Re: [PATCH 5/6] sched/preempt: add PREEMPT_DYNAMIC using static keys

[Mark Rutland]

On Thu, Feb 03, 2022 at 12:34:53PM +0100, Frederic Weisbecker wrote:
> On Thu, Feb 03, 2022 at 09:51:46AM +0000, Mark Rutland wrote:
> > On Thu, Feb 03, 2022 at 12:21:45AM +0100, Frederic Weisbecker wrote:
> > > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > > index 78c351e35fec..7710b6593c72 100644
> > > > --- a/include/linux/sched.h
> > > > +++ b/include/linux/sched.h
> > > > @@ -2008,7 +2008,7 @@ static inline int test_tsk_need_resched(struct task_struct *tsk)
> > > >  #if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
> > > >  extern int __cond_resched(void);
> > > >  
> > > > -#ifdef CONFIG_PREEMPT_DYNAMIC
> > > > +#if defined(CONFIG_PREEMPT_DYNAMIC) && defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
> > > >  
> > > >  DECLARE_STATIC_CALL(cond_resched, __cond_resched);
> > > >  
> > > > @@ -2017,6 +2017,14 @@ static __always_inline int _cond_resched(void)
> > > >  	return static_call_mod(cond_resched)();
> > > >  }
> > > >  
> > > > +#elif defined(CONFIG_PREEMPT_DYNAMIC) && defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
> > > > +extern int dynamic_cond_resched(void);
> > > > +
> > > > +static __always_inline int _cond_resched(void)
> > > > +{
> > > > +	return dynamic_cond_resched();
> > > 
> > > So in the end this is creating an indirect call for every preemption entrypoint.
> > 
> > Huh? "indirect call" usually means a branch to a function pointer, and I don't
> > think that's what you mean here. Do you just mean that we add a (direct)
> > call+return?
> 
> Right, basic terminology and me...

No problem; just wanted to make sure we were talking about the same thing! :)

> > This gets inlined, and will be just a direct call to dynamic_cond_resched().
> > e,g. on arm64 this will be a single instruction:
> > 
> > 	bl	dynamic_cond_resched
> > 
> > ... and (as the commit message desribes) then the implementation of
> > dynamic_cond_resched will be the same as the regular __cond_resched *but* the
> > static key trampoline is inlined at the start, e.g.
> > 
> > | <dynamic_cond_resched>:
> > |        bti     c
> > |        b       <dynamic_cond_resched+0x10>
> > |        mov     w0, #0x0                        // #0
> > |        ret
> > |        mrs     x0, sp_el0
> > |        ldr     x0, [x0, #8]
> > |        cbnz    x0, <dynamic_cond_resched+0x8>
> > |        paciasp
> > |        stp     x29, x30, [sp, #-16]!
> > |        mov     x29, sp
> > |        bl      <preempt_schedule_common>
> > |        mov     w0, #0x1                        // #1
> > |        ldp     x29, x30, [sp], #16
> > |        autiasp
> > |        ret
> > 
> > ... compared to the regular form of the function:
> > 
> > | <__cond_resched>:
> > |        bti     c
> > |        mrs     x0, sp_el0
> > |        ldr     x1, [x0, #8]
> > |        cbz     x1, <__cond_resched+0x18>
> > |        mov     w0, #0x0                        // #0
> > |        ret
> > |        paciasp
> > |        stp     x29, x30, [sp, #-16]!
> > |        mov     x29, sp
> > |        bl      <preempt_schedule_common>
> > |        mov     w0, #0x1                        // #1
> > |        ldp     x29, x30, [sp], #16
> > |        autiasp
> > |        ret
> 
> Who reads changelogs anyway? ;-)
> 
> Ok I didn't know about that. Is this a guaranteed behaviour everywhere?

For anyone with static keys based on jump labels it should look roughly as
above. The *precise* codegen will depend on a bunch of details, but the whole
point of jump labels and static keys is to permit codegen like this.

> Perhaps put a big fat comment below HAVE_PREEMPT_DYNAMIC_KEY help to tell
> about this expectation as I guess it depends on arch/compiler?

Sure; I'll come up with something for v2.

Thanks,
Mark.