<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0//EN" "http://www.w3.org/TR/REC-html40/strict.dtd"> <html><head><meta name="qrichtext" content="1" /><style type="text/css"> p, li { white-space: pre-wrap; } </style></head><body style=" font-family:'Monospace'; font-size:9pt; font-weight:400; font-style:normal;"> Hi, On Friday, August 07, 2015 10:28:10 Ilia Mirkin wrote: > > One alternative I can see is to call _tnl_InvalidateState without the > > _NEW_VIEWPORT and > > > > _NEW_BUFFERS flag set instead of _tnl_wakeup in the two nouveau early > > initialization > > > > code paths in nouveau_swtnl_t.c and nv04_render.c. That should also bring > > us back to > > > > the original behavior. > > > > That seems like a bit of a hack. If you think it's OK, I'm just going to > introduce the DrawBuffer check. Brian, does that seem OK to you? I'm mostly > unfamiliar with the TnL module, its APIs, and when/how they're supposed to > be used. But nouveau_vieux calls tnl_wakeup right in its context creation > pass. nv04/nv05 don't have hwtnl at all, nv10+ have it, but it has various > limitations, with which we end up using swtnl. To be honest, I did not expect the _tnl_InvalidateState function to be called much differently than through _mesa_update_state where it is usually called through a driver provided UpdateState callback. If you look where _tnl_InvalidateState is used you can see this pattern broadly used. The only exception is that it is also called through _tnl_wakeup which sounds to me like something that is not meant for initialization purpose. Actually, only nouveau uses this call for initialization stuff. So, to me the question looks like: what is the right way to initialize nouveau. Either way, just skipping this part of _tnl_InvalidateState either by a pointer check or something else that makes that part of _tnl_InvalidateState happen only for the succeeding calls, should be safe in terms of crashes as well as correctness. As I read the calling code we should end up calling _tnl_InvalidateState a second time past we have a bound a DrawBuffer and before the first draw through nouveau_update_state. So the matrix should be properly set when it is needed for the first draw. The only aspect I can add to that is: The additional pointer check happens on each Driver::UpdateState call. May be it's not worth to care for that, but if we customize nouveaus startup procedure we do not introduce any additional if code in state validation. But, I also won't seriously object against your patch! Thanks for taking care of the problem! Greetings Mathias</body></html>