lib/librte_latencystats/rte_latencystats.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <unistd.h>
   6 #include <sys/types.h>
   7 #include <stdbool.h>
   8 #include <math.h>
   9
  10 #include <rte_string_fns.h>
  11 #include <rte_mbuf.h>
  12 #include <rte_log.h>
  13 #include <rte_cycles.h>
  14 #include <rte_ethdev.h>
  15 #include <rte_metrics.h>
  16 #include <rte_memzone.h>
  17 #include <rte_lcore.h>
  18
  19 #include "rte_latencystats.h"
  20
  21 /** Nano seconds per second */
  22 #define NS_PER_SEC 1E9
  23
  24 /** Clock cycles per nano second */
  25 static uint64_t
  26 latencystat_cycles_per_ns(void)
  27 {
  28         return rte_get_timer_hz() / NS_PER_SEC;
  29 }
  30
  31 /* Macros for printing using RTE_LOG */
  32 #define RTE_LOGTYPE_LATENCY_STATS RTE_LOGTYPE_USER1
  33
  34 static const char *MZ_RTE_LATENCY_STATS = "rte_latencystats";
  35 static int latency_stats_index;
  36 static uint64_t samp_intvl;
  37 static uint64_t timer_tsc;
  38 static uint64_t prev_tsc;
  39
  40 struct rte_latency_stats {
  41         float min_latency; /**< Minimum latency in nano seconds */
  42         float avg_latency; /**< Average latency in nano seconds */
  43         float max_latency; /**< Maximum latency in nano seconds */
  44         float jitter; /** Latency variation */
  45 };
  46
  47 static struct rte_latency_stats *glob_stats;
  48
  49 struct rxtx_cbs {
  50         const struct rte_eth_rxtx_callback *cb;
  51 };
  52
  53 static struct rxtx_cbs rx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
  54 static struct rxtx_cbs tx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
  55
  56 struct latency_stats_nameoff {
  57         char name[RTE_ETH_XSTATS_NAME_SIZE];
  58         unsigned int offset;
  59 };
  60
  61 static const struct latency_stats_nameoff lat_stats_strings[] = {
  62         {"min_latency_ns", offsetof(struct rte_latency_stats, min_latency)},
  63         {"avg_latency_ns", offsetof(struct rte_latency_stats, avg_latency)},
  64         {"max_latency_ns", offsetof(struct rte_latency_stats, max_latency)},
  65         {"jitter_ns", offsetof(struct rte_latency_stats, jitter)},
  66 };
  67
  68 #define NUM_LATENCY_STATS (sizeof(lat_stats_strings) / \
  69                                 sizeof(lat_stats_strings[0]))
  70
  71 int32_t
  72 rte_latencystats_update(void)
  73 {
  74         unsigned int i;
  75         float *stats_ptr = NULL;
  76         uint64_t values[NUM_LATENCY_STATS] = {0};
  77         int ret;
  78
  79         for (i = 0; i < NUM_LATENCY_STATS; i++) {
  80                 stats_ptr = RTE_PTR_ADD(glob_stats,
  81                                 lat_stats_strings[i].offset);
  82                 values[i] = (uint64_t)floor((*stats_ptr)/
  83                                 latencystat_cycles_per_ns());
  84         }
  85
  86         ret = rte_metrics_update_values(RTE_METRICS_GLOBAL,
  87                                         latency_stats_index,
  88                                         values, NUM_LATENCY_STATS);
  89         if (ret < 0)
  90                 RTE_LOG(INFO, LATENCY_STATS, "Failed to push the stats\n");
  91
  92         return ret;
  93 }
  94
  95 static void
  96 rte_latencystats_fill_values(struct rte_metric_value *values)
  97 {
  98         unsigned int i;
  99         float *stats_ptr = NULL;
 100
 101         for (i = 0; i < NUM_LATENCY_STATS; i++) {
 102                 stats_ptr = RTE_PTR_ADD(glob_stats,
 103                                 lat_stats_strings[i].offset);
 104                 values[i].key = i;
 105                 values[i].value = (uint64_t)floor((*stats_ptr)/
 106                                                 latencystat_cycles_per_ns());
 107         }
 108 }
 109
 110 static uint16_t
 111 add_time_stamps(uint16_t pid __rte_unused,
 112                 uint16_t qid __rte_unused,
 113                 struct rte_mbuf **pkts,
 114                 uint16_t nb_pkts,
 115                 uint16_t max_pkts __rte_unused,
 116                 void *user_cb __rte_unused)
 117 {
 118         unsigned int i;
 119         uint64_t diff_tsc, now;
 120
 121         /*
 122          * For every sample interval,
 123          * time stamp is marked on one received packet.
 124          */
 125         now = rte_rdtsc();
 126         for (i = 0; i < nb_pkts; i++) {
 127                 diff_tsc = now - prev_tsc;
 128                 timer_tsc += diff_tsc;
 129
 130                 if ((pkts[i]->ol_flags & PKT_RX_TIMESTAMP) == 0
 131                                 && (timer_tsc >= samp_intvl)) {
 132                         pkts[i]->timestamp = now;
 133                         pkts[i]->ol_flags |= PKT_RX_TIMESTAMP;
 134                         timer_tsc = 0;
 135                 }
 136                 prev_tsc = now;
 137                 now = rte_rdtsc();
 138         }
 139
 140         return nb_pkts;
 141 }
 142
 143 static uint16_t
 144 calc_latency(uint16_t pid __rte_unused,
 145                 uint16_t qid __rte_unused,
 146                 struct rte_mbuf **pkts,
 147                 uint16_t nb_pkts,
 148                 void *_ __rte_unused)
 149 {
 150         unsigned int i, cnt = 0;
 151         uint64_t now;
 152         float latency[nb_pkts];
 153         static float prev_latency;
 154         /*
 155          * Alpha represents degree of weighting decrease in EWMA,
 156          * a constant smoothing factor between 0 and 1. The value
 157          * is used below for measuring average latency.
 158          */
 159         const float alpha = 0.2;
 160
 161         now = rte_rdtsc();
 162         for (i = 0; i < nb_pkts; i++) {
 163                 if (pkts[i]->ol_flags & PKT_RX_TIMESTAMP)
 164                         latency[cnt++] = now - pkts[i]->timestamp;
 165         }
 166
 167         for (i = 0; i < cnt; i++) {
 168                 /*
 169                  * The jitter is calculated as statistical mean of interpacket
 170                  * delay variation. The "jitter estimate" is computed by taking
 171                  * the absolute values of the ipdv sequence and applying an
 172                  * exponential filter with parameter 1/16 to generate the
 173                  * estimate. i.e J=J+(|D(i-1,i)|-J)/16. Where J is jitter,
 174                  * D(i-1,i) is difference in latency of two consecutive packets
 175                  * i-1 and i.
 176                  * Reference: Calculated as per RFC 5481, sec 4.1,
 177                  * RFC 3393 sec 4.5, RFC 1889 sec.
 178                  */
 179                 glob_stats->jitter +=  (fabsf(prev_latency - latency[i])
 180                                         - glob_stats->jitter)/16;
 181                 if (glob_stats->min_latency == 0)
 182                         glob_stats->min_latency = latency[i];
 183                 else if (latency[i] < glob_stats->min_latency)
 184                         glob_stats->min_latency = latency[i];
 185                 else if (latency[i] > glob_stats->max_latency)
 186                         glob_stats->max_latency = latency[i];
 187                 /*
 188                  * The average latency is measured using exponential moving
 189                  * average, i.e. using EWMA
 190                  * https://en.wikipedia.org/wiki/Moving_average
 191                  */
 192                 glob_stats->avg_latency +=
 193                         alpha * (latency[i] - glob_stats->avg_latency);
 194                 prev_latency = latency[i];
 195         }
 196
 197         return nb_pkts;
 198 }
 199
 200 int
 201 rte_latencystats_init(uint64_t app_samp_intvl,
 202                 rte_latency_stats_flow_type_fn user_cb)
 203 {
 204         unsigned int i;
 205         uint16_t pid;
 206         uint16_t qid;
 207         struct rxtx_cbs *cbs = NULL;
 208         const char *ptr_strings[NUM_LATENCY_STATS] = {0};
 209         const struct rte_memzone *mz = NULL;
 210         const unsigned int flags = 0;
 211         int ret;
 212
 213         if (rte_memzone_lookup(MZ_RTE_LATENCY_STATS))
 214                 return -EEXIST;
 215
 216         /** Allocate stats in shared memory fo multi process support */
 217         mz = rte_memzone_reserve(MZ_RTE_LATENCY_STATS, sizeof(*glob_stats),
 218                                         rte_socket_id(), flags);
 219         if (mz == NULL) {
 220                 RTE_LOG(ERR, LATENCY_STATS, "Cannot reserve memory: %s:%d\n",
 221                         __func__, __LINE__);
 222                 return -ENOMEM;
 223         }
 224
 225         glob_stats = mz->addr;
 226         samp_intvl = app_samp_intvl * latencystat_cycles_per_ns();
 227
 228         /** Register latency stats with stats library */
 229         for (i = 0; i < NUM_LATENCY_STATS; i++)
 230                 ptr_strings[i] = lat_stats_strings[i].name;
 231
 232         latency_stats_index = rte_metrics_reg_names(ptr_strings,
 233                                                         NUM_LATENCY_STATS);
 234         if (latency_stats_index < 0) {
 235                 RTE_LOG(DEBUG, LATENCY_STATS,
 236                         "Failed to register latency stats names\n");
 237                 return -1;
 238         }
 239
 240         /** Register Rx/Tx callbacks */
 241         RTE_ETH_FOREACH_DEV(pid) {
 242                 struct rte_eth_dev_info dev_info;
 243
 244                 ret = rte_eth_dev_info_get(pid, &dev_info);
 245                 if (ret != 0) {
 246                         RTE_LOG(INFO, LATENCY_STATS,
 247                                 "Error during getting device (port %u) info: %s\n",
 248                                 pid, strerror(-ret));
 249
 250                         continue;
 251                 }
 252
 253                 for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
 254                         cbs = &rx_cbs[pid][qid];
 255                         cbs->cb = rte_eth_add_first_rx_callback(pid, qid,
 256                                         add_time_stamps, user_cb);
 257                         if (!cbs->cb)
 258                                 RTE_LOG(INFO, LATENCY_STATS, "Failed to "
 259                                         "register Rx callback for pid=%d, "
 260                                         "qid=%d\n", pid, qid);
 261                 }
 262                 for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
 263                         cbs = &tx_cbs[pid][qid];
 264                         cbs->cb =  rte_eth_add_tx_callback(pid, qid,
 265                                         calc_latency, user_cb);
 266                         if (!cbs->cb)
 267                                 RTE_LOG(INFO, LATENCY_STATS, "Failed to "
 268                                         "register Tx callback for pid=%d, "
 269                                         "qid=%d\n", pid, qid);
 270                 }
 271         }
 272         return 0;
 273 }
 274
 275 int
 276 rte_latencystats_uninit(void)
 277 {
 278         uint16_t pid;
 279         uint16_t qid;
 280         int ret = 0;
 281         struct rxtx_cbs *cbs = NULL;
 282         const struct rte_memzone *mz = NULL;
 283
 284         /** De register Rx/Tx callbacks */
 285         RTE_ETH_FOREACH_DEV(pid) {
 286                 struct rte_eth_dev_info dev_info;
 287
 288                 ret = rte_eth_dev_info_get(pid, &dev_info);
 289                 if (ret != 0) {
 290                         RTE_LOG(INFO, LATENCY_STATS,
 291                                 "Error during getting device (port %u) info: %s\n",
 292                                 pid, strerror(-ret));
 293
 294                         continue;
 295                 }
 296
 297                 for (qid = 0; qid < dev_info.nb_rx_queues; qid++) {
 298                         cbs = &rx_cbs[pid][qid];
 299                         ret = rte_eth_remove_rx_callback(pid, qid, cbs->cb);
 300                         if (ret)
 301                                 RTE_LOG(INFO, LATENCY_STATS, "failed to "
 302                                         "remove Rx callback for pid=%d, "
 303                                         "qid=%d\n", pid, qid);
 304                 }
 305                 for (qid = 0; qid < dev_info.nb_tx_queues; qid++) {
 306                         cbs = &tx_cbs[pid][qid];
 307                         ret = rte_eth_remove_tx_callback(pid, qid, cbs->cb);
 308                         if (ret)
 309                                 RTE_LOG(INFO, LATENCY_STATS, "failed to "
 310                                         "remove Tx callback for pid=%d, "
 311                                         "qid=%d\n", pid, qid);
 312                 }
 313         }
 314
 315         /* free up the memzone */
 316         mz = rte_memzone_lookup(MZ_RTE_LATENCY_STATS);
 317         if (mz)
 318                 rte_memzone_free(mz);
 319
 320         return 0;
 321 }
 322
 323 int
 324 rte_latencystats_get_names(struct rte_metric_name *names, uint16_t size)
 325 {
 326         unsigned int i;
 327
 328         if (names == NULL || size < NUM_LATENCY_STATS)
 329                 return NUM_LATENCY_STATS;
 330
 331         for (i = 0; i < NUM_LATENCY_STATS; i++)
 332                 strlcpy(names[i].name, lat_stats_strings[i].name,
 333                         sizeof(names[i].name));
 334
 335         return NUM_LATENCY_STATS;
 336 }
 337
 338 int
 339 rte_latencystats_get(struct rte_metric_value *values, uint16_t size)
 340 {
 341         if (size < NUM_LATENCY_STATS || values == NULL)
 342                 return NUM_LATENCY_STATS;
 343
 344         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 345                 const struct rte_memzone *mz;
 346                 mz = rte_memzone_lookup(MZ_RTE_LATENCY_STATS);
 347                 if (mz == NULL) {
 348                         RTE_LOG(ERR, LATENCY_STATS,
 349                                 "Latency stats memzone not found\n");
 350                         return -ENOMEM;
 351                 }
 352                 glob_stats =  mz->addr;
 353         }
 354
 355         /* Retrieve latency stats */
 356         rte_latencystats_fill_values(values);
 357
 358         return NUM_LATENCY_STATS;
 359 }