Bug 73856

Right, that confirms that is the gc->pCompositeClip that is screwed up.

More debugging, hopefully a step in the right direction:

commit b9ebb016833df5b70336c66993a871e850fd0f61
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Jan 21 09:05:54 2014 +0000

    sna: More assertions for tracking gc->pCompositeClip
    
    Tracking down who leaves it modified...
    
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

Sadly that was a bogus assert. However, there was a bit of extra DBG that should should still be useful. (Baby steps.)

Created attachment 92516 [details]
Xorg.0.log-2.99.907-44-g720d131

Just to be sure, it's a bug that resides inside a driver and not an effect of some misconfiguration, i.e. in kernel/xorg, on my side? Sometimes I learn to have too much power over my config the hard way.

(In reply to comment #3) 
> Just to be sure, it's a bug that resides inside a driver and not an effect
> of some misconfiguration, i.e. in kernel/xorg, on my side? Sometimes I learn
> to have too much power over my config the hard way.

It is definitely a bug in the driver/xserver. For some reason it believes that the gc has already been validated for the drawable, but it still have the clip from the previous drawable.

Ok, I think this is the culprit:

diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 4a2b614..b1744ae 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -4021,7 +4021,8 @@ static bool must_check sna_gc_move_to_cpu(GCPtr gc,
        } else
                changes &= ~GCClipMask;
 
-       if (changes || drawable->serialNumber != sgc->serial) {
+       if (changes || drawable->serialNumber != (sgc->serial & DRAWABLE_SERIAL_BITS)) {
+               long tmp = gc->serialNumber;
                gc->serialNumber = sgc->serial;
 
                if (fb_gc(gc)->bpp != drawable->bitsPerPixel) {
@@ -4042,8 +4043,7 @@ static bool must_check sna_gc_move_to_cpu(GCPtr gc,
                }
 
                fbValidateGC(gc, changes, drawable);
-
-               sgc->serial = drawable->serialNumber;
+               gc->serialNumber = tmp;
        }
        sgc->changes = 0;
 
@@ -15733,6 +15733,7 @@ sna_validate_gc(GCPtr gc, unsigned long changes, DrawablePtr drawable)
        assert(RegionNil(gc->pCompositeClip) || gc->pCompositeClip->extents.y2 - drawable->y <= drawable->height);
 
        sna_gc(gc)->changes |= changes;
+       sna_gc(gc)->serial = gc->serialNumber;
 }
 
 static const GCFuncs sna_gc_funcs = {

Created attachment 92550 [details] [review]
Restore gc->serialNumber correctly after fallback

Applied patch over current git. Looks like the bug is pinned - it made Bruce Springsteen sing. ;)

Thank you very much for your bug report and help,

commit 29e11e2e6f80f6485ed69ea72e09e8d9b31fd8e8
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Jan 21 22:40:11 2014 +0000

    sna: Restore gc->serialNumber correctly after falling back
    
    Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=73856
    Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>