On 10/29/2015 09:02 AM, Jorge Ramirez-Ortiz wrote:
On 10/29/2015 01:39 AM, Kazumoto Kojima wrote:
$ smokey --verbose=2 --run=2 duration=0
you will see the fpu registers being trashed and not correctly saved/restored while cobalt context switches between the realtime and the non realtime threads.
Thank you for the heads up! I'm using a bit old linux-ipipe/xenomai trees and haven't experienced that issue. I'll try it with the updated trees.
With the updated trees, I can see the suggested errors for xenomai smokey test. Although I'm afraid to do a stupid thing, with the change below, smokey test has been running ~50min successfully now.
I think this is exactly what we have been looking for (many thanks for sharing it) testing in UP mode I can see a negative impact on the minimum latency figures but it certainly addresses the fpu stress switch and seems to address the stalls as well. it might be a bit early to say this but this seems like good news :)
having said that, the patch however doesnt seem to take into account the fpu activity of the regular kernel so there might be some other changes coming on top of this one.