On 26/04/17 02:57, Zhi Wang wrote:
Uh...sorry for not mentioning that before:), and stolen memory is not my
business. :(
Actually we root-caused it.
This is how we found this case:
The story is W driver directly allocated the ring buffer after the
context image, and the context image size in W driver is 19 pages. GVT-g
will do shadow context during submission, we copy 20 pages from guest
context image, so you can see, an extra page is copied here as the
context image size is actual 19 pages. The extra page belows to ring
buffer. When guest updates that page with new commands during GVT-g
executing the workload, the extra page ( which is ring buffer page) will
be over-written with old content, since GVT-g will copy the shadow
context (20 pages) back to guest at this time.
That's the full story. I send another email to Harsh. He should know if
the context image size of CHV is also 19 pages.
Thanks,
Zhi.
I did a quick check and according to the specs both the BDW and the CHV
lrcs are formed by 18096 dwords plus the per-context HWSP, which
converts to 19 pages for both platforms.
Regards,
Daniele
于 04/26/17 17:52, Joonas Lahtinen 写道:
On ke, 2017-04-26 at 17:10 +0800, Zhi Wang wrote:
Hi Joonas:
Can you change GEN8_LR_CONTEXT_RENDER_SIZE = (19 * PAGE_SIZE)?
Then we don't need the hack in GVT-g. :P Actually it's 19 pages not
20 pages on BDW.
The exception is only made for BDW, not Gen8 overall. Has the change
been verified for CHV too?
Why hasn't a patch to fix above been sent for i915 in the past? Just
like in the stolen memory disabling case, bugs should be root caused
and then fixed, not just worked around quickly.
Regards, Joonas
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
