On Tue, 7 Jun 2011, Tim Sander wrote:

Nevertheless are there still scheduling latencies for the usermode
handler thread which then runs at rt prio 99 for about 1ms. I know
this is not an preempt-rt kernel but i hoped to get better values out
of this configuration. 

So if anybody has an idea how to get better latencies out of a 2.6.39
kernel, please let me know.

The mainline forced irq thread handling is no guarantee for lower
latencies. We need to disable preemption via local_bh_disable() for
the forced threaded interrupts to satisfy the handler vs. softirq
assumptions. So you might get long lasting preempt disabled regions
due to long running interrupt handlers. Aside of that you still can
get long periods due to code which runs with preemption or interrupts
disabled. The forced threaded option has no way to change that.

It would be interesting to find the root cause for those >1ms
latencies. Tracing is your friend.

Thanks,

	tglx