Re: [PATCH RFC bpf-next 3/4] bpf/selftests: add tests to validate proper arguments alignment on ARM64

[Eduard Zingerman <eddyz87@gmail.com>]

On Fri, 2025-04-11 at 22:32 +0200, Alexis Lothoré (eBPF Foundation) wrote:
> When dealing with large types (>8 bytes), ARM64 trampolines need to take
> extra care about the arguments alignment to respect the calling
> convention set by AAPCS64.
> 
> Add two tests ensuring that the BPF trampoline arranges arguments with
> the relevant layout. The two new tests involve almost the same
> arguments, except that the second one requires a more specific alignment
> to be set by the trampoline when preparing arguments before calling the
> the target function.
> 
> Signed-off-by: Alexis Lothoré (eBPF Foundation) <alexis.lothore@bootlin.com>
> ---

[...]

> +SEC("fentry/bpf_testmod_test_struct_arg_11")
> +int BPF_PROG2(test_struct_many_args_9, struct bpf_testmod_struct_arg_5, a,
> +	      struct bpf_testmod_struct_arg_5, b,
> +	      struct bpf_testmod_struct_arg_5, c,
> +	      struct bpf_testmod_struct_arg_5, d, int, e,
> +	      struct bpf_testmod_struct_arg_5, f)

Hello Alexis,

I'm trying to double check the error you've seen for x86.
I see that tracing_struct/struct_many_args fails with assertion:
"test_struct_many_args:FAIL:t11:f unexpected t11:f: actual 35 != expected 43".
Could you please help me understand this test?
The function listened to is defined as accepting 'struct bpf_testmod_struct_arg_7',
at the same time this function uses 'struct bpf_testmod_struct_arg_5'.
Nevertheless, the assertion persists even with correct types.

> +{
> +	t11_a = a.a;
> +	t11_b = b.a;
> +	t11_c = c.a;
> +	t11_d = d.a;
> +	t11_e = e;
> +	t11_f = f.a;
> +	return 0;
> +}

[...]