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 int zfs_pd_blks_max = 100;
44 typedef struct prefetch_data {
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_is_before(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 mutex_enter(&pd->pd_mtx);
254 ASSERT(pd->pd_blks_fetched >= 0);
255 while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
256 cv_wait(&pd->pd_cv, &pd->pd_mtx);
257 pd->pd_blks_fetched--;
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 for (j = 0; j < dnp->dn_nblkptr; j++) {
429 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
430 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
435 if (err == 0 && dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
436 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
437 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
444 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
445 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
447 prefetch_data_t *pfd = arg;
448 arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
450 ASSERT(pfd->pd_blks_fetched >= 0);
452 return (SET_ERROR(EINTR));
454 if (!prefetch_needed(pfd, bp))
457 mutex_enter(&pfd->pd_mtx);
458 while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
459 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
460 pfd->pd_blks_fetched++;
461 cv_broadcast(&pfd->pd_cv);
462 mutex_exit(&pfd->pd_mtx);
464 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
465 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
471 traverse_prefetch_thread(void *arg)
473 traverse_data_t *td_main = arg;
474 traverse_data_t td = *td_main;
475 zbookmark_phys_t czb;
477 td.td_func = traverse_prefetcher;
478 td.td_arg = td_main->td_pfd;
481 SET_BOOKMARK(&czb, td.td_objset,
482 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
483 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
485 mutex_enter(&td_main->td_pfd->pd_mtx);
486 td_main->td_pfd->pd_exited = B_TRUE;
487 cv_broadcast(&td_main->td_pfd->pd_cv);
488 mutex_exit(&td_main->td_pfd->pd_mtx);
492 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
493 * in syncing context).
496 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
497 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
498 blkptr_cb_t func, void *arg)
501 prefetch_data_t pd = { 0 };
502 zbookmark_phys_t czb;
505 ASSERT(ds == NULL || objset == ds->ds_object);
506 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
509 * The data prefetching mechanism (the prefetch thread) is incompatible
510 * with resuming from a bookmark.
512 ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
515 td.td_objset = objset;
516 td.td_rootbp = rootbp;
517 td.td_min_txg = txg_start;
518 td.td_resume = resume;
523 td.td_paused = B_FALSE;
525 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
526 VERIFY(spa_feature_enabled_txg(spa,
527 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
529 td.td_hole_birth_enabled_txg = 0;
532 pd.pd_blks_max = zfs_pd_blks_max;
534 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
535 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
537 /* See comment on ZIL traversal in dsl_scan_visitds. */
538 if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
539 arc_flags_t flags = ARC_FLAG_WAIT;
543 err = arc_read(NULL, td.td_spa, rootbp,
544 arc_getbuf_func, &buf,
545 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
550 traverse_zil(&td, &osp->os_zil_header);
551 (void) arc_buf_remove_ref(buf, &buf);
554 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
555 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
557 pd.pd_exited = B_TRUE;
559 SET_BOOKMARK(&czb, td.td_objset,
560 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
561 err = traverse_visitbp(&td, NULL, rootbp, &czb);
563 mutex_enter(&pd.pd_mtx);
564 pd.pd_cancel = B_TRUE;
565 cv_broadcast(&pd.pd_cv);
566 while (!pd.pd_exited)
567 cv_wait(&pd.pd_cv, &pd.pd_mtx);
568 mutex_exit(&pd.pd_mtx);
570 mutex_destroy(&pd.pd_mtx);
571 cv_destroy(&pd.pd_cv);
577 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
578 * in syncing context).
581 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
582 blkptr_cb_t func, void *arg)
584 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
585 &dsl_dataset_phys(ds)->ds_bp, txg_start, NULL, flags, func, arg));
589 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
590 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
591 blkptr_cb_t func, void *arg)
593 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
594 blkptr, txg_start, resume, flags, func, arg));
598 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
601 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
602 blkptr_cb_t func, void *arg)
606 dsl_pool_t *dp = spa_get_dsl(spa);
607 objset_t *mos = dp->dp_meta_objset;
608 boolean_t hard = (flags & TRAVERSE_HARD);
611 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
612 txg_start, NULL, flags, func, arg);
616 /* visit each dataset */
617 for (obj = 1; err == 0;
618 err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
619 dmu_object_info_t doi;
621 err = dmu_object_info(mos, obj, &doi);
628 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
630 uint64_t txg = txg_start;
632 dsl_pool_config_enter(dp, FTAG);
633 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
634 dsl_pool_config_exit(dp, FTAG);
640 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
641 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
642 err = traverse_dataset(ds, txg, flags, func, arg);
643 dsl_dataset_rele(ds, FTAG);