The irregular Nouveau Development Companion
Issue #5 for October, 10th
Well, it had to happen. After a hiatus of 4 weeks, the TiNDC is back. Hopefully, there won't be any pauses in the future as long as this one. This companion not only covers the IRC logs but for the first time the DRI mailing list.
Criticism, ideas and questions can be send in my (KoalaBR) direction on the #nouveau IRC channel, I will try to address them in the next TiNDC. Don't hesitate to post, even if I'm not online, I read the backlogs.
Shortly after the release of the last TiNDC we found arlied and marcheu discussing the correct way to design the driver. The nvidia cards maintains hardware contexts for each program which uses the nvidia card. The first context serves X11, any other OpenGL get another context. The card now switches between all those contexts via interrrupt.
As marcheu had some problems getting the card to emit interrupts, arlied and marcheu were talking, whether using only one context would be feasable. Result: doable - yes, but far from efficient, as memory management would get much more complicated in comparison to many different contexts. Besides that we would need to lock the surfaces / memory chunks to certain parts in the logical address space, as user space would react "annoyed" if it's data would move beneath it, without getting any notification. This feature of course would need Kernel based support (kernel module), something which nouveau currently does not need (which makes debugging easy!)
[[!format txt """
The result was again two options:
- _For the memory manager, I guess there are two choices: 1) make the driver use a command-buffer approach even though the hardware supports per-context ring buffers, or 2) extend the memory manager. _
- Keith Whitwell Some other ideas were brought up to (playing MMU tricks), but nothing groundbreaking new came out of it, "only" an evaluation of each possibility, it's advantages and disadvantages.
At that point, nouveau and its developers faced two major problems:
- memory management (which needs understanding the object stuff fully)
- context switching (I think it fails because fifo #2 is actually not correctly setup, some of the magic values should probably be understood reused for that fifo) After some more development and testing, it was found, that mode switching (from X11 to terminal or back) was broken. This again lead to a discussion between darktama, arlied and marcheu whether mode switching should be moved into the kernel. It would have the advantage of sharing code with the nvidia framebuffer but debugging would be more difficult. But as darktama found the problem, this decision was of no importance and arlied submitted (unrelated to nouveau) a patch to the Linux kernel, which would make management of graphics drivers (DRI,X11 and framebuffer) much easier. So it was decided to ignore that small problem for now.
A few days after that, marcheu finally got interrupts working. This means, that nouveau now needs to react on the interrupt correctly otherwise the card will stall (or even crash), which is what promptly happened.
This is the current state of nouveau: Marcheu is searching for the reason of the card's stall and tries to compare it to the binary driver. As we are speaking of interrrupts, this is no easy task and progress is slow (as is the traffic on #nouveau).
Darktama has started implementing the state functions for >=NV30 cards, and has been rewriting the shader code so that most of it is shared between the various implementations (vertex/fragment,NV20/30/40). He also took a brief look at translating Mesa's GLSL bytecode into hw instructions.
We still are looking for C developers who aren't afraid to get their hands dirty. Writing of the driver could benefit from developers, who already know their way around graphics drivers or aren't afraid to learn.