Re: [PATCH v9 00/12] AMD broadcast TLB invalidation

[Peter Zijlstra]

On Thu, Feb 06, 2025 at 09:48:25AM -0500, Rik van Riel wrote:
> On Thu, 2025-02-06 at 15:23 +0100, Peter Zijlstra wrote:
> > On Thu, Feb 06, 2025 at 09:16:35AM -0500, Rik van Riel wrote:
> > 
> > > This can result in flush_tlb_mm_range being called
> > > with a stride_shift for 2MB pages, but a range ending
> > > on a 4kB aligned (not 2MB aligned) boundary.
> > > 
> > > Peter, how should we solve this one?
> > 
> > I don't think that's wrong per-se, since all we really need is for
> > end
> > to be past the end, one byte, one page, or one stride don't matter
> > much.
> > 
> > Anyway, I'm in desperate need of a break, so I'm not quite sure what
> > the
> > best way forward is.
> > 
> Given that the tlb_flush() code is used only for
> page table freeing, 

mmu_gather is used for both page and page-table freeing.

> we can just round up the
> end address to the nearest stride boundary
> there, with a comment explaining why?

Well, why are we rounding at all? I don't think I've seen an explanation
for that anywhere yet.

What made you do this?

> Alternatively, we could just change the page
> sizes used in pmd_free_tlb, pud_free_tlb,
> and p4d_free_tlb, given that the functions
> called now have parameters they didn't seem
> to have back in 2014, when the linked email
> in the comment was written?

Well, it really shouldn't matter.

Notably, stride is simply the smallest size encountered during the
gather; not something that can be determined a priory.

Like I said, end is exclusive, so it doesn't really matter how much
further it is, and PAGE_SIZE is a natural enough step, seeing it is the
smallest granularity.

So start and end not being page aligned is a definite fail.
And start should be stride aligned.

But other than that, I don't see it matters much.