Skip to contentSkip to conversation

drm/msm: Fix race of GPU init vs timestamp power management.

PATCH Superseded Eric Anholt 3 participants 10 messages Jan 27, 2021 – Jan 29, 2021

Subsystems (5)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

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

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.

---
 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gmu.c +22 -3
index 78836b4fb98e..378dc7f190c3 100644
@@ -264,6 +264,16 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 		}
 		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;
@@ -302,9 +312,14 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 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;
 	}
 
@@ -313,6 +328,10 @@ 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);
drivers/gpu/drm/msm/adreno/a6xx_gmu.h +8 -0
index c6d2bced8e5d..9fa278de2106 100644
@@ -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);
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2 -2
index c8a9010c1a1d..7424a70b9d35 100644
@@ -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;
 }
 
Signature · 6 lines
2.30.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#1 lore raw

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

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

---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4 -0
index 7424a70b9d35..e8f0b5325a7f 100644
@@ -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;
 }
 
Signature · 6 lines
2.30.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Reviewed-by on v1 2/3
#2 lore raw

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

Quoted 10 lines · Eric Anholt

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

The joys of not having a global mutex locking everything.

Quoted 29 lines
---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4 -0
index 7424a70b9d35..e8f0b5325a7f 100644
@@ -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;
 }
 
Signature · 1 line
2.30.0
Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#3 lore raw

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

In reply to Eric Anholt
Quoted 8 lines · Eric Anholt

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

Could you add the error msg to the commit msg, to make it more easily searchable?

BR,
-R

Quoted 29 lines
---
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +4 -1
index 7424a70b9d35..e8f0b5325a7f 100644
@@ -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;
 }
 
--
Quoted 2 lines

2.30.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#4 lore raw

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

In reply to Eric Anholt

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

---
 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gmu.c +62 -62
index 378dc7f190c3..c497e0942141 100644
@@ -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, char *file, int line)
 {
 	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 */
@@ -299,7 +318,7 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 		DRM_DEV_ERROR(gmu->dev,
 			"%s:%d Timeout waiting for GMU OOB set %s: 0x%x\n",
 			file, line,
-				name,
+				a6xx_gmu_oob_bits[state].name,
 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
 
 	/* Clear the acknowledge interrupt */
@@ -311,36 +330,17 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 /* 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 */
drivers/gpu/drm/msm/adreno/a6xx_gmu.h +15 -40

67 lines not on this page · raw message

Signature · 6 lines
2.30.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#5 lore raw

PATCH v2 3/3 drm/msm: Clean up GMU OOB set/clear handling.

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

---

Previous version hadn't been rebased off of a bit of debug code I had, so it wouldn't cleanly apply.

 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gmu.c +62 -62
index b3318f86aabc..9066e98eb8ef 100644
@@ -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 */
drivers/gpu/drm/msm/adreno/a6xx_gmu.h +15 -40

67 lines not on this page · raw message

Signature · 6 lines
2.30.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Reviewed-by on v1 3/3
#6 lore raw

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

In reply to Eric Anholt
Quoted 4 lines · Eric Anholt

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

There shouldn't be too many more OOB bits, but this is a good cleanup regardless.

Quoted 239 lines
---
 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gmu.c +62 -62
index 378dc7f190c3..c497e0942141 100644
@@ -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, char *file, int line)
 {
 	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 */
@@ -299,7 +318,7 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 		DRM_DEV_ERROR(gmu->dev,
 			"%s:%d Timeout waiting for GMU OOB set %s: 0x%x\n",
 			file, line,
-				name,
+				a6xx_gmu_oob_bits[state].name,
 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
 
 	/* Clear the acknowledge interrupt */
@@ -311,36 +330,17 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 /* 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 */
drivers/gpu/drm/msm/adreno/a6xx_gmu.h +15 -40
index 9fa278de2106..71dfa60070cc 100644
@@ -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);
Signature · 1 line
2.30.0
Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Reviewed-by on v1 1/3
#7 lore raw
In reply to Eric Anholt
Quoted 5 lines · 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.

You can add:

Fixes 4b565ca5a2cb ("drm/msm: Add A6XX device support")

Because that was my ugly.

Quoted 104 lines
---
 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(-)
drivers/gpu/drm/msm/adreno/a6xx_gmu.c +22 -3
index 78836b4fb98e..378dc7f190c3 100644
@@ -264,6 +264,16 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 		}
 		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;
@@ -302,9 +312,14 @@ int _a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state, char
 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;
 	}
 
@@ -313,6 +328,10 @@ 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);
drivers/gpu/drm/msm/adreno/a6xx_gmu.h +8 -0
index c6d2bced8e5d..9fa278de2106 100644
@@ -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);
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2 -2
index c8a9010c1a1d..7424a70b9d35 100644
@@ -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;
 }
 
Signature · 1 line
2.30.0
Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#8 lore raw
Quoted 8 lines · Jordan Crouse
Quoted 5 lines · 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.

Quoted 6 lines

You can add:

Fixes 4b565ca5a2cb ("drm/msm: Add A6XX device support")

Because that was my ugly.

I only pointed it at 5.9 because it looked like it would probably conflict against older branches. I can add the fixes tag if you'd like, though.
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#9 lore raw
Quoted 12 lines · Eric Anholt
Quoted 11 lines · Jordan Crouse
Quoted 5 lines · 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.

You can add:

Fixes 4b565ca5a2cb ("drm/msm: Add A6XX device support")
Quoted 7 lines
Quoted 3 lines

Because that was my ugly.

I only pointed it at 5.9 because it looked like it would probably conflict against older branches. I can add the fixes tag if you'd like, though.

Fair enough. It is a good bug to fix but not if there are a lot of conflicts to deal with.

Jordan

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel