[RFC PATCH 0/7] sunrpc: Reduce lock contention for NFSD TCP sockets

[Chuck Lever <cel@kernel.org>]

From: Chuck Lever <chuck.lever@oracle.com>

High-throughput NFSD workloads exhibit significant lock contention on
TCP connections. Worker threads compete for the socket lock during
receives and serialize on xpt_mutex during sends, limiting scalability.

This series addresses both paths:

 - Receive: A dedicated kernel thread per TCP connection owns all
   sock_recvmsg() calls and queues complete RPC messages for workers
   via lock-free llist. This eliminates socket lock contention among
   workers.

 - Transmit: Flat combining allows one thread to send on behalf of
   multiple waiters. Threads enqueue requests; the mutex holder
   ("combiner") processes the batch, amortizing lock acquisition and
   enabling TCP segment coalescing via MSG_MORE.

Supporting changes include a workqueue affinity fix for single-LLC
systems, a page recycling pool for receive buffers, and explicit TCP
buffer sizing for high bandwidth-delay product networks.

Base commit: v6.19-rc8

---

Chuck Lever (7):
  workqueue: Automatic affinity scope fallback for single-pod topologies
  sunrpc: split svc_data_ready into protocol-specific callbacks
  sunrpc: add per-transport page recycling pool
  sunrpc: add dedicated TCP receiver thread
  sunrpc: implement flat combining for TCP socket sends
  sunrpc: unify fore and backchannel server TCP send paths
  SUNRPC: Set explicit TCP socket buffer sizes for NFSD

 include/linux/sunrpc/svc.h      |   1 +
 include/linux/sunrpc/svc_xprt.h |  32 ++
 include/linux/sunrpc/svcsock.h  |  40 ++
 include/linux/workqueue.h       |   8 +-
 kernel/workqueue.c              |  68 ++-
 net/sunrpc/svc.c                |  13 +
 net/sunrpc/svc_xprt.c           | 151 ++++++
 net/sunrpc/svcsock.c            | 797 +++++++++++++++++++++++++++++---
 net/sunrpc/xprtsock.c           |  60 +--
 9 files changed, 1044 insertions(+), 126 deletions(-)

-- 
2.52.0