Re: [Qemu-devel] [RFC 03/10] cpus: Introduce async_wait_run_on_cpu()

[Alex Bennée]

alvise rigo <address@hidden> writes:

> I think that async_safe_run_on_cpu() does a different thing: it
> queries a job to the target vCPU and wants all the other to "observe"
> the submitted task. However, we will have the certainty that only the
> target vCPU observed the task, the other might still be running in the
> guest code.

For the code to have run every will have come out of the run loop and
synced up at that point. No safe work is run with guest code executing.

>
> alvise
>
> On Wed, Jun 8, 2016 at 5:20 PM, Alex Bennée <address@hidden> wrote:
>>
>> Sergey Fedorov <address@hidden> writes:
>>
>>> On 08/06/16 17:10, alvise rigo wrote:
>>>> Using run_on_cpu() we might deadlock QEMU if other vCPUs are waiting
>>>> for the current vCPU. We need to exit from the vCPU loop in order to
>>>> avoid this.
>>>
>>> I see, we could deadlock indeed. Alternatively, we may want fix
>>> run_on_cpu() to avoid waiting for completion by itself when called from
>>> CPU loop.
>>
>> async_safe_run_on_cpu can't deadlock as all vCPUs are suspended (or
>> waiting) for the work to complete. The tasks are run in strict order so
>> if you queued async tasks for other vCPUs first you could ensure
>> everything is in the state you want it when you finally service the
>> calling vCPU.
>>
>>>
>>> Kind regards,
>>> Sergey
>>
>>
>> --
>> Alex Bennée


--
Alex Bennée