Re: [RFC 04/10] kmod: provide wrappers for kmod_concurrent inc/dec

[Luis R. Rodriguez]

On Thu, Dec 15, 2016 at 01:46:25PM +0100, Petr Mladek wrote:
> On Thu 2016-12-08 22:08:59, Luis R. Rodriguez wrote:
> > On Thu, Dec 08, 2016 at 12:29:42PM -0800, Kees Cook wrote:
> > > On Thu, Dec 8, 2016 at 11:48 AM, Luis R. Rodriguez <mcgrof@xxxxxxxxxx> wrote:
> > > > kmod_concurrent is used as an atomic counter for enabling
> > > > the allowed limit of modprobe calls, provide wrappers for it
> > > > to enable this to be expanded on more easily. This will be done
> > > > later.
> > > >
> > > > Signed-off-by: Luis R. Rodriguez <mcgrof@xxxxxxxxxx>
> > > > ---
> > > >  kernel/kmod.c | 27 +++++++++++++++++++++------
> > > >  1 file changed, 21 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/kernel/kmod.c b/kernel/kmod.c
> > > > index cb6f7ca7b8a5..049d7eabda38 100644
> > > > --- a/kernel/kmod.c
> > > > +++ b/kernel/kmod.c
> > > > @@ -108,6 +111,20 @@ static int call_modprobe(char *module_name, int wait)
> > > >         return -ENOMEM;
> > > >  }
> > > >
> > > > +static int kmod_umh_threads_get(void)
> > > > +{
> > > > +       atomic_inc(&kmod_concurrent);
> 
> This approach might actually cause false failures. If we
> are on the limit and more processes do this increment
> in parallel, it makes the number bigger that it should be.

This approach is *exactly* what the existing code does :P
I just provided wrappers. I agree with the old approach though,
reason is it acts as a lock in for the bump. What seems rather
stupid though is to just reject with an error on limit without first
taking a breather. I've now added a little clutch so that we first
take some fresh air when close to the limit, this reduces the chances
of going fatal.

With a clutch in place we can still go over the limit, its just we'd
have a few threads waiting until previous calls clear out. If there
is enough calls waiting eventually we'll fail.

Note though that __request_module() can wait, but here is an option
to not wait so such a clutch can only wait if we're allowed to.

> > > > +       if (atomic_read(&kmod_concurrent) < max_modprobes)
> > > > +               return 0;
> > > > +       atomic_dec(&kmod_concurrent);
> > > > +       return -ENOMEM;
> > > > +}
> > > > +
> > > > +static void kmod_umh_threads_put(void)
> > > > +{
> > > > +       atomic_dec(&kmod_concurrent);
> > > > +}
> > > 
> > > Can you use a kref here instead? We're trying to kill raw use of
> > > atomic_t for reference counting...
> > 
> > That's a much broader functional change than I was looking for, but I am up for
> > it. Can you describe the benefit of using kref you expect or why this is an
> > ongoing crusade? Since its a larger functional change how about doing this
> > change later, and we can test impact with the tress test driver. In theory if
> > there are benefits can't we add a test case to prove the gains?
> 
> Kees probably refers to the kref improvements that Peter Zijlstra
> is working on, see
> https://lkml.kernel.org/r/20161114174446.832175072@xxxxxxxxxxxxx
> 
> The advantage is that the new refcount API handles over and
> underflow.
> 
> Another advantage is that it increments/decrements the value
> only when it is safe. It uses cmpxchg to make sure that
> the checks are valid.

Great thanks, will look into that.

  Luis