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) 2013 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>
40 int zfs_pd_blks_max = 100;
42 typedef struct prefetch_data {
52 typedef struct traverse_data {
57 zbookmark_t *td_resume;
59 prefetch_data_t *td_pfd;
64 static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
65 uint64_t objset, uint64_t object);
66 static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *,
67 uint64_t objset, uint64_t object);
70 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
72 traverse_data_t *td = arg;
75 if (bp->blk_birth == 0)
78 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(td->td_spa))
81 SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
82 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
84 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg);
90 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
92 traverse_data_t *td = arg;
94 if (lrc->lrc_txtype == TX_WRITE) {
95 lr_write_t *lr = (lr_write_t *)lrc;
96 blkptr_t *bp = &lr->lr_blkptr;
99 if (bp->blk_birth == 0)
102 if (claim_txg == 0 || bp->blk_birth < claim_txg)
105 SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid,
106 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
108 (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL,
115 traverse_zil(traverse_data_t *td, zil_header_t *zh)
117 uint64_t claim_txg = zh->zh_claim_txg;
121 * We only want to visit blocks that have been claimed but not yet
122 * replayed; plus, in read-only mode, blocks that are already stable.
124 if (claim_txg == 0 && spa_writeable(td->td_spa))
127 zilog = zil_alloc(spa_get_dsl(td->td_spa)->dp_meta_objset, zh);
129 (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, td,
135 typedef enum resume_skip {
142 * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and
143 * the block indicated by zb does not need to be visited at all. Returns
144 * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the
145 * resume point. This indicates that this block should be visited but not its
146 * children (since they must have been visited in a previous traversal).
147 * Otherwise returns RESUME_SKIP_NONE.
150 resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp,
151 const zbookmark_t *zb)
153 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) {
155 * If we already visited this bp & everything below,
156 * don't bother doing it again.
158 if (zbookmark_is_before(dnp, zb, td->td_resume))
159 return (RESUME_SKIP_ALL);
162 * If we found the block we're trying to resume from, zero
163 * the bookmark out to indicate that we have resumed.
165 ASSERT3U(zb->zb_object, <=, td->td_resume->zb_object);
166 if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) {
167 bzero(td->td_resume, sizeof (*zb));
168 if (td->td_flags & TRAVERSE_POST)
169 return (RESUME_SKIP_CHILDREN);
172 return (RESUME_SKIP_NONE);
176 traverse_pause(traverse_data_t *td, const zbookmark_t *zb)
178 ASSERT(td->td_resume != NULL);
179 ASSERT0(zb->zb_level);
180 bcopy(zb, td->td_resume, sizeof (*td->td_resume));
184 traverse_prefetch_metadata(traverse_data_t *td,
185 const blkptr_t *bp, const zbookmark_t *zb)
187 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
189 if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA))
192 * If we are in the process of resuming, don't prefetch, because
193 * some children will not be needed (and in fact may have already
196 if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume))
198 if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg)
200 if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)
203 (void) arc_read(NULL, td->td_spa, bp, NULL, NULL,
204 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
208 traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
209 const blkptr_t *bp, const zbookmark_t *zb)
212 int err = 0, lasterr = 0;
213 arc_buf_t *buf = NULL;
214 prefetch_data_t *pd = td->td_pfd;
215 boolean_t hard = td->td_flags & TRAVERSE_HARD;
216 boolean_t pause = B_FALSE;
218 switch (resume_skip_check(td, dnp, zb)) {
219 case RESUME_SKIP_ALL:
221 case RESUME_SKIP_CHILDREN:
223 case RESUME_SKIP_NONE:
229 if (BP_IS_HOLE(bp)) {
230 err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg);
234 if (bp->blk_birth <= td->td_min_txg)
237 if (pd && !pd->pd_exited &&
238 ((pd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
239 BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) {
240 mutex_enter(&pd->pd_mtx);
241 ASSERT(pd->pd_blks_fetched >= 0);
242 while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
243 cv_wait(&pd->pd_cv, &pd->pd_mtx);
244 pd->pd_blks_fetched--;
245 cv_broadcast(&pd->pd_cv);
246 mutex_exit(&pd->pd_mtx);
249 if (td->td_flags & TRAVERSE_PRE) {
250 err = td->td_func(td->td_spa, NULL, bp, zb, dnp,
252 if (err == TRAVERSE_VISIT_NO_CHILDREN)
255 pause = B_TRUE; /* handle pausing at a common point */
260 if (BP_GET_LEVEL(bp) > 0) {
261 uint32_t flags = ARC_WAIT;
264 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
266 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
267 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
272 for (i = 0; i < epb; i++) {
273 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
275 zb->zb_blkid * epb + i);
276 traverse_prefetch_metadata(td, &cbp[i], &czb);
279 /* recursively visitbp() blocks below this */
280 for (i = 0; i < epb; i++) {
281 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
283 zb->zb_blkid * epb + i);
284 err = traverse_visitbp(td, dnp, &cbp[i], &czb);
291 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
292 uint32_t flags = ARC_WAIT;
294 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
296 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
297 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
302 for (i = 0; i < epb; i++) {
303 prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset,
304 zb->zb_blkid * epb + i);
307 /* recursively visitbp() blocks below this */
308 for (i = 0; i < epb; i++) {
309 err = traverse_dnode(td, &dnp[i], zb->zb_objset,
310 zb->zb_blkid * epb + i);
317 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
318 uint32_t flags = ARC_WAIT;
322 err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf,
323 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
328 dnp = &osp->os_meta_dnode;
329 prefetch_dnode_metadata(td, dnp, zb->zb_objset,
330 DMU_META_DNODE_OBJECT);
331 if (arc_buf_size(buf) >= sizeof (objset_phys_t)) {
332 prefetch_dnode_metadata(td, &osp->os_userused_dnode,
333 zb->zb_objset, DMU_USERUSED_OBJECT);
334 prefetch_dnode_metadata(td, &osp->os_groupused_dnode,
335 zb->zb_objset, DMU_USERUSED_OBJECT);
338 err = traverse_dnode(td, dnp, zb->zb_objset,
339 DMU_META_DNODE_OBJECT);
344 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
345 dnp = &osp->os_userused_dnode;
346 err = traverse_dnode(td, dnp, zb->zb_objset,
347 DMU_USERUSED_OBJECT);
353 if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) {
354 dnp = &osp->os_groupused_dnode;
355 err = traverse_dnode(td, dnp, zb->zb_objset,
356 DMU_GROUPUSED_OBJECT);
361 (void) arc_buf_remove_ref(buf, &buf);
364 if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST)) {
365 err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg);
370 if (pause && td->td_resume != NULL) {
371 ASSERT3U(err, ==, ERESTART);
373 traverse_pause(td, zb);
376 return (err != 0 ? err : lasterr);
380 prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp,
381 uint64_t objset, uint64_t object)
386 for (j = 0; j < dnp->dn_nblkptr; j++) {
387 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
388 traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb);
391 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
392 SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID);
393 traverse_prefetch_metadata(td, &dnp->dn_spill, &czb);
398 traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp,
399 uint64_t objset, uint64_t object)
401 int j, err = 0, lasterr = 0;
403 boolean_t hard = (td->td_flags & TRAVERSE_HARD);
405 for (j = 0; j < dnp->dn_nblkptr; j++) {
406 SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j);
407 err = traverse_visitbp(td, dnp, &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 err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb);
424 return (err != 0 ? err : lasterr);
429 traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
430 const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg)
432 prefetch_data_t *pfd = arg;
433 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
435 ASSERT(pfd->pd_blks_fetched >= 0);
437 return (SET_ERROR(EINTR));
439 if (bp == NULL || !((pfd->pd_flags & TRAVERSE_PREFETCH_DATA) ||
440 BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0) ||
441 BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG)
444 mutex_enter(&pfd->pd_mtx);
445 while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
446 cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
447 pfd->pd_blks_fetched++;
448 cv_broadcast(&pfd->pd_cv);
449 mutex_exit(&pfd->pd_mtx);
451 (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
452 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb);
458 traverse_prefetch_thread(void *arg)
460 traverse_data_t *td_main = arg;
461 traverse_data_t td = *td_main;
464 td.td_func = traverse_prefetcher;
465 td.td_arg = td_main->td_pfd;
468 SET_BOOKMARK(&czb, td.td_objset,
469 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
470 (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb);
472 mutex_enter(&td_main->td_pfd->pd_mtx);
473 td_main->td_pfd->pd_exited = B_TRUE;
474 cv_broadcast(&td_main->td_pfd->pd_cv);
475 mutex_exit(&td_main->td_pfd->pd_mtx);
479 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
480 * in syncing context).
483 traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
484 uint64_t txg_start, zbookmark_t *resume, int flags,
485 blkptr_cb_t func, void *arg)
488 prefetch_data_t pd = { 0 };
492 ASSERT(ds == NULL || objset == ds->ds_object);
493 ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST));
496 * The data prefetching mechanism (the prefetch thread) is incompatible
497 * with resuming from a bookmark.
499 ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA));
502 td.td_objset = objset;
503 td.td_rootbp = rootbp;
504 td.td_min_txg = txg_start;
505 td.td_resume = resume;
511 pd.pd_blks_max = zfs_pd_blks_max;
513 mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL);
514 cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL);
516 /* See comment on ZIL traversal in dsl_scan_visitds. */
517 if (ds != NULL && !dsl_dataset_is_snapshot(ds) && !BP_IS_HOLE(rootbp)) {
518 uint32_t flags = ARC_WAIT;
522 err = arc_read(NULL, td.td_spa, rootbp,
523 arc_getbuf_func, &buf,
524 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, NULL);
529 traverse_zil(&td, &osp->os_zil_header);
530 (void) arc_buf_remove_ref(buf, &buf);
533 if (!(flags & TRAVERSE_PREFETCH_DATA) ||
534 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread,
536 pd.pd_exited = B_TRUE;
538 SET_BOOKMARK(&czb, td.td_objset,
539 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
540 err = traverse_visitbp(&td, NULL, rootbp, &czb);
542 mutex_enter(&pd.pd_mtx);
543 pd.pd_cancel = B_TRUE;
544 cv_broadcast(&pd.pd_cv);
545 while (!pd.pd_exited)
546 cv_wait(&pd.pd_cv, &pd.pd_mtx);
547 mutex_exit(&pd.pd_mtx);
549 mutex_destroy(&pd.pd_mtx);
550 cv_destroy(&pd.pd_cv);
556 * NB: dataset must not be changing on-disk (eg, is a snapshot or we are
557 * in syncing context).
560 traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags,
561 blkptr_cb_t func, void *arg)
563 return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object,
564 &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg));
568 traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr,
569 uint64_t txg_start, zbookmark_t *resume, int flags,
570 blkptr_cb_t func, void *arg)
572 return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET,
573 blkptr, txg_start, resume, flags, func, arg));
577 * NB: pool must not be changing on-disk (eg, from zdb or sync context).
580 traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
581 blkptr_cb_t func, void *arg)
583 int err, lasterr = 0;
585 dsl_pool_t *dp = spa_get_dsl(spa);
586 objset_t *mos = dp->dp_meta_objset;
587 boolean_t hard = (flags & TRAVERSE_HARD);
590 err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa),
591 txg_start, NULL, flags, func, arg);
595 /* visit each dataset */
596 for (obj = 1; err == 0 || (err != ESRCH && hard);
597 err = dmu_object_next(mos, &obj, FALSE, txg_start)) {
598 dmu_object_info_t doi;
600 err = dmu_object_info(mos, obj, &doi);
608 if (doi.doi_type == DMU_OT_DSL_DATASET) {
610 uint64_t txg = txg_start;
612 dsl_pool_config_enter(dp, FTAG);
613 err = dsl_dataset_hold_obj(dp, obj, FTAG, &ds);
614 dsl_pool_config_exit(dp, FTAG);
621 if (ds->ds_phys->ds_prev_snap_txg > txg)
622 txg = ds->ds_phys->ds_prev_snap_txg;
623 err = traverse_dataset(ds, txg, flags, func, arg);
624 dsl_dataset_rele(ds, FTAG);
634 return (err != 0 ? err : lasterr);