Fix a hole in the scheduler where collisions between waking blocked threads in mutexes with interrupts waking blocked threads in message queues can occur. Queue posts will put the threads on a separate list that is then added to the running list with IRQs disabled on the next task switch or CPU wakeup. Basically no overhead for other operations. Seems a likely cause of my occasional observation of the backlight fade causing playback threads to stop running and a recently reported blocking violation upon USB plugging. Time will tell but banging the backlight on and off frequently hasn't hiccuped again for me on H120.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@12915 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 5 insertions, 2 deletions

diff --git a/firmware/kernel.c b/firmware/kernel.c
index c5e47a8..e09edef 100644
--- a/firmware/kernel.c
+++ b/firmware/kernel.c
@@ -126,7 +126,7 @@ static void queue_release_sender(struct thread_entry **sender,
                                  intptr_t retval)
 {
     (*sender)->retval = retval;
-    wakeup_thread(sender);
+    wakeup_thread_irq_safe(sender);
 #if 0
     /* This should _never_ happen - there must never be multiple
        threads in this list and it is a corrupt state */
@@ -289,11 +289,14 @@ void queue_post(struct event_queue *q, long id, intptr_t data)
     }
 #endif
 
-    wakeup_thread(&q->thread);
+    wakeup_thread_irq_safe(&q->thread);
     set_irq_level(oldlevel);
 }
 
 #ifdef HAVE_EXTENDED_MESSAGING_AND_NAME
+/* No wakeup_thread_irq_safe here because IRQ handlers are not allowed
+   use of this function - we only aim to protect the queue integrity by
+   turning them off. */
 intptr_t queue_send(struct event_queue *q, long id, intptr_t data)
 {
     int oldlevel = set_irq_level(HIGHEST_IRQ_LEVEL);