1. Patchew
  2. QEMU
  3. [RFC PATCH v2 0/5] target/arm: Add MTE support to KVM guest
  4. View patch

[RFC PATCH v2 3/5] Add APIs to get/set MTE tags

Haibo Xu posted 5 patches 4 years, 8 months ago
Maintainers: Cornelia Huck <cohuck@redhat.com>, Peter Maydell <peter.maydell@linaro.org>, "Michael S. Tsirkin" <mst@redhat.com>, Paolo Bonzini <pbonzini@redhat.com>, "Dr. David Alan Gilbert" <dgilbert@redhat.com>, Juan Quintela <quintela@redhat.com>
[RFC PATCH v2 3/5] Add APIs to get/set MTE tags
Posted by Haibo Xu 4 years, 8 months ago 
MTE spec provide instructions to retrieve the memory tags:
(1) LDG, at 16 bytes granularity, and available in both user
    and kernel space;
(2) LDGM, at 256 bytes granularity in maximum, and only
    available in kernel space

To improve the performance, KVM has exposed the LDGM capability
to user space by providing a new APIs. This patch is just a
wrapper for the KVM APIs.

Signed-off-by: Haibo Xu <haibo.xu@linaro.org>
---
 target/arm/kvm64.c   | 24 ++++++++++++++++++++++++
 target/arm/kvm_arm.h |  2 ++
 2 files changed, 26 insertions(+)

diff --git a/target/arm/kvm64.c b/target/arm/kvm64.c
index 73a191f8e1..3157025316 100644
--- a/target/arm/kvm64.c
+++ b/target/arm/kvm64.c
@@ -1606,3 +1606,27 @@ bool kvm_arm_verify_ext_dabt_pending(CPUState *cs)
     }
     return false;
 }
+
+int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
+{
+    struct kvm_arm_copy_mte_tags args = {
+        .guest_ipa = ipa,
+        .length = len,
+        .addr = buf,
+        .flags = KVM_ARM_TAGS_FROM_GUEST,
+    };
+
+    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);
+}
+
+int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
+{
+    struct kvm_arm_copy_mte_tags args = {
+        .guest_ipa = ipa,
+        .length = len,
+        .addr = buf,
+        .flags = KVM_ARM_TAGS_TO_GUEST,
+    };
+
+    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);
+}
diff --git a/target/arm/kvm_arm.h b/target/arm/kvm_arm.h
index 34f8daa377..bbb833d6c6 100644
--- a/target/arm/kvm_arm.h
+++ b/target/arm/kvm_arm.h
@@ -360,6 +360,8 @@ int kvm_arm_vgic_probe(void);
 
 void kvm_arm_pmu_set_irq(CPUState *cs, int irq);
 void kvm_arm_pmu_init(CPUState *cs);
+int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
+int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
 
 /**
  * kvm_arm_pvtime_init:
-- 
2.17.1
Re: [RFC PATCH v2 3/5] Add APIs to get/set MTE tags
Posted by Juan Quintela 4 years, 7 months ago 
Haibo Xu <haibo.xu@linaro.org> wrote:
> MTE spec provide instructions to retrieve the memory tags:
> (1) LDG, at 16 bytes granularity, and available in both user
>     and kernel space;
> (2) LDGM, at 256 bytes granularity in maximum, and only
>     available in kernel space
>
> To improve the performance, KVM has exposed the LDGM capability
> to user space by providing a new APIs. This patch is just a
> wrapper for the KVM APIs.
>
> Signed-off-by: Haibo Xu <haibo.xu@linaro.org>
> ---
>  target/arm/kvm64.c   | 24 ++++++++++++++++++++++++
>  target/arm/kvm_arm.h |  2 ++
>  2 files changed, 26 insertions(+)
>
> diff --git a/target/arm/kvm64.c b/target/arm/kvm64.c
> index 73a191f8e1..3157025316 100644
> --- a/target/arm/kvm64.c
> +++ b/target/arm/kvm64.c
> @@ -1606,3 +1606,27 @@ bool kvm_arm_verify_ext_dabt_pending(CPUState *cs)
>      }
>      return false;
>  }
> +
> +int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
> +{
> +    struct kvm_arm_copy_mte_tags args = {
> +        .guest_ipa = ipa,
> +        .length = len,
> +        .addr = buf,
> +        .flags = KVM_ARM_TAGS_FROM_GUEST,
> +    };
> +
> +    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);

Just a question, how fast/slow are this calls?

My understanding is that we are making a kvm call for each page that we
want to migrate, right?

Each time that we want to send it.

Later, Juan.


> +}
> +
> +int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
> +{
> +    struct kvm_arm_copy_mte_tags args = {
> +        .guest_ipa = ipa,
> +        .length = len,
> +        .addr = buf,
> +        .flags = KVM_ARM_TAGS_TO_GUEST,
> +    };
> +
> +    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);
> +}
> diff --git a/target/arm/kvm_arm.h b/target/arm/kvm_arm.h
> index 34f8daa377..bbb833d6c6 100644
> --- a/target/arm/kvm_arm.h
> +++ b/target/arm/kvm_arm.h
> @@ -360,6 +360,8 @@ int kvm_arm_vgic_probe(void);
>  
>  void kvm_arm_pmu_set_irq(CPUState *cs, int irq);
>  void kvm_arm_pmu_init(CPUState *cs);
> +int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
> +int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
>  
>  /**
>   * kvm_arm_pvtime_init:
Re: [RFC PATCH v2 3/5] Add APIs to get/set MTE tags
Posted by Haibo Xu 4 years, 7 months ago 
On Thu, 25 Mar 2021 at 20:18, Juan Quintela <quintela@redhat.com> wrote:
>
> Haibo Xu <haibo.xu@linaro.org> wrote:
> > MTE spec provide instructions to retrieve the memory tags:
> > (1) LDG, at 16 bytes granularity, and available in both user
> >     and kernel space;
> > (2) LDGM, at 256 bytes granularity in maximum, and only
> >     available in kernel space
> >
> > To improve the performance, KVM has exposed the LDGM capability
> > to user space by providing a new APIs. This patch is just a
> > wrapper for the KVM APIs.
> >
> > Signed-off-by: Haibo Xu <haibo.xu@linaro.org>
> > ---
> >  target/arm/kvm64.c   | 24 ++++++++++++++++++++++++
> >  target/arm/kvm_arm.h |  2 ++
> >  2 files changed, 26 insertions(+)
> >
> > diff --git a/target/arm/kvm64.c b/target/arm/kvm64.c
> > index 73a191f8e1..3157025316 100644
> > --- a/target/arm/kvm64.c
> > +++ b/target/arm/kvm64.c
> > @@ -1606,3 +1606,27 @@ bool kvm_arm_verify_ext_dabt_pending(CPUState *cs)
> >      }
> >      return false;
> >  }
> > +
> > +int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
> > +{
> > +    struct kvm_arm_copy_mte_tags args = {
> > +        .guest_ipa = ipa,
> > +        .length = len,
> > +        .addr = buf,
> > +        .flags = KVM_ARM_TAGS_FROM_GUEST,
> > +    };
> > +
> > +    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);
>
> Just a question, how fast/slow are this calls?
>

There is no performance data for this API yet, but at least it's more
efficient than that
to only be able to access a single tag by the EL0 "LDG" instruction.
We will try to
collect some performance data for this KVM API later.

> My understanding is that we are making a kvm call for each page that we
> want to migrate, right?
>

Yes, currently I chose to append the tag values to the page data
during the migration.

> Each time that we want to send it.
>
> Later, Juan.
>
>
> > +}
> > +
> > +int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf)
> > +{
> > +    struct kvm_arm_copy_mte_tags args = {
> > +        .guest_ipa = ipa,
> > +        .length = len,
> > +        .addr = buf,
> > +        .flags = KVM_ARM_TAGS_TO_GUEST,
> > +    };
> > +
> > +    return kvm_vm_ioctl(kvm_state, KVM_ARM_MTE_COPY_TAGS, &args);
> > +}
> > diff --git a/target/arm/kvm_arm.h b/target/arm/kvm_arm.h
> > index 34f8daa377..bbb833d6c6 100644
> > --- a/target/arm/kvm_arm.h
> > +++ b/target/arm/kvm_arm.h
> > @@ -360,6 +360,8 @@ int kvm_arm_vgic_probe(void);
> >
> >  void kvm_arm_pmu_set_irq(CPUState *cs, int irq);
> >  void kvm_arm_pmu_init(CPUState *cs);
> > +int kvm_arm_mte_get_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
> > +int kvm_arm_mte_set_tags(uint64_t ipa, uint64_t len, uint8_t *buf);
> >
> >  /**
> >   * kvm_arm_pvtime_init:
>

Patchew 2.1-devel-b67e4c2

© 2016 - 2025 Red Hat, Inc.