Re: [PATCH v1 2/2] mm: don't install PMD mappings when THPs are disabled by the hw/process/vma

[David Hildenbrand]

On 11.10.24 13:36, Ryan Roberts wrote:
> On 11/10/2024 12:33, David Hildenbrand wrote:
> > On 11.10.24 13:29, Ryan Roberts wrote:
> > > On 11/10/2024 11:24, David Hildenbrand wrote:
> > > > We (or rather, readahead logic :) ) might be allocating a THP in the
> > > > pagecache and then try mapping it into a process that explicitly disabled
> > > > THP: we might end up installing PMD mappings.
> > > >
> > > > This is a problem for s390x KVM, which explicitly remaps all PMD-mapped
> > > > THPs to be PTE-mapped in s390_enable_sie()->thp_split_mm(), before
> > > > starting the VM.
> > > >
> > > > For example, starting a VM backed on a file system with large folios
> > > > supported makes the VM crash when the VM tries accessing such a mapping
> > > > using KVM.
> > > >
> > > > Is it also a problem when the HW disabled THP using
> > > > TRANSPARENT_HUGEPAGE_UNSUPPORTED? At least on x86 this would be the case
> > > > without X86_FEATURE_PSE.
> > > >
> > > > In the future, we might be able to do better on s390x and only disallow
> > > > PMD mappings -- what s390x and likely TRANSPARENT_HUGEPAGE_UNSUPPORTED
> > > > really wants. For now, fix it by essentially performing the same check as
> > > > would be done in __thp_vma_allowable_orders() or in shmem code, where this
> > > > works as expected, and disallow PMD mappings, making us fallback to PTE
> > > > mappings.
> > > >
> > > > Reported-by: Leo Fu <bfu@xxxxxxxxxx>
> > > > Fixes: 793917d997df ("mm/readahead: Add large folio readahead")
> > >
> > > Will this patch be difficult to backport given it depends on the previous patch
> > > and that doesn't have a Fixes tag?
> >
> > "difficult" -- not really. Andrew might want to tag patch #1  with "Fixes:" as
> > well, but I can also send simple stable backports that avoid patch #1.
> >
> > (Thinking again, I assume we want to Cc:stable)
> >
> > > > Cc: Thomas Huth <thuth@xxxxxxxxxx>
> > > > Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > > > Cc: Ryan Roberts <ryan.roberts@xxxxxxx>
> > > > Cc: Christian Borntraeger <borntraeger@xxxxxxxxxxxxx>
> > > > Cc: Janosch Frank <frankja@xxxxxxxxxxxxx>
> > > > Cc: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx>
> > > > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > > > ---
> > > >    mm/memory.c | 9 +++++++++
> > > >    1 file changed, 9 insertions(+)
> > > >
> > > > diff --git a/mm/memory.c b/mm/memory.c
> > > > index 2366578015ad..a2e501489517 100644
> > > > --- a/mm/memory.c
> > > > +++ b/mm/memory.c
> > > > @@ -4925,6 +4925,15 @@ vm_fault_t do_set_pmd(struct vm_fault *vmf, struct
> > > > page *page)
> > > >        pmd_t entry;
> > > >        vm_fault_t ret = VM_FAULT_FALLBACK;
> > > >    +    /*
> > > > +     * It is too late to allocate a small folio, we already have a large
> > > > +     * folio in the pagecache: especially s390 KVM cannot tolerate any
> > > > +     * PMD mappings, but PTE-mapped THP are fine. So let's simply refuse any
> > > > +     * PMD mappings if THPs are disabled.
> > > > +     */
> > > > +    if (thp_disabled_by_hw() || vma_thp_disabled(vma, vma->vm_flags))
> > > > +        return ret;
> > >
> > > Why not just call thp_vma_allowable_orders()?
> >
> > Why call thp_vma_allowable_orders() that does a lot more work that doesn't
> > really apply here? :)
>
> Yeah fair enough, I was just thinking it makes the code simpler to keep all the
> checks in one place. But no strong opinion.
>
> Either way:
>
> Reviewed-by: Ryan Roberts <ryan.roberts@xxxxxxx>

Thanks!

Also, I decided to not use "thp_vma_allowable_orders" because we are 
past the allocation phase (as indicated in the comment) and can really 
just change the way how we map the folio (PMD vs. PTE), not really 
*what* folio to use.

Ideally, in the future we have a different way of just saying "no PMD 
mappings please", decoupling the mapping from the allocation granularity.

--
Cheers,

David / dhildenb