Re: [RFC PATCH v5 4/8] mm: tlb: Pass struct mmu_gather to flush_pmd_tlb_range

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Apr 02, 2020 at 07:24:04PM +0800, Zhenyu Ye wrote:
> Thanks for your detailed explanation.  I notice that you used
> `tlb_end_vma` replace `flush_tlb_range`, which will call `tlb_flush`,
> then finally call `flush_tlb_range` in generic code.  However, some
> architectures define tlb_end_vma|tlb_flush|flush_tlb_range themselves,
> so this may cause problems.
> 
> For example, in s390, it defines:
> 
> #define tlb_end_vma(tlb, vma)			do { } while (0)
> 
> And it doesn't define it's own flush_pmd_tlb_range().  So there will be
> a mistake if we changed flush_pmd_tlb_range() using tlb_end_vma().
> 
> Is this really a problem or something I understand wrong ?

If tlb_end_vma() is a no-op, then tlb_finish_mmu() will do:
tlb_flush_mmu() -> tlb_flush_mmu_tlbonly() -> tlb_flush()

And s390 has tlb_flush().

If tlb_end_vma() is not a no-op and it calls tlb_flush_mmu_tlbonly(),
then tlb_finish_mmu()'s invocation of tlb_flush_mmu_tlbonly() will
terniate early due o no flags set.

IOW, it should all just work.


FYI the whole tlb_{start,end}_vma() thing is a only needed when the
architecture doesn't implement tlb_flush() and instead default to using
flush_tlb_range(), at which point we need to provide a 'fake' vma.

At the time I audited all architectures and they only look at VM_EXEC
(to do $I invalidation) and VM_HUGETLB (for pmd level invalidations),
but I forgot which architectures that were.

But that is all legacy code; eventually we'll get all archs a native
tlb_flush() and this can go away.