Re: [RFC 3/3] virtio-gpu: make the IO handler reentrant

[Li Qiang]

Michael Tokarev <mjt@tls.msk.ru> 于2020年9月3日周四 下午1:12写道：
>
> 02.09.2020 19:22, Li Qiang wrote:
> ..
> > @@ -809,6 +809,10 @@ void virtio_gpu_process_cmdq(VirtIOGPU *g)
> >  {
> >      struct virtio_gpu_ctrl_command *cmd;
> >
> > +    if (atomic_read(&g->in_io)) {
> > +        return;
> > +    }
> > +    atomic_set(&g->in_io, 1);
>
> Can't we race in these two instructions? Both
> threads atomic_reads at the same time, both see zero,
> and both are trying to set it to 1, no?
>
> Just asking really, b/c despite of the atomic_ prefix,
> to me this look a bit unsafe..

Yes you're right. My patch is wrong. Here I try to address race
condition and DMA to MMIO issue at the same time.

I will first focus the DMA to MMIO issue in the revision patch.

Thanks,
Li Qiang

>
> Thanks,
>
> /mjt