hugepage_vma_check is called during khugepaged_scan_mm_slot to ensure
that khugepaged doesn't try to allocate THPs in vmas where they are
disallowed, either due to THPs being disabled system-wide, or through
MADV_NOHUGEPAGE.
The logic that hugepage_vma_check uses doesn't seem to cover all cases,
in my opinion. Looking at the original code:
if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
(vma->vm_flags & VM_NOHUGEPAGE))
We can see that it's possible to have THP disabled system-wide, but still
receive THPs in this vma. It seems that it's assumed that just because
khugepaged_always == false, TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG must be
set, which is not the case. We could have VM_HUGEPAGE set, but have THP
set to "never" system-wide, in which case, the condition presented in the
if will evaluate to false, and (provided the other checks pass) we can
end up giving out a THP even though the behavior is set to "never."
While we do properly check these flags in khugepaged_has_work, it looks
like it's possible to sleep after we check khugepaged_hask_work, but
before hugepage_vma_check, during which time, hugepages could have been
disabled system-wide, in which case, we could hand out THPs when we
shouldn't be.
This small fix makes hugepage_vma_check work more like
transparent_hugepage_enabled, checking if THPs are set to "always"
system-wide, then checking if THPs are set to "madvise," as well as
making sure that VM_HUGEPAGE is set for this vma.
Signed-off-by: Alex Thorlton <[email protected]>
Reported-by: Alex Thorlton <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
mm/huge_memory.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 95d1acb..f62fba9 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2394,7 +2394,8 @@ static struct page
static bool hugepage_vma_check(struct vm_area_struct *vma)
{
- if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
+ if ((!khugepaged_always() ||
+ (!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_req_madv())) ||
(vma->vm_flags & VM_NOHUGEPAGE))
return false;
--
1.7.12.4
On Tue, 21 Jan 2014, Alex Thorlton wrote:
> hugepage_vma_check is called during khugepaged_scan_mm_slot to ensure
> that khugepaged doesn't try to allocate THPs in vmas where they are
> disallowed, either due to THPs being disabled system-wide, or through
> MADV_NOHUGEPAGE.
>
> The logic that hugepage_vma_check uses doesn't seem to cover all cases,
> in my opinion. Looking at the original code:
>
> if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
> (vma->vm_flags & VM_NOHUGEPAGE))
>
> We can see that it's possible to have THP disabled system-wide, but still
> receive THPs in this vma. It seems that it's assumed that just because
> khugepaged_always == false, TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG must be
> set, which is not the case. We could have VM_HUGEPAGE set, but have THP
> set to "never" system-wide, in which case, the condition presented in the
> if will evaluate to false, and (provided the other checks pass) we can
> end up giving out a THP even though the behavior is set to "never."
>
You should be able to add a
BUG_ON(current != khugepaged_thread);
here since khugepaged is supposed to be the only caller to the function.
> While we do properly check these flags in khugepaged_has_work, it looks
> like it's possible to sleep after we check khugepaged_hask_work, but
> before hugepage_vma_check, during which time, hugepages could have been
> disabled system-wide, in which case, we could hand out THPs when we
> shouldn't be.
>
You're talking about when thp is set to "never" and before khugepaged has
stopped, correct?
That doesn't seem like a bug to me or anything that needs to be fixed, the
sysfs knob could be switched even after hugepage_vma_check() is called and
before a hugepage is actually collapsed so you have the same race.
The only thing that's guaranteed is that, upon writing "never" to
/sys/kernel/mm/transparent_hugepage/enabled, no more thp memory will be
collapsed after khugepaged has stopped.
On Tue, Jan 21, 2014 at 03:24:08PM -0800, David Rientjes wrote:
> On Tue, 21 Jan 2014, Alex Thorlton wrote:
>
> > hugepage_vma_check is called during khugepaged_scan_mm_slot to ensure
> > that khugepaged doesn't try to allocate THPs in vmas where they are
> > disallowed, either due to THPs being disabled system-wide, or through
> > MADV_NOHUGEPAGE.
> >
> > The logic that hugepage_vma_check uses doesn't seem to cover all cases,
> > in my opinion. Looking at the original code:
> >
> > if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
> > (vma->vm_flags & VM_NOHUGEPAGE))
> >
> > We can see that it's possible to have THP disabled system-wide, but still
> > receive THPs in this vma. It seems that it's assumed that just because
> > khugepaged_always == false, TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG must be
> > set, which is not the case. We could have VM_HUGEPAGE set, but have THP
> > set to "never" system-wide, in which case, the condition presented in the
> > if will evaluate to false, and (provided the other checks pass) we can
> > end up giving out a THP even though the behavior is set to "never."
> >
>
> You should be able to add a
>
> BUG_ON(current != khugepaged_thread);
>
> here since khugepaged is supposed to be the only caller to the function.
>
> > While we do properly check these flags in khugepaged_has_work, it looks
> > like it's possible to sleep after we check khugepaged_hask_work, but
> > before hugepage_vma_check, during which time, hugepages could have been
> > disabled system-wide, in which case, we could hand out THPs when we
> > shouldn't be.
> >
>
> You're talking about when thp is set to "never" and before khugepaged has
> stopped, correct?
Yes, that's correct.
> That doesn't seem like a bug to me or anything that needs to be fixed, the
> sysfs knob could be switched even after hugepage_vma_check() is called and
> before a hugepage is actually collapsed so you have the same race.
>
> The only thing that's guaranteed is that, upon writing "never" to
> /sys/kernel/mm/transparent_hugepage/enabled, no more thp memory will be
> collapsed after khugepaged has stopped.
That makes sense, I wasn't aware that that's the expected behavior here.
I suppose this isn't something that needs to be changed, in that case.
I needed the logic broken out a bit more explicitly (madvise/never case
need to be handled separately) for a patch that I'm working on - that's
when this caught my attention. Good to know that a change to the
system-wide switch shouldn't affect khugepaged if it's already running.
I would've screwed up that behavior with my patch :)
Thanks, David!
- Alex