Re: [PATCH 01/23] net, sunrpc: convert rpc_cred.cr_count from atomic_t to refcount_t

[Trond Myklebust <trondmy-7I+n7zu2hftEKMMhf/gKZA@public.gmane.org>]

On Fri, 2017-03-17 at 09:02 -0400, Jeff Layton wrote:
> On Fri, 2017-03-17 at 12:50 +0000, Trond Myklebust wrote:
> > On Fri, 2017-03-17 at 14:10 +0200, Elena Reshetova wrote:
> > > refcount_t type and corresponding API should be
> > > used instead of atomic_t when the variable is used as
> > > a reference counter. This allows to avoid accidental
> > > refcounter overflows that might lead to use-after-free
> > > situations.
> > > 
> > > Signed-off-by: Elena Reshetova <elena.reshetova@intel.com>
> > > Signed-off-by: Hans Liljestrand <ishkamiel@gmail.com>
> > > Signed-off-by: Kees Cook <keescook@chromium.org>
> > > Signed-off-by: David Windsor <dwindsor@gmail.com>
> > > ---
> > >  include/linux/sunrpc/auth.h |  8 ++++----
> > >  net/sunrpc/auth.c           | 12 ++++++------
> > >  2 files changed, 10 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/include/linux/sunrpc/auth.h
> > > b/include/linux/sunrpc/auth.h
> > > index b1bc62b..bd36e0b 100644
> > > --- a/include/linux/sunrpc/auth.h
> > > +++ b/include/linux/sunrpc/auth.h
> > > @@ -15,7 +15,7 @@
> > >  #include <linux/sunrpc/msg_prot.h>
> > >  #include <linux/sunrpc/xdr.h>
> > >  
> > > -#include <linux/atomic.h>
> > > +#include <linux/refcount.h>
> > >  #include <linux/rcupdate.h>
> > >  #include <linux/uidgid.h>
> > >  #include <linux/utsname.h>
> > > @@ -68,7 +68,7 @@ struct rpc_cred {
> > >  #endif
> > >  	unsigned long		cr_expire;	/* when
> > > to gc
> > > */
> > >  	unsigned long		cr_flags;	/* various
> > > flags */
> > > -	atomic_t		cr_count;	/* ref count */
> > > +	refcount_t		cr_count;	/* ref count
> > > */
> > > 
> > 
> > NACK. That's going to be hitting
> > WARN_ONCE(!refcount_inc_not_zero(r),
> > "refcount_t: increment on 0; use-after-free.\n") like there's no
> > tomorrow...
> > 
> > Please stop with these automated conversions. They are going to
> > cause a
> > lot more bugs than they fix.
> > 
> 
> Agreed. These patchsets are touching places where we've already
> banged
> out most of the refcounting bugs. I'm against doing large scale
> conversions like this without a damned good reason.
> 
> I think it may be best to do this sort of thing in a more piecemeal
> fashion. Pick a subsystem or two and do the conversions there to
> prove
> that they're better than what we have. If the subsystem already has
> problems with its refcounting, then so much the better. Point to bugs
> that this new infrastructure helped find.
> 
> Encourage people to adopt your new infrastructure as new refcounted
> objects are introduced into the kernel. You might even consider a LWN
> article about this.
> 
> Eventually we'll get around to changing existing code to use it, once
> there is a sufficient advantage to doing so. Most likely when we're
> reworking the code for other reasons, or when we're chasing some
> horrid
> refcounting bug and think that this might help find it.

The main issue is that this "refcount_t" implementation appears to be
assuming that there is one and only one model for refcounts (the one
where a value of "0" means "free me immediately").

The kernel has a plethora of object caching implementations where this
is simply not the case; the dcache is a prime example, and this cache
is another. In both these implementation, the atomic_t variable is
being used more as a semaphore-style lock that prevents freeing of the
object while it is in active use as opposed to being freeable, but
cached. This is why these automated conversions are a nuisance and a
source of bugs.

-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@primarydata.com