We have usecase to use tmpfs as QEMU memory backend and we would like to
take the advantage of THP as well. But, our test shows the EPT is not
PMD mapped even though the underlying THP are PMD mapped on host.
The number showed by /sys/kernel/debug/kvm/largepage is much less than
the number of PMD mapped shmem pages as the below:
7f2778200000-7f2878200000 rw-s 00000000 00:14 262232 /dev/shm/qemu_back_mem.mem.Hz2hSf (deleted)
Size: 4194304 kB
[snip]
AnonHugePages: 0 kB
ShmemPmdMapped: 579584 kB
[snip]
Locked: 0 kB
cat /sys/kernel/debug/kvm/largepages
12
And some benchmarks do worse than with anonymous THPs.
By digging into the code we figured out that commit 127393fbe597 ("mm:
thp: kvm: fix memory corruption in KVM with THP enabled") checks if
there is a single PTE mapping on the page for anonymous THP when
setting up EPT map. But, the _mapcount < 0 check doesn't fit to page
cache THP since every subpage of page cache THP would get _mapcount
inc'ed once it is PMD mapped, so PageTransCompoundMap() always returns
false for page cache THP. This would prevent KVM from setting up PMD
mapped EPT entry.
So we need handle page cache THP correctly. However, when page cache
THP's PMD gets split, kernel just remove the map instead of setting up
PTE map like what anonymous THP does. Before KVM calls get_user_pages()
the subpages may get PTE mapped even though it is still a THP since the
page cache THP may be mapped by other processes at the mean time.
Checking its _mapcount and whether the THP has PTE mapped or not.
Although this may report some false negative cases (PTE mapped by other
processes), it looks not trivial to make this accurate.
With this fix /sys/kernel/debug/kvm/largepage would show reasonable
pages are PMD mapped by EPT as the below:
7fbeaee00000-7fbfaee00000 rw-s 00000000 00:14 275464 /dev/shm/qemu_back_mem.mem.SKUvat (deleted)
Size: 4194304 kB
[snip]
AnonHugePages: 0 kB
ShmemPmdMapped: 557056 kB
[snip]
Locked: 0 kB
cat /sys/kernel/debug/kvm/largepages
271
And the benchmarks are as same as anonymous THPs.
Signed-off-by: Yang Shi <[email protected]>
Reported-by: Gang Deng <[email protected]>
Tested-by: Gang Deng <[email protected]>
Suggested-by: Hugh Dickins <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Cc: <[email protected]> 4.8+
---
v2: Adopted the suggestion from Hugh to use _mapcount and compound_mapcount.
But I just open coding compound_mapcount to avoid duplicating the
definition of compound_mapcount_ptr in two different files. Since
"compound_mapcount" looks self-explained so I'm supposed the open
coding would not affect the readability.
include/linux/page-flags.h | 22 ++++++++++++++++++++--
1 file changed, 20 insertions(+), 2 deletions(-)
diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index f91cb88..954a877 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -622,12 +622,30 @@ static inline int PageTransCompound(struct page *page)
*
* Unlike PageTransCompound, this is safe to be called only while
* split_huge_pmd() cannot run from under us, like if protected by the
- * MMU notifier, otherwise it may result in page->_mapcount < 0 false
+ * MMU notifier, otherwise it may result in page->_mapcount check false
* positives.
+ *
+ * We have to treat page cache THP differently since every subpage of it
+ * would get _mapcount inc'ed once it is PMD mapped. But, it may be PTE
+ * mapped in the current process so comparing subpage's _mapcount to
+ * compound_mapcount ot filter out PTE mapped case.
*/
static inline int PageTransCompoundMap(struct page *page)
{
- return PageTransCompound(page) && atomic_read(&page->_mapcount) < 0;
+ struct page *head;
+ int map_count;
+
+ if (!PageTransCompound(page))
+ return 0;
+
+ if (PageAnon(page))
+ return atomic_read(&page->_mapcount) < 0;
+
+ head = compound_head(page);
+ map_count = atomic_read(&page->_mapcount);
+ /* File THP is PMD mapped and not double mapped */
+ return map_count >= 0 &&
+ map_count == atomic_read(&head[1].compound_mapcount);
}
/*
--
1.8.3.1
On 10/23/19 10:24 AM, Matthew Wilcox wrote:
> On Thu, Oct 24, 2019 at 01:05:04AM +0800, Yang Shi wrote:
>> + return map_count >= 0 &&
>> + map_count == atomic_read(&head[1].compound_mapcount);
>> }
> I didn't like Hugh's duplicate definition either. May I suggest:
Thanks, Willy. It is fine to me. Will take it in v3.
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 2f2199a51941..3d0efd937d2b 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -695,11 +695,6 @@ static inline void *kvcalloc(size_t n, size_t size, gfp_t flags)
>
> extern void kvfree(const void *addr);
>
> -static inline atomic_t *compound_mapcount_ptr(struct page *page)
> -{
> - return &page[1].compound_mapcount;
> -}
> -
> static inline int compound_mapcount(struct page *page)
> {
> VM_BUG_ON_PAGE(!PageCompound(page), page);
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 2222fa795284..270aa8fd2800 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -221,6 +221,11 @@ struct page {
> #endif
> } _struct_page_alignment;
>
> +static inline atomic_t *compound_mapcount_ptr(struct page *page)
> +{
> + return &page[1].compound_mapcount;
> +}
> +
> /*
> * Used for sizing the vmemmap region on some architectures
> */
On Wed, 23 Oct 2019, Yang Shi wrote:
> On 10/23/19 12:28 PM, Hugh Dickins wrote:
> >
> > > + return map_count >= 0 &&
> > You have added a map_count >= 0 test there. Okay, not wrong, but not
> > necessary, and not consistent with what's returned in the PageAnon
> > case (if this were called for an unmapped page).
>
> I was thinking about this too. I'm wondering there might be a case that the
> PMD is split and it was the last PMD map, in this case subpage's _mapcount is
> also equal to compound_mapcount (both is -1). So, it would return true, then
> KVM may setup PMD map in EPT, but it might be PTE mapped later on the host.
> But, I'm not quite sure if this is really possible or if this is really a
> integrity problem. So, I thought it might be safer to add this check.
The mmu_notifier_invalidate_range_start.._end() in __split_huge_pmd(),
with KVM's locking and sequence counting, is required to protect
against such races.
Hugh
On Thu, Oct 24, 2019 at 01:05:04AM +0800, Yang Shi wrote:
> + return map_count >= 0 &&
> + map_count == atomic_read(&head[1].compound_mapcount);
> }
I didn't like Hugh's duplicate definition either. May I suggest:
diff --git a/include/linux/mm.h b/include/linux/mm.h
index 2f2199a51941..3d0efd937d2b 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -695,11 +695,6 @@ static inline void *kvcalloc(size_t n, size_t size, gfp_t flags)
extern void kvfree(const void *addr);
-static inline atomic_t *compound_mapcount_ptr(struct page *page)
-{
- return &page[1].compound_mapcount;
-}
-
static inline int compound_mapcount(struct page *page)
{
VM_BUG_ON_PAGE(!PageCompound(page), page);
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 2222fa795284..270aa8fd2800 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -221,6 +221,11 @@ struct page {
#endif
} _struct_page_alignment;
+static inline atomic_t *compound_mapcount_ptr(struct page *page)
+{
+ return &page[1].compound_mapcount;
+}
+
/*
* Used for sizing the vmemmap region on some architectures
*/
On Wed, 23 Oct 2019, Yang Shi wrote:
> On 10/23/19 10:24 AM, Matthew Wilcox wrote:
> > On Thu, Oct 24, 2019 at 01:05:04AM +0800, Yang Shi wrote:
> > > + return map_count >= 0 &&
> > > + map_count == atomic_read(&head[1].compound_mapcount);
> > > }
> > I didn't like Hugh's duplicate definition either. May I suggest:
>
> Thanks, Willy. It is fine to me. Will take it in v3.
Agreed, that will be better.
Hugh
On Thu, 24 Oct 2019, Yang Shi wrote:
> We have usecase to use tmpfs as QEMU memory backend and we would like to
> take the advantage of THP as well. But, our test shows the EPT is not
> PMD mapped even though the underlying THP are PMD mapped on host.
> The number showed by /sys/kernel/debug/kvm/largepage is much less than
> the number of PMD mapped shmem pages as the below:
>
> 7f2778200000-7f2878200000 rw-s 00000000 00:14 262232 /dev/shm/qemu_back_mem.mem.Hz2hSf (deleted)
> Size: 4194304 kB
> [snip]
> AnonHugePages: 0 kB
> ShmemPmdMapped: 579584 kB
> [snip]
> Locked: 0 kB
>
> cat /sys/kernel/debug/kvm/largepages
> 12
>
> And some benchmarks do worse than with anonymous THPs.
>
> By digging into the code we figured out that commit 127393fbe597 ("mm:
> thp: kvm: fix memory corruption in KVM with THP enabled") checks if
> there is a single PTE mapping on the page for anonymous THP when
> setting up EPT map. But, the _mapcount < 0 check doesn't fit to page
> cache THP since every subpage of page cache THP would get _mapcount
> inc'ed once it is PMD mapped, so PageTransCompoundMap() always returns
> false for page cache THP. This would prevent KVM from setting up PMD
> mapped EPT entry.
>
> So we need handle page cache THP correctly. However, when page cache
> THP's PMD gets split, kernel just remove the map instead of setting up
> PTE map like what anonymous THP does. Before KVM calls get_user_pages()
> the subpages may get PTE mapped even though it is still a THP since the
> page cache THP may be mapped by other processes at the mean time.
>
> Checking its _mapcount and whether the THP has PTE mapped or not.
> Although this may report some false negative cases (PTE mapped by other
> processes), it looks not trivial to make this accurate.
>
> With this fix /sys/kernel/debug/kvm/largepage would show reasonable
> pages are PMD mapped by EPT as the below:
>
> 7fbeaee00000-7fbfaee00000 rw-s 00000000 00:14 275464 /dev/shm/qemu_back_mem.mem.SKUvat (deleted)
> Size: 4194304 kB
> [snip]
> AnonHugePages: 0 kB
> ShmemPmdMapped: 557056 kB
> [snip]
> Locked: 0 kB
>
> cat /sys/kernel/debug/kvm/largepages
> 271
>
> And the benchmarks are as same as anonymous THPs.
>
Fixes: dd78fedde4b9 ("rmap: support file thp")
> Signed-off-by: Yang Shi <[email protected]>
> Reported-by: Gang Deng <[email protected]>
> Tested-by: Gang Deng <[email protected]>
> Suggested-by: Hugh Dickins <[email protected]>
> Cc: Andrea Arcangeli <[email protected]>
> Cc: Kirill A. Shutemov <[email protected]>
> Cc: <[email protected]> 4.8+
> ---
> v2: Adopted the suggestion from Hugh to use _mapcount and compound_mapcount.
> But I just open coding compound_mapcount to avoid duplicating the
> definition of compound_mapcount_ptr in two different files. Since
> "compound_mapcount" looks self-explained so I'm supposed the open
> coding would not affect the readability.
No, relying on head[1] is not nice: Matthew's suggestion better.
>
> include/linux/page-flags.h | 22 ++++++++++++++++++++--
> 1 file changed, 20 insertions(+), 2 deletions(-)
>
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index f91cb88..954a877 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -622,12 +622,30 @@ static inline int PageTransCompound(struct page *page)
> *
> * Unlike PageTransCompound, this is safe to be called only while
> * split_huge_pmd() cannot run from under us, like if protected by the
> - * MMU notifier, otherwise it may result in page->_mapcount < 0 false
> + * MMU notifier, otherwise it may result in page->_mapcount check false
> * positives.
> + *
> + * We have to treat page cache THP differently since every subpage of it
> + * would get _mapcount inc'ed once it is PMD mapped. But, it may be PTE
> + * mapped in the current process so comparing subpage's _mapcount to
> + * compound_mapcount ot filter out PTE mapped case.
s/ ot / to /
> */
> static inline int PageTransCompoundMap(struct page *page)
> {
> - return PageTransCompound(page) && atomic_read(&page->_mapcount) < 0;
> + struct page *head;
> + int map_count;
> +
> + if (!PageTransCompound(page))
> + return 0;
> +
> + if (PageAnon(page))
> + return atomic_read(&page->_mapcount) < 0;
> +
> + head = compound_head(page);
> + map_count = atomic_read(&page->_mapcount);
> + /* File THP is PMD mapped and not double mapped */
s/ double / PTE /
> + return map_count >= 0 &&
You have added a map_count >= 0 test there. Okay, not wrong, but not
necessary, and not consistent with what's returned in the PageAnon
case (if this were called for an unmapped page).
But asking for consistency in this function is asking for too much.
It is *very* specialized to the particular places from which it is
called (doesn't really belong in page-flags.h at all), and just so
long as it gives them the right answer most of the time, and errs
on the safe side the rest of the time, it'll do.
(I don't know if it's ever called on a tail page: it would not crash
if it were, but different tail pages may give different answers.)
> + map_count == atomic_read(&head[1].compound_mapcount);
> }
>
> /*
> --
> 1.8.3.1
>
>
On 10/23/19 12:28 PM, Hugh Dickins wrote:
> On Thu, 24 Oct 2019, Yang Shi wrote:
>
>> We have usecase to use tmpfs as QEMU memory backend and we would like to
>> take the advantage of THP as well. But, our test shows the EPT is not
>> PMD mapped even though the underlying THP are PMD mapped on host.
>> The number showed by /sys/kernel/debug/kvm/largepage is much less than
>> the number of PMD mapped shmem pages as the below:
>>
>> 7f2778200000-7f2878200000 rw-s 00000000 00:14 262232 /dev/shm/qemu_back_mem.mem.Hz2hSf (deleted)
>> Size: 4194304 kB
>> [snip]
>> AnonHugePages: 0 kB
>> ShmemPmdMapped: 579584 kB
>> [snip]
>> Locked: 0 kB
>>
>> cat /sys/kernel/debug/kvm/largepages
>> 12
>>
>> And some benchmarks do worse than with anonymous THPs.
>>
>> By digging into the code we figured out that commit 127393fbe597 ("mm:
>> thp: kvm: fix memory corruption in KVM with THP enabled") checks if
>> there is a single PTE mapping on the page for anonymous THP when
>> setting up EPT map. But, the _mapcount < 0 check doesn't fit to page
>> cache THP since every subpage of page cache THP would get _mapcount
>> inc'ed once it is PMD mapped, so PageTransCompoundMap() always returns
>> false for page cache THP. This would prevent KVM from setting up PMD
>> mapped EPT entry.
>>
>> So we need handle page cache THP correctly. However, when page cache
>> THP's PMD gets split, kernel just remove the map instead of setting up
>> PTE map like what anonymous THP does. Before KVM calls get_user_pages()
>> the subpages may get PTE mapped even though it is still a THP since the
>> page cache THP may be mapped by other processes at the mean time.
>>
>> Checking its _mapcount and whether the THP has PTE mapped or not.
>> Although this may report some false negative cases (PTE mapped by other
>> processes), it looks not trivial to make this accurate.
>>
>> With this fix /sys/kernel/debug/kvm/largepage would show reasonable
>> pages are PMD mapped by EPT as the below:
>>
>> 7fbeaee00000-7fbfaee00000 rw-s 00000000 00:14 275464 /dev/shm/qemu_back_mem.mem.SKUvat (deleted)
>> Size: 4194304 kB
>> [snip]
>> AnonHugePages: 0 kB
>> ShmemPmdMapped: 557056 kB
>> [snip]
>> Locked: 0 kB
>>
>> cat /sys/kernel/debug/kvm/largepages
>> 271
>>
>> And the benchmarks are as same as anonymous THPs.
>>
> Fixes: dd78fedde4b9 ("rmap: support file thp")
OK, though it might be the best blame target.
>
>> Signed-off-by: Yang Shi <[email protected]>
>> Reported-by: Gang Deng <[email protected]>
>> Tested-by: Gang Deng <[email protected]>
>> Suggested-by: Hugh Dickins <[email protected]>
>> Cc: Andrea Arcangeli <[email protected]>
>> Cc: Kirill A. Shutemov <[email protected]>
>> Cc: <[email protected]> 4.8+
>> ---
>> v2: Adopted the suggestion from Hugh to use _mapcount and compound_mapcount.
>> But I just open coding compound_mapcount to avoid duplicating the
>> definition of compound_mapcount_ptr in two different files. Since
>> "compound_mapcount" looks self-explained so I'm supposed the open
>> coding would not affect the readability.
> No, relying on head[1] is not nice: Matthew's suggestion better.
Done in v3.
>
>> include/linux/page-flags.h | 22 ++++++++++++++++++++--
>> 1 file changed, 20 insertions(+), 2 deletions(-)
>>
>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> index f91cb88..954a877 100644
>> --- a/include/linux/page-flags.h
>> +++ b/include/linux/page-flags.h
>> @@ -622,12 +622,30 @@ static inline int PageTransCompound(struct page *page)
>> *
>> * Unlike PageTransCompound, this is safe to be called only while
>> * split_huge_pmd() cannot run from under us, like if protected by the
>> - * MMU notifier, otherwise it may result in page->_mapcount < 0 false
>> + * MMU notifier, otherwise it may result in page->_mapcount check false
>> * positives.
>> + *
>> + * We have to treat page cache THP differently since every subpage of it
>> + * would get _mapcount inc'ed once it is PMD mapped. But, it may be PTE
>> + * mapped in the current process so comparing subpage's _mapcount to
>> + * compound_mapcount ot filter out PTE mapped case.
> s/ ot / to /
oops, will fix it.
>
>> */
>> static inline int PageTransCompoundMap(struct page *page)
>> {
>> - return PageTransCompound(page) && atomic_read(&page->_mapcount) < 0;
>> + struct page *head;
>> + int map_count;
>> +
>> + if (!PageTransCompound(page))
>> + return 0;
>> +
>> + if (PageAnon(page))
>> + return atomic_read(&page->_mapcount) < 0;
>> +
>> + head = compound_head(page);
>> + map_count = atomic_read(&page->_mapcount);
>> + /* File THP is PMD mapped and not double mapped */
> s/ double / PTE /
Will fix it.
>
>> + return map_count >= 0 &&
> You have added a map_count >= 0 test there. Okay, not wrong, but not
> necessary, and not consistent with what's returned in the PageAnon
> case (if this were called for an unmapped page).
I was thinking about this too. I'm wondering there might be a case that
the PMD is split and it was the last PMD map, in this case subpage's
_mapcount is also equal to compound_mapcount (both is -1). So, it would
return true, then KVM may setup PMD map in EPT, but it might be PTE
mapped later on the host. But, I'm not quite sure if this is really
possible or if this is really a integrity problem. So, I thought it
might be safer to add this check.
>
> But asking for consistency in this function is asking for too much.
> It is *very* specialized to the particular places from which it is
> called (doesn't really belong in page-flags.h at all), and just so
> long as it gives them the right answer most of the time, and errs
> on the safe side the rest of the time, it'll do.
>
> (I don't know if it's ever called on a tail page: it would not crash
> if it were, but different tail pages may give different answers.)
>
>> + map_count == atomic_read(&head[1].compound_mapcount);
>> }
>>
>> /*
>> --
>> 1.8.3.1
>>
>>
On 10/23/19 1:00 PM, Hugh Dickins wrote:
> On Wed, 23 Oct 2019, Yang Shi wrote:
>> On 10/23/19 12:28 PM, Hugh Dickins wrote:
>>>> + return map_count >= 0 &&
>>> You have added a map_count >= 0 test there. Okay, not wrong, but not
>>> necessary, and not consistent with what's returned in the PageAnon
>>> case (if this were called for an unmapped page).
>> I was thinking about this too. I'm wondering there might be a case that the
>> PMD is split and it was the last PMD map, in this case subpage's _mapcount is
>> also equal to compound_mapcount (both is -1). So, it would return true, then
>> KVM may setup PMD map in EPT, but it might be PTE mapped later on the host.
>> But, I'm not quite sure if this is really possible or if this is really a
>> integrity problem. So, I thought it might be safer to add this check.
> The mmu_notifier_invalidate_range_start.._end() in __split_huge_pmd(),
> with KVM's locking and sequence counting, is required to protect
> against such races.
OK, it sounds safe. Thanks for confirming. Will post v4 soon.
>
> Hugh