fix(glx): match GetDrawableAttributes and MakeCurrent to Xorg

[ariel3259]

fix(glx): match GetDrawableAttributes and MakeCurrent to Xorg

Branch: glx-reply-xorg-alignment, opened against master. Merged with current master (2252 tests, clippy --all-targets -- -D warnings and +nightly fmt --check clean).

The hardware numbers below were measured on this branch before the merge with current master. master's new commits touch the Present/TFP wait paths rather than the GLX reply arms, and the tree is green, but the probe has not been re-run against the exact merged tree.

Problem

NVIDIA's libGLX_nvidia could not bind a GL context against yserver. Clients died with:

X Error of failed request:  BadAlloc (insufficient resources for operation)
Major opcode of failed request:  148 (GLX)
Minor opcode of failed request:  5 (X_GLXMakeCurrent)
Serial number of failed request:  0

Serial 0 was the clue: yserver never emits that error. Its MAKE_CURRENT handler always replies with a contextTag and cannot fail, and libglvnd 1.7.0 does not synthesise BadAlloc either (it produces BadAccess/BadMatch/GLXBadContext/GLXBadDrawable). The error was fabricated client-side by libGLX_nvidia after reading a reply it considered impossible.

This hit KWin (QGLXContext: Failed to create dummy context), Steam, and games.

Root cause

GetDrawableAttributes reported GLX_FBCONFIG_ID = 0 — a nonexistent id — for a bare X window. drawable_attributes_for looked the XID up in glx_drawables and fell back to 0 when absent, but a GLX 1.2 "naked window" is not in that table by definition.

Xorg has an explicit path for this case (glx/glxcmds.c:1873-1914) and omits the attribute entirely rather than sending a bogus one. It also always sends GLX_DRAWABLE_TYPE, which yserver never sent, and never sends GLX_RENDER_TYPE, which yserver did.

Fix

Six commits, aligning the two replies with Xorg's observable behaviour:

• D1 — geometry resolves from the backing X drawable. The GLXWindow/GLXPixmap XID is a fresh client-allocated id with no X resource behind it, so the old lookup always missed and reported 0×0. This was a live defect against Mesa's loader_dri3 too, which sizes its buffers from this reply.
• D2-D6 — attribute set and error arms matched to Xorg, including the naked-window omission above.
• MakeCurrent returns contextTag = 0 when releasing (context == 0), matching glx/vndcmds.c:232-234,272. yserver previously returned a fresh incrementing tag.

Verification

Measured on an RTX 5060 Ti, NVIDIA 610.57.04, open kernel modules, over a real KMS session.

The causal proof is a single difference in the client trace against the pre-fix baseline: GetDrawableAttributes went from 7 attribs to 5 attribs. Everything before that point is byte-identical. Immediately after, the baseline did MakeCurrent and died; with the fix, NVIDIA continues into DRI3::SetDRMDeviceInUse → GetSupportedModifiers → PixmapFromBuffers → MakeCurrent.

Probe result after the fix:

GL_VERSION:   4.6.0 NVIDIA 610.57.04
GL_RENDERER:  NVIDIA GeForce RTX 5060 Ti/PCIe/SSE2
direct=1
RESULT: GL works

The BadAlloc is gone. Steam — a 32-bit client that previously crashed on launch with the exact signature above — now starts, as does Hollow Knight.

Notes

• docs/status.md updated (entry dated 2026-08-04).
• Design doc and implementation plan included under docs/superpowers/.
• Two divergences from Xorg were found and deliberately not fixed here, recorded in the design's "Deferred" section so they are not rediscovered: QUERY_CONTEXT answers 0 attributes where Xorg answers FBCONFIG_ID/RENDER_TYPE/SCREEN, and fbconfig synthesis produces 2 configs against XWayland's 168.
• Workspace green: 2241 tests, cargo clippy --all-targets -- -D warnings clean, cargo +nightly fmt --check clean. All commits signed.

🤖 Generated with Claude Code

[joske]

Thx.