[PATCH] asm-generic: simd: allow SIMD in process context with BH disabled

[Ard Biesheuvel]

On 31 May 2017 at 15:32, Dave Martin <Dave.Martin at arm.com> wrote:
> On Wed, May 31, 2017 at 12:57:01PM +0000, Ard Biesheuvel wrote:
>> asm-generic supplies a header asm/simd.h which exports a single function
>> may_use_simd(), which conveys whether the current context allows the SIMD
>> register file or instructions to be used.
>>
>> This header is included by crypto code shared between x86 and ARM/arm64,
>> and which offloads SIMD processing to process context if required. The
>> generic asm/simd.h is shared between ARM and arm64 at the moment, while
>> x86 has its own implementation.
>>
>> On arm64, we currently mostly ignore may_use_simd(), because arm64 allows
>> kernel mode NEON in any context. However, this is due to change shortly
>> when support for SVE is merged, at which point we will introduce an arm64
>> specific implementation of asm/simd.h as well.
>>
>> That leaves ARM, which only allows kernel mode NEON in process context,
>> which makes the current generic implementation of may_use_simd() seem
>> appropriate. However, given that in_interrupt() will return true when
>> running in process context with bottom halves disabled, we may end up
>> falling back to less optimized code unnecessarily, given that kernel
>> mode NEON is perfectly usable in that case.
>>
>> So redefine may_use_simd() to disallow SIMD only when running in hardirq
>> or softirq context.
>>
>> While we're at it, add some missing header file decorations such as
>> a license header and include guards.
>>
>> Reported-by: "Jason A. Donenfeld" <Jason at zx2c4.com>
>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>> ---
>>  include/asm-generic/simd.h | 23 ++++++++++++++++++++---
>>  1 file changed, 20 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/asm-generic/simd.h b/include/asm-generic/simd.h
>> index f57eb7b5c23b..a3e5ebe6b2b2 100644
>> --- a/include/asm-generic/simd.h
>> +++ b/include/asm-generic/simd.h
>> @@ -1,14 +1,31 @@
>> +/*
>> + * Copyright (C) 2013 - 2017 Linaro Ltd. <ard.biesheuvel at linaro.org>
>> + *
>> + * This program is free software; you can redistribute it and/or modify it
>> + * under the terms of the GNU General Public License version 2 as published
>> + * by the Free Software Foundation.
>> + */
>>
>> -#include <linux/hardirq.h>
>> +#ifndef __ASM_SIMD_H
>> +#define __ASM_SIMD_H
>> +
>> +#include <linux/types.h>
>> +#include <linux/preempt.h>
>>
>>  /*
>>   * may_use_simd - whether it is allowable at this time to issue SIMD
>>   *                instructions or access the SIMD register file
>>   *
>>   * As architectures typically don't preserve the SIMD register file when
>> - * taking an interrupt, !in_interrupt() should be a reasonable default.
>> + * taking an interrupt, it is reasonable to define the default behavior
>> + * of 'may_use_simd()' to be 'SIMD is only allowed when not handling an
>> + * IRQ or softIRQ'. Since 'in_interrupt()' will also return true when
>> + * running in process context with bottom halves disabled, we have to
>> + * spell out that condition as shown.
>
> Minor nit: do we need the comment about in_interrupt() here?
>
> It makes more sense to explain the change in the commit message (which
> you do) than to explain in-line the behaviour of a function that the
> code doesn't use.
>
> <linux/preempt.h> already hints at the caveats of in_interrupt().
>

Fair enough. I tend to err on the verbose side when it comes to
comments, but this could indeed be omitted.

>
> For this comment block, it may be more helpful to note that SIMD is
> permitted in task context even if bottom halves are enabled.
>
>>   */
>>  static __must_check inline bool may_use_simd(void)
>>  {
>> -     return !in_interrupt();
>> +     return !in_irq() && !in_serving_softirq();
>
> Previously, in_nmi() implied !may_use_simd().
>
> Now, may_use_simd() can return true if in_nmi().
>
> Code in NMI context probably shouldn't be touching this interface at
> all, but we may want to close this hole by adding && !in_nmi()
> explicitly.  I did that in my kernel-mode-neon simplification series,
> but couldn't decide whether it was superfluous.
>
> Any thoughts?
>

I agree. I will add that as well.