Emulating CPUs with larger atomic accesses

Emulating CPUs with larger atomic accesses

[Florian Weimer]

What's QEMU's approach to emulating CPU instructions that atomatically
operate on values larger than what is supported by the host CPU?

I assume that for full system emulation, this is not a problem, but
qemu-user will not achieve atomic behavior on shared memory mappings.
How much of a problem is this in practice?

Thanks,
Florian

Re: Emulating CPUs with larger atomic accesses

[Richard Henderson]

On 5/13/22 03:00, Florian Weimer wrote:
> What's QEMU's approach to emulating CPU instructions that atomatically
> operate on values larger than what is supported by the host CPU?
> 
> I assume that for full system emulation, this is not a problem, but
> qemu-user will not achieve atomic behavior on shared memory mappings.
> How much of a problem is this in practice?

Well, it doesn't work, no.  In practice, x86_64 supports 128-bit atomic operations, and 
guest requires more than that.  No one really cares anymore about 32-bit hosts with 
smaller atomic operations.


r~

Re: Emulating CPUs with larger atomic accesses

[Florian Weimer]

* Richard Henderson:

> On 5/13/22 03:00, Florian Weimer wrote:
>> What's QEMU's approach to emulating CPU instructions that atomatically
>> operate on values larger than what is supported by the host CPU?
>> I assume that for full system emulation, this is not a problem, but
>> qemu-user will not achieve atomic behavior on shared memory mappings.
>> How much of a problem is this in practice?
>
> Well, it doesn't work, no.  In practice, x86_64 supports 128-bit
> atomic operations, and guest requires more than that.  No one really
> cares anymore about 32-bit hosts with smaller atomic operations.

Which part doesn't work?  Full-system emulation?

Do guests really require wider-than-128 atomics?  That's quite
surprising?

Thanks,
Florian

Re: Emulating CPUs with larger atomic accesses

[Richard Henderson]

On 5/24/22 02:27, Florian Weimer wrote:
> * Richard Henderson:
> 
>> On 5/13/22 03:00, Florian Weimer wrote:
>>> What's QEMU's approach to emulating CPU instructions that atomatically
>>> operate on values larger than what is supported by the host CPU?
>>> I assume that for full system emulation, this is not a problem, but
>>> qemu-user will not achieve atomic behavior on shared memory mappings.
>>> How much of a problem is this in practice?
>>
>> Well, it doesn't work, no.  In practice, x86_64 supports 128-bit
>> atomic operations, and guest requires more than that.  No one really
>> cares anymore about 32-bit hosts with smaller atomic operations.
> 
> Which part doesn't work?  Full-system emulation?

No, user-only.

> Do guests really require wider-than-128 atomics?  That's quite
> surprising?

Typo there -- "and no guest requires...".


r~

Re: Emulating CPUs with larger atomic accesses

[Florian Weimer]

* Richard Henderson:

> On 5/24/22 02:27, Florian Weimer wrote:
>> * Richard Henderson:
>> 
>>> On 5/13/22 03:00, Florian Weimer wrote:
>>>> What's QEMU's approach to emulating CPU instructions that atomatically
>>>> operate on values larger than what is supported by the host CPU?
>>>> I assume that for full system emulation, this is not a problem, but
>>>> qemu-user will not achieve atomic behavior on shared memory mappings.
>>>> How much of a problem is this in practice?
>>>
>>> Well, it doesn't work, no.  In practice, x86_64 supports 128-bit
>>> atomic operations, and guest requires more than that.  No one really
>>> cares anymore about 32-bit hosts with smaller atomic operations.
>> Which part doesn't work?  Full-system emulation?
>
> No, user-only.
>
>> Do guests really require wider-than-128 atomics?  That's quite
>> surprising?
>
> Typo there -- "and no guest requires...".

Okay, thanks.  So the overall situation is okay even if we end up with
x86 guests that require CPU support for 128-bit loads.

Florian