[Nouveau] [PATCH] fifo/gk104: kick channel upon removal

[Alexandre Courbot]

A channel may still be processed by the PBDMA even after removal, unless
it is properly kicked. Some chips are more sensible to this than others,
with GM20B triggering the issue very easily (the PBDMA will try to fetch
methods from the previously-removed channel after a new one is added).

Make sure this cannot happen by kicking the channel right after it is
disabled, and before the new runlist is submitted.

Signed-off-by: Alexandre Courbot <acourbot at nvidia.com>
---
Second attempt at this - the first version was reverted after causing a
regression: https://lists.freedesktop.org/archives/nouveau/2015-August/021842.html

I am pretty confident this one will pass though: we are now kicking the channel
only if it has a chance of being scheduled at that time, and in a path from which
we *will* access the FIFO registers after the kick, so if nothing complained until
then, kicking the channel should be unconsequential.

Eric, since you reported the regression on the first version, would you mind trying
this one and giving us your Tested-by? I have tested it on dGPU but unfortunately
could not set it up in an optimus-like manner, with power management kicking in.

 drm/nouveau/nvkm/engine/fifo/gpfifogk104.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c b/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
index 2e1df01bd928..8b4a5e01829c 100644
--- a/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
+++ b/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
@@ -154,6 +154,7 @@ gk104_fifo_gpfifo_fini(struct nvkm_fifo_chan *base)
 	if (!list_empty(&chan->head)) {
 		gk104_fifo_runlist_remove(fifo, chan);
 		nvkm_mask(device, 0x800004 + coff, 0x00000800, 0x00000800);
+		gk104_fifo_gpfifo_kick(chan);
 		gk104_fifo_runlist_commit(fifo, chan->engine);
 	}
 
-- 
2.7.2