Re: [PATCH bpf-next] bpf: remove is_return in struct bpf_session_run_ctx

[menglong.dong@linux.dev]

On 2025/9/24 03:23 Alexei Starovoitov <alexei.starovoitov@gmail.com> write:
> On Mon, Sep 22, 2025 at 7:11 AM Menglong Dong <menglong8.dong@gmail.com> wrote:
> >
> > On Mon, Sep 22, 2025 at 10:08 PM Song Liu <song@kernel.org> wrote:
> > >
> > > On Mon, Sep 22, 2025 at 11:57 AM Menglong Dong <menglong8.dong@gmail.com> wrote:
> > > >
> > > > The "data" in struct bpf_session_run_ctx is always 8-bytes aligned.
> > > > Therefore, we can store the "is_return" to the last bit of the "data",
> > > > which can make bpf_session_run_ctx 8-bytes aligned and save memory.
> > >
> > > Does this really save anything? AFAICT, bpf_session_run_ctx is
> > > only allocated on the stack. Therefore, we don't save any memory
> > > unless there is potential risk of stack overflow.
> >
> > Hi, Song. My original intention is to save the usage of the
> > stack to prevent potential stack overflow,
> 
> 8 bytes won't matter, but wasting 8 bytes for 1 bit is indeed annoying.
> 
> > especially when we
> > trace all the kernel functions with kprobe-multi.
> 
> What do you mean? kprobe-multi won't recurse,
> so tracing all or a few functions is the same concern
> from stack overflow pov, no ?

You are right, I made something wrong. I mixed it with origin
call case of the bpf trampoline, which will store all the things
in the stack for every function call.

> 
> > The most thing for me is that the unaligned field in the struct
> > looks very awkward, and it consumes 8-bytes only for a bit.
> 
> let's keep it as-is. If stack overflow is indeed an issue we need
> a generic way to detect it and prevent it.
> We've been thinking whether vmap stack guard pages
> can become JIT's extable-like things, so when stack overflow

Interesting

> happens we unwind stack and stop bpf prog instead of panicing.

Yeah, I think it's OK to keep it still.

Thanks!
Menglong Dong

> 
>