Re: [PATCH 2.6] serialization of kernelcapi work

[Armin Schindler]

On Thu, 18 Mar 2004, Andrew Morton wrote:
> I would suggest that you look at avoiding the global semaphore.  Suppose
> someone has 64 interfaces or something.  Is that possible?  It might be
> better to put the semaphore into struct capi_ctr so you can at least
> process frames from separate cards in parallel.
>
> > Is there a better way to do user-context work serialized ?
>
> Not really - you've been bitten by the compulsory per-cpuness of the
> workqueue handlers.
>
> You could have a standalone kernel thread or always queue the work onto CPU
> #0 (the function to do this isn't merged, but exists).  But both these are
> unscalable.
>
> So apart from moving recv_handler_lock into struct capi_ctr I can't think
> of anything clever.

I think an atomic counter in the workqueue-function would be more efficient.
What do you think about this?


static atomic_t recv_work_count;

static void recv_handler(void *data)
{
	if (atomic_inc_and_test(&recv_work_count)) {
		do {
			/* work to do */
		} while (!atomic_add_negative(-1, &recv_work_count));
	}
}

static int __init kcapi_init(void)
{
	atomic_set(&recv_work_count, -1);
}


A second call to the recv_handler() just results in an additional loop
of the first one and the second CPU goes on with its work.

Unfortunately, the atomic functions are not available on all architechtures.

atomic_dec_and_test() is the only atomic function with return value on all
platforms, right? Which isn't enough for the loop above.

Armin


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/