Re: [PATCH v1] x86/mm/pat: fix VM_PAT handling in COW mappings

[David Hildenbrand]

On 01.04.24 11:45, Ingo Molnar wrote:
> * David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> > > > > try the trivial restriction approach first, and only go with your original
> > > > > patch if that fails?
> > > >
> > > > Which version would you prefer, I had two alternatives (excluding comment
> > > > changes, white-space expected to be broken).
> > > >
> > > >
> > > > 1) Disallow when we would have set VM_PAT on is_cow_mapping()
> > > >
> > > > diff --git a/arch/x86/mm/pat/memtype.c b/arch/x86/mm/pat/memtype.c
> > > > index 0d72183b5dd0..6979912b1a5d 100644
> > > > --- a/arch/x86/mm/pat/memtype.c
> > > > +++ b/arch/x86/mm/pat/memtype.c
> > > > @@ -994,6 +994,9 @@ int track_pfn_remap(struct vm_area_struct *vma, pgprot_t *prot,
> > > >                                   && size == (vma->vm_end - vma->vm_start))) {
> > > >                   int ret;
> > > > +               if (is_cow_mapping(vma->vm_flags))
> > > > +                       return -EINVAL;
> > > > +
> > > >                   ret = reserve_pfn_range(paddr, size, prot, 0);
> > > >                   if (ret == 0 && vma)
> > > >                           vm_flags_set(vma, VM_PAT);
> > > >
> > > >
> > > > 2) Fallback to !VM_PAT
> > > >
> > > > diff --git a/arch/x86/mm/pat/memtype.c b/arch/x86/mm/pat/memtype.c
> > > > index 0d72183b5dd0..8e97156c9be8 100644
> > > > --- a/arch/x86/mm/pat/memtype.c
> > > > +++ b/arch/x86/mm/pat/memtype.c
> > > > @@ -990,8 +990,8 @@ int track_pfn_remap(struct vm_area_struct *vma, pgprot_t *prot,
> > > >           enum page_cache_mode pcm;
> > > >           /* reserve the whole chunk starting from paddr */
> > > > -       if (!vma || (addr == vma->vm_start
> > > > -                               && size == (vma->vm_end - vma->vm_start))) {
> > > > +       if (!vma || (!is_cow_mapping(vma->vm_flags) && addr == vma->vm_start &&
> > > > +                    size == (vma->vm_end - vma->vm_start))) {
> > > >                   int ret;
> > > >                   ret = reserve_pfn_range(paddr, size, prot, 0);
> > > >
> > > >
> > > >
> > > > Personally, I'd go for 2).
> > >
> > > So what's the advantage of #2? This is clearly something the user didn't
> > > really intend or think about much. Isn't explicitly failing that mapping a
> > > better option than silently downgrading it to !VM_PAT?
> > >
> > > (If I'm reading it right ...)
> >
> > I think a simple mmap(MAP_PRIVATE) of /dev/mem will unconditionally fail
> > with 1), while it keeps on working for 2).
> >
> > Note that I think we currently set VM_PAT on each and every system if
> > remap_pfn_range() will cover the whole VMA, even if pat is not actually
> > enabled.
> >
> > It's all a bit of a mess TBH, but I got my hands dirty enough on that.
> >
> > So 1) can be rather destructive ... 2) at least somehow keeps it working.
> >
> > For that reason I went with the current patch, because it's hard to tell
> > which use case you will end up breaking ... :/

Hi,

> Yeah, so I think you make valid observations, i.e. your first patch is
> probably the best one.

okay, so the original patch, thanks.

> But since it changes mm/memory.c, I'd like to pass that over to Andrew
> and the MM folks.
>
> The x86 bits:
>
>    Acked-by: Ingo Molnar <mingo@xxxxxxxxxx>


Thanks, there is now a conflict with other stuff that already landed in 
mm-unstable that moves follow_phys() to arch/x86/mm/pat/memtype.c.


@Andrew, this here is a fix, how should we best handle that? Likely the 
fix should go in first, and the fixup of Christoph's patch should be 
easy. Just let me know how you want to handle that.

--
Cheers,

David / dhildenb