Re-embed multilist_t storage

This commit partially reverts changes to multilists in PR 7968
(multi-threaded spa-sync()) and adds some cache line alignments to
separate read-only multilists and heavily modified refcount's to different
cache lines.

Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored-by: iXsystems, Inc.
Closes #12158

8 files changed, 88 insertions, 94 deletions

diff --git a/module/zfs/arc.c b/module/zfs/arc.c
index 5526cae37..737904f34 100644
--- a/module/zfs/arc.c
+++ b/module/zfs/arc.c
@@ -2327,7 +2327,7 @@ add_reference(arc_buf_hdr_t *hdr, void *tag)
 	    (state != arc_anon)) {
 		/* We don't use the L2-only state list. */
 		if (state != arc_l2c_only) {
-			multilist_remove(state->arcs_list[arc_buf_type(hdr)],
+			multilist_remove(&state->arcs_list[arc_buf_type(hdr)],
 			    hdr);
 			arc_evictable_space_decrement(hdr, state);
 		}
@@ -2361,7 +2361,7 @@ remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
 	 */
 	if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
 	    (state != arc_anon)) {
-		multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
+		multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
 		arc_evictable_space_increment(hdr, state);
 	}
@@ -2464,7 +2464,7 @@ arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
 	if (refcnt == 0) {
 		if (old_state != arc_anon && old_state != arc_l2c_only) {
 			ASSERT(HDR_HAS_L1HDR(hdr));
-			multilist_remove(old_state->arcs_list[buftype], hdr);
+			multilist_remove(&old_state->arcs_list[buftype], hdr);
 
 			if (GHOST_STATE(old_state)) {
 				ASSERT0(bufcnt);
@@ -2481,7 +2481,7 @@ arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
 			 * beforehand.
 			 */
 			ASSERT(HDR_HAS_L1HDR(hdr));
-			multilist_insert(new_state->arcs_list[buftype], hdr);
+			multilist_insert(&new_state->arcs_list[buftype], hdr);
 
 			if (GHOST_STATE(new_state)) {
 				ASSERT0(bufcnt);
@@ -2633,8 +2633,8 @@ arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
 	 * L2 headers should never be on the L2 state list since they don't
 	 * have L1 headers allocated.
 	 */
-	ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
-	    multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
+	ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
+	    multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
 }
 
 void
@@ -4200,7 +4200,7 @@ arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
     arc_buf_contents_t type)
 {
 	uint64_t total_evicted = 0;
-	multilist_t *ml = state->arcs_list[type];
+	multilist_t *ml = &state->arcs_list[type];
 	int num_sublists;
 	arc_buf_hdr_t **markers;
 
@@ -4534,8 +4534,8 @@ arc_evict_meta(uint64_t meta_used)
 static arc_buf_contents_t
 arc_evict_type(arc_state_t *state)
 {
-	multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
-	multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
+	multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
+	multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
 	int data_idx = multilist_get_random_index(data_ml);
 	int meta_idx = multilist_get_random_index(meta_ml);
 	multilist_sublist_t *data_mls;
@@ -7455,44 +7455,44 @@ arc_state_init(void)
 	arc_mfu_ghost = &ARC_mfu_ghost;
 	arc_l2c_only = &ARC_l2c_only;
 
-	arc_mru->arcs_list[ARC_BUFC_METADATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mru->arcs_list[ARC_BUFC_DATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mfu->arcs_list[ARC_BUFC_METADATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mfu->arcs_list[ARC_BUFC_DATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
-	arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
-	    multilist_create(sizeof (arc_buf_hdr_t),
+	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
+	    sizeof (arc_buf_hdr_t),
 	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
 	    arc_state_multilist_index_func);
 
@@ -7558,16 +7558,16 @@ arc_state_fini(void)
 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
 	zfs_refcount_destroy(&arc_l2c_only->arcs_size);
 
-	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
-	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
-	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
-	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
-	multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
-	multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
-	multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
-	multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
-	multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
-	multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
 
 	aggsum_fini(&arc_meta_used);
 	aggsum_fini(&arc_size);
@@ -8624,16 +8624,16 @@ l2arc_sublist_lock(int list_num)
 
 	switch (list_num) {
 	case 0:
-		ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
+		ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
 		break;
 	case 1:
-		ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
+		ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
 		break;
 	case 2:
-		ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
+		ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
 		break;
 	case 3:
-		ml = arc_mru->arcs_list[ARC_BUFC_DATA];
+		ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
 		break;
 	default:
 		return (NULL);
diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index 764383b2d..9d741545f 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -203,8 +203,8 @@ static boolean_t dbuf_evict_thread_exit;
  * by those caches' matching enum values (from dbuf_cached_state_t).
  */
 typedef struct dbuf_cache {
-	multilist_t *cache;
-	zfs_refcount_t size;
+	multilist_t cache;
+	zfs_refcount_t size ____cacheline_aligned;
 } dbuf_cache_t;
 dbuf_cache_t dbuf_caches[DB_CACHE_MAX];
 
@@ -667,9 +667,9 @@ dbuf_cache_above_lowater(void)
 static void
 dbuf_evict_one(void)
 {
-	int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
+	int idx = multilist_get_random_index(&dbuf_caches[DB_DBUF_CACHE].cache);
 	multilist_sublist_t *mls = multilist_sublist_lock(
-	    dbuf_caches[DB_DBUF_CACHE].cache, idx);
+	    &dbuf_caches[DB_DBUF_CACHE].cache, idx);
 
 	ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
 
@@ -833,8 +833,8 @@ retry:
 	dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
 
 	for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
-		dbuf_caches[dcs].cache =
-		    multilist_create(sizeof (dmu_buf_impl_t),
+		multilist_create(&dbuf_caches[dcs].cache,
+		    sizeof (dmu_buf_impl_t),
 		    offsetof(dmu_buf_impl_t, db_cache_link),
 		    dbuf_cache_multilist_index_func);
 		zfs_refcount_create(&dbuf_caches[dcs].size);
@@ -901,7 +901,7 @@ dbuf_fini(void)
 
 	for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
 		zfs_refcount_destroy(&dbuf_caches[dcs].size);
-		multilist_destroy(dbuf_caches[dcs].cache);
+		multilist_destroy(&dbuf_caches[dcs].cache);
 	}
 
 	if (dbuf_ksp != NULL) {
@@ -2755,7 +2755,7 @@ dbuf_destroy(dmu_buf_impl_t *db)
 		ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
 		    db->db_caching_status == DB_DBUF_METADATA_CACHE);
 
-		multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
+		multilist_remove(&dbuf_caches[db->db_caching_status].cache, db);
 		(void) zfs_refcount_remove_many(
 		    &dbuf_caches[db->db_caching_status].size,
 		    db->db.db_size, db);
@@ -3465,7 +3465,7 @@ dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
 		ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
 		    db->db_caching_status == DB_DBUF_METADATA_CACHE);
 
-		multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
+		multilist_remove(&dbuf_caches[db->db_caching_status].cache, db);
 		(void) zfs_refcount_remove_many(
 		    &dbuf_caches[db->db_caching_status].size,
 		    db->db.db_size, db);
@@ -3707,7 +3707,7 @@ dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
 				    DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
 				db->db_caching_status = dcs;
 
-				multilist_insert(dbuf_caches[dcs].cache, db);
+				multilist_insert(&dbuf_caches[dcs].cache, db);
 				size = zfs_refcount_add_many(
 				    &dbuf_caches[dcs].size,
 				    db->db.db_size, db);
diff --git a/module/zfs/dmu_objset.c b/module/zfs/dmu_objset.c
index bfb4adf26..8c244dc4c 100644
--- a/module/zfs/dmu_objset.c
+++ b/module/zfs/dmu_objset.c
@@ -601,7 +601,7 @@ dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 	os->os_zil = zil_alloc(os, &os->os_zil_header);
 
 	for (i = 0; i < TXG_SIZE; i++) {
-		os->os_dirty_dnodes[i] = multilist_create(sizeof (dnode_t),
+		multilist_create(&os->os_dirty_dnodes[i], sizeof (dnode_t),
 		    offsetof(dnode_t, dn_dirty_link[i]),
 		    dnode_multilist_index_func);
 	}
@@ -995,9 +995,8 @@ dmu_objset_evict_done(objset_t *os)
 	mutex_destroy(&os->os_obj_lock);
 	mutex_destroy(&os->os_user_ptr_lock);
 	mutex_destroy(&os->os_upgrade_lock);
-	for (int i = 0; i < TXG_SIZE; i++) {
-		multilist_destroy(os->os_dirty_dnodes[i]);
-	}
+	for (int i = 0; i < TXG_SIZE; i++)
+		multilist_destroy(&os->os_dirty_dnodes[i]);
 	spa_evicting_os_deregister(os->os_spa, os);
 	kmem_free(os, sizeof (objset_t));
 }
@@ -1520,7 +1519,7 @@ dmu_objset_sync_dnodes(multilist_sublist_t *list, dmu_tx_t *tx)
 		 * of why this dnode hold is always needed (even when not
 		 * doing user accounting).
 		 */
-		multilist_t *newlist = dn->dn_objset->os_synced_dnodes;
+		multilist_t *newlist = &dn->dn_objset->os_synced_dnodes;
 		(void) dnode_add_ref(dn, newlist);
 		multilist_insert(newlist, dn);
 
@@ -1689,17 +1688,16 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 	 * dn_dirty_link[] of this txg.  But it may already
 	 * exist because we call dsl_dataset_sync() twice per txg.
 	 */
-	if (os->os_synced_dnodes == NULL) {
-		os->os_synced_dnodes =
-		    multilist_create(sizeof (dnode_t),
+	if (os->os_synced_dnodes.ml_sublists == NULL) {
+		multilist_create(&os->os_synced_dnodes, sizeof (dnode_t),
 		    offsetof(dnode_t, dn_dirty_link[txgoff]),
 		    dnode_multilist_index_func);
 	} else {
-		ASSERT3U(os->os_synced_dnodes->ml_offset, ==,
+		ASSERT3U(os->os_synced_dnodes.ml_offset, ==,
 		    offsetof(dnode_t, dn_dirty_link[txgoff]));
 	}
 
-	ml = os->os_dirty_dnodes[txgoff];
+	ml = &os->os_dirty_dnodes[txgoff];
 	num_sublists = multilist_get_num_sublists(ml);
 	for (int i = 0; i < num_sublists; i++) {
 		if (multilist_sublist_is_empty_idx(ml, i))
@@ -1738,7 +1736,7 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 boolean_t
 dmu_objset_is_dirty(objset_t *os, uint64_t txg)
 {
-	return (!multilist_is_empty(os->os_dirty_dnodes[txg & TXG_MASK]));
+	return (!multilist_is_empty(&os->os_dirty_dnodes[txg & TXG_MASK]));
 }
 
 static file_info_cb_t *file_cbs[DMU_OST_NUMTYPES];
@@ -1949,7 +1947,7 @@ userquota_updates_task(void *arg)
 	userquota_cache_t cache = { { 0 } };
 
 	multilist_sublist_t *list =
-	    multilist_sublist_lock(os->os_synced_dnodes, uua->uua_sublist_idx);
+	    multilist_sublist_lock(&os->os_synced_dnodes, uua->uua_sublist_idx);
 
 	ASSERT(multilist_sublist_head(list) == NULL ||
 	    dmu_objset_userused_enabled(os));
@@ -2006,7 +2004,7 @@ userquota_updates_task(void *arg)
 		mutex_exit(&dn->dn_mtx);
 
 		multilist_sublist_remove(list, dn);
-		dnode_rele(dn, os->os_synced_dnodes);
+		dnode_rele(dn, &os->os_synced_dnodes);
 	}
 	do_userquota_cacheflush(os, &cache, tx);
 	multilist_sublist_unlock(list);
@@ -2032,12 +2030,12 @@ dnode_rele_task(void *arg)
 	objset_t *os = uua->uua_os;
 
 	multilist_sublist_t *list =
-	    multilist_sublist_lock(os->os_synced_dnodes, uua->uua_sublist_idx);
+	    multilist_sublist_lock(&os->os_synced_dnodes, uua->uua_sublist_idx);
 
 	dnode_t *dn;
 	while ((dn = multilist_sublist_head(list)) != NULL) {
 		multilist_sublist_remove(list, dn);
-		dnode_rele(dn, os->os_synced_dnodes);
+		dnode_rele(dn, &os->os_synced_dnodes);
 	}
 	multilist_sublist_unlock(list);
 	kmem_free(uua, sizeof (*uua));
@@ -2093,7 +2091,7 @@ dmu_objset_sync_done(objset_t *os, dmu_tx_t *tx)
 {
 	boolean_t need_userquota = dmu_objset_do_userquota_updates_prep(os, tx);
 
-	int num_sublists = multilist_get_num_sublists(os->os_synced_dnodes);
+	int num_sublists = multilist_get_num_sublists(&os->os_synced_dnodes);
 	for (int i = 0; i < num_sublists; i++) {
 		userquota_updates_arg_t *uua =
 		    kmem_alloc(sizeof (*uua), KM_SLEEP);
diff --git a/module/zfs/dnode.c b/module/zfs/dnode.c
index 0fc788e28..8434e72aa 100644
--- a/module/zfs/dnode.c
+++ b/module/zfs/dnode.c
@@ -1671,7 +1671,7 @@ dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
 	 */
 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
 
-	multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
+	multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK];
 	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
 
 	/*
diff --git a/module/zfs/dsl_dataset.c b/module/zfs/dsl_dataset.c
index 6da5faf01..9b9bb4228 100644
--- a/module/zfs/dsl_dataset.c
+++ b/module/zfs/dsl_dataset.c
@@ -2267,8 +2267,7 @@ dsl_dataset_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
 
 	dsl_bookmark_sync_done(ds, tx);
 
-	multilist_destroy(os->os_synced_dnodes);
-	os->os_synced_dnodes = NULL;
+	multilist_destroy(&os->os_synced_dnodes);
 
 	if (os->os_encrypted)
 		os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_FALSE;
diff --git a/module/zfs/dsl_pool.c b/module/zfs/dsl_pool.c
index c770eafa7..e66c136a9 100644
--- a/module/zfs/dsl_pool.c
+++ b/module/zfs/dsl_pool.c
@@ -568,8 +568,7 @@ dsl_pool_sync_mos(dsl_pool_t *dp, dmu_tx_t *tx)
 	VERIFY0(zio_wait(zio));
 	dmu_objset_sync_done(dp->dp_meta_objset, tx);
 	taskq_wait(dp->dp_sync_taskq);
-	multilist_destroy(dp->dp_meta_objset->os_synced_dnodes);
-	dp->dp_meta_objset->os_synced_dnodes = NULL;
+	multilist_destroy(&dp->dp_meta_objset->os_synced_dnodes);
 
 	dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
 	spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
diff --git a/module/zfs/metaslab.c b/module/zfs/metaslab.c
index 3b2b79b2f..e588765b3 100644
--- a/module/zfs/metaslab.c
+++ b/module/zfs/metaslab.c
@@ -416,7 +416,7 @@ metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
 	mc->mc_spa = spa;
 	mc->mc_ops = ops;
 	mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
-	mc->mc_metaslab_txg_list = multilist_create(sizeof (metaslab_t),
+	multilist_create(&mc->mc_metaslab_txg_list, sizeof (metaslab_t),
 	    offsetof(metaslab_t, ms_class_txg_node), metaslab_idx_func);
 	for (int i = 0; i < spa->spa_alloc_count; i++) {
 		metaslab_class_allocator_t *mca = &mc->mc_allocator[i];
@@ -443,7 +443,7 @@ metaslab_class_destroy(metaslab_class_t *mc)
 		zfs_refcount_destroy(&mca->mca_alloc_slots);
 	}
 	mutex_destroy(&mc->mc_lock);
-	multilist_destroy(mc->mc_metaslab_txg_list);
+	multilist_destroy(&mc->mc_metaslab_txg_list);
 	kmem_free(mc, offsetof(metaslab_class_t,
 	    mc_allocator[spa->spa_alloc_count]));
 }
@@ -639,7 +639,7 @@ metaslab_class_expandable_space(metaslab_class_t *mc)
 void
 metaslab_class_evict_old(metaslab_class_t *mc, uint64_t txg)
 {
-	multilist_t *ml = mc->mc_metaslab_txg_list;
+	multilist_t *ml = &mc->mc_metaslab_txg_list;
 	for (int i = 0; i < multilist_get_num_sublists(ml); i++) {
 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
 		metaslab_t *msp = multilist_sublist_head(mls);
@@ -1139,7 +1139,7 @@ metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
 
 	metaslab_class_t *mc = msp->ms_group->mg_class;
 	multilist_sublist_t *mls =
-	    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
+	    multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
 	if (multilist_link_active(&msp->ms_class_txg_node))
 		multilist_sublist_remove(mls, msp);
 	multilist_sublist_unlock(mls);
@@ -2175,20 +2175,20 @@ metaslab_potentially_evict(metaslab_class_t *mc)
 	uint64_t size =	spl_kmem_cache_entry_size(zfs_btree_leaf_cache);
 	int tries = 0;
 	for (; allmem * zfs_metaslab_mem_limit / 100 < inuse * size &&
-	    tries < multilist_get_num_sublists(mc->mc_metaslab_txg_list) * 2;
+	    tries < multilist_get_num_sublists(&mc->mc_metaslab_txg_list) * 2;
 	    tries++) {
 		unsigned int idx = multilist_get_random_index(
-		    mc->mc_metaslab_txg_list);
+		    &mc->mc_metaslab_txg_list);
 		multilist_sublist_t *mls =
-		    multilist_sublist_lock(mc->mc_metaslab_txg_list, idx);
+		    multilist_sublist_lock(&mc->mc_metaslab_txg_list, idx);
 		metaslab_t *msp = multilist_sublist_head(mls);
 		multilist_sublist_unlock(mls);
 		while (msp != NULL && allmem * zfs_metaslab_mem_limit / 100 <
 		    inuse * size) {
 			VERIFY3P(mls, ==, multilist_sublist_lock(
-			    mc->mc_metaslab_txg_list, idx));
+			    &mc->mc_metaslab_txg_list, idx));
 			ASSERT3U(idx, ==,
-			    metaslab_idx_func(mc->mc_metaslab_txg_list, msp));
+			    metaslab_idx_func(&mc->mc_metaslab_txg_list, msp));
 
 			if (!multilist_link_active(&msp->ms_class_txg_node)) {
 				multilist_sublist_unlock(mls);
@@ -2535,7 +2535,7 @@ metaslab_unload(metaslab_t *msp)
 	if (msp->ms_group != NULL) {
 		metaslab_class_t *mc = msp->ms_group->mg_class;
 		multilist_sublist_t *mls =
-		    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
+		    multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
 		if (multilist_link_active(&msp->ms_class_txg_node))
 			multilist_sublist_remove(mls, msp);
 		multilist_sublist_unlock(mls);
@@ -2600,7 +2600,7 @@ metaslab_set_selected_txg(metaslab_t *msp, uint64_t txg)
 	ASSERT(MUTEX_HELD(&msp->ms_lock));
 	metaslab_class_t *mc = msp->ms_group->mg_class;
 	multilist_sublist_t *mls =
-	    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
+	    multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
 	if (multilist_link_active(&msp->ms_class_txg_node))
 		multilist_sublist_remove(mls, msp);
 	msp->ms_selected_txg = txg;
@@ -5682,7 +5682,7 @@ metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
 	if (spa_writeable(spa)) {	/* don't dirty if we're zdb(8) */
 		metaslab_class_t *mc = msp->ms_group->mg_class;
 		multilist_sublist_t *mls =
-		    multilist_sublist_lock_obj(mc->mc_metaslab_txg_list, msp);
+		    multilist_sublist_lock_obj(&mc->mc_metaslab_txg_list, msp);
 		if (!multilist_link_active(&msp->ms_class_txg_node)) {
 			msp->ms_selected_txg = txg;
 			multilist_sublist_insert_head(mls, msp);
diff --git a/module/zfs/multilist.c b/module/zfs/multilist.c
index 36c0d33bf..eeac73bd7 100644
--- a/module/zfs/multilist.c
+++ b/module/zfs/multilist.c
@@ -68,8 +68,8 @@ multilist_d2l(multilist_t *ml, void *obj)
  *     requirement, but a general rule of thumb in order to garner the
  *     best multi-threaded performance out of the data structure.
  */
-static multilist_t *
-multilist_create_impl(size_t size, size_t offset,
+static void
+multilist_create_impl(multilist_t *ml, size_t size, size_t offset,
     unsigned int num, multilist_sublist_index_func_t *index_func)
 {
 	ASSERT3U(size, >, 0);
@@ -77,7 +77,6 @@ multilist_create_impl(size_t size, size_t offset,
 	ASSERT3U(num, >, 0);
 	ASSERT3P(index_func, !=, NULL);
 
-	multilist_t *ml = kmem_alloc(sizeof (*ml), KM_SLEEP);
 	ml->ml_offset = offset;
 	ml->ml_num_sublists = num;
 	ml->ml_index_func = index_func;
@@ -92,7 +91,6 @@ multilist_create_impl(size_t size, size_t offset,
 		mutex_init(&mls->mls_lock, NULL, MUTEX_NOLOCKDEP, NULL);
 		list_create(&mls->mls_list, size, offset);
 	}
-	return (ml);
 }
 
 /*
@@ -103,8 +101,8 @@ multilist_create_impl(size_t size, size_t offset,
  * reserve the RAM necessary to create the extra slots for additional CPUs up
  * front, and dynamically adding them is a complex task.
  */
-multilist_t *
-multilist_create(size_t size, size_t offset,
+void
+multilist_create(multilist_t *ml, size_t size, size_t offset,
     multilist_sublist_index_func_t *index_func)
 {
 	int num_sublists;
@@ -115,7 +113,7 @@ multilist_create(size_t size, size_t offset,
 		num_sublists = MAX(boot_ncpus, 4);
 	}
 
-	return (multilist_create_impl(size, offset, num_sublists, index_func));
+	multilist_create_impl(ml, size, offset, num_sublists, index_func);
 }
 
 /*
@@ -141,7 +139,7 @@ multilist_destroy(multilist_t *ml)
 
 	ml->ml_num_sublists = 0;
 	ml->ml_offset = 0;
-	kmem_free(ml, sizeof (multilist_t));
+	ml->ml_sublists = NULL;
 }
 
 /*