Quoting Chris Wilson (2018-05-08 17:30:41)
> We assume that the CSB is written using the normal ringbuffer
> coherency protocols, as outlined in kernel/events/ring_buffer.c:
>
> * (HW) (DRIVER)
> *
> * if (LOAD ->data_tail) { LOAD ->data_head
> * (A) smp_rmb() (C)
> * STORE $data LOAD $data
> * smp_wmb() (B) smp_mb() (D)
> * STORE ->data_head STORE ->data_tail
> * }
>
> So we assume that the HW fulfils its ordering requirements (B), and so
> we should use a complimentary rmb (C) to ensure that our read of its
> WRITE pointer is completed before we start accessing the data.
>
> The final mb (D) is implied by the uncached mmio we perform to inform
> the HW of our READ pointer.
>
> References: https://bugs.freedesktop.org/show_bug.cgi?id=105064
-References: https://bugs.freedesktop.org/show_bug.cgi?id=105064
Resolved as ba74cb10c775 ("drm/i915/execlists: Delay writing to ELSP
until HW has processed the previous write")
So the other possibility is that this fixes up the issue with VT-d
latency.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
