Skip to contentSkip to conversation

nouveau: hack fix for regression from 6.2

RFC PATCH New Dave Airlie 1 participant 1 message Jan 7, 2025

Subsystems (3)
lore
Share & tools
lore

Dave Airlie

author maintainer
#0 lore raw

RFC PATCH 1/1 nouveau: hack fix for regression from 6.2

From: Dave Airlie <[email protected]>

I recently got a regression report for RHEL 8.10 with a multi-card GP106 setup. This led me down a rabbit hole of a few problems, I've sent the cross-fence device fix but this is also needed to make things work properly.

Hopefully I can at least work out how to restrict this to pre-Ampere maybe? or find something else

This reverts 6762510bc8447dc4eb4c3d99541de6b31843e649

Author: Ben Skeggs <[email protected]>
Date:   Wed Jun 1 20:48:08 2022 +1000

drm/nouveau/gr/gf100-: call FECS WFI_GOLDEN_SAVE method

The symptoms are on a dual GPU (turing and pascal) gnome-shell gets a lot of

nouveau 0000:01:00.0: gr: DATA_ERROR 0000009c [] ch 2 [017fd2f000 gnome-shell[1554]] subc 0 class c597 mthd 0d78 data 00000004

and nothing renders on the second GPU (the data errors are on the primary GPU).

---
 drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c +1 -1
index cb390e0134a23..fa4c2174ea089 100644
@@ -1514,7 +1514,7 @@ gf100_grctx_generate(struct gf100_gr *gr, struct gf100_gr_chan *chan, struct nvk
 
 	grctx->main(chan);
 
-	if (!gr->firmware) {
+	if (1) {// {!gr->firmware) {
 		/* Trigger a context unload by unsetting the "next channel valid" bit
 		 * and faking a context switch interrupt.
 		 */
Signature · 1 line
2.43.0