Re: [PATCH v3 15/17] mm: pgtable: remove tlb_remove_page_ptdesc()

[Peter Zijlstra <peterz@infradead.org>]

On Sat, Dec 28, 2024 at 11:26:22AM +0200, Mike Rapoport wrote:
> On Mon, Dec 23, 2024 at 05:41:01PM +0800, Qi Zheng wrote:
> > Here we are explicitly dealing with struct page, and the following logic
> > semms strange:
> > 
> > tlb_remove_page_ptdesc((tlb), (page_ptdesc(pte)));
> > 
> > tlb_remove_page_ptdesc
> > --> tlb_remove_page(tlb, ptdesc_page(pt));
> > 
> > So remove tlb_remove_page_ptdesc() and make callers call tlb_remove_page()
> > directly.
> 
> Please don't. The ptdesc wrappers are there as a part of reducing the size
> of struct page project [1]. 
> 
> For now struct ptdesc overlaps struct page, but the goal is to have them
> separate and always operate on struct ptdesc when working with page tables.

I don't see how the current idiotic code helps with that at all.

Fundamentally tlb_remove_page() is about removing *pages* as from a PTE,
there should not be a page-table anywhere near here *ever*.

Yes, some architectures use tlb_remove_page() for page-tables too, but
that is more or less an implementation detail that can be fixed.

So no, please keep these patches and kill this utterly idiotic code.

The only thing that should eventually care about page-tables is
tlb_remove_table(), and that takes a 'void *' and is expected to match
whatever __tlb_remove_table() does.

Flipping those to pgdesc, once its actually a thing, should be fairly
straight forward.