Re: [PATCH 0/2] arm64: Only call into preempt_schedule() if need_resched()

[Will Deacon <will.deacon@arm.com>]

On Wed, Nov 28, 2018 at 10:01:46AM +0100, Peter Zijlstra wrote:
> On Wed, Nov 28, 2018 at 09:56:40AM +0100, Peter Zijlstra wrote:
> > On Tue, Nov 27, 2018 at 07:45:00PM +0000, Will Deacon wrote:
> > > This pair of patches improves our preempt_enable() implementation slightly
> > > on arm64 by making the resulting call to preempt_schedule() conditional
> > > on need_resched(), which is tracked in bit 32 of the preempt count. The
> > > logic is inverted so that we can detect the preempt count going to zero
> > > and need_resched being set with a single CBZ instruction.
> > 
> > >   40:   a9bf7bfd        stp     x29, x30, [sp, #-16]!
> > >   44:   910003fd        mov     x29, sp
> > >   48:   d5384101        mrs     x1, sp_el0
> > >   4c:   f9400820        ldr     x0, [x1, #16]
> > 
> > We load x0 which is a u64, right?
> > 
> > >   50:   d1000400        sub     x0, x0, #0x1
> > >   54:   b9001020        str     w0, [x1, #16]
> > 
> > But we store w0, which is the low u32, such as to not touch the high
> > word which contains the preempt bit.
> > 
> > >   58:   b4000060        cbz     x0, 64 <will+0x24>
> > >   5c:   a8c17bfd        ldp     x29, x30, [sp], #16
> > >   60:   d65f03c0        ret
> > >   64:   94000000        bl      0 <preempt_schedule>
> > >   68:   a8c17bfd        ldp     x29, x30, [sp], #16
> > >   6c:   d65f03c0        ret
> > 
> > Why not?
> > 
> >    58:   b4000060        cbnz    x0, 60 <will+0x24>
> >    5c:   94000000        bl      0 <preempt_schedule>
> >    60:   a8c17bfd        ldp     x29, x30, [sp], #16
> >    64:   d65f03c0        ret
> > 
> > which seems shorter.
> > 
> > 
> > So it's still early, and I haven't finished (or really even started) my
> > pot 'o tea, but what about:
> > 
> > 
> > 	ldr x0, [x1, #16]	// seees the high bit set -- no preempt needed
> > 	sub x0, x0, #1
> > 
> > 	<interrupt>
> > 	  ...
> > 	  resched_curr()
> > 	    set_tsk_need_resched();
> > 	    set_preempt_need_resched();
> > 	</interrupt> // sees preempt_count != 0, does not preempt
> > 
> > 	str w0, [x1, #16] // stores preempt_count == 0
> > 	cbnz x0, 1f // taken, we still observe the high word from before
> > 
> > 1:	ret
> > 
> > 
> > Which then ends with preempt_count==0, need_resched==0 and no actual
> > preemption afaict.
> > 
> > Can you use mis-matched ll x0 / sc w0 to do this same thing and detector
> > the intermediate write on the high word?
> 
> That is, something along these here lines:
> 
> 1:	ldxr x0, [x1, #16]
> 	sub  x0, x0, #1
> 	stxr w1, w0, [x1, #16]

^^ This guy needs a different encoding but, to be honest, I reckon we're
better off just reloading the need_resched flag in the case where the count
has hit zero. I'll have a play. The assembly I posted is all generated by
GCC, so I can't comment on why it didn't chose your shorter sequence :)

Will