Skip to contentSkip to conversation

drm/amdgpu: dequeue mes scheduler during fini

PATCH Superseded YuBiao Wang 3 participants 4 messages Oct 14, 2022 – Oct 18, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 drm/amdgpu: dequeue mes scheduler during fini

Update: Remove redundant comments as Christian suggests.

[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.

[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 ++
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c  |  3 ++
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 41 +++++++++++++++++++++++--
 3 files changed, 44 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index ad980f4b66e1..ea8efa52503b 100644
@@ -130,6 +130,9 @@ struct amdgpu_mes {
 	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
 	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
 
+	/* dequeue sched pipe via kiq */
+	void                            (*kiq_dequeue_sched)(struct amdgpu_device *adev);
+
 	/* ip specific functions */
 	const struct amdgpu_mes_funcs   *funcs;
 };
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -0
index 257b2e4de600..7c75758f58e1 100644
@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)
 		if (amdgpu_gfx_disable_kcq(adev))
 			DRM_ERROR("KCQ disable failed\n");
 
+		if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched)
+			adev->mes.kiq_dequeue_sched(adev);
+
 		amdgpu_mes_kiq_hw_fini(adev);
 	}
 
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +38 -3

65 lines not on this page · raw message

Signature · 1 line
2.25.1
maintainer Acked-by on v3 1/1
#1 lore raw

Am 14.10.22 um 05:34 schrieb YuBiao Wang:

Quoted 7 lines

Update: Remove redundant comments as Christian suggests.

[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.

Quoted 5 lines

[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.

Quoted 107 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 ++
  drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c  |  3 ++
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 41 +++++++++++++++++++++++--
  3 files changed, 44 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index ad980f4b66e1..ea8efa52503b 100644
@@ -130,6 +130,9 @@ struct amdgpu_mes {
  	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
  	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
 
+	/* dequeue sched pipe via kiq */
+	void                            (*kiq_dequeue_sched)(struct amdgpu_device *adev);
+
  	/* ip specific functions */
  	const struct amdgpu_mes_funcs   *funcs;
  };
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -0
index 257b2e4de600..7c75758f58e1 100644
@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)
  		if (amdgpu_gfx_disable_kcq(adev))
  			DRM_ERROR("KCQ disable failed\n");
 
+		if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched)
+			adev->mes.kiq_dequeue_sched(adev);
+
  		amdgpu_mes_kiq_hw_fini(adev);
  	}
 
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +38 -3
index b48684db2832..eef29646b074 100644
@@ -44,6 +44,7 @@ MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");
  static int mes_v11_0_hw_fini(void *handle);
  static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
  static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
+static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev);
 
  #define MES_EOP_SIZE   2048
 
@@ -1078,6 +1079,7 @@ static int mes_v11_0_sw_init(void *handle)
  	adev->mes.funcs = &mes_v11_0_funcs;
  	adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
  	adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
+	adev->mes.kiq_dequeue_sched = &mes_v11_0_kiq_dequeue_sched;
 
  	r = amdgpu_mes_init(adev);
  	if (r)
@@ -1151,6 +1153,42 @@ static int mes_v11_0_sw_fini(void *handle)
  	return 0;
  }
 
+static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev)
+{
+	uint32_t data;
+	int i;
+
+	mutex_lock(&adev->srbm_mutex);
+	soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
+
+	/* disable the queue if it's active */
+	if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
+		WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
+		for (i = 0; i < adev->usec_timeout; i++) {
+			if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
+				break;
+			udelay(1);
+		}
+	}
+	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
+	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
+				DOORBELL_EN, 0);
+	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
+				DOORBELL_HIT, 1);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
+
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
+
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
+
+	soc21_grbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+
+	adev->mes.ring.sched.ready = false;
+}
+
  static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
  {
  	uint32_t tmp;
@@ -1257,9 +1295,6 @@ static int mes_v11_0_hw_init(void *handle)
 
  static int mes_v11_0_hw_fini(void *handle)
  {
-	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
-	adev->mes.ring.sched.ready = false;
  	return 0;
  }
 
author Acked-by Christian König
#2 lore raw

Hi Alex,

Could you help review this patch? Thanks.

Best Regards,
Yubiao Wang

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Friday, October 14, 2022 3:08 PM
To: Wang, YuBiao <[email protected]>; [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Quan, Evan <[email protected]>; Chen, Horace <[email protected]>; Tuikov, Luben <[email protected]>; Deucher, Alexander <[email protected]>; Xiao, Jack <[email protected]>; Zhang, Hawking <[email protected]>; Liu, Monk <[email protected]>; Xu, Feifei <[email protected]>; Wang, Yang(Kevin) <[email protected]>
Subject: Re: [PATCH v3] drm/amdgpu: dequeue mes scheduler during fini

Am 14.10.22 um 05:34 schrieb YuBiao Wang:

Quoted 7 lines

Update: Remove redundant comments as Christian suggests.

[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.

Quoted 5 lines

[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.

Quoted 109 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 ++
  drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c  |  3 ++
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 41 +++++++++++++++++++++++--
  3 files changed, 44 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index ad980f4b66e1..ea8efa52503b 100644
@@ -130,6 +130,9 @@ struct amdgpu_mes {
  	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
  	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
 
+	/* dequeue sched pipe via kiq */
+	void                            (*kiq_dequeue_sched)(struct amdgpu_device *adev);
+
  	/* ip specific functions */
  	const struct amdgpu_mes_funcs   *funcs;
  };
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -0
index 257b2e4de600..7c75758f58e1 100644
@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)
  		if (amdgpu_gfx_disable_kcq(adev))
  			DRM_ERROR("KCQ disable failed\n");
 
+		if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched)
+			adev->mes.kiq_dequeue_sched(adev);
+
  		amdgpu_mes_kiq_hw_fini(adev);
  	}
 
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +37 -3
index b48684db2832..eef29646b074 100644
@@ -44,6 +44,7 @@ MODULE_FIRMWARE("amdgpu/gc_11_0_3_mes1.bin");
  static int mes_v11_0_hw_fini(void *handle);
  static int mes_v11_0_kiq_hw_init(struct amdgpu_device *adev);
  static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
+static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev);
 
  #define MES_EOP_SIZE   2048
 
@@ -1078,6 +1079,7 @@ static int mes_v11_0_sw_init(void *handle)
  	adev->mes.funcs = &mes_v11_0_funcs;
  	adev->mes.kiq_hw_init = &mes_v11_0_kiq_hw_init;
  	adev->mes.kiq_hw_fini = &mes_v11_0_kiq_hw_fini;
+	adev->mes.kiq_dequeue_sched = &mes_v11_0_kiq_dequeue_sched;
 
  	r = amdgpu_mes_init(adev);
  	if (r)
@@ -1151,6 +1153,42 @@ static int mes_v11_0_sw_fini(void *handle)
  	return 0;
  }
 
+static void mes_v11_0_kiq_dequeue_sched(struct amdgpu_device *adev) {
+	uint32_t data;
+	int i;
+
+	mutex_lock(&adev->srbm_mutex);
+	soc21_grbm_select(adev, 3, AMDGPU_MES_SCHED_PIPE, 0, 0);
+
+	/* disable the queue if it's active */
+	if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
+		WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
+		for (i = 0; i < adev->usec_timeout; i++) {
+			if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
+				break;
+			udelay(1);
+		}
+	}
+	data = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL);
+	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
+				DOORBELL_EN, 0);
+	data = REG_SET_FIELD(data, CP_HQD_PQ_DOORBELL_CONTROL,
+				DOORBELL_HIT, 1);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, data);
+
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 0);
+
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 0);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 0);
+	WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 0);
+
+	soc21_grbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+
+	adev->mes.ring.sched.ready = false;
+}
+
  static void mes_v11_0_kiq_setting(struct amdgpu_ring *ring)
  {
  	uint32_t tmp;
@@ -1257,9 +1295,6 @@ static int mes_v11_0_hw_init(void *handle)
 
  static int mes_v11_0_hw_fini(void *handle)
  {
-	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
-	adev->mes.ring.sched.ready = false;
  	return 0;
  }
 
#3 lore raw
In reply to YuBiao Wang

[AMD Official Use Only - General]

Quoted 3 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+       /* dequeue sched pipe via kiq */
+       void                            (*kiq_dequeue_sched)(struct amdgpu_device *adev);
+

It's unnecessary to expose a new callback function proto to perform dequeuing scheduler queue,
for kiq_fini or mes_hw_fini callback had got the propriate sequence to do this during fini procedure.
In addition, it seems that the issue is caused by that unclear CP_HQD_PQ_WPTR/CONTROL registers would trigger scheduler reading from unexpected address once CP_MES_CNTL is enabled:
a. for dequeue operation uses MMIO rather than kiq, can the dequeue operation move to mes_v11_hw_fini?
b. if something went wrong, it's better move the dequeue operation inside mes_v11_0_kiq_hw_fini before disabling mes controller.

Regards,
Jack

________________________________
From: YuBiao Wang <[email protected]>
Sent: Friday, 14 October 2022 11:34
To: [email protected] <[email protected]>
Cc: Grodzovsky, Andrey <[email protected]>; Quan, Evan <[email protected]>; Chen, Horace <[email protected]>; Tuikov, Luben <[email protected]>; Koenig, Christian <[email protected]>; Deucher, Alexander <[email protected]>; Xiao, Jack <[email protected]>; Zhang, Hawking <[email protected]>; Liu, Monk <[email protected]>; Xu, Feifei <[email protected]>; Wang, Yang(Kevin) <[email protected]>; Wang, YuBiao <[email protected]>
Subject: [PATCH v3] drm/amdgpu: dequeue mes scheduler during fini

Update: Remove redundant comments as Christian suggests.

[Why]
If mes is not dequeued during fini, mes will be in an uncleaned state during reload, then mes couldn't receive some commands which leads to reload failure.

[How]
Perform MES dequeue via MMIO after all the unmap jobs are done by mes and before kiq fini.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |  3 ++
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c  |  3 ++
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 41 +++++++++++++++++++++++--
 3 files changed, 44 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +3 -0
index ad980f4b66e1..ea8efa52503b 100644
@@ -130,6 +130,9 @@ struct amdgpu_mes {
         int                             (*kiq_hw_init)(struct amdgpu_device *adev);
         int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
 
+       /* dequeue sched pipe via kiq */
+       void                            (*kiq_dequeue_sched)(struct amdgpu_device *adev);
+
         /* ip specific functions */
         const struct amdgpu_mes_funcs   *funcs;
 };
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -0
index 257b2e4de600..7c75758f58e1 100644
@@ -4406,6 +4406,9 @@ static int gfx_v11_0_hw_fini(void *handle)
                 if (amdgpu_gfx_disable_kcq(adev))
                         DRM_ERROR("KCQ disable failed\n");
 
+               if (adev->mes.ring.sched.ready && adev->mes.kiq_dequeue_sched)
+                       adev->mes.kiq_dequeue_sched(adev);
+
                 amdgpu_mes_kiq_hw_fini(adev);
         }
 
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +38 -3

65 lines not on this page · raw message

--
2.25.1