[PATCH v3 0/3] drm/msm: fix for "Timeout waiting for GMU OOB set GPU_SET: 0x0"

[PATCH v3 0/3] drm/msm: fix for "Timeout waiting for GMU OOB set GPU_SET: 0x0"

[Eric Anholt]

Updated commit messages over v2, no code changes.

Eric Anholt (3):
  drm/msm: Fix race of GPU init vs timestamp power management.
  drm/msm: Fix races managing the OOB state for timestamp vs timestamps.
  drm/msm: Clean up GMU OOB set/clear handling.

 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 105 +++++++++++++++-----------
 drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  49 ++++--------
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c |   8 +-
 3 files changed, 84 insertions(+), 78 deletions(-)

-- 
2.30.0

[PATCH v3 1/3] drm/msm: Fix race of GPU init vs timestamp power management.

[Eric Anholt]

We were using the same force-poweron bit in the two codepaths, so they
could race to have one of them lose GPU power early.

freedreno CI was seeing intermittent errors like:
[drm:_a6xx_gmu_set_oob] *ERROR* Timeout waiting for GMU OOB set GPU_SET: 0x0
and this issue could have contributed to it.

Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 4b565ca5a2cb ("drm/msm: Add A6XX device support")
Reviewed-by: Jordan Crouse <jcrouse@codeaurora.org>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 25 ++++++++++++++++++++++---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  8 ++++++++
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c |  4 ++--
 3 files changed, 32 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index e6703ae98760..b3318f86aabc 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -264,6 +264,16 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 		}
 		name = "GPU_SET";
 		break;
+	case GMU_OOB_PERFCOUNTER_SET:
+		if (gmu->legacy) {
+			request = GMU_OOB_PERFCOUNTER_REQUEST;
+			ack = GMU_OOB_PERFCOUNTER_ACK;
+		} else {
+			request = GMU_OOB_PERFCOUNTER_REQUEST_NEW;
+			ack = GMU_OOB_PERFCOUNTER_ACK_NEW;
+		}
+		name = "PERFCOUNTER";
+		break;
 	case GMU_OOB_BOOT_SLUMBER:
 		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
 		ack = GMU_OOB_BOOT_SLUMBER_ACK;
@@ -301,9 +311,14 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 {
 	if (!gmu->legacy) {
-		WARN_ON(state != GMU_OOB_GPU_SET);
-		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-			1 << GMU_OOB_GPU_SET_CLEAR_NEW);
+		if (state == GMU_OOB_GPU_SET) {
+			gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+				1 << GMU_OOB_GPU_SET_CLEAR_NEW);
+		} else {
+			WARN_ON(state != GMU_OOB_PERFCOUNTER_SET);
+			gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+				1 << GMU_OOB_PERFCOUNTER_CLEAR_NEW);
+		}
 		return;
 	}
 
@@ -312,6 +327,10 @@ void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
 			1 << GMU_OOB_GPU_SET_CLEAR);
 		break;
+	case GMU_OOB_PERFCOUNTER_SET:
+		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
+			1 << GMU_OOB_PERFCOUNTER_CLEAR);
+		break;
 	case GMU_OOB_BOOT_SLUMBER:
 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
 			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
index c6d2bced8e5d..9fa278de2106 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
@@ -156,6 +156,7 @@ enum a6xx_gmu_oob_state {
 	GMU_OOB_BOOT_SLUMBER = 0,
 	GMU_OOB_GPU_SET,
 	GMU_OOB_DCVS_SET,
+	GMU_OOB_PERFCOUNTER_SET,
 };
 
 /* These are the interrupt / ack bits for each OOB request that are set
@@ -190,6 +191,13 @@ enum a6xx_gmu_oob_state {
 #define GMU_OOB_GPU_SET_ACK_NEW		31
 #define GMU_OOB_GPU_SET_CLEAR_NEW	31
 
+#define GMU_OOB_PERFCOUNTER_REQUEST	17
+#define GMU_OOB_PERFCOUNTER_ACK		25
+#define GMU_OOB_PERFCOUNTER_CLEAR	25
+
+#define GMU_OOB_PERFCOUNTER_REQUEST_NEW	28
+#define GMU_OOB_PERFCOUNTER_ACK_NEW	30
+#define GMU_OOB_PERFCOUNTER_CLEAR_NEW	30
 
 void a6xx_hfi_init(struct a6xx_gmu *gmu);
 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state);
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index c8a9010c1a1d..7424a70b9d35 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -1177,12 +1177,12 @@ static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
 
 	/* Force the GPU power on so we can read this register */
-	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
 
 	*value = gpu_read64(gpu, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
 		REG_A6XX_RBBM_PERFCTR_CP_0_HI);
 
-	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
+	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
 	return 0;
 }
 
-- 
2.30.0

[PATCH v3 2/3] drm/msm: Fix races managing the OOB state for timestamp vs timestamps.

[Eric Anholt]

Now that we're not racing with GPU setup, also fix races of timestamps
against other timestamps.  In freedreno CI, we were seeing this path trigger
timeouts on setting the GMU bit, producing:

[drm:_a6xx_gmu_set_oob] *ERROR* Timeout waiting for GMU OOB set GPU_SET: 0x0

and this triggered especially on the first set of tests right after
boot (it's probably easier to lose the race than one might think,
given that we start many tests in parallel, and waiting for NFS to
page in code probably means that lots of tests hit the same point of
screen init at the same time).  As of this patch, the message seems to
have completely gone away.

Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 4b565ca5a2cb ("drm/msm: Add A6XX device support")
Reviewed-by: Jordan Crouse <jcrouse@codeaurora.org>
---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index 7424a70b9d35..e8f0b5325a7f 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -1175,6 +1175,9 @@ static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
 {
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
+	static DEFINE_MUTEX(perfcounter_oob);
+
+	mutex_lock(&perfcounter_oob);
 
 	/* Force the GPU power on so we can read this register */
 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
@@ -1183,6 +1186,7 @@ static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
 		REG_A6XX_RBBM_PERFCTR_CP_0_HI);
 
 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
+	mutex_unlock(&perfcounter_oob);
 	return 0;
 }
 
-- 
2.30.0

[PATCH v3 3/3] drm/msm: Clean up GMU OOB set/clear handling.

[Eric Anholt]

Now that the bug is fixed in the minimal way for stable, go make the
code table-driven.

Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jordan Crouse <jcrouse@codeaurora.org>
---
 drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 124 +++++++++++++-------------
 drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  55 ++++--------
 2 files changed, 77 insertions(+), 102 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
index b3318f86aabc..9066e98eb8ef 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
@@ -245,47 +245,66 @@ static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
 	return ret;
 }
 
+struct a6xx_gmu_oob_bits {
+	int set, ack, set_new, ack_new;
+	const char *name;
+};
+
+/* These are the interrupt / ack bits for each OOB request that are set
+ * in a6xx_gmu_set_oob and a6xx_clear_oob
+ */
+static const struct a6xx_gmu_oob_bits a6xx_gmu_oob_bits[] = {
+	[GMU_OOB_GPU_SET] = {
+		.name = "GPU_SET",
+		.set = 16,
+		.ack = 24,
+		.set_new = 30,
+		.ack_new = 31,
+	},
+
+	[GMU_OOB_PERFCOUNTER_SET] = {
+		.name = "PERFCOUNTER",
+		.set = 17,
+		.ack = 25,
+		.set_new = 28,
+		.ack_new = 30,
+	},
+
+	[GMU_OOB_BOOT_SLUMBER] = {
+		.name = "BOOT_SLUMBER",
+		.set = 22,
+		.ack = 30,
+	},
+
+	[GMU_OOB_DCVS_SET] = {
+		.name = "GPU_DCVS",
+		.set = 23,
+		.ack = 31,
+	},
+};
+
 /* Trigger a OOB (out of band) request to the GMU */
 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 {
 	int ret;
 	u32 val;
 	int request, ack;
-	const char *name;
 
-	switch (state) {
-	case GMU_OOB_GPU_SET:
-		if (gmu->legacy) {
-			request = GMU_OOB_GPU_SET_REQUEST;
-			ack = GMU_OOB_GPU_SET_ACK;
-		} else {
-			request = GMU_OOB_GPU_SET_REQUEST_NEW;
-			ack = GMU_OOB_GPU_SET_ACK_NEW;
-		}
-		name = "GPU_SET";
-		break;
-	case GMU_OOB_PERFCOUNTER_SET:
-		if (gmu->legacy) {
-			request = GMU_OOB_PERFCOUNTER_REQUEST;
-			ack = GMU_OOB_PERFCOUNTER_ACK;
-		} else {
-			request = GMU_OOB_PERFCOUNTER_REQUEST_NEW;
-			ack = GMU_OOB_PERFCOUNTER_ACK_NEW;
-		}
-		name = "PERFCOUNTER";
-		break;
-	case GMU_OOB_BOOT_SLUMBER:
-		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
-		ack = GMU_OOB_BOOT_SLUMBER_ACK;
-		name = "BOOT_SLUMBER";
-		break;
-	case GMU_OOB_DCVS_SET:
-		request = GMU_OOB_DCVS_REQUEST;
-		ack = GMU_OOB_DCVS_ACK;
-		name = "GPU_DCVS";
-		break;
-	default:
+	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
 		return -EINVAL;
+
+	if (gmu->legacy) {
+		request = a6xx_gmu_oob_bits[state].set;
+		ack = a6xx_gmu_oob_bits[state].ack;
+	} else {
+		request = a6xx_gmu_oob_bits[state].set_new;
+		ack = a6xx_gmu_oob_bits[state].ack_new;
+		if (!request || !ack) {
+			DRM_DEV_ERROR(gmu->dev,
+				      "Invalid non-legacy GMU request %s\n",
+				      a6xx_gmu_oob_bits[state].name);
+			return -EINVAL;
+		}
 	}
 
 	/* Trigger the equested OOB operation */
@@ -298,7 +317,7 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 	if (ret)
 		DRM_DEV_ERROR(gmu->dev,
 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
-				name,
+				a6xx_gmu_oob_bits[state].name,
 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
 
 	/* Clear the acknowledge interrupt */
@@ -310,36 +329,17 @@ int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 /* Clear a pending OOB state in the GMU */
 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
 {
-	if (!gmu->legacy) {
-		if (state == GMU_OOB_GPU_SET) {
-			gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-				1 << GMU_OOB_GPU_SET_CLEAR_NEW);
-		} else {
-			WARN_ON(state != GMU_OOB_PERFCOUNTER_SET);
-			gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-				1 << GMU_OOB_PERFCOUNTER_CLEAR_NEW);
-		}
+	int bit;
+
+	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
 		return;
-	}
 
-	switch (state) {
-	case GMU_OOB_GPU_SET:
-		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-			1 << GMU_OOB_GPU_SET_CLEAR);
-		break;
-	case GMU_OOB_PERFCOUNTER_SET:
-		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-			1 << GMU_OOB_PERFCOUNTER_CLEAR);
-		break;
-	case GMU_OOB_BOOT_SLUMBER:
-		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
-		break;
-	case GMU_OOB_DCVS_SET:
-		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
-			1 << GMU_OOB_DCVS_CLEAR);
-		break;
-	}
+	if (gmu->legacy)
+		bit = a6xx_gmu_oob_bits[state].ack;
+	else
+		bit = a6xx_gmu_oob_bits[state].ack_new;
+
+	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, bit);
 }
 
 /* Enable CPU control of SPTP power power collapse */
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
index 9fa278de2106..71dfa60070cc 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
@@ -153,52 +153,27 @@ static inline void gmu_write_rscc(struct a6xx_gmu *gmu, u32 offset, u32 value)
  */
 
 enum a6xx_gmu_oob_state {
+	/*
+	 * Let the GMU know that a boot or slumber operation has started. The value in
+	 * REG_A6XX_GMU_BOOT_SLUMBER_OPTION lets the GMU know which operation we are
+	 * doing
+	 */
 	GMU_OOB_BOOT_SLUMBER = 0,
+	/*
+	 * Let the GMU know to not turn off any GPU registers while the CPU is in a
+	 * critical section
+	 */
 	GMU_OOB_GPU_SET,
+	/*
+	 * Set a new power level for the GPU when the CPU is doing frequency scaling
+	 */
 	GMU_OOB_DCVS_SET,
+	/*
+	 * Used to keep the GPU on for CPU-side reads of performance counters.
+	 */
 	GMU_OOB_PERFCOUNTER_SET,
 };
 
-/* These are the interrupt / ack bits for each OOB request that are set
- * in a6xx_gmu_set_oob and a6xx_clear_oob
- */
-
-/*
- * Let the GMU know that a boot or slumber operation has started. The value in
- * REG_A6XX_GMU_BOOT_SLUMBER_OPTION lets the GMU know which operation we are
- * doing
- */
-#define GMU_OOB_BOOT_SLUMBER_REQUEST	22
-#define GMU_OOB_BOOT_SLUMBER_ACK	30
-#define GMU_OOB_BOOT_SLUMBER_CLEAR	30
-
-/*
- * Set a new power level for the GPU when the CPU is doing frequency scaling
- */
-#define GMU_OOB_DCVS_REQUEST	23
-#define GMU_OOB_DCVS_ACK	31
-#define GMU_OOB_DCVS_CLEAR	31
-
-/*
- * Let the GMU know to not turn off any GPU registers while the CPU is in a
- * critical section
- */
-#define GMU_OOB_GPU_SET_REQUEST	16
-#define GMU_OOB_GPU_SET_ACK	24
-#define GMU_OOB_GPU_SET_CLEAR	24
-
-#define GMU_OOB_GPU_SET_REQUEST_NEW	30
-#define GMU_OOB_GPU_SET_ACK_NEW		31
-#define GMU_OOB_GPU_SET_CLEAR_NEW	31
-
-#define GMU_OOB_PERFCOUNTER_REQUEST	17
-#define GMU_OOB_PERFCOUNTER_ACK		25
-#define GMU_OOB_PERFCOUNTER_CLEAR	25
-
-#define GMU_OOB_PERFCOUNTER_REQUEST_NEW	28
-#define GMU_OOB_PERFCOUNTER_ACK_NEW	30
-#define GMU_OOB_PERFCOUNTER_CLEAR_NEW	30
-
 void a6xx_hfi_init(struct a6xx_gmu *gmu);
 int a6xx_hfi_start(struct a6xx_gmu *gmu, int boot_state);
 void a6xx_hfi_stop(struct a6xx_gmu *gmu);
-- 
2.30.0