diff --git a/include/linux/closure.h b/include/linux/closure.h
index 99155df162d0..59b8c06b11ff 100644
--- a/include/linux/closure.h
+++ b/include/linux/closure.h
@@ -284,6 +284,21 @@ static inline void closure_get(struct closure *cl)
 #endif
 }
 
+/**
+ * closure_get_not_zero
+ */
+static inline bool closure_get_not_zero(struct closure *cl)
+{
+	unsigned old = atomic_read(&cl->remaining);
+	do {
+		if (!(old & CLOSURE_REMAINING_MASK))
+			return false;
+
+	} while (!atomic_try_cmpxchg_acquire(&cl->remaining, &old, old + 1));
+
+	return true;
+}
+
 /**
  * closure_init - Initialize a closure, setting the refcount to 1
  * @cl:		closure to initialize
@@ -310,6 +325,12 @@ static inline void closure_init_stack(struct closure *cl)
 	atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
 }
 
+static inline void closure_init_stack_release(struct closure *cl)
+{
+	memset(cl, 0, sizeof(struct closure));
+	atomic_set_release(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
+}
+
 /**
  * closure_wake_up - wake up all closures on a wait list,
  *		     with memory barrier
@@ -355,6 +376,8 @@ do {									\
  */
 #define closure_return(_cl)	continue_at((_cl), NULL, NULL)
 
+void closure_return_sync(struct closure *cl);
+
 /**
  * continue_at_nobarrier - jump to another function without barrier
  *
diff --git a/lib/closure.c b/lib/closure.c
index 2e1ee9fdec08..c971216d9d77 100644
--- a/lib/closure.c
+++ b/lib/closure.c
@@ -13,7 +13,7 @@
 #include <linux/seq_file.h>
 #include <linux/sched/debug.h>
 
-static inline void closure_put_after_sub(struct closure *cl, int flags)
+static inline void closure_put_after_sub_checks(int flags)
 {
 	int r = flags & CLOSURE_REMAINING_MASK;
 
@@ -22,12 +22,17 @@ static inline void closure_put_after_sub(struct closure *cl, int flags)
 		 flags & CLOSURE_GUARD_MASK, (unsigned) __fls(r)))
 		r &= ~CLOSURE_GUARD_MASK;
 
-	if (!r) {
-		smp_acquire__after_ctrl_dep();
+	WARN(!r && (flags & ~CLOSURE_DESTRUCTOR),
+	     "closure ref hit 0 with incorrect flags set: %x (%u)",
+	     flags & ~CLOSURE_DESTRUCTOR, (unsigned) __fls(flags));
+}
 
-		WARN(flags & ~CLOSURE_DESTRUCTOR,
-		     "closure ref hit 0 with incorrect flags set: %x (%u)",
-		     flags & ~CLOSURE_DESTRUCTOR, (unsigned) __fls(flags));
+static inline void closure_put_after_sub(struct closure *cl, int flags)
+{
+	closure_put_after_sub_checks(flags);
+
+	if (!(flags & CLOSURE_REMAINING_MASK)) {
+		smp_acquire__after_ctrl_dep();
 
 		cl->closure_get_happened = false;
 
@@ -145,6 +150,41 @@ void __sched __closure_sync(struct closure *cl)
 }
 EXPORT_SYMBOL(__closure_sync);
 
+/*
+ * closure_return_sync - finish running a closure, synchronously (i.e. waiting
+ * for outstanding get()s to finish) and returning once closure refcount is 0.
+ *
+ * Unlike closure_sync() this doesn't reinit the ref to 1; subsequent
+ * closure_get_not_zero() calls waill fail.
+ */
+void __sched closure_return_sync(struct closure *cl)
+{
+	struct closure_syncer s = { .task = current };
+
+	cl->s = &s;
+	set_closure_fn(cl, closure_sync_fn, NULL);
+
+	unsigned flags = atomic_sub_return_release(1 + CLOSURE_RUNNING - CLOSURE_DESTRUCTOR,
+						   &cl->remaining);
+
+	closure_put_after_sub_checks(flags);
+
+	if (unlikely(flags & CLOSURE_REMAINING_MASK)) {
+		while (1) {
+			set_current_state(TASK_UNINTERRUPTIBLE);
+			if (s.done)
+				break;
+			schedule();
+		}
+
+		__set_current_state(TASK_RUNNING);
+	}
+
+	if (cl->parent)
+		closure_put(cl->parent);
+}
+EXPORT_SYMBOL(closure_return_sync);
+
 int __sched __closure_sync_timeout(struct closure *cl, unsigned long timeout)
 {
 	struct closure_syncer s = { .task = current };