On 05/31/2018 11:44 AM, Toke Høiland-Jørgensen wrote:
> Song Liu <liu.song....@gmail.com> writes:
>
>> On Wed, May 30, 2018 at 9:45 AM, Toke Høiland-Jørgensen <t...@toke.dk> wrote:
>>> This adds an example program showing how to sample packets from XDP using
>>> the perf event buffer. The example userspace program just prints the
>>> ethernet header for every packet sampled.
>>>
>>> Most of the userspace code is borrowed from other examples, most notably
>>> trace_output.
>>>
>>> Note that the example only works when everything runs on CPU0; so
>>> suitable smp_affinity needs to be set on the device. Some drivers seem
>>> to reset smp_affinity when loading an XDP program, so it may be
>>> necessary to change it after starting the example userspace program.
>>
>> Why does this only works when everything runs on CPU0? Is this
>> something we can improve?
>
> Yeah, good question. Basically, the call from XDP to
> bpf_perf_event_output() will fail with -EOPNOTSUPP. I tracked this down
> to this if statement in __bpf_perf_event_output() in bpf_trace.c:
>
>> if (unlikely(event->oncpu != cpu))
>> return -EOPNOTSUPP;
>
> I *think* that the way to fix this is for the userspace program to open
> a perf file descriptor for each CPU in the system and poll all of them,
> in which case the XDP program can pass the BPF_F_CURRENT_CPU flag to
> access the right one.
That is correct, you need one perf fd per cpu, and map them accordingly
into the map slots when you use BPF_F_CURRENT_CPU.
