Re: [PATCHv2 bpf-next 3/3] bpf, x86: Add support for jit dry run

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

On Thu Sep 3, 2026 at 2:20 AM PDT, Jiri Olsa wrote:
> Adding support to run jit code generation in dry_run mode that won't
> store any code and only returns the jir code size.
>
> The dry_run is enabled when __arch_prepare_bpf_trampoline is called
> with rw_image argument as NULL.
>
> It's used in arch_bpf_trampoline_size where it allows to skip the
> image allocation, that gives speed up for tracing_multi attachment.
>
> With current code:
>
>   # ./test_progs -t tracing_multi_bench_attach -v
>   ...
>   serial_test_tracing_multi_bench_attach: found 55227 functions
>   serial_test_tracing_multi_bench_attach: attached in   1.563s
>   serial_test_tracing_multi_bench_attach: detached in   0.256s
>
> With the fix:
>
>   # ./test_progs -t tracing_multi_bench_attach -v
>   ...
>   serial_test_tracing_multi_bench_attach: found 55235 functions
>   serial_test_tracing_multi_bench_attach: attached in   0.798s
>   serial_test_tracing_multi_bench_attach: detached in   0.258s
>
> Signed-off-by: Jiri Olsa <jolsa@kernel.org>
> ---
>  arch/x86/net/bpf_jit_comp.c | 42 ++++++++++++++++++-------------------
>  1 file changed, 20 insertions(+), 22 deletions(-)
>
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index ee2ddeba3de1..d1af7dc5c5ce 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -25,6 +25,7 @@ static bool all_callee_regs_used[4] = {true, true, true, true};
>  
>  struct jit_emit_context {
>  	u8 *prog;
> +	bool dry_run;
>  };
>  
>  static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
> @@ -42,7 +43,10 @@ static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
>  
>  static void emit_code_jit(struct jit_emit_context *jit, u32 bytes, unsigned int len)
>  {
> -	jit->prog = emit_code(jit->prog, bytes, len);
> +	if (jit->dry_run)
> +		jit->prog += len;
> +	else
> +		jit->prog = emit_code(jit->prog, bytes, len);

Sorry, I don't believe that skipping emit makes that much
of runtime difference.
I suspect jit_alloc + jit_free are costly.
In such case the earlier patches are not needed.
Point emit logic to some scratch area and discard it.
Same effect with half of the changes.

pw-bot: cr