |
b128c09f | 549 | |
294f6806 | 550 | err = traverse_visitbp(td, NULL, rootbp, czb); |
47050a88 BB |
551 | |
552 | mutex_enter(&pd->pd_mtx); | |
553 | pd->pd_cancel = B_TRUE; | |
554 | cv_broadcast(&pd->pd_cv); | |
555 | while (!pd->pd_exited) | |
556 | cv_wait(&pd->pd_cv, &pd->pd_mtx); | |
557 | mutex_exit(&pd->pd_mtx); | |
b128c09f | 558 | |
47050a88 BB |
559 | mutex_destroy(&pd->pd_mtx); |
560 | cv_destroy(&pd->pd_cv); | |
b128c09f | 561 | |
a1687880 BB |
562 | kmem_free(czb, sizeof (zbookmark_t)); |
563 | kmem_free(pd, sizeof (struct prefetch_data)); | |
564 | kmem_free(td, sizeof (struct traverse_data)); | |
34dc7c2f | 565 | |
b128c09f | 566 | return (err); |
34dc7c2f BB |
567 | } |
568 | ||
b128c09f BB |
569 | /* |
570 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
571 | * in syncing context). | |
572 | */ | |
573 | int | |
574 | traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags, | |
575 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 576 | { |
9ae529ec CS |
577 | return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, |
578 | &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg)); | |
579 | } | |
580 | ||
581 | int | |
582 | traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, | |
583 | uint64_t txg_start, zbookmark_t *resume, int flags, | |
584 | blkptr_cb_t func, void *arg) | |
585 | { | |
586 | return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, | |
587 | blkptr, txg_start, resume, flags, func, arg)); | |
34dc7c2f BB |
588 | } |
589 | ||
b128c09f BB |
590 | /* |
591 | * NB: pool must not be changing on-disk (eg, from zdb or sync context). | |
592 | */ | |
593 | int | |
428870ff BB |
594 | traverse_pool(spa_t *spa, uint64_t txg_start, int flags, |
595 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 596 | { |
428870ff | 597 | int err, lasterr = 0; |
b128c09f BB |
598 | uint64_t obj; |
599 | dsl_pool_t *dp = spa_get_dsl(spa); | |
600 | objset_t *mos = dp->dp_meta_objset; | |
428870ff | 601 | boolean_t hard = (flags & TRAVERSE_HARD); |
b128c09f BB |
602 | |
603 | /* visit the MOS */ | |
9ae529ec CS |
604 | err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), |
605 | txg_start, NULL, flags, func, arg); | |
13fe0198 | 606 | if (err != 0) |
b128c09f BB |
607 | return (err); |
608 | ||
609 | /* visit each dataset */ | |
428870ff BB |
610 | for (obj = 1; err == 0 || (err != ESRCH && hard); |
611 | err = dmu_object_next(mos, &obj, FALSE, txg_start)) { | |
b128c09f BB |
612 | dmu_object_info_t doi; |
613 | ||
614 | err = dmu_object_info(mos, obj, &doi); | |
13fe0198 | 615 | if (err != 0) { |
428870ff BB |
616 | if (!hard) |
617 | return (err); | |
618 | lasterr = err; | |
619 | continue; | |
620 | } | |
b128c09f BB |
621 | |
622 | if (doi.doi_type == DMU_OT_DSL_DATASET) { | |
623 | dsl_dataset_t *ds; | |
428870ff BB |
624 | uint64_t txg = txg_start; |
625 | ||
13fe0198 | 626 | dsl_pool_config_enter(dp, FTAG); |
b128c09f | 627 | err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds); |
13fe0198 MA |
628 | dsl_pool_config_exit(dp, FTAG); |
629 | if (err != 0) { | |
428870ff BB |
630 | if (!hard) |
631 | return (err); | |
632 | lasterr = err; | |
633 | continue; | |
634 | } | |
635 | if (ds->ds_phys->ds_prev_snap_txg > txg) | |
636 | txg = ds->ds_phys->ds_prev_snap_txg; | |
637 | err = traverse_dataset(ds, txg, flags, func, arg); | |
b128c09f | 638 | dsl_dataset_rele(ds, FTAG); |
13fe0198 | 639 | if (err != 0) { |
428870ff BB |
640 | if (!hard) |
641 | return (err); | |
642 | lasterr = err; | |
643 | } | |
b128c09f | 644 | } |
34dc7c2f | 645 | } |
b128c09f BB |
646 | if (err == ESRCH) |
647 | err = 0; | |
428870ff | 648 | return (err != 0 ? err : lasterr); |
34dc7c2f | 649 | } |
c28b2279 BB |
650 | |
651 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
652 | EXPORT_SYMBOL(traverse_dataset); | |
653 | EXPORT_SYMBOL(traverse_pool); | |
c409e464 BB |
654 | |
655 | module_param(zfs_pd_blks_max, int, 0644); | |
656 | MODULE_PARM_DESC(zfs_pd_blks_max, "Max number of blocks to prefetch"); | |
c28b2279 | 657 | #endif |