On Thu 23-06-16 08:58:16, Dave Chinner wrote:
> On Wed, Jun 22, 2016 at 02:38:22PM +0200, Michal Hocko wrote:
> > On Wed 22-06-16 11:03:20, Dave Chinner wrote:
> > > On Tue, Jun 21, 2016 at 04:26:28PM +0200, Michal Hocko wrote:
> > > > On Wed 15-06-16 17:21:54, Dave Chinner wrote:
> > [...]
> > > > > There are allocations outside transaction context which need to be
> > > > > GFP_NOFS - this is what KM_NOFS was originally intended for.
> > > >
> > > > Is it feasible to mark those by the scope NOFS api as well and drop
> > > > the direct KM_NOFS usage? This should help to identify those that are
> > > > lockdep only and use the annotation to prevent from the false positives.
> > >
> > > I don't understand what you are suggesting here. This all started
> > > because we use GFP_NOFS in a handful of places to shut up lockdep
> > > and you didn't want us to use GFP_NOFS like that. Now it sounds to
> > > me like you are advocating setting unconditional GFP_NOFS allocation
> > > contexts for entire XFS code paths - whether it's necessary or
> > > not - to avoid problems with lockdep false positives.
> >
> > No, I meant only those paths which need GFP_NOFS for other than lockdep
> > purposes would use the scope api.
> >
> > Anyway, it seems that we are not getting closer to a desired solution
> > here. Or I am not following it at least...
> >
> > It seems that we have effectively two possibilities (from the
> > MM/lockdep) POV. Either add an explicit API to disable the reclaim
> > lockdep machinery for all allocation in a certain scope or a GFP mask
> > to to achieve the same for a particular allocation. Which one would work
> > better for the xfs usecase?
>
> As I've said - if we annotate the XFS call sites appropriately (e.g.
> KM_NOLOCKDEP rather than KM_NOFS), we don't care what lockdep
> mechanism is used to turn off warnings as it will be wholly
> encapsulated inside kmem_alloc() and friends. This will end up
> similar to how we are currently encapsulate the memalloc_noio_save()
> wrappers in kmem_zalloc_large().
OK, I see. So which way do we go Peter? Are you going to send the GFP
one or is there a way to bribe you to go with a thread flag?
--
Michal Hocko
SUSE Labs
|