We have a unit test that invokes an XDP program with 1m different
inputs, aka 1m BPF_PROG_RUN syscalls. We run this test concurrently
with slight variations in how we generated the input.
Since commit
f23c4b3924d2 ("bpf: Start using the BPF dispatcher in BPF_TEST_RUN")
the unit test has slowed down significantly. Digging deeper reveals that
the concurrent tests are serialised in the kernel on the XDP dispatcher.
This is a global resource that is protected by a mutex, on which we contend.
Fix this by not calling into the XDP dispatcher if we only want to perform
a single run of the BPF program.
See: https://lore.kernel.org/bpf/CACAyw9_y4QumOW35qpgTbLsJ532uGq-kVW-VESJzGyiZkypnvw@mail.gmail.com/
Signed-off-by: Lorenz Bauer <lmb@cloudflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210928093100.27124-1-lmb@cloudflare.com
if (ret)
goto free_data;
- bpf_prog_change_xdp(NULL, prog);
+ if (repeat > 1)
+ bpf_prog_change_xdp(NULL, prog);
ret = bpf_test_run(prog, &xdp, repeat, &retval, &duration, true);
/* We convert the xdp_buff back to an xdp_md before checking the return
* code so the reference count of any held netdevice will be decremented
sizeof(struct xdp_md));
out:
- bpf_prog_change_xdp(prog, NULL);
+ if (repeat > 1)
+ bpf_prog_change_xdp(prog, NULL);
free_data:
kfree(data);
free_ctx: