Garbage textures and a crash in Mesa after GART ran out (2026-10-26)

The driver mechanism of the fix, Mesa 0049, is in docs/DRIVER.md. It follows the out-of-GART paths Mesa 0047 and 0048 closed for Sauerbraten (online-crashes).

Symptom

The first played session (2026-10-25, Scrap) ran at 7 fps on the old default graphics, and every texture was garbage after a restart. The first rules-2.0 session (2026-10-26, package ps3.833, Mesa ps3.827) ended in a crash on Ravenbridge Mansion, with a 21.7 s freeze in play before it, and objects popping in and out on that track (the player, side by side with the PC); the last shot is black. The segfault came while loading the track’s sky, after 414 “out of GART” lines in dmesg.

After the segfault the RSX stopped drawing: X kept running but nothing drew, a fresh X couldn’t draw either (the 3D state still the old X’s target, 0x900000; the TV showed the game’s last frame), the ssh reboot hung, and the user power-cycled the console. That driver question stays in docs/TESTS.md “Next”.

Cause

A segfault at 0x20: nv->fence was NULL in _nouveau_fence_next() after a 4 KiB fence wait object could not be had with GART exhausted; the failed kick (kick_notify false) also made pushbuf_submit drop that whole batch. In the same state the texture transfer map returned NULL (GL_OUT_OF_MEMORY, which games don’t check), so textures kept their old pixels: the garbage after a restart and the popping objects.

Fix

Mesa 0049 (commit 712dcf6): out of GART, nothing crashes or is lost silently. A context always has a current fence: the next one is made before the old one goes, its wait object comes after waiting for the GPU’s submitted work, else the fence goes without one and is polled. A texture upload with no staging buffer waits for the GPU, drains the staging pools and tries again, else goes through malloc memory and a CPU copy into the CPU-mapped VRAM. Buffer maps wait and retry for their staging area; a fragment program or the blit’s program with no buffer keeps the bound one. Host tests alloc-fail-test.c and transfer-cpu-test.c, in mesa-query-steal.

The next session of the same day, on Mesa ps3.850 (the GART fix), had every track right: the desert tutorial, Ravenbridge Mansion and its restart, the mines, the castle, snow, subsea; kart shadows and textures as on the PC, no crash. It also had the package’s lean defaults (supertuxkart 0006) and pad scheme (0007), pushed at the same time.