The miComputeClips VTMove case doesn't properly calculate the borderClip of a composite-redirected window. It leaves the old borderClip in place. Specifically, the visibility of composite-redirected windows computed in this routine is VisibilityFullyObscured. Therefore the region translations of the borderClip and clipList are skipped by the following code: case VTMove: if ((oldVis == newVis) && ((oldVis == VisibilityFullyObscured) || (oldVis == VisibilityUnobscured))) { pChild = pParent; while (1) { if (pChild->viewable) { if (pChild->visibility != VisibilityFullyObscured) { REGION_TRANSLATE( pScreen, &pChild->borderClip, dx, dy); REGION_TRANSLATE( pScreen, &pChild->clipList, dx, dy); My suggested fix is to force the region translation to happen for composite-redirected windows, ignoring their visibility, viz. case VTMove: if ((oldVis == newVis) && ((oldVis == VisibilityFullyObscured) || (oldVis == VisibilityUnobscured))) { pChild = pParent; while (1) { if (pChild->viewable) { if (pChild->visibility != VisibilityFullyObscured #ifdef COMPOSITE || pChild->redirectDraw #endif ) { REGION_TRANSLATE( pScreen, &pChild->borderClip, dx, dy); REGION_TRANSLATE( pScreen, &pChild->clipList, dx, dy); Here is mail from Keith in which he approves of my suggested fix: On Sun, 2006-02-05 at 16:04 -0800, Deron Johnson wrote: >> The problem was that there is a bug in the way miComputeClips treats >> composite-redirected windows during MoveWindow. The visibility of >> composite-redirected windows is FullyObscured in this routine, so the >> following border clip translation code is skipped. So the borderClip >> of a mapped composite-redirected window is stale after it is moved. Good catch. >> This messes up the sending of damage events. And more than that; having regions with the wrong offset should also scramble drawing. Oops. >> My suggested fix is to force the region translation to happen for >> composite-redirected windows, viz. Yeah, I think that's safest. You realize, of course, that the visibility check is just an optimization; FullyObscured windows are supposed to have empty clip lists. We could get away with always translating the regions, but I think having the check in place will cause less confusion in the future.
Keith P. recommends the following fix: move the composite-specific code in miComputeClips to be BEFORE the visibility computation. I have tested this and it works. Here is the code: diff -c mivaltree.c mivaltree.c.fixed *** mivaltree.c 2006-02-23 16:55:19.060842000 -0800 --- mivaltree.c.fixed 2006-02-23 16:56:58.633837000 -0800 *************** *** 266,271 **** --- 266,283 ---- dy = 32767; borderSize.y2 = dy; + #ifdef COMPOSITE + /* + * In redirected drawing case, reset universe to borderSize + */ + if (pParent->redirectDraw) + { + if (miSetRedirectBorderClipProc) + (*miSetRedirectBorderClipProc) (pParent, universe); + REGION_COPY(pScreen, universe, &pParent->borderSize); + } + #endif + oldVis = pParent->visibility; switch (RECT_IN_REGION( pScreen, universe, &borderSize)) { *************** *** 305,322 **** ((pParent->eventMask | wOtherEventMasks(pParent)) & VisibilityChangeMask)) SendVisibilityNotify(pParent); - #ifdef COMPOSITE - /* - * In redirected drawing case, reset universe to borderSize - */ - if (pParent->redirectDraw) - { - if (miSetRedirectBorderClipProc) - (*miSetRedirectBorderClipProc) (pParent, universe); - REGION_COPY(pScreen, universe, &pParent->borderSize); - } - #endif - dx = pParent->drawable.x - pParent->valdata->before.oldAbsCorner.x; dy = pParent->drawable.y - pParent->valdata->before.oldAbsCorner.y;
This patch is in CVS since Wed Mar 22 13:42:44 2006 Søren Sandmann <sandmann@redhat.com> * mi/mivaltree.c (miComputeClips): Patch by Keith Packard to make sure redirected windows don't get considered "FullyObscured".
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.