Re: [PATCH bpf-next v4 12/14] selftests/bpf: test_xdp_veth: Add XDP broadcast redirection tests

[Bastien Curutchet]

Hi Martin,

On 2/1/25 2:33 AM, Martin KaFai Lau wrote:
> On 1/30/25 11:21 PM, Bastien Curutchet (eBPF Foundation) wrote:
> > +#define BROADCAST_REDIRECT_SKEL_NB    2
> > +static void xdp_veth_broadcast_redirect(u32 attach_flags, u64

[...]

> > +
> > +    group_map = bpf_map__fd(xdp_redirect_multi_kern->maps.map_all);
> > +    if (!ASSERT_OK_FD(group_map, "open map_all"))
> > +        goto destroy_xdp_redirect_map;
> > +
> > +    flags_map = bpf_map__fd(xdp_redirect_multi_kern- 
> > >maps.redirect_flags);
> > +    if (!ASSERT_OK_FD(group_map, "open map_all"))
> > +        goto destroy_xdp_redirect_map;
> > +
> > +    err = bpf_map_update_elem(flags_map, &protocol, &redirect_flags, 
> > BPF_NOEXIST);
> > +    if (!ASSERT_OK(err, "init IP count"))
> > +        goto destroy_xdp_redirect_map;
> > +
> > +    cnt_map = bpf_map__fd(xdp_redirect_map->maps.rxcnt);
> > +    if (!ASSERT_OK_FD(cnt_map, "open rxcnt map"))
> > +        goto destroy_xdp_redirect_map;
> > +
> > +    bpf_objs[0] = xdp_redirect_multi_kern->obj;
> > +    bpf_objs[1] = xdp_redirect_map->obj;
> > +    for (i = 0; i < VETH_PAIRS_COUNT; i++) {
> > +        int ifindex = if_nametoindex(net_config[i].local_veth);
> > +
> > +        if (attach_programs_to_veth_pair(bpf_objs, 
> > BROADCAST_REDIRECT_SKEL_NB,
> > +                         net_config, prog_cfg, i))
> > +            goto destroy_xdp_redirect_map;
> > +
> > +        SYS(destroy_xdp_redirect_map,
> > +            "ip -n %s neigh add %s lladdr 00:00:00:00:00:01 dev %s",
> > +            net_config[i].namespace, IP_NEIGH, 
> > net_config[i].remote_veth);
> > +
> > +        devmap_val.ifindex = ifindex;
> > +        err = bpf_map_update_elem(group_map, &ifindex, &devmap_val, 0);
>
> I ran this test in a loop and failed at this line permanently (errno 
> E2BIG -7) after enough iterations. I believe the problem is the 
> group_map (aka "map_all" in the BPF program) has a max_entries 1024 and 
> ifindex can go beyond 1024 after some "./test_progs" iterations. 
> Understood that it is likely an existing assumption in the "map_all" 
> definition but it needs to be addressed first before moving to test_progs.
>
> > +        if (!ASSERT_OK(err, "bpf_map_update_elem"))
> > +            goto destroy_xdp_redirect_map;
> > +
> > +    }
> > +

[...]

> > diff --git a/tools/testing/selftests/bpf/progs/xdp_redirect_map.c b/ 
> > tools/testing/selftests/bpf/progs/xdp_redirect_map.c
> > index 
> > 682dda8dabbc9abbb5d1b0b22dd5f81124142e79..14385df71d7fc40c3b0ee5c6ea0760d0e7336d71 100644
> > --- a/tools/testing/selftests/bpf/progs/xdp_redirect_map.c
> > +++ b/tools/testing/selftests/bpf/progs/xdp_redirect_map.c
> > @@ -1,7 +1,11 @@
> >   // SPDX-License-Identifier: GPL-2.0
> > +#include <linux/if_ether.h>
> > +#include <linux/ip.h>
>
> I have compiler error complaining about __always_inline not defined. 
> Likely ordering issue and environment specific. Regardless, I believe 
> the ip.h is unnecessary, so better clean it up.
>
> I am going to land the patch 1-10 and change the cover letter a little 
> to reflect the fact that patch 11-14 will be a followup. Please post the 
> remaining patches after fixing the bpf_map_update_elem issue. Thanks.

Ok thank you, I plan to solve this issue by using a dedicated namespace 
instead of the root namespace so the index won't be incremented every 
time `test_progs` is launched.
I'll send the new iteration once tested on my side.

Best regards,
Bastien