xref: /dpdk/lib/latencystats/rte_latencystats.c (revision bc2dfa4433ce695854781bdbc5f3587eb64d4487)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2018 Intel Corporation
3  */
4 
5 #include <math.h>
6 
7 #include <rte_string_fns.h>
8 #include <rte_mbuf_dyn.h>
9 #include <rte_log.h>
10 #include <rte_cycles.h>
11 #include <rte_ethdev.h>
12 #include <rte_metrics.h>
13 #include <rte_memzone.h>
14 #include <rte_lcore.h>
15 
16 #include "rte_latencystats.h"
17 
18 /** Nano seconds per second */
19 #define NS_PER_SEC 1E9
20 
21 /** Clock cycles per nano second */
22 static uint64_t
latencystat_cycles_per_ns(void)23 latencystat_cycles_per_ns(void)
24 {
25 	return rte_get_timer_hz() / NS_PER_SEC;
26 }
27 
28 RTE_LOG_REGISTER_DEFAULT(latencystat_logtype, INFO);
29 #define RTE_LOGTYPE_LATENCY_STATS latencystat_logtype
30 #define LATENCY_STATS_LOG(level, ...) \
31 	RTE_LOG_LINE(level, LATENCY_STATS, "" __VA_ARGS__)
32 
33 static uint64_t timestamp_dynflag;
34 static int timestamp_dynfield_offset = -1;
35 
36 static inline rte_mbuf_timestamp_t *
timestamp_dynfield(struct rte_mbuf * mbuf)37 timestamp_dynfield(struct rte_mbuf *mbuf)
38 {
39 	return RTE_MBUF_DYNFIELD(mbuf,
40 			timestamp_dynfield_offset, rte_mbuf_timestamp_t *);
41 }
42 
43 static const char *MZ_RTE_LATENCY_STATS = "rte_latencystats";
44 static int latency_stats_index;
45 static uint64_t samp_intvl;
46 static uint64_t timer_tsc;
47 static uint64_t prev_tsc;
48 
49 struct rte_latency_stats {
50 	float min_latency; /**< Minimum latency in nano seconds */
51 	float avg_latency; /**< Average latency in nano seconds */
52 	float max_latency; /**< Maximum latency in nano seconds */
53 	float jitter; /** Latency variation */
54 	rte_spinlock_t lock; /** Latency calculation lock */
55 };
56 
57 static struct rte_latency_stats *glob_stats;
58 
59 struct rxtx_cbs {
60 	const struct rte_eth_rxtx_callback *cb;
61 };
62 
63 static struct rxtx_cbs rx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
64 static struct rxtx_cbs tx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
65 
66 struct latency_stats_nameoff {
67 	char name[RTE_ETH_XSTATS_NAME_SIZE];
68 	unsigned int offset;
69 };
70 
71 static const struct latency_stats_nameoff lat_stats_strings[] = {
72 	{"min_latency_ns", offsetof(struct rte_latency_stats, min_latency)},
73 	{"avg_latency_ns", offsetof(struct rte_latency_stats, avg_latency)},
74 	{"max_latency_ns", offsetof(struct rte_latency_stats, max_latency)},
75 	{"jitter_ns", offsetof(struct rte_latency_stats, jitter)},
76 };
77 
78 #define NUM_LATENCY_STATS (sizeof(lat_stats_strings) / \
79 				sizeof(lat_stats_strings[0]))
80 
81 int32_t
rte_latencystats_update(void)82 rte_latencystats_update(void)
83 {
84 	unsigned int i;
85 	float *stats_ptr = NULL;
86 	uint64_t values[NUM_LATENCY_STATS] = {0};
87 	int ret;
88 
89 	for (i = 0; i < NUM_LATENCY_STATS; i++) {
90 		stats_ptr = RTE_PTR_ADD(glob_stats,
91 				lat_stats_strings[i].offset);
92 		values[i] = (uint64_t)floor((*stats_ptr)/
93 				latencystat_cycles_per_ns());
94 	}
95 
96 	ret = rte_metrics_update_values(RTE_METRICS_GLOBAL,
97 					latency_stats_index,
98 					values, NUM_LATENCY_STATS);
99 	if (ret < 0)
100 		LATENCY_STATS_LOG(INFO, "Failed to push the stats");
101 
102 	return ret;
103 }
104 
105 static void
rte_latencystats_fill_values(struct rte_metric_value * values)106 rte_latencystats_fill_values(struct rte_metric_value *values)
107 {
108 	unsigned int i;
109 	float *stats_ptr = NULL;
110 
111 	for (i = 0; i < NUM_LATENCY_STATS; i++) {
112 		stats_ptr = RTE_PTR_ADD(glob_stats,
113 				lat_stats_strings[i].offset);
114 		values[i].key = i;
115 		values[i].value = (uint64_t)floor((*stats_ptr)/
116 						latencystat_cycles_per_ns());
117 	}
118 }
119 
120 static uint16_t
add_time_stamps(uint16_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)121 add_time_stamps(uint16_t pid __rte_unused,
122 		uint16_t qid __rte_unused,
123 		struct rte_mbuf **pkts,
124 		uint16_t nb_pkts,
125 		uint16_t max_pkts __rte_unused,
126 		void *user_cb __rte_unused)
127 {
128 	unsigned int i;
129 	uint64_t diff_tsc, now;
130 
131 	/*
132 	 * For every sample interval,
133 	 * time stamp is marked on one received packet.
134 	 */
135 	now = rte_rdtsc();
136 	for (i = 0; i < nb_pkts; i++) {
137 		diff_tsc = now - prev_tsc;
138 		timer_tsc += diff_tsc;
139 
140 		if ((pkts[i]->ol_flags & timestamp_dynflag) == 0
141 				&& (timer_tsc >= samp_intvl)) {
142 			*timestamp_dynfield(pkts[i]) = now;
143 			pkts[i]->ol_flags |= timestamp_dynflag;
144 			timer_tsc = 0;
145 		}
146 		prev_tsc = now;
147 		now = rte_rdtsc();
148 	}
149 
150 	return nb_pkts;
151 }
152 
153 static uint16_t
calc_latency(uint16_t pid __rte_unused,uint16_t qid __rte_unused,struct rte_mbuf ** pkts,uint16_t nb_pkts,void * _ __rte_unused)154 calc_latency(uint16_t pid __rte_unused,
155 		uint16_t qid __rte_unused,
156 		struct rte_mbuf **pkts,
157 		uint16_t nb_pkts,
158 		void *_ __rte_unused)
159 {
160 	unsigned int i;
161 	uint64_t now;
162 	float latency;
163 	static float prev_latency;
164 	/*
165 	 * Alpha represents degree of weighting decrease in EWMA,
166 	 * a constant smoothing factor between 0 and 1. The value
167 	 * is used below for measuring average latency.
168 	 */
169 	const float alpha = 0.2f;
170 
171 	now = rte_rdtsc();
172 
173 	rte_spinlock_lock(&glob_stats->lock);
174 	for (i = 0; i < nb_pkts; i++) {
175 		if (!(pkts[i]->ol_flags & timestamp_dynflag))
176 			continue;
177 
178 		latency = now - *timestamp_dynfield(pkts[i]);
179 
180 		/*
181 		 * The jitter is calculated as statistical mean of interpacket
182 		 * delay variation. The "jitter estimate" is computed by taking
183 		 * the absolute values of the ipdv sequence and applying an
184 		 * exponential filter with parameter 1/16 to generate the
185 		 * estimate. i.e J=J+(|D(i-1,i)|-J)/16. Where J is jitter,
186 		 * D(i-1,i) is difference in latency of two consecutive packets
187 		 * i-1 and i.
188 		 * Reference: Calculated as per RFC 5481, sec 4.1,
189 		 * RFC 3393 sec 4.5, RFC 1889 sec.
190 		 */
191 		glob_stats->jitter +=  (fabsf(prev_latency - latency)
192 					- glob_stats->jitter)/16;
193 		if (glob_stats->min_latency == 0)
194 			glob_stats->min_latency = latency;
195 		else if (latency < glob_stats->min_latency)
196 			glob_stats->min_latency = latency;
197 		else if (latency > glob_stats->max_latency)
198 			glob_stats->max_latency = latency;
199 		/*
200 		 * The average latency is measured using exponential moving
201 		 * average, i.e. using EWMA
202 		 * https://en.wikipedia.org/wiki/Moving_average
203 		 */
204 		glob_stats->avg_latency +=
205 			alpha * (latency - glob_stats->avg_latency);
206 		prev_latency = latency;
207 	}
208 	rte_spinlock_unlock(&glob_stats->lock);
209 
210 	return nb_pkts;
211 }
212 
213 int
rte_latencystats_init(uint64_t app_samp_intvl,rte_latency_stats_flow_type_fn user_cb)214 rte_latencystats_init(uint64_t app_samp_intvl,
215 		rte_latency_stats_flow_type_fn user_cb)
216 {
217 	unsigned int i;
218 	uint16_t pid;
219 	uint16_t qid;
220 	struct rxtx_cbs *cbs = NULL;
221 	const char *ptr_strings[NUM_LATENCY_STATS] = {0};
222 	const struct rte_memzone *mz = NULL;
223 	const unsigned int flags = 0;
224 	int ret;
225 
226 	if (rte_memzone_lookup(MZ_RTE_LATENCY_STATS))
227 		return -EEXIST;
228 
229 	/** Allocate stats in shared memory fo multi process support */
230 	mz = rte_memzone_reserve(MZ_RTE_LATENCY_STATS, sizeof(*glob_stats),
231 					rte_socket_id(), flags);
232 	if (mz == NULL) {
233 		LATENCY_STATS_LOG(ERR, "Cannot reserve memory: %s:%d",
234 			__func__, __LINE__);
235 		return -ENOMEM;
236 	}
237 
238 	glob_stats = mz->addr;
239 	rte_spinlock_init(&glob_stats->lock);
240 	samp_intvl = app_samp_intvl * latencystat_cycles_per_ns();
241 
242 	/** Register latency stats with stats library */
243 	for (i = 0; i < NUM_LATENCY_STATS; i++)
244 		ptr_strings[i] = lat_stats_strings[i].name;
245 
246 	latency_stats_index = rte_metrics_reg_names(ptr_strings,
247 							NUM_LATENCY_STATS);
248 	if (latency_stats_index < 0) {
249 		LATENCY_STATS_LOG(DEBUG,
250 			"Failed to register latency stats names");
251 		return -1;
252 	}
253 
254 	/* Register mbuf field and flag for Rx timestamp */
255 	ret = rte_mbuf_dyn_rx_timestamp_register(&timestamp_dynfield_offset,
256 			&timestamp_dynflag);
257 	if (ret != 0) {
258 		LATENCY_STATS_LOG(ERR,
259 			"Cannot register mbuf field/flag for timestamp");
260 		return -rte_errno;
261 	}
262 
263 	/** Register Rx/Tx callbacks */
264 	RTE_ETH_FOREACH_DEV(pid) {
265 		struct rte_eth_dev_info dev_info;
266 
267 		ret = rte_eth_dev_info_get(pid, &dev_info);
268 		if (ret != 0) {
269 			LATENCY_STATS_LOG(INFO,
270 				"Error during getting device (port %u) info: %s",
271 				pid, strerror(-ret));
272 
273 			continue;
274 		}
275 
276 		for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
277 			cbs = &rx_cbs[pid][qid];
278 			cbs->cb = rte_eth_add_first_rx_callback(pid, qid,
279 					add_time_stamps, user_cb);
280 			if (!cbs->cb)
281 				LATENCY_STATS_LOG(INFO, "Failed to "
282 					"register Rx callback for pid=%d, "
283 					"qid=%d", pid, qid);
284 		}
285 		for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
286 			cbs = &tx_cbs[pid][qid];
287 			cbs->cb =  rte_eth_add_tx_callback(pid, qid,
288 					calc_latency, user_cb);
289 			if (!cbs->cb)
290 				LATENCY_STATS_LOG(INFO, "Failed to "
291 					"register Tx callback for pid=%d, "
292 					"qid=%d", pid, qid);
293 		}
294 	}
295 	return 0;
296 }
297 
298 int
rte_latencystats_uninit(void)299 rte_latencystats_uninit(void)
300 {
301 	uint16_t pid;
302 	uint16_t qid;
303 	int ret = 0;
304 	struct rxtx_cbs *cbs = NULL;
305 	const struct rte_memzone *mz = NULL;
306 
307 	/** De register Rx/Tx callbacks */
308 	RTE_ETH_FOREACH_DEV(pid) {
309 		struct rte_eth_dev_info dev_info;
310 
311 		ret = rte_eth_dev_info_get(pid, &dev_info);
312 		if (ret != 0) {
313 			LATENCY_STATS_LOG(INFO,
314 				"Error during getting device (port %u) info: %s",
315 				pid, strerror(-ret));
316 
317 			continue;
318 		}
319 
320 		for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
321 			cbs = &rx_cbs[pid][qid];
322 			ret = rte_eth_remove_rx_callback(pid, qid, cbs->cb);
323 			if (ret)
324 				LATENCY_STATS_LOG(INFO, "failed to "
325 					"remove Rx callback for pid=%d, "
326 					"qid=%d", pid, qid);
327 		}
328 		for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
329 			cbs = &tx_cbs[pid][qid];
330 			ret = rte_eth_remove_tx_callback(pid, qid, cbs->cb);
331 			if (ret)
332 				LATENCY_STATS_LOG(INFO, "failed to "
333 					"remove Tx callback for pid=%d, "
334 					"qid=%d", pid, qid);
335 		}
336 	}
337 
338 	/* free up the memzone */
339 	mz = rte_memzone_lookup(MZ_RTE_LATENCY_STATS);
340 	rte_memzone_free(mz);
341 
342 	return 0;
343 }
344 
345 int
rte_latencystats_get_names(struct rte_metric_name * names,uint16_t size)346 rte_latencystats_get_names(struct rte_metric_name *names, uint16_t size)
347 {
348 	unsigned int i;
349 
350 	if (names == NULL || size < NUM_LATENCY_STATS)
351 		return NUM_LATENCY_STATS;
352 
353 	for (i = 0; i < NUM_LATENCY_STATS; i++)
354 		strlcpy(names[i].name, lat_stats_strings[i].name,
355 			sizeof(names[i].name));
356 
357 	return NUM_LATENCY_STATS;
358 }
359 
360 int
rte_latencystats_get(struct rte_metric_value * values,uint16_t size)361 rte_latencystats_get(struct rte_metric_value *values, uint16_t size)
362 {
363 	if (size < NUM_LATENCY_STATS || values == NULL)
364 		return NUM_LATENCY_STATS;
365 
366 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
367 		const struct rte_memzone *mz;
368 		mz = rte_memzone_lookup(MZ_RTE_LATENCY_STATS);
369 		if (mz == NULL) {
370 			LATENCY_STATS_LOG(ERR,
371 				"Latency stats memzone not found");
372 			return -ENOMEM;
373 		}
374 		glob_stats =  mz->addr;
375 	}
376 
377 	/* Retrieve latency stats */
378 	rte_latencystats_fill_values(values);
379 
380 	return NUM_LATENCY_STATS;
381 }
382