SUNRPC: provide a mechanism for collecting stats in the RPC client

Add a simple mechanism for collecting stats in the RPC client.  Stats are
tabulated during xprt_release.  Note that per_cpu shenanigans are not
required here because the RPC client already serializes on the transport
write lock.

Test plan:
Compile kernel with CONFIG_NFS enabled.  Basic performance regression
testing with high-speed networking and high performance server.

Signed-off-by: Chuck Lever <cel@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

2 files changed, 79 insertions, 1 deletions

diff --git a/include/linux/sunrpc/clnt.h b/include/linux/sunrpc/clnt.h
index f147e6b84332..0f3662002ffc 100644
--- a/include/linux/sunrpc/clnt.h
+++ b/include/linux/sunrpc/clnt.h
@@ -45,7 +45,8 @@ struct rpc_clnt {
 	char *			cl_server;	/* server machine name */
 	char *			cl_protname;	/* protocol name */
 	struct rpc_auth *	cl_auth;	/* authenticator */
-	struct rpc_stat *	cl_stats;	/* statistics */
+	struct rpc_stat *	cl_stats;	/* per-program statistics */
+	struct rpc_iostats *	cl_metrics;	/* per-client statistics */
 
 	unsigned int		cl_softrtry : 1,/* soft timeouts */
 				cl_intr     : 1,/* interruptible */
diff --git a/include/linux/sunrpc/metrics.h b/include/linux/sunrpc/metrics.h
new file mode 100644
index 000000000000..8f96e9dc369a
--- /dev/null
+++ b/include/linux/sunrpc/metrics.h
@@ -0,0 +1,77 @@
+/*
+ *  linux/include/linux/sunrpc/metrics.h
+ *
+ *  Declarations for RPC client per-operation metrics
+ *
+ *  Copyright (C) 2005	Chuck Lever <cel@netapp.com>
+ *
+ *  RPC client per-operation statistics provide latency and retry
+ *  information about each type of RPC procedure in a given RPC program.
+ *  These statistics are not for detailed problem diagnosis, but simply
+ *  to indicate whether the problem is local or remote.
+ *
+ *  These counters are not meant to be human-readable, but are meant to be
+ *  integrated into system monitoring tools such as "sar" and "iostat".  As
+ *  such, the counters are sampled by the tools over time, and are never
+ *  zeroed after a file system is mounted.  Moving averages can be computed
+ *  by the tools by taking the difference between two instantaneous samples
+ *  and dividing that by the time between the samples.
+ *
+ *  The counters are maintained in a single array per RPC client, indexed
+ *  by procedure number.  There is no need to maintain separate counter
+ *  arrays per-CPU because these counters are always modified behind locks.
+ */
+
+#ifndef _LINUX_SUNRPC_METRICS_H
+#define _LINUX_SUNRPC_METRICS_H
+
+#include <linux/seq_file.h>
+
+#define RPC_IOSTATS_VERS	"1.0"
+
+struct rpc_iostats {
+	/*
+	 * These counters give an idea about how many request
+	 * transmissions are required, on average, to complete that
+	 * particular procedure.  Some procedures may require more
+	 * than one transmission because the server is unresponsive,
+	 * the client is retransmitting too aggressively, or the
+	 * requests are large and the network is congested.
+	 */
+	unsigned long		om_ops,		/* count of operations */
+				om_ntrans,	/* count of RPC transmissions */
+				om_timeouts;	/* count of major timeouts */
+
+	/*
+	 * These count how many bytes are sent and received for a
+	 * given RPC procedure type.  This indicates how much load a
+	 * particular procedure is putting on the network.  These
+	 * counts include the RPC and ULP headers, and the request
+	 * payload.
+	 */
+	unsigned long long      om_bytes_sent,	/* count of bytes out */
+				om_bytes_recv;	/* count of bytes in */
+
+	/*
+	 * The length of time an RPC request waits in queue before
+	 * transmission, the network + server latency of the request,
+	 * and the total time the request spent from init to release
+	 * are measured.
+	 */
+	unsigned long long	om_queue,	/* jiffies queued for xmit */
+				om_rtt,		/* jiffies for RPC RTT */
+				om_execute;	/* jiffies for RPC execution */
+} ____cacheline_aligned;
+
+struct rpc_task;
+struct rpc_clnt;
+
+/*
+ * EXPORTed functions for managing rpc_iostats structures
+ */
+struct rpc_iostats *	rpc_alloc_iostats(struct rpc_clnt *);
+void			rpc_count_iostats(struct rpc_task *);
+void			rpc_print_iostats(struct seq_file *, struct rpc_clnt *);
+void			rpc_free_iostats(struct rpc_iostats *);
+
+#endif /* _LINUX_SUNRPC_METRICS_H */