X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fdmu_traverse.c;h=1c39723714379b2596c233f9b1ec6f0b192df70f;hb=refs%2Fheads%2Frertzinger%2Ffeature-zpool-get--p;hp=775bb83583787313835ce2d83241d64a6ce87ab6;hpb=c28b227942b421ebdc03c9df9a012642fb517223;p=zfs.git diff --git a/module/zfs/dmu_traverse.c b/module/zfs/dmu_traverse.c index 775bb83..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 @@ -53,39 +54,17 @@ typedef struct traverse_data { uint64_t td_objset; blkptr_t *td_rootbp; uint64_t td_min_txg; + zbookmark_t *td_resume; int td_flags; prefetch_data_t *td_pfd; blkptr_cb_t *td_func; void *td_arg; } traverse_data_t; -typedef struct traverse_visitbp_data { - /* Function arguments */ - traverse_data_t *tv_td; - const dnode_phys_t *tv_dnp; - arc_buf_t *tv_pbuf; - blkptr_t *tv_bp; - const zbookmark_t *tv_zb; - /* Local variables */ - prefetch_data_t *tv_pd; - zbookmark_t tv_czb; - arc_buf_t *tv_buf; - boolean_t tv_hard; - objset_phys_t *tv_osp; - dnode_phys_t *tv_ldnp; - blkptr_t *tv_cbp; - uint32_t tv_flags; - int tv_err; - int tv_lasterr; - int tv_i; - int tv_epb; - int tv_depth; -} traverse_visitbp_data_t; - -static inline int traverse_visitbp(traverse_data_t *td, const - dnode_phys_t *dnp, arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb); static int traverse_dnode(traverse_data_t *td, const dnode_phys_t *dnp, - arc_buf_t *buf, uint64_t objset, uint64_t object); + 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); static int traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) @@ -102,7 +81,7 @@ 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); } @@ -126,7 +105,7 @@ traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) 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); @@ -153,196 +132,271 @@ traverse_zil(traverse_data_t *td, zil_header_t *zh) zil_free(zilog); } -#define TRAVERSE_VISITBP_MAX_DEPTH 20 +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_visitbp_init(traverse_visitbp_data_t *tv, - traverse_data_t *td, const dnode_phys_t *dnp, - arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb, int depth) +traverse_pause(traverse_data_t *td, const zbookmark_t *zb) { - tv->tv_td = td; - tv->tv_dnp = dnp; - tv->tv_pbuf = pbuf; - tv->tv_bp = bp; - tv->tv_zb = zb; - tv->tv_err = 0; - tv->tv_lasterr = 0; - tv->tv_buf = NULL; - tv->tv_pd = td->td_pfd; - tv->tv_hard = td->td_flags & TRAVERSE_HARD; - tv->tv_flags = ARC_WAIT; - tv->tv_depth = depth; + 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 noinline int -__traverse_visitbp(traverse_visitbp_data_t *tv) +static int +traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, + const blkptr_t *bp, const zbookmark_t *zb) { - ASSERT3S(tv->tv_depth, <, TRAVERSE_VISITBP_MAX_DEPTH); + zbookmark_t czb; + int err = 0, lasterr = 0; + arc_buf_t *buf = NULL; + prefetch_data_t *pd = td->td_pfd; + boolean_t hard = td->td_flags & TRAVERSE_HARD; + boolean_t pause = B_FALSE; + + 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 (tv->tv_bp->blk_birth == 0) { - tv->tv_err = tv->tv_td->td_func(tv->tv_td->td_spa, NULL, NULL, - tv->tv_pbuf, tv->tv_zb, tv->tv_dnp, tv->tv_td->td_arg); - return (tv->tv_err); + if (BP_IS_HOLE(bp)) { + err = td->td_func(td->td_spa, NULL, NULL, zb, dnp, td->td_arg); + return (err); } - if (tv->tv_bp->blk_birth <= tv->tv_td->td_min_txg) + if (bp->blk_birth <= td->td_min_txg) return (0); - if (tv->tv_pd && !tv->tv_pd->pd_exited && - ((tv->tv_pd->pd_flags & TRAVERSE_PREFETCH_DATA) || - BP_GET_TYPE(tv->tv_bp) == DMU_OT_DNODE || - BP_GET_LEVEL(tv->tv_bp) > 0)) { - mutex_enter(&tv->tv_pd->pd_mtx); - ASSERT(tv->tv_pd->pd_blks_fetched >= 0); - while (tv->tv_pd->pd_blks_fetched == 0 && !tv->tv_pd->pd_exited) - cv_wait(&tv->tv_pd->pd_cv, &tv->tv_pd->pd_mtx); - tv->tv_pd->pd_blks_fetched--; - cv_broadcast(&tv->tv_pd->pd_cv); - mutex_exit(&tv->tv_pd->pd_mtx); + if (pd && !pd->pd_exited && + ((pd->pd_flags & TRAVERSE_PREFETCH_DATA) || + BP_GET_TYPE(bp) == DMU_OT_DNODE || BP_GET_LEVEL(bp) > 0)) { + mutex_enter(&pd->pd_mtx); + ASSERT(pd->pd_blks_fetched >= 0); + while (pd->pd_blks_fetched == 0 && !pd->pd_exited) + cv_wait(&pd->pd_cv, &pd->pd_mtx); + pd->pd_blks_fetched--; + cv_broadcast(&pd->pd_cv); + mutex_exit(&pd->pd_mtx); } - if (tv->tv_td->td_flags & TRAVERSE_PRE) { - tv->tv_err = tv->tv_td->td_func(tv->tv_td->td_spa, NULL, - tv->tv_bp, tv->tv_pbuf, tv->tv_zb, tv->tv_dnp, - tv->tv_td->td_arg); - if (tv->tv_err == TRAVERSE_VISIT_NO_CHILDREN) + if (td->td_flags & TRAVERSE_PRE) { + err = td->td_func(td->td_spa, NULL, bp, zb, dnp, + td->td_arg); + if (err == TRAVERSE_VISIT_NO_CHILDREN) return (0); - if (tv->tv_err) - return (tv->tv_err); + if (err == ERESTART) + pause = B_TRUE; /* handle pausing at a common point */ + if (err != 0) + goto post; } - if (BP_GET_LEVEL(tv->tv_bp) > 0) { - tv->tv_epb = BP_GET_LSIZE(tv->tv_bp) >> SPA_BLKPTRSHIFT; + if (BP_GET_LEVEL(bp) > 0) { + uint32_t flags = ARC_WAIT; + int i; + blkptr_t *cbp; + int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; - tv->tv_err = dsl_read(NULL, tv->tv_td->td_spa, tv->tv_bp, - tv->tv_pbuf, arc_getbuf_func, &tv->tv_buf, - ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, - &tv->tv_flags, tv->tv_zb); - if (tv->tv_err) - return (tv->tv_err); + 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 */ - tv->tv_cbp = tv->tv_buf->b_data; - for (tv->tv_i = 0; tv->tv_i < tv->tv_epb; - tv->tv_i++, tv->tv_cbp++) { - SET_BOOKMARK(&tv->tv_czb, tv->tv_zb->zb_objset, - tv->tv_zb->zb_object, tv->tv_zb->zb_level - 1, - tv->tv_zb->zb_blkid * tv->tv_epb + tv->tv_i); - __traverse_visitbp_init(tv + 1, tv->tv_td, - tv->tv_dnp, tv->tv_buf, tv->tv_cbp, - &tv->tv_czb, tv->tv_depth + 1); - tv->tv_err = __traverse_visitbp(tv + 1); - if (tv->tv_err) { - if (!tv->tv_hard) + 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, &cbp[i], &czb); + if (err) { + if (!hard) break; - tv->tv_lasterr = tv->tv_err; + lasterr = err; } } - } else if (BP_GET_TYPE(tv->tv_bp) == DMU_OT_DNODE) { - tv->tv_epb = BP_GET_LSIZE(tv->tv_bp) >> DNODE_SHIFT; + } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { + uint32_t flags = ARC_WAIT; + int i; + int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; - tv->tv_err = dsl_read(NULL, tv->tv_td->td_spa, tv->tv_bp, - tv->tv_pbuf, arc_getbuf_func, &tv->tv_buf, - ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, - &tv->tv_flags, tv->tv_zb); - if (tv->tv_err) - return (tv->tv_err); + 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 */ - tv->tv_dnp = tv->tv_buf->b_data; - for (tv->tv_i = 0; tv->tv_i < tv->tv_epb; - tv->tv_i++, tv->tv_dnp++) { - tv->tv_err = traverse_dnode(tv->tv_td, tv->tv_dnp, - tv->tv_buf, tv->tv_zb->zb_objset, - tv->tv_zb->zb_blkid * tv->tv_epb + tv->tv_i); - if (tv->tv_err) { - if (!tv->tv_hard) + for (i = 0; i < epb; i++) { + err = traverse_dnode(td, &dnp[i], zb->zb_objset, + zb->zb_blkid * epb + i); + if (err) { + if (!hard) break; - tv->tv_lasterr = tv->tv_err; + lasterr = err; } } - } else if (BP_GET_TYPE(tv->tv_bp) == DMU_OT_OBJSET) { - - tv->tv_err = dsl_read_nolock(NULL, tv->tv_td->td_spa, - tv->tv_bp, arc_getbuf_func, &tv->tv_buf, - ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, - &tv->tv_flags, tv->tv_zb); - if (tv->tv_err) - return (tv->tv_err); - - tv->tv_osp = tv->tv_buf->b_data; - tv->tv_ldnp = &tv->tv_osp->os_meta_dnode; - tv->tv_err = traverse_dnode(tv->tv_td, tv->tv_ldnp, tv->tv_buf, - tv->tv_zb->zb_objset, DMU_META_DNODE_OBJECT); - if (tv->tv_err && tv->tv_hard) { - tv->tv_lasterr = tv->tv_err; - tv->tv_err = 0; + } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { + uint32_t flags = ARC_WAIT; + objset_phys_t *osp; + dnode_phys_t *dnp; + + 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; + dnp = &osp->os_meta_dnode; + 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; + err = 0; } - if (tv->tv_err == 0 && - arc_buf_size(tv->tv_buf) >= sizeof (objset_phys_t)) { - tv->tv_ldnp = &tv->tv_osp->os_userused_dnode; - tv->tv_err = traverse_dnode(tv->tv_td, tv->tv_ldnp, - tv->tv_buf, tv->tv_zb->zb_objset, + if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { + dnp = &osp->os_userused_dnode; + err = traverse_dnode(td, dnp, zb->zb_objset, DMU_USERUSED_OBJECT); } - if (tv->tv_err && tv->tv_hard) { - tv->tv_lasterr = tv->tv_err; - tv->tv_err = 0; + if (err && hard) { + lasterr = err; + err = 0; } - if (tv->tv_err == 0 && - arc_buf_size(tv->tv_buf) >= sizeof (objset_phys_t)) { - tv->tv_ldnp = &tv->tv_osp->os_groupused_dnode; - tv->tv_err = traverse_dnode(tv->tv_td, tv->tv_ldnp, - tv->tv_buf, tv->tv_zb->zb_objset, + if (err == 0 && arc_buf_size(buf) >= sizeof (objset_phys_t)) { + dnp = &osp->os_groupused_dnode; + err = traverse_dnode(td, dnp, zb->zb_objset, DMU_GROUPUSED_OBJECT); } } - if (tv->tv_buf) - (void) arc_buf_remove_ref(tv->tv_buf, &tv->tv_buf); + if (buf) + (void) arc_buf_remove_ref(buf, &buf); - if (tv->tv_err == 0 && tv->tv_lasterr == 0 && - (tv->tv_td->td_flags & TRAVERSE_POST)) { - tv->tv_err = tv->tv_td->td_func(tv->tv_td->td_spa, NULL, - tv->tv_bp, tv->tv_pbuf, tv->tv_zb, tv->tv_dnp, - tv->tv_td->td_arg); +post: + if (err == 0 && lasterr == 0 && (td->td_flags & TRAVERSE_POST)) { + err = td->td_func(td->td_spa, NULL, bp, zb, dnp, td->td_arg); + if (err == ERESTART) + pause = B_TRUE; } - return (tv->tv_err != 0 ? tv->tv_err : tv->tv_lasterr); + if (pause && td->td_resume != NULL) { + ASSERT3U(err, ==, ERESTART); + ASSERT(!hard); + traverse_pause(td, zb); + } + + return (err != 0 ? err : lasterr); } -/* - * Due to limited stack space recursive functions are frowned upon in - * the Linux kernel. However, they often are the most elegant solution - * to a problem. The following code preserves the recursive function - * traverse_visitbp() but moves the local variables AND function - * arguments to the heap to minimize the stack frame size. Enough - * space is initially allocated on the stack for 16 levels of recursion. - * This change does ugly-up-the-code but it reduces the worst case - * usage from roughly 2496 bytes to 576 bytes on x86_64 archs. - */ -static int -traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp, - arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb) +static void +prefetch_dnode_metadata(traverse_data_t *td, const dnode_phys_t *dnp, + uint64_t objset, uint64_t object) { - traverse_visitbp_data_t *tv; - int error; - - tv = kmem_zalloc(sizeof(traverse_visitbp_data_t) * - TRAVERSE_VISITBP_MAX_DEPTH, KM_SLEEP); - __traverse_visitbp_init(tv, td, dnp, pbuf, bp, zb, 0); - - error = __traverse_visitbp(tv); + int j; + zbookmark_t czb; - kmem_free(tv, sizeof(traverse_visitbp_data_t) * - TRAVERSE_VISITBP_MAX_DEPTH); + 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); + } - return (error); + 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(traverse_data_t *td, const dnode_phys_t *dnp, - arc_buf_t *buf, uint64_t objset, uint64_t object) + uint64_t objset, uint64_t object) { int j, err = 0, lasterr = 0; zbookmark_t czb; @@ -350,8 +404,7 @@ traverse_dnode(traverse_data_t *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; @@ -360,10 +413,8 @@ traverse_dnode(traverse_data_t *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); @@ -376,8 +427,7 @@ traverse_dnode(traverse_data_t *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) { prefetch_data_t *pfd = arg; uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; @@ -398,10 +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); } @@ -419,7 +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; @@ -432,22 +480,33 @@ traverse_prefetch_thread(void *arg) * in syncing context). */ static int -traverse_impl(spa_t *spa, dsl_dataset_t *ds, 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) { traverse_data_t *td; prefetch_data_t *pd; zbookmark_t *czb; int err; - td = kmem_alloc(sizeof(traverse_data_t), KM_SLEEP); - pd = kmem_zalloc(sizeof(prefetch_data_t), KM_SLEEP); - czb = kmem_alloc(sizeof(zbookmark_t), KM_SLEEP); + 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 = ds ? ds->ds_object : 0; + 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; @@ -469,14 +528,14 @@ traverse_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *rootbp, traverse_zil(td, &os->os_zil_header); } - if (!(flags & TRAVERSE_PREFETCH) || + if (!(flags & TRAVERSE_PREFETCH_DATA) || 0 == taskq_dispatch(system_taskq, traverse_prefetch_thread, td, TQ_NOQUEUE)) pd->pd_exited = B_TRUE; 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; @@ -503,8 +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->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)); } /* @@ -521,8 +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, NULL, 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); @@ -571,4 +639,7 @@ traverse_pool(spa_t *spa, uint64_t txg_start, int flags, #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