[Mesa-dev] [PATCH 00/12] render reordering for optimized tile buffer usage

[Rob Clark]

On Fri, Jul 8, 2016 at 12:14 PM, Rob Clark <robdclark at gmail.com> wrote:
> One of the annoying things, since pipe_resource is per-screen, not
> per-context, I end up having to push batch_cache down into screen.
> Which means that, for example, one context switching fb state could
> force flushing a batch from another context.  Eventually if I push of
> gmem+submit to a per-context helper thread, that should help keep
> things properly serialized.  Although I still need some (currently
> missing) mutexes to serialize batch_cache access, etc.  Also it means
> that the upper limit on # of batches is per-screen, not per-context.
> Not really sure what to do about that.  I really wish resources were
> not shared across contexts (but rather just use flink or dmabuf when
> you need to share), but I guess it is too late for that now :-(

so what sorts of scenarios do I have to care about multiple contexts
sharing a pipe_resource?  Pushing batch_cache down into screen is
awfully painful.  I more or less have it working, but not terribly
happy with all the extra locking and complexity.

If I restrict things so that a pipe_resource could only be active (ie.
pending read or write) by a single context at a time (without blocking
on a fence when switching between contexts) I can move batch_cache
back up to the context, and make a whole lot of pain go away..

BR,
-R