Module pump 

File-read → megakernel ring-slot pump. Linux-only.

The two halves needed for mapped-read → GPU-visible-memory → compute already existed separately before this module: AsyncUringStream owns the io_uring submission + completion queue and the GPU-mapped DMA buffer, while crate::megakernel::Megakernel::publish_slot owns the host-side ring-slot writer that signals a persistent GPU kernel. Nothing composed them - a caller had to manually reach into both every dispatch. UringMegakernelPump wires them together so a host thread can run one compact loop:

pump.submit_file_scan(fd, offset, len, tenant, opcode, [a0,a1,a2])?;
pump.drain_into_ring(&mut ring_bytes)?;
// …later…
let epoch = pump.observe_epoch(&control_bytes);

Flow

1. submit_file_scan posts an IORING_OP_READ_FIXED that targets GpuMappedBuffer[chunk_idx * slot_len..]. The bytes land in host-visible GPU memory, so the kernel sees them the moment the ring-slot status flips to PUBLISHED.
2. The (tenant, opcode, args) payload is staged in pending: Vec<PendingPublish> keyed by chunk_idx.
3. drain_into_ring polls the io_uring CQ and, for each success, writes the staged slot into the caller-supplied ring buffer via Megakernel::publish_slot. Errors surface with a structured PipelineError that names the failing chunk.

Backpressure

The pump does not allocate new ring slots on its own - submit_file_scan takes a caller-assigned slot_idx. The host thread is responsible for slot bookkeeping (e.g., round-robin over slot_count published slots with the kernel draining them).

Linux-only

This module only compiles on target_os = "linux"; the io_uring surface itself is Linux-specific. Callers gate their pipeline code the same way.

Structs

UringMegakernelPump

Compose an AsyncUringStream with the megakernel ring-slot writer so the host can drive the compatibility mapped-read ingest loop with one compact pump. Native NVMe → BAR1 ingest is owned by super::driver::NvmeGpuIngestDriver::new_gpudirect.