[PATCH v5] arm64: fpsimd: improve stacking logic in non-interruptible context

[Ard Biesheuvel]

On 13 December 2016 at 11:11, Dave Martin <Dave.Martin at arm.com> wrote:
> On Mon, Dec 12, 2016 at 05:56:27PM +0000, Ard Biesheuvel wrote:
>> Currently, we allow kernel mode NEON in softirq or hardirq context by
>> stacking and unstacking a slice of the NEON register file for each call
>> to kernel_neon_begin() and kernel_neon_end(), respectively.
>>
>> Given that
>> a) a CPU typically spends most of its time in userland, during which time
>>    no kernel mode NEON in process context is in progress,
>> b) a CPU spends most of its time in the kernel doing other things than
>>    kernel mode NEON when it gets interrupted to perform kernel mode NEON
>>    in softirq context
>>
>> the stacking and subsequent unstacking is only necessary if we are
>> interrupting a thread while it is performing kernel mode NEON in process
>> context, which means that in all other cases, we can simply preserve the
>> userland FPSIMD state once, and only restore it upon return to userland,
>> even if we are being invoked from softirq or hardirq context.
>>
>> So instead of checking whether we are running in interrupt context, keep
>> track of the level of nested kernel mode NEON calls in progress, and only
>> perform the eager stack/unstack if the level exceeds 1.
>>
>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>> ---
>>  arch/arm64/kernel/fpsimd.c | 64 +++++++++++++++++++++++++++++++++-------------
>>  1 file changed, 46 insertions(+), 18 deletions(-)
>>
>> diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
>> index 394c61db5566..c19363775436 100644
>> --- a/arch/arm64/kernel/fpsimd.c
>> +++ b/arch/arm64/kernel/fpsimd.c
>> @@ -220,45 +220,73 @@ void fpsimd_flush_task_state(struct task_struct *t)
>>
>>  #ifdef CONFIG_KERNEL_MODE_NEON
>>
>> -static DEFINE_PER_CPU(struct fpsimd_partial_state, hardirq_fpsimdstate);
>> -static DEFINE_PER_CPU(struct fpsimd_partial_state, softirq_fpsimdstate);
>> +/*
>> + * Although unlikely, it is possible for three kernel mode NEON contexts to
>> + * be live at the same time: process context, softirq context and hardirq
>> + * context. So while the userland context is stashed in the thread's fpsimd
>> + * state structure, we need two additional levels of storage.
>> + */
>> +static DEFINE_PER_CPU(struct fpsimd_partial_state, nested_fpsimdstate[2]);
>> +static DEFINE_PER_CPU(int, kernel_neon_nesting_level);
>>
>>  /*
>>   * Kernel-side NEON support functions
>>   */
>>  void kernel_neon_begin_partial(u32 num_regs)
>>  {
>> -     if (in_interrupt()) {
>> -             struct fpsimd_partial_state *s = this_cpu_ptr(
>> -                     in_irq() ? &hardirq_fpsimdstate : &softirq_fpsimdstate);
>> +     struct fpsimd_partial_state *s;
>> +     int level;
>>
>> -             BUG_ON(num_regs > 32);
>> -             fpsimd_save_partial_state(s, roundup(num_regs, 2));
>> -     } else {
>> +     preempt_disable();
>> +
>> +     if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
>>               /*
>>                * Save the userland FPSIMD state if we have one and if we
>>                * haven't done so already. Clear fpsimd_last_state to indicate
>>                * that there is no longer userland FPSIMD state in the
>>                * registers.
>>                */
>> -             preempt_disable();
>> -             if (current->mm &&
>> -                 !test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
>> -                     fpsimd_save_state(&current->thread.fpsimd_state);
>> +             if (current->mm) {
>> +                     unsigned long flags;
>> +
>> +                     local_irq_save(flags);
>
> For non-SVE hardware (i.e., all hardware for now), this raises mean IRQ
> latency if KERNEL_MODE_NEON is used, to fix a bug that doesn't exist.
>
> For SVE hardware, we would be disabling interrupts around typically a
> few KB of stores.  I don't know what the actual hardware numbers would
> look like, but this feels like a disproportionate cost to the problem
> being solved...
>

That is a fair point. We'd be better off doing a spin_trylock)_, and
whoever gets the lock first can perform the fpsimd_save_state() until
completion.

> After all, why do this here...
>
>> +                     if (!test_and_set_thread_flag(TIF_FOREIGN_FPSTATE))
>> +                             fpsimd_save_state(&current->thread.fpsimd_state);
>> +                     local_irq_restore(flags);
>> +             } else {
>> +                     set_thread_flag(TIF_FOREIGN_FPSTATE);
>> +             }
>>               this_cpu_write(fpsimd_last_state, NULL);
>>       }
>> +
>> +     level = this_cpu_inc_return(kernel_neon_nesting_level);
>
> ...and yet go to all this trouble to avoid disabling IRQs for a _nested_
> kernel_neon_begin()?
>

I am not sure I follow here: do you mean the this_cpu_inc_return would
be simpler if we would en-/disable interrupts?

> Did I miss something, or does increasing the count around the outermost
> case too without IRQ disabling not work, in the non-SVE case?
>

I think it does work for the non-SVE case, since it will restore the
FP/SIMD state after returning.