[PATCH 1/2] drm/amdgpu: fix using the reserved VMID with gang submit

[Christian König]

We need to ensure that even when using a reserved VMID that the gang
members can still run in parallel.

Signed-off-by: Christian König <christian.koenig at amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 21 ++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c    | 36 ++++++++++++++++------
 3 files changed, 45 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 1f71c7b98d77..e28fc07c7dbc 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1450,6 +1450,7 @@ u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
 				u32 reg);
 void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
 				u32 reg, u32 v);
+struct dma_fence *amdgpu_device_get_gang(struct amdgpu_device *adev);
 struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
 					    struct dma_fence *gang);
 bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index d549de26f931..94a6c0b1ae8e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -6532,6 +6532,22 @@ void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
 	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 }
 
+/**
+ * amdgpu_device_get_gang - return a reference to the current gang
+ * @adev: amdgpu_device pointer
+ *
+ * Returns: A new reference to the current gang leader.
+ */
+struct dma_fence *amdgpu_device_get_gang(struct amdgpu_device *adev)
+{
+	struct dma_fence *fence;
+
+	rcu_read_lock();
+	fence = dma_fence_get_rcu_safe(&adev->gang_submit);
+	rcu_read_unlock();
+	return fence;
+}
+
 /**
  * amdgpu_device_switch_gang - switch to a new gang
  * @adev: amdgpu_device pointer
@@ -6548,10 +6564,7 @@ struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
 
 	do {
 		dma_fence_put(old);
-		rcu_read_lock();
-		old = dma_fence_get_rcu_safe(&adev->gang_submit);
-		rcu_read_unlock();
-
+		old = amdgpu_device_get_gang(adev);
 		if (old == gang)
 			break;
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
index 3d7fcdeaf8cf..b5b9d4f40f53 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
@@ -290,18 +290,36 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
 	     !dma_fence_is_signaled((*id)->last_flush))) {
 		struct dma_fence *tmp;
 
-		/* Don't use per engine and per process VMID at the same time */
-		if (adev->vm_manager.concurrent_flush)
-			ring = NULL;
-
-		/* to prevent one context starved by another context */
-		(*id)->pd_gpu_addr = 0;
-		tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
-		if (tmp) {
+		/* Wait for the gang to be assembled before using a
+		 * reserved VMID or otherwise the gang could deadlock.
+		 */
+		tmp = amdgpu_device_get_gang(adev);
+		if (!dma_fence_is_signaled(tmp) && tmp != job->gang_submit) {
 			*id = NULL;
-			*fence = dma_fence_get(tmp);
+			*fence = tmp;
 			return 0;
 		}
+		dma_fence_put(tmp);
+
+		/* Make sure the id is owned by the gang before proceeding */
+		if (!job->gang_submit ||
+		    (*id)->owner != vm->immediate.fence_context) {
+
+			/* Don't use per engine and per process VMID at the
+			 * same time
+			 */
+			if (adev->vm_manager.concurrent_flush)
+				ring = NULL;
+
+			/* to prevent one context starved by another context */
+			(*id)->pd_gpu_addr = 0;
+			tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
+			if (tmp) {
+				*id = NULL;
+				*fence = dma_fence_get(tmp);
+				return 0;
+			}
+		}
 		needs_flush = true;
 	}
 
-- 
2.34.1