[dpdk-dev] [PATCH v6 4/4] latencystats: added new library for latency stats

[Remy Horton <remy.horton@intel.com>]

Add a library designed to calculate latency statistics and report them
to the application when queried. The library measures minimum, average and
maximum latencies, and jitter in nano seconds. The current implementation
supports global latency stats, i.e. per application stats.

Added new field to mbuf struct to mark the packet arrival time on Rx.

Modify testpmd code to initialize/uninitialize latency statistics
calulation.

Modify the dpdk-procinfo process to display the newly added metrics.
Added new command line option "--metrics" to display metrics.

APIs:

* Added APIs to initialize and un initialize latency stats
  calculation.
* Added API to retrieve latency stats names and values.

Functionality:

* The library will register ethdev Rx/Tx callbacks for each active port,
  queue combinations.
* The library will register latency stats names with new metrics library.
* Rx packets will be marked with time stamp on each sampling interval.
* On Tx side, packets with time stamp will be considered for calculating
  the minimum, maximum, average latencies and also jitter.
* Average latency is calculated using exponential weighted moving average
  method.
* Minimum and maximum latencies will be low and high latency values
  observed so far.
* Jitter calculation is done based on inter packet delay variation.
* Measured stats are reported to the metrics library in a separate
  pthread.
* Measured stats can be retrieved via get API of the libray (or)
  by calling generic get API of the new metrics library.

Signed-off-by: Reshma Pattan <reshma.pattan@intel.com>
Signed-off-by: Remy Horton <remy.horton@intel.com>
---
 MAINTAINERS                                        |   4 +
 app/proc_info/main.c                               |  73 ++++
 app/test-pmd/testpmd.c                             |  10 +
 config/common_base                                 |   5 +
 doc/api/doxy-api-index.md                          |   1 +
 doc/api/doxy-api.conf                              |   1 +
 doc/guides/rel_notes/release_17_02.rst             |   5 +
 lib/Makefile                                       |   1 +
 lib/librte_latencystats/Makefile                   |  57 +++
 lib/librte_latencystats/rte_latencystats.c         | 389 +++++++++++++++++++++
 lib/librte_latencystats/rte_latencystats.h         | 146 ++++++++
 .../rte_latencystats_version.map                   |  10 +
 lib/librte_mbuf/rte_mbuf.h                         |   3 +
 mk/rte.app.mk                                      |   2 +-
 14 files changed, 706 insertions(+), 1 deletion(-)
 create mode 100644 lib/librte_latencystats/Makefile
 create mode 100644 lib/librte_latencystats/rte_latencystats.c
 create mode 100644 lib/librte_latencystats/rte_latencystats.h
 create mode 100644 lib/librte_latencystats/rte_latencystats_version.map

diff --git a/MAINTAINERS b/MAINTAINERS
index 6cd9896..0a41fe5 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -714,3 +714,7 @@ F: examples/tep_termination/
 F: examples/vmdq/
 F: examples/vmdq_dcb/
 F: doc/guides/sample_app_ug/vmdq_dcb_forwarding.rst
+
+Latency Stats
+M: Reshma Pattan <reshma.pattan@intel.com>
+F: lib/librte_latencystats/
diff --git a/app/proc_info/main.c b/app/proc_info/main.c
index 2c56d10..5f801a9 100644
--- a/app/proc_info/main.c
+++ b/app/proc_info/main.c
@@ -57,6 +57,7 @@
 #include <rte_atomic.h>
 #include <rte_branch_prediction.h>
 #include <rte_string_fns.h>
+#include <rte_metrics.h>
 
 /* Maximum long option length for option parsing. */
 #define MAX_LONG_OPT_SZ 64
@@ -68,6 +69,8 @@ static uint32_t enabled_port_mask;
 static uint32_t enable_stats;
 /**< Enable xstats. */
 static uint32_t enable_xstats;
+/**< Enable metrics. */
+static uint32_t enable_metrics;
 /**< Enable stats reset. */
 static uint32_t reset_stats;
 /**< Enable xstats reset. */
@@ -85,6 +88,8 @@ proc_info_usage(const char *prgname)
 		"  --stats: to display port statistics, enabled by default\n"
 		"  --xstats: to display extended port statistics, disabled by "
 			"default\n"
+		"  --metrics: to display derived metrics of the ports, disabled by "
+			"default\n"
 		"  --stats-reset: to reset port statistics\n"
 		"  --xstats-reset: to reset port extended statistics\n",
 		prgname);
@@ -127,6 +132,7 @@ proc_info_parse_args(int argc, char **argv)
 		{"stats", 0, NULL, 0},
 		{"stats-reset", 0, NULL, 0},
 		{"xstats", 0, NULL, 0},
+		{"metrics", 0, NULL, 0},
 		{"xstats-reset", 0, NULL, 0},
 		{NULL, 0, 0, 0}
 	};
@@ -159,6 +165,10 @@ proc_info_parse_args(int argc, char **argv)
 			else if (!strncmp(long_option[option_index].name, "xstats",
 					MAX_LONG_OPT_SZ))
 				enable_xstats = 1;
+			else if (!strncmp(long_option[option_index].name,
+					"metrics",
+					MAX_LONG_OPT_SZ))
+				enable_metrics = 1;
 			/* Reset stats */
 			if (!strncmp(long_option[option_index].name, "stats-reset",
 					MAX_LONG_OPT_SZ))
@@ -301,6 +311,63 @@ nic_xstats_clear(uint8_t port_id)
 	printf("\n  NIC extended statistics for port %d cleared\n", port_id);
 }
 
+static void
+metrics_display(int port_id)
+{
+	struct rte_metric_value *metrics;
+	struct rte_metric_name *names;
+	int len, ret;
+	static const char *nic_stats_border = "########################";
+
+	memset(&metrics, 0, sizeof(struct rte_metric_value));
+	len = rte_metrics_get_names(NULL, 0);
+	if (len < 0) {
+		printf("Cannot get metrics count\n");
+		return;
+	}
+
+	metrics = malloc(sizeof(struct rte_metric_value) * len);
+	if (metrics == NULL) {
+		printf("Cannot allocate memory for metrics\n");
+		return;
+	}
+
+	names =  malloc(sizeof(struct rte_metric_name) * len);
+	if (names == NULL) {
+		printf("Cannot allocate memory for metrcis names\n");
+		free(metrics);
+		return;
+	}
+
+	if (len != rte_metrics_get_names(names, len)) {
+		printf("Cannot get metrics names\n");
+		free(metrics);
+		free(names);
+		return;
+	}
+
+	if (port_id == RTE_METRICS_NONPORT)
+		printf("###### Non port specific metrics  #########\n");
+	else
+		printf("###### metrics for port %-2d #########\n", port_id);
+	printf("%s############################\n", nic_stats_border);
+	ret = rte_metrics_get_values(port_id, metrics, len);
+	if (ret < 0 || ret > len) {
+		printf("Cannot get metrics values\n");
+		free(metrics);
+		free(names);
+		return;
+	}
+
+	int i;
+	for (i = 0; i < len; i++)
+		printf("%s: %"PRIu64"\n", names[i].name, metrics[i].value);
+
+	printf("%s############################\n", nic_stats_border);
+	free(metrics);
+	free(names);
+}
+
 int
 main(int argc, char **argv)
 {
@@ -360,8 +427,14 @@ main(int argc, char **argv)
 				nic_stats_clear(i);
 			else if (reset_xstats)
 				nic_xstats_clear(i);
+			else if (enable_metrics)
+				metrics_display(i);
 		}
 	}
 
+	/* print port independent stats */
+	if (enable_metrics)
+		metrics_display(RTE_METRICS_NONPORT);
+
 	return 0;
 }
diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
index a0b7430..2874ce4 100644
--- a/app/test-pmd/testpmd.c
+++ b/app/test-pmd/testpmd.c
@@ -83,6 +83,10 @@
 #ifdef RTE_LIBRTE_BITRATE
 #include <rte_bitrate.h>
 #endif
+#include <rte_metrics.h>
+#ifdef RTE_LIBRTE_LATENCY_STATS
+#include <rte_latencystats.h>
+#endif
 
 #include "testpmd.h"
 
@@ -2108,6 +2112,9 @@ signal_handler(int signum)
 		/* uninitialize packet capture framework */
 		rte_pdump_uninit();
 #endif
+#ifdef RTE_LIBRTE_LATENCY_STATS
+		rte_latencystats_uninit();
+#endif
 		force_quit();
 		/* exit with the expected status */
 		signal(signum, SIG_DFL);
@@ -2165,6 +2172,9 @@ main(int argc, char** argv)
 	/* set all ports to promiscuous mode by default */
 	FOREACH_PORT(port_id, ports)
 		rte_eth_promiscuous_enable(port_id);
+#ifdef RTE_LIBRTE_LATENCY_STATS
+	rte_latencystats_init(1, NULL);
+#endif
 
 	/* Setup bitrate stats */
 #ifdef RTE_LIBRTE_BITRATE
diff --git a/config/common_base b/config/common_base
index decebe5..762a54c 100644
--- a/config/common_base
+++ b/config/common_base
@@ -603,3 +603,8 @@ CONFIG_RTE_LIBRTE_METRICS=y
 # Compile the bitrate statistics library
 #
 CONFIG_RTE_LIBRTE_BITRATE=y
+
+#
+# Compile the latency statistics library
+#
+CONFIG_RTE_LIBRTE_LATENCY_STATS=y
diff --git a/doc/api/doxy-api-index.md b/doc/api/doxy-api-index.md
index 5e194b0..1cacacc 100644
--- a/doc/api/doxy-api-index.md
+++ b/doc/api/doxy-api-index.md
@@ -152,4 +152,5 @@ There are many libraries, so their headers may be grouped by topics:
   [keepalive]          (@ref rte_keepalive.h),
   [Device Metrics]     (@ref rte_metrics.h),
   [Bitrate Statistics] (@ref rte_bitrate.h),
+  [Latency stats]      (@ref rte_latencystats.h),
   [version]            (@ref rte_version.h)
diff --git a/doc/api/doxy-api.conf b/doc/api/doxy-api.conf
index ff15f5b..1e55da4 100644
--- a/doc/api/doxy-api.conf
+++ b/doc/api/doxy-api.conf
@@ -46,6 +46,7 @@ INPUT                   = doc/api/doxy-api-index.md \
                           lib/librte_jobstats \
                           lib/librte_kni \
                           lib/librte_kvargs \
+                          lib/librte_latencystats \
                           lib/librte_lpm \
                           lib/librte_mbuf \
                           lib/librte_mempool \
diff --git a/doc/guides/rel_notes/release_17_02.rst b/doc/guides/rel_notes/release_17_02.rst
index 70f93e1..52c7142 100644
--- a/doc/guides/rel_notes/release_17_02.rst
+++ b/doc/guides/rel_notes/release_17_02.rst
@@ -63,6 +63,10 @@ New Features
   See the :ref:`Generic flow API <Generic_flow_API>` documentation for more
   information.
 
+* **Added latency stats library.**
+  A library that facilitates latency stats measurment of the dpdk based applications.
+  The library measures minimum, average, maximum latencies and also jitter
+  in nano seconds.
 
 Resolved Issues
 ---------------
@@ -179,6 +183,7 @@ The libraries prepended with a plus sign were incremented in this version.
      librte_jobstats.so.1
      librte_kni.so.2
      librte_kvargs.so.1
+   + librte_latencystats.so.1
      librte_lpm.so.2
      librte_mbuf.so.2
      librte_mempool.so.2
diff --git a/lib/Makefile b/lib/Makefile
index e211bc0..6fffa88 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -60,6 +60,7 @@ DIRS-$(CONFIG_RTE_LIBRTE_REORDER) += librte_reorder
 DIRS-$(CONFIG_RTE_LIBRTE_PDUMP) += librte_pdump
 DIRS-$(CONFIG_RTE_LIBRTE_METRICS) += librte_metrics
 DIRS-$(CONFIG_RTE_LIBRTE_BITRATE) += librte_bitratestats
+DIRS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) += librte_latencystats
 
 ifeq ($(CONFIG_RTE_EXEC_ENV_LINUXAPP),y)
 DIRS-$(CONFIG_RTE_LIBRTE_KNI) += librte_kni
diff --git a/lib/librte_latencystats/Makefile b/lib/librte_latencystats/Makefile
new file mode 100644
index 0000000..f744da6
--- /dev/null
+++ b/lib/librte_latencystats/Makefile
@@ -0,0 +1,57 @@
+#   BSD LICENSE
+#
+#   Copyright(c) 2016 Intel Corporation. All rights reserved.
+#   All rights reserved.
+#
+#   Redistribution and use in source and binary forms, with or without
+#   modification, are permitted provided that the following conditions
+#   are met:
+#
+#     * Redistributions of source code must retain the above copyright
+#       notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above copyright
+#       notice, this list of conditions and the following disclaimer in
+#       the documentation and/or other materials provided with the
+#       distribution.
+#     * Neither the name of Intel Corporation nor the names of its
+#       contributors may be used to endorse or promote products derived
+#       from this software without specific prior written permission.
+#
+#   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+#   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+#   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+#   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+#   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+#   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+#   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+#   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+#   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+#   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+#   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+include $(RTE_SDK)/mk/rte.vars.mk
+
+# library name
+LIB = librte_latencystats.a
+
+CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -O3
+LDLIBS += -lm
+LDLIBS += -lpthread
+
+EXPORT_MAP := rte_latencystats_version.map
+
+LIBABIVER := 1
+
+# all source are stored in SRCS-y
+SRCS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) := rte_latencystats.c
+
+# install this header file
+SYMLINK-$(CONFIG_RTE_LIBRTE_LATENCY_STATS)-include := rte_latencystats.h
+
+# this lib depends upon:
+DEPDIRS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) += lib/librte_mbuf
+DEPDIRS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) += lib/librte_eal
+DEPDIRS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) += lib/librte_ether
+DEPDIRS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS) += lib/librte_metrics
+
+include $(RTE_SDK)/mk/rte.lib.mk
diff --git a/lib/librte_latencystats/rte_latencystats.c b/lib/librte_latencystats/rte_latencystats.c
new file mode 100644
index 0000000..dcde7f6
--- /dev/null
+++ b/lib/librte_latencystats/rte_latencystats.c
@@ -0,0 +1,389 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2016 Intel Corporation. All rights reserved.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include <unistd.h>
+#include <sys/types.h>
+#include <stdbool.h>
+#include <math.h>
+#include <pthread.h>
+
+#include <rte_mbuf.h>
+#include <rte_log.h>
+#include <rte_cycles.h>
+#include <rte_ethdev.h>
+#include <rte_metrics.h>
+#include <rte_memzone.h>
+#include <rte_lcore.h>
+#include <rte_timer.h>
+
+#include "rte_latencystats.h"
+
+/** Nano seconds per second */
+#define NS_PER_SEC 1E9
+
+/** Clock cycles per nano second */
+#define CYCLES_PER_NS (rte_get_timer_hz() / NS_PER_SEC)
+
+/* Macros for printing using RTE_LOG */
+#define RTE_LOGTYPE_LATENCY_STATS RTE_LOGTYPE_USER1
+
+static pthread_t latency_stats_thread;
+static const char *MZ_RTE_LATENCY_STATS = "rte_latencystats";
+static int latency_stats_index;
+static uint64_t samp_intvl;
+static uint64_t timer_tsc;
+static uint64_t prev_tsc;
+
+static struct rte_latency_stats {
+	float min_latency; /**< Minimum latency in nano seconds */
+	float avg_latency; /**< Average latency in nano seconds */
+	float max_latency; /**< Maximum latency in nano seconds */
+	float jitter; /** Latency variation */
+} *glob_stats;
+
+static struct rxtx_cbs {
+	struct rte_eth_rxtx_callback *cb;
+} rx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT],
+	tx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
+
+struct latency_stats_nameoff {
+	char name[RTE_ETH_XSTATS_NAME_SIZE];
+	unsigned int offset;
+};
+
+static const struct latency_stats_nameoff lat_stats_strings[] = {
+	{"min_latency_ns", offsetof(struct rte_latency_stats, min_latency)},
+	{"avg_latency_ns", offsetof(struct rte_latency_stats, avg_latency)},
+	{"max_latency_ns", offsetof(struct rte_latency_stats, max_latency)},
+	{"jitter_ns", offsetof(struct rte_latency_stats, jitter)},
+};
+
+#define NUM_LATENCY_STATS (sizeof(lat_stats_strings) / \
+				sizeof(lat_stats_strings[0]))
+
+static __attribute__((noreturn)) void *
+report_latency_stats(__rte_unused void *arg)
+{
+	for (;;) {
+		unsigned int i;
+		float *stats_ptr = NULL;
+		uint64_t values[NUM_LATENCY_STATS] = {0};
+		int ret;
+
+		for (i = 0; i < NUM_LATENCY_STATS; i++) {
+			stats_ptr = RTE_PTR_ADD(glob_stats,
+					lat_stats_strings[i].offset);
+			values[i] = (uint64_t)floor((*stats_ptr)/
+					CYCLES_PER_NS);
+		}
+
+		ret = rte_metrics_update_metrics(RTE_METRICS_NONPORT,
+						latency_stats_index,
+						values, NUM_LATENCY_STATS);
+		if (ret < 0)
+			RTE_LOG(INFO, LATENCY_STATS,
+				"Failed to push the stats\n");
+	}
+}
+
+static void
+rte_latencystats_fill_values(struct rte_metric_value *values)
+{
+	unsigned int i;
+	float *stats_ptr = NULL;
+
+	for (i = 0; i < NUM_LATENCY_STATS; i++) {
+		stats_ptr = RTE_PTR_ADD(glob_stats,
+				lat_stats_strings[i].offset);
+		values[i].key = i;
+		values[i].value = (uint64_t)floor((*stats_ptr)/
+						CYCLES_PER_NS);
+	}
+}
+
+static uint16_t
+add_time_stamps(uint8_t pid __rte_unused,
+		uint16_t qid __rte_unused,
+		struct rte_mbuf **pkts,
+		uint16_t nb_pkts,
+		uint16_t max_pkts __rte_unused,
+		void *user_cb __rte_unused)
+{
+	unsigned int i;
+	uint64_t diff_tsc, now;
+
+	/*
+	 * For every sample interval,
+	 * time stamp is marked on one received packet.
+	 */
+	now = rte_rdtsc();
+	for (i = 0; i < nb_pkts; i++) {
+		diff_tsc = now - prev_tsc;
+		timer_tsc += diff_tsc;
+		if (timer_tsc >= samp_intvl) {
+			/*
+			 * TBD: Mark the timestamp only
+			 * if not already marked by the
+			 * hardware or the PMD.
+			 */
+			pkts[i]->timestamp = now;
+			timer_tsc = 0;
+		}
+		prev_tsc = now;
+		now = rte_rdtsc();
+	}
+
+	return nb_pkts;
+}
+
+static uint16_t
+calc_latency(uint8_t pid __rte_unused,
+		uint16_t qid __rte_unused,
+		struct rte_mbuf **pkts,
+		uint16_t nb_pkts,
+		void *_ __rte_unused)
+{
+	unsigned int i, cnt = 0;
+	uint64_t now;
+	float latency[nb_pkts];
+	static float prev_latency;
+	/*
+	 * Alpha represents degree of weighting decrease in EWMA,
+	 * a constant smoothing factor between 0 and 1. The value
+	 * is used below for measuring average latency.
+	 */
+	const float alpha = 0.2;
+
+	now = rte_rdtsc();
+	for (i = 0; i < nb_pkts; i++) {
+		if (pkts[i]->timestamp)
+			latency[cnt++] = now - pkts[i]->timestamp;
+	}
+
+	for (i = 0; i < cnt; i++) {
+		/*
+		 * The jitter is calculated as statistical mean of interpacket
+		 * delay variation. The "jitter estimate" is computed by taking
+		 * the absolute values of the ipdv sequence and applying an
+		 * exponential filter with parameter 1/16 to generate the
+		 * estimate. i.e J=J+(|D(i-1,i)|-J)/16. Where J is jitter,
+		 * D(i-1,i) is difference in latency of two consecutive packets
+		 * i-1 and i.
+		 * Reference: Calculated as per RFC 5481, sec 4.1,
+		 * RFC 3393 sec 4.5, RFC 1889 sec.
+		 */
+		glob_stats->jitter +=  (abs(prev_latency - latency[i])
+					- glob_stats->jitter)/16;
+		if (glob_stats->min_latency == 0)
+			glob_stats->min_latency = latency[i];
+		else if (latency[i] < glob_stats->min_latency)
+			glob_stats->min_latency = latency[i];
+		else if (latency[i] > glob_stats->max_latency)
+			glob_stats->max_latency = latency[i];
+		/*
+		 * The average latency is measured using exponential moving
+		 * average, i.e. using EWMA
+		 * https://en.wikipedia.org/wiki/Moving_average
+		 */
+		glob_stats->avg_latency +=
+			alpha * (latency[i] - glob_stats->avg_latency);
+		prev_latency = latency[i];
+	}
+
+	return nb_pkts;
+}
+
+int
+rte_latencystats_init(uint64_t samp_intvl,
+		rte_latency_stats_flow_type_fn user_cb)
+{
+	unsigned int i;
+	uint8_t pid;
+	uint16_t qid;
+	struct rxtx_cbs *cbs = NULL;
+	const uint8_t nb_ports = rte_eth_dev_count();
+	const char *ptr_strings[NUM_LATENCY_STATS] = {0};
+	const struct rte_memzone *mz = NULL;
+	const unsigned int flags = 0;
+
+	/** Allocate stats in shared memory fo muliti process support */
+	mz = rte_memzone_reserve(MZ_RTE_LATENCY_STATS, sizeof(*glob_stats),
+					rte_socket_id(), flags);
+	if (mz == NULL) {
+		RTE_LOG(ERR, LATENCY_STATS, "Cannot reserve memory: %s:%d\n",
+			__func__, __LINE__);
+		return -ENOMEM;
+	}
+
+	glob_stats = mz->addr;
+	samp_intvl *= CYCLES_PER_NS;
+
+	/** Register latency stats with stats library */
+	for (i = 0; i < NUM_LATENCY_STATS; i++)
+		ptr_strings[i] = lat_stats_strings[i].name;
+
+	latency_stats_index = rte_metrics_reg_metrics(ptr_strings,
+							NUM_LATENCY_STATS);
+	if (latency_stats_index < 0) {
+		RTE_LOG(DEBUG, LATENCY_STATS,
+			"Failed to register latency stats names\n");
+		return -1;
+	}
+
+	/** Register Rx/Tx callbacks */
+	for (pid = 0; pid < nb_ports; pid++) {
+		struct rte_eth_dev_info dev_info;
+		rte_eth_dev_info_get(pid, &dev_info);
+		for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
+			cbs = &rx_cbs[pid][qid];
+			cbs->cb = rte_eth_add_first_rx_callback(pid, qid,
+					add_time_stamps, user_cb);
+			if (!cbs->cb)
+				RTE_LOG(INFO, LATENCY_STATS, "Failed to "
+					"register Rx callback for pid=%d, "
+					"qid=%d\n", pid, qid);
+		}
+		for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
+			cbs = &tx_cbs[pid][qid];
+			cbs->cb =  rte_eth_add_tx_callback(pid, qid,
+					calc_latency, user_cb);
+			if (!cbs->cb)
+				RTE_LOG(INFO, LATENCY_STATS, "Failed to "
+					"register Tx callback for pid=%d, "
+					"qid=%d\n", pid, qid);
+		}
+	}
+
+	int ret = 0;
+	char thread_name[RTE_MAX_THREAD_NAME_LEN];
+
+	/** Create the host thread to update latency stats to stats library */
+	ret = pthread_create(&latency_stats_thread, NULL, report_latency_stats,
+				NULL);
+	if (ret != 0) {
+		RTE_LOG(ERR, LATENCY_STATS,
+			"Failed to create the latency stats thread:%s, %s:%d\n",
+			strerror(errno), __func__, __LINE__);
+		return -1;
+	}
+	/** Set thread_name for aid in debugging */
+	snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN, "latency-stats-thread");
+	ret = rte_thread_setname(latency_stats_thread, thread_name);
+	if (ret != 0)
+		RTE_LOG(DEBUG, LATENCY_STATS,
+			"Failed to set thread name for latency stats handling\n");
+
+	return 0;
+}
+
+int
+rte_latencystats_uninit(void)
+{
+	uint8_t pid;
+	uint16_t qid;
+	int ret = 0;
+	struct rxtx_cbs *cbs = NULL;
+	const uint8_t nb_ports = rte_eth_dev_count();
+
+	/** De register Rx/Tx callbacks */
+	for (pid = 0; pid < nb_ports; pid++) {
+		struct rte_eth_dev_info dev_info;
+		rte_eth_dev_info_get(pid, &dev_info);
+		for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
+			cbs = &rx_cbs[pid][qid];
+			ret = rte_eth_remove_rx_callback(pid, qid, cbs->cb);
+			if (ret)
+				RTE_LOG(INFO, LATENCY_STATS, "failed to "
+					"remove Rx callback for pid=%d, "
+					"qid=%d\n", pid, qid);
+		}
+		for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
+			cbs = &tx_cbs[pid][qid];
+			ret = rte_eth_remove_tx_callback(pid, qid, cbs->cb);
+			if (ret)
+				RTE_LOG(INFO, LATENCY_STATS, "failed to "
+					"remove Tx callback for pid=%d, "
+					"qid=%d\n", pid, qid);
+		}
+	}
+
+	/** Cancel the thread */
+	ret = pthread_cancel(latency_stats_thread);
+	if (ret != 0) {
+		RTE_LOG(ERR, LATENCY_STATS,
+			"Failed to cancel latency stats update thread:"
+			"%s,%s:%d\n",
+			strerror(errno), __func__, __LINE__);
+		return -1;
+	}
+
+	return 0;
+}
+
+int
+rte_latencystats_get_names(struct rte_metric_name *names, uint16_t size)
+{
+	unsigned int i;
+
+	if (names == NULL || size < NUM_LATENCY_STATS)
+		return NUM_LATENCY_STATS;
+
+	for (i = 0; i < NUM_LATENCY_STATS; i++)
+		snprintf(names[i].name, sizeof(names[i].name),
+				"%s", lat_stats_strings[i].name);
+
+	return NUM_LATENCY_STATS;
+}
+
+int
+rte_latencystats_get(struct rte_metric_value *values, uint16_t size)
+{
+	if (size < NUM_LATENCY_STATS || values == NULL)
+		return NUM_LATENCY_STATS;
+
+	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
+		const struct rte_memzone *mz;
+		mz = rte_memzone_lookup(MZ_RTE_LATENCY_STATS);
+		if (mz == NULL) {
+			RTE_LOG(ERR, LATENCY_STATS,
+				"Latency stats memzone not found\n");
+			return -ENOMEM;
+		}
+		glob_stats =  mz->addr;
+	}
+
+	/* Retrieve latency stats */
+	rte_latencystats_fill_values(values);
+
+	return NUM_LATENCY_STATS;
+}
diff --git a/lib/librte_latencystats/rte_latencystats.h b/lib/librte_latencystats/rte_latencystats.h
new file mode 100644
index 0000000..405b878
--- /dev/null
+++ b/lib/librte_latencystats/rte_latencystats.h
@@ -0,0 +1,146 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2016 Intel Corporation. All rights reserved.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef _RTE_LATENCYSTATS_H_
+#define _RTE_LATENCYSTATS_H_
+
+/**
+ * @file
+ * RTE latency stats
+ *
+ * library to provide application and flow based latency stats.
+ */
+
+#include <rte_metrics.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ *  Note: This function pointer is for future flow based latency stats
+ *  implementation.
+ *
+ * Function type used for identifting flow types of a Rx packet.
+ *
+ * The callback function is called on Rx for each packet.
+ * This function is used for flow based latency calculations.
+ *
+ * @param pkt
+ *   Packet that has to be identified with its flow types.
+ * @param user_param
+ *   The arbitrary user parameter passed in by the application when
+ *   the callback was originally configured.
+ * @return
+ *   The flow_mask, representing the multiple flow types of a packet.
+ */
+typedef uint16_t (*rte_latency_stats_flow_type_fn)(struct rte_mbuf *pkt,
+							void *user_param);
+
+/**
+ *  Registers Rx/Tx callbacks for each active port, queue.
+ *
+ * @param samp_intvl
+ *  Sampling time period in nano seconds, at which packet
+ *  should be marked with time stamp.
+ * @param user_cb
+ *  Note: This param is for future flow based latency stats
+ *  implementation.
+ *  User callback to be called to get flow types of a packet.
+ *  Used for flow based latency calculation.
+ *  If the value is NULL, global stats will be calculated,
+ *  else flow based latency stats will be calculated.
+ *  For now just pass on the NULL value to this param.
+ *  @return
+ *   -1     : On error
+ *   -ENOMEM: On error
+ *    0     : On success
+ */
+int rte_latencystats_init(uint64_t samp_intvl,
+			rte_latency_stats_flow_type_fn user_cb);
+
+/**
+ *  Removes registered Rx/Tx callbacks for each active port, queue.
+ *
+ *  @return
+ *   -1: On error
+ *    0: On success
+ */
+int rte_latencystats_uninit(void);
+
+/**
+ * Retrieve names of latency statistics
+ *
+ * @param names
+ *  Block of memory to insert names into. Must be at least size in capacity.
+ *  If set to NULL, function returns required capacity.
+ * @param size
+ *  Capacity of latency stats names (number of names).
+ * @return
+ *   - positive value lower or equal to size: success. The return value
+ *     is the number of entries filled in the stats table.
+ *   - positive value higher than size: error, the given statistics table
+ *     is too small. The return value corresponds to the size that should
+ *     be given to succeed. The entries in the table are not valid and
+ *     shall not be used by the caller.
+ */
+int rte_latencystats_get_names(struct rte_metric_name *names,
+				uint16_t size);
+
+/**
+ * Retrieve latency statistics.
+ *
+ * @param values
+ *   A pointer to a table of structure of type *rte_metric_value*
+ *   to be filled with latency statistics ids and values.
+ *   This parameter can be set to NULL if size is 0.
+ * @param size
+ *   The size of the stats table, which should be large enough to store
+ *   all the latency stats.
+ * @return
+ *   - positive value lower or equal to size: success. The return value
+ *     is the number of entries filled in the stats table.
+ *   - positive value higher than size: error, the given statistics table
+ *     is too small. The return value corresponds to the size that should
+ *     be given to succeed. The entries in the table are not valid and
+ *     shall not be used by the caller.
+ *   -ENOMEM: On failure.
+ */
+int rte_latencystats_get(struct rte_metric_value *values,
+			uint16_t size);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _RTE_LATENCYSTATS_H_ */
diff --git a/lib/librte_latencystats/rte_latencystats_version.map b/lib/librte_latencystats/rte_latencystats_version.map
new file mode 100644
index 0000000..502018e
--- /dev/null
+++ b/lib/librte_latencystats/rte_latencystats_version.map
@@ -0,0 +1,10 @@
+DPDK_17.02 {
+	global:
+
+	rte_latencystats_get;
+	rte_latencystats_get_names;
+	rte_latencystats_init;
+	rte_latencystats_uninit;
+
+	local: *;
+};
diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
index 4476d75..df85d48 100644
--- a/lib/librte_mbuf/rte_mbuf.h
+++ b/lib/librte_mbuf/rte_mbuf.h
@@ -506,6 +506,9 @@ struct rte_mbuf {
 
 	/** Timesync flags for use with IEEE1588. */
 	uint16_t timesync;
+
+	/** Timestamp for measuring latency. */
+	uint64_t timestamp;
 } __rte_cache_aligned;
 
 /**
diff --git a/mk/rte.app.mk b/mk/rte.app.mk
index 6aac5ac..1d36fad 100644
--- a/mk/rte.app.mk
+++ b/mk/rte.app.mk
@@ -100,7 +100,7 @@ _LDLIBS-$(CONFIG_RTE_LIBRTE_CMDLINE)        += -lrte_cmdline
 _LDLIBS-$(CONFIG_RTE_LIBRTE_CFGFILE)        += -lrte_cfgfile
 _LDLIBS-$(CONFIG_RTE_LIBRTE_METRICS)        += -lrte_metrics
 _LDLIBS-$(CONFIG_RTE_LIBRTE_BITRATE)        += -lrte_bitratestats
-
+_LDLIBS-$(CONFIG_RTE_LIBRTE_LATENCY_STATS)  += -lrte_latencystats
 
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_BOND)       += -lrte_pmd_bond
 _LDLIBS-$(CONFIG_RTE_LIBRTE_PMD_XENVIRT)    += -lrte_pmd_xenvirt -lxenstore
-- 
2.5.5