diff --git a/liberty.c b/liberty.c
index e004464..fd9ccc7 100644
--- a/liberty.c
+++ b/liberty.c
@@ -1073,6 +1073,7 @@ struct async
 	// "cancelled" may not be accesed or modified by the worker thread
 
 	pthread_t worker;                   ///< Worker thread ID
+	bool started;                       ///< Worker thread ID is valid
 	bool cancelled;                     ///< Task has been cancelled
 
 	async_fn execute;                   ///< Worker main function
@@ -1087,6 +1088,8 @@ async_init (struct async *self, struct async_manager *manager)
 	self->manager = manager;
 }
 
+static bool async_run (struct async *self);
+
 // - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 
 struct async_manager
@@ -1095,6 +1098,12 @@ struct async_manager
 	struct async *running;              ///< Queue of running jobs
 	struct async *finished;             ///< Queue of completed/cancelled jobs
 
+	// It's upon the user to call async_manager_dispatch() to retry the delayed.
+	// It's somewhat questionable if this feature is of any use. Possibly if we
+	// provide a means of actively limiting the amount of running async jobs.
+
+	struct async *delayed;              ///< Resource exhaustion queue
+
 	// We need the pipe in order to abort polling (instead of using EINTR)
 
 	pthread_cond_t finished_cond;       ///< Signals that a task has finished
@@ -1114,6 +1123,18 @@ async_manager_init (struct async_manager *self)
 	set_cloexec (self->finished_pipe[1]);
 }
 
+static bool
+async_manager_retry (struct async_manager *self, struct async *async)
+{
+	if (async->cancelled)
+	{
+		if (async->destroy)
+			async->destroy (async);
+		return true;
+	}
+	return async_run (async);
+}
+
 static struct async *
 async_manager_dispatch_fetch (struct async_manager *self)
 {
@@ -1145,6 +1166,13 @@ async_manager_dispatch (struct async_manager *self)
 		if (iter->destroy)
 			iter->destroy (iter);
 	}
+
+	LIST_FOR_EACH (struct async, iter, self->delayed)
+	{
+		LIST_UNLINK (self->delayed, iter);
+		if (!async_manager_retry (self, iter))
+			break;
+	}
 }
 
 static void
@@ -1164,6 +1192,8 @@ async_manager_cancel_all (struct async_manager *self)
 	// Mark everything cancelled so that it's not actually dispatched
 	LIST_FOR_EACH (struct async, iter, self->finished)
 		iter->cancelled = true;
+	LIST_FOR_EACH (struct async, iter, self->delayed)
+		iter->cancelled = true;
 
 	hard_assert (!pthread_mutex_unlock (&self->lock));
 	async_manager_dispatch (self);
@@ -1187,7 +1217,8 @@ async_manager_free (struct async_manager *self)
 static void
 async_cancel (struct async *self)
 {
-	soft_assert (!pthread_cancel (self->worker));
+	if (self->started)
+		soft_assert (!pthread_cancel (self->worker));
 	self->cancelled = true;
 }
 
@@ -1219,7 +1250,7 @@ async_routine (void *user_data)
 	return NULL;
 }
 
-static void
+static bool
 async_run (struct async *self)
 {
 	hard_assert (!pthread_mutex_lock (&self->manager->lock));
@@ -1231,11 +1262,24 @@ async_run (struct async *self)
 	hard_assert (!sigfillset (&all_blocked));
 	hard_assert (!pthread_sigmask (SIG_SETMASK, &all_blocked, &old_blocked));
 
-	hard_assert (!pthread_create (&self->worker, NULL,
-		async_routine, self));
+	int error = pthread_create (&self->worker, NULL, async_routine, self);
 
 	// Now that we've created the thread, resume signal processing as usual
 	hard_assert (!pthread_sigmask (SIG_SETMASK, &old_blocked, NULL));
+
+	if (error)
+	{
+		hard_assert (error == EAGAIN);
+
+		hard_assert (!pthread_mutex_lock (&self->manager->lock));
+		LIST_UNLINK (self->manager->running, self);
+		hard_assert (!pthread_mutex_unlock (&self->manager->lock));
+
+		// FIXME: we probably want to have some kind of a limit on the queue
+		LIST_PREPEND (self->manager->delayed, self);
+		return false;
+	}
+	return (self->started = true);
 }
 
 #endif // LIBERTY_WANT_ASYNC
@@ -2138,7 +2182,17 @@ poller_common_free (struct poller_common *self)
 static int
 poller_common_get_timeout (struct poller_common *self)
 {
-	return self->idle ? 0 : poller_timers_get_poll_timeout (&self->timers);
+	if (self->idle)
+		return 0;
+
+	int timeout = poller_timers_get_poll_timeout (&self->timers);
+#ifdef LIBERTY_WANT_ASYNC
+	// This is completely arbitrary, in general we have no idea when to retry,
+	// however one second doesn't sound like a particularly bad number
+	if (self->async.delayed)
+		timeout = MIN (timeout, 1000);
+#endif // LIBERTY_WANT_ASYNC
+	return timeout;
 }
 
 static void