[Intel-gfx] [PATCH 1/2] sna: Disable userspace w/a for i830/845 cs incoherency

Daniel Vetter daniel.vetter at ffwll.ch
Tue Oct 8 10:02:38 CEST 2013


commit c7f7dd61fd07dbf938fc6ba711de07986d35ce1f
Author: Chris Wilson <chris at chris-wilson.co.uk>
Date:   Wed Dec 12 19:43:19 2012 +0000

    sna: Pin some batches to avoid CS incoherence on 830/845

the in-kernel w/a of pinning batches was duplicated for older

Unfortunately the userspace w/a isn't as good as the kernel one - when
restarting X we need to pin a new set of batch buffer objects. Those
new batches could potentially have stale tlb entries still in the CS,
leading to gpu hangs right after X started.

The other problem is that when support for detecting the kernel-based
w/a added to be able to opt out of it using the EXEC_IS_PINNED in

commit 5b0572503eab235bc7eff20d369241330c41e630
Author: Chris Wilson <chris at chris-wilson.co.uk>
Date:   Sun Dec 16 23:04:55 2012 +0000

    sna: Enable support for opting out of the kernel CS workaround

the logic in kgem_create_batch was fumbled: Since we always opt out we
should unconditionally pin the batches. The end result is that we've
ended up with unpinned batches if the kernel implements the w/a :(

Instead of fixing the userspace w/a to work even when the kernel w/a
is available I've opted to just disable the EXEC_IS_PINNED
optimization completely - given the above issues around restarting X
it's imo better to be safe than sorry.

Cc: Chris Wilson <chris at chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter at ffwll.ch>
 src/sna/kgem.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/src/sna/kgem.c b/src/sna/kgem.c
index 7da05dd..934ab3d 100644
--- a/src/sna/kgem.c
+++ b/src/sna/kgem.c
@@ -1381,8 +1381,6 @@ void kgem_init(struct kgem *kgem, int fd, struct pci_device *dev, unsigned gen)
 		kgem->batch_flags_base |= LOCAL_I915_EXEC_NO_RELOC;
 	if (kgem->has_handle_lut)
 		kgem->batch_flags_base |= LOCAL_I915_EXEC_HANDLE_LUT;
-	if (kgem->has_pinned_batches)
-		kgem->batch_flags_base |= LOCAL_I915_EXEC_IS_PINNED;

More information about the Intel-gfx mailing list