在 2023/2/17 03:34, Oliver Upton 写道:
> On Tue, Feb 14, 2023 at 09:00:50PM +0800, Tianrui Zhao wrote:
>> 在 2023年02月14日 17:58, Greg Kroah-Hartman 写道:
>>> On Tue, Feb 14, 2023 at 05:00:56PM +0800, Tianrui Zhao wrote:
>>>>>> +#define KVM_GET_CSRS _IOWR(KVMIO, 0xc5, struct kvm_csrs)
>>>>>> +#define KVM_SET_CSRS _IOW(KVMIO, 0xc6, struct kvm_csrs)
>>>>> Why does this arch need new ioctls?
>>>> We want to use this ioctl to access multiple csrs at one time. If without
>>>> this, we only access one csr.
>>> What is wrong with accessing only one csr at a time? Isn't this what
>>> other architectures do?
>>
>> Generally, using KVM_GET_ONE ioctl to get one reg, but we want a
>> more convenient interface to get serial regs at one time, so we add this
>> ioctl.
>
> Have you found register accesses through the KVM_{GET,SET}_ONE_REG
> ioctls to actually be a bounding issue? I'd be surprised if that were
> actually the case.
>
> An architecture-neutral implementation was entertained a few years ago
> [*], but even then it saved an inconsequential amount of time relative
> to the rest of VM serialization (at least for arm64). The one thing that
> series got right was to share the plumbing across all architectures that
> use the ONE_REG interface (i.e. everyone but x86).
>
> If you have data that supports the thesis that a batched ioctl is
> useful then please do share. But in any case this should not use an ioctl
> tied down to a single architecture.
we found this issue during migration downtime stage, all vcpus will get
vcpu hw registers by KVM_GET_ONE_REG interface. If the virtual machine
has many vcpus, it will increase downtime time and reduce migration success
ratio; however it is just in theory and there is no actual data supporting
the point.
I am newbie for kvm:), and we will use KVM_GET_ONE_REG interface for loongarch
kvm for updated patches, which is common for all architectures. I just think
that it is normal to get many registers just like ptrace function, ptrace
regset API can be used to get many registers at one time.
Regards
Bibo, Mao
>
> [*] https://lore.kernel.org/kvm/20201120125616.14436-1-darkhan@xxxxxxxxxx/
>
