ipc: do not hold a spinlock when calling schedule_ipc_worker()

Calling schedule_ipc_worker() with spinlock held is not stricly
required. schedule_ipc_worker() calls k_work_schedule_for_queue()
and the Zephyr workqueue has an internal lock protecting
the workqueue structures. Keeping spinlock held during the call
does lead to complex scenarios to debug and verify as schedule_ipc_worker()
itself is dispatched from a work queue.

Simplify the flow and release spinlock before rescheduling the worker.
In ipc_work_handler(), this leaves a small window where the message is
just sent after we release the spinlock, and before we call
schedule_ipc_worker(). This is however harmless, as in worst case the
IPC worker is woken up one extra time.

Signed-off-by: Kai Vehmanen <kai.vehmanen@linux.intel.com>

Author: kv2019i

src/ipc/ipc-common.c

@@ -254,9 +254,9 @@ void ipc_msg_send(struct ipc_msg *msg, void *data, bool high_priority)
 			list_item_append(&msg->list, &ipc->msg_list);
 	}
 
-	schedule_ipc_worker();
-
 	k_spin_unlock(&ipc->lock, key);
+
+	schedule_ipc_worker();
 }
 EXPORT_SYMBOL(ipc_msg_send);
 
@@ -265,15 +265,25 @@ static void ipc_work_handler(struct k_work *work)
 {
 	struct ipc *ipc = ipc_get();
 	k_spinlock_key_t key;
+	bool more_to_send = false;
 
 	ipc_send_queued_msg();
 
 	key = k_spin_lock(&ipc->lock);
 
 	if (!list_is_empty(&ipc->msg_list) && !ipc->pm_prepare_D3)
-		schedule_ipc_worker();
+		more_to_send = true;
 
 	k_spin_unlock(&ipc->lock, key);
+
+	/*
+	 * If there's a race and the message is already sent and/or
+	 * prepare_D3 is set, we may do an extra wake-up of the
+	 * worker, but it is harmless as the worker will take
+	 * the lock and check the status again.
+	 */
+	if (more_to_send)
+		schedule_ipc_worker();
 }
 #endif