On Thu, Nov 15, 2012 at 01:52:44PM -0800, David Rientjes wrote:
> On Thu, 15 Nov 2012, Kirill A. Shutemov wrote:
>
> > > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > > > index f36bc7d..41f05f1 100644
> > > > --- a/mm/huge_memory.c
> > > > +++ b/mm/huge_memory.c
> > > > @@ -726,6 +726,16 @@ int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
> > > > return VM_FAULT_OOM;
> > > > if (unlikely(khugepaged_enter(vma)))
> > > > return VM_FAULT_OOM;
> > > > + if (!(flags & FAULT_FLAG_WRITE)) {
> > > > + pgtable_t pgtable;
> > > > + pgtable = pte_alloc_one(mm, haddr);
> > > > + if (unlikely(!pgtable))
> > > > + goto out;
> > >
> > > No use in retrying, just return VM_FAULT_OOM.
> >
> > Hm. It's consistent with non-hzp path: if pte_alloc_one() in
> > __do_huge_pmd_anonymous_page() fails __do_huge_pmd_anonymous_page()
> > returns VM_FAULT_OOM which leads to "goto out".
> >
>
> If the pte_alloc_one(), which wraps __pte_alloc(), you're adding fails,
> it's pointless to "goto out" to try __pte_alloc() which we know won't
> succeed.
>
> > Should it be fixed too?
> >
>
> It's done for maintainablility because although
> __do_huge_pmd_anonymous_page() will only return VM_FAULT_OOM today when
> pte_alloc_one() fails, if it were to ever fail in a different way then the
> caller is already has a graceful failure.
Okay, here's fixlet. Andrew, please squash it to the patch 08/11.
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 1f6c6de..d7ab890 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -793,7 +793,7 @@ int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
unsigned long zero_pfn;
pgtable = pte_alloc_one(mm, haddr);
if (unlikely(!pgtable))
- goto out;
+ return VM_FAULT_OOM;
zero_pfn = get_huge_zero_page();
if (unlikely(!zero_pfn)) {
pte_free(mm, pgtable);
--
Kirill A. Shutemov
Attachment:
signature.asc
Description: Digital signature