On Wed, Dec 12, 2018 at 11:54:20AM -0800, Stanislav Fomichev wrote:
> On 12/12, Stanislav Fomichev wrote:
> > On 12/12, Alexei Starovoitov wrote:
> > > On Wed, Dec 12, 2018 at 10:59:13AM -0800, Stanislav Fomichev wrote:
> > > > On 12/12, Alexei Starovoitov wrote:
> > > > > On Wed, Dec 12, 2018 at 10:27:24AM -0800, Stanislav Fomichev wrote:
> > > > > > The following prog types don't make sense without bpf cgroup:
> > > > > > * BPF_PROG_TYPE_CGROUP_SKB
> > > > > > * BPF_PROG_TYPE_CGROUP_SOCK
> > > > > > * BPF_PROG_TYPE_CGROUP_SOCK_ADDR
> > > > > >
> > > > > > Skip running verifier tests that exercise these prog types if
> > > > > > kernel is built without proper support.
> > > > > >
> > > > > > See commit e5c504858a18 ("selftests/bpf: skip verifier sockmap tests
> > > > > > on kernels without support") for original motivation.
> > > > > >
> > > > > > Signed-off-by: Stanislav Fomichev <s...@google.com>
> > > > > > ---
> > > > > > tools/testing/selftests/bpf/test_verifier.c | 35
> > > > > > +++++++++++++++++++++
> > > > > > 1 file changed, 35 insertions(+)
> > > > > >
> > > > > > diff --git a/tools/testing/selftests/bpf/test_verifier.c
> > > > > > b/tools/testing/selftests/bpf/test_verifier.c
> > > > > > index f5015566ae1b..b5470a399996 100644
> > > > > > --- a/tools/testing/selftests/bpf/test_verifier.c
> > > > > > +++ b/tools/testing/selftests/bpf/test_verifier.c
> > > > > > @@ -44,6 +44,7 @@
> > > > > > /* fallback to all features enabled */
> > > > > > # define CONFIG_BPF_STREAM_PARSER 1
> > > > > > # define CONFIG_XDP_SOCKETS 1
> > > > > > +# define CONFIG_CGROUP_BPF 1
> > > > >
> > > > > I really don't like where these is going.
> > > > > I think previous set should be reverted.
> > > > > This is not a scalable approach.
> > > > > Use libbpf probing approach to check whether feature is present
> > > > > instead.
> > > > I can probably add runtime probing instead of depending on compile-time
> > > > config, but I think that we would still need some per-test mechanism
> > > > to say that it depends on feature X (per-test .config_disabled or
> > > > similar).
> > > > Will moving these checks to runtime address your concern? (there is
> > > > sill a
> > > > scalability issue though)
> > >
> > > you said it youself. config_disabled doesn't scale.
> > > hence it's not a solution regardless of macro or runtime probing.
> > Let me see if I can use per-test prog_type for this type of probing.
> > Load 'return 0' program per-prog_type and use it as an indication of
> > runtime support. This should probably work in the majority of the cases.
> > I'll get back to you shortly.
>
> How about doing something like the (very work-in-progress) patch below?
> It mostly gets the job done, I think I need to do similar probing for maps
> and skip the tests that have fixup_map_{sockmap,sockhash,xskmap,stacktrace} in
> case appropriate map type is not supported.
>
> ---
>
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index aa582cd5bfcf..e55c116ff971 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -157,8 +157,11 @@ enum bpf_prog_type {
> BPF_PROG_TYPE_LIRC_MODE2,
> BPF_PROG_TYPE_SK_REUSEPORT,
> BPF_PROG_TYPE_FLOW_DISSECTOR,
> + __MAX_BPF_PROG_TYPE
> };
>
> +#define MAX_BPF_PROG_TYPE __MAX_BPF_PROG_TYPE
> +
that won't work either.
"bpf feature set" is a lot more than number of program
and map types the kernel supports. There are all sorts of
helper combinations, hooks, and verifier improvements.
test_verifier.c must test all that.
I don't think there is a way to make usptream test_verfier.c
not to report failure on older kernels.
I think you have to backport and enable everything
if you want to run the latest test_verifier.c on your kernel.
Other options are: deal with test failures in post processing
or have test_verifier.c specific for your kernel.
Since you've decided not to backport kernel side of sockmap,
don't enable these tests in _your_ test_verifier.c.
I've dropped two 'skip * sockmap' patches in bpf-next.
