diff mbox series

[v8,1/9] KVM: arm/arm64: Ensure only THP is candidate for adjustment

Message ID 20181001155443.23032-2-punit.agrawal@arm.com (mailing list archive)
State Superseded, archived
Headers show
Series KVM: Support PUD hugepage at stage 2 | expand

Commit Message

Punit Agrawal Oct. 1, 2018, 3:54 p.m. UTC
PageTransCompoundMap() returns true for hugetlbfs and THP
hugepages. This behaviour incorrectly leads to stage 2 faults for
unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
treated as THP faults.

Tighten the check to filter out hugetlbfs pages. This also leads to
consistently mapping all unsupported hugepage sizes as PTE level
entries at stage 2.

Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
Cc: Christoffer Dall <christoffer.dall@arm.com>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Cc: stable@vger.kernel.org # v4.13+
---
 virt/kvm/arm/mmu.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

Comments

Marc Zyngier Oct. 3, 2018, 10:50 a.m. UTC | #1
On 01/10/18 16:54, Punit Agrawal wrote:
> PageTransCompoundMap() returns true for hugetlbfs and THP
> hugepages. This behaviour incorrectly leads to stage 2 faults for
> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
> treated as THP faults.
> 
> Tighten the check to filter out hugetlbfs pages. This also leads to
> consistently mapping all unsupported hugepage sizes as PTE level
> entries at stage 2.
> 
> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
> Cc: Christoffer Dall <christoffer.dall@arm.com>
> Cc: Marc Zyngier <marc.zyngier@arm.com>
> Cc: stable@vger.kernel.org # v4.13+

FWIW, I've cherry-picked that single patch from the series and queued it 
for 4.20.

Thanks,

	M.
Punit Agrawal Oct. 3, 2018, 10:59 a.m. UTC | #2
Marc Zyngier <marc.zyngier@arm.com> writes:

> On 01/10/18 16:54, Punit Agrawal wrote:
>> PageTransCompoundMap() returns true for hugetlbfs and THP
>> hugepages. This behaviour incorrectly leads to stage 2 faults for
>> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
>> treated as THP faults.
>>
>> Tighten the check to filter out hugetlbfs pages. This also leads to
>> consistently mapping all unsupported hugepage sizes as PTE level
>> entries at stage 2.
>>
>> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
>> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
>> Cc: Christoffer Dall <christoffer.dall@arm.com>
>> Cc: Marc Zyngier <marc.zyngier@arm.com>
>> Cc: stable@vger.kernel.org # v4.13+
>
> FWIW, I've cherry-picked that single patch from the series and queued
> it for 4.20.

Thanks for picking up the fix.
Christoffer Dall Oct. 31, 2018, 2:36 p.m. UTC | #3
On Mon, Oct 01, 2018 at 04:54:35PM +0100, Punit Agrawal wrote:
> PageTransCompoundMap() returns true for hugetlbfs and THP
> hugepages. This behaviour incorrectly leads to stage 2 faults for
> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
> treated as THP faults.
> 
> Tighten the check to filter out hugetlbfs pages. This also leads to
> consistently mapping all unsupported hugepage sizes as PTE level
> entries at stage 2.
> 
> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
> Cc: Christoffer Dall <christoffer.dall@arm.com>
> Cc: Marc Zyngier <marc.zyngier@arm.com>
> Cc: stable@vger.kernel.org # v4.13+


Hmm, this function is only actually called from user_mem_abort() if we
have (!hugetlb), so I'm not sure the cc stable here was actually
warranted, nor that this patch is strictly necessary.

It doesn't hurt, and makes the code potentially more robust for the
future though.

Am I missing something?

Thanks,

    Christoffer

> ---
>  virt/kvm/arm/mmu.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index 7e477b3cae5b..c23a1b323aad 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -1231,8 +1231,14 @@ static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
>  {
>  	kvm_pfn_t pfn = *pfnp;
>  	gfn_t gfn = *ipap >> PAGE_SHIFT;
> +	struct page *page = pfn_to_page(pfn);
>  
> -	if (PageTransCompoundMap(pfn_to_page(pfn))) {
> +	/*
> +	 * PageTransCompoungMap() returns true for THP and
> +	 * hugetlbfs. Make sure the adjustment is done only for THP
> +	 * pages.
> +	 */
> +	if (!PageHuge(page) && PageTransCompoundMap(page)) {
>  		unsigned long mask;
>  		/*
>  		 * The address we faulted on is backed by a transparent huge
> -- 
> 2.18.0
>
Punit Agrawal Oct. 31, 2018, 2:52 p.m. UTC | #4
Christoffer Dall <christoffer.dall@arm.com> writes:

> On Mon, Oct 01, 2018 at 04:54:35PM +0100, Punit Agrawal wrote:
>> PageTransCompoundMap() returns true for hugetlbfs and THP
>> hugepages. This behaviour incorrectly leads to stage 2 faults for
>> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
>> treated as THP faults.
>> 
>> Tighten the check to filter out hugetlbfs pages. This also leads to
>> consistently mapping all unsupported hugepage sizes as PTE level
>> entries at stage 2.
>> 
>> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
>> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
>> Cc: Christoffer Dall <christoffer.dall@arm.com>
>> Cc: Marc Zyngier <marc.zyngier@arm.com>
>> Cc: stable@vger.kernel.org # v4.13+
>
>
> Hmm, this function is only actually called from user_mem_abort() if we
> have (!hugetlb), so I'm not sure the cc stable here was actually
> warranted, nor that this patch is strictly necessary.
>
> It doesn't hurt, and makes the code potentially more robust for the
> future though.
>
> Am I missing something?

!hugetlb is only true for hugepage sizes supported at stage 2. The
function also got called for unsupported hugepage size at stage 2, e.g.,
64k hugepage with 4k page size, which then ended up doing the wrong
thing.

Hope that adds some context. I should've added this to the commit log.

>
> Thanks,
>
>     Christoffer
>
>> ---
>>  virt/kvm/arm/mmu.c | 8 +++++++-
>>  1 file changed, 7 insertions(+), 1 deletion(-)
>> 
>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>> index 7e477b3cae5b..c23a1b323aad 100644
>> --- a/virt/kvm/arm/mmu.c
>> +++ b/virt/kvm/arm/mmu.c
>> @@ -1231,8 +1231,14 @@ static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
>>  {
>>  	kvm_pfn_t pfn = *pfnp;
>>  	gfn_t gfn = *ipap >> PAGE_SHIFT;
>> +	struct page *page = pfn_to_page(pfn);
>>  
>> -	if (PageTransCompoundMap(pfn_to_page(pfn))) {
>> +	/*
>> +	 * PageTransCompoungMap() returns true for THP and
>> +	 * hugetlbfs. Make sure the adjustment is done only for THP
>> +	 * pages.
>> +	 */
>> +	if (!PageHuge(page) && PageTransCompoundMap(page)) {
>>  		unsigned long mask;
>>  		/*
>>  		 * The address we faulted on is backed by a transparent huge
>> -- 
>> 2.18.0
>>
Punit Agrawal Oct. 31, 2018, 5:15 p.m. UTC | #5
Punit Agrawal <punit.agrawal@arm.com> writes:

> Christoffer Dall <christoffer.dall@arm.com> writes:
>
>> On Mon, Oct 01, 2018 at 04:54:35PM +0100, Punit Agrawal wrote:
>>> PageTransCompoundMap() returns true for hugetlbfs and THP
>>> hugepages. This behaviour incorrectly leads to stage 2 faults for
>>> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
>>> treated as THP faults.
>>> 
>>> Tighten the check to filter out hugetlbfs pages. This also leads to
>>> consistently mapping all unsupported hugepage sizes as PTE level
>>> entries at stage 2.
>>> 
>>> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
>>> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
>>> Cc: Christoffer Dall <christoffer.dall@arm.com>
>>> Cc: Marc Zyngier <marc.zyngier@arm.com>
>>> Cc: stable@vger.kernel.org # v4.13+
>>
>>
>> Hmm, this function is only actually called from user_mem_abort() if we
>> have (!hugetlb), so I'm not sure the cc stable here was actually
>> warranted, nor that this patch is strictly necessary.
>>
>> It doesn't hurt, and makes the code potentially more robust for the
>> future though.
>>
>> Am I missing something?
>
> !hugetlb is only true for hugepage sizes supported at stage 2. 

Of course I meant "hugetlb" above (Note the lack of "!").

> The function also got called for unsupported hugepage size at stage 2,
> e.g., 64k hugepage with 4k page size, which then ended up doing the
> wrong thing.
>
> Hope that adds some context. I should've added this to the commit log.
>
>>
>> Thanks,
>>
>>     Christoffer
>>
>>> ---
>>>  virt/kvm/arm/mmu.c | 8 +++++++-
>>>  1 file changed, 7 insertions(+), 1 deletion(-)
>>> 
>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>> index 7e477b3cae5b..c23a1b323aad 100644
>>> --- a/virt/kvm/arm/mmu.c
>>> +++ b/virt/kvm/arm/mmu.c
>>> @@ -1231,8 +1231,14 @@ static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
>>>  {
>>>  	kvm_pfn_t pfn = *pfnp;
>>>  	gfn_t gfn = *ipap >> PAGE_SHIFT;
>>> +	struct page *page = pfn_to_page(pfn);
>>>  
>>> -	if (PageTransCompoundMap(pfn_to_page(pfn))) {
>>> +	/*
>>> +	 * PageTransCompoungMap() returns true for THP and
>>> +	 * hugetlbfs. Make sure the adjustment is done only for THP
>>> +	 * pages.
>>> +	 */
>>> +	if (!PageHuge(page) && PageTransCompoundMap(page)) {
>>>  		unsigned long mask;
>>>  		/*
>>>  		 * The address we faulted on is backed by a transparent huge
>>> -- 
>>> 2.18.0
>>> 
> _______________________________________________
> kvmarm mailing list
> kvmarm@lists.cs.columbia.edu
> https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Christoffer Dall Nov. 1, 2018, 8:31 a.m. UTC | #6
On Wed, Oct 31, 2018 at 02:52:20PM +0000, Punit Agrawal wrote:
> Christoffer Dall <christoffer.dall@arm.com> writes:
> 
> > On Mon, Oct 01, 2018 at 04:54:35PM +0100, Punit Agrawal wrote:
> >> PageTransCompoundMap() returns true for hugetlbfs and THP
> >> hugepages. This behaviour incorrectly leads to stage 2 faults for
> >> unsupported hugepage sizes (e.g., 64K hugepage with 4K pages) to be
> >> treated as THP faults.
> >> 
> >> Tighten the check to filter out hugetlbfs pages. This also leads to
> >> consistently mapping all unsupported hugepage sizes as PTE level
> >> entries at stage 2.
> >> 
> >> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
> >> Reviewed-by: Suzuki Poulose <suzuki.poulose@arm.com>
> >> Cc: Christoffer Dall <christoffer.dall@arm.com>
> >> Cc: Marc Zyngier <marc.zyngier@arm.com>
> >> Cc: stable@vger.kernel.org # v4.13+
> >
> >
> > Hmm, this function is only actually called from user_mem_abort() if we
> > have (!hugetlb), so I'm not sure the cc stable here was actually
> > warranted, nor that this patch is strictly necessary.
> >
> > It doesn't hurt, and makes the code potentially more robust for the
> > future though.
> >
> > Am I missing something?
> 
> !hugetlb is only true for hugepage sizes supported at stage 2. The
> function also got called for unsupported hugepage size at stage 2, e.g.,
> 64k hugepage with 4k page size, which then ended up doing the wrong
> thing.
> 
> Hope that adds some context. I should've added this to the commit log.
> 

To be fair you did say that this was for unsupported hugepage sizes.

Thanks for the explanation.


    Christoffer
diff mbox series

Patch

diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
index 7e477b3cae5b..c23a1b323aad 100644
--- a/virt/kvm/arm/mmu.c
+++ b/virt/kvm/arm/mmu.c
@@ -1231,8 +1231,14 @@  static bool transparent_hugepage_adjust(kvm_pfn_t *pfnp, phys_addr_t *ipap)
 {
 	kvm_pfn_t pfn = *pfnp;
 	gfn_t gfn = *ipap >> PAGE_SHIFT;
+	struct page *page = pfn_to_page(pfn);
 
-	if (PageTransCompoundMap(pfn_to_page(pfn))) {
+	/*
+	 * PageTransCompoungMap() returns true for THP and
+	 * hugetlbfs. Make sure the adjustment is done only for THP
+	 * pages.
+	 */
+	if (!PageHuge(page) && PageTransCompoundMap(page)) {
 		unsigned long mask;
 		/*
 		 * The address we faulted on is backed by a transparent huge