On Tue, Nov 18, 2025 at 2:41 PM Alexei Starovoitov
[off-list ref] wrote:

On Mon, Nov 17, 2025 at 10:34 PM Menglong Dong [off-list ref] wrote:

quoted

On 2025/11/18 14:31, Alexei Starovoitov wrote:

quoted

On Sun, Nov 16, 2025 at 7:49 PM Menglong Dong [off-list ref] wrote:

quoted

For now, the bpf trampoline is called by the "call" instruction. However,
it break the RSB and introduce extra overhead in x86_64 arch.

Please include performance numbers in the cover letter when you respin.

Hmm...I included a little performance, do you mean more performance
data? Current description:

As we can see above, the RSB is totally balanced. After the modification,
the performance of fexit increases from 76M/s to 130M/s.

I saw that. I meant full comparison with fentry and fmodret.
I suspect fmodret improved as well, right?
And include the command line that you used to measure.
selftests/bpf/bench...
so there is a way to reproduce what patchset claims.

I see. "fmodret" improved too, and all the BPF prog that based on
bpf trampoline origin call have a performance improvement.

I'll add the full comparison results in the next version.

Thanks!
Menglong Dong