Print this page
3740 Poor ZFS send / receive performance due to snapshot hold / release processing
Submitted by: Steven Hartland <steven.hartland@multiplay.co.uk>

*** 19,28 **** --- 19,29 ---- * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2013 by Delphix. All rights reserved. + * Copyright (c) 2013 Steven Hartland. All rights reserved. */ #include <sys/zfs_context.h> #include <sys/dsl_userhold.h> #include <sys/dsl_dataset.h>
*** 35,44 **** --- 36,46 ---- #include <sys/zfs_ioctl.h> #include <sys/zap.h> typedef struct dsl_dataset_user_hold_arg { nvlist_t *dduha_holds; + nvlist_t *dduha_chkholds; nvlist_t *dduha_errlist; minor_t dduha_minor; } dsl_dataset_user_hold_arg_t; /*
*** 51,135 **** { dsl_pool_t *dp = dmu_tx_pool(tx); objset_t *mos = dp->dp_meta_objset; int error = 0; if (strlen(htag) > MAXNAMELEN) return (E2BIG); /* Tempholds have a more restricted length */ if (temphold && strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN) return (E2BIG); /* tags must be unique (if ds already exists) */ ! if (ds != NULL) { ! mutex_enter(&ds->ds_lock); ! if (ds->ds_phys->ds_userrefs_obj != 0) { uint64_t value; error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag, 8, 1, &value); if (error == 0) error = SET_ERROR(EEXIST); else if (error == ENOENT) error = 0; } - mutex_exit(&ds->ds_lock); - } return (error); } static int dsl_dataset_user_hold_check(void *arg, dmu_tx_t *tx) { dsl_dataset_user_hold_arg_t *dduha = arg; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; - int rv = 0; if (spa_version(dp->dp_spa) < SPA_VERSION_USERREFS) return (SET_ERROR(ENOTSUP)); for (pair = nvlist_next_nvpair(dduha->dduha_holds, NULL); pair != NULL; pair = nvlist_next_nvpair(dduha->dduha_holds, pair)) { - int error = 0; dsl_dataset_t *ds; ! char *htag; /* must be a snapshot */ ! if (strchr(nvpair_name(pair), '@') == NULL) error = SET_ERROR(EINVAL); if (error == 0) error = nvpair_value_string(pair, &htag); ! if (error == 0) { ! error = dsl_dataset_hold(dp, ! nvpair_name(pair), FTAG, &ds); ! } if (error == 0) { error = dsl_dataset_user_hold_check_one(ds, htag, dduha->dduha_minor != 0, tx); dsl_dataset_rele(ds, FTAG); } ! if (error != 0) { ! rv = error; ! fnvlist_add_int32(dduha->dduha_errlist, ! nvpair_name(pair), error); } } ! return (rv); } ! void ! dsl_dataset_user_hold_sync_one(dsl_dataset_t *ds, const char *htag, ! minor_t minor, uint64_t now, dmu_tx_t *tx) { dsl_pool_t *dp = ds->ds_dir->dd_pool; objset_t *mos = dp->dp_meta_objset; uint64_t zapobj; ! mutex_enter(&ds->ds_lock); if (ds->ds_phys->ds_userrefs_obj == 0) { /* * This is the first user hold for this dataset. Create * the userrefs zap object. */ --- 53,152 ---- { dsl_pool_t *dp = dmu_tx_pool(tx); objset_t *mos = dp->dp_meta_objset; int error = 0; + ASSERT(RRW_READ_HELD(&dp->dp_config_rwlock)); + if (strlen(htag) > MAXNAMELEN) return (E2BIG); /* Tempholds have a more restricted length */ if (temphold && strlen(htag) + MAX_TAG_PREFIX_LEN >= MAXNAMELEN) return (E2BIG); /* tags must be unique (if ds already exists) */ ! if (ds != NULL && ds->ds_phys->ds_userrefs_obj != 0) { uint64_t value; + error = zap_lookup(mos, ds->ds_phys->ds_userrefs_obj, htag, 8, 1, &value); if (error == 0) error = SET_ERROR(EEXIST); else if (error == ENOENT) error = 0; } return (error); } static int dsl_dataset_user_hold_check(void *arg, dmu_tx_t *tx) { dsl_dataset_user_hold_arg_t *dduha = arg; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; if (spa_version(dp->dp_spa) < SPA_VERSION_USERREFS) return (SET_ERROR(ENOTSUP)); + if (!dmu_tx_is_syncing(tx)) + return (0); + for (pair = nvlist_next_nvpair(dduha->dduha_holds, NULL); pair != NULL; pair = nvlist_next_nvpair(dduha->dduha_holds, pair)) { dsl_dataset_t *ds; ! int error = 0; ! char *htag, *name; /* must be a snapshot */ ! name = nvpair_name(pair); ! if (strchr(name, '@') == NULL) error = SET_ERROR(EINVAL); if (error == 0) error = nvpair_value_string(pair, &htag); ! ! if (error == 0) ! error = dsl_dataset_hold(dp, name, FTAG, &ds); ! if (error == 0) { error = dsl_dataset_user_hold_check_one(ds, htag, dduha->dduha_minor != 0, tx); dsl_dataset_rele(ds, FTAG); } ! if (error == 0) { ! fnvlist_add_string(dduha->dduha_chkholds, name, htag); ! } else { ! /* ! * We register ENOENT errors so they can be correctly ! * reported if needed, such as when all holds fail. ! */ ! fnvlist_add_int32(dduha->dduha_errlist, name, error); ! if (error != ENOENT) ! return (error); } } ! ! /* Return ENOENT if no holds would be created. */ ! if (nvlist_next_nvpair(dduha->dduha_chkholds, NULL) == NULL) ! return (ENOENT); ! ! return (0); } ! ! static void ! dsl_dataset_user_hold_sync_one_impl(nvlist_t *tmpholds, dsl_dataset_t *ds, ! const char *htag, minor_t minor, uint64_t now, dmu_tx_t *tx) { dsl_pool_t *dp = ds->ds_dir->dd_pool; objset_t *mos = dp->dp_meta_objset; uint64_t zapobj; ! ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock)); ! if (ds->ds_phys->ds_userrefs_obj == 0) { /* * This is the first user hold for this dataset. Create * the userrefs zap object. */
*** 138,504 **** zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx); } else { zapobj = ds->ds_phys->ds_userrefs_obj; } ds->ds_userrefs++; - mutex_exit(&ds->ds_lock); VERIFY0(zap_add(mos, zapobj, htag, 8, 1, &now, tx)); if (minor != 0) { VERIFY0(dsl_pool_user_hold(dp, ds->ds_object, htag, now, tx)); ! dsl_register_onexit_hold_cleanup(ds, htag, minor); } spa_history_log_internal_ds(ds, "hold", tx, "tag=%s temp=%d refs=%llu", htag, minor != 0, ds->ds_userrefs); } static void dsl_dataset_user_hold_sync(void *arg, dmu_tx_t *tx) { dsl_dataset_user_hold_arg_t *dduha = arg; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; uint64_t now = gethrestime_sec(); ! for (pair = nvlist_next_nvpair(dduha->dduha_holds, NULL); pair != NULL; ! pair = nvlist_next_nvpair(dduha->dduha_holds, pair)) { dsl_dataset_t *ds; VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds)); ! dsl_dataset_user_hold_sync_one(ds, fnvpair_value_string(pair), ! dduha->dduha_minor, now, tx); dsl_dataset_rele(ds, FTAG); } } /* * holds is nvl of snapname -> holdname * errlist will be filled in with snapname -> error - * if cleanup_minor is not 0, the holds will be temporary, cleaned up - * when the process exits. * ! * if any fails, all will fail. */ int dsl_dataset_user_hold(nvlist_t *holds, minor_t cleanup_minor, nvlist_t *errlist) { dsl_dataset_user_hold_arg_t dduha; nvpair_t *pair; pair = nvlist_next_nvpair(holds, NULL); if (pair == NULL) return (0); dduha.dduha_holds = holds; dduha.dduha_errlist = errlist; dduha.dduha_minor = cleanup_minor; ! return (dsl_sync_task(nvpair_name(pair), dsl_dataset_user_hold_check, ! dsl_dataset_user_hold_sync, &dduha, fnvlist_num_pairs(holds))); } typedef struct dsl_dataset_user_release_arg { nvlist_t *ddura_holds; nvlist_t *ddura_todelete; nvlist_t *ddura_errlist; } dsl_dataset_user_release_arg_t; static int ! dsl_dataset_user_release_check_one(dsl_dataset_t *ds, ! nvlist_t *holds, boolean_t *todelete) { uint64_t zapobj; nvpair_t *pair; objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset; ! int error; ! int numholds = 0; ! ! *todelete = B_FALSE; if (!dsl_dataset_is_snapshot(ds)) return (SET_ERROR(EINVAL)); zapobj = ds->ds_phys->ds_userrefs_obj; if (zapobj == 0) return (SET_ERROR(ESRCH)); for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; pair = nvlist_next_nvpair(holds, pair)) { - /* Make sure the hold exists */ uint64_t tmp; ! error = zap_lookup(mos, zapobj, nvpair_name(pair), 8, 1, &tmp); if (error == ENOENT) ! error = SET_ERROR(ESRCH); ! if (error != 0) return (error); numholds++; } if (DS_IS_DEFER_DESTROY(ds) && ds->ds_phys->ds_num_children == 1 && ds->ds_userrefs == numholds) { /* we need to destroy the snapshot as well */ ! ! if (dsl_dataset_long_held(ds)) return (SET_ERROR(EBUSY)); - *todelete = B_TRUE; } ! return (0); } static int dsl_dataset_user_release_check(void *arg, dmu_tx_t *tx) { ! dsl_dataset_user_release_arg_t *ddura = arg; ! dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; - int rv = 0; if (!dmu_tx_is_syncing(tx)) return (0); for (pair = nvlist_next_nvpair(ddura->ddura_holds, NULL); pair != NULL; pair = nvlist_next_nvpair(ddura->ddura_holds, pair)) { ! const char *name = nvpair_name(pair); int error; dsl_dataset_t *ds; nvlist_t *holds; error = nvpair_value_nvlist(pair, &holds); if (error != 0) ! return (SET_ERROR(EINVAL)); ! ! error = dsl_dataset_hold(dp, name, FTAG, &ds); if (error == 0) { ! boolean_t deleteme; ! error = dsl_dataset_user_release_check_one(ds, ! holds, &deleteme); ! if (error == 0 && deleteme) { ! fnvlist_add_boolean(ddura->ddura_todelete, ! name); ! } dsl_dataset_rele(ds, FTAG); } if (error != 0) { if (ddura->ddura_errlist != NULL) { ! fnvlist_add_int32(ddura->ddura_errlist, ! name, error); } ! rv = error; } } ! return (rv); } static void ! dsl_dataset_user_release_sync_one(dsl_dataset_t *ds, nvlist_t *holds, ! dmu_tx_t *tx) { dsl_pool_t *dp = ds->ds_dir->dd_pool; objset_t *mos = dp->dp_meta_objset; - uint64_t zapobj; - int error; nvpair_t *pair; for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; pair = nvlist_next_nvpair(holds, pair)) { ! ds->ds_userrefs--; ! error = dsl_pool_user_release(dp, ds->ds_object, ! nvpair_name(pair), tx); VERIFY(error == 0 || error == ENOENT); ! zapobj = ds->ds_phys->ds_userrefs_obj; ! VERIFY0(zap_remove(mos, zapobj, nvpair_name(pair), tx)); spa_history_log_internal_ds(ds, "release", tx, ! "tag=%s refs=%lld", nvpair_name(pair), ! (longlong_t)ds->ds_userrefs); } } static void dsl_dataset_user_release_sync(void *arg, dmu_tx_t *tx) { dsl_dataset_user_release_arg_t *ddura = arg; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; ! for (pair = nvlist_next_nvpair(ddura->ddura_holds, NULL); pair != NULL; ! pair = nvlist_next_nvpair(ddura->ddura_holds, pair)) { dsl_dataset_t *ds; ! VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds)); ! dsl_dataset_user_release_sync_one(ds, fnvpair_value_nvlist(pair), tx); ! if (nvlist_exists(ddura->ddura_todelete, ! nvpair_name(pair))) { ASSERT(ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 && DS_IS_DEFER_DESTROY(ds)); dsl_destroy_snapshot_sync_impl(ds, B_FALSE, tx); } dsl_dataset_rele(ds, FTAG); } } /* * holds is nvl of snapname -> { holdname, ... } * errlist will be filled in with snapname -> error * ! * if any fails, all will fail. */ ! int ! dsl_dataset_user_release(nvlist_t *holds, nvlist_t *errlist) { dsl_dataset_user_release_arg_t ddura; nvpair_t *pair; int error; pair = nvlist_next_nvpair(holds, NULL); if (pair == NULL) return (0); ! ddura.ddura_holds = holds; ! ddura.ddura_errlist = errlist; ! ddura.ddura_todelete = fnvlist_alloc(); ! ! error = dsl_sync_task(nvpair_name(pair), dsl_dataset_user_release_check, ! dsl_dataset_user_release_sync, &ddura, fnvlist_num_pairs(holds)); ! fnvlist_free(ddura.ddura_todelete); ! return (error); ! } ! ! typedef struct dsl_dataset_user_release_tmp_arg { ! uint64_t ddurta_dsobj; ! nvlist_t *ddurta_holds; ! boolean_t ddurta_deleteme; ! } dsl_dataset_user_release_tmp_arg_t; ! ! static int ! dsl_dataset_user_release_tmp_check(void *arg, dmu_tx_t *tx) ! { ! dsl_dataset_user_release_tmp_arg_t *ddurta = arg; ! dsl_pool_t *dp = dmu_tx_pool(tx); ! dsl_dataset_t *ds; ! int error; ! ! if (!dmu_tx_is_syncing(tx)) ! return (0); ! ! error = dsl_dataset_hold_obj(dp, ddurta->ddurta_dsobj, FTAG, &ds); ! if (error) ! return (error); ! ! error = dsl_dataset_user_release_check_one(ds, ! ddurta->ddurta_holds, &ddurta->ddurta_deleteme); ! dsl_dataset_rele(ds, FTAG); ! return (error); ! } ! ! static void ! dsl_dataset_user_release_tmp_sync(void *arg, dmu_tx_t *tx) ! { ! dsl_dataset_user_release_tmp_arg_t *ddurta = arg; ! dsl_pool_t *dp = dmu_tx_pool(tx); ! dsl_dataset_t *ds; ! ! VERIFY0(dsl_dataset_hold_obj(dp, ddurta->ddurta_dsobj, FTAG, &ds)); ! dsl_dataset_user_release_sync_one(ds, ddurta->ddurta_holds, tx); ! if (ddurta->ddurta_deleteme) { ! ASSERT(ds->ds_userrefs == 0 && ! ds->ds_phys->ds_num_children == 1 && ! DS_IS_DEFER_DESTROY(ds)); ! dsl_destroy_snapshot_sync_impl(ds, B_FALSE, tx); ! } ! dsl_dataset_rele(ds, FTAG); ! } ! ! /* ! * Called at spa_load time to release a stale temporary user hold. ! * Also called by the onexit code. */ ! void ! dsl_dataset_user_release_tmp(dsl_pool_t *dp, uint64_t dsobj, const char *htag) ! { ! dsl_dataset_user_release_tmp_arg_t ddurta; dsl_dataset_t *ds; - int error; ! #ifdef _KERNEL ! /* Make sure it is not mounted. */ ! dsl_pool_config_enter(dp, FTAG); ! error = dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds); if (error == 0) { char name[MAXNAMELEN]; dsl_dataset_name(ds, name); dsl_dataset_rele(ds, FTAG); - dsl_pool_config_exit(dp, FTAG); zfs_unmount_snap(name); } else { ! dsl_pool_config_exit(dp, FTAG); } #endif ! ddurta.ddurta_dsobj = dsobj; ! ddurta.ddurta_holds = fnvlist_alloc(); ! fnvlist_add_boolean(ddurta.ddurta_holds, htag); ! ! (void) dsl_sync_task(spa_name(dp->dp_spa), ! dsl_dataset_user_release_tmp_check, ! dsl_dataset_user_release_tmp_sync, &ddurta, 1); ! fnvlist_free(ddurta.ddurta_holds); ! } ! ! typedef struct zfs_hold_cleanup_arg { ! char zhca_spaname[MAXNAMELEN]; ! uint64_t zhca_spa_load_guid; ! uint64_t zhca_dsobj; ! char zhca_htag[MAXNAMELEN]; ! } zfs_hold_cleanup_arg_t; ! static void ! dsl_dataset_user_release_onexit(void *arg) ! { ! zfs_hold_cleanup_arg_t *ca = arg; ! spa_t *spa; ! int error; ! error = spa_open(ca->zhca_spaname, &spa, FTAG); ! if (error != 0) { ! zfs_dbgmsg("couldn't release hold on pool=%s ds=%llu tag=%s " ! "because pool is no longer loaded", ! ca->zhca_spaname, ca->zhca_dsobj, ca->zhca_htag); ! return; ! } ! if (spa_load_guid(spa) != ca->zhca_spa_load_guid) { ! zfs_dbgmsg("couldn't release hold on pool=%s ds=%llu tag=%s " ! "because pool is no longer loaded (guid doesn't match)", ! ca->zhca_spaname, ca->zhca_dsobj, ca->zhca_htag); ! spa_close(spa, FTAG); ! return; ! } ! dsl_dataset_user_release_tmp(spa_get_dsl(spa), ! ca->zhca_dsobj, ca->zhca_htag); ! kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t)); ! spa_close(spa, FTAG); } void ! dsl_register_onexit_hold_cleanup(dsl_dataset_t *ds, const char *htag, ! minor_t minor) { ! zfs_hold_cleanup_arg_t *ca = kmem_alloc(sizeof (*ca), KM_SLEEP); ! spa_t *spa = dsl_dataset_get_spa(ds); ! (void) strlcpy(ca->zhca_spaname, spa_name(spa), ! sizeof (ca->zhca_spaname)); ! ca->zhca_spa_load_guid = spa_load_guid(spa); ! ca->zhca_dsobj = ds->ds_object; ! (void) strlcpy(ca->zhca_htag, htag, sizeof (ca->zhca_htag)); ! VERIFY0(zfs_onexit_add_cb(minor, ! dsl_dataset_user_release_onexit, ca, NULL)); } int dsl_dataset_get_holds(const char *dsname, nvlist_t *nvl) { --- 155,635 ---- zap_create(mos, DMU_OT_USERREFS, DMU_OT_NONE, 0, tx); } else { zapobj = ds->ds_phys->ds_userrefs_obj; } ds->ds_userrefs++; VERIFY0(zap_add(mos, zapobj, htag, 8, 1, &now, tx)); if (minor != 0) { + char name[MAXNAMELEN]; + nvlist_t *tags; + VERIFY0(dsl_pool_user_hold(dp, ds->ds_object, htag, now, tx)); ! (void) snprintf(name, sizeof (name), "%llx", ! (u_longlong_t)ds->ds_object); ! ! if (nvlist_lookup_nvlist(tmpholds, name, &tags) != 0) { ! tags = fnvlist_alloc(); ! fnvlist_add_boolean(tags, htag); ! fnvlist_add_nvlist(tmpholds, name, tags); ! fnvlist_free(tags); ! } else { ! fnvlist_add_boolean(tags, htag); ! } } spa_history_log_internal_ds(ds, "hold", tx, "tag=%s temp=%d refs=%llu", htag, minor != 0, ds->ds_userrefs); } + typedef struct zfs_hold_cleanup_arg { + char zhca_spaname[MAXNAMELEN]; + uint64_t zhca_spa_load_guid; + nvlist_t *zhca_holds; + } zfs_hold_cleanup_arg_t; + + static void + dsl_dataset_user_release_onexit(void *arg) + { + zfs_hold_cleanup_arg_t *ca = (zfs_hold_cleanup_arg_t *)arg; + spa_t *spa; + int error; + + error = spa_open(ca->zhca_spaname, &spa, FTAG); + if (error != 0) { + zfs_dbgmsg("couldn't release holds on pool=%s " + "because pool is no longer loaded", + ca->zhca_spaname); + return; + } + if (spa_load_guid(spa) != ca->zhca_spa_load_guid) { + zfs_dbgmsg("couldn't release holds on pool=%s " + "because pool is no longer loaded (guid doesn't match)", + ca->zhca_spaname); + spa_close(spa, FTAG); + return; + } + + (void) dsl_dataset_user_release_tmp(spa_get_dsl(spa), ca->zhca_holds); + fnvlist_free(ca->zhca_holds); + kmem_free(ca, sizeof (zfs_hold_cleanup_arg_t)); + spa_close(spa, FTAG); + } + + static void + dsl_onexit_hold_cleanup(spa_t *spa, nvlist_t *holds, minor_t minor) + { + zfs_hold_cleanup_arg_t *ca; + + if (minor == 0 || nvlist_next_nvpair(holds, NULL) == NULL) { + fnvlist_free(holds); + return; + } + + ASSERT(spa != NULL); + ca = kmem_alloc(sizeof (*ca), KM_SLEEP); + + (void) strlcpy(ca->zhca_spaname, spa_name(spa), + sizeof (ca->zhca_spaname)); + ca->zhca_spa_load_guid = spa_load_guid(spa); + ca->zhca_holds = holds; + VERIFY0(zfs_onexit_add_cb(minor, + dsl_dataset_user_release_onexit, ca, NULL)); + } + + void + dsl_dataset_user_hold_sync_one(dsl_dataset_t *ds, const char *htag, + minor_t minor, uint64_t now, dmu_tx_t *tx) + { + nvlist_t *tmpholds; + + if (minor != 0) + tmpholds = fnvlist_alloc(); + else + tmpholds = NULL; + dsl_dataset_user_hold_sync_one_impl(tmpholds, ds, htag, minor, now, tx); + dsl_onexit_hold_cleanup(dsl_dataset_get_spa(ds), tmpholds, minor); + } + static void dsl_dataset_user_hold_sync(void *arg, dmu_tx_t *tx) { dsl_dataset_user_hold_arg_t *dduha = arg; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; + nvlist_t *tmpholds; uint64_t now = gethrestime_sec(); ! if (dduha->dduha_minor != 0) ! tmpholds = fnvlist_alloc(); ! else ! tmpholds = NULL; ! for (pair = nvlist_next_nvpair(dduha->dduha_chkholds, NULL); ! pair != NULL; ! pair = nvlist_next_nvpair(dduha->dduha_chkholds, pair)) { dsl_dataset_t *ds; + VERIFY0(dsl_dataset_hold(dp, nvpair_name(pair), FTAG, &ds)); ! dsl_dataset_user_hold_sync_one_impl(tmpholds, ds, ! fnvpair_value_string(pair), dduha->dduha_minor, now, tx); dsl_dataset_rele(ds, FTAG); } + dsl_onexit_hold_cleanup(dp->dp_spa, tmpholds, dduha->dduha_minor); } /* + * The full semantics of this function are described in the comment above + * lzc_hold(). + * + * To summarize: * holds is nvl of snapname -> holdname * errlist will be filled in with snapname -> error * ! * The snaphosts must all be in the same pool. ! * ! * Holds for snapshots that don't exist will be skipped. ! * ! * If none of the snapshots for requested holds exist then ENOENT will be ! * returned. ! * ! * If cleanup_minor is not 0, the holds will be temporary, which will be cleaned ! * up when the process exits. ! * ! * On success all the holds, for snapshots that existed, will be created and 0 ! * will be returned. ! * ! * On failure no holds will be created, the errlist will be filled in, ! * and an errno will returned. ! * ! * In all cases the errlist will contain entries for holds where the snapshot ! * didn't exist. */ int dsl_dataset_user_hold(nvlist_t *holds, minor_t cleanup_minor, nvlist_t *errlist) { dsl_dataset_user_hold_arg_t dduha; nvpair_t *pair; + int ret; pair = nvlist_next_nvpair(holds, NULL); if (pair == NULL) return (0); dduha.dduha_holds = holds; + dduha.dduha_chkholds = fnvlist_alloc(); dduha.dduha_errlist = errlist; dduha.dduha_minor = cleanup_minor; ! ret = dsl_sync_task(nvpair_name(pair), dsl_dataset_user_hold_check, ! dsl_dataset_user_hold_sync, &dduha, fnvlist_num_pairs(holds)); ! fnvlist_free(dduha.dduha_chkholds); ! ! return (ret); } + typedef int (dsl_holdfunc_t)(dsl_pool_t *dp, const char *name, void *tag, + dsl_dataset_t **dsp); + typedef struct dsl_dataset_user_release_arg { + dsl_holdfunc_t *ddura_holdfunc; nvlist_t *ddura_holds; nvlist_t *ddura_todelete; nvlist_t *ddura_errlist; + nvlist_t *ddura_chkholds; } dsl_dataset_user_release_arg_t; + /* Place a dataset hold on the snapshot identified by passed dsobj string */ + static int + dsl_dataset_hold_obj_string(dsl_pool_t *dp, const char *dsobj, void *tag, + dsl_dataset_t **dsp) + { + return (dsl_dataset_hold_obj(dp, strtonum(dsobj, NULL), tag, dsp)); + } + static int ! dsl_dataset_user_release_check_one(dsl_dataset_user_release_arg_t *ddura, ! dsl_dataset_t *ds, nvlist_t *holds, const char *name) { uint64_t zapobj; nvpair_t *pair; + nvlist_t *holds_found; objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset; ! int ret, numholds; if (!dsl_dataset_is_snapshot(ds)) return (SET_ERROR(EINVAL)); zapobj = ds->ds_phys->ds_userrefs_obj; if (zapobj == 0) return (SET_ERROR(ESRCH)); + ret = 0; + numholds = 0; + holds_found = fnvlist_alloc(); + for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; pair = nvlist_next_nvpair(holds, pair)) { uint64_t tmp; ! int error; ! const char *name; ! ! name = nvpair_name(pair); ! error = zap_lookup(mos, zapobj, name, 8, 1, &tmp); ! ! /* Non-existent holds aren't always an error. */ if (error == ENOENT) ! continue; ! ! if (error != 0) { ! fnvlist_free(holds_found); return (error); + } + + fnvlist_add_boolean(holds_found, name); numholds++; } if (DS_IS_DEFER_DESTROY(ds) && ds->ds_phys->ds_num_children == 1 && ds->ds_userrefs == numholds) { /* we need to destroy the snapshot as well */ ! if (dsl_dataset_long_held(ds)) { ! fnvlist_free(holds_found); return (SET_ERROR(EBUSY)); } ! fnvlist_add_boolean(ddura->ddura_todelete, name); ! } ! ! if (numholds == 0) ! ret = ENOENT; ! else ! fnvlist_add_nvlist(ddura->ddura_chkholds, name, holds_found); ! fnvlist_free(holds_found); ! ! return (ret); } static int dsl_dataset_user_release_check(void *arg, dmu_tx_t *tx) { ! dsl_dataset_user_release_arg_t *ddura; ! dsl_holdfunc_t *holdfunc; ! dsl_pool_t *dp; nvpair_t *pair; if (!dmu_tx_is_syncing(tx)) return (0); + ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock)); + + dp = dmu_tx_pool(tx); + ddura = (dsl_dataset_user_release_arg_t *)arg; + holdfunc = ddura->ddura_holdfunc; + for (pair = nvlist_next_nvpair(ddura->ddura_holds, NULL); pair != NULL; pair = nvlist_next_nvpair(ddura->ddura_holds, pair)) { ! const char *name; int error; dsl_dataset_t *ds; nvlist_t *holds; + name = nvpair_name(pair); error = nvpair_value_nvlist(pair, &holds); if (error != 0) ! error = (SET_ERROR(EINVAL)); ! if (error == 0) ! error = holdfunc(dp, name, FTAG, &ds); if (error == 0) { ! error = dsl_dataset_user_release_check_one(ddura, ds, ! holds, name); dsl_dataset_rele(ds, FTAG); } if (error != 0) { if (ddura->ddura_errlist != NULL) { ! fnvlist_add_int32(ddura->ddura_errlist, name, ! error); } ! /* Non-existent holds aren't always an error. */ ! if (error != ENOENT) ! return (error); } } ! ! /* ! * Return ENOENT if none of the holds existed avoiding the overhead ! * of a sync. ! */ ! if (nvlist_next_nvpair(ddura->ddura_chkholds, NULL) == NULL) ! return (ENOENT); ! ! return (0); } static void ! dsl_dataset_user_release_sync_one(dsl_dataset_user_release_arg_t *ddura, ! dsl_dataset_t *ds, nvlist_t *holds, dmu_tx_t *tx) { dsl_pool_t *dp = ds->ds_dir->dd_pool; objset_t *mos = dp->dp_meta_objset; nvpair_t *pair; for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; pair = nvlist_next_nvpair(holds, pair)) { ! int error; ! const char *name; ! ! name = nvpair_name(pair); ! ! /* Remove temporary hold if one exists. */ ! error = dsl_pool_user_release(dp, ds->ds_object, name, tx); VERIFY(error == 0 || error == ENOENT); ! ! VERIFY0(zap_remove(mos, ds->ds_phys->ds_userrefs_obj, name, ! tx)); ! ds->ds_userrefs--; spa_history_log_internal_ds(ds, "release", tx, ! "tag=%s refs=%lld", name, (longlong_t)ds->ds_userrefs); } } static void dsl_dataset_user_release_sync(void *arg, dmu_tx_t *tx) { dsl_dataset_user_release_arg_t *ddura = arg; + dsl_holdfunc_t *holdfunc = ddura->ddura_holdfunc; dsl_pool_t *dp = dmu_tx_pool(tx); nvpair_t *pair; ! ASSERT(RRW_WRITE_HELD(&dp->dp_config_rwlock)); ! ! for (pair = nvlist_next_nvpair(ddura->ddura_chkholds, NULL); ! pair != NULL; pair = nvlist_next_nvpair(ddura->ddura_chkholds, ! pair)) { dsl_dataset_t *ds; + const char *name; ! name = nvpair_name(pair); ! VERIFY0(holdfunc(dp, name, FTAG, &ds)); ! ! dsl_dataset_user_release_sync_one(ddura, ds, fnvpair_value_nvlist(pair), tx); ! if (nvlist_exists(ddura->ddura_todelete, name)) { ASSERT(ds->ds_userrefs == 0 && ds->ds_phys->ds_num_children == 1 && DS_IS_DEFER_DESTROY(ds)); dsl_destroy_snapshot_sync_impl(ds, B_FALSE, tx); } dsl_dataset_rele(ds, FTAG); } } /* + * The full semantics of this function are described in the comment above + * lzc_release(). + * + * To summarize: + * Releases holds specified in the nvl holds. + * * holds is nvl of snapname -> { holdname, ... } * errlist will be filled in with snapname -> error * ! * If tmpdp is not NULL the names for holds should be the dsobj's of snapshots, ! * otherwise they should be the names of shapshots. ! * ! * As a release may cause snapshots to be destroyed this trys to ensure they ! * aren't mounted. ! * ! * The release of non-existent holds are skipped. ! * ! * At least one hold must have been released for the this function to succeed ! * and return 0. */ ! static int ! dsl_dataset_user_release_impl(nvlist_t *holds, nvlist_t *errlist, ! dsl_pool_t *tmpdp) { dsl_dataset_user_release_arg_t ddura; nvpair_t *pair; + char *pool; int error; pair = nvlist_next_nvpair(holds, NULL); if (pair == NULL) return (0); ! #ifdef _KERNEL ! /* ! * The release may cause snapshots to be destroyed; make sure they ! * are not mounted. */ ! if (tmpdp != NULL) { ! /* Temporary holds are specified by dsobj string. */ ! ddura.ddura_holdfunc = dsl_dataset_hold_obj_string; ! pool = spa_name(tmpdp->dp_spa); ! ! dsl_pool_config_enter(tmpdp, FTAG); ! for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; ! pair = nvlist_next_nvpair(holds, pair)) { dsl_dataset_t *ds; ! error = dsl_dataset_hold_obj_string(tmpdp, ! nvpair_name(pair), FTAG, &ds); if (error == 0) { char name[MAXNAMELEN]; dsl_dataset_name(ds, name); dsl_dataset_rele(ds, FTAG); zfs_unmount_snap(name); + } + } + dsl_pool_config_exit(tmpdp, FTAG); } else { ! /* Non-temporary holds are specified by name. */ ! ddura.ddura_holdfunc = dsl_dataset_hold; ! pool = nvpair_name(pair); ! ! for (pair = nvlist_next_nvpair(holds, NULL); pair != NULL; ! pair = nvlist_next_nvpair(holds, pair)) { ! zfs_unmount_snap(nvpair_name(pair)); ! } } #endif ! ddura.ddura_holds = holds; ! ddura.ddura_errlist = errlist; ! ddura.ddura_todelete = fnvlist_alloc(); ! ddura.ddura_chkholds = fnvlist_alloc(); ! error = dsl_sync_task(pool, dsl_dataset_user_release_check, ! dsl_dataset_user_release_sync, &ddura, ! fnvlist_num_pairs(holds)); ! fnvlist_free(ddura.ddura_todelete); ! fnvlist_free(ddura.ddura_chkholds); ! return (error); ! } ! /* ! * holds is nvl of snapname -> { holdname, ... } ! * errlist will be filled in with snapname -> error ! */ ! int ! dsl_dataset_user_release(nvlist_t *holds, nvlist_t *errlist) ! { ! return (dsl_dataset_user_release_impl(holds, errlist, NULL)); } + /* + * holds is nvl of snapdsobj -> { holdname, ... } + */ void ! dsl_dataset_user_release_tmp(struct dsl_pool *dp, nvlist_t *holds) { ! ASSERT(dp != NULL); ! (void) dsl_dataset_user_release_impl(holds, NULL, dp); } int dsl_dataset_get_holds(const char *dsname, nvlist_t *nvl) {