SIGN IN SIGN UP

UCP/FT: rebuild failed lanes on recovery via LANES_ADDR exchange (#11562)

UCP/FT: rebuild failed lanes on recovery via LANES_ADDR exchange

- Rebuild failed p2p/iface lanes: install wireup proxies over failed stubs,
  recreate the inner UCT endpoints, and drive them to READY so the lane
  resumes serving traffic.
- Wire the LANES_ADDR_REQUEST/REPLY handlers to exchange peer addresses and
  rebuild the requested lanes, including local failover for asymmetric
  failures, and free the recovery state once all lanes recover.
- Pin a recovered iface lane to its original peer device.
- Add end-to-end fault-tolerance gtests (initiator/target failure incl. the
  all-lanes-failed case) using the keepalive-driven recovery cadence.
- UCX_RESOLVE_REMOTE_EP_ID = auto by default.
E
Evgeny Leksikov committed
9db2e48836c1a8c4a15c3e4abeab5f0754f0032d
Parent: ce07f5c
Committed by GitHub <noreply@github.com> on 7/6/2026, 10:49:02 AM