Re: [PATCH bpf-next] bpf: use kvmalloc in map_lookup_elem

[sdf@google.com]

On 08/16, Andrii Nakryiko wrote:
> On Mon, Aug 16, 2021 at 2:43 PM Daniel Borkmann <daniel@iogearbox.net>  
> wrote:
> >
> > On 8/16/21 6:48 PM, Stanislav Fomichev wrote:
> > > Use kvmalloc/kvfree for temporary value when looking up a map.
> > > kmalloc might not be sufficient for percpu maps where the value is  
> big.
> > >
> > > Can be reproduced with netcnt test on qemu with "-smp 255".
> > >
> > > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> > > ---
> > >   kernel/bpf/syscall.c | 4 ++--
> > >   1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > > index 9a2068e39d23..ae0b1c1c8ece 100644
> > > --- a/kernel/bpf/syscall.c
> > > +++ b/kernel/bpf/syscall.c
> > > @@ -1076,7 +1076,7 @@ static int map_lookup_elem(union bpf_attr *attr)
> > >       value_size = bpf_map_value_size(map);
> > >
> > >       err = -ENOMEM;
> > > -     value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
> > > +     value = kvmalloc(value_size, GFP_USER | __GFP_NOWARN);
> > >       if (!value)
> > >               goto free_key;
> >
> > What about other cases like map_update_elem(), shouldn't they be adapted
> > similarly?

> And in the same vein (with keys potentially being big as well), should
> we switch __bpf_copy_key() to use vmemdup_user() instead of
> memdup_user()?

Those are good questions :-)

I'm assuming that whatever is doing kmalloc on top of
bpf_map_value_size() should definitely use kvmalloc since
bpf_map_value_size() can blow up the value size. (will resend)

For __bpf_copy_key I'm less sure, but it might be a good idea as well.
Let me try to look at bit more into it, but feels like there shouldn't
be any downsides?