Re: [EXTERNAL] [PATCH] mm/thp: fix "mm: thp: kill __transhuge_page_enabled()"

From: Zach O'Keefe
Date: Mon Aug 21 2023 - 11:09:24 EST


On Fri, Aug 18, 2023 at 2:21 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> On Thu, Aug 17, 2023 at 11:29 AM Zach O'Keefe <zokeefe@xxxxxxxxxx> wrote:
> >
> > On Thu, Aug 17, 2023 at 10:47 AM Yang Shi <shy828301@xxxxxxxxx> wrote:
> > >
> > > On Wed, Aug 16, 2023 at 2:48 PM Zach O'Keefe <zokeefe@xxxxxxxxxx> wrote:
> > > >
> > > > > We have a out of tree driver that maps huge pages through a file handle and
> > > > > relies on -> huge_fault. It used to work in 5.19 kernels but 6.1 changed this
> > > > > behaviour.
> > > > >
> > > > > I don’t think reverting the earlier behaviour of fault_path for huge pages should
> > > > > impact kernel negatively.
> > > > >
> > > > > Do you think we can restore this earlier behaviour of kernel to allow page fault
> > > > > for huge pages via ->huge_fault.
> > > >
> > > > That seems reasonable to me. I think using the existence of a
> > > > ->huge_fault() handler as a predicate to return "true" makes sense to
> > > > me. The "normal" flow for file-backed memory along fault path still
> > > > needs to return "false", so that we correctly fallback to ->fault()
> > > > handler. Unless there are objections, I can do that in a v2.
> > >
> > > Sorry for chiming in late. I'm just back from vacation and trying to catch up...
> > >
> > > IIUC the out-of-tree driver tries to allocate huge page and install
> > > PMD mapping via huge_fault() handler, but the cleanup of
> > > hugepage_vma_check() prevents this due to the check to
> > > VM_NO_KHUGEPAGED?
> > >
> > > So you would like to check whether a huge_fault() handler existed
> > > instead of vma_is_dax()?
> >
> > Sorry for the multiple threads here. There are two problems: (a) the
> > VM_NO_KHUGEPAGED check along fault path, and (b) we don't give
> > ->huge_fault() a fair shake, if it exists, along fault path. The
> > current code assumes vma_is_dax() iff ->huge_fault() exists.
> >
> > (a) is easy enough to fix. For (b), I'm currently looking at the
> > possibility of not worrying about ->huge_fault() in
> > hugepage_vma_check(), and just letting create_huge_pud() /
> > create_huge_pmd() check and fallback as necessary. I think we'll need
> > the explicit DAX check still, since we want to keep khugepaged and
> > MADV_COLLAPSE away, and the presence / absence of ->huge_fault() isn't
> > enough to know that (well.. today it kind of is, but we shouldn't
> > depend on it).
>
> You meant something like:
>
> if (vma->vm_ops->huge_fault) {
> if (vma_is_dax(vma))
> return in_pf;
>
> /Fall through */
> }

I don't think this will work for Saurabh's case, since IIUC, they
aren't using dax, but are using VM_HUGEPAGE|VM_MIXEDMAP, faulted in
using ->huge_fault()

The old (v5.19) fault path looked like:

static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
unsigned long vm_flags)
{
/* Explicitly disabled through madvise. */
if ((vm_flags & VM_NOHUGEPAGE) ||
test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
return false;
return true;
}

/*
* to be used on vmas which are known to support THP.
* Use transparent_hugepage_active otherwise
*/
static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
{

/*
* If the hardware/firmware marked hugepage support disabled.
*/
if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
return false;

if (!transhuge_vma_enabled(vma, vma->vm_flags))
return false;

if (vma_is_temporary_stack(vma))
return false;

if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
return true;

if (vma_is_dax(vma))
return true;

if (transparent_hugepage_flags &
(1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
return !!(vma->vm_flags & VM_HUGEPAGE);

return false;
}

For non-anonymous, the next check (in create_huge_*) would be for that
->huge_fault handler, falling back as necessary if it didn't exist.

The patch I sent out last week[1] somewhat restores this logic -- the
only difference being we do the check for ->huge_fault in
hugepage_vma_check() as well. This is so smaps can surface this
possibility with some accuracy. I just realized it will erroneously
return "true" for the collapse path, however..

Maybe Matthew was right about unifying everything here :P That's 2
mistakes I've made in trying to fix this issue (but maybe that's just
me).

[1] https://lore.kernel.org/linux-mm/20230818211533.2523697-1-zokeefe@xxxxxxxxxx/