[Intel-gfx] [RFC 08/14] drm/i915/pxp: Implement arb session teardown

[Chris Wilson]

Quoting Daniele Ceraolo Spurio (2021-02-08 19:43:23)
> 
> 
> On 2/6/2021 4:59 AM, Chris Wilson wrote:
> > Is there any reason at all to use the batch and not just emit directly
> > into the ring? The command sequence is short. And you probably want to
> > disable arbitration.
> 
> Future proofing - with multiple sessions in place we'd need to emit a 
> termination for each of them (pxp_emit_session_selection + 
> pxp_emit_inline_termination), so the sequence would be longer. It'd 
> still be below a page, so it should still be possible to fit it in the 
> ring if you believe that works better.

In terms of complexity and assurance (pre-allocated space), emitting from
the ring buffer is far simpler. You can make the ring upto 512KiB, but
if a single page is all that is needed for the most complicated
invalidate command, 8/16KiB should be plenty. (Number of pages then
equals number of invalidates that can in flight at any time, which
realistically is going to be a small number.)
-Chris