1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2015 Chunwei Chen. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/dmu_objset.h>
29#include <sys/dmu_traverse.h>
30#include <sys/dsl_dataset.h>
31#include <sys/dsl_dir.h>
32#include <sys/dsl_pool.h>
33#include <sys/dnode.h>
34#include <sys/spa.h>
35#include <sys/spa_impl.h>
36#include <sys/zio.h>
37#include <sys/dmu_impl.h>
38#include <sys/sa.h>
39#include <sys/sa_impl.h>
40#include <sys/callb.h>
41#include <sys/zfeature.h>
42
43int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */
44boolean_t send_holes_without_birth_time = B_TRUE;
45
46#ifdef _KERNEL
47SYSCTL_DECL(_vfs_zfs);
48SYSCTL_UINT(_vfs_zfs, OID_AUTO, send_holes_without_birth_time, CTLFLAG_RWTUN,
49 &send_holes_without_birth_time, 0, "Send holes without birth time");
50#endif
51
52typedef struct prefetch_data {
53 kmutex_t pd_mtx;
54 kcondvar_t pd_cv;
55 int32_t pd_bytes_fetched;
56 int pd_flags;
57 boolean_t pd_cancel;
58 boolean_t pd_exited;
59 zbookmark_phys_t pd_resume;
60} prefetch_data_t;
61
62typedef struct traverse_data {
63 spa_t *td_spa;
64 uint64_t td_objset;
65 blkptr_t *td_rootbp;
66 uint64_t td_min_txg;
67 zbookmark_phys_t *td_resume;
68 int td_flags;
69 prefetch_data_t *td_pfd;
70 boolean_t td_paused;
71 uint64_t td_hole_birth_enabled_txg;
72 blkptr_cb_t *td_func;
73 void *td_arg;
74 boolean_t td_realloc_possible;
75} traverse_data_t;
76
77static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
78 uint64_t objset, uint64_t object);
79static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
80 uint64_t objset, uint64_t object);
81
82static int
83traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
84{
85 traverse_data_t *td = arg;
86 zbookmark_phys_t zb;
87
88 if (BP_IS_HOLE(bp))
89 return (0);
90
91 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa))
92 return (-1);
93
94 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
95 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
96
97 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
98
99 return (0);
100}
101
102static int
103traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
104{
105 traverse_data_t *td = arg;
106
107 if (lrc->lrc_txtype == TX_WRITE) {
108 lr_write_t *lr = (lr_write_t *)lrc;
109 blkptr_t *bp = &lr->lr_blkptr;
110 zbookmark_phys_t zb;
111
112 if (BP_IS_HOLE(bp))
113 return (0);
114
115 if (claim_txg == 0 || bp->blk_birth < claim_txg)
116 return (0);
117
118 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
119 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
120
121 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
122 td->td_arg);
123 }
124 return (0);
125}
126
127static void
128traverse_zil(traverse_data_t *td, zil_header_t *zh)
129{
130 uint64_t claim_txg = zh->zh_claim_txg;
131
132 /*
133 * We only want to visit blocks that have been claimed but not yet
134 * replayed; plus blocks that are already stable in read-only mode.
135 */
136 if (claim_txg == 0 && spa_writeable(td->td_spa))
137 return;
138
139 zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
140 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
141 claim_txg);
142 zil_free(zilog);
143}
144
145typedef enum resume_skip {
146 RESUME_SKIP_ALL,
147 RESUME_SKIP_NONE,
148 RESUME_SKIP_CHILDREN
149} resume_skip_t;
150
151/*
152 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
153 * the block indicated by zb does not need to be visited at all. Returns
154 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
155 * resume point. This indicates that this block should be visited but not its
156 * children (since they must have been visited in a previous traversal).
157 * Otherwise returns RESUME_SKIP_NONE.
158 */
159static resume_skip_t
160resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
161 const zbookmark_phys_t *zb)
162{
163 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
164 /*
165 * If we already visited this bp & everything below,
166 * don't bother doing it again.
167 */
168 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
169 return (RESUME_SKIP_ALL);
170
171 /*
172 * If we found the block we're trying to resume from, zero
173 * the bookmark out to indicate that we have resumed.
174 */
175 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
176 bzero(td->td_resume, sizeof (*zb));
177 if (td->td_flags & TRAVERSE_POST)
178 return (RESUME_SKIP_CHILDREN);
179 }
180 }
181 return (RESUME_SKIP_NONE);
182}
183
184static void
185traverse_prefetch_metadata(traverse_data_t *td,
186 const blkptr_t *bp, const zbookmark_phys_t *zb)
187{
188 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
189
190 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
191 return;
192 /*
193 * If we are in the process of resuming, don't prefetch, because
194 * some children will not be needed (and in fact may have already
195 * been freed).
196 */
197 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
198 return;
199 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
200 return;
201 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
202 return;
203
204 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
205 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
206}
207
208static boolean_t
209prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
210{
211 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
212 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
213 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
214 return (B_FALSE);
215 return (B_TRUE);
216}
217
218static int
219traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
220 const blkptr_t *bp, const zbookmark_phys_t *zb)
221{
222 zbookmark_phys_t czb;
223 int err = 0;
224 arc_buf_t *buf = NULL;
225 prefetch_data_t *pd = td->td_pfd;
226 boolean_t hard = td->td_flags & TRAVERSE_HARD;
227
228 switch (resume_skip_check(td, dnp, zb)) {
229 case RESUME_SKIP_ALL:
230 return (0);
231 case RESUME_SKIP_CHILDREN:
232 goto post;
233 case RESUME_SKIP_NONE:
234 break;
235 default:
236 ASSERT(0);
237 }
238
239 if (bp->blk_birth == 0) {
240 /*
241 * Since this block has a birth time of 0 it must be one of
242 * two things: a hole created before the
243 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
244 * which has always been a hole in an object.
245 *
246 * If a file is written sparsely, then the unwritten parts of
247 * the file were "always holes" -- that is, they have been
248 * holes since this object was allocated. However, we (and
249 * our callers) can not necessarily tell when an object was
250 * allocated. Therefore, if it's possible that this object
251 * was freed and then its object number reused, we need to
252 * visit all the holes with birth==0.
253 *
254 * If it isn't possible that the object number was reused,
255 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
256 * all the blocks we will visit as part of this traversal,
257 * then this hole must have always existed, so we can skip
258 * it. We visit blocks born after (exclusive) td_min_txg.
259 *
260 * Note that the meta-dnode cannot be reallocated.
261 */
262 if (!send_holes_without_birth_time &&
263 (!td->td_realloc_possible ||
264 zb->zb_object == DMU_META_DNODE_OBJECT) &&
265 td->td_hole_birth_enabled_txg <= td->td_min_txg)
266 return (0);
267 } else if (bp->blk_birth <= td->td_min_txg) {
268 return (0);
269 }
270
271 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
272 uint64_t size = BP_GET_LSIZE(bp);
273 mutex_enter(&pd->pd_mtx);
274 ASSERT(pd->pd_bytes_fetched >= 0);
275 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
276 cv_wait(&pd->pd_cv, &pd->pd_mtx);
277 pd->pd_bytes_fetched -= size;
278 cv_broadcast(&pd->pd_cv);
279 mutex_exit(&pd->pd_mtx);
280 }
281
282 if (BP_IS_HOLE(bp)) {
283 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
284 if (err != 0)
285 goto post;
286 return (0);
287 }
288
289 if (td->td_flags & TRAVERSE_PRE) {
290 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
291 td->td_arg);
292 if (err == TRAVERSE_VISIT_NO_CHILDREN)
293 return (0);
294 if (err != 0)
295 goto post;
296 }
297
298 if (BP_GET_LEVEL(bp) > 0) {
299 arc_flags_t flags = ARC_FLAG_WAIT;
300 int i;
301 blkptr_t *cbp;
302 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
303
304 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
305 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
306 if (err != 0)
307 goto post;
308 cbp = buf->b_data;
309
310 for (i = 0; i < epb; i++) {
311 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
312 zb->zb_level - 1,
313 zb->zb_blkid * epb + i);
314 traverse_prefetch_metadata(td, &cbp[i], &czb);
315 }
316
317 /* recursively visitbp() blocks below this */
318 for (i = 0; i < epb; i++) {
319 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
320 zb->zb_level - 1,
321 zb->zb_blkid * epb + i);
322 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
323 if (err != 0)
324 break;
325 }
326 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
327 arc_flags_t flags = ARC_FLAG_WAIT;
328 int i;
329 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
330
331 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
332 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
333 if (err != 0)
334 goto post;
335 dnode_phys_t *child_dnp = buf->b_data;
336
337 for (i = 0; i < epb; i++) {
338 prefetch_dnode_metadata(td, &child_dnp[i],
339 zb->zb_objset, zb->zb_blkid * epb + i);
340 }
341
342 /* recursively visitbp() blocks below this */
343 for (i = 0; i < epb; i++) {
344 err = traverse_dnode(td, &child_dnp[i],
345 zb->zb_objset, zb->zb_blkid * epb + i);
346 if (err != 0)
347 break;
348 }
349 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
350 arc_flags_t flags = ARC_FLAG_WAIT;
351
352 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
353 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
354 if (err != 0)
355 goto post;
356
357 objset_phys_t *osp = buf->b_data;
358 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
359 DMU_META_DNODE_OBJECT);
360 /*
361 * See the block comment above for the goal of this variable.
362 * If the maxblkid of the meta-dnode is 0, then we know that
363 * we've never had more than DNODES_PER_BLOCK objects in the
364 * dataset, which means we can't have reused any object ids.
365 */
366 if (osp->os_meta_dnode.dn_maxblkid == 0)
367 td->td_realloc_possible = B_FALSE;
368
369 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
370 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
371 zb->zb_objset, DMU_GROUPUSED_OBJECT);
372 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
373 zb->zb_objset, DMU_USERUSED_OBJECT);
374 }
375
376 err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
377 DMU_META_DNODE_OBJECT);
378 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
379 err = traverse_dnode(td, &osp->os_groupused_dnode,
380 zb->zb_objset, DMU_GROUPUSED_OBJECT);
381 }
382 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
383 err = traverse_dnode(td, &osp->os_userused_dnode,
384 zb->zb_objset, DMU_USERUSED_OBJECT);
385 }
386 }
387
388 if (buf)
389 arc_buf_destroy(buf, &buf);
390
391post:
392 if (err == 0 && (td->td_flags & TRAVERSE_POST))
393 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
394
395 if (hard && (err == EIO || err == ECKSUM)) {
396 /*
397 * Ignore this disk error as requested by the HARD flag,
398 * and continue traversal.
399 */
400 err = 0;
401 }
402
403 /*
404 * If we are stopping here, set td_resume.
405 */
406 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
407 td->td_resume->zb_objset = zb->zb_objset;
408 td->td_resume->zb_object = zb->zb_object;
409 td->td_resume->zb_level = 0;
410 /*
411 * If we have stopped on an indirect block (e.g. due to
412 * i/o error), we have not visited anything below it.
413 * Set the bookmark to the first level-0 block that we need
414 * to visit. This way, the resuming code does not need to
415 * deal with resuming from indirect blocks.
416 *
417 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
418 * to dereference it.
419 */
420 td->td_resume->zb_blkid = zb->zb_blkid;
421 if (zb->zb_level > 0) {
422 td->td_resume->zb_blkid <<= zb->zb_level *
423 (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
424 }
425 td->td_paused = B_TRUE;
426 }
427
428 return (err);
429}
430
431static void
432prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
433 uint64_t objset, uint64_t object)
434{
435 int j;
436 zbookmark_phys_t czb;
437
438 for (j = 0; j < dnp->dn_nblkptr; j++) {
439 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
440 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
441 }
442
443 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
444 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
445 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
446 }
447}
448
449static int
450traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
451 uint64_t objset, uint64_t object)
452{
453 int j, err = 0;
454 zbookmark_phys_t czb;
455
456 if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
457 object < td->td_resume->zb_object)
458 return (0);
459
460 if (td->td_flags & TRAVERSE_PRE) {
461 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
462 ZB_DNODE_BLKID);
463 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
464 td->td_arg);
465 if (err == TRAVERSE_VISIT_NO_CHILDREN)
466 return (0);
467 if (err != 0)
468 return (err);
469 }
470
471 for (j = 0; j < dnp->dn_nblkptr; j++) {
472 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
473 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
474 if (err != 0)
475 break;
476 }
477
478 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
479 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
480 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
481 }
482
483 if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
484 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
485 ZB_DNODE_BLKID);
486 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
487 td->td_arg);
488 if (err == TRAVERSE_VISIT_NO_CHILDREN)
489 return (0);
490 if (err != 0)
491 return (err);
492 }
493 return (err);
494}
495
496/* ARGSUSED */
497static int
498traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
499 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
500{
501 prefetch_data_t *pfd = arg;
| 1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2015 Chunwei Chen. All rights reserved.
25 */
26
27#include <sys/zfs_context.h>
28#include <sys/dmu_objset.h>
29#include <sys/dmu_traverse.h>
30#include <sys/dsl_dataset.h>
31#include <sys/dsl_dir.h>
32#include <sys/dsl_pool.h>
33#include <sys/dnode.h>
34#include <sys/spa.h>
35#include <sys/spa_impl.h>
36#include <sys/zio.h>
37#include <sys/dmu_impl.h>
38#include <sys/sa.h>
39#include <sys/sa_impl.h>
40#include <sys/callb.h>
41#include <sys/zfeature.h>
42
43int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */
44boolean_t send_holes_without_birth_time = B_TRUE;
45
46#ifdef _KERNEL
47SYSCTL_DECL(_vfs_zfs);
48SYSCTL_UINT(_vfs_zfs, OID_AUTO, send_holes_without_birth_time, CTLFLAG_RWTUN,
49 &send_holes_without_birth_time, 0, "Send holes without birth time");
50#endif
51
52typedef struct prefetch_data {
53 kmutex_t pd_mtx;
54 kcondvar_t pd_cv;
55 int32_t pd_bytes_fetched;
56 int pd_flags;
57 boolean_t pd_cancel;
58 boolean_t pd_exited;
59 zbookmark_phys_t pd_resume;
60} prefetch_data_t;
61
62typedef struct traverse_data {
63 spa_t *td_spa;
64 uint64_t td_objset;
65 blkptr_t *td_rootbp;
66 uint64_t td_min_txg;
67 zbookmark_phys_t *td_resume;
68 int td_flags;
69 prefetch_data_t *td_pfd;
70 boolean_t td_paused;
71 uint64_t td_hole_birth_enabled_txg;
72 blkptr_cb_t *td_func;
73 void *td_arg;
74 boolean_t td_realloc_possible;
75} traverse_data_t;
76
77static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
78 uint64_t objset, uint64_t object);
79static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
80 uint64_t objset, uint64_t object);
81
82static int
83traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
84{
85 traverse_data_t *td = arg;
86 zbookmark_phys_t zb;
87
88 if (BP_IS_HOLE(bp))
89 return (0);
90
91 if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(td->td_spa))
92 return (-1);
93
94 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
95 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
96
97 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
98
99 return (0);
100}
101
102static int
103traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
104{
105 traverse_data_t *td = arg;
106
107 if (lrc->lrc_txtype == TX_WRITE) {
108 lr_write_t *lr = (lr_write_t *)lrc;
109 blkptr_t *bp = &lr->lr_blkptr;
110 zbookmark_phys_t zb;
111
112 if (BP_IS_HOLE(bp))
113 return (0);
114
115 if (claim_txg == 0 || bp->blk_birth < claim_txg)
116 return (0);
117
118 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
119 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
120
121 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
122 td->td_arg);
123 }
124 return (0);
125}
126
127static void
128traverse_zil(traverse_data_t *td, zil_header_t *zh)
129{
130 uint64_t claim_txg = zh->zh_claim_txg;
131
132 /*
133 * We only want to visit blocks that have been claimed but not yet
134 * replayed; plus blocks that are already stable in read-only mode.
135 */
136 if (claim_txg == 0 && spa_writeable(td->td_spa))
137 return;
138
139 zilog_t *zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
140 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
141 claim_txg);
142 zil_free(zilog);
143}
144
145typedef enum resume_skip {
146 RESUME_SKIP_ALL,
147 RESUME_SKIP_NONE,
148 RESUME_SKIP_CHILDREN
149} resume_skip_t;
150
151/*
152 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
153 * the block indicated by zb does not need to be visited at all. Returns
154 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
155 * resume point. This indicates that this block should be visited but not its
156 * children (since they must have been visited in a previous traversal).
157 * Otherwise returns RESUME_SKIP_NONE.
158 */
159static resume_skip_t
160resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
161 const zbookmark_phys_t *zb)
162{
163 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
164 /*
165 * If we already visited this bp & everything below,
166 * don't bother doing it again.
167 */
168 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
169 return (RESUME_SKIP_ALL);
170
171 /*
172 * If we found the block we're trying to resume from, zero
173 * the bookmark out to indicate that we have resumed.
174 */
175 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
176 bzero(td->td_resume, sizeof (*zb));
177 if (td->td_flags & TRAVERSE_POST)
178 return (RESUME_SKIP_CHILDREN);
179 }
180 }
181 return (RESUME_SKIP_NONE);
182}
183
184static void
185traverse_prefetch_metadata(traverse_data_t *td,
186 const blkptr_t *bp, const zbookmark_phys_t *zb)
187{
188 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
189
190 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
191 return;
192 /*
193 * If we are in the process of resuming, don't prefetch, because
194 * some children will not be needed (and in fact may have already
195 * been freed).
196 */
197 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
198 return;
199 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
200 return;
201 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
202 return;
203
204 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
205 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
206}
207
208static boolean_t
209prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
210{
211 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
212 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
213 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
214 return (B_FALSE);
215 return (B_TRUE);
216}
217
218static int
219traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
220 const blkptr_t *bp, const zbookmark_phys_t *zb)
221{
222 zbookmark_phys_t czb;
223 int err = 0;
224 arc_buf_t *buf = NULL;
225 prefetch_data_t *pd = td->td_pfd;
226 boolean_t hard = td->td_flags & TRAVERSE_HARD;
227
228 switch (resume_skip_check(td, dnp, zb)) {
229 case RESUME_SKIP_ALL:
230 return (0);
231 case RESUME_SKIP_CHILDREN:
232 goto post;
233 case RESUME_SKIP_NONE:
234 break;
235 default:
236 ASSERT(0);
237 }
238
239 if (bp->blk_birth == 0) {
240 /*
241 * Since this block has a birth time of 0 it must be one of
242 * two things: a hole created before the
243 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
244 * which has always been a hole in an object.
245 *
246 * If a file is written sparsely, then the unwritten parts of
247 * the file were "always holes" -- that is, they have been
248 * holes since this object was allocated. However, we (and
249 * our callers) can not necessarily tell when an object was
250 * allocated. Therefore, if it's possible that this object
251 * was freed and then its object number reused, we need to
252 * visit all the holes with birth==0.
253 *
254 * If it isn't possible that the object number was reused,
255 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
256 * all the blocks we will visit as part of this traversal,
257 * then this hole must have always existed, so we can skip
258 * it. We visit blocks born after (exclusive) td_min_txg.
259 *
260 * Note that the meta-dnode cannot be reallocated.
261 */
262 if (!send_holes_without_birth_time &&
263 (!td->td_realloc_possible ||
264 zb->zb_object == DMU_META_DNODE_OBJECT) &&
265 td->td_hole_birth_enabled_txg <= td->td_min_txg)
266 return (0);
267 } else if (bp->blk_birth <= td->td_min_txg) {
268 return (0);
269 }
270
271 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
272 uint64_t size = BP_GET_LSIZE(bp);
273 mutex_enter(&pd->pd_mtx);
274 ASSERT(pd->pd_bytes_fetched >= 0);
275 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
276 cv_wait(&pd->pd_cv, &pd->pd_mtx);
277 pd->pd_bytes_fetched -= size;
278 cv_broadcast(&pd->pd_cv);
279 mutex_exit(&pd->pd_mtx);
280 }
281
282 if (BP_IS_HOLE(bp)) {
283 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
284 if (err != 0)
285 goto post;
286 return (0);
287 }
288
289 if (td->td_flags & TRAVERSE_PRE) {
290 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
291 td->td_arg);
292 if (err == TRAVERSE_VISIT_NO_CHILDREN)
293 return (0);
294 if (err != 0)
295 goto post;
296 }
297
298 if (BP_GET_LEVEL(bp) > 0) {
299 arc_flags_t flags = ARC_FLAG_WAIT;
300 int i;
301 blkptr_t *cbp;
302 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
303
304 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
305 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
306 if (err != 0)
307 goto post;
308 cbp = buf->b_data;
309
310 for (i = 0; i < epb; i++) {
311 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
312 zb->zb_level - 1,
313 zb->zb_blkid * epb + i);
314 traverse_prefetch_metadata(td, &cbp[i], &czb);
315 }
316
317 /* recursively visitbp() blocks below this */
318 for (i = 0; i < epb; i++) {
319 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
320 zb->zb_level - 1,
321 zb->zb_blkid * epb + i);
322 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
323 if (err != 0)
324 break;
325 }
326 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
327 arc_flags_t flags = ARC_FLAG_WAIT;
328 int i;
329 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
330
331 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
332 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
333 if (err != 0)
334 goto post;
335 dnode_phys_t *child_dnp = buf->b_data;
336
337 for (i = 0; i < epb; i++) {
338 prefetch_dnode_metadata(td, &child_dnp[i],
339 zb->zb_objset, zb->zb_blkid * epb + i);
340 }
341
342 /* recursively visitbp() blocks below this */
343 for (i = 0; i < epb; i++) {
344 err = traverse_dnode(td, &child_dnp[i],
345 zb->zb_objset, zb->zb_blkid * epb + i);
346 if (err != 0)
347 break;
348 }
349 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
350 arc_flags_t flags = ARC_FLAG_WAIT;
351
352 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
353 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
354 if (err != 0)
355 goto post;
356
357 objset_phys_t *osp = buf->b_data;
358 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
359 DMU_META_DNODE_OBJECT);
360 /*
361 * See the block comment above for the goal of this variable.
362 * If the maxblkid of the meta-dnode is 0, then we know that
363 * we've never had more than DNODES_PER_BLOCK objects in the
364 * dataset, which means we can't have reused any object ids.
365 */
366 if (osp->os_meta_dnode.dn_maxblkid == 0)
367 td->td_realloc_possible = B_FALSE;
368
369 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
370 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
371 zb->zb_objset, DMU_GROUPUSED_OBJECT);
372 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
373 zb->zb_objset, DMU_USERUSED_OBJECT);
374 }
375
376 err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
377 DMU_META_DNODE_OBJECT);
378 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
379 err = traverse_dnode(td, &osp->os_groupused_dnode,
380 zb->zb_objset, DMU_GROUPUSED_OBJECT);
381 }
382 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
383 err = traverse_dnode(td, &osp->os_userused_dnode,
384 zb->zb_objset, DMU_USERUSED_OBJECT);
385 }
386 }
387
388 if (buf)
389 arc_buf_destroy(buf, &buf);
390
391post:
392 if (err == 0 && (td->td_flags & TRAVERSE_POST))
393 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
394
395 if (hard && (err == EIO || err == ECKSUM)) {
396 /*
397 * Ignore this disk error as requested by the HARD flag,
398 * and continue traversal.
399 */
400 err = 0;
401 }
402
403 /*
404 * If we are stopping here, set td_resume.
405 */
406 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
407 td->td_resume->zb_objset = zb->zb_objset;
408 td->td_resume->zb_object = zb->zb_object;
409 td->td_resume->zb_level = 0;
410 /*
411 * If we have stopped on an indirect block (e.g. due to
412 * i/o error), we have not visited anything below it.
413 * Set the bookmark to the first level-0 block that we need
414 * to visit. This way, the resuming code does not need to
415 * deal with resuming from indirect blocks.
416 *
417 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
418 * to dereference it.
419 */
420 td->td_resume->zb_blkid = zb->zb_blkid;
421 if (zb->zb_level > 0) {
422 td->td_resume->zb_blkid <<= zb->zb_level *
423 (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
424 }
425 td->td_paused = B_TRUE;
426 }
427
428 return (err);
429}
430
431static void
432prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
433 uint64_t objset, uint64_t object)
434{
435 int j;
436 zbookmark_phys_t czb;
437
438 for (j = 0; j < dnp->dn_nblkptr; j++) {
439 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
440 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
441 }
442
443 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
444 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
445 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
446 }
447}
448
449static int
450traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
451 uint64_t objset, uint64_t object)
452{
453 int j, err = 0;
454 zbookmark_phys_t czb;
455
456 if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
457 object < td->td_resume->zb_object)
458 return (0);
459
460 if (td->td_flags & TRAVERSE_PRE) {
461 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
462 ZB_DNODE_BLKID);
463 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
464 td->td_arg);
465 if (err == TRAVERSE_VISIT_NO_CHILDREN)
466 return (0);
467 if (err != 0)
468 return (err);
469 }
470
471 for (j = 0; j < dnp->dn_nblkptr; j++) {
472 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
473 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
474 if (err != 0)
475 break;
476 }
477
478 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
479 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
480 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
481 }
482
483 if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
484 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
485 ZB_DNODE_BLKID);
486 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
487 td->td_arg);
488 if (err == TRAVERSE_VISIT_NO_CHILDREN)
489 return (0);
490 if (err != 0)
491 return (err);
492 }
493 return (err);
494}
495
496/* ARGSUSED */
497static int
498traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
499 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
500{
501 prefetch_data_t *pfd = arg;
|
504 ASSERT(pfd->pd_bytes_fetched >= 0);
505 if (bp == NULL)
506 return (0);
507 if (pfd->pd_cancel)
508 return (SET_ERROR(EINTR));
509
510 if (!prefetch_needed(pfd, bp))
511 return (0);
512
513 mutex_enter(&pfd->pd_mtx);
514 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
515 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
516 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
517 cv_broadcast(&pfd->pd_cv);
518 mutex_exit(&pfd->pd_mtx);
519
520 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
521 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
522
523 return (0);
524}
525
526static void
527traverse_prefetch_thread(void *arg)
528{
529 traverse_data_t *td_main = arg;
530 traverse_data_t td = *td_main;
531 zbookmark_phys_t czb;
532
533 td.td_func = traverse_prefetcher;
534 td.td_arg = td_main->td_pfd;
535 td.td_pfd = NULL;
536 td.td_resume = &td_main->td_pfd->pd_resume;
537
538 SET_BOOKMARK(&czb, td.td_objset,
539 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
540 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
541
542 mutex_enter(&td_main->td_pfd->pd_mtx);
543 td_main->td_pfd->pd_exited = B_TRUE;
544 cv_broadcast(&td_main->td_pfd->pd_cv);
545 mutex_exit(&td_main->td_pfd->pd_mtx);
546}
547
548/*
549 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
550 * in syncing context).
551 */
552static int
553traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
554 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
555 blkptr_cb_t func, void *arg)
556{
557 traverse_data_t td;
558 prefetch_data_t pd = { 0 };
559 zbookmark_phys_t czb;
560 int err;
561
562 ASSERT(ds == NULL || objset == ds->ds_object);
563 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
564
565 td.td_spa = spa;
566 td.td_objset = objset;
567 td.td_rootbp = rootbp;
568 td.td_min_txg = txg_start;
569 td.td_resume = resume;
570 td.td_func = func;
571 td.td_arg = arg;
572 td.td_pfd = &pd;
573 td.td_flags = flags;
574 td.td_paused = B_FALSE;
575 td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
576
577 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
578 VERIFY(spa_feature_enabled_txg(spa,
579 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
580 } else {
581 td.td_hole_birth_enabled_txg = UINT64_MAX;
582 }
583
584 pd.pd_flags = flags;
585 if (resume != NULL)
586 pd.pd_resume = *resume;
587 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
588 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
589
590 /* See comment on ZIL traversal in dsl_scan_visitds. */
591 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
592 arc_flags_t flags = ARC_FLAG_WAIT;
593 objset_phys_t *osp;
594 arc_buf_t *buf;
595
596 err = arc_read(NULL, td.td_spa, rootbp,
597 arc_getbuf_func, &buf,
598 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
599 if (err != 0)
600 return (err);
601
602 osp = buf->b_data;
603 traverse_zil(&td, &osp->os_zil_header);
604 arc_buf_destroy(buf, &buf);
605 }
606
607 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
608 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
609 &td, TQ_NOQUEUE))
610 pd.pd_exited = B_TRUE;
611
612 SET_BOOKMARK(&czb, td.td_objset,
613 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
614 err = traverse_visitbp(&td, NULL, rootbp, &czb);
615
616 mutex_enter(&pd.pd_mtx);
617 pd.pd_cancel = B_TRUE;
618 cv_broadcast(&pd.pd_cv);
619 while (!pd.pd_exited)
620 cv_wait(&pd.pd_cv, &pd.pd_mtx);
621 mutex_exit(&pd.pd_mtx);
622
623 mutex_destroy(&pd.pd_mtx);
624 cv_destroy(&pd.pd_cv);
625
626 return (err);
627}
628
629/*
630 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
631 * in syncing context).
632 */
633int
634traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
635 zbookmark_phys_t *resume,
636 int flags, blkptr_cb_t func, void *arg)
637{
638 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
639 &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
640}
641
642int
643traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
644 int flags, blkptr_cb_t func, void *arg)
645{
646 return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
647}
648
649int
650traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
651 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
652 blkptr_cb_t func, void *arg)
653{
654 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
655 blkptr, txg_start, resume, flags, func, arg));
656}
657
658/*
659 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
660 */
661int
662traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
663 blkptr_cb_t func, void *arg)
664{
665 int err;
666 dsl_pool_t *dp = spa_get_dsl(spa);
667 objset_t *mos = dp->dp_meta_objset;
668 boolean_t hard = (flags & TRAVERSE_HARD);
669
670 /* visit the MOS */
671 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
672 txg_start, NULL, flags, func, arg);
673 if (err != 0)
674 return (err);
675
676 /* visit each dataset */
677 for (uint64_t obj = 1; err == 0;
678 err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
679 dmu_object_info_t doi;
680
681 err = dmu_object_info(mos, obj, &doi);
682 if (err != 0) {
683 if (hard)
684 continue;
685 break;
686 }
687
688 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
689 dsl_dataset_t *ds;
690 uint64_t txg = txg_start;
691
692 dsl_pool_config_enter(dp, FTAG);
693 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
694 dsl_pool_config_exit(dp, FTAG);
695 if (err != 0) {
696 if (hard)
697 continue;
698 break;
699 }
700 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
701 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
702 err = traverse_dataset(ds, txg, flags, func, arg);
703 dsl_dataset_rele(ds, FTAG);
704 if (err != 0)
705 break;
706 }
707 }
708 if (err == ESRCH)
709 err = 0;
710 return (err);
711}
| 505 ASSERT(pfd->pd_bytes_fetched >= 0);
506 if (bp == NULL)
507 return (0);
508 if (pfd->pd_cancel)
509 return (SET_ERROR(EINTR));
510
511 if (!prefetch_needed(pfd, bp))
512 return (0);
513
514 mutex_enter(&pfd->pd_mtx);
515 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
516 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
517 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
518 cv_broadcast(&pfd->pd_cv);
519 mutex_exit(&pfd->pd_mtx);
520
521 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
522 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
523
524 return (0);
525}
526
527static void
528traverse_prefetch_thread(void *arg)
529{
530 traverse_data_t *td_main = arg;
531 traverse_data_t td = *td_main;
532 zbookmark_phys_t czb;
533
534 td.td_func = traverse_prefetcher;
535 td.td_arg = td_main->td_pfd;
536 td.td_pfd = NULL;
537 td.td_resume = &td_main->td_pfd->pd_resume;
538
539 SET_BOOKMARK(&czb, td.td_objset,
540 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
541 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
542
543 mutex_enter(&td_main->td_pfd->pd_mtx);
544 td_main->td_pfd->pd_exited = B_TRUE;
545 cv_broadcast(&td_main->td_pfd->pd_cv);
546 mutex_exit(&td_main->td_pfd->pd_mtx);
547}
548
549/*
550 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
551 * in syncing context).
552 */
553static int
554traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
555 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
556 blkptr_cb_t func, void *arg)
557{
558 traverse_data_t td;
559 prefetch_data_t pd = { 0 };
560 zbookmark_phys_t czb;
561 int err;
562
563 ASSERT(ds == NULL || objset == ds->ds_object);
564 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
565
566 td.td_spa = spa;
567 td.td_objset = objset;
568 td.td_rootbp = rootbp;
569 td.td_min_txg = txg_start;
570 td.td_resume = resume;
571 td.td_func = func;
572 td.td_arg = arg;
573 td.td_pfd = &pd;
574 td.td_flags = flags;
575 td.td_paused = B_FALSE;
576 td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
577
578 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
579 VERIFY(spa_feature_enabled_txg(spa,
580 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
581 } else {
582 td.td_hole_birth_enabled_txg = UINT64_MAX;
583 }
584
585 pd.pd_flags = flags;
586 if (resume != NULL)
587 pd.pd_resume = *resume;
588 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
589 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
590
591 /* See comment on ZIL traversal in dsl_scan_visitds. */
592 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
593 arc_flags_t flags = ARC_FLAG_WAIT;
594 objset_phys_t *osp;
595 arc_buf_t *buf;
596
597 err = arc_read(NULL, td.td_spa, rootbp,
598 arc_getbuf_func, &buf,
599 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
600 if (err != 0)
601 return (err);
602
603 osp = buf->b_data;
604 traverse_zil(&td, &osp->os_zil_header);
605 arc_buf_destroy(buf, &buf);
606 }
607
608 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
609 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
610 &td, TQ_NOQUEUE))
611 pd.pd_exited = B_TRUE;
612
613 SET_BOOKMARK(&czb, td.td_objset,
614 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
615 err = traverse_visitbp(&td, NULL, rootbp, &czb);
616
617 mutex_enter(&pd.pd_mtx);
618 pd.pd_cancel = B_TRUE;
619 cv_broadcast(&pd.pd_cv);
620 while (!pd.pd_exited)
621 cv_wait(&pd.pd_cv, &pd.pd_mtx);
622 mutex_exit(&pd.pd_mtx);
623
624 mutex_destroy(&pd.pd_mtx);
625 cv_destroy(&pd.pd_cv);
626
627 return (err);
628}
629
630/*
631 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
632 * in syncing context).
633 */
634int
635traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
636 zbookmark_phys_t *resume,
637 int flags, blkptr_cb_t func, void *arg)
638{
639 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
640 &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
641}
642
643int
644traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
645 int flags, blkptr_cb_t func, void *arg)
646{
647 return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
648}
649
650int
651traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
652 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
653 blkptr_cb_t func, void *arg)
654{
655 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
656 blkptr, txg_start, resume, flags, func, arg));
657}
658
659/*
660 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
661 */
662int
663traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
664 blkptr_cb_t func, void *arg)
665{
666 int err;
667 dsl_pool_t *dp = spa_get_dsl(spa);
668 objset_t *mos = dp->dp_meta_objset;
669 boolean_t hard = (flags & TRAVERSE_HARD);
670
671 /* visit the MOS */
672 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
673 txg_start, NULL, flags, func, arg);
674 if (err != 0)
675 return (err);
676
677 /* visit each dataset */
678 for (uint64_t obj = 1; err == 0;
679 err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
680 dmu_object_info_t doi;
681
682 err = dmu_object_info(mos, obj, &doi);
683 if (err != 0) {
684 if (hard)
685 continue;
686 break;
687 }
688
689 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
690 dsl_dataset_t *ds;
691 uint64_t txg = txg_start;
692
693 dsl_pool_config_enter(dp, FTAG);
694 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
695 dsl_pool_config_exit(dp, FTAG);
696 if (err != 0) {
697 if (hard)
698 continue;
699 break;
700 }
701 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
702 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
703 err = traverse_dataset(ds, txg, flags, func, arg);
704 dsl_dataset_rele(ds, FTAG);
705 if (err != 0)
706 break;
707 }
708 }
709 if (err == ESRCH)
710 err = 0;
711 return (err);
712}
|