[Intel-gfx] Patches for i830 flicker on panning

[Thomas Richter]

Hi Daniel, dear intel experts,

the following is a "solution" for the flicker on i830 panning. To remind 
you, when panning is enabled on an i830,
the display shows an annoying 30Hz flicker on some "forbidden" 
positions. On such positions, half of the frames
is black, and the other half shows a screen that is "ripped in half", 
with one half scrolled to the wrong
position, and the other half displayed correctly, with a discontinuity 
in the middle.
That is, every other frame is black, and the ones that are not black are 
"wrong".

The patch is, arguably, ugly as hell, and creates a certain "yerkiness", 
though it does its job. One of the side
effects this flicker causes is complete crashes and lockups with video 
overlays which are avoided by the patch.
That is, if you play a video with xine, and then pan over a forbidden 
position, the machine crashes and comes
to a complete stop.

Sorry if the patch format isn't quite right, and the patch is not quite 
in a final state (not yet fully tested on all possible
combinations of panning dimensions and display depths), but it's a start 
and "works good enough for me".

For that, patch: i9xx_update_plane(...) in intel_display.c,

     if (INTEL_INFO(dev)->gen >= 4) {
         I915_MODIFY_DISPBASE(DSPSURF(plane),
                      i915_gem_obj_ggtt_offset(obj) + 
intel_crtc->dspaddr_offset);
         I915_WRITE(DSPTILEOFF(plane), (y << 16) | x);
         I915_WRITE(DSPLINOFF(plane), linear_offset);
+    } else if (INTEL_INFO(dev)->gen == 2) {
+      int xflick = (x >> 4) & 0x03;
+      if (xflick == 3) {       /* in forbidden column, probably a i830 
bug */
+        int xfix  = 16 + (x & ~(0x3f));
+        int fix_offset = y * fb->pitches[0] + xfix * 
(fb->bits_per_pixel / 8);
+        I915_WRITE(DSPADDR(plane), i915_gem_obj_ggtt_offset(obj) + 
fix_offset);
+        POSTING_READ(reg);
+        intel_wait_for_vblank(dev, intel_crtc->pipe);
+      }
+      I915_WRITE(DSPADDR(plane), i915_gem_obj_ggtt_offset(obj) + 
linear_offset);
     } else
         I915_WRITE(DSPADDR(plane), i915_gem_obj_ggtt_offset(obj) + 
linear_offset);
     POSTING_READ(reg);

The trick is, as you see, first to display the one frame the display 
scrolled off *before* the forbidden column, and
then scroll to the right position, going to the right. )-:

*Why* that works is unclear to me. Don't ask, just observation.

An improvement would be to check whether the x position did change, and 
if it did not, just avoid all the hack to
allow smooth vertical scrolling. Still to be done. Maybe later the week 
if the above also works on the R31
(affected by the same problem).

Greetings,
     Thomas