Incorrect BPF stats accounting for fentry on arm64

[Ivan Babrou]

On Thu, Apr 11, 2024 at 7:30 PM Xu Kuohai <xukuohai at huaweicloud.com> wrote:
>
> On 4/12/2024 2:09 AM, Ivan Babrou wrote:
> > Hello,
> >
> > We're seeing incorrect data for bpf runtime stats on arm64. Here's an example:
> >
> > $ sudo bpftool prog show id 693110
> > 693110: tracing  name __tcp_retransmit_skb  tag e37be2fbe8be4726  gpl
> > run_time_ns 2493581964213176 run_cnt 1133532 recursion_misses 1
> >      loaded_at 2024-04-10T22:33:09+0000  uid 62727
> >      xlated 312B  jited 344B  memlock 4096B  map_ids 8550445,8550441
> >      btf_id 8726522
> >      pids prometheus-ebpf(2224907)
> >
> > According to bpftool, this program reported 66555800ns of runtime at
> > one point and then it jumped to 2493581675247416ns just 53s later when
> > we looked at it again. This is happening only on arm64 nodes in our
> > fleet on both v6.1.82 and v6.6.25.
> >
> > We have two services that are involved:
> >
> > * ebpf_exporter attaches bpf programs to the kernel and exports
> > prometheus metrics and opentelementry traces driven by its probes
> > * bpf_stats_exporter runs bpftool every 53s to capture bpf runtime metrics
> >
> > The problematic fentry is attached to __tcp_retransmit_skb, but an
> > identical one is also attached to tcp_send_loss_probe, which does not
> > exhibit the same issue:
> >
> > SEC("fentry/__tcp_retransmit_skb")
> > int BPF_PROG(__tcp_retransmit_skb, struct sock *sk)
> > {
> >    return handle_sk((struct pt_regs *) ctx, sk, sk_kind_tcp_retransmit_skb);
> > }
> >
> > SEC("fentry/tcp_send_loss_probe")
> > int BPF_PROG(tcp_send_loss_probe, struct sock *sk)
> > {
> >    return handle_sk((struct pt_regs *) ctx, sk, sk_kind_tcp_send_loss_probe);
> > }
> >
> > In handle_sk we do a map lookup and an optional ringbuf push. There is
> > no sleeping (I don't think it's even allowed on v6.1). It's
> > interesting that it only happens for the retransmit, but not for the
> > loss probe.
> >
> > The issue manifests some time after we restart ebpf_exporter and
> > reattach the probes. It doesn't happen immediately, as we need to
> > capture metrics 53s apart to produce a visible spike in metrics.
> >
> > There is no corresponding spike in execution count, only in execution time.
> >
> > It doesn't happen deterministically. Some ebpf_exporter restarts show
> > it, some don't.
> >
> > It doesn't keep happening after ebpf_exporter restart. It happens once
> > and that's it.
> >
> > Maybe recursion_misses plays a role here? We see none for
> > tcp_send_loss_probe. We do see some for inet_sk_error_report
> > tracepoint, but it doesn't spike like __tcp_retransmit_skb does.
> >
> > The biggest smoking gun is that it only happens on arm64.
> >
> > I'm happy to try out patches to figure this one out.
> >
>
> I guess the issue is caused by the not setting of x20 register
> when __bpf_prog_enter(prog) returns zero.

Yes, I think this is it. Your patch makes it match x86_64 and it seems logical.

I'm building a kernel with it to put it into production to make sure.

> The following patch may help:
>
> --- a/arch/arm64/net/bpf_jit_comp.c
> +++ b/arch/arm64/net/bpf_jit_comp.c
> @@ -1905,15 +1905,15 @@ static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_link *l,
>
>          emit_call(enter_prog, ctx);
>
> +       /* save return value to callee saved register x20 */
> +       emit(A64_MOV(1, A64_R(20), A64_R(0)), ctx);
> +
>          /* if (__bpf_prog_enter(prog) == 0)
>           *         goto skip_exec_of_prog;
>           */
>          branch = ctx->image + ctx->idx;
>          emit(A64_NOP, ctx);
>
> -       /* save return value to callee saved register x20 */
> -       emit(A64_MOV(1, A64_R(20), A64_R(0)), ctx);
> -
>          emit(A64_ADD_I(1, A64_R(0), A64_SP, args_off), ctx);
>          if (!p->jited)
>                  emit_addr_mov_i64(A64_R(1), (const u64)p->insnsi, ctx);
>