Re: [kvm-devel] [PATCH 3/3] KVM: Expose MSRs to userspace

[Avi Kivity <avi@qumranet.com>]

Arnd Bergmann wrote:
> On Thursday 16 November 2006 19:04, Avi Kivity wrote:
>   
>> +struct kvm_msr_entry {
>> +       __u32 index;
>> +       __u32 reserved;
>> +       __u64 data;
>> +};
>> +
>> +/* for KVM_GET_MSRS and KVM_SET_MSRS */
>> +struct kvm_msrs {
>> +       __u32 vcpu;
>> +       __u32 nmsrs; /* number of msrs in entries */
>> +
>> +       union {
>> +               struct kvm_msr_entry __user *entries;
>> +               __u64 padding;
>> +       };
>> +};
>>     
>
> ioctl interfaces with pointers in them are generally a bad idea,
> though you handle most of the points against them fine here
> (endianess doesn't matter, padding is correct).
>
> Still, it might be better not to set a bad example. Is accessing
> the MSRs actually performance critical? If not, you could
> define the ioctl to take only a single entry argument.
>
>   

But then you can't dynamically determine which MSRs are available.

And no, reading/setting MSRs isn't performance critical for the current 
use cases.

> A possible alternative could also be to have a variable length
> argument like below, but that creates other problems:
>
> +struct kvm_msrs {
> +       __u32 vcpu;
> +       __u32 nmsrs; /* number of msrs in entries */
> +       struct kvm_msr_entry entries[0]; /* followed by actual msrs */
> +};
>
> This would mean that you can't tell the transfer size from the
> ioctl number, but you can't do that in your code either, because
> you do two separate transfers.
>
>   

Heh.  That was the original implementation by Uri.  I felt that was 
wrong because _IOW() encodes the size in the ioctl number, bit the 
actual size is different.


> 	Arnd <><
>   


-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.