Re: [PATCH 3/3] rust: add tracepoint support

[Alice Ryhl]

On Thu, Jun 6, 2024 at 5:29 PM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> On 2024-06-06 11:05, Alice Ryhl wrote:
> > Make it possible to have Rust code call into tracepoints defined by C
> > code. It is still required that the tracepoint is declared in a C
> > header, and that this header is included in the input to bindgen.
> >
> > Signed-off-by: Alice Ryhl <aliceryhl@xxxxxxxxxx>
>
> [...]
>
> > diff --git a/rust/helpers.c b/rust/helpers.c
> > index 2c37a0f5d7a8..0560cc2a512a 100644
> > --- a/rust/helpers.c
> > +++ b/rust/helpers.c
> > @@ -165,6 +165,30 @@ rust_helper_krealloc(const void *objp, size_t new_size, gfp_t flags)
> >   }
> >   EXPORT_SYMBOL_GPL(rust_helper_krealloc);
> >
> > +void rust_helper_preempt_enable_notrace(void)
> > +{
> > +     preempt_enable_notrace();
> > +}
> > +EXPORT_SYMBOL_GPL(rust_helper_preempt_enable_notrace);
> > +
> > +void rust_helper_preempt_disable_notrace(void)
> > +{
> > +     preempt_disable_notrace();
> > +}
> > +EXPORT_SYMBOL_GPL(rust_helper_preempt_disable_notrace);
> > +
> > +bool rust_helper_current_cpu_online(void)
> > +{
> > +     return cpu_online(raw_smp_processor_id());
> > +}
> > +EXPORT_SYMBOL_GPL(rust_helper_current_cpu_online);
> > +
> > +void *rust_helper___rcu_dereference_raw(void **p)
> > +{
> > +     return rcu_dereference_raw(p);
> > +}
> > +EXPORT_SYMBOL_GPL(rust_helper___rcu_dereference_raw);
>
> Ouch. Doing a function call for each of those small operations will
> have a rather large performance impact when tracing is active. If it is
> not possible to inline those in Rust, then implementing __DO_TRACE in
> a C function would at least allow Rust to only do a single call to C
> rather than go back and forth between Rust and C.
>
> What prevents inlining those helpers in Rust ?

There's nothing fundamental that prevents it. But they contain a large
amount of architecture specific code, which makes them a significant
amount of work to reimplement in Rust.

For example, rcu_dereference_raw calls into READ_ONCE. READ_ONCE is
usually a volatile load, but under arm64+LTO, you get a bunch of
inline assembly that relies on ALTERNATIVE for feature detection:
https://elixir.bootlin.com/linux/v6.9/source/arch/arm64/include/asm/rwonce.h#L36

And preempt_enable_notrace has a similar story.

The solution that Boqun mentions is nice, but it relies on rustc and
clang using the same version of LLVM. You are unlikely to have
compilers with matching LLVMs unless you intentionally take steps to
make that happen.

But yes, perhaps these helpers are an argument to have a single call
for __DO_TRACE instead.

Alice