Re: [PATCH RT v2] crypto: limit more FPU-enabled sections

[Sebastian Andrzej Siewior]

On 2017-12-01 12:32:35 [+0100], Peter Zijlstra wrote:
> On Fri, Dec 01, 2017 at 11:44:22AM +0100, Sebastian Andrzej Siewior wrote:
â
> > +static void camellia_fpu_sched_rt(struct crypt_priv *ctx)
> > +{
> > +       bool fpu_enabled = ctx->fpu_enabled;
> > +
> > +       if (!fpu_enabled || !tif_need_resched_now())
> > +               return;
> > +       camellia_fpu_end(fpu_enabled);
> > +       kernel_fpu_end();
> > +       /* schedule due to preemptible */
> > +       kernel_fpu_begin();
> > +}
> 
> There's a ton of duplication in there; you're not nearly lazy enough.
> 
> Why can't we do something simple like kernel_fpu_resched() ?
> 
> diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
> index f92a6593de1e..05321b98a55a 100644
> --- a/arch/x86/kernel/fpu/core.c
> +++ b/arch/x86/kernel/fpu/core.c
> @@ -130,6 +130,18 @@ void kernel_fpu_begin(void)
>  }
>  EXPORT_SYMBOL_GPL(kernel_fpu_begin);
>  
> +void kernel_fpu_resched(void)
> +{
> +	WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
> +
> +	if (should_resched(PREEMPT_OFFSET)) {
> +		kernel_fpu_end();
> +		cond_resched();
> +		kernel_fpu_begin();

I can do that but I would still keep it RT only to avoid the
kernel_fpu_begin/end to be invoked more often on !RT.
But why that cond_resched()? kernel_fpu_end() ends with preempt_enable()
and this one should do the trick.

> +	}
> +}
> +EXPORT_SYMBOL_GPL(kernel_fpu_resched);
> +
>  void kernel_fpu_end(void)
>  {
>  	__kernel_fpu_end();

Sebastian