Module gpu_sync 

Process-global lock serializing GPU/window surface lifecycle against active rendering.

With M3 several hardware surfaces live in one process, each rendering on its own thread. On Wayland the wl_surface and the Vulkan swapchain are driven from different threads — the main thread owns the winit window, the render thread acquires/presents. Creating or destroying one window’s surface on the main thread while another window’s render thread is inside vkAcquireNextImageKHR corrupts the shared driver connection and segfaults (reproduced on the NVIDIA driver with two hardware surfaces). Render threads run concurrently under a shared read guard; a surface’s creation or teardown takes the exclusive write guard, which waits for every in-flight frame to finish and blocks new ones — so lifecycle can never overlap a live acquire/present.

Functions

lifecycle_guard

Held while a window/renderer surface is created or destroyed; exclusive against every render thread. The caller must not already hold a render guard on this thread (would self-deadlock).

render_guard

Held for the duration of a render thread’s frame (acquire → present); concurrent with other render threads, but mutually exclusive with surface lifecycle. The () payload has no invariants, so a panic poisoning the lock is irrelevant — recover and continue.