Re: [Qemu-devel] [PATCH v5 3/4] Plumb the HAXM-based hardware acceleration support

[Vincent Palatin <vpalatin@chromium.org>]

On Thu, Jan 5, 2017 at 10:38 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
>
>
> On 05/01/2017 15:01, Paolo Bonzini wrote:
>>
>>
>> On 05/01/2017 14:50, Vincent Palatin wrote:
>>> Sorry I missed it.
>>> I move it to qemu_cpu_kick() as asked in the Darwin patch.
>>>
>>>> Apart from the above change, can you check if there are some less
>>>> heavyeight methods to force an exit?  I can think of QueueUserAPC with
>>>> an empty pfnAPC here, and SleepEx(0, TRUE) in qemu_hax_cpu_thread_fn
>>>> before qemu_wait_io_event_common.
>>>
>>> Actually I don't know a good test case to verify such a change, any advice ?
>
> In fact there is a race anyway:

Thanks for the detailed examples and thoughts.
The timing/benchmarking code might actually need some kind of per-vcpu
time storage, but that's a detail.
I have experimented with it and so far, I have mainly generated random
numbers ...
I have yet to find a use-case where the current code (with
SuspendThread/ResumeThread) yields a better latency than just nothing
instead :(



>
>         if (cpu->exit_request) {
>             ret = 1;
>             break;
>         }
>                                         cpu->exit_request
>                                         SuspendThread
>                                         ResumeThread
>         hax_vcpu_interrupt(env);
>         qemu_mutex_unlock_iothread();
>         hax_ret = hax_vcpu_run(vcpu);
>
> and the same race is true for QueueUserAPC.  It's rare enough that I
> guess we can accept the patches with just a FIXME comment, but...  Yu
> Ning, can you tell us what user_event_pending is for? :)  My hunch is
> that we should call hax_raise_event after setting cpu->exit_request, like
>
>         hax_raise_event();
>         /* write user_event_pending before exit_request */
>         smp_wmb();
>         cpu->exit_request = 1;
>         SuspendThread/ResumeThread
>                 (or QueueUserAPC)
>
> and in the hax thread:
>
>         if (cpu->exit_request) {
>             cpu->hax_vcpu->tunnel->user_event_pending = 0;
>             ret = 1;
>             break;
>         }
>
>         hax_vcpu_interrupt(env);
>         qemu_mutex_unlock_iothread();
>
>         /* read exit_request before user_event_pending */
>         smp_rmb();
>         hax_ret = hax_vcpu_run(vcpu);
>
> but I would like some more official documentation than my own reverse
> engineering of the brain of whoever wrote the interface (I have not
> looked at the HAXM driver binary).
>
> Paolo