[Intel-gfx] [PATCH v4 4/8] drm/i915: vgpu context submission pv optimization

[Zhang, Xiaolin]

On 03/30/2019 03:14 AM, Chris Wilson wrote:
> Quoting Xiaolin Zhang (2019-03-29 13:32:40)
>> +       spin_lock(&engine->i915->vgpu.shared_page_lock);
>> +       shared_page->ring_id = engine->id;
>> +       for (n = 0; n < execlists_num_ports(execlists); n++)
>> +               shared_page->descs[n] = descs[n];
>> +
>> +       __raw_i915_write32(uncore, vgtif_reg(g2v_notify),
>> +                       VGT_G2V_PV_SUBMISSION);
> There's no serialisation with the consumer here. The
> shared_page->descs[] will be clobbered by a second engine before they
> are read by pv. There should be a wait-for-ack here, or separate
> channels for each engine. In execlists, we avoid this by not dequeuing
> until after we get a response/ack from HW.
>
> Preemption? I guess preemption was turned off, or else you have to worry
> about the same engine overwriting the desc[] (if no ack).
> -Chris
>
Chris, thanks you to point it out my flaw here. will use the per-engine
channel for notification.  preemption is turned off and no support in
current version.