Bug 31532 - [RADEON:KMS:RV670:AGP:R600C-R600G] PCIE gart stalls
Summary: [RADEON:KMS:RV670:AGP:R600C-R600G] PCIE gart stalls
Status: RESOLVED INVALID
Alias: None
Product: Mesa
Classification: Unclassified
Component: Drivers/Gallium/r600 (show other bugs)
Version: git
Hardware: Other All
: medium normal
Assignee: Default DRI bug account
QA Contact:
URL:
Whiteboard:
Keywords:
Depends on:
Blocks:
 
Reported: 2010-11-10 12:09 UTC by Andy Furniss
Modified: 2014-11-19 22:36 UTC (History)
0 users

See Also:
i915 platform:
i915 features:


Attachments
dmesg agp gart showing a reset (30.76 KB, text/plain)
2010-12-08 10:02 UTC, Andy Furniss
Details

Description Andy Furniss 2010-11-10 12:09:51 UTC
Running d-r-t kernels, xserver 1.9 and git mesa, ddx, lib drm.

Any non trivial mesa demo with r600g will cause a gpu lockup/reset unless I use PCIE gart and turn off writebacks.

Tiling on/off makes no difference.

Not a regression AFAICT, this box was retired a while back, before 600g was really usable, so I don't have and have been unable to find a time when it ever worked.

Classic KMS + AGP gart works OK.

Both Classic KMS and Gallium with agpmode=-1 have stalling problems.

Classic runs demos fine, but will stall for 1/2 sec on some games and some levels eg. openarena anholt is OK but other levels will stutter a bit.
These do not show with latencytop.

Classic + UMS does not stall.

Gallium has far worse stalling problems than classic.

Some mesa demos will stall, gearbox shows the most - up to a second when running sometimes more at start.
Latency top will show gallium stalls - 

radeon_fence_wait often shows 488ms but never more.

Long stalls such as when starting games/demos seem to be in -

radeon_cs_ioctl I have seen 3000ms for games and 1000ms for demos.

Openarena anholt demo 600g behaves strangely - it will stall so much while loading and rendering the first few frames that this takes minutes.
After this the demo will run at full speed to completion (sometimes stalls again midway). Benchmark results have shown it took 8 seconds to render one of the early frames - 840 frames 45.0 seconds 18.7 fps 7.0/53.5/8004.0/23.7 ms.
Comment 1 Andy Furniss 2010-12-08 10:02:06 UTC
Created attachment 40928 [details]
dmesg agp gart showing a reset
Comment 2 Alex Deucher 2010-12-08 11:21:51 UTC
Can you try r600g from git master?  Make sure it has this patch:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=b7617346dcff50a66a10c61b95c33682cf629c9e
Comment 3 Andy Furniss 2010-12-08 13:35:54 UTC
(In reply to comment #2)
> Can you try r600g from git master?  Make sure it has this patch:
> http://cgit.freedesktop.org/mesa/mesa/commit/?id=b7617346dcff50a66a10c61b95c33682cf629c9e

It doesn't help this issue. I was already on master tip and just re-pulled/built to be sure, but no luck.
Comment 4 Andy Furniss 2011-05-25 08:11:21 UTC
Running agpgart with gallium is now fixed and runs without stalls since

r600g: bump domain selection up one layer.
ecc051d65b6e17115439fb3609cccfd59f6272bf

Running with agpmode=-1 + gallium is better than it was, all demos are stall free apart from fbo_firecube, but nexuiz still stalls badly.
Comment 5 Andy Furniss 2014-11-19 22:36:00 UTC
This is ancient and my AGP box is long dead so no point leaving it open.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.