Re: [PATCH v2 07/11] mm/memory-failure: Convert memory_failure() to use a folio

[Matthew Wilcox <willy@infradead.org>]

On Wed, Apr 10, 2024 at 06:27:38PM -0700, Jane Chu wrote:
> On 4/10/2024 4:15 PM, Jane Chu wrote:
> 
> > On 4/10/2024 3:21 AM, Oscar Salvador wrote:
> > 
> > > On Mon, Apr 08, 2024 at 05:34:29PM -0700, Jane Chu wrote:
> > > > At this stage,  'p' is not expected to be a large page since a
> > > > _refcount has
> > > > been taken, right?  So is it reasonable to replace the "if
> > > > (retry)" block
> > > > with a VM_BUG_ON warning?  otherwise, if 'p' became part of a different
> > > > large folio,  how to recover from here ?
> > > We might have split the THP (if it was part of), but AFAICS
> > > nothing stops the kernel to coallesce this page again into a new THP
> > > via e.g:
> > > madvise(MADV_HUGEPAGE) ?
> > 
> > Good point, thanks!
> 
> Come to think of it a bit more, maybe the check could be
> 
>     if (folio_test_large(folio) || (folio != page_folio(p))) {
> 
> ? Because, 'p' could have become part of a THP after shake_folio() and
> before folio_lock(), right?

What is the mechanism for reassembling a large folio?  I don't see that
code anywhere.