Re: [PATCH v4 2/3] LoongArch: KVM: Add LBT feature detection function

From: maobibo
Date: Mon Jul 08 2024 - 21:46:11 EST

Next message: Jiaxun Yang: "Re: [PATCH v4 07/14] clocksource: mips-gic-timer: Always use cluster 0 counter as clocksource"
Previous message: Barry Song: "[PATCH v1] mm: shmem: Rename mTHP shmem counters"
In reply to: Huacai Chen: "Re: [PATCH v4 2/3] LoongArch: KVM: Add LBT feature detection function"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 2024/7/8 下午5:51, Huacai Chen wrote:

On Thu, Jul 4, 2024 at 9:24 AM maobibo <maobibo@xxxxxxxxxxx> wrote:

On 2024/7/3 下午11:35, Huacai Chen wrote:

On Wed, Jul 3, 2024 at 11:15 AM maobibo <maobibo@xxxxxxxxxxx> wrote:

On 2024/7/2 下午11:43, Huacai Chen wrote:

On Tue, Jul 2, 2024 at 4:42 PM maobibo <maobibo@xxxxxxxxxxx> wrote:

On 2024/7/2 下午3:28, Huacai Chen wrote:

On Tue, Jul 2, 2024 at 12:13 PM maobibo <maobibo@xxxxxxxxxxx> wrote:

On 2024/7/2 上午10:34, Huacai Chen wrote:

On Tue, Jul 2, 2024 at 10:25 AM maobibo <maobibo@xxxxxxxxxxx> wrote:

On 2024/7/2 上午9:59, Huacai Chen wrote:

On Tue, Jul 2, 2024 at 9:51 AM maobibo <maobibo@xxxxxxxxxxx> wrote:

Huacai,

On 2024/7/1 下午6:26, Huacai Chen wrote:

On Mon, Jul 1, 2024 at 9:27 AM maobibo <maobibo@xxxxxxxxxxx> wrote:

Huacai,

On 2024/6/30 上午10:07, Huacai Chen wrote:

Hi, Bibo,

On Wed, Jun 26, 2024 at 2:32 PM Bibo Mao <maobibo@xxxxxxxxxxx> wrote:

Two kinds of LBT feature detection are added here, one is VCPU
feature, the other is VM feature. VCPU feature dection can only
work with VCPU thread itself, and requires VCPU thread is created
already. So LBT feature detection for VM is added also, it can
be done even if VM is not created, and also can be done by any
thread besides VCPU threads.

Loongson Binary Translation (LBT) feature is defined in register
cpucfg2. Here LBT capability detection for VCPU is added.

Here ioctl command KVM_HAS_DEVICE_ATTR is added for VM, and macro
KVM_LOONGARCH_VM_FEAT_CTRL is added to check supported feature. And
three sub-features relative with LBT are added as following:
KVM_LOONGARCH_VM_FEAT_X86BT
KVM_LOONGARCH_VM_FEAT_ARMBT
KVM_LOONGARCH_VM_FEAT_MIPSBT

Signed-off-by: Bibo Mao <maobibo@xxxxxxxxxxx>
---
arch/loongarch/include/uapi/asm/kvm.h | 6 ++++
arch/loongarch/kvm/vcpu.c | 6 ++++
arch/loongarch/kvm/vm.c | 44 ++++++++++++++++++++++++++-
3 files changed, 55 insertions(+), 1 deletion(-)

diff --git a/arch/loongarch/include/uapi/asm/kvm.h b/arch/loongarch/include/uapi/asm/kvm.h
index ddc5cab0ffd0..c40f7d9ffe13 100644
--- a/arch/loongarch/include/uapi/asm/kvm.h
+++ b/arch/loongarch/include/uapi/asm/kvm.h
@@ -82,6 +82,12 @@ struct kvm_fpu {
#define KVM_IOC_CSRID(REG) LOONGARCH_REG_64(KVM_REG_LOONGARCH_CSR, REG)
#define KVM_IOC_CPUCFG(REG) LOONGARCH_REG_64(KVM_REG_LOONGARCH_CPUCFG, REG)

+/* Device Control API on vm fd */
+#define KVM_LOONGARCH_VM_FEAT_CTRL 0
+#define KVM_LOONGARCH_VM_FEAT_X86BT 0
+#define KVM_LOONGARCH_VM_FEAT_ARMBT 1
+#define KVM_LOONGARCH_VM_FEAT_MIPSBT 2
+
/* Device Control API on vcpu fd */
#define KVM_LOONGARCH_VCPU_CPUCFG 0
#define KVM_LOONGARCH_VCPU_PVTIME_CTRL 1

If you insist that LBT should be a vm feature, then I suggest the
above two also be vm features. Though this is an UAPI change, but
CPUCFG is upstream in 6.10-rc1 and 6.10-final hasn't been released. We
have a chance to change it now.

KVM_LOONGARCH_VCPU_PVTIME_CTRL need be attr percpu since every vcpu
has is own different gpa address.

Then leave this as a vm feature.

For KVM_LOONGARCH_VCPU_CPUCFG attr, it will not changed. We cannot break
the API even if it is 6.10-rc1, VMM has already used this. Else there is
uapi breaking now, still will be in future if we cannot control this.

UAPI changing before the first release is allowed, which means, we can
change this before the 6.10-final, but cannot change it after
6.10-final.

Now QEMU has already synced uapi to its own directory, also I never hear
about this, with my experience with uapi change, there is only newly
added or removed deprecated years ago.

Is there any documentation about UAPI change rules?

No document, but learn from my more than 10 years upstream experience.

Can you show me an example about with your rich upstream experience?

A simple example,
e877d705704d7c8fe17b6b5ebdfdb14b84c revert
1dccdba084897443d116508a8ed71e0ac8a0 and it changes UAPI.
1dccdba084897443d116508a8ed71e0ac8a0 is upstream in 6.9-rc1, and
e877d705704d7c8fe17b6b5ebdfdb14b84c can revert the behavior before
6.9-final, but not after that.

Before the first release, the code status is treated as "unstable", so
revert, modify is allowed. But after the first release, even if an
"error" should also be treated as a "bad feature".

Huacai,

Thanks for showing the example.

For this issue, Can we adding new uapi and mark the old as deprecated?
so that it can be removed after years.

Unnecessary, just remove the old one. Deprecation is for the usage
after the first release.

For me, it is too frequent to revert the old uapi, it is not bug and
only that we have better method now. Also QEMU has synchronized the uapi
to its directory already.

QEMU also hasn't been released after synchronizing the uapi, so it is
OK to remove the old api now.

No, I will not do such thing. It is just a joke to revert the uapi.

So just create new world and old world on Loongarch system again?

Again, code status before the first release is *unstable*, that status
is not enough to be a "world".

It's your responsibility to make a good design at the beginning, but
you fail to do that. Fortunately we are before the first release;
unfortunately you don't want to do that.

Yes, this is flaw at the beginning, however it can works and new abi can
be added.

If there is no serious bug and it is synced to QEMU already, I am not
willing to revert uabi. Different projects have its own schedule plan,
that is one reason. The most important reason may be that different
peoples have different ways handling these issues.

In another thread I found that Jiaxun said he has a solution to make
LBT be a vcpu feature and still works well. However, that may take
some time and is too late for 6.11.

It is welcome if Jiaxun provide patch for host machine type, I have no
time give any feedback with suggestion of Jianxun now.

But we have another choice now: just remove the UAPI and vm.c parts in
this series, let the LBT main parts be upstream in 6.11, and then
solve other problems after 6.11. Even if Jiaxun's solution isn't
usable, we can still use this old vm feature solution then.

There is not useul if only LBT main part goes upstream. VMM cannot use
LBT if control part is not merged.

There is no control part UAPI for LSX/LASX, but it works.

LSX/LASX feature probing comes from CPU feature detecting.

If you insist that all should be merged together, there is probably
not enough time for the 6.11 merge window.

It is not so hurry for LBT feature, only that PMU function uses the same method, that will be delayed also. Almost all user visible features need feature detection.

Regards
Bibo Mao

Huacai

From another side, what do you like to do? Reviewing patch of others
and give comments whatever grammar spelling or useful suggestions, or
Writing patch which needs much efforts rather than somethings like
feature configuration, BSP drivers.

Regards
Bibo Mao

Huacai

Regards
Bibo, Mao

Huacai

Regards
Bibo, Mao

Huacai

Regards
Bibo, Mao

Huacai

How about adding new extra features capability for VM such as?
+#define KVM_LOONGARCH_VM_FEAT_LSX 3
+#define KVM_LOONGARCH_VM_FEAT_LASX 4

They should be similar as LBT, if LBT is vcpu feature, they should
also be vcpu features; if LBT is vm feature, they should also be vm
features.

On other architectures, with function kvm_vm_ioctl_check_extension()
KVM_CAP_XSAVE2/KVM_CAP_PMU_CAPABILITY on x86
KVM_CAP_ARM_PMU_V3/KVM_CAP_ARM_SVE on arm64
These features are all cpu features, at the same time they are VM features.

If they are cpu features, how does VMM detect validity of these features
passing from command line? After all VCPUs are created and send bootup
command to these VCPUs? That is too late, VMM main thread is easy to
detect feature validity if they are VM features also.

To be honest, I am not familiar with KVM still, only get further
understanding after actual problems solving. Welcome to give comments,
however please read more backgroud if you insist on, else there will be
endless argument again.

I just say CPUCFG/LSX/LASX and LBT should be in the same class, I
haven't insisted on whether they should be vcpu features or vm
features.

It is reasonable if LSX/LASX/LBT should be in the same class, since
there is feature options such as lsx=on/off,lasx=on/off,lbt=on/off.

What is the usage about CPUCFG capability used for VM feature? It is not
a detailed feature, it is only feature-set indicator like cpuid.

Regards
Bibo Mao

Huacai

Regards
Bibo, Mao

Huacai

Regards
Bibo Mao

Huacai

diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
index 233d28d0e928..9734b4d8db05 100644
--- a/arch/loongarch/kvm/vcpu.c
+++ b/arch/loongarch/kvm/vcpu.c
@@ -565,6 +565,12 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
*v |= CPUCFG2_LSX;
if (cpu_has_lasx)
*v |= CPUCFG2_LASX;
+ if (cpu_has_lbt_x86)
+ *v |= CPUCFG2_X86BT;
+ if (cpu_has_lbt_arm)
+ *v |= CPUCFG2_ARMBT;
+ if (cpu_has_lbt_mips)
+ *v |= CPUCFG2_MIPSBT;

return 0;
case LOONGARCH_CPUCFG3:
diff --git a/arch/loongarch/kvm/vm.c b/arch/loongarch/kvm/vm.c
index 6b2e4f66ad26..09e05108c68b 100644
--- a/arch/loongarch/kvm/vm.c
+++ b/arch/loongarch/kvm/vm.c
@@ -99,7 +99,49 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
return r;
}

+static int kvm_vm_feature_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
+{
+ switch (attr->attr) {
+ case KVM_LOONGARCH_VM_FEAT_X86BT:
+ if (cpu_has_lbt_x86)
+ return 0;
+ return -ENXIO;
+ case KVM_LOONGARCH_VM_FEAT_ARMBT:
+ if (cpu_has_lbt_arm)
+ return 0;
+ return -ENXIO;
+ case KVM_LOONGARCH_VM_FEAT_MIPSBT:
+ if (cpu_has_lbt_mips)
+ return 0;
+ return -ENXIO;
+ default:
+ return -ENXIO;
+ }
+}
+
+static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
+{
+ switch (attr->group) {
+ case KVM_LOONGARCH_VM_FEAT_CTRL:
+ return kvm_vm_feature_has_attr(kvm, attr);
+ default:
+ return -ENXIO;
+ }
+}
+
int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
{
- return -ENOIOCTLCMD;
+ struct kvm *kvm = filp->private_data;
+ void __user *argp = (void __user *)arg;
+ struct kvm_device_attr attr;
+
+ switch (ioctl) {
+ case KVM_HAS_DEVICE_ATTR:
+ if (copy_from_user(&attr, argp, sizeof(attr)))
+ return -EFAULT;
+
+ return kvm_vm_has_attr(kvm, &attr);
+ default:
+ return -EINVAL;
+ }
}
--
2.39.3

Next message: Jiaxun Yang: "Re: [PATCH v4 07/14] clocksource: mips-gic-timer: Always use cluster 0 counter as clocksource"
Previous message: Barry Song: "[PATCH v1] mm: shmem: Rename mTHP shmem counters"
In reply to: Huacai Chen: "Re: [PATCH v4 2/3] LoongArch: KVM: Add LBT feature detection function"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]