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.
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.
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]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2015 Chunwei Chen. All rights reserved.
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>
36 #include <sys/dmu_impl.h>
38 #include <sys/sa_impl.h>
39 #include <sys/callb.h>
40 #include <sys/zfeature.h>
42 int32_t zfs_pd_bytes_max = 50 * 1024 * 1024; /* 50MB */
44 typedef struct prefetch_data {
47 int32_t pd_bytes_fetched;
53 typedef struct traverse_data {
58 zbookmark_phys_t *td_resume;
60 prefetch_data_t *td_pfd;
62 uint64_t td_hole_birth_enabled_txg;
67 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
68 uint64_t objset, uint64_t object);
69 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
70 uint64_t objset, uint64_t object);
73 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
75 traverse_data_t *td = arg;
81 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
84 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
85 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
87 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
93 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
95 traverse_data_t *td = arg;
97 if (lrc->lrc_txtype == TX_WRITE) {
98 lr_write_t *lr = (lr_write_t *)lrc;
99 blkptr_t *bp = &lr->lr_blkptr;
105 if (claim_txg == 0 || bp->blk_birth < claim_txg)
108 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
109 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
111 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
118 traverse_zil(traverse_data_t *td, zil_header_t *zh)
120 uint64_t claim_txg = zh->zh_claim_txg;
124 * We only want to visit blocks that have been claimed but not yet
125 * replayed; plus, in read-only mode, blocks that are already stable.
127 if (claim_txg == 0 && spa_writeable(td->td_spa))
130 zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
132 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
138 typedef enum resume_skip {
145 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
146 * the block indicated by zb does not need to be visited at all. Returns
147 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
148 * resume point. This indicates that this block should be visited but not its
149 * children (since they must have been visited in a previous traversal).
150 * Otherwise returns RESUME_SKIP_NONE.
153 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
154 const zbookmark_phys_t *zb)
156 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
158 * If we already visited this bp & everything below,
159 * don't bother doing it again.
161 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
162 return (RESUME_SKIP_ALL);
165 * If we found the block we're trying to resume from, zero
166 * the bookmark out to indicate that we have resumed.
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);
174 return (RESUME_SKIP_NONE);
178 traverse_prefetch_metadata(traverse_data_t *td,
179 const blkptr_t *bp, const zbookmark_phys_t *zb)
181 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
183 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
186 * If we are in the process of resuming, don't prefetch, because
187 * some children will not be needed (and in fact may have already
190 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
192 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
194 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
197 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
198 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
202 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
204 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
205 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
206 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
212 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
213 const blkptr_t *bp, const zbookmark_phys_t *zb)
215 zbookmark_phys_t czb;
217 arc_buf_t *buf = NULL;
218 prefetch_data_t *pd = td->td_pfd;
219 boolean_t hard = td->td_flags & TRAVERSE_HARD;
221 switch (resume_skip_check(td, dnp, zb)) {
222 case RESUME_SKIP_ALL:
224 case RESUME_SKIP_CHILDREN:
226 case RESUME_SKIP_NONE:
232 if (bp->blk_birth == 0) {
234 * Since this block has a birth time of 0 it must be a
235 * hole created before the SPA_FEATURE_HOLE_BIRTH
236 * feature was enabled. If SPA_FEATURE_HOLE_BIRTH
237 * was enabled before the min_txg for this traveral we
238 * know the hole must have been created before the
239 * min_txg for this traveral, so we can skip it. If
240 * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg
241 * for this traveral we cannot tell if the hole was
242 * created before or after the min_txg for this
243 * traversal, so we cannot skip it.
245 if (td->td_hole_birth_enabled_txg < td->td_min_txg)
247 } else if (bp->blk_birth <= td->td_min_txg) {
251 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
252 uint64_t size = BP_GET_LSIZE(bp);
253 mutex_enter(&pd->pd_mtx);
254 ASSERT(pd->pd_bytes_fetched >= 0);
255 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
256 cv_wait(&pd->pd_cv, &pd->pd_mtx);
257 pd->pd_bytes_fetched -= size;
258 cv_broadcast(&pd->pd_cv);
259 mutex_exit(&pd->pd_mtx);
262 if (BP_IS_HOLE(bp)) {
263 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
269 if (td->td_flags & TRAVERSE_PRE) {
270 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
272 if (err == TRAVERSE_VISIT_NO_CHILDREN)
278 if (BP_GET_LEVEL(bp) > 0) {
279 arc_flags_t flags = ARC_FLAG_WAIT;
282 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
284 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
285 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
290 for (i = 0; i < epb; i++) {
291 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
293 zb->zb_blkid * epb + i);
294 traverse_prefetch_metadata(td, &cbp[i], &czb);
297 /* recursively visitbp() blocks below this */
298 for (i = 0; i < epb; i++) {
299 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
301 zb->zb_blkid * epb + i);
302 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
306 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
307 arc_flags_t flags = ARC_FLAG_WAIT;
309 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
312 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
313 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
318 for (i = 0; i < epb; i++) {
319 prefetch_dnode_metadata(td, &cdnp[i], zb->zb_objset,
320 zb->zb_blkid * epb + i);
323 /* recursively visitbp() blocks below this */
324 for (i = 0; i < epb; i++) {
325 err = traverse_dnode(td, &cdnp[i], zb->zb_objset,
326 zb->zb_blkid * epb + i);
330 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
331 arc_flags_t flags = ARC_FLAG_WAIT;
333 dnode_phys_t *mdnp, *gdnp, *udnp;
335 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
336 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
341 mdnp = &osp->os_meta_dnode;
342 gdnp = &osp->os_groupused_dnode;
343 udnp = &osp->os_userused_dnode;
345 prefetch_dnode_metadata(td, mdnp, zb->zb_objset,
346 DMU_META_DNODE_OBJECT);
347 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
348 prefetch_dnode_metadata(td, gdnp, zb->zb_objset,
349 DMU_GROUPUSED_OBJECT);
350 prefetch_dnode_metadata(td, udnp, zb->zb_objset,
351 DMU_USERUSED_OBJECT);
354 err = traverse_dnode(td, mdnp, zb->zb_objset,
355 DMU_META_DNODE_OBJECT);
356 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
357 err = traverse_dnode(td, gdnp, zb->zb_objset,
358 DMU_GROUPUSED_OBJECT);
360 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
361 err = traverse_dnode(td, udnp, zb->zb_objset,
362 DMU_USERUSED_OBJECT);
367 (void) arc_buf_remove_ref(buf, &buf);
370 if (err == 0 && (td->td_flags & TRAVERSE_POST))
371 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
373 if (hard && (err == EIO || err == ECKSUM)) {
375 * Ignore this disk error as requested by the HARD flag,
376 * and continue traversal.
382 * If we are stopping here, set td_resume.
384 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
385 td->td_resume->zb_objset = zb->zb_objset;
386 td->td_resume->zb_object = zb->zb_object;
387 td->td_resume->zb_level = 0;
389 * If we have stopped on an indirect block (e.g. due to
390 * i/o error), we have not visited anything below it.
391 * Set the bookmark to the first level-0 block that we need
392 * to visit. This way, the resuming code does not need to
393 * deal with resuming from indirect blocks.
395 td->td_resume->zb_blkid = zb->zb_blkid <<
396 (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
397 td->td_paused = B_TRUE;
404 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
405 uint64_t objset, uint64_t object)
408 zbookmark_phys_t czb;
410 for (j = 0; j < dnp->dn_nblkptr; j++) {
411 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
412 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
415 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
416 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
417 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
422 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
423 uint64_t objset, uint64_t object)
426 zbookmark_phys_t czb;
428 if (td->td_flags & TRAVERSE_PRE) {
429 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
431 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
433 if (err == TRAVERSE_VISIT_NO_CHILDREN)
439 for (j = 0; j < dnp->dn_nblkptr; j++) {
440 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
441 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
446 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
447 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
448 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
451 if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
452 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
454 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
456 if (err == TRAVERSE_VISIT_NO_CHILDREN)
466 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
467 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
469 prefetch_data_t *pfd = arg;
470 arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
472 ASSERT(pfd->pd_bytes_fetched >= 0);
476 return (SET_ERROR(EINTR));
478 if (!prefetch_needed(pfd, bp))
481 mutex_enter(&pfd->pd_mtx);
482 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
483 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
484 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
485 cv_broadcast(&pfd->pd_cv);
486 mutex_exit(&pfd->pd_mtx);
488 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
489 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
495 traverse_prefetch_thread(void *arg)
497 traverse_data_t *td_main = arg;
498 traverse_data_t td = *td_main;
499 zbookmark_phys_t czb;
501 td.td_func = traverse_prefetcher;
502 td.td_arg = td_main->td_pfd;
505 SET_BOOKMARK(&czb, td.td_objset,
506 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
507 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
509 mutex_enter(&td_main->td_pfd->pd_mtx);
510 td_main->td_pfd->pd_exited = B_TRUE;
511 cv_broadcast(&td_main->td_pfd->pd_cv);
512 mutex_exit(&td_main->td_pfd->pd_mtx);
516 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
517 * in syncing context).
520 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
521 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
522 blkptr_cb_t func, void *arg)
525 prefetch_data_t pd = { 0 };
526 zbookmark_phys_t czb;
529 ASSERT(ds == NULL || objset == ds->ds_object);
530 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
533 * The data prefetching mechanism (the prefetch thread) is incompatible
534 * with resuming from a bookmark.
536 ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
539 td.td_objset = objset;
540 td.td_rootbp = rootbp;
541 td.td_min_txg = txg_start;
542 td.td_resume = resume;
547 td.td_paused = B_FALSE;
549 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
550 VERIFY(spa_feature_enabled_txg(spa,
551 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
553 td.td_hole_birth_enabled_txg = 0;
557 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
558 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
560 /* See comment on ZIL traversal in dsl_scan_visitds. */
561 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
562 arc_flags_t flags = ARC_FLAG_WAIT;
566 err = arc_read(NULL, td.td_spa, rootbp,
567 arc_getbuf_func, &buf,
568 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
573 traverse_zil(&td, &osp->os_zil_header);
574 (void) arc_buf_remove_ref(buf, &buf);
577 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
578 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
580 pd.pd_exited = B_TRUE;
582 SET_BOOKMARK(&czb, td.td_objset,
583 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
584 err = traverse_visitbp(&td, NULL, rootbp, &czb);
586 mutex_enter(&pd.pd_mtx);
587 pd.pd_cancel = B_TRUE;
588 cv_broadcast(&pd.pd_cv);
589 while (!pd.pd_exited)
590 cv_wait(&pd.pd_cv, &pd.pd_mtx);
591 mutex_exit(&pd.pd_mtx);
593 mutex_destroy(&pd.pd_mtx);
594 cv_destroy(&pd.pd_cv);
600 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
601 * in syncing context).
604 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
605 blkptr_cb_t func, void *arg)
607 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
608 &dsl_dataset_phys(ds)->ds_bp, txg_start, NULL, flags, func, arg));
612 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
613 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
614 blkptr_cb_t func, void *arg)
616 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
617 blkptr, txg_start, resume, flags, func, arg));
621 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
624 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
625 blkptr_cb_t func, void *arg)
629 dsl_pool_t *dp = spa_get_dsl(spa);
630 objset_t *mos = dp->dp_meta_objset;
631 boolean_t hard = (flags & TRAVERSE_HARD);
634 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
635 txg_start, NULL, flags, func, arg);
639 /* visit each dataset */
640 for (obj = 1; err == 0;
641 err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
642 dmu_object_info_t doi;
644 err = dmu_object_info(mos, obj, &doi);
651 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
653 uint64_t txg = txg_start;
655 dsl_pool_config_enter(dp, FTAG);
656 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
657 dsl_pool_config_exit(dp, FTAG);
663 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
664 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
665 err = traverse_dataset(ds, txg, flags, func, arg);
666 dsl_dataset_rele(ds, FTAG);