[OpenAFS-devel] kernel locking

[John S. Bucy]

On Fri, Nov 05, 2004 at 03:56:52PM -0500, Chaskiel M Grundman wrote:
> --On Friday, November 05, 2004 15:19:39 -0500 "John S. Bucy"
> <bucy-openafs-devel@gloop.org> wrote:
> 
> > 
> > Our experience with PAG problems has been that when the system goes to
> > the wall and runs out of PAGs that the cache manager effectively
> > deadlocks.  afs_setpag() is protected by the GLOCK and so it seems
> > plausible that reducing its use of the GLOCK might be beneficial. It
> > seems like it shouldn't be much work to add some finer-grained locking
> > here; I'm just not sure what locks I should be using.  The afs/lock.h
> > ones?
> 
> you can only call ObtainXXXLock or afs_osi_Wait()/afs_osi_Sleep while
> holding the GLOCK. The lock functions do not actually contain a mutual
> exclusion mechanism - they rely on the GLOCK to provide one. the GLOCK is
> automatically released by afs_osi_Wait/afs_osi_Sleep, so I don't think that
> is the problem you are having.

I noticed that looking at the code.  This may be a dumb question but
it seems like this locking scheme will lead to a lot of contention for
GLOCK.  Is it known how this affects the concurrency of the client?
The obvious thing to want to do is have a struct afs_mutex which
contains whatever the underlying system gives us.  It would probably
be a lot of work to change; I guess we'd need to establish that the
client is serializing around GLOCK needlessly, first.



john