On Sat, May 12, 2007 at 08:32:36PM -0400, shivers_at_ccs.neu.edu wrote:
> You should probably add the "reintegrate" option to the cfs man page; it's not
> there (it is in the "cfs help" output, though). My man page 

There are quite a few cfs commands that are still undocumented in the
man page. I have an open ticket in the bug tracker that lists others
that are also missing.

>    > Ten minutes later, it's still not visible on client B. During this ten
>    > minutes, venus on client A spins, maxing out my load average.
> 
>    It may be that the new reintegration parameters set by cfs strong are
>    causing some sort of a problem. I'll have to see if I can reproduce the
>    problem.
> 
> Yeah, the venus-at-100%-of-cpu thing is pretty common right after
> I get back on the net; it usually lasts for about 10-15 minutes.
> During this time, by the way, codacon is pretty calm -- it's not
> blasting out "validate" messages or anything.

Interesting, in that case the 100% cpu usage probably doesn't have
anything to do with reintegration. I guess it is the demotion of all
cached objects as a result of the server/volume state change(s).

I guess that code path may be missing a yield in the outer loop. This
wouldn't fix the CPU usage, but make the system a little more responsive
again. A better fix may be to use some sort of an epoch/event counter
when the volume state changes and use that to detect which objects need
to be revalidated. Not sure if such a solution would merge well into the
existing revalidation mechanism.

Jan