Struct DirectSwapFile 

pub struct DirectSwapFile { /* private fields */ }

The generic paging core, lifted to atlas-tier (CUDA- and verbs-free). Re-exported under the atlas_tier names (no historical aliases). O_DIRECT fixed-stride swap file on NVMe (the peer’s cold tier). record_bytes MUST be a 4 KiB multiple (O_DIRECT) — the SSM snapshot blob (66,846,720 B = 16,320 × 4 KiB) already is. Records are addressed by disk_slot at disk_slot * record_bytes; the file grows sparsely as slots are allocated.

Buffers passed to read/write must be page-aligned for O_DIRECT; read_record / write_record stage through an internal aligned bounce (a full extra record memcpy) when they aren’t. Both in-tree callers are aligned: the peer passes its mmap’d arena, and crate::Residency’s scratch is a PageAlignedBuf — it used to be a plain Vec, which took the bounce every single time.

A 0-byte swap file is not evidence of a broken write. Records are only written when the residency’s hot arena has no free slot, so the first num_slots PUTs of distinct keys never touch this file; with the default 64-slot SSM arena the first pwrite happens on spill #65. The size then jumps to (highest disk_slot + 1) × record_bytes — and never shrinks, since discard_record is the documented no-op (freed records are reused by index, their blocks are never punched back out).

Implementations

impl DirectSwapFile

pub fn create(path: &Path, record_bytes: usize) -> Result<DirectSwapFile, Error>

Trait Implementations

impl SwapStore for DirectSwapFile

fn record_bytes(&self) -> usize

fn write_record(&mut self, disk_slot: usize, bytes: &[u8]) -> Result<(), Error>

fn read_record(&self, disk_slot: usize, out: &mut [u8]) -> Result<(), Error>

fn discard_record(&mut self, _disk_slot: usize)

Optional: reclaim disk space for a freed slot (default no-op; a hole in a preallocated file is fine — the free-list reuses the index).