[Intel-wired-lan] FW: [PATCH bpf-next 2/4] xsk: allow AF_XDP sockets to receive packets directly from a queue

[Samudrala, Sridhar]

On 10/9/2019 10:17 AM, Alexei Starovoitov wrote:
> On Wed, Oct 9, 2019 at 9:53 AM Samudrala, Sridhar
> <sridhar.samudrala at intel.com> wrote:
>>
>>
>>>> +
>>>> +u32 bpf_direct_xsk(const struct bpf_prog *prog, struct xdp_buff *xdp)
>>>> +{
>>>> +       struct xdp_sock *xsk;
>>>> +
>>>> +       xsk = xdp_get_xsk_from_qid(xdp->rxq->dev, xdp->rxq->queue_index);
>>>> +       if (xsk) {
>>>> +               struct bpf_redirect_info *ri =
>>>> + this_cpu_ptr(&bpf_redirect_info);
>>>> +
>>>> +               ri->xsk = xsk;
>>>> +               return XDP_REDIRECT;
>>>> +       }
>>>> +
>>>> +       return XDP_PASS;
>>>> +}
>>>> +EXPORT_SYMBOL(bpf_direct_xsk);
>>>
>>> So you're saying there is a:
>>> """
>>> xdpsock rxdrop 1 core (both app and queue's irq pinned to the same core)
>>>      default : taskset -c 1 ./xdpsock -i enp66s0f0 -r -q 1
>>>      direct-xsk :taskset -c 1 ./xdpsock -i enp66s0f0 -r -q 1 6.1x improvement in drop rate """
>>>
>>> 6.1x gain running above C code vs exactly equivalent BPF code?
>>> How is that possible?
>>
>> It seems to be due to the overhead of __bpf_prog_run on older processors
>> (Ivybridge). The overhead is smaller on newer processors, but even on
>> skylake i see around 1.5x improvement.
>>
>> perf report with default xdpsock
>> ================================
>> Samples: 2K of event 'cycles:ppp', Event count (approx.): 8437658090
>> Overhead  Command          Shared Object     Symbol
>>     34.57%  xdpsock          xdpsock           [.] main
>>     17.19%  ksoftirqd/1      [kernel.vmlinux]  [k] ___bpf_prog_run
>>     13.12%  xdpsock          [kernel.vmlinux]  [k] ___bpf_prog_run
> 
> That must be a bad joke.
> The whole patch set is based on comparing native code to interpreter?!
> It's pretty awesome that interpreter is only 1.5x slower than native x86.
> Just turn the JIT on.

Thanks Alexei for pointing out that i didn't have JIT on.
When i turn it on, the performance improvement is a more modest 1.5x 
with rxdrop and 1.2x with l2fwd.

> 
> Obvious Nack to the patch set.
> 

Will update the patchset with the right performance data and address 
feedback from Bjorn.
Hope you are not totally against direct XDP approach as it does provide 
value when an AF_XDP socket is bound to a queue and a HW filter can 
direct packets targeted for that queue.