Re: suspicious __GFP_NOMEMALLOC in selinux

[Michal Hocko]

On Thu 03-08-17 19:44:46, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Thu 03-08-17 19:02:57, Tetsuo Handa wrote:
> > > On 2017/08/03 17:11, Michal Hocko wrote:
> > > > [CC Mel]
> > > > 
> > > > On Wed 02-08-17 17:45:56, Paul Moore wrote:
> > > >> On Wed, Aug 2, 2017 at 6:50 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > >>> Hi,
> > > >>> while doing something completely unrelated to selinux I've noticed a
> > > >>> really strange __GFP_NOMEMALLOC usage pattern in selinux, especially
> > > >>> GFP_ATOMIC | __GFP_NOMEMALLOC doesn't make much sense to me. GFP_ATOMIC
> > > >>> on its own allows to access memory reserves while the later flag tells
> > > >>> we cannot use memory reserves at all. The primary usecase for
> > > >>> __GFP_NOMEMALLOC is to override a global PF_MEMALLOC should there be a
> > > >>> need.
> > > >>>
> > > >>> It all leads to fa1aa143ac4a ("selinux: extended permissions for
> > > >>> ioctls") which doesn't explain this aspect so let me ask. Why is the
> > > >>> flag used at all? Moreover shouldn't GFP_ATOMIC be actually GFP_NOWAIT.
> > > >>> What makes this path important to access memory reserves?
> > > >>
> > > >> [NOTE: added the SELinux list to the CC line, please include that list
> > > >> when asking SELinux questions]
> > > > 
> > > > Sorry about that. Will keep it in mind for next posts
> > > >  
> > > >> The GFP_ATOMIC|__GFP_NOMEMALLOC use in SELinux appears to be limited
> > > >> to security/selinux/avc.c, and digging a bit, I'm guessing commit
> > > >> fa1aa143ac4a copied the combination from 6290c2c43973 ("selinux: tag
> > > >> avc cache alloc as non-critical") and the avc_alloc_node() function.
> > > > 
> > > > Thanks for the pointer. That makes much more sense now. Back in 2012 we
> > > > really didn't have a good way to distinguish non sleeping and atomic
> > > > with reserves allocations.
> > > >  
> > > >> I can't say that I'm an expert at the vm subsystem and the variety of
> > > >> different GFP_* flags, but your suggestion of moving to GFP_NOWAIT in
> > > >> security/selinux/avc.c seems reasonable and in keeping with the idea
> > > >> behind commit 6290c2c43973.
> > > > 
> > > > What do you think about the following? I haven't tested it but it should
> > > > be rather straightforward.
> > > 
> > > Why not at least __GFP_NOWARN ?
> > 
> > This would require an additional justification.
> 
> If allocation failure is not a problem, printing allocation failure messages
> is nothing but noisy.

That alone is not a sufficient justification. An allocation warning
might still tell you that something is not configured properly. Note
that I am not objecting that __GFP_NOWARN is wrong it should just not be
added blindly withtout deep understanding of the code which I do not
have.

> > > And why not also __GFP_NOMEMALLOC ?
> > 
> > What would be the purpose of __GFP_NOMEMALLOC? In other words which
> > context would set PF_NOMEMALLOC so that the flag would override it?
> > 
> 
> When allocating thread is selected as an OOM victim, it gets TIF_MEMDIE.
> Since that function might be called from !in_interrupt() context, it is
> possible that gfp_pfmemalloc_allowed() returns true due to TIF_MEMDIE and
> the OOM victim will dip into memory reserves even when allocation failure
> is not a problem.

Yes this is possible but I do not see any major problem with that.
I wouldn't add __GFP_NOMEMALLOC unless there is a real runaway of some
sort that could be abused.

> Thus, I think that majority of plain GFP_NOWAIT users want to use
> GFP_NOWAIT | __GFP_NOWARN | __GFP_NOMEMALLOC.

-- 
Michal Hocko
SUSE Labs