On Thu, Sep 08, 2016 at 07:45:36AM +1000, Dave Chinner wrote:
> On Mon, Sep 05, 2016 at 05:15:29PM +0200, Christoph Hellwig wrote:
> > Hi Dave,
> >
> > I looked into killing the mrlock and ran into an unexpected problem.
> >
> > Currently mr_writer tracks that there is someone holding a write lock,
> > lockdep on the other hand checks if the calling thread has that lock.
> >
> > While that generally is the right semantic, our hack to offload
> > btree splits to a work item offends lockdep. E.g. this callstack
> > now asserts:
>
> It's a semaphore, not a mutex. Semaphore locking is independent of
> task context, the lock follows the object it protects, not the task
> that took the lock. i.e. Lockdep is wrong to assume the "owner" of a
> rw_sem will not change between lock and unlock.
We've added strict owner semantics to rwsem a long time ago.
If you want the actual semaphore semantics (which we greatly discourage,
because you cannot do validation on it) you should use
{down,up}_read_non_owner().
I'm not sure we've got write_non_owner() variants for this.
Turns out, there really are very few 'semaphore' users.
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
