perf(cache): use foyer io_uring I/O engine on Linux

[jaredLunde]

What

Both foyer-backed caches (clean block cache, pack-index cache) defaulted to foyer's psync engine, which dispatches every SSD-tier read to a tokio blocking thread (spawn_blocking → pread). This switches to foyer's io_uring engine on Linux, submitting the read inline instead.

• Selection is automatic: io_uring on Linux, falling back to psync if io_uring setup fails (e.g. seccomp/gVisor sandboxes) or on non-Linux. No config knob.
• Shared helper build_preferring_uring() (block/foyer_engine.rs) centralizes the prefer-uring / fallback-to-psync logic; both cache constructors route their builder through it.

Why

psync's blocking-pool hop adds latency on top of the actual pread for every clean-cache / pack-index SSD-tier read. io_uring removes that hop.

Measured

New bench benches/foyer_ssd_read.rs isolates the SSD-tier read (psync vs io_uring in one run):

Engine	Median latency	Throughput
psync (old default)	55.6 µs	2.20 GiB/s
io_uring (new)	40.3 µs	3.03 GiB/s

~27% lower latency, ~37% higher throughput on cached reads.

Verification

• cargo build -p glidefs clean; clippy clean on the lib.
• block::cache (7), block::pack_index_cache (16, incl. test_ssd_persistence), block::write_cache (61) tests all pass through the io_uring path on a Linux host.
• cargo bench --features test-utils --bench foyer_ssd_read produces the numbers above.

Notes

Also bundles an unrelated benches/gc.rs tweak and a ublk-core/examples/cleanup_stale.rs example.

🤖 Generated with Claude Code