On Fri, 16 Feb 2007 08:43:37 +1100
Benjamin Herrenschmidt [off-list ref] wrote:

On Thu, 2007-02-15 at 18:50 +0100, Christoph Hellwig wrote:

quoted

On Thu, Feb 15, 2007 at 04:23:02PM +0100, Geert.Uytterhoeven@sonycom.com wrote:

quoted

+	do {
+		try_to_freeze();
+		error = down_interruptible(&ps3fb.sem);
+		if (!error && !atomic_read(&ps3fb.ext_flip))
 			ps3fb_sync(0);	/* single buffer */

this still can deadlock when calling kthread_stop.  You really want
to use wake_up_process to kick this thread or use a workqueue.

kthread_stop does wake_up_process no ? However, that might not get you
out of interruptible if you don't also have signal_pending...

No, it won't get you out of down_interruptible().  But the code would have
failed trivial testing so perhaps we're missing something.

It seems crufty to use semaphores in this manner.  afaict all we're doing
here is poking a kernel thread and asking it to do a bit of work.  The
standard way of doing this is to go to sleep on a waitqueue_head.

	DEFINE_WAIT(wait);

	while (!kthread_should_stop()) {
		prepare_to_wait(&wq, &wait, TASK_INTERRUPTIBLE);
		if (!atomic_read(&ps3fb.ext_flip))
			schedule();
		finish_wait(&wq, &wait);
		if (!atomic_read(&ps3fb.ext_flip))
			WARN_ON(1);
		else
			ps3fb_sync(0);
	}

and, elsewhere,

	atomic_inc(&ps3fb.ext_flip);
	wake_up_process(my_kernel_therad);