Re: [PATCH 0/5] Rust abstractions for network device drivers

From: "Andreas Hindborg (Samsung)" <nmi@metaspace.dk>
To: Miguel Ojeda <miguel.ojeda.sandonis@gmail.com>
Cc: "Emilio Cobos Álvarez" <emilio@crisal.io>,
	"Andrew Lunn" <andrew@lunn.ch>,
	"FUJITA Tomonori" <fujita.tomonori@gmail.com>,
	kuba@kernel.org, netdev@vger.kernel.org,
	rust-for-linux@vger.kernel.org, aliceryhl@google.com
Subject: Re: [PATCH 0/5] Rust abstractions for network device drivers
Date: Tue, 20 Jun 2023 21:12:49 +0200	[thread overview]
Message-ID: <877cryx74x.fsf@metaspace.dk> (raw)
In-Reply-To: <CANiq72=x9kEniX78vA7fLu+6wiwDKEr=BYy+aCMZ5S+eSRFf+A@mail.gmail.com>

Hi All,

Miguel Ojeda <miguel.ojeda.sandonis@gmail.com> writes:

> On Mon, Jun 19, 2023 at 1:27 PM Emilio Cobos Álvarez <emilio@crisal.io> wrote:
>>
>> Hi Andrew, Miguel,
>>
>> On 6/16/23 16:43, Andrew Lunn wrote:
>> > I said in another email, i don't want to suggest premature
>> > optimisation, before profiling is done. But in C, these functions are
>> > inline for a reason. We don't want the cost of a subroutine call. We
>> > want the compiler to be able to inline the code, and the optimiser to
>> > be able to see it and generate the best code it can.
>> >
>> > Can the rust compile inline the binding including the FFI call?
>>
>> This is possible, with cross-language LTO, see:
>>
>>    https://blog.llvm.org/2019/09/closing-gap-cross-language-lto-between.html
>>
>> There are some requirements that need to happen for that to work
>> (mainly, I believe, that the LLVM version used by rustc and clang agree).
>>
>> But in general it is possible. We use it extensively on Firefox. Of
>> course the requirements of Firefox and the kernel might be different.
>>
>> I think we rely heavily on PGO instrumentation to make the linker inline
>> ffi functions, but there might be other ways of forcing the linker to
>> inline particular calls that bindgen could generate or what not.
>
> Thanks Emilio! It is nice to hear cross-language LTO is working well
> for Firefox.
>
> Andreas took a look at cross-language LTO some weeks ago, if I
> remember correctly (Cc'd).
>
> I am not sure about the latest status on kernel PGO, though.

I hacked it to work a while back for the NVMe and null_blk drivers. You
need to build the C and Rust parts of the kernel with compatible
clang/rustc compilers (same major version I believe), and then pass the
right compiler flags to get rustc and clang to emit llvm bitcode instead
of ELF files with machine code. As far as I recall, some infrastructure
is present in kbuild, but I had to add some bits to make it build.

Also, I had some issues with LLVM not doing the inline properly and I
had to use `llvm-link` on the bitcode of my module to link in the
bitcode of the C code I wanted to inline. Without that step, inlining
did not happen.

I was able to build and execute in qemu like this, but I was not able to
boot on bare metal. I think the LTO breaks something in C land. I did
not investigate that further.

Eventually I paused the work because I did not observe conclusive
speedups in my benchmarks. I plan to resume the work at some point and
do more rigorous benchmarking to see if I can observe a statistically
significant speedup. Not sure when that will happen though.

Best regards,
Andreas