drm/vmwgfx: Move irq bottom half processing to threads

This gets rid of the irq bottom half tasklets and instead performs the
work needed in process context. We also convert irq-disabling spinlocks to
ordinary spinlocks.

This should decrease system latency for other system components, like
sound for example but has the potential to increase latency for processes
that wait on the GPU.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Sinclair Yeh <syeh@vmware.com>

1 files changed, 57 insertions, 11 deletions

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
index 9b519c4b4ec2..b9239ba067c4 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
@@ -30,11 +30,56 @@
 
 #define VMW_FENCE_WRAP (1 << 24)
 
+/**
+ * vmw_thread_fn - Deferred (process context) irq handler
+ *
+ * @irq: irq number
+ * @arg: Closure argument. Pointer to a struct drm_device cast to void *
+ *
+ * This function implements the deferred part of irq processing.
+ * The function is guaranteed to run at least once after the
+ * vmw_irq_handler has returned with IRQ_WAKE_THREAD.
+ *
+ */
+static irqreturn_t vmw_thread_fn(int irq, void *arg)
+{
+	struct drm_device *dev = (struct drm_device *)arg;
+	struct vmw_private *dev_priv = vmw_priv(dev);
+	irqreturn_t ret = IRQ_NONE;
+
+	if (test_and_clear_bit(VMW_IRQTHREAD_FENCE,
+			       dev_priv->irqthread_pending)) {
+		vmw_fences_update(dev_priv->fman);
+		wake_up_all(&dev_priv->fence_queue);
+		ret = IRQ_HANDLED;
+	}
+
+	if (test_and_clear_bit(VMW_IRQTHREAD_CMDBUF,
+			       dev_priv->irqthread_pending)) {
+		vmw_cmdbuf_irqthread(dev_priv->cman);
+		ret = IRQ_HANDLED;
+	}
+
+	return ret;
+}
+
+/**
+ * vmw_irq_handler irq handler
+ *
+ * @irq: irq number
+ * @arg: Closure argument. Pointer to a struct drm_device cast to void *
+ *
+ * This function implements the quick part of irq processing.
+ * The function performs fast actions like clearing the device interrupt
+ * flags and also reasonably quick actions like waking processes waiting for
+ * FIFO space. Other IRQ actions are deferred to the IRQ thread.
+ */
 static irqreturn_t vmw_irq_handler(int irq, void *arg)
 {
 	struct drm_device *dev = (struct drm_device *)arg;
 	struct vmw_private *dev_priv = vmw_priv(dev);
 	uint32_t status, masked_status;
+	irqreturn_t ret = IRQ_HANDLED;
 
 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 	masked_status = status & READ_ONCE(dev_priv->irq_mask);
@@ -45,20 +90,21 @@ static irqreturn_t vmw_irq_handler(int irq, void *arg)
 	if (!status)
 		return IRQ_NONE;
 
-	if (masked_status & (SVGA_IRQFLAG_ANY_FENCE |
-			     SVGA_IRQFLAG_FENCE_GOAL)) {
-		vmw_fences_update(dev_priv->fman);
-		wake_up_all(&dev_priv->fence_queue);
-	}
-
 	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
 		wake_up_all(&dev_priv->fifo_queue);
 
-	if (masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
-			     SVGA_IRQFLAG_ERROR))
-		vmw_cmdbuf_tasklet_schedule(dev_priv->cman);
+	if ((masked_status & (SVGA_IRQFLAG_ANY_FENCE |
+			      SVGA_IRQFLAG_FENCE_GOAL)) &&
+	    !test_and_set_bit(VMW_IRQTHREAD_FENCE, dev_priv->irqthread_pending))
+		ret = IRQ_WAKE_THREAD;
+
+	if ((masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
+			      SVGA_IRQFLAG_ERROR)) &&
+	    !test_and_set_bit(VMW_IRQTHREAD_CMDBUF,
+			      dev_priv->irqthread_pending))
+		ret = IRQ_WAKE_THREAD;
 
-	return IRQ_HANDLED;
+	return ret;
 }
 
 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
@@ -326,7 +372,7 @@ int vmw_irq_install(struct drm_device *dev, int irq)
 
 	vmw_irq_preinstall(dev);
 
-	ret = request_threaded_irq(irq, vmw_irq_handler, NULL,
+	ret = request_threaded_irq(irq, vmw_irq_handler, vmw_thread_fn,
 				   IRQF_SHARED, VMWGFX_DRIVER_NAME, dev);
 	if (ret < 0)
 		return ret;