On 08/05/2018 20:17, Marc Zyngier wrote:
> On 08/05/18 15:56, Paolo Bonzini wrote:
>> On 08/05/2018 07:16, Christian Borntraeger wrote:
>>> On 05/08/2018 07:14 AM, Paul Mackerras wrote:
>>>> On Mon, May 07, 2018 at 03:18:46PM -0700, Marc Orr wrote:
>>>>> The kvm struct is (currently) tens of kilo-bytes, which turns out to be a
>>>>> large amount of memory to allocate contiguously via kzalloc. Thus, this
>>>>> patch changes the kzalloc to a vzalloc, so that the memory allocation is
>>>>> less likely to fail in resource-constrained environments.
>>>>
>>>> This will break HV KVM on powerpc, which needs the KVM struct to be
>>>> physically contiguous and in the linear mapping. We'll need to add
>>>> #define __KVM_HAVE_ARCH_VM_ALLOC to arch/powerpc/include/asm/kvm_host.h
>>>> and the kzalloc/kfree variant to arch/powerpc/kvm/powerpc.c much like
>>>> you did on arm.
>>>
>>> In the end I also want kmalloc for s390 (since we do not need the vmalloc for
>>> s390 as we are < 16kb).
>>>
>>> So Paolo,
>>> can we turn things around and only use vmalloc for x86?
>>>
>>
>> The other idea that the #define would pick kzalloc vs. vzalloc (instead
>> of enabling a custom kvm_arch_alloc_vm) would be even better I think.
>
> I'm not sure it is better, as it forces the architecture to make that
> choice at compile time, while I'd like to be able to do it at runtime.
>
> Something like this (untested):
>
> diff --git a/arch/arm/include/asm/kvm_host.h b/arch/arm/include/asm/kvm_host.h
> index c6a749568dd6..b6b1bdafd08e 100644
> --- a/arch/arm/include/asm/kvm_host.h
> +++ b/arch/arm/include/asm/kvm_host.h
> @@ -315,4 +315,8 @@ static inline bool kvm_arm_harden_branch_predictor(void)
> static inline void kvm_vcpu_load_sysregs(struct kvm_vcpu *vcpu) {}
> static inline void kvm_vcpu_put_sysregs(struct kvm_vcpu *vcpu) {}
>
> +#define __KVM_HAVE_ARCH_VM_ALLOC
> +struct kvm *kvm_arch_alloc_vm(void);
> +void kvm_arch_free_vm(struct kvm *kvm);
> +
> #endif /* __ARM_KVM_HOST_H__ */
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index ab46bc70add6..51a64ff6965e 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -452,4 +452,8 @@ static inline bool kvm_arm_harden_branch_predictor(void)
> void kvm_vcpu_load_sysregs(struct kvm_vcpu *vcpu);
> void kvm_vcpu_put_sysregs(struct kvm_vcpu *vcpu);
>
> +#define __KVM_HAVE_ARCH_VM_ALLOC
> +struct kvm *kvm_arch_alloc_vm(void);
> +void kvm_arch_free_vm(struct kvm *kvm);
> +
> #endif /* __ARM64_KVM_HOST_H__ */
> diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c
> index dba629c5f8ac..21d74e215519 100644
> --- a/virt/kvm/arm/arm.c
> +++ b/virt/kvm/arm/arm.c
> @@ -249,6 +249,21 @@ long kvm_arch_dev_ioctl(struct file *filp,
> return -EINVAL;
> }
>
> +struct kvm *kvm_arch_alloc_vm(void)
> +{
> + if (!has_vhe())
> + return kzalloc(sizeof(struct kvm), GFP_KERNEL);
> +
> + return vzalloc(sizeof(struct kvm));
> +}
> +
> +void kvm_arch_free_vm(struct kvm *kvm)
> +{
> + if (!has_vhe())
> + kfree(kvm);
> + else
> + vfree(kvm);
> +}
>
> struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm, unsigned int id)
> {
>
> Thanks,
>
> M.
>
Sounds good then.
Paolo
