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;
51 zbookmark_phys_t pd_resume;
54 typedef struct traverse_data {
59 zbookmark_phys_t *td_resume;
61 prefetch_data_t *td_pfd;
63 uint64_t td_hole_birth_enabled_txg;
68 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
69 uint64_t objset, uint64_t object);
70 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
71 uint64_t objset, uint64_t object);
74 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
76 traverse_data_t *td = arg;
82 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
85 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
86 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
88 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
94 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
96 traverse_data_t *td = arg;
98 if (lrc->lrc_txtype == TX_WRITE) {
99 lr_write_t *lr = (lr_write_t *)lrc;
100 blkptr_t *bp = &lr->lr_blkptr;
106 if (claim_txg == 0 || bp->blk_birth < claim_txg)
109 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
110 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
112 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
119 traverse_zil(traverse_data_t *td, zil_header_t *zh)
121 uint64_t claim_txg = zh->zh_claim_txg;
125 * We only want to visit blocks that have been claimed but not yet
126 * replayed; plus, in read-only mode, blocks that are already stable.
128 if (claim_txg == 0 && spa_writeable(td->td_spa))
131 zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
133 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
139 typedef enum resume_skip {
146 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
147 * the block indicated by zb does not need to be visited at all. Returns
148 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
149 * resume point. This indicates that this block should be visited but not its
150 * children (since they must have been visited in a previous traversal).
151 * Otherwise returns RESUME_SKIP_NONE.
154 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
155 const zbookmark_phys_t *zb)
157 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
159 * If we already visited this bp & everything below,
160 * don't bother doing it again.
162 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
163 return (RESUME_SKIP_ALL);
166 * If we found the block we're trying to resume from, zero
167 * the bookmark out to indicate that we have resumed.
169 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
170 bzero(td->td_resume, sizeof (*zb));
171 if (td->td_flags & TRAVERSE_POST)
172 return (RESUME_SKIP_CHILDREN);
175 return (RESUME_SKIP_NONE);
179 traverse_prefetch_metadata(traverse_data_t *td,
180 const blkptr_t *bp, const zbookmark_phys_t *zb)
182 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
184 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
187 * If we are in the process of resuming, don't prefetch, because
188 * some children will not be needed (and in fact may have already
191 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
193 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
195 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
198 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
199 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
203 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
205 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
206 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
207 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
213 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
214 const blkptr_t *bp, const zbookmark_phys_t *zb)
216 zbookmark_phys_t czb;
218 arc_buf_t *buf = NULL;
219 prefetch_data_t *pd = td->td_pfd;
220 boolean_t hard = td->td_flags & TRAVERSE_HARD;
222 switch (resume_skip_check(td, dnp, zb)) {
223 case RESUME_SKIP_ALL:
225 case RESUME_SKIP_CHILDREN:
227 case RESUME_SKIP_NONE:
233 if (bp->blk_birth == 0) {
235 * Since this block has a birth time of 0 it must be a
236 * hole created before the SPA_FEATURE_HOLE_BIRTH
237 * feature was enabled. If SPA_FEATURE_HOLE_BIRTH
238 * was enabled before the min_txg for this traveral we
239 * know the hole must have been created before the
240 * min_txg for this traveral, so we can skip it. If
241 * SPA_FEATURE_HOLE_BIRTH was enabled after the min_txg
242 * for this traveral we cannot tell if the hole was
243 * created before or after the min_txg for this
244 * traversal, so we cannot skip it.
246 if (td->td_hole_birth_enabled_txg < td->td_min_txg)
248 } else if (bp->blk_birth <= td->td_min_txg) {
252 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
253 uint64_t size = BP_GET_LSIZE(bp);
254 mutex_enter(&pd->pd_mtx);
255 ASSERT(pd->pd_bytes_fetched >= 0);
256 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
257 cv_wait(&pd->pd_cv, &pd->pd_mtx);
258 pd->pd_bytes_fetched -= size;
259 cv_broadcast(&pd->pd_cv);
260 mutex_exit(&pd->pd_mtx);
263 if (BP_IS_HOLE(bp)) {
264 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
270 if (td->td_flags & TRAVERSE_PRE) {
271 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
273 if (err == TRAVERSE_VISIT_NO_CHILDREN)
279 if (BP_GET_LEVEL(bp) > 0) {
280 arc_flags_t flags = ARC_FLAG_WAIT;
283 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
285 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
286 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
291 for (i = 0; i < epb; i++) {
292 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
294 zb->zb_blkid * epb + i);
295 traverse_prefetch_metadata(td, &cbp[i], &czb);
298 /* recursively visitbp() blocks below this */
299 for (i = 0; i < epb; i++) {
300 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
302 zb->zb_blkid * epb + i);
303 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
307 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
308 arc_flags_t flags = ARC_FLAG_WAIT;
310 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);
316 dnode_phys_t *child_dnp = buf->b_data;
318 for (i = 0; i < epb; i++) {
319 prefetch_dnode_metadata(td, &child_dnp[i],
320 zb->zb_objset, zb->zb_blkid * epb + i);
323 /* recursively visitbp() blocks below this */
324 for (i = 0; i < epb; i++) {
325 err = traverse_dnode(td, &child_dnp[i],
326 zb->zb_objset, zb->zb_blkid * epb + i);
330 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
331 arc_flags_t flags = ARC_FLAG_WAIT;
333 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
334 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
338 objset_phys_t *osp = buf->b_data;
339 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
340 DMU_META_DNODE_OBJECT);
341 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
342 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
343 zb->zb_objset, DMU_GROUPUSED_OBJECT);
344 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
345 zb->zb_objset, DMU_USERUSED_OBJECT);
348 err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
349 DMU_META_DNODE_OBJECT);
350 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
351 err = traverse_dnode(td, &osp->os_groupused_dnode,
352 zb->zb_objset, DMU_GROUPUSED_OBJECT);
354 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
355 err = traverse_dnode(td, &osp->os_userused_dnode,
356 zb->zb_objset, DMU_USERUSED_OBJECT);
361 (void) arc_buf_remove_ref(buf, &buf);
364 if (err == 0 && (td->td_flags & TRAVERSE_POST))
365 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
367 if (hard && (err == EIO || err == ECKSUM)) {
369 * Ignore this disk error as requested by the HARD flag,
370 * and continue traversal.
376 * If we are stopping here, set td_resume.
378 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
379 td->td_resume->zb_objset = zb->zb_objset;
380 td->td_resume->zb_object = zb->zb_object;
381 td->td_resume->zb_level = 0;
383 * If we have stopped on an indirect block (e.g. due to
384 * i/o error), we have not visited anything below it.
385 * Set the bookmark to the first level-0 block that we need
386 * to visit. This way, the resuming code does not need to
387 * deal with resuming from indirect blocks.
389 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
392 td->td_resume->zb_blkid = zb->zb_blkid;
393 if (zb->zb_level > 0) {
394 td->td_resume->zb_blkid <<= zb->zb_level *
395 (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 (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
429 object < td->td_resume->zb_object)
432 if (td->td_flags & TRAVERSE_PRE) {
433 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
435 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
437 if (err == TRAVERSE_VISIT_NO_CHILDREN)
443 for (j = 0; j < dnp->dn_nblkptr; j++) {
444 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
445 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
450 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
451 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
452 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
455 if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
456 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
458 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
460 if (err == TRAVERSE_VISIT_NO_CHILDREN)
470 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
471 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
473 prefetch_data_t *pfd = arg;
474 arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
476 ASSERT(pfd->pd_bytes_fetched >= 0);
480 return (SET_ERROR(EINTR));
482 if (!prefetch_needed(pfd, bp))
485 mutex_enter(&pfd->pd_mtx);
486 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
487 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
488 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
489 cv_broadcast(&pfd->pd_cv);
490 mutex_exit(&pfd->pd_mtx);
492 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
493 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
499 traverse_prefetch_thread(void *arg)
501 traverse_data_t *td_main = arg;
502 traverse_data_t td = *td_main;
503 zbookmark_phys_t czb;
505 td.td_func = traverse_prefetcher;
506 td.td_arg = td_main->td_pfd;
508 td.td_resume = &td_main->td_pfd->pd_resume;
510 SET_BOOKMARK(&czb, td.td_objset,
511 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
512 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
514 mutex_enter(&td_main->td_pfd->pd_mtx);
515 td_main->td_pfd->pd_exited = B_TRUE;
516 cv_broadcast(&td_main->td_pfd->pd_cv);
517 mutex_exit(&td_main->td_pfd->pd_mtx);
521 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
522 * in syncing context).
525 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
526 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
527 blkptr_cb_t func, void *arg)
530 prefetch_data_t pd = { 0 };
531 zbookmark_phys_t czb;
534 ASSERT(ds == NULL || objset == ds->ds_object);
535 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
538 td.td_objset = objset;
539 td.td_rootbp = rootbp;
540 td.td_min_txg = txg_start;
541 td.td_resume = resume;
546 td.td_paused = B_FALSE;
548 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
549 VERIFY(spa_feature_enabled_txg(spa,
550 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
552 td.td_hole_birth_enabled_txg = 0;
557 pd.pd_resume = *resume;
558 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
559 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
561 /* See comment on ZIL traversal in dsl_scan_visitds. */
562 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
563 arc_flags_t flags = ARC_FLAG_WAIT;
567 err = arc_read(NULL, td.td_spa, rootbp,
568 arc_getbuf_func, &buf,
569 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
574 traverse_zil(&td, &osp->os_zil_header);
575 (void) arc_buf_remove_ref(buf, &buf);
578 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
579 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
581 pd.pd_exited = B_TRUE;
583 SET_BOOKMARK(&czb, td.td_objset,
584 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
585 err = traverse_visitbp(&td, NULL, rootbp, &czb);
587 mutex_enter(&pd.pd_mtx);
588 pd.pd_cancel = B_TRUE;
589 cv_broadcast(&pd.pd_cv);
590 while (!pd.pd_exited)
591 cv_wait(&pd.pd_cv, &pd.pd_mtx);
592 mutex_exit(&pd.pd_mtx);
594 mutex_destroy(&pd.pd_mtx);
595 cv_destroy(&pd.pd_cv);
601 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
602 * in syncing context).
605 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
606 zbookmark_phys_t *resume,
607 int flags, blkptr_cb_t func, void *arg)
609 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
610 &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
614 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
615 int flags, blkptr_cb_t func, void *arg)
617 return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
621 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
622 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
623 blkptr_cb_t func, void *arg)
625 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
626 blkptr, txg_start, resume, flags, func, arg));
630 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
633 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
634 blkptr_cb_t func, void *arg)
637 dsl_pool_t *dp = spa_get_dsl(spa);
638 objset_t *mos = dp->dp_meta_objset;
639 boolean_t hard = (flags & TRAVERSE_HARD);
642 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
643 txg_start, NULL, flags, func, arg);
647 /* visit each dataset */
648 for (uint64_t obj = 1; err == 0;
649 err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
650 dmu_object_info_t doi;
652 err = dmu_object_info(mos, obj, &doi);
659 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
661 uint64_t txg = txg_start;
663 dsl_pool_config_enter(dp, FTAG);
664 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
665 dsl_pool_config_exit(dp, FTAG);
671 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
672 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
673 err = traverse_dataset(ds, txg, flags, func, arg);
674 dsl_dataset_rele(ds, FTAG);