Quoting Tvrtko Ursulin (2018-06-27 14:03:07)
>
> On 27/06/2018 11:35, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2018-06-27 10:52:45)
> >>
> >> On 25/06/2018 10:48, Chris Wilson wrote:
> >>> @@ -1109,16 +1089,11 @@ static void process_csb(struct intel_engine_cs *engine)
> >>> } else {
> >>> port_set(port, port_pack(rq, count));
> >>> }
> >>> - }
> >>> + } while (head != tail);
> >>>
> >>> - if (head != execlists->csb_head) {
> >>> - execlists->csb_head = head;
> >>> - writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, head << 8),
> >>> - i915->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine)));
> >>> - }
> >>> -
> >>> - if (unlikely(fw))
> >>> - intel_uncore_forcewake_put(i915, execlists->fw_domains);
> >>> + writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK, head << 8),
> >>> + execlists->csb_read);
> >>
> >> Continuing from the last round - so what to do with this one? It does
> >> need forcewake. So I think it needs to go if we are claiming there is no
> >> mmio any longer.
> >
> > From last round, we decided it didn't, or at least concluded the
> > (from the lack of) evidence that it does not, because we are not using
> > forcewake right now...
>
> But we are not sure if our writes stick 100% of the time due using the
> HWSP path. And we are wasting time on MMIO for nothing. Put an "if
> (execlists->csb_use_mmio)" on it?
We only ever read from the status buffer. We always commit to HW with a
mmio write to let the HW know how far we read up to (there's no slot in
the HWSP for our read pointer, just the HW write pointer).
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
