Skip to contentSkip to conversation

drm/amdgpu/mes: Simplify hqd mask initialization

PATCH Superseded Lang Yu 2 participants 3 messages Dec 22, 2025 – Jan 9

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Lang Yu

author
#0 lore raw

PATCH 1/1 drm/amdgpu/mes: Simplify hqd mask initialization

Let's get rid of version check and hardcode,
calculate hqd mask with number of queues per pipe
and number of gfx/compute queues kernel used.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 42 ++++++++++++-------------
 1 file changed, 20 insertions(+), 22 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +20 -22
index dffa0f7276b7..5b5962a727f2 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES	2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
 {
@@ -115,8 +114,10 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 	adev->mes.vmid_mask_mmhub = 0xFF00;
 	adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
 
-	queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
-	reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
+	queue_mask = (u32)(1UL << adev->gfx.me.num_queue_per_pipe) - 1;
+	reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_gfx_rings,
+					    adev->gfx.me.num_pipe_per_me) /
+					    adev->gfx.me.num_pipe_per_me)) - 1;
 
 	num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
 	if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
@@ -126,34 +127,26 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 	for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
 		if (i >= num_pipes)
 			break;
-		if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-		    IP_VERSION(12, 0, 0))
-			/*
-			 * GFX V12 has only one GFX pipe, but 8 queues in it.
-			 * GFX pipe 0 queue 0 is being used by Kernel queue.
-			 * Set GFX pipe 0 queue 1-7 for MES scheduling
-			 * mask = 1111 1110b
-			 */
-			adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-		else
-			/*
-			 * GFX pipe 0 queue 0 is being used by Kernel queue.
-			 * Set GFX pipe 0 queue 1 for MES scheduling
-			 * mask = 10b
-			 */
-			adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+		adev->mes.gfx_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
 	}
 
+	queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
+	reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_compute_rings,
+					    adev->gfx.mec.num_pipe_per_mec) /
+					    adev->gfx.mec.num_pipe_per_mec)) - 1;
+
 	num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
 	if (num_pipes > AMDGPU_MES_MAX_COMPUTE_PIPES)
 		dev_warn(adev->dev, "more compute pipes than supported by MES! (%d vs %d)\n",
 			 num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
 
 	for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-		if (i >= num_pipes)
+		/* skip MEC2, since only MEC1 is used */
+		if (i >= adev->gfx.mec.num_pipe_per_mec)
 			break;
-		adev->mes.compute_hqd_mask[i] =
-			adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+		adev->mes.compute_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
 	}
 
 	num_pipes = adev->sdma.num_instances;
@@ -167,6 +160,11 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 		adev->mes.sdma_hqd_mask[i] = 0xfc;
 	}
 
+	dev_info(adev->dev, "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+		adev->mes.vmid_mask_mmhub, adev->mes.vmid_mask_gfxhub);
+	dev_info(adev->dev, "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+		adev->mes.gfx_hqd_mask[0], adev->mes.compute_hqd_mask[0], adev->mes.sdma_hqd_mask[0]);
+
 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
 		r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
 		if (r) {
Signature · 1 line
2.34.1
maintainer Reviewed-by on v1 1/1
#1 lore raw

[Public]

________________________________
From: Yu, Lang <[email protected]>
Sent: Sunday, December 21, 2025 10:45 PM
To: Deucher, Alexander <[email protected]>; [email protected] <[email protected]>
Cc: Zhang, Hawking <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH] drm/amdgpu/mes: Simplify hqd mask initialization

Let's get rid of version check and hardcode,
calculate hqd mask with number of queues per pipe
and number of gfx/compute queues kernel used.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 42 ++++++++++++-------------
 1 file changed, 20 insertions(+), 22 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +20 -22
index dffa0f7276b7..5b5962a727f2 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES     2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
 {
@@ -115,8 +114,10 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
         adev->mes.vmid_mask_mmhub = 0xFF00;
         adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
 
-       queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
-       reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
+       queue_mask = (u32)(1UL << adev->gfx.me.num_queue_per_pipe) - 1;
+       reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_gfx_rings,
+                                           adev->gfx.me.num_pipe_per_me) /
+                                           adev->gfx.me.num_pipe_per_me)) - 1;
 
         num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
         if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
@@ -126,34 +127,26 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
         for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
                 if (i >= num_pipes)
                         break;
-               if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-                   IP_VERSION(12, 0, 0))
-                       /*
-                        * GFX V12 has only one GFX pipe, but 8 queues in it.
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1-7 for MES scheduling
-                        * mask = 1111 1110b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-               else
-                       /*
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1 for MES scheduling
-                        * mask = 10b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+               adev->mes.gfx_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
         }
 
+       queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
+       reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_compute_rings,
+                                           adev->gfx.mec.num_pipe_per_mec) /
+                                           adev->gfx.mec.num_pipe_per_mec)) - 1;
+
         num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
         if (num_pipes > AMDGPU_MES_MAX_COMPUTE_PIPES)
                 dev_warn(adev->dev, "more compute pipes than supported by MES! (%d vs %d)\n",
                          num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
 
         for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-               if (i >= num_pipes)
+               /* skip MEC2, since only MEC1 is used */
+               if (i >= adev->gfx.mec.num_pipe_per_mec)
                         break;
-               adev->mes.compute_hqd_mask[i] =
-                       adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+               adev->mes.compute_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
         }
 
         num_pipes = adev->sdma.num_instances;
@@ -167,6 +160,11 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                 adev->mes.sdma_hqd_mask[i] = 0xfc;
         }
 
+       dev_info(adev->dev, "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+               adev->mes.vmid_mask_mmhub, adev->mes.vmid_mask_gfxhub);
+       dev_info(adev->dev, "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+               adev->mes.gfx_hqd_mask[0], adev->mes.compute_hqd_mask[0], adev->mes.sdma_hqd_mask[0]);
+
         for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
                 r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
                 if (r) {

--
2.34.1

author Reviewed-by Alex Deucher
#2 lore raw

[Public]

Hi Alex,

Thanks for your review. This patch is buggy. Will send a v2 for review. 😊

I noticed that when kernel queue is explicitly disabled, gfx11 still sets num_gfx_rings to 1 to set up the clear state. This patch doesn’t work under such a case.

Regards,
Lang

From: Deucher, Alexander <[email protected]>
Sent: Friday, January 9, 2026 1:23 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Zhang, Hawking <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/mes: Simplify hqd mask initialization

[Public]

________________________________
From: Yu, Lang <[email protected]<mailto:[email protected]>>
Sent: Sunday, December 21, 2025 10:45 PM
To: Deucher, Alexander <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Zhang, Hawking <[email protected]<mailto:[email protected]>>; Yu, Lang <[email protected]<mailto:[email protected]>>
Subject: [PATCH] drm/amdgpu/mes: Simplify hqd mask initialization

Let's get rid of version check and hardcode,
calculate hqd mask with number of queues per pipe
and number of gfx/compute queues kernel used.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 42 ++++++++++++-------------
 1 file changed, 20 insertions(+), 22 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +20 -22
index dffa0f7276b7..5b5962a727f2 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES     2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
 {
@@ -115,8 +114,10 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
         adev->mes.vmid_mask_mmhub = 0xFF00;
         adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
 
-       queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
-       reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
+       queue_mask = (u32)(1UL << adev->gfx.me.num_queue_per_pipe) - 1;
+       reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_gfx_rings,
+                                           adev->gfx.me.num_pipe_per_me) /
+                                           adev->gfx.me.num_pipe_per_me)) - 1;
 
         num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
         if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
@@ -126,34 +127,26 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
         for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
                 if (i >= num_pipes)
                         break;
-               if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-                   IP_VERSION(12, 0, 0))
-                       /*
-                        * GFX V12 has only one GFX pipe, but 8 queues in it.
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1-7 for MES scheduling
-                        * mask = 1111 1110b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-               else
-                       /*
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1 for MES scheduling
-                        * mask = 10b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+               adev->mes.gfx_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
         }
 
+       queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
+       reserved_queue_mask = (u32)(1UL << (ALIGN(adev->gfx.num_compute_rings,
+                                           adev->gfx.mec.num_pipe_per_mec) /
+                                           adev->gfx.mec.num_pipe_per_mec)) - 1;
+
         num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
         if (num_pipes > AMDGPU_MES_MAX_COMPUTE_PIPES)
                 dev_warn(adev->dev, "more compute pipes than supported by MES! (%d vs %d)\n",
                          num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
 
         for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-               if (i >= num_pipes)
+               /* skip MEC2, since only MEC1 is used */
+               if (i >= adev->gfx.mec.num_pipe_per_mec)
                         break;
-               adev->mes.compute_hqd_mask[i] =
-                       adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+               adev->mes.compute_hqd_mask[i] = (queue_mask & ~reserved_queue_mask);
         }
 
         num_pipes = adev->sdma.num_instances;
@@ -167,6 +160,11 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                 adev->mes.sdma_hqd_mask[i] = 0xfc;
         }
 
+       dev_info(adev->dev, "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+               adev->mes.vmid_mask_mmhub, adev->mes.vmid_mask_gfxhub);
+       dev_info(adev->dev, "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+               adev->mes.gfx_hqd_mask[0], adev->mes.compute_hqd_mask[0], adev->mes.sdma_hqd_mask[0]);
+
         for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
                 r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
                 if (r) {

--
2.34.1