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

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Wed, 22 Aug 2018 22:20:32 -0700
Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Wed, Aug 22, 2018 at 10:11 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > 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.  
> 
> Note that this obviously works fine for a hashed table model too - you
> just ignore the "freed_tables" bit entirely and continue to do
> whatever you always did.
> 
> And we can ignore it on x86 too, because we just see the range, and we
> invalidate the range, and that will always invalidate the mid-level
> caching too.
> 
> So the new bit is literally for arm and powerpc-radix (and maybe
> s390), but we want to make the actual VM interface truly generic and
> not have one set of code with five different behaviors (which we
> _currently_ kind of have with the whole in addition to all the
> HAVE_RCU_TABLE_FREE etc config options that modify how the code works.
> 
> It would be good to also cut down on the millions of functions that
> each architecture can override, because Christ, it got very confusing
> at times to follow just how the code flowed from generic code to
> architecture-specific macros back to generic code and then
> arch-specific inline helper functions.
> 
> It's a maze of underscores and "page" vs "table", and "flush" vs "remove" etc.
> 
> But that "it would be good to really make everybody to use as much of
> the generic code as possible" and everybody have the same pattern,
> that's a future thing. But the whole "let's just add that
> "freed_tables" thing would be part of trying to get people to use the
> same overall pattern, even if some architectures might not care about
> that detail.

For s390 the new freed_tables bit looks to be step into the right
direction. Right now there is the flush_mm bit in the mm->context
that I use to keep track of the need for a global flush of the mm.
The flush_mm bit is currently used for both lazy PTE flushing and
TLB flushes for freed page table pages.

I'll look into it once the generic patch is merged. The switch to
the generic mmu_gather is certainly not a change that can be done
in a hurry, we had too many subtle TLB flush issues in the past.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.