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, 2016 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;
66 boolean_t td_realloc_possible;
69 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
70 uint64_t objset, uint64_t object);
71 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
72 uint64_t objset, uint64_t object);
75 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
77 traverse_data_t *td = arg;
83 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
86 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
87 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
89 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
95 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
97 traverse_data_t *td = arg;
99 if (lrc->lrc_txtype == TX_WRITE) {
100 lr_write_t *lr = (lr_write_t *)lrc;
101 blkptr_t *bp = &lr->lr_blkptr;
107 if (claim_txg == 0 || bp->blk_birth < claim_txg)
110 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
111 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
113 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
120 traverse_zil(traverse_data_t *td, zil_header_t *zh)
122 uint64_t claim_txg = zh->zh_claim_txg;
126 * We only want to visit blocks that have been claimed but not yet
127 * replayed; plus, in read-only mode, blocks that are already stable.
129 if (claim_txg == 0 && spa_writeable(td->td_spa))
132 zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
134 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
140 typedef enum resume_skip {
147 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
148 * the block indicated by zb does not need to be visited at all. Returns
149 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
150 * resume point. This indicates that this block should be visited but not its
151 * children (since they must have been visited in a previous traversal).
152 * Otherwise returns RESUME_SKIP_NONE.
155 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
156 const zbookmark_phys_t *zb)
158 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
160 * If we already visited this bp & everything below,
161 * don't bother doing it again.
163 if (zbookmark_subtree_completed(dnp, zb, td->td_resume))
164 return (RESUME_SKIP_ALL);
167 * If we found the block we're trying to resume from, zero
168 * the bookmark out to indicate that we have resumed.
170 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
171 bzero(td->td_resume, sizeof (*zb));
172 if (td->td_flags & TRAVERSE_POST)
173 return (RESUME_SKIP_CHILDREN);
176 return (RESUME_SKIP_NONE);
180 traverse_prefetch_metadata(traverse_data_t *td,
181 const blkptr_t *bp, const zbookmark_phys_t *zb)
183 arc_flags_t flags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
185 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
188 * If we are in the process of resuming, don't prefetch, because
189 * some children will not be needed (and in fact may have already
192 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
194 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
196 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
199 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
200 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
204 prefetch_needed(prefetch_data_t *pfd, const blkptr_t *bp)
206 ASSERT(pfd->pd_flags & TRAVERSE_PREFETCH_DATA);
207 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) ||
208 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
214 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
215 const blkptr_t *bp, const zbookmark_phys_t *zb)
217 zbookmark_phys_t czb;
219 arc_buf_t *buf = NULL;
220 prefetch_data_t *pd = td->td_pfd;
221 boolean_t hard = td->td_flags & TRAVERSE_HARD;
223 switch (resume_skip_check(td, dnp, zb)) {
224 case RESUME_SKIP_ALL:
226 case RESUME_SKIP_CHILDREN:
228 case RESUME_SKIP_NONE:
234 if (bp->blk_birth == 0) {
236 * Since this block has a birth time of 0 it must be one of
237 * two things: a hole created before the
238 * SPA_FEATURE_HOLE_BIRTH feature was enabled, or a hole
239 * which has always been a hole in an object.
241 * If a file is written sparsely, then the unwritten parts of
242 * the file were "always holes" -- that is, they have been
243 * holes since this object was allocated. However, we (and
244 * our callers) can not necessarily tell when an object was
245 * allocated. Therefore, if it's possible that this object
246 * was freed and then its object number reused, we need to
247 * visit all the holes with birth==0.
249 * If it isn't possible that the object number was reused,
250 * then if SPA_FEATURE_HOLE_BIRTH was enabled before we wrote
251 * all the blocks we will visit as part of this traversal,
252 * then this hole must have always existed, so we can skip
253 * it. We visit blocks born after (exclusive) td_min_txg.
255 * Note that the meta-dnode cannot be reallocated.
257 if ((!td->td_realloc_possible ||
258 zb->zb_object == DMU_META_DNODE_OBJECT) &&
259 td->td_hole_birth_enabled_txg <= td->td_min_txg)
261 } else if (bp->blk_birth <= td->td_min_txg) {
265 if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
266 uint64_t size = BP_GET_LSIZE(bp);
267 mutex_enter(&pd->pd_mtx);
268 ASSERT(pd->pd_bytes_fetched >= 0);
269 while (pd->pd_bytes_fetched < size && !pd->pd_exited)
270 cv_wait(&pd->pd_cv, &pd->pd_mtx);
271 pd->pd_bytes_fetched -= size;
272 cv_broadcast(&pd->pd_cv);
273 mutex_exit(&pd->pd_mtx);
276 if (BP_IS_HOLE(bp)) {
277 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
283 if (td->td_flags & TRAVERSE_PRE) {
284 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
286 if (err == TRAVERSE_VISIT_NO_CHILDREN)
292 if (BP_GET_LEVEL(bp) > 0) {
293 arc_flags_t flags = ARC_FLAG_WAIT;
296 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
298 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
299 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
304 for (i = 0; i < epb; i++) {
305 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
307 zb->zb_blkid * epb + i);
308 traverse_prefetch_metadata(td, &cbp[i], &czb);
311 /* recursively visitbp() blocks below this */
312 for (i = 0; i < epb; i++) {
313 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
315 zb->zb_blkid * epb + i);
316 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
320 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
321 arc_flags_t flags = ARC_FLAG_WAIT;
323 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
325 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
326 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
329 dnode_phys_t *child_dnp = buf->b_data;
331 for (i = 0; i < epb; i++) {
332 prefetch_dnode_metadata(td, &child_dnp[i],
333 zb->zb_objset, zb->zb_blkid * epb + i);
336 /* recursively visitbp() blocks below this */
337 for (i = 0; i < epb; i++) {
338 err = traverse_dnode(td, &child_dnp[i],
339 zb->zb_objset, zb->zb_blkid * epb + i);
343 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
344 arc_flags_t flags = ARC_FLAG_WAIT;
346 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
347 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
351 objset_phys_t *osp = buf->b_data;
352 prefetch_dnode_metadata(td, &osp->os_meta_dnode, zb->zb_objset,
353 DMU_META_DNODE_OBJECT);
355 * See the block comment above for the goal of this variable.
356 * If the maxblkid of the meta-dnode is 0, then we know that
357 * we've never had more than DNODES_PER_BLOCK objects in the
358 * dataset, which means we can't have reused any object ids.
360 if (osp->os_meta_dnode.dn_maxblkid == 0)
361 td->td_realloc_possible = B_FALSE;
363 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
364 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
365 zb->zb_objset, DMU_GROUPUSED_OBJECT);
366 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
367 zb->zb_objset, DMU_USERUSED_OBJECT);
370 err = traverse_dnode(td, &osp->os_meta_dnode, zb->zb_objset,
371 DMU_META_DNODE_OBJECT);
372 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
373 err = traverse_dnode(td, &osp->os_groupused_dnode,
374 zb->zb_objset, DMU_GROUPUSED_OBJECT);
376 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
377 err = traverse_dnode(td, &osp->os_userused_dnode,
378 zb->zb_objset, DMU_USERUSED_OBJECT);
383 (void) arc_buf_remove_ref(buf, &buf);
386 if (err == 0 && (td->td_flags & TRAVERSE_POST))
387 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
389 if (hard && (err == EIO || err == ECKSUM)) {
391 * Ignore this disk error as requested by the HARD flag,
392 * and continue traversal.
398 * If we are stopping here, set td_resume.
400 if (td->td_resume != NULL && err != 0 && !td->td_paused) {
401 td->td_resume->zb_objset = zb->zb_objset;
402 td->td_resume->zb_object = zb->zb_object;
403 td->td_resume->zb_level = 0;
405 * If we have stopped on an indirect block (e.g. due to
406 * i/o error), we have not visited anything below it.
407 * Set the bookmark to the first level-0 block that we need
408 * to visit. This way, the resuming code does not need to
409 * deal with resuming from indirect blocks.
411 * Note, if zb_level <= 0, dnp may be NULL, so we don't want
414 td->td_resume->zb_blkid = zb->zb_blkid;
415 if (zb->zb_level > 0) {
416 td->td_resume->zb_blkid <<= zb->zb_level *
417 (dnp->dn_indblkshift - SPA_BLKPTRSHIFT);
419 td->td_paused = B_TRUE;
426 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
427 uint64_t objset, uint64_t object)
430 zbookmark_phys_t czb;
432 for (j = 0; j < dnp->dn_nblkptr; j++) {
433 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
434 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
437 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
438 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
439 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
444 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
445 uint64_t objset, uint64_t object)
448 zbookmark_phys_t czb;
450 if (object != DMU_META_DNODE_OBJECT && td->td_resume != NULL &&
451 object < td->td_resume->zb_object)
454 if (td->td_flags & TRAVERSE_PRE) {
455 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
457 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
459 if (err == TRAVERSE_VISIT_NO_CHILDREN)
465 for (j = 0; j < dnp->dn_nblkptr; j++) {
466 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
467 err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb);
472 if (err == 0 && (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
473 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
474 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
477 if (err == 0 && (td->td_flags & TRAVERSE_POST)) {
478 SET_BOOKMARK(&czb, objset, object, ZB_DNODE_LEVEL,
480 err = td->td_func(td->td_spa, NULL, NULL, &czb, dnp,
482 if (err == TRAVERSE_VISIT_NO_CHILDREN)
492 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
493 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
495 prefetch_data_t *pfd = arg;
496 arc_flags_t aflags = ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
498 ASSERT(pfd->pd_bytes_fetched >= 0);
502 return (SET_ERROR(EINTR));
504 if (!prefetch_needed(pfd, bp))
507 mutex_enter(&pfd->pd_mtx);
508 while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
509 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
510 pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
511 cv_broadcast(&pfd->pd_cv);
512 mutex_exit(&pfd->pd_mtx);
514 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
515 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
521 traverse_prefetch_thread(void *arg)
523 traverse_data_t *td_main = arg;
524 traverse_data_t td = *td_main;
525 zbookmark_phys_t czb;
527 td.td_func = traverse_prefetcher;
528 td.td_arg = td_main->td_pfd;
530 td.td_resume = &td_main->td_pfd->pd_resume;
532 SET_BOOKMARK(&czb, td.td_objset,
533 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
534 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
536 mutex_enter(&td_main->td_pfd->pd_mtx);
537 td_main->td_pfd->pd_exited = B_TRUE;
538 cv_broadcast(&td_main->td_pfd->pd_cv);
539 mutex_exit(&td_main->td_pfd->pd_mtx);
543 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
544 * in syncing context).
547 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
548 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
549 blkptr_cb_t func, void *arg)
552 prefetch_data_t pd = { 0 };
553 zbookmark_phys_t czb;
556 ASSERT(ds == NULL || objset == ds->ds_object);
557 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
560 td.td_objset = objset;
561 td.td_rootbp = rootbp;
562 td.td_min_txg = txg_start;
563 td.td_resume = resume;
568 td.td_paused = B_FALSE;
569 td.td_realloc_possible = (txg_start == 0 ? B_FALSE : B_TRUE);
571 if (spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
572 VERIFY(spa_feature_enabled_txg(spa,
573 SPA_FEATURE_HOLE_BIRTH, &td.td_hole_birth_enabled_txg));
575 td.td_hole_birth_enabled_txg = UINT64_MAX;
580 pd.pd_resume = *resume;
581 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
582 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
584 /* See comment on ZIL traversal in dsl_scan_visitds. */
585 if (ds != NULL && !ds->ds_is_snapshot && !BP_IS_HOLE(rootbp)) {
586 arc_flags_t flags = ARC_FLAG_WAIT;
590 err = arc_read(NULL, td.td_spa, rootbp,
591 arc_getbuf_func, &buf,
592 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
597 traverse_zil(&td, &osp->os_zil_header);
598 (void) arc_buf_remove_ref(buf, &buf);
601 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
602 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
604 pd.pd_exited = B_TRUE;
606 SET_BOOKMARK(&czb, td.td_objset,
607 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
608 err = traverse_visitbp(&td, NULL, rootbp, &czb);
610 mutex_enter(&pd.pd_mtx);
611 pd.pd_cancel = B_TRUE;
612 cv_broadcast(&pd.pd_cv);
613 while (!pd.pd_exited)
614 cv_wait(&pd.pd_cv, &pd.pd_mtx);
615 mutex_exit(&pd.pd_mtx);
617 mutex_destroy(&pd.pd_mtx);
618 cv_destroy(&pd.pd_cv);
624 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
625 * in syncing context).
628 traverse_dataset_resume(dsl_dataset_t *ds, uint64_t txg_start,
629 zbookmark_phys_t *resume,
630 int flags, blkptr_cb_t func, void *arg)
632 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
633 &dsl_dataset_phys(ds)->ds_bp, txg_start, resume, flags, func, arg));
637 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start,
638 int flags, blkptr_cb_t func, void *arg)
640 return (traverse_dataset_resume(ds, txg_start, NULL, flags, func, arg));
644 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
645 uint64_t txg_start, zbookmark_phys_t *resume, int flags,
646 blkptr_cb_t func, void *arg)
648 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
649 blkptr, txg_start, resume, flags, func, arg));
653 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
656 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
657 blkptr_cb_t func, void *arg)
660 dsl_pool_t *dp = spa_get_dsl(spa);
661 objset_t *mos = dp->dp_meta_objset;
662 boolean_t hard = (flags & TRAVERSE_HARD);
665 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
666 txg_start, NULL, flags, func, arg);
670 /* visit each dataset */
671 for (uint64_t obj = 1; err == 0;
672 err = dmu_object_next(mos, &obj, B_FALSE, txg_start)) {
673 dmu_object_info_t doi;
675 err = dmu_object_info(mos, obj, &doi);
682 if (doi.doi_bonus_type == DMU_OT_DSL_DATASET) {
684 uint64_t txg = txg_start;
686 dsl_pool_config_enter(dp, FTAG);
687 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
688 dsl_pool_config_exit(dp, FTAG);
694 if (dsl_dataset_phys(ds)->ds_prev_snap_txg > txg)
695 txg = dsl_dataset_phys(ds)->ds_prev_snap_txg;
696 err = traverse_dataset(ds, txg, flags, func, arg);
697 dsl_dataset_rele(ds, FTAG);