Skip to contentSkip to conversation

drm/amdgpu/gfx10: implement iqueue reset via MMIO

PATCH In review [email protected] 2 participants 4 messages Jan 10, 2025 – Jan 14, 2025

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

PATCH v2 1/2 drm/amdgpu/gfx10: implement iqueue reset via MMIO

From: "[email protected]" <[email protected]>

Using mmio to do queue reset.

v2: Alignment this function with gfx9/gfx9.4.3.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 34 ++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +34 -0
index 5ba263fe5512..6157e1126030 100644
@@ -3790,12 +3790,46 @@ static void gfx10_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
 	gfx_v10_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1);
 }
 
+static void gfx_v10_0_kiq_reset_hw_queue(struct amdgpu_ring *kiq_ring, uint32_t queue_type,
+					uint32_t me_id, uint32_t pipe_id, uint32_t queue_id,
+					uint32_t xcc_id, uint32_t vmid)
+{
+	struct amdgpu_device *adev = kiq_ring->adev;
+	unsigned i;
+
+	/* enter save mode */
+	amdgpu_gfx_rlc_enter_safe_mode(adev, xcc_id);
+	mutex_lock(&adev->srbm_mutex);
+	nv_grbm_select(adev, me_id, pipe_id, queue_id, 0);
+
+	if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
+		WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 0x2);
+		WREG32_SOC15(GC, 0, mmSPI_COMPUTE_QUEUE_RESET, 0x1);
+		/* wait till dequeue take effects */
+		for (i = 0; i < adev->usec_timeout; i++) {
+			if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
+				break;
+			udelay(1);
+		}
+		if (i >= adev->usec_timeout)
+			dev_err(adev->dev, "fail to wait on hqd deactive\n");
+	} else {
+		dev_err(adev->dev, "reset queue_type(%d) not supported\n", queue_type);
+	}
+
+	nv_grbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+	/* exit safe mode */
+	amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
+}
+
 static const struct kiq_pm4_funcs gfx_v10_0_kiq_pm4_funcs = {
 	.kiq_set_resources = gfx10_kiq_set_resources,
 	.kiq_map_queues = gfx10_kiq_map_queues,
 	.kiq_unmap_queues = gfx10_kiq_unmap_queues,
 	.kiq_query_status = gfx10_kiq_query_status,
 	.kiq_invalidate_tlbs = gfx10_kiq_invalidate_tlbs,
+	.kiq_reset_hw_queue = gfx_v10_0_kiq_reset_hw_queue,
 	.set_resources_size = 8,
 	.map_queues_size = 7,
 	.unmap_queues_size = 6,
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 2/2 drm/amdgpu/gfx10: implement gfx queue reset via MMIO

From: "[email protected]" <[email protected]>

Using mmio to do queue reset

v2: Alignment the function with gfx9/gfx9.4.3.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +20 -1
index 6157e1126030..a3d726dd6fe5 100644
@@ -3796,6 +3796,7 @@ static void gfx_v10_0_kiq_reset_hw_queue(struct amdgpu_ring *kiq_ring, uint32_t
 {
 	struct amdgpu_device *adev = kiq_ring->adev;
 	unsigned i;
+	uint32_t tmp;
 
 	/* enter save mode */
 	amdgpu_gfx_rlc_enter_safe_mode(adev, xcc_id);
@@ -3813,7 +3814,25 @@ static void gfx_v10_0_kiq_reset_hw_queue(struct amdgpu_ring *kiq_ring, uint32_t
 		}
 		if (i >= adev->usec_timeout)
 			dev_err(adev->dev, "fail to wait on hqd deactive\n");
-	} else {
+	} else if (queue_type == AMDGPU_RING_TYPE_GFX) {
+		WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX,
+			     (uint32_t)(0x1 << GRBM_GFX_INDEX__SE_BROADCAST_WRITES__SHIFT));
+		tmp = REG_SET_FIELD(0, CP_VMID_RESET, RESET_REQUEST, 1 << vmid);
+		if (pipe_id == 0)
+			tmp = REG_SET_FIELD(tmp, CP_VMID_RESET, PIPE0_QUEUES, 1 << queue_id);
+		else
+			tmp = REG_SET_FIELD(tmp, CP_VMID_RESET, PIPE1_QUEUES, 1 << queue_id);
+		WREG32_SOC15(GC, 0, mmCP_VMID_RESET, tmp);
+
+		/* wait till dequeue take effects */
+		for (i = 0; i < adev->usec_timeout; i++) {
+			if (!(RREG32_SOC15(GC, 0, mmCP_GFX_HQD_ACTIVE) & 1))
+				break;
+			udelay(1);
+		}
+		if (i >= adev->usec_timeout)
+			dev_err(adev->dev, "failed to wait on gfx hqd deactivate\n");
+	}else {
 		dev_err(adev->dev, "reset queue_type(%d) not supported\n", queue_type);
 	}
 
Signature · 1 line
2.25.1
#2 lore raw
In reply to [email protected]

[AMD Official Use Only - AMD Internal Distribution Only]

Ping ...

-----Original Message-----
From: [email protected] <[email protected]>
Sent: Friday, January 10, 2025 1:22 PM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Huang, Tim <[email protected]>; Zhu, Jiadong <[email protected]>; Zhang, Jesse(Jie) <[email protected]>; Zhang, Jesse(Jie) <[email protected]>
Subject: [PATCH 1/2 V2] drm/amdgpu/gfx10: implement iqueue reset via MMIO

From: "[email protected]" <[email protected]>

Using mmio to do queue reset.

v2: Alignment this function with gfx9/gfx9.4.3.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 34 ++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +33 -1
index 5ba263fe5512..6157e1126030 100644
@@ -3790,12 +3790,46 @@ static void gfx10_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
        gfx_v10_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1);  }
 
+static void gfx_v10_0_kiq_reset_hw_queue(struct amdgpu_ring *kiq_ring, uint32_t queue_type,
+                                       uint32_t me_id, uint32_t pipe_id, uint32_t queue_id,
+                                       uint32_t xcc_id, uint32_t vmid)
+{
+       struct amdgpu_device *adev = kiq_ring->adev;
+       unsigned i;
+
+       /* enter save mode */
+       amdgpu_gfx_rlc_enter_safe_mode(adev, xcc_id);
+       mutex_lock(&adev->srbm_mutex);
+       nv_grbm_select(adev, me_id, pipe_id, queue_id, 0);
+
+       if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
+               WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 0x2);
+               WREG32_SOC15(GC, 0, mmSPI_COMPUTE_QUEUE_RESET, 0x1);
+               /* wait till dequeue take effects */
+               for (i = 0; i < adev->usec_timeout; i++) {
+                       if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
+                               break;
+                       udelay(1);
+               }
+               if (i >= adev->usec_timeout)
+                       dev_err(adev->dev, "fail to wait on hqd deactive\n");
+       } else {
+               dev_err(adev->dev, "reset queue_type(%d) not supported\n", queue_type);
+       }
+
+       nv_grbm_select(adev, 0, 0, 0, 0);
+       mutex_unlock(&adev->srbm_mutex);
+       /* exit safe mode */
+       amdgpu_gfx_rlc_exit_safe_mode(adev, 0); }
+
 static const struct kiq_pm4_funcs gfx_v10_0_kiq_pm4_funcs = {
        .kiq_set_resources = gfx10_kiq_set_resources,
        .kiq_map_queues = gfx10_kiq_map_queues,
        .kiq_unmap_queues = gfx10_kiq_unmap_queues,
        .kiq_query_status = gfx10_kiq_query_status,
        .kiq_invalidate_tlbs = gfx10_kiq_invalidate_tlbs,
+       .kiq_reset_hw_queue = gfx_v10_0_kiq_reset_hw_queue,
        .set_resources_size = 8,
        .map_queues_size = 7,
        .unmap_queues_size = 6,
--

2.25.1

Reviewed-by
#3 lore raw
Quoted 5 lines · Jesse(Jie)

[AMD Official Use Only - AMD Internal Distribution Only]

Ping ...

Series is:

Quoted 70 lines

-----Original Message-----
From: [email protected] <[email protected]>
Sent: Friday, January 10, 2025 1:22 PM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Huang, Tim <[email protected]>; Zhu, Jiadong <[email protected]>; Zhang, Jesse(Jie) <[email protected]>; Zhang, Jesse(Jie) <[email protected]>
Subject: [PATCH 1/2 V2] drm/amdgpu/gfx10: implement iqueue reset via MMIO

From: "[email protected]" <[email protected]>

Using mmio to do queue reset.

v2: Alignment this function with gfx9/gfx9.4.3.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 34 ++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +33 -1
index 5ba263fe5512..6157e1126030 100644
@@ -3790,12 +3790,46 @@ static void gfx10_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
        gfx_v10_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1);  }
 
+static void gfx_v10_0_kiq_reset_hw_queue(struct amdgpu_ring *kiq_ring, uint32_t queue_type,
+                                       uint32_t me_id, uint32_t pipe_id, uint32_t queue_id,
+                                       uint32_t xcc_id, uint32_t vmid)
+{
+       struct amdgpu_device *adev = kiq_ring->adev;
+       unsigned i;
+
+       /* enter save mode */
+       amdgpu_gfx_rlc_enter_safe_mode(adev, xcc_id);
+       mutex_lock(&adev->srbm_mutex);
+       nv_grbm_select(adev, me_id, pipe_id, queue_id, 0);
+
+       if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
+               WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 0x2);
+               WREG32_SOC15(GC, 0, mmSPI_COMPUTE_QUEUE_RESET, 0x1);
+               /* wait till dequeue take effects */
+               for (i = 0; i < adev->usec_timeout; i++) {
+                       if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
+                               break;
+                       udelay(1);
+               }
+               if (i >= adev->usec_timeout)
+                       dev_err(adev->dev, "fail to wait on hqd deactive\n");
+       } else {
+               dev_err(adev->dev, "reset queue_type(%d) not supported\n", queue_type);
+       }
+
+       nv_grbm_select(adev, 0, 0, 0, 0);
+       mutex_unlock(&adev->srbm_mutex);
+       /* exit safe mode */
+       amdgpu_gfx_rlc_exit_safe_mode(adev, 0); }
+
 static const struct kiq_pm4_funcs gfx_v10_0_kiq_pm4_funcs = {
        .kiq_set_resources = gfx10_kiq_set_resources,
        .kiq_map_queues = gfx10_kiq_map_queues,
        .kiq_unmap_queues = gfx10_kiq_unmap_queues,
        .kiq_query_status = gfx10_kiq_query_status,
        .kiq_invalidate_tlbs = gfx10_kiq_invalidate_tlbs,
+       .kiq_reset_hw_queue = gfx_v10_0_kiq_reset_hw_queue,
        .set_resources_size = 8,
        .map_queues_size = 7,
        .unmap_queues_size = 6,
--

2.25.1