Re: [PATCH 3/4] mm/tlb, x86/mm: Support invalidating TLB caches for RCU_TABLE_FREE

[Will Deacon]

On Wed, Aug 22, 2018 at 10:11:41PM -0700, Linus Torvalds wrote:
> On Wed, Aug 22, 2018 at 9:54 PM Benjamin Herrenschmidt <benh@xxxxxxxxxxx> wrote:
> >
> >
> > So we do need a different flush instruction for the page tables vs. the
> > normal TLB pages.
> 
> Right. ARM wants it too. x86 is odd in that a regular "invlpg" already
> invalidates all the internal tlb cache nodes.
> 
> So the "new world order" is exactly that patch that PeterZ sent you, that adds a
> 
> +       unsigned int            freed_tables : 1;
> 
> to the 'struct mmu_gather', and then makes all those
> pte/pmd/pud/p4d_free_tlb() functions set that bit.
> 
> So I'm referring to the email PeterZ sent you in this thread that said:
> 
>   Nick, Will is already looking at using this to remove the synchronous
>   invalidation from __p*_free_tlb() for ARM, could you have a look to see
>   if PowerPC-radix could benefit from that too?
> 
>   Basically, using a patch like the below, would give your tlb_flush()
>   information on if tables were removed or not.
> 
> then, in that model, you do *not* need to override these
> pte/pmd/pud/p4d_free_tlb() macros at all (well, you *can* if you want
> to, for doing games with the range modification, but let's sayt that
> you don't need that right now).
> 
> So instead, when you get to the actual "tlb_flush(tlb)", you do
> exactly that - flush the tlb. And the mmu_gather structure shows you
> how much you need to flush. If you see that "freed_tables" is set,
> then you know that you need to also do the special instruction to
> flush the inner level caches. The range continues to show the page
> range.

The only problem with this approach is that we've lost track of the granule
size by the point we get to the tlb_flush(), so we can't adjust the stride of
the TLB invalidations for huge mappings, which actually works nicely in the
synchronous case (e.g. we perform a single invalidation for a 2MB mapping,
rather than iterating over it at a 4k granule).

One thing we could do is switch to synchronous mode if we detect a change in
granule (i.e. treat it like a batch failure).

Will