Re: [PATCH v3] kasan: speed up mte_set_mem_tag_range

[Evgenii Stepanov]

On Wed, May 19, 2021 at 11:13 AM Catalin Marinas
<catalin.marinas@xxxxxxx> wrote:
>
> On Tue, May 18, 2021 at 11:11:52AM -0700, Peter Collingbourne wrote:
> > On Tue, May 18, 2021 at 10:44 AM Catalin Marinas
> > <catalin.marinas@xxxxxxx> wrote:
> > > If we want to get the best performance out of this, we should look at
> > > the memset implementation and do something similar. In principle it's
> > > not that far from a memzero, though depending on the microarchitecture
> > > it may behave slightly differently.
> >
> > For Scudo I compared our storeTags implementation linked above against
> > __mtag_tag_zero_region from the arm-optimized-routines repository
> > (which I think is basically an improved version of that memset
> > implementation rewritten to use STG and DC GZVA), and our
> > implementation performed better on the hardware that we have access
> > to.
>
> That's the advantage of having hardware early ;).
>
> > > Anyway, before that I wonder if we wrote all this in C + inline asm
> > > (three while loops or maybe two and some goto), what's the performance
> > > difference? It has the advantage of being easier to maintain even if we
> > > used some C macros to generate gva/gzva variants.
> >
> > I'm not sure I agree that it will be easier to maintain. Due to the
> > number of "unusual" instructions required here it seems more readable
> > to have the code in pure assembly than to require readers to switch
> > contexts between C and asm. If we did move it to inline asm then I
> > think it should basically be a large blob of asm like the Scudo code
> > that I linked.
>
> I was definitely not thinking of a big asm block, that's even less
> readable than separate .S file. It's more like adding dedicated macros
> for single STG or DC GVA uses and using them in while loops.

I've got a C version with 4 single-instruction asm blocks, and it
looks pretty nice. The assembly is almost identical to the hand
written variant, and performance is 3% better, presumably because of
the inlining. Also, the C version allows more potential optimizations,
like specialization on the value of "init" - which is not happening
right now because it is not constant in any of the callers.

I'll upload a v4 shortly.

>
> Anyway, let's see a better commented .S implementation first. Given that
> tagging is very sensitive to the performance of this function, we'd
> probably benefit from a (few percent I suspect) perf improvement with
> the hand-coded assembly.
>
> --
> Catalin