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.
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>
35 #include <sys/dmu_impl.h>
37 #include <sys/sa_impl.h>
38 #include <sys/callb.h>
39 #include <sys/zfeature.h>
41 int zfs_pd_blks_max = 100;
43 typedef struct prefetch_data {
53 typedef struct traverse_data {
58 zbookmark_phys_t *td_resume;
60 prefetch_data_t *td_pfd;
66 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
67 uint64_t objset, uint64_t object);
68 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
69 uint64_t objset, uint64_t object);
72 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
74 traverse_data_t *td = arg;
80 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
83 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
84 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
86 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
92 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
94 traverse_data_t *td = arg;
96 if (lrc->lrc_txtype == TX_WRITE) {
97 lr_write_t *lr = (lr_write_t *)lrc;
98 blkptr_t *bp = &lr->lr_blkptr;
104 if (claim_txg == 0 || bp->blk_birth < claim_txg)
107 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
108 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
110 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
117 traverse_zil(traverse_data_t *td, zil_header_t *zh)
119 uint64_t claim_txg = zh->zh_claim_txg;
123 * We only want to visit blocks that have been claimed but not yet
124 * replayed; plus, in read-only mode, blocks that are already stable.
126 if (claim_txg == 0 && spa_writeable(td->td_spa))
129 zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
131 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
137 typedef enum resume_skip {
144 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
145 * the block indicated by zb does not need to be visited at all. Returns
146 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
147 * resume point. This indicates that this block should be visited but not its
148 * children (since they must have been visited in a previous traversal).
149 * Otherwise returns RESUME_SKIP_NONE.
152 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
153 const zbookmark_phys_t *zb)
155 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
157 * If we already visited this bp & everything below,
158 * don't bother doing it again.
160 if (zbookmark_is_before(dnp, zb, td->td_resume))
161 return (RESUME_SKIP_ALL);
164 * If we found the block we're trying to resume from, zero
165 * the bookmark out to indicate that we have resumed.
167 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
168 bzero(td->td_resume, sizeof (*zb));
169 if (td->td_flags & TRAVERSE_POST)
170 return (RESUME_SKIP_CHILDREN);
173 return (RESUME_SKIP_NONE);
177 traverse_prefetch_metadata(traverse_data_t *td,
178 const blkptr_t *bp, const zbookmark_phys_t *zb)
180 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
182 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
185 * If we are in the process of resuming, don't prefetch, because
186 * some children will not be needed (and in fact may have already
189 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
191 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
193 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
196 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
197 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
201 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
203 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
204 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
205 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
211 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
212 const blkptr_t *bp, const zbookmark_phys_t *zb)
214 zbookmark_phys_t czb;
216 arc_buf_t *buf = NULL;
217 prefetch_data_t *pd = td->td_pfd;
218 boolean_t hard = td->td_flags & TRAVERSE_HARD;
220 switch (resume_skip_check(td, dnp, zb)) {
221 case RESUME_SKIP_ALL:
223 case RESUME_SKIP_CHILDREN:
225 case RESUME_SKIP_NONE:
231 if (bp->blk_birth == 0) {
232 if (spa_feature_is_active(td->td_spa, SPA_FEATURE_HOLE_BIRTH)) {
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 uint64_t hole_birth_enabled_txg;
246 VERIFY(spa_feature_enabled_txg(td->td_spa,
247 SPA_FEATURE_HOLE_BIRTH, &hole_birth_enabled_txg));
248 if (hole_birth_enabled_txg < td->td_min_txg)
251 } else if (bp->blk_birth <= td->td_min_txg) {
255 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
256 mutex_enter(&pd->pd_mtx);
257 ASSERT(pd->pd_blks_fetched >= 0);
258 while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
259 cv_wait(&pd->pd_cv, &pd->pd_mtx);
260 pd->pd_blks_fetched--;
261 cv_broadcast(&pd->pd_cv);
262 mutex_exit(&pd->pd_mtx);
265 if (BP_IS_HOLE(bp)) {
266 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
272 if (td->td_flags & TRAVERSE_PRE) {
273 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
275 if (err == TRAVERSE_VISIT_NO_CHILDREN)
281 if (BP_GET_LEVEL(bp) > 0) {
282 uint32_t flags = ARC_WAIT;
285 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
287 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
288 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
293 for (i = 0; i < epb; i++) {
294 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
296 zb->zb_blkid * epb + i);
297 traverse_prefetch_metadata(td, &cbp[i], &czb);
300 /* recursively visitbp() blocks below this */
301 for (i = 0; i < epb; i++) {
302 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
304 zb->zb_blkid * epb + i);
305 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
309 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
310 uint32_t flags = ARC_WAIT;
312 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
314 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
315 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
320 for (i = 0; i < epb; i++) {
321 prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
322 zb->zb_blkid * epb + i);
325 /* recursively visitbp() blocks below this */
326 for (i = 0; i < epb; i++) {
327 err = traverse_dnode(td, &dnp[i], zb->zb_objset,
328 zb->zb_blkid * epb + i);
332 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
333 uint32_t flags = ARC_WAIT;
337 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
338 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
343 dnp = &osp->os_meta_dnode;
344 prefetch_dnode_metadata(td, dnp, zb->zb_objset,
345 DMU_META_DNODE_OBJECT);
346 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
347 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
348 zb->zb_objset, DMU_GROUPUSED_OBJECT);
349 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
350 zb->zb_objset, DMU_USERUSED_OBJECT);
353 err = traverse_dnode(td, dnp, zb->zb_objset,
354 DMU_META_DNODE_OBJECT);
355 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
356 dnp = &osp->os_groupused_dnode;
357 err = traverse_dnode(td, dnp, zb->zb_objset,
358 DMU_GROUPUSED_OBJECT);
360 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
361 dnp = &osp->os_userused_dnode;
362 err = traverse_dnode(td, dnp, zb->zb_objset,
363 DMU_USERUSED_OBJECT);
368 (void) arc_buf_remove_ref(buf, &buf);
371 if (err == 0 && (td->td_flags & TRAVERSE_POST))
372 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
374 if (hard && (err == EIO || err == ECKSUM)) {
376 * Ignore this disk error as requested by the HARD flag,
377 * and continue traversal.
383 * If we are stopping here, set td_resume.
385 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
386 td->td_resume->zb_objset = zb->zb_objset;
387 td->td_resume->zb_object = zb->zb_object;
388 td->td_resume->zb_level = 0;
390 * If we have stopped on an indirect block (e.g. due to
391 * i/o error), we have not visited anything below it.
392 * Set the bookmark to the first level-0 block that we need
393 * to visit. This way, the resuming code does not need to
394 * deal with resuming from indirect blocks.
396 td->td_resume->zb_blkid = zb->zb_blkid <<
397 (zb->zb_level * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
398 td->td_paused = B_TRUE;
405 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
406 uint64_t objset, uint64_t object)
409 zbookmark_phys_t czb;
411 for (j = 0; j < dnp->dn_nblkptr; j++) {
412 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
413 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
416 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
417 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
418 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
423 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
424 uint64_t objset, uint64_t object)
427 zbookmark_phys_t czb;
429 for (j = 0; j < dnp->dn_nblkptr; j++) {
430 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
431 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
436 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
437 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
438 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
445 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
446 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
448 prefetch_data_t *pfd = arg;
449 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
451 ASSERT(pfd->pd_blks_fetched >= 0);
453 return (SET_ERROR(EINTR));
455 if (!prefetch_needed(pfd, bp))
458 mutex_enter(&pfd->pd_mtx);
459 while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
460 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
461 pfd->pd_blks_fetched++;
462 cv_broadcast(&pfd->pd_cv);
463 mutex_exit(&pfd->pd_mtx);
465 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
466 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
472 traverse_prefetch_thread(void *arg)
474 traverse_data_t *td_main = arg;
475 traverse_data_t td = *td_main;
476 zbookmark_phys_t czb;
478 td.td_func = traverse_prefetcher;
479 td.td_arg = td_main->td_pfd;
482 SET_BOOKMARK(&czb, td.td_objset,
483 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
484 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
486 mutex_enter(&td_main->td_pfd->pd_mtx);
487 td_main->td_pfd->pd_exited = B_TRUE;
488 cv_broadcast(&td_main->td_pfd->pd_cv);
489 mutex_exit(&td_main->td_pfd->pd_mtx);
493 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
494 * in syncing context).
497 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
498 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
499 blkptr_cb_t func, void *arg)
502 prefetch_data_t pd = { 0 };
503 zbookmark_phys_t czb;
506 ASSERT(ds == NULL || objset == ds->ds_object);
507 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
510 * The data prefetching mechanism (the prefetch thread) is incompatible
511 * with resuming from a bookmark.
513 ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
516 td.td_objset = objset;
517 td.td_rootbp = rootbp;
518 td.td_min_txg = txg_start;
519 td.td_resume = resume;
524 td.td_paused = B_FALSE;
526 pd.pd_blks_max = zfs_pd_blks_max;
528 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
529 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
531 /* See comment on ZIL traversal in dsl_scan_visitds. */
532 if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
533 uint32_t flags = ARC_WAIT;
537 err = arc_read(NULL, td.td_spa, rootbp,
538 arc_getbuf_func, &buf,
539 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
544 traverse_zil(&td, &osp->os_zil_header);
545 (void) arc_buf_remove_ref(buf, &buf);
548 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
549 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
551 pd.pd_exited = B_TRUE;
553 SET_BOOKMARK(&czb, td.td_objset,
554 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
555 err = traverse_visitbp(&td, NULL, rootbp, &czb);
557 mutex_enter(&pd.pd_mtx);
558 pd.pd_cancel = B_TRUE;
559 cv_broadcast(&pd.pd_cv);
560 while (!pd.pd_exited)
561 cv_wait(&pd.pd_cv, &pd.pd_mtx);
562 mutex_exit(&pd.pd_mtx);
564 mutex_destroy(&pd.pd_mtx);
565 cv_destroy(&pd.pd_cv);
571 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
572 * in syncing context).
575 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
576 blkptr_cb_t func, void *arg)
578 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
579 &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
583 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
584 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
585 blkptr_cb_t func, void *arg)
587 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
588 blkptr, txg_start, resume, flags, func, arg));
592 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
595 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
596 blkptr_cb_t func, void *arg)
600 dsl_pool_t *dp = spa_get_dsl(spa);
601 objset_t *mos = dp->dp_meta_objset;
602 boolean_t hard = (flags & TRAVERSE_HARD);
605 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
606 txg_start, NULL, flags, func, arg);
610 /* visit each dataset */
611 for (obj = 1; err == 0;
612 err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
613 dmu_object_info_t doi;
615 err = dmu_object_info(mos, obj, &doi);
622 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
624 uint64_t txg = txg_start;
626 dsl_pool_config_enter(dp, FTAG);
627 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
628 dsl_pool_config_exit(dp, FTAG);
634 if (ds->ds_phys->ds_prev_snap_txg > txg)
635 txg = ds->ds_phys->ds_prev_snap_txg;
636 err = traverse_dataset(ds, txg, flags, func, arg);
637 dsl_dataset_rele(ds, FTAG);