Re: [PATCH v7 4/5] rust: add find_bit_benchmark_rust module.

[Boqun Feng <boqun.feng@gmail.com>]

On Thu, Apr 24, 2025 at 09:48:17AM -0700, Boqun Feng wrote:
> On Thu, Apr 24, 2025 at 06:45:33PM +0200, Burak Emir wrote:
> > On Wed, Apr 23, 2025 at 6:56 PM Yury Norov <yury.norov@gmail.com> wrote:
> > > So? Can you show your numbers?
> > 
> > For now, I only have numbers that may not be very interesting:
> > 
> > - for find_next_bit,  find_next_zero_bit and find_next_zero_bit (sparse):
> >   22 ns/iteration in C, 32 ns/iteration in Rust.
> > 
> > - for sparse find_next_bit (sparse):
> >   60 ns/iteration in C, 70 ns/iteration in Rust.
> > 
> > This is a VM running nested in a VM. More importantly: the C helper
> > method is not inlined.
> > So we are likely measuring the overhead (plus the extra bounds checking).
> > 
> > I would like to get cross-language inlining to work with thinLTO to
> > have a more realistic comparison.
> > However, that is not something that works out of the box.
> > I am looking at Gary Guo's patch for this:
> > https://lore.kernel.org/all/20250319205141.3528424-1-gary@garyguo.net/
> > Currently, I get duplicate symbol errors.
> > 
> 
> You will need to add __rust_helper attribute for the new rust helpers
> introduce in your patches. See Gary's patch #2 for example.
> 

Here you go ;-)

	https://github.com/fbq/linux/tree/rust-inline-bitmap

I rebased on the top of rust-next and applied your patches onto it. The
last one is the necessary bits that enables helper inlining for the new
APIs in your patch. There is also a "TMP" patch in-between, otherwise
INLINE_HELPERS won't work, we need to dig more of it. But anyway, it
works on my machine (TM):

(on x86, in your test function)

000000000028b090 <_RNvNtNtCs3KHxpmQFgFb_6kernel6bitmap5tests40kunit_rust_wrapper_bitmap_set_clear_find>:
  ...
  28b0dd: 48 0f ba 2b 11                btsq    $0x11, (%rbx)

^ this is the "b.set_bit(17);" in bitmap_set_clear_find() test.

Regards,
Boqun
	

> Regards,
> Boqun
> 
> > > Can you print the existing C test output back to back with the new one?
> > > Can you also ask 0-day folks to enable your test in their rust config?
> > 
> > Will look into these. Rong (hi!) is working on LTO for kernel and will
> > know a lot more than me how Rust will fit in eventually.
> > IMHO, making cross-language inlining work out of the box will be a
> > necessary baseline to get Rust performance for hot code.
> > 
> > > > We add a fill_random() method protected by the config in order to
> > > > maintain the abstraction.
> > > >
> > > > Minor fix to the documentation of the corresponding C config
> > > > FIND_BIT_BENCHMARK, it was mentioning the wrong module name.
> > >
> > > Indeed. Can you make it a separate patch, please?
> > 
> > Will do.