Re: hyperthreading and tstile

[Jean-Yves Migeon <jeanyves.migeon%free.fr@localhost>]

Le 19/10/2013 20:28, Patrick Welche a écrit :
> Could the tstiles have been because of the cpu which failed to become
> ready?

I don't think so, CPUs that failed to init cannot take part in the 
scheduling.

However, if another cpu starts acting in the same way (ie. fails to 
wake up after being signaled), LWPs may become blocked in tstile. They 
are waiting for a ressource that will probably never be available again 
if the CPU went away.

It would be more useful to see the ressources on which all LWPs are 
waiting on though.

> Is that really a hardware problem?

We have to test it more before saying anything. If you cannot reproduce 
the same issue with another OS while keeping HT enabled (4 CPUs 
available and working), NetBSD is missing something. Caching issues 
probably, HT play dirty tricks with those.

Otherwise, the hardware issue is a good candidate; but I doubt that.

--
Jean-Yves Migeon