Files
seaweedfs/seaweedfs-rdma-sidecar/rdma-engine/src
Chris Lu 96722ebf4b rdma-engine: real libibverbs/rdma_cm RDMA READ datapath (proof of plumbing)
The engine backend was a mock: post_read fabricated byte=i%256 and
never touched a wire (the Go sidecar even discarded the engine reply
and re-faked the same pattern). Add a real datapath behind a new
'real-rdma' cargo feature:

- src/rdma_real.rs: RealRdmaContext (connect via rdma_cm, register_memory,
  a genuine IBV_WR_RDMA_READ via rdma_post_read, poll_completion,
  device_info) + RealResponder (rdma_reg_read REMOTE_READ source that
  hands the client {addr,rkey,len}).
- src/bin/rdma_loopback.rs: proof binary. A responder registers a buffer
  with a verifiable pattern (REALRDMA magic header, not the mock's i%256);
  the client RDMA-READs it and verifies the bytes match.
- REAL_RDMA.md: build/run + design notes.

Verified on a SoftRoCE (rxe) loopback (Linux 6.12.90, rdma_rxe):
1 MiB IBV_WR_RDMA_READ completes status=Success, dest head 'REALRDMA',
1048576 bytes match -> REAL-RDMA-READ-SUCCESS. (rxe on a 2-vCPU VM is a
correctness proof, not a benchmark.)

vendor/rdma-sys: rdma-sys 0.3.0 with bindgen bumped to 0.69.4 (the
crates.io 0.59.2 pin cannot parse rdma-core 56 headers); redirected via
[patch.crates-io]. Temporary workaround; see REAL_RDMA.md.

Phase-1 proof of plumbing only. A demonstrable zero-copy read of real
volume data still needs a needle-backed RDMA target in the volume
server, a control-plane handshake, an IPC-contract change, and mount
wiring (multi-week; deferred).
2026-06-14 17:54:42 -07:00
..
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00
2025-08-17 20:45:44 -07:00