Re: [PATCH] xen: Introduce cmpxchg64() and guest_cmpxchg64()

[Julien Grall <julien@xen.org>]

On 19/08/2020 10:22, Jan Beulich wrote:
> On 17.08.2020 15:03, Julien Grall wrote:
>> On 17/08/2020 12:50, Roger Pau Monné wrote:
>>> On Mon, Aug 17, 2020 at 12:05:54PM +0100, Julien Grall wrote:
>>>> The only way I could see to make it work would be to use the same trick as
>>>> we do for {read, write}_atomic() (see asm-arm/atomic.h). We are using union
>>>> and void pointer to prevent explicit cast.
>>>
>>> I'm mostly worried about common code having assumed that cmpxchg
>>> does also handle 64bit sized parameters, and thus failing to use
>>> cmpxchg64 when required. I assume this is not much of a deal as then
>>> the Arm 32 build would fail, so it should be fairly easy to catch
>>> those.
>> FWIW, this is not very different to the existing approach. If one would
>> use cmpxchg() with 64-bit, then it would fail to compile.
> 
> A somewhat related question then: Do you really need both the
> guest_* and the non-guest variants? Limiting things to plain
> cmpxchg() would further reduce the risk of someone picking the
> wrong one without right away noticing the build issue on Arm32.
> For guest_cmpxchg{,64}() I think there's less of a risk.

For the IOREQ code, we will need the guest_* version that is built on 
top of the non-guest variant.

I would like at least consistency between the two variants. IOW, if we 
decide to use the name guest_cmpxchg64(), then I would like to use 
cmpxchg64().

I still need to explore the code generated by cmpxchg() if I include 
support for 64-bit.

Cheers,

-- 
Julien Grall