Monday, September 07, 2009

Proximate And Ultimate Crashes

One last comment about driver crashes: I try to be fair about the distinction between causation and correlation when it comes to driver bugs; the piece of software that "fixes" a bug may not be the piece of software that contains a bug, and work-arounds happen in both directions between drivers and apps.

Similarly, seeing "ATI" or "NVOGL" in your crash report does not imply a driver bug. It does imply that the crash happened in the driver, but the driver crash may be a proximate cause, while the ultimate cause may be in X-Plane. (Sometimes the stack crawl is also just plain wrong, depending on the severity of the crash.)

OpenGL drivers need to be very, very fast; the whole point of buying that shiny new GPU was to get faster framerates. So driver writers have to make trade-offs about how much error checking they do when an application passes in junk data. If I were writing a driver and debating an error check that would slow performance and prevent a crash of an application that passed illegal data (but not the machine), I might very well skip the error check and say "if the application doesn't want to crash, it should pass better data."

Of course, it's not up to me, and it's not always up to the driver writers; the OpenGL specification actually dictates what outcome are acceptable in various error conditions. Quite often, "program termination" (read: the application crashes but the machine does not) is an allowed response to certain kinds of bad data from the application to the driver, particularly where checking the data would slow performance for all applications and the application should know better.

No comments: