We cannot use the pte value used in set_pte_at for pte_same comparison,
because archs like ppc64, filter/add new pte flag in set_pte_at. Instead
fetch the pte value inside hugetlb_cow. We are comparing pte value to
make sure the pte didn't change since we dropped the page table lock.
hugetlb_cow get called with page table lock held, and we can take a copy
of the pte value before we drop the page table lock.
With hugetlbfs, we optimize the MAP_PRIVATE write fault path with no
previous mapping (huge_pte_none entries), by forcing a cow in the fault
path. This avoid take an addition fault to covert a read-only mapping
to read/write. Here we were comparing a recently instantiated pte (via
set_pte_at) to the pte values from linux page table. As explained above
on ppc64 such pte_same check returned wrong result, resulting in us
taking an additional fault on ppc64.
Fixes: 6a119eae942c ("powerpc/mm: Add a _PAGE_PTE bit")
Reported-by: Jan Stancek <[email protected]>
Signed-off-by: Aneesh Kumar K.V <[email protected]>
---
mm/hugetlb.c | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index ec49d9ef1eef..da8fbd02b92e 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3386,15 +3386,17 @@ static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
* Keep the pte_same checks anyway to make transition from the mutex easier.
*/
static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
- unsigned long address, pte_t *ptep, pte_t pte,
- struct page *pagecache_page, spinlock_t *ptl)
+ unsigned long address, pte_t *ptep,
+ struct page *pagecache_page, spinlock_t *ptl)
{
+ pte_t pte;
struct hstate *h = hstate_vma(vma);
struct page *old_page, *new_page;
int ret = 0, outside_reserve = 0;
unsigned long mmun_start; /* For mmu_notifiers */
unsigned long mmun_end; /* For mmu_notifiers */
+ pte = huge_ptep_get(ptep);
old_page = pte_page(pte);
retry_avoidcopy:
@@ -3668,7 +3670,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
hugetlb_count_add(pages_per_huge_page(h), mm);
if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
/* Optimization, do the COW without a second fault */
- ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
+ ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
}
spin_unlock(ptl);
@@ -3822,8 +3824,8 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
if (flags & FAULT_FLAG_WRITE) {
if (!huge_pte_write(entry)) {
- ret = hugetlb_cow(mm, vma, address, ptep, entry,
- pagecache_page, ptl);
+ ret = hugetlb_cow(mm, vma, address, ptep,
+ pagecache_page, ptl);
goto out_put_page;
}
entry = huge_pte_mkdirty(entry);
--
2.10.1
On Tue, 18 Oct 2016 21:12:45 +0530 "Aneesh Kumar K.V" <[email protected]> wrote:
> We cannot use the pte value used in set_pte_at for pte_same comparison,
> because archs like ppc64, filter/add new pte flag in set_pte_at. Instead
> fetch the pte value inside hugetlb_cow. We are comparing pte value to
> make sure the pte didn't change since we dropped the page table lock.
> hugetlb_cow get called with page table lock held, and we can take a copy
> of the pte value before we drop the page table lock.
>
> With hugetlbfs, we optimize the MAP_PRIVATE write fault path with no
> previous mapping (huge_pte_none entries), by forcing a cow in the fault
> path. This avoid take an addition fault to covert a read-only mapping
> to read/write. Here we were comparing a recently instantiated pte (via
> set_pte_at) to the pte values from linux page table. As explained above
> on ppc64 such pte_same check returned wrong result, resulting in us
> taking an additional fault on ppc64.
>From my reading this is a minor performance improvement and a -stable
backport isn't needed. But it is unclear whether the impact warrants a
4.9 merge.
Please be careful about describing end-user visible impacts when fixing
bugs, thanks.
On Tuesday, October 18, 2016 11:43 PM Aneesh Kumar K.V wrote:
>
> We cannot use the pte value used in set_pte_at for pte_same comparison,
> because archs like ppc64, filter/add new pte flag in set_pte_at. Instead
> fetch the pte value inside hugetlb_cow. We are comparing pte value to
> make sure the pte didn't change since we dropped the page table lock.
> hugetlb_cow get called with page table lock held, and we can take a copy
> of the pte value before we drop the page table lock.
>
> With hugetlbfs, we optimize the MAP_PRIVATE write fault path with no
> previous mapping (huge_pte_none entries), by forcing a cow in the fault
> path. This avoid take an addition fault to covert a read-only mapping
> to read/write. Here we were comparing a recently instantiated pte (via
> set_pte_at) to the pte values from linux page table. As explained above
> on ppc64 such pte_same check returned wrong result, resulting in us
> taking an additional fault on ppc64.
>
> Fixes: 6a119eae942c ("powerpc/mm: Add a _PAGE_PTE bit")
>
> Reported-by: Jan Stancek <[email protected]>
> Signed-off-by: Aneesh Kumar K.V <[email protected]>
> ---
Acked-by: Hillf Danton <[email protected]>
> mm/hugetlb.c | 12 +++++++-----
> 1 file changed, 7 insertions(+), 5 deletions(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index ec49d9ef1eef..da8fbd02b92e 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -3386,15 +3386,17 @@ static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
> * Keep the pte_same checks anyway to make transition from the mutex easier.
> */
> static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
> - unsigned long address, pte_t *ptep, pte_t pte,
> - struct page *pagecache_page, spinlock_t *ptl)
> + unsigned long address, pte_t *ptep,
> + struct page *pagecache_page, spinlock_t *ptl)
> {
> + pte_t pte;
> struct hstate *h = hstate_vma(vma);
> struct page *old_page, *new_page;
> int ret = 0, outside_reserve = 0;
> unsigned long mmun_start; /* For mmu_notifiers */
> unsigned long mmun_end; /* For mmu_notifiers */
>
> + pte = huge_ptep_get(ptep);
> old_page = pte_page(pte);
>
> retry_avoidcopy:
> @@ -3668,7 +3670,7 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
> hugetlb_count_add(pages_per_huge_page(h), mm);
> if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
> /* Optimization, do the COW without a second fault */
> - ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
> + ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
> }
>
> spin_unlock(ptl);
> @@ -3822,8 +3824,8 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>
> if (flags & FAULT_FLAG_WRITE) {
> if (!huge_pte_write(entry)) {
> - ret = hugetlb_cow(mm, vma, address, ptep, entry,
> - pagecache_page, ptl);
> + ret = hugetlb_cow(mm, vma, address, ptep,
> + pagecache_page, ptl);
> goto out_put_page;
> }
> entry = huge_pte_mkdirty(entry);
> --
> 2.10.1
Andrew Morton <[email protected]> writes:
> On Tue, 18 Oct 2016 21:12:45 +0530 "Aneesh Kumar K.V" <[email protected]> wrote:
>
>> We cannot use the pte value used in set_pte_at for pte_same comparison,
>> because archs like ppc64, filter/add new pte flag in set_pte_at. Instead
>> fetch the pte value inside hugetlb_cow. We are comparing pte value to
>> make sure the pte didn't change since we dropped the page table lock.
>> hugetlb_cow get called with page table lock held, and we can take a copy
>> of the pte value before we drop the page table lock.
>>
>> With hugetlbfs, we optimize the MAP_PRIVATE write fault path with no
>> previous mapping (huge_pte_none entries), by forcing a cow in the fault
>> path. This avoid take an addition fault to covert a read-only mapping
>> to read/write. Here we were comparing a recently instantiated pte (via
>> set_pte_at) to the pte values from linux page table. As explained above
>> on ppc64 such pte_same check returned wrong result, resulting in us
>> taking an additional fault on ppc64.
>
> From my reading this is a minor performance improvement and a -stable
> backport isn't needed. But it is unclear whether the impact warrants a
> 4.9 merge.
This patch workaround the issue reported at https://lkml.kernel.org/r/[email protected]
The reason for that OOM was a reserve count accounting issue which
happens in the error path of hugetlb_cow. Not this patch avoid us taking
the error path and hence we don't have the reported OOM.
An actual fix for that issue is being worked on by Mike Kravetz.
>
> Please be careful about describing end-user visible impacts when fixing
> bugs, thanks.
-aneesh