may_use_simd on aarch64, chacha20

[Ard Biesheuvel]

On 26 May 2017 at 15:28, Dave Martin <Dave.Martin at arm.com> wrote:
> On Sun, May 21, 2017 at 10:55:20PM +0200, Ard Biesheuvel wrote:
>> (+ Dave)
>
> Apologies for the slow reply -- hopefully this is still useful.
>
>> > On 21 May 2017, at 19:02, Jason A. Donenfeld <Jason at zx2c4.com> wrote:
>> >
>> > Hi folks,
>> >
>> > I noticed that the ARM implementation [1] of chacha20 makes a check to
>> > may_use_simd(), but the ARM64 implementation [2] does not. Question 1:
>> > is this a bug, in which case I'll submit a patch shortly, or is this
>> > intentional? In case of the latter, could somebody explain the
>> > reasoning?
>>
>> This is intentional. arm64 supports kernel mode NEON in any context,
>> whereas ARM only supports it in process context. This is due to the way
>> lazy FP restore is implemented on ARM. However, we are about to change
>> this on arm64 to only allow non-nested kernel mode NEON, similar to
>> x86. This is necessary to support SVE.
>>
>> > On a similar note, the only ARM64 glue code that uses
>> > may_use_simd() is sha256; everything else does not. Shall I submit a
>> > substantial patch series to fix this up everywhere?
>> >
>>
>> Currently, may_use_simd() is only used as a hint on arm64 whether it
>> makes sense to offload crypto to process context. In the sha256 code,
>> whose arm64 neon implementation is only marginally faster than scalar
>> on some micro-architectures, it is used to prefer the scalar code in
>> interrupt context, because the NEON code preserves/restores the NEON
>> state of the interrupted context eagerly, which is costly.
>>
>> > Secondly, I noticed that may_use_simd() is essentially aliased to
>> > !in_interrupt(), since it uses the asm-generic variety. Question 2:
>> > Isn't this overkill? Couldn't we make an arm/arm64 variant of this
>> > that only checks in_irq()?
>> >
>>
>> No. ARM does not support kernel mode NEON in softirq context, and
>> arm64 will soon have its own override that only allows non-nested use
>> in softirq context.
>>
>> > Lastly, APIs like pcrypts and padata execute with bottom halves
>> > disabled, even though their actual execution environment is process
>> > context, via a workqueue. Thus, here, in_interrupt() will always be
>> > true, even though this is likely a place where we want to use simd.
>> > Question 3: is there something better that could be done?
>>
>> I guess we should switch to in_serving_softirq() instead.
>
> For context, the arm64 kernel-mode NEON refactoring I've been working
> on [1] defines may_use_simd() to indicate whether calling
> kernel_neon_begin() is safe or not, rather then being a hint about
> whether it is desirable to do so.
>

I should mention again, perhaps redundantly, that may_use_simd() was
not intended as a hint. It only became this way after we removed the
limitation on arm64 that SIMD may only be used in process context.

> This changes the definition of may_use_simd() to something that is
> probably appropriate for your scenario.  From process context without
> any outer kernel_neon_begin()...kernel_neon_end(), it should return
> true.
>
> In general, callers in any context should do something like:
>
>         if (may_use_simd()) {
>                 kernel_neon_begin();
>                 /* SIMD implementation */
>                 kernel_neon_end();
>         } else {
>                 /* fallback implementation, or defer to another context */
>         }
>
> Even from task context where may_use_simd() should always return true
> and thus where the fallback path can be omitted, it may be a good idea
> to do something like
>
>         if (!may_use_simd()) {
>                 WARN_ON(1);
>                 return -EBUSY;
>         }
>
> ... or similar.
>

I don't think it is generally a good idea to make inferences about
whether may_use_simd() is going to return true or false in code that
is not tightly coupled to the FP/SIMD arch code itself, although I do
understand that there may be cases where it is cumbersome to provide a
fallback, especially if you know it is never going to be used.