Re: [PATCH] arm64/mm: Avoid direct referencing page table enties in map_range()

[Will Deacon <will@kernel.org>]

On Thu, Aug 01, 2024 at 01:48:17PM +0100, Ryan Roberts wrote:
> On 01/08/2024 12:34, Will Deacon wrote:
> > On Thu, Jul 25, 2024 at 11:36:56AM +0100, Ryan Roberts wrote:
> >> On 25/07/2024 10:10, Anshuman Khandual wrote:
> >>> Like else where in arm64 platform, use WRITE_ONCE() in map_range() while
> >>> creating page table entries. This avoids referencing page table entries
> >>> directly.
> >>
> >> I could be wrong, but I don't think this code is ever operating on live
> >> pgtables? So there is never a potential to race with the HW walker and therefore
> >> no need to guarrantee copy atomicity? As long as the correct barriers are placed
> >> at the point where you load the pgdir into the TTBRx there should be no problem?
> >>
> >> If my assertion is correct, I don't think there is any need for this change.
> > 
> > Agreed.
> 
> I think I need to row back on this. It looks like map_range() does act on live
> pgtables; see map_kernel() where twopass == true. init_pg_dir is populated then
> installed in TTBR1, then permissions are modified with 3 [un]map_segment()
> calls, which call through to map_range().

I think the permission part is fine, but I hadn't spotted that
unmap_segment() uses map_range() to zap the ptes mapping the text. That
*does* need single-copy atomicity, so should probably use
__set_pte_nosync().

> So on that basis, I think the WRITE_ONCE() calls are warranted. And to be
> consistent, I'd additionally recommend adding a READ_ONCE() around the:
> 
> if (pte_none(*tbl)) {

Why? I don't think we need that.

Will