Re: [PATCH bpf-next] xsk, net: make sock_def_readable() have external linkage

On Tue, 21 Jan 2020 at 01:48, John Fastabend [off-list ref] wrote:

Björn Töpel wrote:

quoted

From: Björn Töpel <redacted>

XDP sockets use the default implementation of struct sock's
sk_data_ready callback, which is sock_def_readable(). This function is
called in the XDP socket fast-path, and involves a retpoline. By
letting sock_def_readable() have external linkage, and being called
directly, the retpoline can be avoided.

Signed-off-by: Björn Töpel <redacted>
---
 include/net/sock.h | 2 ++
 net/core/sock.c    | 2 +-
 net/xdp/xsk.c      | 2 +-
 3 files changed, 4 insertions(+), 2 deletions(-)

I think this is fine but curious were you able to measure the
difference with before/after pps or something?

Ugh, yeah, of course I've should have added that. Sorry for that! Here
goes; Benchmark is xdpsock rxdrop, NAPI running on core 20:

**Pre-patch: xdpsock rxdrop: 22.8 Mpps
 Performance counter stats for 'CPU(s) 20':

         10,000.58 msec cpu-clock                 #    1.000 CPUs
utilized
                12      context-switches          #    0.001 K/sec
                 1      cpu-migrations            #    0.000 K/sec
                 0      page-faults               #    0.000 K/sec
    29,931,407,416      cycles                    #    2.993 GHz
    82,538,852,331      instructions              #    2.76  insn per
cycle
    15,894,169,979      branches                  # 1589.324 M/sec
        30,916,486      branch-misses             #    0.19% of all
branches

      10.000636027 seconds time elapsed

**Post-patch: xdpsock rxdrop: 23.2 Mpps
         10,000.90 msec cpu-clock                 #    1.000 CPUs
utilized
                12      context-switches          #    0.001 K/sec
                 1      cpu-migrations            #    0.000 K/sec
                 0      page-faults               #    0.000 K/sec
    29,932,353,067      cycles                    #    2.993 GHz
    84,299,636,827      instructions              #    2.82  insn per
cycle
    16,228,795,437      branches                  # 1622.733 M/sec
        28,113,847      branch-misses             #    0.17% of all
branches

      10.000596454 seconds time elapsed

This could fall into the category of noise. :-) PPS and IPC is up a
bit. OTOH, maybe UDP can benefit from this as well?


Björn