I came in with this giant task list I'd been jotting down in Workflowy (Thanks for the emphatic recommendation of that, Qiaochu!). Each of the tasks I had were things where I'd have been perfectly unsurprised if they'd taken a week or two. Instead, I've knocked out about 5 of them, and by Friday I had phire's "hackdriver" triangle code running on a kernel with a relocations-based GEM interface. Oh, sure, the code's full of XXX comments, insecure, and synchronous, but again, a single triangle rendering in a month would have been OK with me.
I've been incredibly lucky, really -- I think I had reasonable expectations given my knowledge going in. One of the ways I'm lucky is that my new group is extremely helpful. Some of it is things like "oh, just go talk to Dom about how to set up your serial console" (turns out minicom fails hard, use gtkterm instead. Also, someone else will hand you a cable instead of having to order one, and Derek will solder you a connector. Also, we hid your precious dmesg from the console after boot, sorry), but it extends to "Let's go have a chat with Tim about how to get modesetting up and running fast." (We came up with a plan that involves understanding what the firmware does with the code I had written already, and basically whacking a register beyond that. More importantly, they handed me a git tree full of sample code for doing real modesetting, whenever I'm ready.).
But I'm also lucky that there's been this community of outsiders reverse engineering the hardware. It meant that I had this sample "hackdriver" code for drawing a triangle with the hardware entirely from userspace, that I could incrementally modify to sit on top of more and more kernel code. Each step of the way I got to just debug that one step to go from "does not render a triangle" back to "renders that one triangle." (Note: When a bug in your command validator results in pointing the framebuffer at physical address 0 and storing the clear color to it, the computer will go away and stop talking to you. Related note: When a bug in your command validator results in reading your triangle from physical address 0, you don't get a triangle. It's like a I need a command validator for my command validator.).
https://github.com/anholt/linux/tree/vc4 is the code I've published so far. Starting Thursday night I've been hacking together the gallium driver. I haven't put it up yet because 1) it doesn't even initialize, but more importantly 2) I've been using freedreno as my main reference, and I need to update copyrights instead of just having my boilerplate at the top of everything. But next week I hope to be incrementally deleting parts of hackdriver's triangle code and replacing it with actual driver code.