diff options
Diffstat (limited to 'module/zfs/vdev.c')
| -rw-r--r-- | module/zfs/vdev.c | 225 |
1 files changed, 200 insertions, 25 deletions
diff --git a/module/zfs/vdev.c b/module/zfs/vdev.c index 7ffe924212da..36001e0a6626 100644 --- a/module/zfs/vdev.c +++ b/module/zfs/vdev.c @@ -59,6 +59,27 @@ #include <sys/zvol.h> #include <sys/zfs_ratelimit.h> +/* + * One metaslab from each (normal-class) vdev is used by the ZIL. These are + * called "embedded slog metaslabs", are referenced by vdev_log_mg, and are + * part of the spa_embedded_log_class. The metaslab with the most free space + * in each vdev is selected for this purpose when the pool is opened (or a + * vdev is added). See vdev_metaslab_init(). + * + * Log blocks can be allocated from the following locations. Each one is tried + * in order until the allocation succeeds: + * 1. dedicated log vdevs, aka "slog" (spa_log_class) + * 2. embedded slog metaslabs (spa_embedded_log_class) + * 3. other metaslabs in normal vdevs (spa_normal_class) + * + * zfs_embedded_slog_min_ms disables the embedded slog if there are fewer + * than this number of metaslabs in the vdev. This ensures that we don't set + * aside an unreasonable amount of space for the ZIL. If set to less than + * 1 << (spa_slop_shift + 1), on small pools the usable space may be reduced + * (by more than 1<<spa_slop_shift) due to the embedded slog metaslab. + */ +int zfs_embedded_slog_min_ms = 64; + /* default target for number of metaslabs per top-level vdev */ int zfs_vdev_default_ms_count = 200; @@ -223,6 +244,22 @@ vdev_getops(const char *type) return (ops); } +/* + * Given a vdev and a metaslab class, find which metaslab group we're + * interested in. All vdevs may belong to two different metaslab classes. + * Dedicated slog devices use only the primary metaslab group, rather than a + * separate log group. For embedded slogs, the vdev_log_mg will be non-NULL. + */ +metaslab_group_t * +vdev_get_mg(vdev_t *vd, metaslab_class_t *mc) +{ + if (mc == spa_embedded_log_class(vd->vdev_spa) && + vd->vdev_log_mg != NULL) + return (vd->vdev_log_mg); + else + return (vd->vdev_mg); +} + /* ARGSUSED */ void vdev_default_xlate(vdev_t *vd, const range_seg64_t *logical_rs, @@ -978,6 +1015,11 @@ vdev_free(vdev_t *vd) metaslab_group_destroy(vd->vdev_mg); vd->vdev_mg = NULL; } + if (vd->vdev_log_mg != NULL) { + ASSERT0(vd->vdev_ms_count); + metaslab_group_destroy(vd->vdev_log_mg); + vd->vdev_log_mg = NULL; + } ASSERT0(vd->vdev_stat.vs_space); ASSERT0(vd->vdev_stat.vs_dspace); @@ -1098,14 +1140,20 @@ vdev_top_transfer(vdev_t *svd, vdev_t *tvd) if (tvd->vdev_mg) ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg); + if (tvd->vdev_log_mg) + ASSERT3P(tvd->vdev_log_mg, ==, svd->vdev_log_mg); tvd->vdev_mg = svd->vdev_mg; + tvd->vdev_log_mg = svd->vdev_log_mg; tvd->vdev_ms = svd->vdev_ms; svd->vdev_mg = NULL; + svd->vdev_log_mg = NULL; svd->vdev_ms = NULL; if (tvd->vdev_mg != NULL) tvd->vdev_mg->mg_vd = tvd; + if (tvd->vdev_log_mg != NULL) + tvd->vdev_log_mg->mg_vd = tvd; tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm; svd->vdev_checkpoint_sm = NULL; @@ -1283,7 +1331,7 @@ vdev_remove_parent(vdev_t *cvd) vdev_free(mvd); } -static void +void vdev_metaslab_group_create(vdev_t *vd) { spa_t *spa = vd->vdev_spa; @@ -1317,6 +1365,11 @@ vdev_metaslab_group_create(vdev_t *vd) vd->vdev_mg = metaslab_group_create(mc, vd, spa->spa_alloc_count); + if (!vd->vdev_islog) { + vd->vdev_log_mg = metaslab_group_create( + spa_embedded_log_class(spa), vd, 1); + } + /* * The spa ashift min/max only apply for the normal metaslab * class. Class destination is late binding so ashift boundry @@ -1340,8 +1393,6 @@ int vdev_metaslab_init(vdev_t *vd, uint64_t txg) { spa_t *spa = vd->vdev_spa; - objset_t *mos = spa->spa_meta_objset; - uint64_t m; uint64_t oldc = vd->vdev_ms_count; uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift; metaslab_t **mspp; @@ -1369,16 +1420,17 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg) vd->vdev_ms = mspp; vd->vdev_ms_count = newc; - for (m = oldc; m < newc; m++) { - uint64_t object = 0; + for (uint64_t m = oldc; m < newc; m++) { + uint64_t object = 0; /* * vdev_ms_array may be 0 if we are creating the "fake" * metaslabs for an indirect vdev for zdb's leak detection. * See zdb_leak_init(). */ if (txg == 0 && vd->vdev_ms_array != 0) { - error = dmu_read(mos, vd->vdev_ms_array, + error = dmu_read(spa->spa_meta_objset, + vd->vdev_ms_array, m * sizeof (uint64_t), sizeof (uint64_t), &object, DMU_READ_PREFETCH); if (error != 0) { @@ -1388,17 +1440,6 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg) } } -#ifndef _KERNEL - /* - * To accommodate zdb_leak_init() fake indirect - * metaslabs, we allocate a metaslab group for - * indirect vdevs which normally don't have one. - */ - if (vd->vdev_mg == NULL) { - ASSERT0(vdev_is_concrete(vd)); - vdev_metaslab_group_create(vd); - } -#endif error = metaslab_init(vd->vdev_mg, m, object, txg, &(vd->vdev_ms[m])); if (error != 0) { @@ -1408,6 +1449,47 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg) } } + /* + * Find the emptiest metaslab on the vdev and mark it for use for + * embedded slog by moving it from the regular to the log metaslab + * group. + */ + if (vd->vdev_mg->mg_class == spa_normal_class(spa) && + vd->vdev_ms_count > zfs_embedded_slog_min_ms && + avl_is_empty(&vd->vdev_log_mg->mg_metaslab_tree)) { + uint64_t slog_msid = 0; + uint64_t smallest = UINT64_MAX; + + /* + * Note, we only search the new metaslabs, because the old + * (pre-existing) ones may be active (e.g. have non-empty + * range_tree's), and we don't move them to the new + * metaslab_t. + */ + for (uint64_t m = oldc; m < newc; m++) { + uint64_t alloc = + space_map_allocated(vd->vdev_ms[m]->ms_sm); + if (alloc < smallest) { + slog_msid = m; + smallest = alloc; + } + } + metaslab_t *slog_ms = vd->vdev_ms[slog_msid]; + /* + * The metaslab was marked as dirty at the end of + * metaslab_init(). Remove it from the dirty list so that we + * can uninitialize and reinitialize it to the new class. + */ + if (txg != 0) { + (void) txg_list_remove_this(&vd->vdev_ms_list, + slog_ms, txg); + } + uint64_t sm_obj = space_map_object(slog_ms->ms_sm); + metaslab_fini(slog_ms); + VERIFY0(metaslab_init(vd->vdev_log_mg, slog_msid, sm_obj, txg, + &vd->vdev_ms[slog_msid])); + } + if (txg == 0) spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER); @@ -1418,6 +1500,8 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg) */ if (!expanding && !vd->vdev_removing) { metaslab_group_activate(vd->vdev_mg); + if (vd->vdev_log_mg != NULL) + metaslab_group_activate(vd->vdev_log_mg); } if (txg == 0) @@ -1453,7 +1537,12 @@ vdev_metaslab_fini(vdev_t *vd) if (vd->vdev_ms != NULL) { metaslab_group_t *mg = vd->vdev_mg; + metaslab_group_passivate(mg); + if (vd->vdev_log_mg != NULL) { + ASSERT(!vd->vdev_islog); + metaslab_group_passivate(vd->vdev_log_mg); + } uint64_t count = vd->vdev_ms_count; for (uint64_t m = 0; m < count; m++) { @@ -1463,11 +1552,13 @@ vdev_metaslab_fini(vdev_t *vd) } vmem_free(vd->vdev_ms, count * sizeof (metaslab_t *)); vd->vdev_ms = NULL; - vd->vdev_ms_count = 0; - for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) + for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) { ASSERT0(mg->mg_histogram[i]); + if (vd->vdev_log_mg != NULL) + ASSERT0(vd->vdev_log_mg->mg_histogram[i]); + } } ASSERT0(vd->vdev_ms_count); ASSERT3U(vd->vdev_pending_fastwrite, ==, 0); @@ -1634,6 +1725,14 @@ vdev_probe(vdev_t *vd, zio_t *zio) } static void +vdev_load_child(void *arg) +{ + vdev_t *vd = arg; + + vd->vdev_load_error = vdev_load(vd); +} + +static void vdev_open_child(void *arg) { vdev_t *vd = arg; @@ -2037,6 +2136,16 @@ vdev_open(vdev_t *vd) return (0); } +static void +vdev_validate_child(void *arg) +{ + vdev_t *vd = arg; + + vd->vdev_validate_thread = curthread; + vd->vdev_validate_error = vdev_validate(vd); + vd->vdev_validate_thread = NULL; +} + /* * Called once the vdevs are all opened, this routine validates the label * contents. This needs to be done before vdev_load() so that we don't @@ -2051,18 +2160,43 @@ int vdev_validate(vdev_t *vd) { spa_t *spa = vd->vdev_spa; + taskq_t *tq = NULL; nvlist_t *label; uint64_t guid = 0, aux_guid = 0, top_guid; uint64_t state; nvlist_t *nvl; uint64_t txg; + int children = vd->vdev_children; if (vdev_validate_skip) return (0); - for (uint64_t c = 0; c < vd->vdev_children; c++) - if (vdev_validate(vd->vdev_child[c]) != 0) + if (children > 0) { + tq = taskq_create("vdev_validate", children, minclsyspri, + children, children, TASKQ_PREPOPULATE); + } + + for (uint64_t c = 0; c < children; c++) { + vdev_t *cvd = vd->vdev_child[c]; + + if (tq == NULL || vdev_uses_zvols(cvd)) { + vdev_validate_child(cvd); + } else { + VERIFY(taskq_dispatch(tq, vdev_validate_child, cvd, + TQ_SLEEP) != TASKQID_INVALID); + } + } + if (tq != NULL) { + taskq_wait(tq); + taskq_destroy(tq); + } + for (int c = 0; c < children; c++) { + int error = vd->vdev_child[c]->vdev_validate_error; + + if (error != 0) return (SET_ERROR(EBADF)); + } + /* * If the device has already failed, or was marked offline, don't do @@ -3259,18 +3393,46 @@ vdev_checkpoint_sm_object(vdev_t *vd, uint64_t *sm_obj) int vdev_load(vdev_t *vd) { + int children = vd->vdev_children; int error = 0; + taskq_t *tq = NULL; + + /* + * It's only worthwhile to use the taskq for the root vdev, because the + * slow part is metaslab_init, and that only happens for top-level + * vdevs. + */ + if (vd->vdev_ops == &vdev_root_ops && vd->vdev_children > 0) { + tq = taskq_create("vdev_load", children, minclsyspri, + children, children, TASKQ_PREPOPULATE); + } /* * Recursively load all children. */ for (int c = 0; c < vd->vdev_children; c++) { - error = vdev_load(vd->vdev_child[c]); - if (error != 0) { - return (error); + vdev_t *cvd = vd->vdev_child[c]; + + if (tq == NULL || vdev_uses_zvols(cvd)) { + cvd->vdev_load_error = vdev_load(cvd); + } else { + VERIFY(taskq_dispatch(tq, vdev_load_child, + cvd, TQ_SLEEP) != TASKQID_INVALID); } } + if (tq != NULL) { + taskq_wait(tq); + taskq_destroy(tq); + } + + for (int c = 0; c < vd->vdev_children; c++) { + int error = vd->vdev_child[c]->vdev_load_error; + + if (error != 0) + return (error); + } + vdev_set_deflate_ratio(vd); /* @@ -3531,8 +3693,11 @@ vdev_sync_done(vdev_t *vd, uint64_t txg) != NULL) metaslab_sync_done(msp, txg); - if (reassess) + if (reassess) { metaslab_sync_reassess(vd->vdev_mg); + if (vd->vdev_log_mg != NULL) + metaslab_sync_reassess(vd->vdev_log_mg); + } } void @@ -3856,6 +4021,7 @@ top: /* * Prevent any future allocations. */ + ASSERT3P(tvd->vdev_log_mg, ==, NULL); metaslab_group_passivate(mg); (void) spa_vdev_state_exit(spa, vd, 0); @@ -4256,6 +4422,12 @@ vdev_get_stats_ex(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx) */ if (vd->vdev_aux == NULL && vd == vd->vdev_top && vdev_is_concrete(vd)) { + /* + * The vdev fragmentation rating doesn't take into + * account the embedded slog metaslab (vdev_log_mg). + * Since it's only one metaslab, it would have a tiny + * impact on the overall fragmentation. + */ vs->vs_fragmentation = (vd->vdev_mg != NULL) ? vd->vdev_mg->mg_fragmentation : 0; } @@ -5234,6 +5406,9 @@ ZFS_MODULE_PARAM(zfs_vdev, vdev_, validate_skip, INT, ZMOD_RW, ZFS_MODULE_PARAM(zfs, zfs_, nocacheflush, INT, ZMOD_RW, "Disable cache flushes"); +ZFS_MODULE_PARAM(zfs, zfs_, embedded_slog_min_ms, INT, ZMOD_RW, + "Minimum number of metaslabs required to dedicate one for log blocks"); + ZFS_MODULE_PARAM_CALL(zfs_vdev, zfs_vdev_, min_auto_ashift, param_set_min_auto_ashift, param_get_ulong, ZMOD_RW, "Minimum ashift used when creating new top-level vdevs"); |