Re: [PATCH v2 bpf 3/5] bpf: bpf_prog_array_free() should take a generic non-rcu pointer

[Roman Gushchin]

On Wed, Jul 18, 2018 at 03:07:48PM +0200, Daniel Borkmann wrote:
> On 07/18/2018 12:55 AM, Roman Gushchin wrote:
> > On Wed, Jul 18, 2018 at 12:38:50AM +0200, Daniel Borkmann wrote:
> >> On 07/17/2018 12:57 AM, Roman Gushchin wrote:
> >>> On Tue, Jul 17, 2018 at 12:30:18AM +0200, Daniel Borkmann wrote:
> >>>> On 07/13/2018 09:41 PM, Roman Gushchin wrote:
> >>>>> bpf_prog_array_free() should take a generic non-rcu pointer
> >>>>> as an argument, as freeing the objects assumes that we're
> >>>>> holding an exclusive rights on it.
> >>>>>
> >>>>> rcu_access_pointer() can be used to convert a __rcu pointer to
> >>>>> a generic pointer before passing it to bpf_prog_array_free(),
> >>>>> if necessary.
> >>>>>
> >>>>> This patch eliminates the following sparse warning:
> >>>>> kernel/bpf/core.c:1556:9: warning: incorrect type in argument 1 (different address spaces)
> >>>>> kernel/bpf/core.c:1556:9:    expected struct callback_head *head
> >>>>> kernel/bpf/core.c:1556:9:    got struct callback_head [noderef] <asn:4>*<noident>
> >>>>>
> >>>>> Fixes: 324bda9e6c5a ("bpf: multi program support for cgroup+bpf")
> >>>>> Signed-off-by: Roman Gushchin <guro@xxxxxx>
> >>>>> Cc: Alexei Starovoitov <ast@xxxxxxxxxx>
> >>>>> Cc: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
> >>>>> ---
> >>>>>  drivers/media/rc/bpf-lirc.c |  6 +++---
> >>>>>  include/linux/bpf.h         |  2 +-
> >>>>>  kernel/bpf/cgroup.c         | 11 ++++++-----
> >>>>>  kernel/bpf/core.c           |  5 ++---
> >>>>>  kernel/trace/bpf_trace.c    |  8 ++++----
> >>>>>  5 files changed, 16 insertions(+), 16 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/media/rc/bpf-lirc.c b/drivers/media/rc/bpf-lirc.c
> >>>>> index fcfab6635f9c..509b262aa0dc 100644
> >>>>> --- a/drivers/media/rc/bpf-lirc.c
> >>>>> +++ b/drivers/media/rc/bpf-lirc.c
> >>>>> @@ -135,7 +135,7 @@ static int lirc_bpf_attach(struct rc_dev *rcdev, struct bpf_prog *prog)
> >>>>>  		goto unlock;
> >>>>>  
> >>>>>  	rcu_assign_pointer(raw->progs, new_array);
> >>>>> -	bpf_prog_array_free(old_array);
> >>>>> +	bpf_prog_array_free(rcu_access_pointer(old_array));
> >>>>
> >>>> Taking this one as an example, why can't we already do the rcu_dereference() on the
> >>>> 'old_array = raw->progs' where we fetch the old_array initially? Then we also wouldn't
> >>>> need the rcu_access_pointer() on bpf_prog_array_free() and yet another rcu_dereference()
> >>>> inside the bpf_prog_array_copy() from your later patch?
> >>>
> >>> We can, but then we have to change bpf_prog_array_copy() args annotation,
> >>> and also all places, where it's called.
> >>> IMO, basically all local variables and function args marked as __rcu
> >>> should be not marked as RCU, but fixing them all is beyond this patchset.
> >>
> >> Right, agree, the __rcu markings seem somewhat arbitrary. :-( I think we need to
> >> investigate this a bit deeper and do a proper audit on the whole bpf prog array's
> >> RCU handling (probably won't get to it in next two weeks but put onto backlog just
> >> in case it's still unresolved till then). That said, given this has been there for
> >> quite a while and it's rc5 now, I think we could start out on bpf-next with the
> >> obvious candidates which should be okay even if it ends up bigger.
> > 
> > Totally agree.
> > 
> >> First two from this series we could already take in if you prefer.
> > 
> > That would be nice!
> 
> Ok, done, applied 1+2 to bpf-next, thanks Roman!

Thanks, Daniel!