From: Yukimasa Sugizaki ysugi@idein.jp
Hi,
The current V3D scheduler has two issues where CSD jobs are resubmitted regardless of the previous timed-out flag, and where the timer is not restarted for timed-out CL/CSD jobs (which we wish to continue running). The second one is due to the DRM scheduler API change and fixed in a similar way to [1]. A kernel command-line option to set the default timeout value is also added.
I tested this patchset with Piglit and our CSD programs in [2]. Because it is hard to get the current upstream kernel to work on BCM2711, I used the kernel from rpi-5.8.y tree [3]. There still are problems where some Piglit tests get longer time to finish running (3610 minutes to 3650 minutes in total), and some ones result in the invalid memory read errors with unknown reasons:
[17086.230959] v3d fec00000.v3d: MMU error from client CLE (4) at 0xac1000, pte invalid [17086.238722] v3d fec00000.v3d: MMU error from client CLE (4) at 0x1b61000, pte invalid [18643.303188] v3d fec00000.v3d: MMU error from client L2T (0) at 0x15bff00, pte invalid [18655.933748] v3d fec00000.v3d: MMU error from client L2T (0) at 0x15bff00, pte invalid
However, most of the CL/CSD programs are now working happily without kernel warnings and errors.
Regards, Sugizaki
[1] https://patchwork.kernel.org/patch/11732895/ [2] https://github.com/Idein/py-videocore6 [3] https://github.com/raspberrypi/linux/tree/rpi-5.8.y
Yukimasa Sugizaki (3): drm/v3d: Don't resubmit guilty CSD jobs drm/v3d: Correctly restart the timer when progress is made drm/v3d: Add job timeout module param
drivers/gpu/drm/v3d/v3d_sched.c | 62 +++++++++++++++++++++++++++++++++-------- 1 file changed, 51 insertions(+), 11 deletions(-)
-- 2.7.4