Struct UringMegakernelPump 

pub struct UringMegakernelPump<'a> { /* private fields */ }

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.

Implementations

impl<'a> UringMegakernelPump<'a>

pub fn new(stream: AsyncUringStream<'a>, chunk_bytes: u32) -> Self

Construct a pump bound to an existing stream. chunk_bytes is the fixed read size - every call to submit_file_scan must request exactly this many bytes.

The pump takes ownership of stream; reclaim it via into_stream on shutdown.

pub fn into_stream(self) -> AsyncUringStream<'a>

Release the underlying stream for explicit shutdown sequences.

pub fn inflight(&self) -> u32

Inflight submissions (submit - drain diff).

pub unsafe fn submit_file_scan( &mut self, fd: i32, file_offset: u64, len: u32, chunk_idx: u32, slot_idx: u32, tenant_id: u32, opcode: u32, args: [u32; 3], ) -> Result<(), PipelineError>

Submit one file-scan read. Destination inside the GPU buffer is chunk_idx * self.chunk_bytes.

On CQE completion, drain_into_ring publishes a megakernel ring slot at slot_idx with tenant_id, opcode, and args. The three args fit in the fixed 3-word prefix of a megakernel slot; callers with more payload use the packed-slot opcode (PACKED_SLOT) out-of- band.

Errors

• PipelineError::QueueFull if the io_uring SQ or the GPU-side destination buffer is out of room.
• Arbitrary PipelineError variants from the underlying syscall wrappers.

Safety

fd must be an open file descriptor the pump’s io_uring ring can read from. The caller retains ownership - the pump does not close it. len must equal self.chunk_bytes; mismatches are rejected with PipelineError::QueueFull.

pub fn drain_into_ring( &mut self, ring_bytes: &mut [u8], ) -> Result<u32, PipelineError>

Drain completions + publish corresponding ring slots into ring_bytes.

Returns the number of completions processed (including those that surfaced errors - those still advance the inflight counter). The first error is returned via Err(PipelineError::IoUringSyscall); subsequent completions keep draining so the ring does not overflow.

Errors

• PipelineError::IoUringSyscall on the first failed CQE.
• PipelineError::QueueFull if Megakernel::publish_slot rejects the published slot (e.g., slot_idx still in-flight on the GPU side - caller must wait for the kernel to drain).

pub fn observe_epoch(&self, control_bytes: &[u8]) -> u32

Host-visible epoch field from the megakernel control buffer. The kernel atomic-adds this on every BATCH_FENCE; callers observe forward progress by polling the field between dispatches.