X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fdmu_traverse.c;h=1c39723714379b2596c233f9b1ec6f0b192df70f;hb=294f68063b49c06d3118d51016811063e69cf97a;hp=429c76ae11f12e2a40c9f55a250aa5643dc04f60;hpb=428870ff734fdaccc342b33fc53cf94724409a46;p=zfs.git diff --git a/module/zfs/dmu_traverse.c b/module/zfs/dmu_traverse.c index 429c76a..1c39723 100644 --- a/module/zfs/dmu_traverse.c +++ b/module/zfs/dmu_traverse.c @@ -20,6 +20,7 @@ */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. + * Copyright (c) 2012 by Delphix. All rights reserved. */ #include @@ -36,7 +37,9 @@ #include #include -struct prefetch_data { +int zfs_pd_blks_max = 100; + +typedef struct prefetch_data { kmutex_t pd_mtx; kcondvar_t pd_cv; int pd_blks_max; @@ -44,27 +47,29 @@ struct prefetch_data { int pd_flags; boolean_t pd_cancel; boolean_t pd_exited; -}; +} prefetch_data_t; -struct traverse_data { +typedef struct traverse_data { spa_t *td_spa; uint64_t td_objset; blkptr_t *td_rootbp; uint64_t td_min_txg; + zbookmark_t *td_resume; int td_flags; - struct prefetch_data *td_pfd; + prefetch_data_t *td_pfd; blkptr_cb_t *td_func; void *td_arg; -}; +} traverse_data_t; -static int traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp, - arc_buf_t *buf, uint64_t objset, uint64_t object); +static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, + uint64_t objset, uint64_t object); +static void prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *, + uint64_t objset, uint64_t object); -/* ARGSUSED */ static int traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) { - struct traverse_data *td = arg; + traverse_data_t *td = arg; zbookmark_t zb; if (bp->blk_birth == 0) @@ -76,16 +81,15 @@ traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) SET_BOOKMARK(&zb, td->td_objset, ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); - (void) td->td_func(td->td_spa, zilog, bp, NULL, &zb, NULL, td->td_arg); + (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); return (0); } -/* ARGSUSED */ static int traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) { - struct traverse_data *td = arg; + traverse_data_t *td = arg; if (lrc->lrc_txtype == TX_WRITE) { lr_write_t *lr = (lr_write_t *)lrc; @@ -98,17 +102,17 @@ traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) if (claim_txg == 0 || bp->blk_birth < claim_txg) return (0); - SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, ZB_ZIL_LEVEL, - lr->lr_offset / BP_GET_LSIZE(bp)); + SET_BOOKMARK(&zb, td->td_objset, lr->lr_foid, + ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); - (void) td->td_func(td->td_spa, zilog, bp, NULL, &zb, NULL, + (void) td->td_func(td->td_spa, zilog, bp, &zb, NULL, td->td_arg); } return (0); } static void -traverse_zil(struct traverse_data *td, zil_header_t *zh) +traverse_zil(traverse_data_t *td, zil_header_t *zh) { uint64_t claim_txg = zh->zh_claim_txg; zilog_t *zilog; @@ -128,19 +132,102 @@ traverse_zil(struct traverse_data *td, zil_header_t *zh) zil_free(zilog); } +typedef enum resume_skip { + RESUME_SKIP_ALL, + RESUME_SKIP_NONE, + RESUME_SKIP_CHILDREN +} resume_skip_t; + +/* + * Returns RESUME_SKIP_ALL if td indicates that we are resuming a traversal and + * the block indicated by zb does not need to be visited at all. Returns + * RESUME_SKIP_CHILDREN if we are resuming a post traversal and we reach the + * resume point. This indicates that this block should be visited but not its + * children (since they must have been visited in a previous traversal). + * Otherwise returns RESUME_SKIP_NONE. + */ +static resume_skip_t +resume_skip_check(traverse_data_t *td, const dnode_phys_t *dnp, + const zbookmark_t *zb) +{ + if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) { + /* + * If we already visited this bp & everything below, + * don't bother doing it again. + */ + if (zbookmark_is_before(dnp, zb, td->td_resume)) + return (RESUME_SKIP_ALL); + + /* + * If we found the block we're trying to resume from, zero + * the bookmark out to indicate that we have resumed. + */ + ASSERT3U(zb->zb_object, <=, td->td_resume->zb_object); + if (bcmp(zb, td->td_resume, sizeof (*zb)) == 0) { + bzero(td->td_resume, sizeof (*zb)); + if (td->td_flags & TRAVERSE_POST) + return (RESUME_SKIP_CHILDREN); + } + } + return (RESUME_SKIP_NONE); +} + +static void +traverse_pause(traverse_data_t *td, const zbookmark_t *zb) +{ + ASSERT(td->td_resume != NULL); + ASSERT0(zb->zb_level); + bcopy(zb, td->td_resume, sizeof (*td->td_resume)); +} + +static void +traverse_prefetch_metadata(traverse_data_t *td, + const blkptr_t *bp, const zbookmark_t *zb) +{ + uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; + + if (!(td->td_flags & TRAVERSE_PREFETCH_METADATA)) + return; + /* + * If we are in the process of resuming, don't prefetch, because + * some children will not be needed (and in fact may have already + * been freed). + */ + if (td->td_resume != NULL && !ZB_IS_ZERO(td->td_resume)) + return; + if (BP_IS_HOLE(bp) || bp->blk_birth <= td->td_min_txg) + return; + if (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE) + return; + + (void) arc_read(NULL, td->td_spa, bp, NULL, NULL, + ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); +} + static int -traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, - arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb) +traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, + const blkptr_t *bp, const zbookmark_t *zb) { zbookmark_t czb; int err = 0, lasterr = 0; arc_buf_t *buf = NULL; - struct prefetch_data *pd = td->td_pfd; + prefetch_data_t *pd = td->td_pfd; boolean_t hard = td->td_flags & TRAVERSE_HARD; + boolean_t pause = B_FALSE; - if (bp->blk_birth == 0) { - err = td->td_func(td->td_spa, NULL, NULL, pbuf, zb, dnp, - td->td_arg); + switch (resume_skip_check(td, dnp, zb)) { + case RESUME_SKIP_ALL: + return (0); + case RESUME_SKIP_CHILDREN: + goto post; + case RESUME_SKIP_NONE: + break; + default: + ASSERT(0); + } + + if (BP_IS_HOLE(bp)) { + err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg); return (err); } @@ -160,10 +247,14 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, } if (td->td_flags & TRAVERSE_PRE) { - err = td->td_func(td->td_spa, NULL, bp, pbuf, zb, dnp, + err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); - if (err) - return (err); + if (err == TRAVERSE_VISIT_NO_CHILDREN) + return (0); + if (err == ERESTART) + pause = B_TRUE; /* handle pausing at a common point */ + if (err != 0) + goto post; } if (BP_GET_LEVEL(bp) > 0) { @@ -172,19 +263,25 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, blkptr_t *cbp; int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; - err = dsl_read(NULL, td->td_spa, bp, pbuf, - arc_getbuf_func, &buf, + err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); if (err) return (err); + cbp = buf->b_data; + + for (i = 0; i < epb; i++) { + SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, + zb->zb_level - 1, + zb->zb_blkid * epb + i); + traverse_prefetch_metadata(td, &cbp[i], &czb); + } /* recursively visitbp() blocks below this */ - cbp = buf->b_data; - for (i = 0; i < epb; i++, cbp++) { + for (i = 0; i < epb; i++) { SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, zb->zb_level - 1, zb->zb_blkid * epb + i); - err = traverse_visitbp(td, dnp, buf, cbp, &czb); + err = traverse_visitbp(td, dnp, &cbp[i], &czb); if (err) { if (!hard) break; @@ -196,16 +293,20 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, int i; int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; - err = dsl_read(NULL, td->td_spa, bp, pbuf, - arc_getbuf_func, &buf, + err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); if (err) return (err); + dnp = buf->b_data; + + for (i = 0; i < epb; i++) { + prefetch_dnode_metadata(td, &dnp[i], zb->zb_objset, + zb->zb_blkid * epb + i); + } /* recursively visitbp() blocks below this */ - dnp = buf->b_data; - for (i = 0; i < epb; i++, dnp++) { - err = traverse_dnode(td, dnp, buf, zb->zb_objset, + for (i = 0; i < epb; i++) { + err = traverse_dnode(td, &dnp[i], zb->zb_objset, zb->zb_blkid * epb + i); if (err) { if (!hard) @@ -218,17 +319,23 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, objset_phys_t *osp; dnode_phys_t *dnp; - err = dsl_read_nolock(NULL, td->td_spa, bp, - arc_getbuf_func, &buf, + err = arc_read(NULL, td->td_spa, bp, arc_getbuf_func, &buf, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb); if (err) return (err); osp = buf->b_data; - traverse_zil(td, &osp->os_zil_header); - dnp = &osp->os_meta_dnode; - err = traverse_dnode(td, dnp, buf, zb->zb_objset, + prefetch_dnode_metadata(td, dnp, zb->zb_objset, + DMU_META_DNODE_OBJECT); + if (arc_buf_size(buf) >= sizeof (objset_phys_t)) { + prefetch_dnode_metadata(td, &osp->os_userused_dnode, + zb->zb_objset, DMU_USERUSED_OBJECT); + prefetch_dnode_metadata(td, &osp->os_groupused_dnode, + zb->zb_objset, DMU_USERUSED_OBJECT); + } + + err = traverse_dnode(td, dnp, zb->zb_objset, DMU_META_DNODE_OBJECT); if (err && hard) { lasterr = err; @@ -236,7 +343,7 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, } if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { dnp = &osp->os_userused_dnode; - err = traverse_dnode(td, dnp, buf, zb->zb_objset, + err = traverse_dnode(td, dnp, zb->zb_objset, DMU_USERUSED_OBJECT); } if (err && hard) { @@ -245,7 +352,7 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, } if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { dnp = &osp->os_groupused_dnode; - err = traverse_dnode(td, dnp, buf, zb->zb_objset, + err = traverse_dnode(td, dnp, zb->zb_objset, DMU_GROUPUSED_OBJECT); } } @@ -253,17 +360,43 @@ traverse_visitbp(struct traverse_data *td, const dnode_phys_t *dnp, if (buf) (void) arc_buf_remove_ref(buf, &buf); +post: if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST)) { - err = td->td_func(td->td_spa, NULL, bp, pbuf, zb, dnp, - td->td_arg); + err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); + if (err == ERESTART) + pause = B_TRUE; + } + + if (pause && td->td_resume != NULL) { + ASSERT3U(err, ==, ERESTART); + ASSERT(!hard); + traverse_pause(td, zb); } return (err != 0 ? err : lasterr); } +static void +prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, + uint64_t objset, uint64_t object) +{ + int j; + zbookmark_t czb; + + for (j = 0; j < dnp->dn_nblkptr; j++) { + SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); + traverse_prefetch_metadata(td, &dnp->dn_blkptr[j], &czb); + } + + if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { + SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); + traverse_prefetch_metadata(td, &dnp->dn_spill, &czb); + } +} + static int -traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp, - arc_buf_t *buf, uint64_t objset, uint64_t object) +traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, + uint64_t objset, uint64_t object) { int j, err = 0, lasterr = 0; zbookmark_t czb; @@ -271,8 +404,7 @@ traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp, for (j = 0; j < dnp->dn_nblkptr; j++) { SET_BOOKMARK(&czb, objset, object, dnp->dn_nlevels - 1, j); - err = traverse_visitbp(td, dnp, buf, - (blkptr_t *)&dnp->dn_blkptr[j], &czb); + err = traverse_visitbp(td, dnp, &dnp->dn_blkptr[j], &czb); if (err) { if (!hard) break; @@ -281,10 +413,8 @@ traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp, } if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { - SET_BOOKMARK(&czb, objset, - object, 0, DMU_SPILL_BLKID); - err = traverse_visitbp(td, dnp, buf, - (blkptr_t *)&dnp->dn_spill, &czb); + SET_BOOKMARK(&czb, objset, object, 0, DMU_SPILL_BLKID); + err = traverse_visitbp(td, dnp, &dnp->dn_spill, &czb); if (err) { if (!hard) return (err); @@ -297,10 +427,9 @@ traverse_dnode(struct traverse_data *td, const dnode_phys_t *dnp, /* ARGSUSED */ static int traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, - arc_buf_t *pbuf, const zbookmark_t *zb, const dnode_phys_t *dnp, - void *arg) + const zbookmark_t *zb, const dnode_phys_t *dnp, void *arg) { - struct prefetch_data *pfd = arg; + prefetch_data_t *pfd = arg; uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; ASSERT(pfd->pd_blks_fetched >= 0); @@ -319,10 +448,8 @@ traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, cv_broadcast(&pfd->pd_cv); mutex_exit(&pfd->pd_mtx); - (void) dsl_read(NULL, spa, bp, pbuf, NULL, NULL, - ZIO_PRIORITY_ASYNC_READ, - ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, - &aflags, zb); + (void) arc_read(NULL, spa, bp, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, + ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, zb); return (0); } @@ -330,8 +457,8 @@ traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, static void traverse_prefetch_thread(void *arg) { - struct traverse_data *td_main = arg; - struct traverse_data td = *td_main; + traverse_data_t *td_main = arg; + traverse_data_t td = *td_main; zbookmark_t czb; td.td_func = traverse_prefetcher; @@ -340,7 +467,7 @@ traverse_prefetch_thread(void *arg) SET_BOOKMARK(&czb, td.td_objset, ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); - (void) traverse_visitbp(&td, NULL, NULL, td.td_rootbp, &czb); + (void) traverse_visitbp(&td, NULL, td.td_rootbp, &czb); mutex_enter(&td_main->td_pfd->pd_mtx); td_main->td_pfd->pd_exited = B_TRUE; @@ -353,46 +480,76 @@ traverse_prefetch_thread(void *arg) * in syncing context). */ static int -traverse_impl(spa_t *spa, uint64_t objset, blkptr_t *rootbp, - uint64_t txg_start, int flags, blkptr_cb_t func, void *arg) +traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp, + uint64_t txg_start, zbookmark_t *resume, int flags, + blkptr_cb_t func, void *arg) { - struct traverse_data td; - struct prefetch_data pd = { 0 }; - zbookmark_t czb; + traverse_data_t *td; + prefetch_data_t *pd; + zbookmark_t *czb; int err; - td.td_spa = spa; - td.td_objset = objset; - td.td_rootbp = rootbp; - td.td_min_txg = txg_start; - td.td_func = func; - td.td_arg = arg; - td.td_pfd = &pd; - td.td_flags = flags; - - pd.pd_blks_max = 100; - pd.pd_flags = flags; - mutex_init(&pd.pd_mtx, NULL, MUTEX_DEFAULT, NULL); - cv_init(&pd.pd_cv, NULL, CV_DEFAULT, NULL); - - if (!(flags & TRAVERSE_PREFETCH) || + ASSERT(ds == NULL || objset == ds->ds_object); + ASSERT(!(flags & TRAVERSE_PRE) || !(flags & TRAVERSE_POST)); + + /* + * The data prefetching mechanism (the prefetch thread) is incompatible + * with resuming from a bookmark. + */ + ASSERT(resume == NULL || !(flags & TRAVERSE_PREFETCH_DATA)); + + td = kmem_alloc(sizeof(traverse_data_t), KM_PUSHPAGE); + pd = kmem_zalloc(sizeof(prefetch_data_t), KM_PUSHPAGE); + czb = kmem_alloc(sizeof(zbookmark_t), KM_PUSHPAGE); + + td->td_spa = spa; + td->td_objset = objset; + td->td_rootbp = rootbp; + td->td_min_txg = txg_start; + td->td_resume = resume; + td->td_func = func; + td->td_arg = arg; + td->td_pfd = pd; + td->td_flags = flags; + + pd->pd_blks_max = zfs_pd_blks_max; + pd->pd_flags = flags; + mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL); + cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL); + + /* See comment on ZIL traversal in dsl_scan_visitds. */ + if (ds != NULL && !dsl_dataset_is_snapshot(ds)) { + objset_t *os; + + err = dmu_objset_from_ds(ds, &os); + if (err) + return (err); + + traverse_zil(td, &os->os_zil_header); + } + + if (!(flags & TRAVERSE_PREFETCH_DATA) || 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread, - &td, TQ_NOQUEUE)) - pd.pd_exited = B_TRUE; + td, TQ_NOQUEUE)) + pd->pd_exited = B_TRUE; - SET_BOOKMARK(&czb, objset, + SET_BOOKMARK(czb, td->td_objset, ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); - err = traverse_visitbp(&td, NULL, NULL, rootbp, &czb); + err = traverse_visitbp(td, NULL, rootbp, czb); + + mutex_enter(&pd->pd_mtx); + pd->pd_cancel = B_TRUE; + cv_broadcast(&pd->pd_cv); + while (!pd->pd_exited) + cv_wait(&pd->pd_cv, &pd->pd_mtx); + mutex_exit(&pd->pd_mtx); - mutex_enter(&pd.pd_mtx); - pd.pd_cancel = B_TRUE; - cv_broadcast(&pd.pd_cv); - while (!pd.pd_exited) - cv_wait(&pd.pd_cv, &pd.pd_mtx); - mutex_exit(&pd.pd_mtx); + mutex_destroy(&pd->pd_mtx); + cv_destroy(&pd->pd_cv); - mutex_destroy(&pd.pd_mtx); - cv_destroy(&pd.pd_cv); + kmem_free(czb, sizeof(zbookmark_t)); + kmem_free(pd, sizeof(struct prefetch_data)); + kmem_free(td, sizeof(struct traverse_data)); return (err); } @@ -405,8 +562,17 @@ int traverse_dataset(dsl_dataset_t *ds, uint64_t txg_start, int flags, blkptr_cb_t func, void *arg) { - return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds->ds_object, - &ds->ds_phys->ds_bp, txg_start, flags, func, arg)); + return (traverse_impl(ds->ds_dir->dd_pool->dp_spa, ds, ds->ds_object, + &ds->ds_phys->ds_bp, txg_start, NULL, flags, func, arg)); +} + +int +traverse_dataset_destroyed(spa_t *spa, blkptr_t *blkptr, + uint64_t txg_start, zbookmark_t *resume, int flags, + blkptr_cb_t func, void *arg) +{ + return (traverse_impl(spa, NULL, ZB_DESTROYED_OBJSET, + blkptr, txg_start, resume, flags, func, arg)); } /* @@ -423,8 +589,8 @@ traverse_pool(spa_t *spa, uint64_t txg_start, int flags, boolean_t hard = (flags & TRAVERSE_HARD); /* visit the MOS */ - err = traverse_impl(spa, 0, spa_get_rootblkptr(spa), - txg_start, flags, func, arg); + err = traverse_impl(spa, NULL, 0, spa_get_rootblkptr(spa), + txg_start, NULL, flags, func, arg); if (err) return (err); @@ -469,3 +635,11 @@ traverse_pool(spa_t *spa, uint64_t txg_start, int flags, err = 0; return (err != 0 ? err : lasterr); } + +#if defined(_KERNEL) && defined(HAVE_SPL) +EXPORT_SYMBOL(traverse_dataset); +EXPORT_SYMBOL(traverse_pool); + +module_param(zfs_pd_blks_max, int, 0644); +MODULE_PARM_DESC(zfs_pd_blks_max, "Max number of blocks to prefetch"); +#endif