Re: [PATCH v3 04/10] KVM: Implement kvm_put_guest()

[Steven Price]

On 22/08/2019 17:24, Sean Christopherson wrote:
> On Thu, Aug 22, 2019 at 04:46:10PM +0100, Steven Price wrote:
>> On 22/08/2019 16:28, Sean Christopherson wrote:
>>> On Wed, Aug 21, 2019 at 04:36:50PM +0100, Steven Price wrote:
>>>> kvm_put_guest() is analogous to put_user() - it writes a single value to
>>>> the guest physical address. The implementation is built upon put_user()
>>>> and so it has the same single copy atomic properties.
>>>
>>> What you mean by "single copy atomic"?  I.e. what guarantees does
>>> put_user() provide that __copy_to_user() does not?
>>
>> Single-copy atomicity is defined by the Arm architecture[1] and I'm not
>> going to try to go into the full details here, so this is a summary.
>>
>> For the sake of this feature what we care about is that the value
>> written/read cannot be "torn". In other words if there is a read (in
>> this case from another VCPU) that is racing with the write then the read
>> will either get the old value or the new value. It cannot return a
>> mixture. (This is of course assuming that the read is using a
>> single-copy atomic safe method).
> 
> Thanks for the explanation.  I assumed that's what you were referring to,
> but wanted to double check.
>  
>> __copy_to_user() is implemented as a memcpy() and as such cannot provide
>> single-copy atomicity in the general case (the buffer could easily be
>> bigger than the architecture can guarantee).
>>
>> put_user() on the other hand is implemented (on arm64) as an explicit
>> store instruction and therefore is guaranteed by the architecture to be
>> single-copy atomic (i.e. another CPU cannot see a half-written value).
> 
> I don't think kvm_put_guest() belongs in generic code, at least not with
> the current changelog explanation about it providing single-copy atomic
> semantics.  AFAICT, the single-copy thing is very much an arm64
> implementation detail, e.g. the vast majority of 32-bit architectures,
> including x86, do not provide any guarantees, and x86-64 generates more
> or less the same code for put_user() and __copy_to_user() for 8-byte and
> smaller accesses.
> 
> As an alternative to kvm_put_guest() entirely, is it an option to change
> arm64's raw_copy_to_user() to redirect to __put_user() for sizes that are
> constant at compile time and can be handled by __put_user()?  That would
> allow using kvm_write_guest() to update stolen time, albeit with
> arguably an even bigger dependency on the uaccess implementation details.

I think it's important to in some way ensure that the desire that this
is a single write is shown. copy_to_user() is effectively
"setup();memcpy();finish();" and while a good memcpy() implementation
would be identical to put_user() there's a lot more room for this being
broken in the future by changes to the memcpy() implementation. (And I
don't want to require that memcpy() has to detect this case).

One suggestion is to call it something like kvm_put_guest_atomic() to
reflect the atomicity requirement. Presumably that would be based on a
new put_user_atomic() which architectures could override as necessary if
put_user() doesn't provide the necessary guarantees.

Steve