From ccede7598588ae344143f82fb763912535648d58 Mon Sep 17 00:00:00 2001
From: Chuck Lever <chuck.lever@oracle.com>
Date: Mon, 4 Dec 2017 14:04:04 -0500
Subject: xprtrdma: Spread reply processing over more CPUs

Commit d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler
directly from RECV completion") introduced a performance regression
for NFS I/O small enough to not need memory registration. In multi-
threaded benchmarks that generate primarily small I/O requests,
IOPS throughput is reduced by nearly a third. This patch restores
the previous level of throughput.

Because workqueues are typically BOUND (in particular ib_comp_wq,
nfsiod_workqueue, and rpciod_workqueue), NFS/RDMA workloads tend
to aggregate on the CPU that is handling Receive completions.

The usual approach to addressing this problem is to create a QP
and CQ for each CPU, and then schedule transactions on the QP
for the CPU where you want the transaction to complete. The
transaction then does not require an extra context switch during
completion to end up on the same CPU where the transaction was
started.

This approach doesn't work for the Linux NFS/RDMA client because
currently the Linux NFS client does not support multiple connections
per client-server pair, and the RDMA core API does not make it
straightforward for ULPs to determine which CPU is responsible for
handling Receive completions for a CQ.

So for the moment, record the CPU number in the rpcrdma_req before
the transport sends each RPC Call. Then during Receive completion,
queue the RPC completion on that same CPU.

Additionally, move all RPC completion processing to the deferred
handler so that even RPCs with simple small replies complete on
the CPU that sent the corresponding RPC Call.

Fixes: d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler ...")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
---
 net/sunrpc/xprtrdma/rpc_rdma.c  | 6 +-----
 net/sunrpc/xprtrdma/transport.c | 2 ++
 net/sunrpc/xprtrdma/verbs.c     | 2 +-
 net/sunrpc/xprtrdma/xprt_rdma.h | 1 +
 4 files changed, 5 insertions(+), 6 deletions(-)

(limited to 'net/sunrpc')

diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index ed34dc0f144c..a3f2ab283aeb 100644
--- a/net/sunrpc/xprtrdma/rpc_rdma.c
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -1408,11 +1408,7 @@ void rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 	dprintk("RPC:       %s: reply %p completes request %p (xid 0x%08x)\n",
 		__func__, rep, req, be32_to_cpu(rep->rr_xid));
 
-	if (list_empty(&req->rl_registered) &&
-	    !test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags))
-		rpcrdma_complete_rqst(rep);
-	else
-		queue_work(rpcrdma_receive_wq, &rep->rr_work);
+	queue_work_on(req->rl_cpu, rpcrdma_receive_wq, &rep->rr_work);
 	return;
 
 out_badstatus:
diff --git a/net/sunrpc/xprtrdma/transport.c b/net/sunrpc/xprtrdma/transport.c
index 646c24494ea7..6ee1ad8978f3 100644
--- a/net/sunrpc/xprtrdma/transport.c
+++ b/net/sunrpc/xprtrdma/transport.c
@@ -52,6 +52,7 @@
 #include <linux/slab.h>
 #include <linux/seq_file.h>
 #include <linux/sunrpc/addr.h>
+#include <linux/smp.h>
 
 #include "xprt_rdma.h"
 
@@ -656,6 +657,7 @@ xprt_rdma_allocate(struct rpc_task *task)
 		task->tk_pid, __func__, rqst->rq_callsize,
 		rqst->rq_rcvsize, req);
 
+	req->rl_cpu = smp_processor_id();
 	req->rl_connect_cookie = 0;	/* our reserved value */
 	rpcrdma_set_xprtdata(rqst, req);
 	rqst->rq_buffer = req->rl_sendbuf->rg_base;
diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index 710b3f77db82..8607c029c0dd 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -83,7 +83,7 @@ rpcrdma_alloc_wq(void)
 	struct workqueue_struct *recv_wq;
 
 	recv_wq = alloc_workqueue("xprtrdma_receive",
-				  WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
+				  WQ_MEM_RECLAIM | WQ_HIGHPRI,
 				  0);
 	if (!recv_wq)
 		return -ENOMEM;
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
index 51686d9eac5f..1342f743f1c4 100644
--- a/net/sunrpc/xprtrdma/xprt_rdma.h
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -342,6 +342,7 @@ enum {
 struct rpcrdma_buffer;
 struct rpcrdma_req {
 	struct list_head	rl_list;
+	int			rl_cpu;
 	unsigned int		rl_connect_cookie;
 	struct rpcrdma_buffer	*rl_buffer;
 	struct rpcrdma_rep	*rl_reply;
-- 
cgit v1.2.3


From 90d91b0cd371193d9dbfa9beacab8ab9a4cb75e0 Mon Sep 17 00:00:00 2001
From: Trond Myklebust <trond.myklebust@primarydata.com>
Date: Thu, 14 Dec 2017 21:24:08 -0500
Subject: SUNRPC: Fix a race in the receive code path

We must ensure that the call to rpc_sleep_on() in xprt_transmit() cannot
race with the call to xprt_complete_rqst().

Reported-by: Chuck Lever <chuck.lever@oracle.com>
Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=317
Fixes: ce7c252a8c74 ("SUNRPC: Add a separate spinlock to protect..")
Cc: stable@vger.kernel.org # 4.14+
Reviewed-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
---
 net/sunrpc/xprt.c | 28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)

(limited to 'net/sunrpc')

diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
index 02a9bacb239b..5b06f6906a27 100644
--- a/net/sunrpc/xprt.c
+++ b/net/sunrpc/xprt.c
@@ -1001,6 +1001,7 @@ void xprt_transmit(struct rpc_task *task)
 {
 	struct rpc_rqst	*req = task->tk_rqstp;
 	struct rpc_xprt	*xprt = req->rq_xprt;
+	unsigned int connect_cookie;
 	int status, numreqs;
 
 	dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
@@ -1024,6 +1025,7 @@ void xprt_transmit(struct rpc_task *task)
 	} else if (!req->rq_bytes_sent)
 		return;
 
+	connect_cookie = xprt->connect_cookie;
 	req->rq_xtime = ktime_get();
 	status = xprt->ops->send_request(task);
 	trace_xprt_transmit(xprt, req->rq_xid, status);
@@ -1047,20 +1049,28 @@ void xprt_transmit(struct rpc_task *task)
 	xprt->stat.bklog_u += xprt->backlog.qlen;
 	xprt->stat.sending_u += xprt->sending.qlen;
 	xprt->stat.pending_u += xprt->pending.qlen;
+	spin_unlock_bh(&xprt->transport_lock);
 
-	/* Don't race with disconnect */
-	if (!xprt_connected(xprt))
-		task->tk_status = -ENOTCONN;
-	else {
+	req->rq_connect_cookie = connect_cookie;
+	if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) {
 		/*
-		 * Sleep on the pending queue since
-		 * we're expecting a reply.
+		 * Sleep on the pending queue if we're expecting a reply.
+		 * The spinlock ensures atomicity between the test of
+		 * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on().
 		 */
-		if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task))
+		spin_lock(&xprt->recv_lock);
+		if (!req->rq_reply_bytes_recvd) {
 			rpc_sleep_on(&xprt->pending, task, xprt_timer);
-		req->rq_connect_cookie = xprt->connect_cookie;
+			/*
+			 * Send an extra queue wakeup call if the
+			 * connection was dropped in case the call to
+			 * rpc_sleep_on() raced.
+			 */
+			if (!xprt_connected(xprt))
+				xprt_wake_pending_tasks(xprt, -ENOTCONN);
+		}
+		spin_unlock(&xprt->recv_lock);
 	}
-	spin_unlock_bh(&xprt->transport_lock);
 }
 
 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
-- 
cgit v1.2.3