Avoid extra taskq_dispatch() calls by DMU

DMU sync code calls taskq_dispatch() for each sublist of os_dirty_dnodes
and os_synced_dnodes.  Since the number of sublists by default is equal
to number of CPUs, it will dispatch equal, potentially large, number of
tasks, waking up many CPUs to handle them, even if only one or few of
sublists actually have any work to do.

This change adds check for empty sublists to avoid this.

Reviewed by: Sean Eric Fagan <[email protected]>
Reviewed by: Matt Ahrens <[email protected]>
Reviewed by: Brian Behlendorf <[email protected]>
Signed-off-by:  Alexander Motin <[email protected]>
Closes #8909 

3 files changed, 38 insertions, 5 deletions

diff --git a/include/sys/multilist.h b/include/sys/multilist.h
index 439540685..0c7b4075d 100644
--- a/include/sys/multilist.h
+++ b/include/sys/multilist.h
@@ -89,6 +89,8 @@ void multilist_sublist_insert_head(multilist_sublist_t *, void *);
 void multilist_sublist_insert_tail(multilist_sublist_t *, void *);
 void multilist_sublist_move_forward(multilist_sublist_t *mls, void *obj);
 void multilist_sublist_remove(multilist_sublist_t *, void *);
+int  multilist_sublist_is_empty(multilist_sublist_t *);
+int  multilist_sublist_is_empty_idx(multilist_t *, unsigned int);
 
 void *multilist_sublist_head(multilist_sublist_t *);
 void *multilist_sublist_tail(multilist_sublist_t *);
diff --git a/module/zfs/dmu_objset.c b/module/zfs/dmu_objset.c
index 4091ac355..8b55469a5 100644
--- a/module/zfs/dmu_objset.c
+++ b/module/zfs/dmu_objset.c
@@ -1598,6 +1598,8 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 	zio_t *zio;
 	list_t *list;
 	dbuf_dirty_record_t *dr;
+	int num_sublists;
+	multilist_t *ml;
 	blkptr_t *blkptr_copy = kmem_alloc(sizeof (*os->os_rootbp), KM_SLEEP);
 	*blkptr_copy = *os->os_rootbp;
 
@@ -1686,10 +1688,13 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 		}
 	}
 
-	for (int i = 0;
-	    i < multilist_get_num_sublists(os->os_dirty_dnodes[txgoff]); i++) {
+	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))
+			continue;
 		sync_dnodes_arg_t *sda = kmem_alloc(sizeof (*sda), KM_SLEEP);
-		sda->sda_list = os->os_dirty_dnodes[txgoff];
+		sda->sda_list = ml;
 		sda->sda_sublist_idx = i;
 		sda->sda_tx = tx;
 		(void) taskq_dispatch(dmu_objset_pool(os)->dp_sync_taskq,
@@ -1992,6 +1997,8 @@ userquota_updates_task(void *arg)
 void
 dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
 {
+	int num_sublists;
+
 	if (!dmu_objset_userused_enabled(os))
 		return;
 
@@ -2024,8 +2031,10 @@ dmu_objset_do_userquota_updates(objset_t *os, dmu_tx_t *tx)
 		    DMU_OT_USERGROUP_USED, DMU_OT_NONE, 0, tx));
 	}
 
-	for (int i = 0;
-	    i < multilist_get_num_sublists(os->os_synced_dnodes); i++) {
+	num_sublists = multilist_get_num_sublists(os->os_synced_dnodes);
+	for (int i = 0; i < num_sublists; i++) {
+		if (multilist_sublist_is_empty_idx(os->os_synced_dnodes, i))
+			continue;
 		userquota_updates_arg_t *uua =
 		    kmem_alloc(sizeof (*uua), KM_SLEEP);
 		uua->uua_os = os;
diff --git a/module/zfs/multilist.c b/module/zfs/multilist.c
index 2a594c56c..b74ee0f06 100644
--- a/module/zfs/multilist.c
+++ b/module/zfs/multilist.c
@@ -363,6 +363,28 @@ multilist_sublist_remove(multilist_sublist_t *mls, void *obj)
 	list_remove(&mls->mls_list, obj);
 }
 
+int
+multilist_sublist_is_empty(multilist_sublist_t *mls)
+{
+	ASSERT(MUTEX_HELD(&mls->mls_lock));
+	return (list_is_empty(&mls->mls_list));
+}
+
+int
+multilist_sublist_is_empty_idx(multilist_t *ml, unsigned int sublist_idx)
+{
+	multilist_sublist_t *mls;
+	int empty;
+
+	ASSERT3U(sublist_idx, <, ml->ml_num_sublists);
+	mls = &ml->ml_sublists[sublist_idx];
+	ASSERT(!MUTEX_HELD(&mls->mls_lock));
+	mutex_enter(&mls->mls_lock);
+	empty = list_is_empty(&mls->mls_list);
+	mutex_exit(&mls->mls_lock);
+	return (empty);
+}
+
 void *
 multilist_sublist_head(multilist_sublist_t *mls)
 {