[RFC PATCH 4/5] arm64: fpsimd: run kernel mode NEON with softirqs disabled

[Ard Biesheuvel]

On Tue, 19 Jan 2021 at 17:01, Dave Martin <Dave.Martin at arm.com> wrote:
>
> On Fri, Dec 18, 2020 at 06:01:05PM +0100, Ard Biesheuvel wrote:
> > Kernel mode NEON can be used in task or softirq context, but only in
> > a non-nesting manner, i.e., softirq context is only permitted if the
> > interrupt was not taken at a point where the kernel was using the NEON
> > in task context.
> >
> > This means all users of kernel mode NEON have to be aware of this
> > limitation, and either need to provide scalar fallbacks that may be much
> > slower (up to 20x for AES instructions) and potentially less safe, or
> > use an asynchronous interface that defers processing to a later time
> > when the NEON is guaranteed to be available.
> >
> > Given that grabbing and releasing the NEON is cheap, we can relax this
> > restriction, by increasing the granularity of kernel mode NEON code, and
> > always disabling softirq processing while the NEON is being used in task
> > context.
> >
> > Signed-off-by: Ard Biesheuvel <ardb at kernel.org>
>
> Sorry for the slow reply on this...  it looks reasonable, but I have a
> few comments below.
>

No worries - thanks for taking a look.

> > ---
> >  arch/arm64/include/asm/assembler.h | 19 +++++++++++++------
> >  arch/arm64/kernel/asm-offsets.c    |  2 ++
> >  arch/arm64/kernel/fpsimd.c         |  4 ++--
> >  3 files changed, 17 insertions(+), 8 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/assembler.h b/arch/arm64/include/asm/assembler.h
> > index ddbe6bf00e33..74ce46ed55ac 100644
> > --- a/arch/arm64/include/asm/assembler.h
> > +++ b/arch/arm64/include/asm/assembler.h
> > @@ -15,6 +15,7 @@
> >  #include <asm-generic/export.h>
> >
> >  #include <asm/asm-offsets.h>
> > +#include <asm/alternative.h>
> >  #include <asm/cpufeature.h>
> >  #include <asm/cputype.h>
> >  #include <asm/debug-monitors.h>
> > @@ -717,17 +718,23 @@ USER(\label, ic ivau, \tmp2)                    // invalidate I line PoU
> >       .endm
> >
> >       .macro          if_will_cond_yield_neon
> > -#ifdef CONFIG_PREEMPTION
> >       get_current_task        x0
> >       ldr             x0, [x0, #TSK_TI_PREEMPT]
> > -     sub             x0, x0, #PREEMPT_DISABLE_OFFSET
> > -     cbz             x0, .Lyield_\@
> > +#ifdef CONFIG_PREEMPTION
> > +     cmp             x0, #PREEMPT_DISABLE_OFFSET
> > +     beq             .Lyield_\@      // yield on need_resched in task context
> > +#endif
> > +     /* never yield while serving a softirq */
> > +     tbnz            x0, #SOFTIRQ_SHIFT, .Lnoyield_\@
>
> Can you explain the rationale here?
>
> Using if_will_cond_yield_neon suggests the algo thinks it may run for
> too long the stall preemption until completion, but we happily stall
> preemption _and_ softirqs here.
>
> Is it actually a bug to use the NEON conditional yield helpers in
> softirq context?
>

No, it is not. But calling kernel_neon_end() from softirq context will
not cause it to finish any faster, so there is really no point in
doing so.

> Ideally, if processing in softirq context takes an unreasonable about of
> time, the work would be handed off to an asynchronous worker, but that
> does seem to conflict rather with the purpose of this series...
>

Agreed, but this is not something we can police at this level. If the
caller does an unreasonable amount of work from a softirq, no amount
of yielding is going to make a difference.

> > +
> > +     adr_l           x0, irq_stat + IRQ_CPUSTAT_SOFTIRQ_PENDING
> > +     this_cpu_offset x1
> > +     ldr             w0, [x0, x1]
> > +     cbnz            w0, .Lyield_\@  // yield on pending softirq in task context
> > +.Lnoyield_\@:
> >       /* fall through to endif_yield_neon */
> >       .subsection     1
> >  .Lyield_\@ :
> > -#else
> > -     .section        ".discard.cond_yield_neon", "ax"
> > -#endif
> >       .endm
> >
> >       .macro          do_cond_yield_neon
> > diff --git a/arch/arm64/kernel/asm-offsets.c b/arch/arm64/kernel/asm-offsets.c
> > index 7d32fc959b1a..34ef70877de4 100644
> > --- a/arch/arm64/kernel/asm-offsets.c
> > +++ b/arch/arm64/kernel/asm-offsets.c
> > @@ -93,6 +93,8 @@ int main(void)
> >    DEFINE(DMA_FROM_DEVICE,    DMA_FROM_DEVICE);
> >    BLANK();
> >    DEFINE(PREEMPT_DISABLE_OFFSET, PREEMPT_DISABLE_OFFSET);
> > +  DEFINE(SOFTIRQ_SHIFT, SOFTIRQ_SHIFT);
> > +  DEFINE(IRQ_CPUSTAT_SOFTIRQ_PENDING, offsetof(irq_cpustat_t, __softirq_pending));
> >    BLANK();
> >    DEFINE(CPU_BOOT_STACK,     offsetof(struct secondary_data, stack));
> >    DEFINE(CPU_BOOT_TASK,              offsetof(struct secondary_data, task));
> > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> > index 062b21f30f94..823e3a8a8871 100644
> > --- a/arch/arm64/kernel/fpsimd.c
> > +++ b/arch/arm64/kernel/fpsimd.c
> > @@ -180,7 +180,7 @@ static void __get_cpu_fpsimd_context(void)
> >   */
> >  static void get_cpu_fpsimd_context(void)
> >  {
> > -     preempt_disable();
> > +     local_bh_disable();
> >       __get_cpu_fpsimd_context();
> >  }
> >
> > @@ -201,7 +201,7 @@ static void __put_cpu_fpsimd_context(void)
> >  static void put_cpu_fpsimd_context(void)
> >  {
> >       __put_cpu_fpsimd_context();
> > -     preempt_enable();
> > +     local_bh_enable();
> >  }
> >
> >  static bool have_cpu_fpsimd_context(void)
>
> I was concerned about catching all the relevant preempt_disable()s, but
> it had slipped my memory that Julien had factored these into one place.
>
> I can't see off the top of my head any reason why this shouldn't work.
>

Thanks.

>
> In threory, switching to local_bh_enable() here will add a check for
> pending softirqs onto context handling fast paths.  I haven't dug into
> how that works, so perhaps this is trivial on top of the preemption
> check in preempt_enable().  Do you see any difference in hackbench or
> similar benchmarks?
>

I haven't tried, tbh. But by context handling fast paths, you mean
managing the FP/SIMD state at context switch time, right? Checking for
pending softirqs amounts to a single per-CPU load plus compare, so
that should be negligible AFAICT. Obviously, actually handling the
softirq may take additional time, but that penalty has to be taken
somewhere - I don't see how that would create extra work that we
wouldn't have to do otherwise.

I'll do some experiments with hackbench once I get back to this series.