[PATCH 1/4] ARM: tlb: don't perform inner-shareable invalidation for local TLB ops

[Catalin Marinas]

On Wed, Mar 27, 2013 at 12:07:37PM +0000, Will Deacon wrote:
> Cheers for looking at this.

Just warming up for Marc's KVM patches ;)

> On Wed, Mar 27, 2013 at 10:34:30AM +0000, Catalin Marinas wrote:
> > On Mon, Mar 25, 2013 at 06:19:38PM +0000, Will Deacon wrote:
> > > @@ -352,22 +369,33 @@ static inline void local_flush_tlb_mm(struct mm_struct *mm)
> > >  		dsb();
> > >  
> > >  	if (possible_tlb_flags & (TLB_V3_FULL|TLB_V4_U_FULL|TLB_V4_D_FULL|TLB_V4_I_FULL)) {
> > > -		if (cpumask_test_cpu(get_cpu(), mm_cpumask(mm))) {
> > > +		if (!cpumask_test_cpu(smp_processor_id(), mm_cpumask(mm))) {
> > >  			tlb_op(TLB_V3_FULL, "c6, c0, 0", zero);
> > >  			tlb_op(TLB_V4_U_FULL, "c8, c7, 0", zero);
> > >  			tlb_op(TLB_V4_D_FULL, "c8, c6, 0", zero);
> > >  			tlb_op(TLB_V4_I_FULL, "c8, c5, 0", zero);
> > >  		}
> > > -		put_cpu();
> > 
> > Why is this change needed? You only flush the local TLB if the mm never
> > wasn't active on this processor?
> 
> Ouch, that's a cock-up, sorry. I'll remove the '!'.

Do we also need to disable preemtion?

> > > @@ -398,6 +426,21 @@ local_flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
> > >  	tlb_op(TLB_V6_U_PAGE, "c8, c7, 1", uaddr);
> > >  	tlb_op(TLB_V6_D_PAGE, "c8, c6, 1", uaddr);
> > >  	tlb_op(TLB_V6_I_PAGE, "c8, c5, 1", uaddr);
> > > +
> > > +	if (tlb_flag(TLB_BARRIER))
> > > +		dsb();
> > > +}
> > > +
> > > +static inline void
> > > +__flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
> > > +{
> > > +	const unsigned int __tlb_flag = __cpu_tlb_flags;
> > > +
> > > +	uaddr = (uaddr & PAGE_MASK) | ASID(vma->vm_mm);
> > > +
> > > +	if (tlb_flag(TLB_WB))
> > > +		dsb();
> > > +
> > 
> > I guess here we could just have a single *_tlb_page() variant. I
> > couldn't find any place where we call the local_flush_tlb_page()
> > explicitly, I guess we don't really need local semantics. On ARMv6 SMP,
> > they are local anyway.
> > 
> > If we have a single *_tlb_page() function, you would need to drop the
> > TLB_V6_*_PAGE from the v8 possible TLB ops.
> 
> Having the local variant doesn't hurt though, and provides the same symmetry
> as other architectures (powerpc, sh, tile, mips, ...).

It's probably harmless to have them, though they may not get used.

> > >  #ifdef CONFIG_ARM_ERRATA_720789
> > >  	tlb_op(TLB_V7_UIS_PAGE, "c8, c3, 3", uaddr & PAGE_MASK);
> > >  #else
> > > @@ -428,6 +471,22 @@ static inline void local_flush_tlb_kernel_page(unsigned long kaddr)
> > >  	tlb_op(TLB_V6_U_PAGE, "c8, c7, 1", kaddr);
> > >  	tlb_op(TLB_V6_D_PAGE, "c8, c6, 1", kaddr);
> > >  	tlb_op(TLB_V6_I_PAGE, "c8, c5, 1", kaddr);
> > > +
> > > +	if (tlb_flag(TLB_BARRIER)) {
> > > +		dsb();
> > > +		isb();
> > > +	}
> > > +}
> > 
> > I have some worries with this function. It is used by set_top_pte() and
> > it really doesn't look like it has local-only semantics. For example,
> > you use it do flush the I-cache aliases and this must target all the
> > CPUs because of speculative prefetches, which means that set_top_pte()
> > must set the new alias on all the CPUs.
> 
> This looks like a bug in set_top_pte when it's called for cache-flushing.
> However, the only core this would affect is 11MPCore, which uses the
> ipi-based flushing anyway, so I think we're ok.

I don't think its 11MPCore only, set_top_pte() is called by
flush_icache_alias() from flush_ptrace_access() even on ARMv7.

> > Highmem mappings need to be revisited as well.
> 
> I think they're ok. Everything is either done in atomic context or under a
> raw spinlock, so the mappings aren't expected to be used by other CPUs.

It's not whether they are used explicitly but whether a speculative TLB
load can bring them in on a different CPU. I don't immediately see a
problem with non-aliasing caches but needs some more thinking.

-- 
Catalin