i915: Attempt to schedule buffer swap on target vertical blank when possible.

This has some advantages over the traditional way of first waiting for the
target vertical blank and then emitting the buffer swap, e.g.

* glXSwapBuffers returns immediately, only the next time the driver needs the
  hardware lock will it block until the target vertical blank. This should
  allow applications that don't intermix rendering and other processing to
  start processing for the next frame right away.
* It's less likely to produce tearing.

src/mesa/drivers/dri/i915/intel_batchbuffer.c

@@ -368,6 +368,7 @@ void intelCopyBuffer( const __DRIdrawablePrivate *dPriv,
 		      const drm_clip_rect_t	 *rect)
 {
    intelContextPtr intel;
+   const intelScreenPrivate *intelScreen;
    GLboolean   missed_target;
    int64_t ust;
 
@@ -382,6 +383,38 @@ void intelCopyBuffer( const __DRIdrawablePrivate *dPriv,
 
    intelFlush( &intel->ctx );
    
+   intelScreen = intel->intelScreen;
+
+   if (!rect && !intel->swap_scheduled && intelScreen->drmMinor >= 6 &&
+       !(intel->vblank_flags & VBLANK_FLAG_NO_IRQ) &&
+       intelScreen->current_rotation == 0) {
+      unsigned int interval = driGetVBlankInterval(dPriv, intel->vblank_flags);
+      unsigned int target;
+      drm_i915_vblank_swap_t swap;
+
+      swap.drawable = dPriv->hHWDrawable;
+      swap.seqtype = DRM_VBLANK_ABSOLUTE;
+      target = swap.sequence = intel->vbl_seq + interval;
+
+      if (intel->vblank_flags & VBLANK_FLAG_SYNC) {
+	 swap.seqtype |= DRM_VBLANK_NEXTONMISS;
+      } else if (interval == 0) {
+	 goto noschedule;
+      }
+
+      if (!drmCommandWriteRead(intel->driFd, DRM_I915_VBLANK_SWAP, &swap,
+                              sizeof(swap))) {
+        intel->swap_scheduled = 1;
+        intel->vbl_seq = swap.sequence;
+        swap.sequence -= target;
+        missed_target = swap.sequence > 0 && swap.sequence <= (1 << 23);
+      }
+   } else {
+      intel->swap_scheduled = 0;
+   }
+noschedule:
+
+   if (!intel->swap_scheduled) {
       intelWaitForFrameCompletion( intel );
       LOCK_HARDWARE( intel );
 
@@ -473,6 +506,7 @@ void intelCopyBuffer( const __DRIdrawablePrivate *dPriv,
       }
       intelFlushBatchLocked( intel, GL_TRUE, GL_TRUE, GL_TRUE );
       UNLOCK_HARDWARE( intel );
+   }
 
    if (!rect)
    {

src/mesa/drivers/dri/i915/intel_context.h

@@ -261,6 +261,8 @@ struct intel_context
 
    GLuint swap_count;
    GLuint swap_missed_count;
+
+   GLuint swap_scheduled;
 };
 
 
@@ -321,6 +323,13 @@ do {							\
     char __ret=0;					\
     DEBUG_CHECK_LOCK();					\
     assert(!(intel)->locked);				\
+    if ((intel)->swap_scheduled) {			\
+        drmVBlank vbl;					\
+        vbl.request.type = DRM_VBLANK_ABSOLUTE;		\
+        vbl.request.sequence = (intel)->vbl_seq;	\
+        drmWaitVBlank((intel)->driFd, &vbl);		\
+        (intel)->swap_scheduled = 0;			\
+    }							\
     DRM_CAS((intel)->driHwLock, (intel)->hHWContext,	\
         (DRM_LOCK_HELD|(intel)->hHWContext), __ret);	\
     if (__ret)						\