Disable threaded rendering if Mesa is used

Mesa and xcb show some bad interaction which leads to frequent crashed
on multithreaded access. Also, the selective approach to blacklisting
only specific chipsets isn't feasible, given the resources available.

The client glx vendor string is used to identify mesa instead of the
server GL vendor and/or renderer string as that is much more reliable.

Task-number: QTBUG-38221
Change-Id: I2d8c037aa4fd9c38eb9537452a5e7e62f72a081d
Reviewed-by: Gunnar Sletta <gunnar.sletta@jollamobile.com>

1 files changed, 24 insertions, 4 deletions

diff --git a/src/plugins/platforms/xcb/qglxintegration.cpp b/src/plugins/platforms/xcb/qglxintegration.cpp
index c183deb3b8..f78d3bcc4e 100644
--- a/src/plugins/platforms/xcb/qglxintegration.cpp
+++ b/src/plugins/platforms/xcb/qglxintegration.cpp
@@ -449,12 +449,32 @@ bool QGLXContext::m_supportsThreading = true;
 // binary search.
 static const char *qglx_threadedgl_blacklist_renderer[] = {
     "Chromium",                             // QTBUG-32225 (initialization fails)
-    "Mesa DRI Intel(R) Sandybridge Mobile", // QTBUG-34492 (flickering in fullscreen)
     0
 };
 
+// This disables threaded rendering on anything using mesa, e.g.
+// - nvidia/nouveau
+// - amd/gallium
+// - intel
+// - some software opengl implementations
+//
+// The client glx vendor string is used to identify those setups as that seems to show the least
+// variance between the bad configurations. It's always "Mesa Project and SGI". There are some
+// configurations which don't use mesa and which can do threaded rendering (amd and nvidia chips
+// with their own proprietary drivers).
+//
+// This, of course, is very broad and disables threaded rendering on a lot of devices which would
+// be able to use it. However, the bugs listed below don't follow any easily recognizable pattern
+// and we should rather be safe.
+//
+// http://cgit.freedesktop.org/xcb/libxcb/commit/?id=be0fe56c3bcad5124dcc6c47a2fad01acd16f71a will
+// fix some of the issues. Basically, the proprietary drivers seem to have a way of working around
+// a fundamental flaw with multithreaded access to xcb, but mesa doesn't. The blacklist should be
+// reevaluated once that patch is released in some version of xcb.
 static const char *qglx_threadedgl_blacklist_vendor[] = {
-    "nouveau",                             // QTCREATORBUG-10875 (crash in creator)
+    "Mesa Project and SGI",                // QTCREATORBUG-10875 (crash in creator)
+                                           // QTBUG-34492 (flickering in fullscreen)
+                                           // QTBUG-38221
     0
 };
 
@@ -502,9 +522,9 @@ void QGLXContext::queryDummyContext()
         }
     }
 
-    if (const char *vendor = (const char *) glGetString(GL_VENDOR)) {
+    if (glxvendor) {
         for (int i = 0; qglx_threadedgl_blacklist_vendor[i]; ++i) {
-            if (strstr(vendor, qglx_threadedgl_blacklist_vendor[i]) != 0) {
+            if (strstr(glxvendor, qglx_threadedgl_blacklist_vendor[i]) != 0) {
                 m_supportsThreading = false;
                 break;
             }