Skip to contentSkip to conversation

drm/amdgpu/gfx11: drop old bring up code

PATCH In review Alex Deucher 3 participants 17 messages Apr 27, 2023 – Apr 28, 2023

Subsystems (2)
lore
Share & tools
lore

Alex Deucher

author maintainer
#0 lore raw

PATCH 1/12 drm/amdgpu/gfx11: drop old bring up code

No longer used. Remove it.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 70 ++------------------------
 1 file changed, 3 insertions(+), 67 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +3 -67
index 8a4c4769e607..646003b2faf8 100644
@@ -3680,55 +3680,6 @@ static int gfx_v11_0_gfx_mqd_init(struct amdgpu_device *adev, void *m,
 	return 0;
 }
 
-#ifdef BRING_UP_DEBUG
-static int gfx_v11_0_gfx_queue_init_register(struct amdgpu_ring *ring)
-{
-	struct amdgpu_device *adev = ring->adev;
-	struct v11_gfx_mqd *mqd = ring->mqd_ptr;
-
-	/* set mmCP_GFX_HQD_WPTR/_HI to 0 */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_WPTR, mqd->cp_gfx_hqd_wptr);
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_WPTR_HI, mqd->cp_gfx_hqd_wptr_hi);
-
-	/* set GFX_MQD_BASE */
-	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr);
-	WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
-
-	/* set GFX_MQD_CONTROL */
-	WREG32_SOC15(GC, 0, regCP_GFX_MQD_CONTROL, mqd->cp_gfx_mqd_control);
-
-	/* set GFX_HQD_VMID to 0 */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_VMID, mqd->cp_gfx_hqd_vmid);
-
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY,
-			mqd->cp_gfx_hqd_queue_priority);
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_QUANTUM, mqd->cp_gfx_hqd_quantum);
-
-	/* set GFX_HQD_BASE, similar as CP_RB_BASE */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_BASE, mqd->cp_gfx_hqd_base);
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_BASE_HI, mqd->cp_gfx_hqd_base_hi);
-
-	/* set GFX_HQD_RPTR_ADDR, similar as CP_RB_RPTR */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_RPTR_ADDR, mqd->cp_gfx_hqd_rptr_addr);
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_RPTR_ADDR_HI, mqd->cp_gfx_hqd_rptr_addr_hi);
-
-	/* set GFX_HQD_CNTL, similar as CP_RB_CNTL */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_CNTL, mqd->cp_gfx_hqd_cntl);
-
-	/* set RB_WPTR_POLL_ADDR */
-	WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO, mqd->cp_rb_wptr_poll_addr_lo);
-	WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI, mqd->cp_rb_wptr_poll_addr_hi);
-
-	/* set RB_DOORBELL_CONTROL */
-	WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL, mqd->cp_rb_doorbell_control);
-
-	/* active the queue */
-	WREG32_SOC15(GC, 0, regCP_GFX_HQD_ACTIVE, mqd->cp_gfx_hqd_active);
-
-	return 0;
-}
-#endif
-
 static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -3740,9 +3691,6 @@ static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 		mutex_lock(&adev->srbm_mutex);
 		soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
 		amdgpu_ring_init_mqd(ring);
-#ifdef BRING_UP_DEBUG
-		gfx_v11_0_gfx_queue_init_register(ring);
-#endif
 		soc21_grbm_select(adev, 0, 0, 0, 0);
 		mutex_unlock(&adev->srbm_mutex);
 		if (adev->gfx.me.mqd_backup[mqd_idx])
@@ -3755,13 +3703,6 @@ static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 		ring->wptr = 0;
 		*ring->wptr_cpu_addr = 0;
 		amdgpu_ring_clear_ring(ring);
-#ifdef BRING_UP_DEBUG
-		mutex_lock(&adev->srbm_mutex);
-		soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
-		gfx_v11_0_gfx_queue_init_register(ring);
-		soc21_grbm_select(adev, 0, 0, 0, 0);
-		mutex_unlock(&adev->srbm_mutex);
-#endif
 	} else {
 		amdgpu_ring_clear_ring(ring);
 	}
@@ -3769,7 +3710,6 @@ static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 	return 0;
 }
 
-#ifndef BRING_UP_DEBUG
 static int gfx_v11_0_kiq_enable_kgq(struct amdgpu_device *adev)
 {
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
@@ -3791,7 +3731,6 @@ static int gfx_v11_0_kiq_enable_kgq(struct amdgpu_device *adev)
 
 	return amdgpu_ring_test_helper(kiq_ring);
 }
-#endif
 
 static int gfx_v11_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 {
@@ -3815,11 +3754,11 @@ static int gfx_v11_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 		if (r)
 			goto done;
 	}
-#ifndef BRING_UP_DEBUG
+
 	r = gfx_v11_0_kiq_enable_kgq(adev);
 	if (r)
 		goto done;
-#endif
+
 	r = gfx_v11_0_cp_gfx_start(adev);
 	if (r)
 		goto done;
@@ -4453,7 +4392,6 @@ static int gfx_v11_0_hw_init(void *handle)
 	return r;
 }
 
-#ifndef BRING_UP_DEBUG
 static int gfx_v11_0_kiq_disable_kgq(struct amdgpu_device *adev)
 {
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
@@ -4476,7 +4414,6 @@ static int gfx_v11_0_kiq_disable_kgq(struct amdgpu_device *adev)
 
 	return r;
 }
-#endif
 
 static int gfx_v11_0_hw_fini(void *handle)
 {
@@ -4488,13 +4425,12 @@ static int gfx_v11_0_hw_fini(void *handle)
 	amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
 
 	if (!adev->no_hw_access) {
-#ifndef BRING_UP_DEBUG
 		if (amdgpu_async_gfx_ring) {
 			r = gfx_v11_0_kiq_disable_kgq(adev);
 			if (r)
 				DRM_ERROR("KGQ disable failed\n");
 		}
-#endif
+
 		if (amdgpu_gfx_disable_kcq(adev, 0))
 			DRM_ERROR("KCQ disable failed\n");
 
Signature · 1 line
2.40.0
author maintainer
#1 lore raw

PATCH 2/12 drm/amdgpu/gfx10: drop old bring up code

No longer used. Remove it.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 70 ++------------------------
 1 file changed, 3 insertions(+), 67 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +3 -67
index 8bd07ff59671..25be4485dcd5 100644
@@ -6423,55 +6423,6 @@ static int gfx_v10_0_gfx_mqd_init(struct amdgpu_device *adev, void *m,
 	return 0;
 }
 
-#ifdef BRING_UP_DEBUG
-static int gfx_v10_0_gfx_queue_init_register(struct amdgpu_ring *ring)
-{
-	struct amdgpu_device *adev = ring->adev;
-	struct v10_gfx_mqd *mqd = ring->mqd_ptr;
-
-	/* set mmCP_GFX_HQD_WPTR/_HI to 0 */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_WPTR, mqd->cp_gfx_hqd_wptr);
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_WPTR_HI, mqd->cp_gfx_hqd_wptr_hi);
-
-	/* set GFX_MQD_BASE */
-	WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr);
-	WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
-
-	/* set GFX_MQD_CONTROL */
-	WREG32_SOC15(GC, 0, mmCP_GFX_MQD_CONTROL, mqd->cp_gfx_mqd_control);
-
-	/* set GFX_HQD_VMID to 0 */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_VMID, mqd->cp_gfx_hqd_vmid);
-
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_QUEUE_PRIORITY,
-			mqd->cp_gfx_hqd_queue_priority);
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_QUANTUM, mqd->cp_gfx_hqd_quantum);
-
-	/* set GFX_HQD_BASE, similar as CP_RB_BASE */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_BASE, mqd->cp_gfx_hqd_base);
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_BASE_HI, mqd->cp_gfx_hqd_base_hi);
-
-	/* set GFX_HQD_RPTR_ADDR, similar as CP_RB_RPTR */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_RPTR_ADDR, mqd->cp_gfx_hqd_rptr_addr);
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_RPTR_ADDR_HI, mqd->cp_gfx_hqd_rptr_addr_hi);
-
-	/* set GFX_HQD_CNTL, similar as CP_RB_CNTL */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_CNTL, mqd->cp_gfx_hqd_cntl);
-
-	/* set RB_WPTR_POLL_ADDR */
-	WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_LO, mqd->cp_rb_wptr_poll_addr_lo);
-	WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_HI, mqd->cp_rb_wptr_poll_addr_hi);
-
-	/* set RB_DOORBELL_CONTROL */
-	WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_CONTROL, mqd->cp_rb_doorbell_control);
-
-	/* active the queue */
-	WREG32_SOC15(GC, 0, mmCP_GFX_HQD_ACTIVE, mqd->cp_gfx_hqd_active);
-
-	return 0;
-}
-#endif
-
 static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -6492,9 +6443,6 @@ static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 		if (ring->doorbell_index == adev->doorbell_index.gfx_ring0 << 1)
 			gfx_v10_0_cp_gfx_set_doorbell(adev, ring);
 
-#ifdef BRING_UP_DEBUG
-		gfx_v10_0_gfx_queue_init_register(ring);
-#endif
 		nv_grbm_select(adev, 0, 0, 0, 0);
 		mutex_unlock(&adev->srbm_mutex);
 		if (adev->gfx.me.mqd_backup[mqd_idx])
@@ -6507,13 +6455,6 @@ static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 		ring->wptr = 0;
 		*ring->wptr_cpu_addr = 0;
 		amdgpu_ring_clear_ring(ring);
-#ifdef BRING_UP_DEBUG
-		mutex_lock(&adev->srbm_mutex);
-		nv_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
-		gfx_v10_0_gfx_queue_init_register(ring);
-		nv_grbm_select(adev, 0, 0, 0, 0);
-		mutex_unlock(&adev->srbm_mutex);
-#endif
 	} else {
 		amdgpu_ring_clear_ring(ring);
 	}
@@ -6521,7 +6462,6 @@ static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 	return 0;
 }
 
-#ifndef BRING_UP_DEBUG
 static int gfx_v10_0_kiq_enable_kgq(struct amdgpu_device *adev)
 {
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
@@ -6543,7 +6483,6 @@ static int gfx_v10_0_kiq_enable_kgq(struct amdgpu_device *adev)
 
 	return amdgpu_ring_test_helper(kiq_ring);
 }
-#endif
 
 static int gfx_v10_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 {
@@ -6567,11 +6506,11 @@ static int gfx_v10_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 		if (r)
 			goto done;
 	}
-#ifndef BRING_UP_DEBUG
+
 	r = gfx_v10_0_kiq_enable_kgq(adev);
 	if (r)
 		goto done;
-#endif
+
 	r = gfx_v10_0_cp_gfx_start(adev);
 	if (r)
 		goto done;
@@ -7239,7 +7178,6 @@ static int gfx_v10_0_hw_init(void *handle)
 	return r;
 }
 
-#ifndef BRING_UP_DEBUG
 static int gfx_v10_0_kiq_disable_kgq(struct amdgpu_device *adev)
 {
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
@@ -7261,7 +7199,6 @@ static int gfx_v10_0_kiq_disable_kgq(struct amdgpu_device *adev)
 	else
 		return 0;
 }
-#endif
 
 static int gfx_v10_0_hw_fini(void *handle)
 {
@@ -7272,13 +7209,12 @@ static int gfx_v10_0_hw_fini(void *handle)
 	amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
 
 	if (!adev->no_hw_access) {
-#ifndef BRING_UP_DEBUG
 		if (amdgpu_async_gfx_ring) {
 			r = gfx_v10_0_kiq_disable_kgq(adev);
 			if (r)
 				DRM_ERROR("KGQ disable failed\n");
 		}
-#endif
+
 		if (amdgpu_gfx_disable_kcq(adev, 0))
 			DRM_ERROR("KCQ disable failed\n");
 	}
Signature · 1 line
2.40.0
author maintainer
#2 lore raw

PATCH 3/12 drm/amdgpu: add [en/dis]able_kgq() functions

In reply to Alex Deucher

To replace the IP specific variants which are largely duplicate.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 68 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h |  2 +
 2 files changed, 70 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +68 -0
index 60bb4bba1994..0560568b3925 100644
@@ -518,6 +518,39 @@ int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
 	return r;
 }
 
+int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
+{
+	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
+	struct amdgpu_ring *kiq_ring = &kiq->ring;
+	int i, r = 0;
+	int j;
+
+	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
+		return -EINVAL;
+
+	spin_lock(&kiq->ring_lock);
+	if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
+		if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
+						adev->gfx.num_gfx_rings)) {
+			spin_unlock(&kiq->ring_lock);
+			return -ENOMEM;
+		}
+
+		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
+			j = i + xcc_id * adev->gfx.num_gfx_rings;
+			kiq->pmf->kiq_unmap_queues(kiq_ring,
+						   &adev->gfx.gfx_ring[i],
+						   PREEMPT_QUEUES, 0, 0);
+		}
+	}
+
+	if (adev->gfx.kiq[0].ring.sched.ready && !adev->job_hang)
+		r = amdgpu_ring_test_helper(kiq_ring);
+	spin_unlock(&kiq->ring_lock);
+
+	return r;
+}
+
 int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev,
 					int queue_bit)
 {
@@ -589,6 +622,41 @@ int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id)
 	return r;
 }
 
+int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
+{
+	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
+	struct amdgpu_ring *kiq_ring = &kiq->ring;
+	int r, i, j;
+
+	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
+		return -EINVAL;
+
+	spin_lock(&kiq->ring_lock);
+	/* No need to map kcq on the slave */
+	if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
+		r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
+						adev->gfx.num_gfx_rings);
+		if (r) {
+			DRM_ERROR("Failed to lock KIQ (%d).\n", r);
+			spin_unlock(&adev->gfx.kiq[0].ring_lock);
+			return r;
+		}
+
+		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
+			j = i + xcc_id * adev->gfx.num_gfx_rings;
+			kiq->pmf->kiq_map_queues(kiq_ring,
+						 &adev->gfx.gfx_ring[i]);
+		}
+	}
+
+	r = amdgpu_ring_test_helper(kiq_ring);
+	spin_unlock(&kiq->ring_lock);
+	if (r)
+		DRM_ERROR("KCQ enable failed\n");
+
+	return r;
+}
+
 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
  *
  * @adev: amdgpu_device pointer
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +2 -0
index f1a2ce7b6aca..2755f00ac19a 100644
@@ -433,6 +433,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id);
 int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id);
 int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id);
+int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id);
+int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id);
 
 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev);
 void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev);
Signature · 1 line
2.40.0
author maintainer
#3 lore raw

PATCH 4/12 drm/amdgpu/gfx10: use generic [en/dis]able_kgq() helpers

In reply to Alex Deucher

And remove the duplicate local variants.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 48 ++------------------------
 1 file changed, 2 insertions(+), 46 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +2 -46
index 25be4485dcd5..24d7134228b0 100644
@@ -6462,28 +6462,6 @@ static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 	return 0;
 }
 
-static int gfx_v10_0_kiq_enable_kgq(struct amdgpu_device *adev)
-{
-	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
-	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
-	int r, i;
-
-	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
-		return -EINVAL;
-
-	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
-					adev->gfx.num_gfx_rings);
-	if (r) {
-		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
-		return r;
-	}
-
-	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
-		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.gfx_ring[i]);
-
-	return amdgpu_ring_test_helper(kiq_ring);
-}
-
 static int gfx_v10_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 {
 	int r, i;
@@ -6507,7 +6485,7 @@ static int gfx_v10_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 			goto done;
 	}
 
-	r = gfx_v10_0_kiq_enable_kgq(adev);
+	r = amdgpu_gfx_enable_kgq(adev, 0);
 	if (r)
 		goto done;
 
@@ -7178,28 +7156,6 @@ static int gfx_v10_0_hw_init(void *handle)
 	return r;
 }
 
-static int gfx_v10_0_kiq_disable_kgq(struct amdgpu_device *adev)
-{
-	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
-	struct amdgpu_ring *kiq_ring = &kiq->ring;
-	int i;
-
-	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
-		return -EINVAL;
-
-	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
-					adev->gfx.num_gfx_rings))
-		return -ENOMEM;
-
-	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
-		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.gfx_ring[i],
-					   PREEMPT_QUEUES, 0, 0);
-	if (!adev->job_hang)
-		return amdgpu_ring_test_helper(kiq_ring);
-	else
-		return 0;
-}
-
 static int gfx_v10_0_hw_fini(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
@@ -7210,7 +7166,7 @@ static int gfx_v10_0_hw_fini(void *handle)
 
 	if (!adev->no_hw_access) {
 		if (amdgpu_async_gfx_ring) {
-			r = gfx_v10_0_kiq_disable_kgq(adev);
+			r = amdgpu_gfx_disable_kgq(adev, 0);
 			if (r)
 				DRM_ERROR("KGQ disable failed\n");
 		}
Signature · 1 line
2.40.0
author maintainer
#4 lore raw

PATCH 5/12 drm/amdgpu/gfx11: use generic [en/dis]able_kgq() helpers

In reply to Alex Deucher

And remove the duplicate local variants.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 49 ++------------------------
 1 file changed, 2 insertions(+), 47 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +2 -47
index 646003b2faf8..d36d365cb582 100644
@@ -3710,28 +3710,6 @@ static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 	return 0;
 }
 
-static int gfx_v11_0_kiq_enable_kgq(struct amdgpu_device *adev)
-{
-	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
-	struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
-	int r, i;
-
-	if (!kiq->pmf || !kiq->pmf->kiq_map_queues)
-		return -EINVAL;
-
-	r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size *
-					adev->gfx.num_gfx_rings);
-	if (r) {
-		DRM_ERROR("Failed to lock KIQ (%d).\n", r);
-		return r;
-	}
-
-	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
-		kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.gfx_ring[i]);
-
-	return amdgpu_ring_test_helper(kiq_ring);
-}
-
 static int gfx_v11_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 {
 	int r, i;
@@ -3755,7 +3733,7 @@ static int gfx_v11_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
 			goto done;
 	}
 
-	r = gfx_v11_0_kiq_enable_kgq(adev);
+	r = amdgpu_gfx_enable_kgq(adev, 0);
 	if (r)
 		goto done;
 
@@ -4392,29 +4370,6 @@ static int gfx_v11_0_hw_init(void *handle)
 	return r;
 }
 
-static int gfx_v11_0_kiq_disable_kgq(struct amdgpu_device *adev)
-{
-	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
-	struct amdgpu_ring *kiq_ring = &kiq->ring;
-	int i, r = 0;
-
-	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
-		return -EINVAL;
-
-	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size *
-					adev->gfx.num_gfx_rings))
-		return -ENOMEM;
-
-	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
-		kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.gfx_ring[i],
-					   PREEMPT_QUEUES, 0, 0);
-
-	if (adev->gfx.kiq[0].ring.sched.ready)
-		r = amdgpu_ring_test_helper(kiq_ring);
-
-	return r;
-}
-
 static int gfx_v11_0_hw_fini(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
@@ -4426,7 +4381,7 @@ static int gfx_v11_0_hw_fini(void *handle)
 
 	if (!adev->no_hw_access) {
 		if (amdgpu_async_gfx_ring) {
-			r = gfx_v11_0_kiq_disable_kgq(adev);
+			r = amdgpu_gfx_disable_kgq(adev, 0);
 			if (r)
 				DRM_ERROR("KGQ disable failed\n");
 		}
Signature · 1 line
2.40.0
author maintainer
#5 lore raw

PATCH 6/12 drm/amdgpu/gfx10: drop unused variable

In reply to Alex Deucher

Just check the return value directly.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -3
index 24d7134228b0..5c67c91c4297 100644
@@ -7159,15 +7159,13 @@ static int gfx_v10_0_hw_init(void *handle)
 static int gfx_v10_0_hw_fini(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-	int r;
 
 	amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
 	amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
 
 	if (!adev->no_hw_access) {
 		if (amdgpu_async_gfx_ring) {
-			r = amdgpu_gfx_disable_kgq(adev, 0);
-			if (r)
+			if (amdgpu_gfx_disable_kgq(adev, 0))
 				DRM_ERROR("KGQ disable failed\n");
 		}
 
Signature · 1 line
2.40.0
Reviewed-by on v1 6/12
#6 lore raw

Re: [PATCH 06/12] drm/amdgpu/gfx10: drop unused variable

Am 27.04.23 um 17:27 schrieb Alex Deucher:

Quoted 3 lines

Just check the return value directly.

Quoted 25 lines
---
  drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 4 +---
  1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -3
index 24d7134228b0..5c67c91c4297 100644
@@ -7159,15 +7159,13 @@ static int gfx_v10_0_hw_init(void *handle)
  static int gfx_v10_0_hw_fini(void *handle)
  {
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-	int r;
 
  	amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
  	amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
 
  	if (!adev->no_hw_access) {
  		if (amdgpu_async_gfx_ring) {
-			r = amdgpu_gfx_disable_kgq(adev, 0);
-			if (r)
+			if (amdgpu_gfx_disable_kgq(adev, 0))
  				DRM_ERROR("KGQ disable failed\n");
  		}
 
author maintainer
#7 lore raw

PATCH 7/12 drm/amdgpu/gfx11: drop unused variable

In reply to Alex Deucher

Just check the return value directly.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +1 -3
index d36d365cb582..256014a8c824 100644
@@ -4373,7 +4373,6 @@ static int gfx_v11_0_hw_init(void *handle)
 static int gfx_v11_0_hw_fini(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-	int r;
 
 	amdgpu_irq_put(adev, &adev->gfx.cp_ecc_error_irq, 0);
 	amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
@@ -4381,8 +4380,7 @@ static int gfx_v11_0_hw_fini(void *handle)
 
 	if (!adev->no_hw_access) {
 		if (amdgpu_async_gfx_ring) {
-			r = amdgpu_gfx_disable_kgq(adev, 0);
-			if (r)
+			if (amdgpu_gfx_disable_kgq(adev, 0))
 				DRM_ERROR("KGQ disable failed\n");
 		}
 
Signature · 1 line
2.40.0
Reviewed-by on v1 7/12
#8 lore raw

Re: [PATCH 07/12] drm/amdgpu/gfx11: drop unused variable

Am 27.04.23 um 17:27 schrieb Alex Deucher:

Quoted 3 lines

Just check the return value directly.

Quoted 26 lines
---
  drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 4 +---
  1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +1 -3
index d36d365cb582..256014a8c824 100644
@@ -4373,7 +4373,6 @@ static int gfx_v11_0_hw_init(void *handle)
  static int gfx_v11_0_hw_fini(void *handle)
  {
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-	int r;
 
  	amdgpu_irq_put(adev, &adev->gfx.cp_ecc_error_irq, 0);
  	amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
@@ -4381,8 +4380,7 @@ static int gfx_v11_0_hw_fini(void *handle)
 
  	if (!adev->no_hw_access) {
  		if (amdgpu_async_gfx_ring) {
-			r = amdgpu_gfx_disable_kgq(adev, 0);
-			if (r)
+			if (amdgpu_gfx_disable_kgq(adev, 0))
  				DRM_ERROR("KGQ disable failed\n");
  		}
 
author maintainer
#9 lore raw

PATCH 8/12 drm/amdgpu/gfx8: always restore kcq MQDs

In reply to Alex Deucher

Always restore the MQD not just when we do a reset. This allows us to move the MQD to VRAM if we want.

v2: always reset ring pointer as well (Christian)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -4
index 5de44d7e92de..2ae7f167985f 100644
@@ -4651,15 +4651,13 @@ static int gfx_v8_0_kcq_init_queue(struct amdgpu_ring *ring)
 
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct vi_mqd_allocation));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(struct vi_mqd_allocation));
 		/* reset ring buffer */
 		ring->wptr = 0;
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 	return 0;
 }
Signature · 1 line
2.40.0
Reviewed-by on v1 8/12
#10 lore raw

Re: [PATCH 08/12] drm/amdgpu/gfx8: always restore kcq MQDs

Am 27.04.23 um 17:27 schrieb Alex Deucher:

Quoted 6 lines

Always restore the MQD not just when we do a reset. This allows us to move the MQD to VRAM if we want.

v2: always reset ring pointer as well (Christian)

Quoted 26 lines
---
  drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 6 ++----
  1 file changed, 2 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -4
index 5de44d7e92de..2ae7f167985f 100644
@@ -4651,15 +4651,13 @@ static int gfx_v8_0_kcq_init_queue(struct amdgpu_ring *ring)
 
  		if (adev->gfx.mec.mqd_backup[mqd_idx])
  			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct vi_mqd_allocation));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
  		if (adev->gfx.mec.mqd_backup[mqd_idx])
  			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(struct vi_mqd_allocation));
  		/* reset ring buffer */
  		ring->wptr = 0;
  		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
  	}
  	return 0;
  }
author maintainer
#11 lore raw

PATCH 9/12 drm/amdgpu/gfx9: always restore kcq MQDs

In reply to Alex Deucher

Always restore the MQD not just when we do a reset. This allows us to move the MQD to VRAM if we want.

v2: always reset ring pointer as well (Christian)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 7 ++-----
 drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 7 ++-----
 2 files changed, 4 insertions(+), 10 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +2 -5
index 4ef3e7d9b108..cb01ba5da1f1 100644
@@ -3591,17 +3591,14 @@ static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring)
 
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(struct v9_mqd_allocation));
-
 		/* reset ring buffer */
 		ring->wptr = 0;
 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0);
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +2 -5
index 56a415e151d4..3e552684f464 100644
@@ -1799,17 +1799,14 @@ static int gfx_v9_4_3_kcq_init_queue(struct amdgpu_ring *ring, int xcc_id)
 
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(struct v9_mqd_allocation));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(struct v9_mqd_allocation));
-
 		/* reset ring buffer */
 		ring->wptr = 0;
 		atomic64_set((atomic64_t *)&adev->wb.wb[ring->wptr_offs], 0);
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
Signature · 1 line
2.40.0
author maintainer
#12 lore raw

PATCH 10/12 drm/amdgpu/gfx10: always restore kcq/kgq MQDs

In reply to Alex Deucher

Always restore the MQD not just when we do a reset. This allows us to move the MQD to VRAM if we want.

v2: always reset ring pointer as well (Christian)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +4 -9
index 5c67c91c4297..2e0234b43f43 100644
@@ -6447,16 +6447,14 @@ static int gfx_v10_0_gfx_init_queue(struct amdgpu_ring *ring)
 		mutex_unlock(&adev->srbm_mutex);
 		if (adev->gfx.me.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
-	} else if (amdgpu_in_reset(adev)) {
-		/* reset mqd with the backup copy */
+	} else {
+		/* restore mqd with the backup copy */
 		if (adev->gfx.me.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd));
 		/* reset the ring */
 		ring->wptr = 0;
 		*ring->wptr_cpu_addr = 0;
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
@@ -6780,17 +6778,14 @@ static int gfx_v10_0_kcq_init_queue(struct amdgpu_ring *ring)
 
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
-
 		/* reset ring buffer */
 		ring->wptr = 0;
 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0);
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
Signature · 1 line
2.40.0
author maintainer
#13 lore raw

PATCH 11/12 drm/amdgpu/gfx11: always restore kcq/kgq MQDs

In reply to Alex Deucher

Always restore the MQD not just when we do a reset. This allows us to move the MQD to VRAM if we want.

v2: always reset ring pointer as well (Christian)

---
 drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +4 -9
index 256014a8c824..f09e2558e73b 100644
@@ -3695,16 +3695,14 @@ static int gfx_v11_0_gfx_init_queue(struct amdgpu_ring *ring)
 		mutex_unlock(&adev->srbm_mutex);
 		if (adev->gfx.me.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
-	} else if (amdgpu_in_reset(adev)) {
-		/* reset mqd with the backup copy */
+	} else {
+		/* restore mqd with the backup copy */
 		if (adev->gfx.me.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd));
 		/* reset the ring */
 		ring->wptr = 0;
 		*ring->wptr_cpu_addr = 0;
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
@@ -4043,17 +4041,14 @@ static int gfx_v11_0_kcq_init_queue(struct amdgpu_ring *ring)
 
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
-	} else if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
-		/* reset MQD to a clean status */
+	} else {
+		/* restore MQD to a clean status */
 		if (adev->gfx.mec.mqd_backup[mqd_idx])
 			memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
-
 		/* reset ring buffer */
 		ring->wptr = 0;
 		atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0);
 		amdgpu_ring_clear_ring(ring);
-	} else {
-		amdgpu_ring_clear_ring(ring);
 	}
 
 	return 0;
Signature · 1 line
2.40.0
author maintainer
#14 lore raw

PATCH 12/12 drm/amdgpu: put MQDs in VRAM

In reply to Alex Deucher

Reduces preemption latency.

v2: move MES MQDs into VRAM as well (YuBiao)
v3: enable on gfx10, 11 only (Alex)

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 4 ++++
 drivers/gpu/drm/amd/amdgpu/mes_v10_1.c  | 1 +
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 1 +
 3 files changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +4 -0
index 0560568b3925..92c5f0ce8bbb 100644
@@ -382,6 +382,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 	int r, i;
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
 	struct amdgpu_ring *ring = &kiq->ring;
+	u32 domain_vram = adev->ip_versions[GC_HWIP][0] >= IP_VERSION(10, 0, 0) ?
+		AMDGPU_GEM_DOMAIN_VRAM : 0;
 
 	/* create MQD for KIQ */
 	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
@@ -413,6 +415,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 			ring = &adev->gfx.gfx_ring[i];
 			if (!ring->mqd_obj) {
 				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+							    domain_vram |
 							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 				if (r) {
@@ -434,6 +437,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 		ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
 		if (!ring->mqd_obj) {
 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+						    domain_vram |
 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 			if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v10_1.c +1 -0
index 0599f8a6813e..4560476c7c31 100644
@@ -901,6 +901,7 @@ static int mes_v10_1_mqd_sw_init(struct amdgpu_device *adev,
 		return 0;
 
 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+				    AMDGPU_GEM_DOMAIN_VRAM |
 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 	if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +1 -0
index e853bcb892fc..3adb450eec07 100644
@@ -999,6 +999,7 @@ static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
 		return 0;
 
 	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+				    AMDGPU_GEM_DOMAIN_VRAM |
 				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
 				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 	if (r) {
Signature · 1 line
2.40.0
#15 lore raw

Re: [PATCH 12/12] drm/amdgpu: put MQDs in VRAM

Am 27.04.23 um 17:27 schrieb Alex Deucher:

Quoted 4 lines

Reduces preemption latency.

v2: move MES MQDs into VRAM as well (YuBiao)
v3: enable on gfx10, 11 only (Alex)

The why we do that not for gfx9 is missing.

Quoted 13 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 4 ++++
  drivers/gpu/drm/amd/amdgpu/mes_v10_1.c  | 1 +
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 1 +
  3 files changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
index 0560568b3925..92c5f0ce8bbb 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, · line 386

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -0
@@ -382,6 +382,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
  	int r, i;
  	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
  	struct amdgpu_ring *ring = &kiq->ring;
+	u32 domain_vram = adev->ip_versions[GC_HWIP][0] >= IP_VERSION(10, 0, 0) ?
+		AMDGPU_GEM_DOMAIN_VRAM : 0;

Maybe cleaner to do something like:

    domain = AMDGPU_GEM_DOMAIN_GTT;
    if (...)
        domain |= AMDGPU_GEM_DOMAIN_VRAM;

Christian.

Quoted 43 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -0
  	/* create MQD for KIQ */
  	if (!adev->enable_mes_kiq && !ring->mqd_obj) {
@@ -413,6 +415,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
  			ring = &adev->gfx.gfx_ring[i];
  			if (!ring->mqd_obj) {
  				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+							    domain_vram |
  							    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
  							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
  				if (r) {
@@ -434,6 +437,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
  		ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
  		if (!ring->mqd_obj) {
  			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+						    domain_vram |
  						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
  						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
  			if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v10_1.c +1 -0
index 0599f8a6813e..4560476c7c31 100644
@@ -901,6 +901,7 @@ static int mes_v10_1_mqd_sw_init(struct amdgpu_device *adev,
  		return 0;
 
  	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+				    AMDGPU_GEM_DOMAIN_VRAM |
  				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
  				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
  	if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +1 -0
index e853bcb892fc..3adb450eec07 100644
@@ -999,6 +999,7 @@ static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
  		return 0;
 
  	r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+				    AMDGPU_GEM_DOMAIN_VRAM |
  				    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
  				    &ring->mqd_gpu_addr, &ring->mqd_ptr);
  	if (r) {
#16 lore raw

Re: [PATCH 12/12] drm/amdgpu: put MQDs in VRAM

Quoted 8 lines · Christian König

Am 27.04.23 um 17:27 schrieb Alex Deucher:

Quoted 4 lines

Reduces preemption latency.

v2: move MES MQDs into VRAM as well (YuBiao)
v3: enable on gfx10, 11 only (Alex)

The why we do that not for gfx9 is missing.

We could do it for gfx8-11. That said, gfx8 and 9 are working fine and there's no reasons to change them at this point. Less chance for regressions. I'll update the commit message.

Quoted 22 lines
Quoted 14 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 4 ++++
  drivers/gpu/drm/amd/amdgpu/mes_v10_1.c  | 1 +
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c  | 1 +
  3 files changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
index 0560568b3925..92c5f0ce8bbb 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, · line 386

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -0
@@ -382,6 +382,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
      int r, i;
      struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
      struct amdgpu_ring *ring = &kiq->ring;
+     u32 domain_vram = adev->ip_versions[GC_HWIP][0] >= IP_VERSION(10, 0, 0) ?
+             AMDGPU_GEM_DOMAIN_VRAM : 0;

Maybe cleaner to do something like:

Quoted 4 lines
     domain = AMDGPU_GEM_DOMAIN_GTT;
     if (...)
         domain |= AMDGPU_GEM_DOMAIN_VRAM;

I can fix that up.

Thanks,

Alex

Quoted 46 lines

Christian.

Quoted 43 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -0
      /* create MQD for KIQ */
      if (!adev->enable_mes_kiq && !ring->mqd_obj) {
@@ -413,6 +415,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
                      ring = &adev->gfx.gfx_ring[i];
                      if (!ring->mqd_obj) {
                              r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+                                                         domain_vram |
                                                          AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
                                                          &ring->mqd_gpu_addr, &ring->mqd_ptr);
                              if (r) {
@@ -434,6 +437,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
              ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
              if (!ring->mqd_obj) {
                      r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+                                                 domain_vram |
                                                  AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
                                                  &ring->mqd_gpu_addr, &ring->mqd_ptr);
                      if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v10_1.c +1 -0
index 0599f8a6813e..4560476c7c31 100644
@@ -901,6 +901,7 @@ static int mes_v10_1_mqd_sw_init(struct amdgpu_device *adev,
              return 0;
 
      r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+                                 AMDGPU_GEM_DOMAIN_VRAM |
                                  AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
                                  &ring->mqd_gpu_addr, &ring->mqd_ptr);
      if (r) {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +1 -0
index e853bcb892fc..3adb450eec07 100644
@@ -999,6 +999,7 @@ static int mes_v11_0_mqd_sw_init(struct amdgpu_device *adev,
              return 0;
 
      r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+                                 AMDGPU_GEM_DOMAIN_VRAM |
                                  AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
                                  &ring->mqd_gpu_addr, &ring->mqd_ptr);
      if (r) {