Comment # 13 on bug 62959 from Marek Olšák

This kernel patch fixes everything:

diff --git a/drivers/gpu/drm/radeon/radeon_cs.c
b/drivers/gpu/drm/radeon/radeon_cs.c
index 70d3824..748a933 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -459,6 +459,7 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device
*rdev,
        if (r) {
                goto out;
        }
+       radeon_fence_wait(vm->fence, false);
        radeon_cs_sync_rings(parser);
        radeon_ib_sync_to(&parser->ib, vm->fence);
        radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(

It's merely a workaround and it kills performance, but it's now pretty clear
there is a synchronization issue in the kernel affecting all NI chips with
virtual memory, and it should now be easier to find the bug. I'm not really
familiar with the kernel code. I had to do some code reading before I found the
right place to put the wait call in.