Large spikes in latencies

[Ken Moffat]

On Fri, May 29, 2015 at 06:14:46PM -0400, Yogi A. Patel wrote:
> 
> > On 05/28/2015 08:54 PM, Michel Dänzer wrote:
> > Can you narrow down when the problem was introduced in the kernel,
> > ideally with git bisect?
> > 

> I did a few experiments and here is an update:
> 
> + It seems that whatever is causing this is from drm/radeon and it came in sometime between 3.8.13 and 3.10.32. The dmesg output from 3.10.32 is attached, too

> 
> + the latencies are worse on 3.14 (40us) than on 3.10 (20-30us) regardless of the card.
> 
> + latencies on 3.8 for both cards are below 10us. 
> 

Just in case you don't realise how stable kernel releases happen:
First, Linus release .0.  For most people, these kernels work ok
most of the time.  Then gradual fixes appear during the development
of a later kernel, and get applied by the stable team.  So, I
suggest that an approach something like this might help:

Begin by testing 3.10.0 and (ideally) 3.8.0.

Is 3.10.0 worse than 3.8 ?  If it is, the problem was introduced
between 3.8 (good) and 3.10 (bad) and you can bisect linus's tree.

If 3.10.0 was good, the problem appeared in the 3.10 _stable_ tree.
For something like that, I have on occasion been tempted to take a
broad-brush initial approach and try the released versions - in this
case I might begin by trying e.g. 3.10.16, 3.10.8 or 3.10.24, etc to
get a better feeling for when it went bad.  But ultimately,
bisecting the 3.10 stable tree will be the way to go.

Also, it sounds as if 3.14 is a different, or extra, problem.

Would it be practical for you to move to a newer version if that
solved the problem ?  If so, you could try 4.0.latest just in case
it is already fixed.  A couple of times, I've had to do bisection in
-rc kernels, only to discover that a commit during the current week
had already removed the problem I had.

ĸen
-- 
Nanny Ogg usually went to bed early. After all, she was an old lady.
Sometimes she went to bed as early as 6 a.m.