Nouveau - Nov 2015 - [PATCH v3] pmu: fix queued messages while getting no IRQ

I encountered while stresstesting the reclocking code, that rarely (1 out of
20.000+ requests) we don't get any IRQ in nvkm_pmu_intr.

This means we have a queued message on the pmu, but nouveau doesn't read it
and
waits infinitely in nvkm_pmu_send:
	if (reply) {
		wait_event(pmu->recv.wait, (pmu->recv.process == 0));

therefore let us use wait_event_timeout with a 1s timeout frame and just check
whether there is a message queued and handle it if there is one.

Return -ETIMEDOUT whenever we timed out and there is no message queued or when
we hit another timeout while trying to read the message without getting any IRQ

The benefit of not using wait_event is, that we don't have a kworker waiting
on an event, which makes it easier to reload the module at runtime, which helps
me developing on nouveau on my laptop a lot, because I don't need to reboot
anymore

Nethertheless, we shouldn't use wait_event here, because we can't
guarantee any
answere at all, can we?

v2: moved it into a new function
v3: moved mutex unlock into nvkm_pmu_send

Signed-off-by: Karol Herbst <nouveau at karolherbst.de>
---
 drm/nouveau/nvkm/subdev/pmu/base.c | 39 ++++++++++++++++++++++++++++++++++----
 1 file changed, 35 insertions(+), 4 deletions(-)

diff --git a/drm/nouveau/nvkm/subdev/pmu/base.c
b/drm/nouveau/nvkm/subdev/pmu/base.c
index 6b2007f..81a5583 100644
--- a/drm/nouveau/nvkm/subdev/pmu/base.c
+++ b/drm/nouveau/nvkm/subdev/pmu/base.c
@@ -43,6 +43,34 @@ nvkm_pmu_handle_reclk_request(struct work_struct *work)
 	nvkm_clk_pmu_reclk_request(clk, pmu->intr.data[0]);
 }
 
+static int
+wait_for_pmu_reply(struct nvkm_pmu *pmu, u32 reply[2])
+{
+	struct nvkm_subdev *subdev = &pmu->subdev;
+	struct nvkm_device *device = subdev->device;
+	unsigned long jiffies = msecs_to_jiffies(1000);
+
+	if (!wait_event_timeout(pmu->recv.wait, (pmu->recv.process == 0),
jiffies)) {
+		u32 addr = nvkm_rd32(device, 0x10a4cc);
+		nvkm_error(subdev, "wait on reply timed out\n");
+
+		if (addr == nvkm_rd32(device, 0x10a4c8))
+			return -ETIMEDOUT;
+
+		nvkm_error(subdev, "found queued message without getting an
interrupt\n");
+		schedule_work(&pmu->recv.work);
+
+		if (!wait_event_timeout(pmu->recv.wait, (pmu->recv.process == 0),
jiffies)) {
+			nvkm_error(subdev, "failed to repair PMU state\n");
+			return -ETIMEDOUT;
+		}
+	}
+
+	reply[0] = pmu->recv.data[0];
+	reply[1] = pmu->recv.data[1];
+	return 0;
+}
+
 int
 nvkm_pmu_send(struct nvkm_pmu *pmu, u32 reply[2],
 	      u32 process, u32 message, u32 data0, u32 data1)
@@ -50,6 +78,7 @@ nvkm_pmu_send(struct nvkm_pmu *pmu, u32 reply[2],
 	struct nvkm_subdev *subdev = &pmu->subdev;
 	struct nvkm_device *device = subdev->device;
 	u32 addr;
+	int ret = 0;
 
 	/* wait for a free slot in the fifo */
 	addr  = nvkm_rd32(device, 0x10a4a0);
@@ -89,13 +118,15 @@ nvkm_pmu_send(struct nvkm_pmu *pmu, u32 reply[2],
 
 	/* wait for reply, if requested */
 	if (reply) {
-		wait_event(pmu->recv.wait, (pmu->recv.process == 0));
-		reply[0] = pmu->recv.data[0];
-		reply[1] = pmu->recv.data[1];
+		ret = wait_for_pmu_reply(pmu, reply);
+		if (ret < 0) {
+			reply[0] = 0;
+			reply[1] = 0;
+		}
 		mutex_unlock(&subdev->mutex);
 	}
 
-	return 0;
+	return ret;
 }
 
 static void
-- 
2.6.3