vmwgfx: Add fence events

Add a way to send DRM events down the gpu fifo by attaching them to
fence objects. This may be useful for Xserver swapbuffer throttling and
page-flip done notifications.

Bump version to 2.2 to signal the availability of the FENCE_EVENT ioctl.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c

@@ -94,6 +94,9 @@
 #define DRM_IOCTL_VMW_FENCE_UNREF				\
 	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_UNREF,		\
 		 struct drm_vmw_fence_arg)
+#define DRM_IOCTL_VMW_FENCE_EVENT				\
+	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_EVENT,		\
+		 struct drm_vmw_fence_event_arg)
 #define DRM_IOCTL_VMW_PRESENT					\
 	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_PRESENT,		\
 		 struct drm_vmw_present_arg)
@@ -150,6 +153,9 @@ static struct drm_ioctl_desc vmw_ioctls[] = {
 		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_FENCE_UNREF, vmw_fence_obj_unref_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
+	VMW_IOCTL_DEF(VMW_FENCE_EVENT,
+		      vmw_fence_event_ioctl,
+		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_GET_3D_CAP, vmw_get_cap_3d_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
 

drivers/gpu/drm/vmwgfx/vmwgfx_drv.h

@@ -40,9 +40,9 @@
 #include "ttm/ttm_module.h"
 #include "vmwgfx_fence.h"
 
-#define VMWGFX_DRIVER_DATE "20110927"
+#define VMWGFX_DRIVER_DATE "20111008"
 #define VMWGFX_DRIVER_MAJOR 2
-#define VMWGFX_DRIVER_MINOR 1
+#define VMWGFX_DRIVER_MINOR 2
 #define VMWGFX_DRIVER_PATCHLEVEL 0
 #define VMWGFX_FILE_PAGE_OFFSET 0x00100000
 #define VMWGFX_FIFO_STATIC_SIZE (1024*1024)
@@ -264,10 +264,12 @@ struct vmw_private {
 	wait_queue_head_t fence_queue;
 	wait_queue_head_t fifo_queue;
 	int fence_queue_waiters; /* Protected by hw_mutex */
+	int goal_queue_waiters; /* Protected by hw_mutex */
 	atomic_t fifo_queue_waiters;
 	uint32_t last_read_seqno;
 	spinlock_t irq_lock;
 	struct vmw_fence_manager *fman;
+	uint32_t irq_mask;
 
 	/*
 	 * Device state
@@ -532,7 +534,13 @@ extern int vmw_execbuf_fence_commands(struct drm_file *file_priv,
 				      struct vmw_private *dev_priv,
 				      struct vmw_fence_obj **p_fence,
 				      uint32_t *p_handle);
-
+extern void vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
+					struct vmw_fpriv *vmw_fp,
+					int ret,
+					struct drm_vmw_fence_rep __user
+					*user_fence_rep,
+					struct vmw_fence_obj *fence,
+					uint32_t fence_handle);
 
 /**
  * IRQs and wating - vmwgfx_irq.c
@@ -557,6 +565,8 @@ extern void vmw_update_seqno(struct vmw_private *dev_priv,
 				struct vmw_fifo_state *fifo_state);
 extern void vmw_seqno_waiter_add(struct vmw_private *dev_priv);
 extern void vmw_seqno_waiter_remove(struct vmw_private *dev_priv);
+extern void vmw_goal_waiter_add(struct vmw_private *dev_priv);
+extern void vmw_goal_waiter_remove(struct vmw_private *dev_priv);
 
 /**
  * Rudimentary fence-like objects currently used only for throttling -

drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

@@ -1057,7 +1057,7 @@ int vmw_execbuf_fence_commands(struct drm_file *file_priv,
  * object so we wait for it immediately, and then unreference the
  * user-space reference.
  */
-static void
+void
 vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
 			    struct vmw_fpriv *vmw_fp,
 			    int ret,

drivers/gpu/drm/vmwgfx/vmwgfx_fence.h

@@ -37,8 +37,14 @@ struct vmw_fence_manager;
  *
  *
  */
+enum vmw_action_type {
+	VMW_ACTION_EVENT = 0,
+	VMW_ACTION_MAX
+};
+
 struct vmw_fence_action {
 	struct list_head head;
+	enum vmw_action_type type;
 	void (*seq_passed) (struct vmw_fence_action *action);
 	void (*cleanup) (struct vmw_fence_action *action);
 };
@@ -66,8 +72,7 @@ extern void vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p);
 extern struct vmw_fence_obj *
 vmw_fence_obj_reference(struct vmw_fence_obj *fence);
 
-extern void vmw_fences_update(struct vmw_fence_manager *fman,
-			      u32 sequence);
+extern void vmw_fences_update(struct vmw_fence_manager *fman);
 
 extern bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
 				   uint32_t flags);
@@ -102,4 +107,7 @@ extern int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
 
 extern int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
 				     struct drm_file *file_priv);
+extern int vmw_fence_event_ioctl(struct drm_device *dev, void *data,
+				 struct drm_file *file_priv);
+
 #endif /* _VMWGFX_FENCE_H_ */

drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c

@@ -247,9 +247,8 @@ static int vmw_fifo_wait(struct vmw_private *dev_priv,
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		outl(SVGA_IRQFLAG_FIFO_PROGRESS,
 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
-			  SVGA_IRQFLAG_FIFO_PROGRESS);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_FIFO_PROGRESS;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);
@@ -271,9 +270,8 @@ static int vmw_fifo_wait(struct vmw_private *dev_priv,
 	mutex_lock(&dev_priv->hw_mutex);
 	if (atomic_dec_and_test(&dev_priv->fifo_queue_waiters)) {
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
-			  ~SVGA_IRQFLAG_FIFO_PROGRESS);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_FIFO_PROGRESS;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);

drivers/gpu/drm/vmwgfx/vmwgfx_irq.c

@@ -34,28 +34,30 @@ irqreturn_t vmw_irq_handler(DRM_IRQ_ARGS)
 {
 	struct drm_device *dev = (struct drm_device *)arg;
 	struct vmw_private *dev_priv = vmw_priv(dev);
-	uint32_t status;
+	uint32_t status, masked_status;
 
 	spin_lock(&dev_priv->irq_lock);
 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
+	masked_status = status & dev_priv->irq_mask;
 	spin_unlock(&dev_priv->irq_lock);
 
-	if (status & SVGA_IRQFLAG_ANY_FENCE) {
-		__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
-		uint32_t seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+	if (likely(status))
+		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
 
-		vmw_fences_update(dev_priv->fman, seqno);
+	if (!masked_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 (status & SVGA_IRQFLAG_FIFO_PROGRESS)
+
+	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
 		wake_up_all(&dev_priv->fifo_queue);
 
-	if (likely(status)) {
-		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		return IRQ_HANDLED;
-	}
 
-	return IRQ_NONE;
+	return IRQ_HANDLED;
 }
 
 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
@@ -78,7 +80,7 @@ void vmw_update_seqno(struct vmw_private *dev_priv,
 	if (dev_priv->last_read_seqno != seqno) {
 		dev_priv->last_read_seqno = seqno;
 		vmw_marker_pull(&fifo_state->marker_queue, seqno);
-		vmw_fences_update(dev_priv->fman, seqno);
+		vmw_fences_update(dev_priv->fman);
 	}
 }
 
@@ -189,9 +191,8 @@ void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
 		outl(SVGA_IRQFLAG_ANY_FENCE,
 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
-			  SVGA_IRQFLAG_ANY_FENCE);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_ANY_FENCE;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);
@@ -204,9 +205,39 @@ void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
 		unsigned long irq_flags;
 
 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
-		vmw_write(dev_priv, SVGA_REG_IRQMASK,
-			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
-			  ~SVGA_IRQFLAG_ANY_FENCE);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_ANY_FENCE;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+	}
+	mutex_unlock(&dev_priv->hw_mutex);
+}
+
+
+void vmw_goal_waiter_add(struct vmw_private *dev_priv)
+{
+	mutex_lock(&dev_priv->hw_mutex);
+	if (dev_priv->goal_queue_waiters++ == 0) {
+		unsigned long irq_flags;
+
+		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+		outl(SVGA_IRQFLAG_FENCE_GOAL,
+		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
+		dev_priv->irq_mask |= SVGA_IRQFLAG_FENCE_GOAL;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
+		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
+	}
+	mutex_unlock(&dev_priv->hw_mutex);
+}
+
+void vmw_goal_waiter_remove(struct vmw_private *dev_priv)
+{
+	mutex_lock(&dev_priv->hw_mutex);
+	if (--dev_priv->goal_queue_waiters == 0) {
+		unsigned long irq_flags;
+
+		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
+		dev_priv->irq_mask &= ~SVGA_IRQFLAG_FENCE_GOAL;
+		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
 	}
 	mutex_unlock(&dev_priv->hw_mutex);

include/drm/vmwgfx_drm.h

@@ -683,6 +683,51 @@ struct drm_vmw_fence_arg {
 };
 
 
+/*************************************************************************/
+/**
+ * DRM_VMW_FENCE_EVENT
+ *
+ * Queues an event on a fence to be delivered on the drm character device
+ * when the fence has signaled the DRM_VMW_FENCE_FLAG_EXEC flag.
+ * Optionally the approximate time when the fence signaled is
+ * given by the event.
+ */
+
+/*
+ * The event type
+ */
+#define DRM_VMW_EVENT_FENCE_SIGNALED 0x80000000
+
+struct drm_vmw_event_fence {
+	struct drm_event base;
+	uint64_t user_data;
+	uint32_t tv_sec;
+	uint32_t tv_usec;
+};
+
+/*
+ * Flags that may be given to the command.
+ */
+/* Request fence signaled time on the event. */
+#define DRM_VMW_FE_FLAG_REQ_TIME (1 << 0)
+
+/**
+ * struct drm_vmw_fence_event_arg
+ *
+ * @fence_rep: Pointer to fence_rep structure cast to uint64_t or 0 if
+ * the fence is not supposed to be referenced by user-space.
+ * @user_info: Info to be delivered with the event.
+ * @handle: Attach the event to this fence only.
+ * @flags: A set of flags as defined above.
+ */
+struct drm_vmw_fence_event_arg {
+	uint64_t fence_rep;
+	uint64_t user_data;
+	uint32_t handle;
+	uint32_t flags;
+};
+
+
 /*************************************************************************/
 /**
  * DRM_VMW_PRESENT
@@ -743,6 +788,4 @@ struct drm_vmw_present_readback_arg {
 	 uint64_t clips_ptr;
 	 uint64_t fence_rep;
 };
-
-
 #endif