Re: [PATCH 01/18] KVM: x86: hyper-v: Introduce XMM output support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Nicolas Saenz Julienne <nsaenz@xxxxxxxxxx> writes:

> Hi Vitaly,
> Thanks for having a look at this.
>
> On Mon Jul 8, 2024 at 2:59 PM UTC, Vitaly Kuznetsov wrote:
>> Nicolas Saenz Julienne <nsaenz@xxxxxxxxxx> writes:
>>
>> > Prepare infrastructure to be able to return data through the XMM
>> > registers when Hyper-V hypercalls are issues in fast mode. The XMM
>> > registers are exposed to user-space through KVM_EXIT_HYPERV_HCALL and
>> > restored on successful hypercall completion.
>> >
>> > Signed-off-by: Nicolas Saenz Julienne <nsaenz@xxxxxxxxxx>
>> >
>> > ---
>> >
>> > There was some discussion in the RFC about whether growing 'struct
>> > kvm_hyperv_exit' is ABI breakage. IMO it isn't:
>> > - There is padding in 'struct kvm_run' that ensures that a bigger
>> >   'struct kvm_hyperv_exit' doesn't alter the offsets within that struct.
>> > - Adding a new field at the bottom of the 'hcall' field within the
>> >   'struct kvm_hyperv_exit' should be fine as well, as it doesn't alter
>> >   the offsets within that struct either.
>> > - Ultimately, previous updates to 'struct kvm_hyperv_exit's hint that
>> >   its size isn't part of the uABI. It already grew when syndbg was
>> >   introduced.
>>
>> Yes but SYNDBG exit comes with KVM_EXIT_HYPERV_SYNDBG. While I don't see
>> any immediate issues with the current approach, we may want to introduce
>> something like KVM_EXIT_HYPERV_HCALL_XMM: the userspace must be prepared
>> to handle this new information anyway and it is better to make
>> unprepared userspace fail with 'unknown exit' then to mishandle a
>> hypercall by ignoring XMM portion of the data.
>
> OK, I'll go that way. Just wanted to get a better understanding of why
> you felt it was necessary.
>

(sorry for delayed reply, I was on vacation)

I don't think it's an absolute must but it appears as a cleaner approach
to me. 

Imagine there's some userspace which handles KVM_EXIT_HYPERV_HCALL today
and we want to add XMM handling there. How would we know if xmm portion
of the data is actually filled by KVM or not? With your patch, we can of
course check for HV_X64_HYPERCALL_XMM_OUTPUT_AVAILABLE in
KVM_GET_SUPPORTED_HV_CPUID but this is not really straightforward, is
it? Checking the size is not good either. E.g. think about downstream
versions of KVM which may or may not have certain backports. In case we
(theoretically) do several additions to 'struct kvm_hyperv_exit', it
will quickly become a nightmare.

On the contrary, KVM_EXIT_HYPERV_HCALL_XMM (or just
KVM_EXIT_HYPERV_HCALL2) approach looks cleaner: once userspace sees it,
it knows that 'xmm' portion of the data can be relied upon.

-- 
Vitaly





[Index of Archives]     [Linux Kernel]     [Kernel Newbies]     [x86 Platform Driver]     [Netdev]     [Linux Wireless]     [Netfilter]     [Bugtraq]     [Linux Filesystems]     [Yosemite Discussion]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]

  Powered by Linux