Skip to contentSkip to conversation

drm:amdgpu: check before setting hw priority

PATCH In review Nirmoy Das 4 participants 6 messages Oct 14, 2020 – Oct 16, 2020

Subsystems (2)
lore
Share & tools
lore

Nirmoy Das

author
#0 lore raw

PATCH 1/2 drm:amdgpu: check before setting hw priority

Check validity of drm_gpu_scheduler before setting hw priority. Also fix a minor indentation issue.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +6 -3
index a03398c87344..c8da0978d4f6 100644
@@ -504,9 +504,9 @@ struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 }
 
 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
-					    struct amdgpu_ctx_entity *aentity,
-					    int hw_ip,
-					    enum drm_sched_priority priority)
+					   struct amdgpu_ctx_entity *aentity,
+					   int hw_ip,
+					   enum drm_sched_priority priority)
 {
 	struct amdgpu_device *adev = ctx->adev;
 	unsigned int hw_prio;
@@ -523,6 +523,9 @@ static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
 		hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
 		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
 		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
+		if (!scheds || !num_scheds)
+			return;
+
 		drm_sched_entity_modify_sched(&aentity->entity, scheds,
 					      num_scheds);
 	}
Signature · 6 lines
2.28.0

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

PATCH 2/2 drm/amdgpu: enable only one compute queue for raven

Because of firmware bug, Raven asics can't handle jobs scheduled to multiple compute queues. So enable only one compute queue till we have a firmware fix.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c |  4 ++++
 drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 11 ++++++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +4 -0
index 8c9bacfdbc30..ca2ac985b300 100644
@@ -195,6 +195,10 @@ static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
 					       int queue)
 {
+	/* We only enable one compute queue for Raven */
+	if (adev->asic_type == CHIP_RAVEN)
+		return false;
+
 	/* Policy: make queue 0 of each pipe as high priority compute queue */
 	return (queue == 0);
 
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +10 -1
index 0d8e203b10ef..f3fc9ad8bc20 100644
@@ -4633,7 +4633,16 @@ static int gfx_v9_0_early_init(void *handle)
 		adev->gfx.num_gfx_rings = 0;
 	else
 		adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
-	adev->gfx.num_compute_rings = amdgpu_num_kcq;
+
+	/* raven firmware currently can not load balance jobs
+	 * among multiple compute queues. Enable only one
+	 * compute queue till we have a firmware fix.
+	 */
+	if (adev->asic_type == CHIP_RAVEN)
+		adev->gfx.num_compute_rings = 1;
+	else
+		adev->gfx.num_compute_rings = amdgpu_num_kcq;
+
 	gfx_v9_0_set_kiq_pm4_funcs(adev);
 	gfx_v9_0_set_ring_funcs(adev);
 	gfx_v9_0_set_irq_funcs(adev);
Signature · 6 lines
2.28.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#2 lore raw

Re: [PATCH 2/2] drm/amdgpu: enable only one compute queue for raven

Please ignore this patch, it didn't workaround the fw bug.

Regards,

Nirmoy

Quoted 37 lines · Nirmoy Das

Because of firmware bug, Raven asics can't handle jobs scheduled to multiple compute queues. So enable only one compute queue till we have a firmware fix.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c |  4 ++++
  drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 11 ++++++++++-
  2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +4 -0
index 8c9bacfdbc30..ca2ac985b300 100644
@@ -195,6 +195,10 @@ static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
  bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
  					       int queue)
  {
+	/* We only enable one compute queue for Raven */
+	if (adev->asic_type == CHIP_RAVEN)
+		return false;
+
  	/* Policy: make queue 0 of each pipe as high priority compute queue */
  	return (queue == 0);
 
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +3 -1
index 0d8e203b10ef..f3fc9ad8bc20 100644
@@ -4633,7 +4633,16 @@ static int gfx_v9_0_early_init(void *handle)
  		adev->gfx.num_gfx_rings = 0;
  	else
  		adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
-	adev->gfx.num_compute_rings = amdgpu_num_kcq;
+
+	/* raven firmware currently can not load balance jobs
+	 * among multiple compute queues. Enable only one

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c static int gfx_v9_0_early_init(void *handle) · line 4648

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +7 -0
+	 * compute queue till we have a firmware fix.
+	 */
+	if (adev->asic_type == CHIP_RAVEN)
+		adev->gfx.num_compute_rings = 1;
+	else
+		adev->gfx.num_compute_rings = amdgpu_num_kcq;
+
  	gfx_v9_0_set_kiq_pm4_funcs(adev);
  	gfx_v9_0_set_ring_funcs(adev);
  	gfx_v9_0_set_irq_funcs(adev);

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

#3 lore raw

Re: [PATCH 2/2] drm/amdgpu: enable only one compute queue for raven

In reply to Nirmoy Das
Quoted 35 lines · Nirmoy Das

Because of firmware bug, Raven asics can't handle jobs scheduled to multiple compute queues. So enable only one compute queue till we have a firmware fix.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c |  4 ++++
 drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 11 ++++++++++-
 2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +4 -0
index 8c9bacfdbc30..ca2ac985b300 100644
@@ -195,6 +195,10 @@ static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
 bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
                                               int queue)
 {
+       /* We only enable one compute queue for Raven */
+       if (adev->asic_type == CHIP_RAVEN)
+               return false;
+
        /* Policy: make queue 0 of each pipe as high priority compute queue */
        return (queue == 0);
 
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +0 -1
index 0d8e203b10ef..f3fc9ad8bc20 100644
@@ -4633,7 +4633,16 @@ static int gfx_v9_0_early_init(void *handle)
                adev->gfx.num_gfx_rings = 0;
        else
                adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
-       adev->gfx.num_compute_rings = amdgpu_num_kcq;

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c static int gfx_v9_0_early_init(void *handle) · line 4645

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +10 -0
+
+       /* raven firmware currently can not load balance jobs
+        * among multiple compute queues. Enable only one
+        * compute queue till we have a firmware fix.
+        */
+       if (adev->asic_type == CHIP_RAVEN)
+               adev->gfx.num_compute_rings = 1;
+       else
+               adev->gfx.num_compute_rings = amdgpu_num_kcq;
+

I would suggest something like this instead so we can override easily for testing:

(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index abddcd9dab3d..a2954b41e59d 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
@@ -1376,6 +1376,12 @@ static int amdgpu_device_check_arguments(struct

amdgpu_device *adev)

        if (amdgpu_num_kcq == -1) {
                amdgpu_num_kcq = 8;
+               /* raven firmware currently can not load balance jobs
+                * among multiple compute queues. Enable only one
+                * compute queue till we have a firmware fix.
+                */
+               if (adev->asic_type == CHIP_RAVEN)
+                       amdgpu_num_kcq = 1;
        } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
                amdgpu_num_kcq = 8;
                dev_warn(adev->dev, "set kernel compute queue number

to 8 due to invalid parameter provided by user\n");

Alex

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +0 -1
        gfx_v9_0_set_kiq_pm4_funcs(adev);
        gfx_v9_0_set_ring_funcs(adev);
        gfx_v9_0_set_irq_funcs(adev);
--

2.28.0

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

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

#4 lore raw

Re: [PATCH 2/2] drm/amdgpu: enable only one compute queue for raven

Quoted 67 lines · Alex Deucher
Quoted 34 lines · Nirmoy Das

Because of firmware bug, Raven asics can't handle jobs scheduled to multiple compute queues. So enable only one compute queue till we have a firmware fix.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c |  4 ++++
  drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 11 ++++++++++-
  2 files changed, 14 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +4 -0
index 8c9bacfdbc30..ca2ac985b300 100644
@@ -195,6 +195,10 @@ static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
  bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
                                                int queue)
  {
+       /* We only enable one compute queue for Raven */
+       if (adev->asic_type == CHIP_RAVEN)
+               return false;
+
         /* Policy: make queue 0 of each pipe as high priority compute queue */
         return (queue == 0);
 
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +0 -1
index 0d8e203b10ef..f3fc9ad8bc20 100644
@@ -4633,7 +4633,16 @@ static int gfx_v9_0_early_init(void *handle)
                 adev->gfx.num_gfx_rings = 0;
         else
                 adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
-       adev->gfx.num_compute_rings = amdgpu_num_kcq;

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c static int gfx_v9_0_early_init(void *handle) · line 4645

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +10 -0
+
+       /* raven firmware currently can not load balance jobs
+        * among multiple compute queues. Enable only one
+        * compute queue till we have a firmware fix.
+        */
+       if (adev->asic_type == CHIP_RAVEN)
+               adev->gfx.num_compute_rings = 1;
+       else
+               adev->gfx.num_compute_rings = amdgpu_num_kcq;
+

I would suggest something like this instead so we can override easily for testing:

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +6 -0
index abddcd9dab3d..a2954b41e59d 100644
@@ -1376,6 +1376,12 @@ static int amdgpu_device_check_arguments(struct
 amdgpu_device *adev)
 
         if (amdgpu_num_kcq == -1) {
                 amdgpu_num_kcq = 8;
+               /* raven firmware currently can not load balance jobs
+                * among multiple compute queues. Enable only one
+                * compute queue till we have a firmware fix.
+                */
+               if (adev->asic_type == CHIP_RAVEN)
+                       amdgpu_num_kcq = 1;
         } else if (amdgpu_num_kcq > 8 || amdgpu_num_kcq < 0) {
                 amdgpu_num_kcq = 8;
                 dev_warn(adev->dev, "set kernel compute queue number
Quoted 2 lines

to 8 due to invalid parameter provided by user\n");

Thanks, this looks much better,

I will update.

Nirmoy

Quoted 12 lines

Alex

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +0 -1
         gfx_v9_0_set_kiq_pm4_funcs(adev);
         gfx_v9_0_set_ring_funcs(adev);
         gfx_v9_0_set_irq_funcs(adev);
--

2.28.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=04%7C01%7Cnirmoy.das%40amd.com%7Cc3012ca19bf149cb880608d871db5494%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637384534119165172%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=Rcd6aMUMxxvDcwi695IYNvvhHfpKAq74KAOT9Vpzvmo%3D&amp;reserved=0

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

Tested-by on v1 1/2
#5 lore raw
In reply to Nirmoy Das

[AMD Public Use]

Series is:

--
Best Regards
Aaron Liu

-----Original Message-----
From: Das, Nirmoy <[email protected]>
Sent: Wednesday, October 14, 2020 9:53 PM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Chen, Guchun <[email protected]>; Liu, Aaron <[email protected]>; Das, Nirmoy <[email protected]>
Subject: [PATCH 1/2] drm:amdgpu: check before setting hw priority

Check validity of drm_gpu_scheduler before setting hw priority. Also fix a minor indentation issue.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +6 -3
index a03398c87344..c8da0978d4f6 100644
@@ -504,9 +504,9 @@ struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,  }
 
 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
-					    struct amdgpu_ctx_entity *aentity,
-					    int hw_ip,
-					    enum drm_sched_priority priority)
+					   struct amdgpu_ctx_entity *aentity,
+					   int hw_ip,
+					   enum drm_sched_priority priority)
 {
 	struct amdgpu_device *adev = ctx->adev;
 	unsigned int hw_prio;
@@ -523,6 +523,9 @@ static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
 		hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
 		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
 		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
+		if (!scheds || !num_scheds)
+			return;
+
 		drm_sched_entity_modify_sched(&aentity->entity, scheds,
 					      num_scheds);
 	}

--
2.28.0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx