Module pinned_hosts 

2026-09-25: Which host buffers are page-locked, so the CUDA backend’s copy_h2d_async can keep its promise that the caller may drop the source on return.

From pageable memory the driver stages the bytes before the call returns; from page-locked memory the DMA engine reads the caller’s pages after it returns. MetraleCudaBackend::alloc_host_pinned registers each buffer it allocates and free_host_pinned unregisters it before the free. copy_h2d_async asks is_pinned and, for a pinned source, synchronises the stream before returning and logs a warning once per process.

Owner: gpu-runtime. Invariants:

• Only memory from alloc_host_pinned is tracked. Page-locked memory from another route, such as the cudaHostAlloc workspace in flashinfer.rs, is not seen.
• Entries stay sorted by base, with at most one entry per base.

Functions

is_pinned

2026-09-25: Whether all of src lies inside one live registered region. An empty slice is never pinned.

live_count

2026-09-25: Number of live registered regions.

register

2026-09-25: Record a page-locked region. A repeated base replaces the stored length; a null pointer or a zero length is ignored.

unregister

2026-09-25: Forget a region. free_host_pinned calls it before the free, so a reused address is not reported as still pinned.