There is already support of enabling dirty log graually in small chunks
for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
small chunks"). This adds support for arm64.
x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
is eanbled. However, for arm64, both huge pages and normal pages can be
write protected gradually by userspace.
Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
Linux VMs with different page size. The memory pressure is 127G in each
case. The time taken of memory_global_dirty_log_start in QEMU is listed
below:
Page Size Before After Optimization
4K 650ms 1.8ms
2M 4ms 1.8ms
1G 2ms 1.8ms
Besides the time reduction, the biggest income is that we will minimize
the performance side effect (because of dissloving huge pages and marking
memslots dirty) on guest after enabling dirty log.
Signed-off-by: Keqian Zhu <[email protected]>
---
Documentation/virt/kvm/api.rst | 2 +-
arch/arm64/include/asm/kvm_host.h | 3 +++
virt/kvm/arm/mmu.c | 12 ++++++++++--
3 files changed, 14 insertions(+), 3 deletions(-)
diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index efbbe570aa9b..0017f63fa44f 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
dirty logging can be enabled gradually in small chunks on the first call
to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
-x86 for now).
+x86 and arm64 for now).
KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index 32c8a675e5a4..a723f84fab83 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h
@@ -46,6 +46,9 @@
#define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
#define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
+#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
+ KVM_DIRTY_LOG_INITIALLY_SET)
+
DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
extern unsigned int kvm_sve_max_vl;
diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
index e3b9ee268823..1077f653a611 100644
--- a/virt/kvm/arm/mmu.c
+++ b/virt/kvm/arm/mmu.c
@@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
* allocated dirty_bitmap[], dirty pages will be be tracked while the
* memory slot is write protected.
*/
- if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
- kvm_mmu_wp_memory_region(kvm, mem->slot);
+ if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
+ /*
+ * If we're with initial-all-set, we don't need to write
+ * protect any pages because they're all reported as dirty.
+ * Huge pages and normal pages will be write protect gradually.
+ */
+ if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
+ kvm_mmu_wp_memory_region(kvm, mem->slot);
+ }
+ }
}
int kvm_arch_prepare_memory_region(struct kvm *kvm,
--
2.19.1
On 13/04/20 14:20, Keqian Zhu wrote:
> There is already support of enabling dirty log graually in small chunks
> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
> small chunks"). This adds support for arm64.
>
> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
> is eanbled. However, for arm64, both huge pages and normal pages can be
> write protected gradually by userspace.
>
> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
> Linux VMs with different page size. The memory pressure is 127G in each
> case. The time taken of memory_global_dirty_log_start in QEMU is listed
> below:
>
> Page Size Before After Optimization
> 4K 650ms 1.8ms
> 2M 4ms 1.8ms
> 1G 2ms 1.8ms
>
> Besides the time reduction, the biggest income is that we will minimize
> the performance side effect (because of dissloving huge pages and marking
> memslots dirty) on guest after enabling dirty log.
>
> Signed-off-by: Keqian Zhu <[email protected]>
> ---
> Documentation/virt/kvm/api.rst | 2 +-
> arch/arm64/include/asm/kvm_host.h | 3 +++
> virt/kvm/arm/mmu.c | 12 ++++++++++--
> 3 files changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
> index efbbe570aa9b..0017f63fa44f 100644
> --- a/Documentation/virt/kvm/api.rst
> +++ b/Documentation/virt/kvm/api.rst
> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
> dirty logging can be enabled gradually in small chunks on the first call
> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
> -x86 for now).
> +x86 and arm64 for now).
>
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 32c8a675e5a4..a723f84fab83 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -46,6 +46,9 @@
> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>
> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
> + KVM_DIRTY_LOG_INITIALLY_SET)
> +
> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>
> extern unsigned int kvm_sve_max_vl;
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index e3b9ee268823..1077f653a611 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
> * allocated dirty_bitmap[], dirty pages will be be tracked while the
> * memory slot is write protected.
> */
> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
> - kvm_mmu_wp_memory_region(kvm, mem->slot);
> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
> + /*
> + * If we're with initial-all-set, we don't need to write
> + * protect any pages because they're all reported as dirty.
> + * Huge pages and normal pages will be write protect gradually.
> + */
> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
> + kvm_mmu_wp_memory_region(kvm, mem->slot);
> + }
> + }
> }
>
> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>
Marc, what is the status of this patch?
Paolo
Hi Marc,
In RFC patch, I still write protect huge pages when DIRTY_LOG_INITIALLY_ALL_SET
is enabled by userspace. I find that both huge pages and normal pages can be
write protected during log clear. So this formal patch is pretty simple now.
Thanks,
Keqian
On 2020/4/13 20:20, Keqian Zhu wrote:
> There is already support of enabling dirty log graually in small chunks
> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
> small chunks"). This adds support for arm64.
>
> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
> is eanbled. However, for arm64, both huge pages and normal pages can be
> write protected gradually by userspace.
>
> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
> Linux VMs with different page size. The memory pressure is 127G in each
> case. The time taken of memory_global_dirty_log_start in QEMU is listed
> below:
>
> Page Size Before After Optimization
> 4K 650ms 1.8ms
> 2M 4ms 1.8ms
> 1G 2ms 1.8ms
>
> Besides the time reduction, the biggest income is that we will minimize
> the performance side effect (because of dissloving huge pages and marking
> memslots dirty) on guest after enabling dirty log.
>
> Signed-off-by: Keqian Zhu <[email protected]>
> ---
> Documentation/virt/kvm/api.rst | 2 +-
> arch/arm64/include/asm/kvm_host.h | 3 +++
> virt/kvm/arm/mmu.c | 12 ++++++++++--
> 3 files changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
> index efbbe570aa9b..0017f63fa44f 100644
> --- a/Documentation/virt/kvm/api.rst
> +++ b/Documentation/virt/kvm/api.rst
> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
> dirty logging can be enabled gradually in small chunks on the first call
> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
> -x86 for now).
> +x86 and arm64 for now).
>
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 32c8a675e5a4..a723f84fab83 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -46,6 +46,9 @@
> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>
> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
> + KVM_DIRTY_LOG_INITIALLY_SET)
> +
> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>
> extern unsigned int kvm_sve_max_vl;
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index e3b9ee268823..1077f653a611 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
> * allocated dirty_bitmap[], dirty pages will be be tracked while the
> * memory slot is write protected.
> */
> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
> - kvm_mmu_wp_memory_region(kvm, mem->slot);
> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
> + /*
> + * If we're with initial-all-set, we don't need to write
> + * protect any pages because they're all reported as dirty.
> + * Huge pages and normal pages will be write protect gradually.
> + */
> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
> + kvm_mmu_wp_memory_region(kvm, mem->slot);
> + }
> + }
> }
>
> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>
On Mon, 13 Apr 2020 20:20:23 +0800
Keqian Zhu <[email protected]> wrote:
> There is already support of enabling dirty log graually in small chunks
gradually
> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
> small chunks"). This adds support for arm64.
>
> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
> is eanbled. However, for arm64, both huge pages and normal pages can be
enabled
> write protected gradually by userspace.
>
> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
> Linux VMs with different page size. The memory pressure is 127G in each
> case. The time taken of memory_global_dirty_log_start in QEMU is listed
> below:
>
> Page Size Before After Optimization
> 4K 650ms 1.8ms
> 2M 4ms 1.8ms
> 1G 2ms 1.8ms
These numbers are different from what you have advertised before. What
changed?
>
> Besides the time reduction, the biggest income is that we will minimize
s/income/improvement/
> the performance side effect (because of dissloving huge pages and marking
dissolving
> memslots dirty) on guest after enabling dirty log.
>
> Signed-off-by: Keqian Zhu <[email protected]>
> ---
> Documentation/virt/kvm/api.rst | 2 +-
> arch/arm64/include/asm/kvm_host.h | 3 +++
> virt/kvm/arm/mmu.c | 12 ++++++++++--
> 3 files changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
> index efbbe570aa9b..0017f63fa44f 100644
> --- a/Documentation/virt/kvm/api.rst
> +++ b/Documentation/virt/kvm/api.rst
> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
> dirty logging can be enabled gradually in small chunks on the first call
> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
> -x86 for now).
> +x86 and arm64 for now).
>
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 32c8a675e5a4..a723f84fab83 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -46,6 +46,9 @@
> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>
> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
> + KVM_DIRTY_LOG_INITIALLY_SET)
> +
> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>
> extern unsigned int kvm_sve_max_vl;
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index e3b9ee268823..1077f653a611 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
> * allocated dirty_bitmap[], dirty pages will be be tracked while the
> * memory slot is write protected.
> */
> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
> - kvm_mmu_wp_memory_region(kvm, mem->slot);
> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
> + /*
> + * If we're with initial-all-set, we don't need to write
> + * protect any pages because they're all reported as dirty.
> + * Huge pages and normal pages will be write protect gradually.
> + */
> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
> + kvm_mmu_wp_memory_region(kvm, mem->slot);
> + }
> + }
> }
>
> int kvm_arch_prepare_memory_region(struct kvm *kvm,
As it is, it is pretty good. The one thing that isn't clear to me is
why we have a difference in behaviour between x86 and arm64. What
prevents x86 from having the same behaviour as arm64?
Thanks,
M.
--
Jazz is not dead. It just smells funny...
On Wed, 15 Apr 2020 18:13:56 +0200
Paolo Bonzini <[email protected]> wrote:
> On 13/04/20 14:20, Keqian Zhu wrote:
> > There is already support of enabling dirty log graually in small chunks
> > for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
> > small chunks"). This adds support for arm64.
> >
> > x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
> > is eanbled. However, for arm64, both huge pages and normal pages can be
> > write protected gradually by userspace.
> >
> > Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
> > Linux VMs with different page size. The memory pressure is 127G in each
> > case. The time taken of memory_global_dirty_log_start in QEMU is listed
> > below:
> >
> > Page Size Before After Optimization
> > 4K 650ms 1.8ms
> > 2M 4ms 1.8ms
> > 1G 2ms 1.8ms
> >
> > Besides the time reduction, the biggest income is that we will minimize
> > the performance side effect (because of dissloving huge pages and marking
> > memslots dirty) on guest after enabling dirty log.
> >
> > Signed-off-by: Keqian Zhu <[email protected]>
> > ---
> > Documentation/virt/kvm/api.rst | 2 +-
> > arch/arm64/include/asm/kvm_host.h | 3 +++
> > virt/kvm/arm/mmu.c | 12 ++++++++++--
> > 3 files changed, 14 insertions(+), 3 deletions(-)
> >
> > diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
> > index efbbe570aa9b..0017f63fa44f 100644
> > --- a/Documentation/virt/kvm/api.rst
> > +++ b/Documentation/virt/kvm/api.rst
> > @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
> > dirty logging can be enabled gradually in small chunks on the first call
> > to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
> > KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
> > -x86 for now).
> > +x86 and arm64 for now).
> >
> > KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
> > KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
> > diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> > index 32c8a675e5a4..a723f84fab83 100644
> > --- a/arch/arm64/include/asm/kvm_host.h
> > +++ b/arch/arm64/include/asm/kvm_host.h
> > @@ -46,6 +46,9 @@
> > #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
> > #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
> >
> > +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
> > + KVM_DIRTY_LOG_INITIALLY_SET)
> > +
> > DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
> >
> > extern unsigned int kvm_sve_max_vl;
> > diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> > index e3b9ee268823..1077f653a611 100644
> > --- a/virt/kvm/arm/mmu.c
> > +++ b/virt/kvm/arm/mmu.c
> > @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
> > * allocated dirty_bitmap[], dirty pages will be be tracked while the
> > * memory slot is write protected.
> > */
> > - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
> > - kvm_mmu_wp_memory_region(kvm, mem->slot);
> > + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
> > + /*
> > + * If we're with initial-all-set, we don't need to write
> > + * protect any pages because they're all reported as dirty.
> > + * Huge pages and normal pages will be write protect gradually.
> > + */
> > + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
> > + kvm_mmu_wp_memory_region(kvm, mem->slot);
> > + }
> > + }
> > }
> >
> > int kvm_arch_prepare_memory_region(struct kvm *kvm,
> >
>
> Marc, what is the status of this patch?
I just had a look at it. Is there any urgency for merging it?
Thanks,
M.
--
Jazz is not dead. It just smells funny...
On 16/04/20 17:09, Marc Zyngier wrote:
> On Wed, 15 Apr 2020 18:13:56 +0200
> Paolo Bonzini <[email protected]> wrote:
>
>> On 13/04/20 14:20, Keqian Zhu wrote:
>>> There is already support of enabling dirty log graually in small chunks
>>> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
>>> small chunks"). This adds support for arm64.
>>>
>>> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
>>> is eanbled. However, for arm64, both huge pages and normal pages can be
>>> write protected gradually by userspace.
>>>
>>> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
>>> Linux VMs with different page size. The memory pressure is 127G in each
>>> case. The time taken of memory_global_dirty_log_start in QEMU is listed
>>> below:
>>>
>>> Page Size Before After Optimization
>>> 4K 650ms 1.8ms
>>> 2M 4ms 1.8ms
>>> 1G 2ms 1.8ms
>>>
>>> Besides the time reduction, the biggest income is that we will minimize
>>> the performance side effect (because of dissloving huge pages and marking
>>> memslots dirty) on guest after enabling dirty log.
>>>
>>> Signed-off-by: Keqian Zhu <[email protected]>
>>> ---
>>> Documentation/virt/kvm/api.rst | 2 +-
>>> arch/arm64/include/asm/kvm_host.h | 3 +++
>>> virt/kvm/arm/mmu.c | 12 ++++++++++--
>>> 3 files changed, 14 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
>>> index efbbe570aa9b..0017f63fa44f 100644
>>> --- a/Documentation/virt/kvm/api.rst
>>> +++ b/Documentation/virt/kvm/api.rst
>>> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
>>> dirty logging can be enabled gradually in small chunks on the first call
>>> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
>>> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
>>> -x86 for now).
>>> +x86 and arm64 for now).
>>>
>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
>>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
>>> index 32c8a675e5a4..a723f84fab83 100644
>>> --- a/arch/arm64/include/asm/kvm_host.h
>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>> @@ -46,6 +46,9 @@
>>> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
>>> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>>>
>>> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
>>> + KVM_DIRTY_LOG_INITIALLY_SET)
>>> +
>>> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>>>
>>> extern unsigned int kvm_sve_max_vl;
>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>> index e3b9ee268823..1077f653a611 100644
>>> --- a/virt/kvm/arm/mmu.c
>>> +++ b/virt/kvm/arm/mmu.c
>>> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
>>> * allocated dirty_bitmap[], dirty pages will be be tracked while the
>>> * memory slot is write protected.
>>> */
>>> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
>>> - kvm_mmu_wp_memory_region(kvm, mem->slot);
>>> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
>>> + /*
>>> + * If we're with initial-all-set, we don't need to write
>>> + * protect any pages because they're all reported as dirty.
>>> + * Huge pages and normal pages will be write protect gradually.
>>> + */
>>> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
>>> + kvm_mmu_wp_memory_region(kvm, mem->slot);
>>> + }
>>> + }
>>> }
>>>
>>> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>>>
>>
>> Marc, what is the status of this patch?
>
> I just had a look at it. Is there any urgency for merging it?
No, I thought I was still replying to the v1.
Paolo
Hi Paolo,
On 2020/4/16 23:55, Paolo Bonzini wrote:
> On 16/04/20 17:09, Marc Zyngier wrote:
>> On Wed, 15 Apr 2020 18:13:56 +0200
>> Paolo Bonzini <[email protected]> wrote:
>>
>>> On 13/04/20 14:20, Keqian Zhu wrote:
>>>> There is already support of enabling dirty log graually in small chunks
>>>> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
>>>> small chunks"). This adds support for arm64.
>>>>
>>>> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
>>>> is eanbled. However, for arm64, both huge pages and normal pages can be
>>>> write protected gradually by userspace.
>>>>
>>>> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
>>>> Linux VMs with different page size. The memory pressure is 127G in each
>>>> case. The time taken of memory_global_dirty_log_start in QEMU is listed
>>>> below:
>>>>
>>>> Page Size Before After Optimization
>>>> 4K 650ms 1.8ms
>>>> 2M 4ms 1.8ms
>>>> 1G 2ms 1.8ms
>>>>
>>>> Besides the time reduction, the biggest income is that we will minimize
>>>> the performance side effect (because of dissloving huge pages and marking
>>>> memslots dirty) on guest after enabling dirty log.
>>>>
>>>> Signed-off-by: Keqian Zhu <[email protected]>
>>>> ---
>>>> Documentation/virt/kvm/api.rst | 2 +-
>>>> arch/arm64/include/asm/kvm_host.h | 3 +++
>>>> virt/kvm/arm/mmu.c | 12 ++++++++++--
>>>> 3 files changed, 14 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
>>>> index efbbe570aa9b..0017f63fa44f 100644
>>>> --- a/Documentation/virt/kvm/api.rst
>>>> +++ b/Documentation/virt/kvm/api.rst
>>>> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
>>>> dirty logging can be enabled gradually in small chunks on the first call
>>>> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
>>>> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
>>>> -x86 for now).
>>>> +x86 and arm64 for now).
>>>>
>>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
>>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
>>>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
>>>> index 32c8a675e5a4..a723f84fab83 100644
>>>> --- a/arch/arm64/include/asm/kvm_host.h
>>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>>> @@ -46,6 +46,9 @@
>>>> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
>>>> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>>>>
>>>> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
>>>> + KVM_DIRTY_LOG_INITIALLY_SET)
>>>> +
>>>> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>>>>
>>>> extern unsigned int kvm_sve_max_vl;
>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>>> index e3b9ee268823..1077f653a611 100644
>>>> --- a/virt/kvm/arm/mmu.c
>>>> +++ b/virt/kvm/arm/mmu.c
>>>> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
>>>> * allocated dirty_bitmap[], dirty pages will be be tracked while the
>>>> * memory slot is write protected.
>>>> */
>>>> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
>>>> - kvm_mmu_wp_memory_region(kvm, mem->slot);
>>>> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
>>>> + /*
>>>> + * If we're with initial-all-set, we don't need to write
>>>> + * protect any pages because they're all reported as dirty.
>>>> + * Huge pages and normal pages will be write protect gradually.
>>>> + */
>>>> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
>>>> + kvm_mmu_wp_memory_region(kvm, mem->slot);
>>>> + }
>>>> + }
>>>> }
>>>>
>>>> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>>>>
>>>
>>> Marc, what is the status of this patch?
>>
>> I just had a look at it. Is there any urgency for merging it?
>
> No, I thought I was still replying to the v1.
Sorry that patch v1 is dropped. Because I realized that stage2 page tables
will be unmapped during VM reboot, or they are not established soon after
migration, so stage2 page tables can not be used to decide whether a page
is needed to migrate.
Thanks,
Keqian
>
> Paolo
>
>
> .
>
Hi Marc,
On 2020/4/16 23:08, Marc Zyngier wrote:
> On Mon, 13 Apr 2020 20:20:23 +0800
> Keqian Zhu <[email protected]> wrote:
>
>> There is already support of enabling dirty log graually in small chunks
>
> gradually
>
>> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually in
>> small chunks"). This adds support for arm64.
>>
>> x86 still writes protect all huge pages when DIRTY_LOG_INITIALLY_ALL_SET
>> is eanbled. However, for arm64, both huge pages and normal pages can be
>
> enabled
>
>> write protected gradually by userspace.
>>
>> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on 128G
>> Linux VMs with different page size. The memory pressure is 127G in each
>> case. The time taken of memory_global_dirty_log_start in QEMU is listed
>> below:
>>
>> Page Size Before After Optimization
>> 4K 650ms 1.8ms
>> 2M 4ms 1.8ms
>> 1G 2ms 1.8ms
>
> These numbers are different from what you have advertised before. What
> changed?
In patch RFC, the numbers is got when memory pressure is 100G, so the number
is bigger here.
>
>>
>> Besides the time reduction, the biggest income is that we will minimize
>
> s/income/improvement/
>
>> the performance side effect (because of dissloving huge pages and marking
>
> dissolving
embarrassed for these misspell :(
>
>> memslots dirty) on guest after enabling dirty log.
>>
>> Signed-off-by: Keqian Zhu <[email protected]>
>> ---
>> Documentation/virt/kvm/api.rst | 2 +-
>> arch/arm64/include/asm/kvm_host.h | 3 +++
>> virt/kvm/arm/mmu.c | 12 ++++++++++--
>> 3 files changed, 14 insertions(+), 3 deletions(-)
>>
>> diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
>> index efbbe570aa9b..0017f63fa44f 100644
>> --- a/Documentation/virt/kvm/api.rst
>> +++ b/Documentation/virt/kvm/api.rst
>> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This also improves performance because
>> dirty logging can be enabled gradually in small chunks on the first call
>> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
>> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
>> -x86 for now).
>> +x86 and arm64 for now).
>>
>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
>> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
>> index 32c8a675e5a4..a723f84fab83 100644
>> --- a/arch/arm64/include/asm/kvm_host.h
>> +++ b/arch/arm64/include/asm/kvm_host.h
>> @@ -46,6 +46,9 @@
>> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
>> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>>
>> +#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
>> + KVM_DIRTY_LOG_INITIALLY_SET)
>> +
>> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>>
>> extern unsigned int kvm_sve_max_vl;
>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>> index e3b9ee268823..1077f653a611 100644
>> --- a/virt/kvm/arm/mmu.c
>> +++ b/virt/kvm/arm/mmu.c
>> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
>> * allocated dirty_bitmap[], dirty pages will be be tracked while the
>> * memory slot is write protected.
>> */
>> - if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES)
>> - kvm_mmu_wp_memory_region(kvm, mem->slot);
>> + if (change != KVM_MR_DELETE && mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
>> + /*
>> + * If we're with initial-all-set, we don't need to write
>> + * protect any pages because they're all reported as dirty.
>> + * Huge pages and normal pages will be write protect gradually.
>> + */
>> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
>> + kvm_mmu_wp_memory_region(kvm, mem->slot);
>> + }
>> + }
>> }
>>
>> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>
> As it is, it is pretty good. The one thing that isn't clear to me is
> why we have a difference in behaviour between x86 and arm64. What
> prevents x86 from having the same behaviour as arm64?
I am also not very clear about the difference. Maybe there is historic reason.
Before introducing DIRTY_LOG_INITIALLY_ALL_SET, all pages will be write
protected when starting dirty log, but only normal pages are needed
to be write protected again during dirty log sync, because huge pages will
be dissolved to normal pages.
For that x86 uses different routine to write protect huge pages and normal pages,
and arm64 uses same routine to do this, so arm64 still write protect all
pages again during dirty log sync, but x86 didn't.
So I think that x86 can write protect huge pages gradually too, just need to add
some code legs in dirty log sync.
Thanks,
Keqian
>
> Thanks,
>
> M.
>
On 2020-04-17 10:46, zhukeqian wrote:
> Hi Marc,
>
> On 2020/4/16 23:08, Marc Zyngier wrote:
>> On Mon, 13 Apr 2020 20:20:23 +0800
>> Keqian Zhu <[email protected]> wrote:
>>
>>> There is already support of enabling dirty log graually in small
>>> chunks
>>
>> gradually
>>
>>> for x86 in commit 3c9bd4006bfc ("KVM: x86: enable dirty log gradually
>>> in
>>> small chunks"). This adds support for arm64.
>>>
>>> x86 still writes protect all huge pages when
>>> DIRTY_LOG_INITIALLY_ALL_SET
>>> is eanbled. However, for arm64, both huge pages and normal pages can
>>> be
>>
>> enabled
>>
>>> write protected gradually by userspace.
>>>
>>> Under the Huawei Kunpeng 920 2.6GHz platform, I did some tests on
>>> 128G
>>> Linux VMs with different page size. The memory pressure is 127G in
>>> each
>>> case. The time taken of memory_global_dirty_log_start in QEMU is
>>> listed
>>> below:
>>>
>>> Page Size Before After Optimization
>>> 4K 650ms 1.8ms
>>> 2M 4ms 1.8ms
>>> 1G 2ms 1.8ms
>>
>> These numbers are different from what you have advertised before. What
>> changed?
> In patch RFC, the numbers is got when memory pressure is 100G, so the
> number
> is bigger here.
OK.
>>
>>>
>>> Besides the time reduction, the biggest income is that we will
>>> minimize
>>
>> s/income/improvement/
>>
>>> the performance side effect (because of dissloving huge pages and
>>> marking
>>
>> dissolving
> embarrassed for these misspell :(
No need to be embarrassed. I do a lot worse, at all times. That is why
you
and I need other people to review our patches and fix things! ;-)
>>
>>> memslots dirty) on guest after enabling dirty log.
>>>
>>> Signed-off-by: Keqian Zhu <[email protected]>
>>> ---
>>> Documentation/virt/kvm/api.rst | 2 +-
>>> arch/arm64/include/asm/kvm_host.h | 3 +++
>>> virt/kvm/arm/mmu.c | 12 ++++++++++--
>>> 3 files changed, 14 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/Documentation/virt/kvm/api.rst
>>> b/Documentation/virt/kvm/api.rst
>>> index efbbe570aa9b..0017f63fa44f 100644
>>> --- a/Documentation/virt/kvm/api.rst
>>> +++ b/Documentation/virt/kvm/api.rst
>>> @@ -5777,7 +5777,7 @@ will be initialized to 1 when created. This
>>> also improves performance because
>>> dirty logging can be enabled gradually in small chunks on the first
>>> call
>>> to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
>>> KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
>>> -x86 for now).
>>> +x86 and arm64 for now).
>>>
>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the
>>> name
>>> KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs
>>> that make
>>> diff --git a/arch/arm64/include/asm/kvm_host.h
>>> b/arch/arm64/include/asm/kvm_host.h
>>> index 32c8a675e5a4..a723f84fab83 100644
>>> --- a/arch/arm64/include/asm/kvm_host.h
>>> +++ b/arch/arm64/include/asm/kvm_host.h
>>> @@ -46,6 +46,9 @@
>>> #define KVM_REQ_RECORD_STEAL KVM_ARCH_REQ(3)
>>> #define KVM_REQ_RELOAD_GICv4 KVM_ARCH_REQ(4)
>>>
>>> +#define KVM_DIRTY_LOG_MANUAL_CAPS
>>> (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
>>> + KVM_DIRTY_LOG_INITIALLY_SET)
>>> +
>>> DECLARE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
>>>
>>> extern unsigned int kvm_sve_max_vl;
>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>> index e3b9ee268823..1077f653a611 100644
>>> --- a/virt/kvm/arm/mmu.c
>>> +++ b/virt/kvm/arm/mmu.c
>>> @@ -2265,8 +2265,16 @@ void kvm_arch_commit_memory_region(struct kvm
>>> *kvm,
>>> * allocated dirty_bitmap[], dirty pages will be be tracked while
>>> the
>>> * memory slot is write protected.
>>> */
>>> - if (change != KVM_MR_DELETE && mem->flags &
>>> KVM_MEM_LOG_DIRTY_PAGES)
>>> - kvm_mmu_wp_memory_region(kvm, mem->slot);
>>> + if (change != KVM_MR_DELETE && mem->flags &
>>> KVM_MEM_LOG_DIRTY_PAGES) {
>>> + /*
>>> + * If we're with initial-all-set, we don't need to write
>>> + * protect any pages because they're all reported as dirty.
>>> + * Huge pages and normal pages will be write protect gradually.
>>> + */
>>> + if (!kvm_dirty_log_manual_protect_and_init_set(kvm)) {
>>> + kvm_mmu_wp_memory_region(kvm, mem->slot);
>>> + }
>>> + }
>>> }
>>>
>>> int kvm_arch_prepare_memory_region(struct kvm *kvm,
>>
>> As it is, it is pretty good. The one thing that isn't clear to me is
>> why we have a difference in behaviour between x86 and arm64. What
>> prevents x86 from having the same behaviour as arm64?
> I am also not very clear about the difference. Maybe there is historic
> reason.
>
> Before introducing DIRTY_LOG_INITIALLY_ALL_SET, all pages will be write
> protected when starting dirty log, but only normal pages are needed
> to be write protected again during dirty log sync, because huge pages
> will
> be dissolved to normal pages.
>
> For that x86 uses different routine to write protect huge pages and
> normal pages,
> and arm64 uses same routine to do this, so arm64 still write protect
> all
> pages again during dirty log sync, but x86 didn't.
>
> So I think that x86 can write protect huge pages gradually too, just
> need to add
> some code legs in dirty log sync.
Fair enough. It'd be good if you could investigate this as well. In the
meantime,
I'll queue this patch for a spin in -next.
Thanks,
M.
--
Jazz is not dead. It just smells funny...