drm/amdgpu: Track pending retry faults in IH and VM (v2)

IH tracks pending retry faults in a hash table for fast lookup in
interrupt context. Each VM has a short FIFO of pending VM faults for
processing in a bottom half.

The IH prescreening stage adds retry faults and filters out repeated
retry interrupts to minimize the impact of interrupt storms.

It's the VM's responsibility remove pending faults once they are
handled. For now this is only done when the VM is destroyed.

v2:
- Made the hash table smaller and the FIFO longer. I never want the
  FIFO to fill up, because that would make prescreen take longer.
  128 pending page faults should be enough to keep migrations busy.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com> (v1)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 77 insertions, 1 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/vega10_ih.c b/drivers/gpu/drm/amd/amdgpu/vega10_ih.c
index eda4771e273f..dd6af2176d3e 100644
--- a/drivers/gpu/drm/amd/amdgpu/vega10_ih.c
+++ b/drivers/gpu/drm/amd/amdgpu/vega10_ih.c
@@ -235,8 +235,73 @@ static u32 vega10_ih_get_wptr(struct amdgpu_device *adev)
  */
 static bool vega10_ih_prescreen_iv(struct amdgpu_device *adev)
 {
-	/* TODO: Filter known pending page faults */
+	u32 ring_index = adev->irq.ih.rptr >> 2;
+	u32 dw0, dw3, dw4, dw5;
+	u16 pasid;
+	u64 addr, key;
+	struct amdgpu_vm *vm;
+	int r;
+
+	dw0 = le32_to_cpu(adev->irq.ih.ring[ring_index + 0]);
+	dw3 = le32_to_cpu(adev->irq.ih.ring[ring_index + 3]);
+	dw4 = le32_to_cpu(adev->irq.ih.ring[ring_index + 4]);
+	dw5 = le32_to_cpu(adev->irq.ih.ring[ring_index + 5]);
+
+	/* Filter retry page faults, let only the first one pass. If
+	 * there are too many outstanding faults, ignore them until
+	 * some faults get cleared.
+	 */
+	switch (dw0 & 0xff) {
+	case AMDGPU_IH_CLIENTID_VMC:
+	case AMDGPU_IH_CLIENTID_UTCL2:
+		break;
+	default:
+		/* Not a VM fault */
+		return true;
+	}
+
+	/* Not a retry fault */
+	if (!(dw5 & 0x80))
+		return true;
+
+	pasid = dw3 & 0xffff;
+	/* No PASID, can't identify faulting process */
+	if (!pasid)
+		return true;
+
+	addr = ((u64)(dw5 & 0xf) << 44) | ((u64)dw4 << 12);
+	key = AMDGPU_VM_FAULT(pasid, addr);
+	r = amdgpu_ih_add_fault(adev, key);
+
+	/* Hash table is full or the fault is already being processed,
+	 * ignore further page faults
+	 */
+	if (r != 0)
+		goto ignore_iv;
+
+	/* Track retry faults in per-VM fault FIFO. */
+	spin_lock(&adev->vm_manager.pasid_lock);
+	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
+	spin_unlock(&adev->vm_manager.pasid_lock);
+	if (WARN_ON_ONCE(!vm)) {
+		/* VM not found, process it normally */
+		amdgpu_ih_clear_fault(adev, key);
+		return true;
+	}
+	/* No locking required with single writer and single reader */
+	r = kfifo_put(&vm->faults, key);
+	if (!r) {
+		/* FIFO is full. Ignore it until there is space */
+		amdgpu_ih_clear_fault(adev, key);
+		goto ignore_iv;
+	}
+
+	/* It's the first fault for this address, process it normally */
 	return true;
+
+ignore_iv:
+	adev->irq.ih.rptr += 32;
+	return false;
 }
 
 /**
@@ -323,6 +388,14 @@ static int vega10_ih_sw_init(void *handle)
 	adev->irq.ih.use_doorbell = true;
 	adev->irq.ih.doorbell_index = AMDGPU_DOORBELL64_IH << 1;
 
+	adev->irq.ih.faults = kmalloc(sizeof(*adev->irq.ih.faults), GFP_KERNEL);
+	if (!adev->irq.ih.faults)
+		return -ENOMEM;
+	INIT_CHASH_TABLE(adev->irq.ih.faults->hash,
+			 AMDGPU_PAGEFAULT_HASH_BITS, 8, 0);
+	spin_lock_init(&adev->irq.ih.faults->lock);
+	adev->irq.ih.faults->count = 0;
+
 	r = amdgpu_irq_init(adev);
 
 	return r;
@@ -335,6 +408,9 @@ static int vega10_ih_sw_fini(void *handle)
 	amdgpu_irq_fini(adev);
 	amdgpu_ih_ring_fini(adev);
 
+	kfree(adev->irq.ih.faults);
+	adev->irq.ih.faults = NULL;
+
 	return 0;
 }