On 2026/2/22 00:51, Alexei Starovoitov wrote:

On Sat, Feb 21, 2026 at 4:45 AM Leon Hwang [off-list ref] wrote:

quoted

[...]

quoted

Good idea. But I concern about the "in kernel disasm". Do you mean we
will build a disassembler for whitelist kfuncs at starting?

I noticed you've mentioned the same direction in "bpf/s390: Implement
get_preempt_count()" [1]. So, I added Ilya here to discuss this direction.

Much better. Keep it human.

"in kernel disasm" already exists for some architectures
(at least x86 and arm64) since it's being used by kprobes.
The ask here is to figure out whether they're usable for such
insn analysis. x86 disasm is likely capable.

After looking into x86&arm insn decoder, they are able to do insn analysis.

re:"whitelist kfunc"
I suspect an additional list is not necessary.
kf_fastcall is a good enough signal that such kfunc should
be inlinable.

I thought it was to build a light-weight custom disassembler, which
would only support limited machine codes (whitelist kfunc).

Obviously, I was wrong.

We can reuse the in-kernel insn decoding ability to validate fastcall
function by checking the registers use.

I'll post RFC after finishing poc, on both x86_64 and arm64 of course.

Thanks,
Leon