]>
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. |
9ae529ec | 23 | * Copyright (c) 2012 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 | |
572e2857 | 64 | static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 65 | uint64_t objset, uint64_t object); |
96b89346 | 66 | static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, |
294f6806 | 67 | uint64_t objset, uint64_t object); |
9babb374 | 68 | |
428870ff | 69 | static int |
34dc7c2f BB |
70 | traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) |
71 | { | |
572e2857 | 72 | traverse_data_t *td = arg; |
b128c09f | 73 | zbookmark_t zb; |
34dc7c2f | 74 | |
b128c09f | 75 | if (bp->blk_birth == 0) |
428870ff | 76 | return (0); |
34dc7c2f | 77 | |
b128c09f | 78 | if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa)) |
428870ff BB |
79 | return (0); |
80 | ||
81 | SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
82 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
83 | ||
294f6806 | 84 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); |
b128c09f | 85 | |
428870ff | 86 | return (0); |
34dc7c2f BB |
87 | } |
88 | ||
428870ff | 89 | static int |
34dc7c2f BB |
90 | traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) |
91 | { | |
572e2857 | 92 | traverse_data_t *td = arg; |
34dc7c2f BB |
93 | |
94 | if (lrc->lrc_txtype == TX_WRITE) { | |
95 | lr_write_t *lr = (lr_write_t *)lrc; | |
96 | blkptr_t *bp = &lr->lr_blkptr; | |
b128c09f | 97 | zbookmark_t zb; |
34dc7c2f | 98 | |
b128c09f | 99 | if (bp->blk_birth == 0) |
428870ff | 100 | return (0); |
34dc7c2f | 101 | |
b128c09f | 102 | if (claim_txg == 0 || bp->blk_birth < claim_txg) |
428870ff | 103 | return (0); |
b128c09f | 104 | |
572e2857 BB |
105 | SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, |
106 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
428870ff | 107 | |
294f6806 | 108 | (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, |
428870ff | 109 | td->td_arg); |
34dc7c2f | 110 | } |
428870ff | 111 | return (0); |
34dc7c2f BB |
112 | } |
113 | ||
114 | static void | |
572e2857 | 115 | traverse_zil(traverse_data_t *td, zil_header_t *zh) |
34dc7c2f | 116 | { |
34dc7c2f BB |
117 | uint64_t claim_txg = zh->zh_claim_txg; |
118 | zilog_t *zilog; | |
119 | ||
34dc7c2f BB |
120 | /* |
121 | * We only want to visit blocks that have been claimed but not yet | |
428870ff | 122 | * replayed; plus, in read-only mode, blocks that are already stable. |
34dc7c2f | 123 | */ |
fb5f0bc8 | 124 | if (claim_txg == 0 && spa_writeable(td->td_spa)) |
34dc7c2f BB |
125 | return; |
126 | ||
b128c09f | 127 | zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh); |
34dc7c2f | 128 | |
b128c09f | 129 | (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td, |
34dc7c2f BB |
130 | claim_txg); |
131 | ||
132 | zil_free(zilog); | |
133 | } | |
134 | ||
9ae529ec CS |
135 | typedef enum resume_skip { |
136 | RESUME_SKIP_ALL, | |
137 | RESUME_SKIP_NONE, | |
138 | RESUME_SKIP_CHILDREN | |
139 | } resume_skip_t; | |
140 | ||
141 | /* | |
142 | * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and | |
143 | * the block indicated by zb does not need to be visited at all. Returns | |
144 | * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the | |
145 | * resume point. This indicates that this block should be visited but not its | |
146 | * children (since they must have been visited in a previous traversal). | |
147 | * Otherwise returns RESUME_SKIP_NONE. | |
148 | */ | |
149 | static resume_skip_t | |
150 | resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, | |
151 | const zbookmark_t *zb) | |
152 | { | |
153 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { | |
154 | /* | |
155 | * If we already visited this bp & everything below, | |
156 | * don't bother doing it again. | |
157 | */ | |
158 | if (zbookmark_is_before(dnp, zb, td->td_resume)) | |
159 | return (RESUME_SKIP_ALL); | |
160 | ||
161 | /* | |
162 | * If we found the block we're trying to resume from, zero | |
163 | * the bookmark out to indicate that we have resumed. | |
164 | */ | |
165 | ASSERT3U(zb->zb_object, <=, td->td_resume->zb_object); | |
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 | ||
175 | static void | |
176 | traverse_pause(traverse_data_t *td, const zbookmark_t *zb) | |
177 | { | |
178 | ASSERT(td->td_resume != NULL); | |
c99c9001 | 179 | ASSERT0(zb->zb_level); |
9ae529ec CS |
180 | bcopy(zb, td->td_resume, sizeof (*td->td_resume)); |
181 | } | |
182 | ||
96b89346 MA |
183 | static void |
184 | traverse_prefetch_metadata(traverse_data_t *td, | |
294f6806 | 185 | const blkptr_t *bp, const zbookmark_t *zb) |
96b89346 MA |
186 | { |
187 | uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; | |
188 | ||
189 | if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) | |
190 | return; | |
191 | /* | |
192 | * If we are in the process of resuming, don't prefetch, because | |
193 | * some children will not be needed (and in fact may have already | |
194 | * been freed). | |
195 | */ | |
196 | if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) | |
197 | return; | |
198 | if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) | |
199 | return; | |
200 | if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) | |
201 | return; | |
202 | ||
294f6806 GW |
203 | (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, |
204 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); | |
96b89346 MA |
205 | } |
206 | ||
c7f8f831 BB |
207 | static int |
208 | traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 209 | const blkptr_t *bp, const zbookmark_t *zb) |
6656bf56 | 210 | { |
c7f8f831 BB |
211 | zbookmark_t czb; |
212 | int err = 0, lasterr = 0; | |
213 | arc_buf_t *buf = NULL; | |
214 | prefetch_data_t *pd = td->td_pfd; | |
215 | boolean_t hard = td->td_flags & TRAVERSE_HARD; | |
9ae529ec CS |
216 | boolean_t pause = B_FALSE; |
217 | ||
218 | switch (resume_skip_check(td, dnp, zb)) { | |
219 | case RESUME_SKIP_ALL: | |
220 | return (0); | |
221 | case RESUME_SKIP_CHILDREN: | |
222 | goto post; | |
223 | case RESUME_SKIP_NONE: | |
224 | break; | |
225 | default: | |
226 | ASSERT(0); | |
227 | } | |
6656bf56 | 228 | |
9ae529ec | 229 | if (BP_IS_HOLE(bp)) { |
294f6806 | 230 | err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg); |
c7f8f831 | 231 | return (err); |
34dc7c2f BB |
232 | } |
233 | ||
c7f8f831 | 234 | if (bp->blk_birth <= td->td_min_txg) |
b128c09f | 235 | return (0); |
34dc7c2f | 236 | |
c7f8f831 BB |
237 | if (pd && !pd->pd_exited && |
238 | ((pd->pd_flags & TRAVERSE_PREFETCH_DATA) || | |
239 | BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) { | |
240 | mutex_enter(&pd->pd_mtx); | |
241 | ASSERT(pd->pd_blks_fetched >= 0); | |
242 | while (pd->pd_blks_fetched == 0 && !pd->pd_exited) | |
243 | cv_wait(&pd->pd_cv, &pd->pd_mtx); | |
244 | pd->pd_blks_fetched--; | |
245 | cv_broadcast(&pd->pd_cv); | |
246 | mutex_exit(&pd->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; | |
262 | int i; | |
263 | blkptr_t *cbp; | |
264 | int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; | |
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); |
96b89346 MA |
270 | cbp = buf->b_data; |
271 | ||
272 | for (i = 0; i < epb; i++) { | |
273 | SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, | |
274 | zb->zb_level - 1, | |
275 | zb->zb_blkid * epb + i); | |
294f6806 | 276 | traverse_prefetch_metadata(td, &cbp[i], &czb); |
96b89346 | 277 | } |
b128c09f BB |
278 | |
279 | /* recursively visitbp() blocks below this */ | |
96b89346 | 280 | for (i = 0; i < epb; i++) { |
c7f8f831 BB |
281 | SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, |
282 | zb->zb_level - 1, | |
283 | zb->zb_blkid * epb + i); | |
294f6806 | 284 | err = traverse_visitbp(td, dnp, &cbp[i], &czb); |
13fe0198 | 285 | if (err != 0) { |
c7f8f831 | 286 | if (!hard) |
428870ff | 287 | break; |
c7f8f831 | 288 | lasterr = err; |
428870ff | 289 | } |
b128c09f | 290 | } |
c7f8f831 BB |
291 | } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { |
292 | uint32_t flags = ARC_WAIT; | |
293 | int i; | |
294 | int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; | |
b128c09f | 295 | |
294f6806 | 296 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 297 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 298 | if (err != 0) |
c7f8f831 | 299 | return (err); |
96b89346 MA |
300 | dnp = buf->b_data; |
301 | ||
302 | for (i = 0; i < epb; i++) { | |
294f6806 | 303 | prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset, |
96b89346 MA |
304 | zb->zb_blkid * epb + i); |
305 | } | |
b128c09f BB |
306 | |
307 | /* recursively visitbp() blocks below this */ | |
96b89346 | 308 | for (i = 0; i < epb; i++) { |
294f6806 | 309 | err = traverse_dnode(td, &dnp[i], zb->zb_objset, |
c7f8f831 | 310 | zb->zb_blkid * epb + i); |
13fe0198 | 311 | if (err != 0) { |
c7f8f831 | 312 | if (!hard) |
428870ff | 313 | break; |
c7f8f831 | 314 | lasterr = err; |
428870ff | 315 | } |
34dc7c2f | 316 | } |
c7f8f831 BB |
317 | } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { |
318 | uint32_t flags = ARC_WAIT; | |
319 | objset_phys_t *osp; | |
320 | dnode_phys_t *dnp; | |
321 | ||
294f6806 | 322 | err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, |
c7f8f831 | 323 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); |
13fe0198 | 324 | if (err != 0) |
c7f8f831 BB |
325 | return (err); |
326 | ||
327 | osp = buf->b_data; | |
328 | dnp = &osp->os_meta_dnode; | |
294f6806 | 329 | prefetch_dnode_metadata(td, dnp, zb->zb_objset, |
96b89346 MA |
330 | DMU_META_DNODE_OBJECT); |
331 | if (arc_buf_size(buf) >= sizeof (objset_phys_t)) { | |
332 | prefetch_dnode_metadata(td, &osp->os_userused_dnode, | |
294f6806 | 333 | zb->zb_objset, DMU_USERUSED_OBJECT); |
96b89346 | 334 | prefetch_dnode_metadata(td, &osp->os_groupused_dnode, |
294f6806 | 335 | zb->zb_objset, DMU_USERUSED_OBJECT); |
96b89346 MA |
336 | } |
337 | ||
294f6806 | 338 | err = traverse_dnode(td, dnp, zb->zb_objset, |
c7f8f831 BB |
339 | DMU_META_DNODE_OBJECT); |
340 | if (err && hard) { | |
341 | lasterr = err; | |
342 | err = 0; | |
428870ff | 343 | } |
c7f8f831 BB |
344 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
345 | dnp = &osp->os_userused_dnode; | |
294f6806 | 346 | err = traverse_dnode(td, dnp, zb->zb_objset, |
9babb374 BB |
347 | DMU_USERUSED_OBJECT); |
348 | } | |
c7f8f831 BB |
349 | if (err && hard) { |
350 | lasterr = err; | |
351 | err = 0; | |
428870ff | 352 | } |
c7f8f831 BB |
353 | if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { |
354 | dnp = &osp->os_groupused_dnode; | |
294f6806 | 355 | err = traverse_dnode(td, dnp, zb->zb_objset, |
9babb374 | 356 | DMU_GROUPUSED_OBJECT); |
34dc7c2f | 357 | } |
34dc7c2f BB |
358 | } |
359 | ||
c7f8f831 BB |
360 | if (buf) |
361 | (void) arc_buf_remove_ref(buf, &buf); | |
34dc7c2f | 362 | |
9ae529ec | 363 | post: |
c7f8f831 | 364 | if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST)) { |
294f6806 | 365 | err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); |
9ae529ec CS |
366 | if (err == ERESTART) |
367 | pause = B_TRUE; | |
368 | } | |
369 | ||
370 | if (pause && td->td_resume != NULL) { | |
371 | ASSERT3U(err, ==, ERESTART); | |
372 | ASSERT(!hard); | |
373 | traverse_pause(td, zb); | |
428870ff | 374 | } |
34dc7c2f | 375 | |
c7f8f831 | 376 | return (err != 0 ? err : lasterr); |
34dc7c2f BB |
377 | } |
378 | ||
96b89346 MA |
379 | static void |
380 | prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, | |
294f6806 | 381 | uint64_t objset, uint64_t object) |
96b89346 MA |
382 | { |
383 | int j; | |
384 | zbookmark_t czb; | |
385 | ||
386 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
387 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 388 | traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); |
96b89346 MA |
389 | } |
390 | ||
391 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
392 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); | |
294f6806 | 393 | traverse_prefetch_metadata(td, &dnp->dn_spill, &czb); |
96b89346 MA |
394 | } |
395 | } | |
396 | ||
9babb374 | 397 | static int |
572e2857 | 398 | traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, |
294f6806 | 399 | uint64_t objset, uint64_t object) |
9babb374 | 400 | { |
428870ff | 401 | int j, err = 0, lasterr = 0; |
9babb374 | 402 | zbookmark_t czb; |
428870ff | 403 | boolean_t hard = (td->td_flags & TRAVERSE_HARD); |
9babb374 BB |
404 | |
405 | for (j = 0; j < dnp->dn_nblkptr; j++) { | |
406 | SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); | |
294f6806 | 407 | err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); |
13fe0198 | 408 | if (err != 0) { |
428870ff BB |
409 | if (!hard) |
410 | break; | |
411 | lasterr = err; | |
412 | } | |
9babb374 | 413 | } |
428870ff BB |
414 | |
415 | if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { | |
96b89346 | 416 | SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); |
294f6806 | 417 | err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb); |
13fe0198 | 418 | if (err != 0) { |
428870ff BB |
419 | if (!hard) |
420 | return (err); | |
421 | lasterr = err; | |
422 | } | |
423 | } | |
424 | return (err != 0 ? err : lasterr); | |
9babb374 BB |
425 | } |
426 | ||
b128c09f BB |
427 | /* ARGSUSED */ |
428 | static int | |
428870ff | 429 | traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, |
294f6806 | 430 | const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg) |
34dc7c2f | 431 | { |
572e2857 | 432 | prefetch_data_t *pfd = arg; |
b128c09f | 433 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
34dc7c2f | 434 | |
b128c09f BB |
435 | ASSERT(pfd->pd_blks_fetched >= 0); |
436 | if (pfd->pd_cancel) | |
437 | return (EINTR); | |
34dc7c2f | 438 | |
b128c09f | 439 | if (bp == NULL || !((pfd->pd_flags & TRAVERSE_PREFETCH_DATA) || |
428870ff BB |
440 | BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0) || |
441 | BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) | |
34dc7c2f BB |
442 | return (0); |
443 | ||
b128c09f BB |
444 | mutex_enter(&pfd->pd_mtx); |
445 | while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max) | |
446 | cv_wait(&pfd->pd_cv, &pfd->pd_mtx); | |
447 | pfd->pd_blks_fetched++; | |
448 | cv_broadcast(&pfd->pd_cv); | |
449 | mutex_exit(&pfd->pd_mtx); | |
34dc7c2f | 450 | |
294f6806 GW |
451 | (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, |
452 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb); | |
34dc7c2f | 453 | |
b128c09f | 454 | return (0); |
34dc7c2f BB |
455 | } |
456 | ||
34dc7c2f | 457 | static void |
b128c09f | 458 | traverse_prefetch_thread(void *arg) |
34dc7c2f | 459 | { |
572e2857 BB |
460 | traverse_data_t *td_main = arg; |
461 | traverse_data_t td = *td_main; | |
b128c09f | 462 | zbookmark_t czb; |
34dc7c2f | 463 | |
b128c09f BB |
464 | td.td_func = traverse_prefetcher; |
465 | td.td_arg = td_main->td_pfd; | |
466 | td.td_pfd = NULL; | |
34dc7c2f | 467 | |
428870ff BB |
468 | SET_BOOKMARK(&czb, td.td_objset, |
469 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); | |
294f6806 | 470 | (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); |
34dc7c2f | 471 | |
b128c09f BB |
472 | mutex_enter(&td_main->td_pfd->pd_mtx); |
473 | td_main->td_pfd->pd_exited = B_TRUE; | |
474 | cv_broadcast(&td_main->td_pfd->pd_cv); | |
475 | mutex_exit(&td_main->td_pfd->pd_mtx); | |
34dc7c2f BB |
476 | } |
477 | ||
b128c09f BB |
478 | /* |
479 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
480 | * in syncing context). | |
481 | */ | |
482 | static int | |
9ae529ec CS |
483 | traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, |
484 | uint64_t txg_start, zbookmark_t *resume, int flags, | |
485 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 486 | { |
47050a88 BB |
487 | traverse_data_t *td; |
488 | prefetch_data_t *pd; | |
489 | zbookmark_t *czb; | |
b128c09f | 490 | int err; |
34dc7c2f | 491 | |
9ae529ec CS |
492 | ASSERT(ds == NULL || objset == ds->ds_object); |
493 | ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); | |
494 | ||
96b89346 MA |
495 | /* |
496 | * The data prefetching mechanism (the prefetch thread) is incompatible | |
497 | * with resuming from a bookmark. | |
498 | */ | |
499 | ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA)); | |
500 | ||
b8d06fca RY |
501 | td = kmem_alloc(sizeof(traverse_data_t), KM_PUSHPAGE); |
502 | pd = kmem_zalloc(sizeof(prefetch_data_t), KM_PUSHPAGE); | |
503 | czb = kmem_alloc(sizeof(zbookmark_t), KM_PUSHPAGE); | |
47050a88 BB |
504 | |
505 | td->td_spa = spa; | |
9ae529ec | 506 | td->td_objset = objset; |
47050a88 BB |
507 | td->td_rootbp = rootbp; |
508 | td->td_min_txg = txg_start; | |
9ae529ec | 509 | td->td_resume = resume; |
47050a88 BB |
510 | td->td_func = func; |
511 | td->td_arg = arg; | |
512 | td->td_pfd = pd; | |
513 | td->td_flags = flags; | |
b128c09f | 514 | |
47050a88 BB |
515 | pd->pd_blks_max = zfs_pd_blks_max; |
516 | pd->pd_flags = flags; | |
517 | mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); | |
518 | cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); | |
b128c09f | 519 | |
572e2857 | 520 | /* See comment on ZIL traversal in dsl_scan_visitds. */ |
13fe0198 MA |
521 | if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) { |
522 | uint32_t flags = ARC_WAIT; | |
523 | objset_phys_t *osp; | |
524 | arc_buf_t *buf; | |
572e2857 | 525 | |
13fe0198 MA |
526 | err = arc_read(NULL, td->td_spa, rootbp, |
527 | arc_getbuf_func, &buf, | |
528 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL); | |
529 | if (err != 0) | |
572e2857 BB |
530 | return (err); |
531 | ||
13fe0198 MA |
532 | osp = buf->b_data; |
533 | traverse_zil(td, &osp->os_zil_header); | |
534 | (void) arc_buf_remove_ref(buf, &buf); | |
572e2857 BB |
535 | } |
536 | ||
96b89346 | 537 | if (!(flags & TRAVERSE_PREFETCH_DATA) || |
b128c09f | 538 | 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread, |
47050a88 BB |
539 | td, TQ_NOQUEUE)) |
540 | pd->pd_exited = B_TRUE; | |
b128c09f | 541 | |
47050a88 | 542 | SET_BOOKMARK(czb, td->td_objset, |
428870ff | 543 | ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); |
294f6806 | 544 | err = traverse_visitbp(td, NULL, rootbp, czb); |
47050a88 BB |
545 | |
546 | mutex_enter(&pd->pd_mtx); | |
547 | pd->pd_cancel = B_TRUE; | |
548 | cv_broadcast(&pd->pd_cv); | |
549 | while (!pd->pd_exited) | |
550 | cv_wait(&pd->pd_cv, &pd->pd_mtx); | |
551 | mutex_exit(&pd->pd_mtx); | |
b128c09f | 552 | |
47050a88 BB |
553 | mutex_destroy(&pd->pd_mtx); |
554 | cv_destroy(&pd->pd_cv); | |
b128c09f | 555 | |
47050a88 BB |
556 | kmem_free(czb, sizeof(zbookmark_t)); |
557 | kmem_free(pd, sizeof(struct prefetch_data)); | |
558 | kmem_free(td, sizeof(struct traverse_data)); | |
34dc7c2f | 559 | |
b128c09f | 560 | return (err); |
34dc7c2f BB |
561 | } |
562 | ||
b128c09f BB |
563 | /* |
564 | * NB: dataset must not be changing on-disk (eg, is a snapshot or we are | |
565 | * in syncing context). | |
566 | */ | |
567 | int | |
568 | traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags, | |
569 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 570 | { |
9ae529ec CS |
571 | return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, |
572 | &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg)); | |
573 | } | |
574 | ||
575 | int | |
576 | traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, | |
577 | uint64_t txg_start, zbookmark_t *resume, int flags, | |
578 | blkptr_cb_t func, void *arg) | |
579 | { | |
580 | return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, | |
581 | blkptr, txg_start, resume, flags, func, arg)); | |
34dc7c2f BB |
582 | } |
583 | ||
b128c09f BB |
584 | /* |
585 | * NB: pool must not be changing on-disk (eg, from zdb or sync context). | |
586 | */ | |
587 | int | |
428870ff BB |
588 | traverse_pool(spa_t *spa, uint64_t txg_start, int flags, |
589 | blkptr_cb_t func, void *arg) | |
34dc7c2f | 590 | { |
428870ff | 591 | int err, lasterr = 0; |
b128c09f BB |
592 | uint64_t obj; |
593 | dsl_pool_t *dp = spa_get_dsl(spa); | |
594 | objset_t *mos = dp->dp_meta_objset; | |
428870ff | 595 | boolean_t hard = (flags & TRAVERSE_HARD); |
b128c09f BB |
596 | |
597 | /* visit the MOS */ | |
9ae529ec CS |
598 | err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), |
599 | txg_start, NULL, flags, func, arg); | |
13fe0198 | 600 | if (err != 0) |
b128c09f BB |
601 | return (err); |
602 | ||
603 | /* visit each dataset */ | |
428870ff BB |
604 | for (obj = 1; err == 0 || (err != ESRCH && hard); |
605 | err = dmu_object_next(mos, &obj, FALSE, txg_start)) { | |
b128c09f BB |
606 | dmu_object_info_t doi; |
607 | ||
608 | err = dmu_object_info(mos, obj, &doi); | |
13fe0198 | 609 | if (err != 0) { |
428870ff BB |
610 | if (!hard) |
611 | return (err); | |
612 | lasterr = err; | |
613 | continue; | |
614 | } | |
b128c09f BB |
615 | |
616 | if (doi.doi_type == DMU_OT_DSL_DATASET) { | |
617 | dsl_dataset_t *ds; | |
428870ff BB |
618 | uint64_t txg = txg_start; |
619 | ||
13fe0198 | 620 | dsl_pool_config_enter(dp, FTAG); |
b128c09f | 621 | err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds); |
13fe0198 MA |
622 | dsl_pool_config_exit(dp, FTAG); |
623 | if (err != 0) { | |
428870ff BB |
624 | if (!hard) |
625 | return (err); | |
626 | lasterr = err; | |
627 | continue; | |
628 | } | |
629 | if (ds->ds_phys->ds_prev_snap_txg > txg) | |
630 | txg = ds->ds_phys->ds_prev_snap_txg; | |
631 | err = traverse_dataset(ds, txg, flags, func, arg); | |
b128c09f | 632 | dsl_dataset_rele(ds, FTAG); |
13fe0198 | 633 | if (err != 0) { |
428870ff BB |
634 | if (!hard) |
635 | return (err); | |
636 | lasterr = err; | |
637 | } | |
b128c09f | 638 | } |
34dc7c2f | 639 | } |
b128c09f BB |
640 | if (err == ESRCH) |
641 | err = 0; | |
428870ff | 642 | return (err != 0 ? err : lasterr); |
34dc7c2f | 643 | } |
c28b2279 BB |
644 | |
645 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
646 | EXPORT_SYMBOL(traverse_dataset); | |
647 | EXPORT_SYMBOL(traverse_pool); | |
c409e464 BB |
648 | |
649 | module_param(zfs_pd_blks_max, int, 0644); | |
650 | MODULE_PARM_DESC(zfs_pd_blks_max, "Max number of blocks to prefetch"); | |
c28b2279 | 651 | #endif |