svcrdma: Don't account for Receive queue "starvation"
authorChuck Lever <chuck.lever@oracle.com>
Fri, 23 Jun 2017 21:18:08 +0000 (17:18 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Wed, 28 Jun 2017 18:21:44 +0000 (14:21 -0400)
>From what I can tell, calling ->recvfrom when there is no work to do
is a normal part of operation. This is the only way svc_recv can
tell when there is no more data ready to receive on the transport.

Neither the TCP nor the UDP transport implementations have a
"starve" metric.

The cost of receive starvation accounting is bumping an atomic, which
results in extra (IMO unnecessary) bus traffic between CPU sockets,
while holding a spin lock.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c

index b48089314f859fbfce98b0c9fcca1ed3fffac22d..1452bd02d857f5f6dcb736747a7da66abf7367ef 100644 (file)
@@ -844,9 +844,9 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
        struct svc_xprt *xprt = rqstp->rq_xprt;
        struct svcxprt_rdma *rdma_xprt =
                container_of(xprt, struct svcxprt_rdma, sc_xprt);
-       struct svc_rdma_op_ctxt *ctxt = NULL;
+       struct svc_rdma_op_ctxt *ctxt;
        struct rpcrdma_msg *rmsgp;
-       int ret = 0;
+       int ret;
 
        dprintk("svcrdma: rqstp=%p\n", rqstp);
 
@@ -863,21 +863,13 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
                                        struct svc_rdma_op_ctxt, list);
                list_del(&ctxt->list);
        } else {
-               atomic_inc(&rdma_stat_rq_starve);
+               /* No new incoming requests, terminate the loop */
                clear_bit(XPT_DATA, &xprt->xpt_flags);
-               ctxt = NULL;
+               spin_unlock(&rdma_xprt->sc_rq_dto_lock);
+               return 0;
        }
        spin_unlock(&rdma_xprt->sc_rq_dto_lock);
-       if (!ctxt) {
-               /* This is the EAGAIN path. The svc_recv routine will
-                * return -EAGAIN, the nfsd thread will go to call into
-                * svc_recv again and we shouldn't be on the active
-                * transport list
-                */
-               if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
-                       goto defer;
-               goto out;
-       }
+
        dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p\n",
                ctxt, rdma_xprt, rqstp);
        atomic_inc(&rdma_stat_recv);
@@ -920,7 +912,6 @@ complete:
                + rqstp->rq_arg.page_len
                + rqstp->rq_arg.tail[0].iov_len;
        svc_rdma_put_context(ctxt, 0);
- out:
        dprintk("svcrdma: ret=%d, rq_arg.len=%u, "
                "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zd\n",
                ret, rqstp->rq_arg.len,