mirror of
https://github.com/versity/scoutfs.git
synced 2026-04-22 22:40:31 +00:00
a3dabb4c79
Previously, client_greeting spliced pending requests back onto send_queue when reconnecting to a new server. Those requests carried state from the old server (sequence numbers, log tree references, lock modes) that was reclaimed at fence time, so resending against the new server was incorrect. Drain pending requests with -ECONNRESET at greeting time, mirroring the forcing_unmount drain in the shutdown worker. Thread the lock pointer through scoutfs_client_lock_request so the response callback can clear request_pending and wake waiters on error; otherwise a lock_key_range waiter would block forever because the new server's lock recovery only reports granted modes, not pending requests. Wrap the sync request senders in client_sync_request so userspace paths (statfs, mkdir, sysfs volopt, resize ioctl, walk-inodes ioctl) retry transparently across failover instead of surfacing a new -ECONNRESET that callers never saw before. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>