Re: Concurrent BPF_PROG_TEST_RUN for XDP contend on dispatcher mutex

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Thu, Aug 26, 2021 at 4:09 AM Lorenz Bauer <lmb@cloudflare.com> wrote:
>
> Hi,
>
> One of the tests for our XDP-based load balancer has gotten quite
> slow, so I dug in. Roughly, it simulates 1m distinct packets arriving
> at the load balancer by calling BPF_PROG_TEST_RUN a million times.
>
>     distribution_test.go:40: 1000000 iterations
>     distribution_test.go:99: Coefficient of variation: 0.52%
> --- PASS: TestLoadBalancerDistribution (0.00s)
>     --- PASS: TestLoadBalancerDistribution/32_endpoints (22.04s)
>
> You can see that the test takes 20s. Running the same test with slight
> variations in three threads results in this:
>
>     distribution_test.go:40: 1000000 iterations
> === CONT  TestLoadBalancerDistribution/32_endpoints
>     distribution_test.go:99: Coefficient of variation: 0.60%
> === CONT  TestLoadBalancerDistribution/64_endpoints
>     distribution_test.go:99: Coefficient of variation: 0.82%
> === CONT  TestLoadBalancerDistribution/128_endpoints
>     distribution_test.go:99: Coefficient of variation: 1.24%
> --- PASS: TestLoadBalancerDistribution (0.00s)
>     --- PASS: TestLoadBalancerDistribution/32_endpoints (55.61s)
>     --- PASS: TestLoadBalancerDistribution/64_endpoints (55.61s)
>     --- PASS: TestLoadBalancerDistribution/128_endpoints (55.61s)
>
> It's pretty clear that something is serialising the threads. Digging
> around in perf reveals that the culprit is bpf_prog_change_xdp called
> from bpf_prog_test_run_xdp. The call was added in f23c4b3924d2 ("bpf:
> Start using the BPF dispatcher in BPF_TEST_RUN").
>
> Is there something we can do about this? Maybe only call into the
> dispatcher when repeat > 1?

Are you doing three parallel test_run commands
with repeat=1 and doing this syscall 1m times?
yeah, that would stress bpf_dispatcher_update() logic nicely :)
3m accesses to the same mutex and flip flop of a single page
with tlb flush and text_poke_bp.
Can your test harness use test_run with repeat = 1m instead?
Or it's not possible, since input data is different every time?
I think avoiding xdp dispatcher for repeat=1 makes sense.
Folks might be using this facility in similar fashion and
paying the dispatcher penalty for a single run is unnecessary.
While at it would be good to add the test_run specific xdp dispatcher.
Since right now all netdevs share a single global xdp dispatcher.
100 parallel xdp test_run threads will probably fail because
they will reach BPF_DISPATCHER_MAX limit.

Bjorn,
could you make such a change?

Other ideas?

Thanks!