]> git.baikalelectronics.ru Git - kernel.git/commit
SUNRPC: Move xpt_mutex into socket xpo_sendto methods
authorChuck Lever <chuck.lever@oracle.com>
Sat, 2 May 2020 14:37:44 +0000 (10:37 -0400)
committerChuck Lever <chuck.lever@oracle.com>
Mon, 18 May 2020 14:21:21 +0000 (10:21 -0400)
commit1b6f678ccf8299219c06608d1320a4a4d5c8b549
tree557c33f5f34050f49e27961857ba4e52ad4945da
parenta20819a848cd39220dc3e8fc9bca50353b49e001
SUNRPC: Move xpt_mutex into socket xpo_sendto methods

It appears that the RPC/RDMA transport does not need serialization
of calls to its xpo_sendto method. Move the mutex into the socket
methods that still need that serialization.

Tail latencies are unambiguously better with this patch applied.
fio randrw 8KB 70/30 on NFSv3, smaller numbers are better:

    clat percentiles (usec):

With xpt_mutex:
r    | 99.99th=[ 8848]
w    | 99.99th=[ 9634]

Without xpt_mutex:
r    | 99.99th=[ 8586]
w    | 99.99th=[ 8979]

Serializing the construction of RPC/RDMA transport headers is not
really necessary at this point, because the Linux NFS server
implementation never changes its credit grant on a connection. If
that should change, then svc_rdma_sendto will need to serialize
access to the transport's credit grant fields.

Reported-by: kbuild test robot <lkp@intel.com>
[ cel: fix uninitialized variable warning ]
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
include/linux/sunrpc/svc_xprt.h
net/sunrpc/svc_xprt.c
net/sunrpc/svcsock.c
net/sunrpc/xprtrdma/svc_rdma_backchannel.c
net/sunrpc/xprtrdma/svc_rdma_sendto.c
net/sunrpc/xprtsock.c