[Intel-gfx] [PATCH 05/13] drm/i915: Insert a flush between batches if the breadcrumb was dropped

Chris Wilson chris at chris-wilson.co.uk
Sat Jul 14 12:24:33 CEST 2012


On Fri, 13 Jul 2012 17:46:20 +0200, Daniel Vetter <daniel at ffwll.ch> wrote:
> On Fri, Jul 13, 2012 at 02:14:08PM +0100, Chris Wilson wrote:
> > If we drop the breadcrumb request after a batch due to a signal for
> > example we aim to fix it up at the next opportunity. In this case we
> > emit a second batchbuffer with no waits upon the first and so no
> > opportunity to insert the missing request, so we need to emit the
> > missing flush for coherency. (Note that that invalidating the render
> > cache is the same as flushing it, so there should have been no
> > observable corruption.)
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> Imo still too meager commit message ;-) As I've said in the previous mail,
> I'd like some mention of the two commits that made this disaster possible
> (put the blame on me where it is due). And I think some more in-detail
> walk-thru of how things blow up would be great. And the Bugzilla link for
> the QA bugreport.

Sure, in the patch I thought I was sending I had an extra paragraph:

    As a side effect this will also paper over issues such as
      https://bugs.freedesktop.org/show_bug.cgi?id=52040
    whereby we clear the write_domain on objects on the defunct
    gpu_write_list.
    
    References: https://bugs.freedesktop.org/show_bug.cgi?id=52040

> Also, I still don't understand why this patch here isn't enough to fix up
> the fallout. So if you can enlighten me where/why stuff blows up even with
> this I'd highly appreciate. Not just because not understanding bugs makes
> me queasy, but also to have a clear picture of what I'd need to send to
> Dave it this -next cycle misses 3.6.

The remaining fallout is that we still end up using the flushing-list,
as revealed by *adding* a WARN.

To end up in that situation we must retire an object with a write-domain
still set. But how can this be possible if we always clear the
write_list prior to the request/retirment?

I thought I had it, being sneaky with the use of INSTRUCTION write
domain for pipe-control. However, looks like I'm going to have to
reproduce with some more debugging.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre



More information about the Intel-gfx mailing list