ceph/src/seastar/dpdk/examples/server_node_efd/server/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2017 Intel Corporation
   3  */
   4
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <unistd.h>
   9 #include <stdint.h>
  10 #include <stdarg.h>
  11 #include <inttypes.h>
  12 #include <sys/queue.h>
  13 #include <errno.h>
  14 #include <netinet/ip.h>
  15
  16 #include <rte_common.h>
  17 #include <rte_memory.h>
  18 #include <rte_eal.h>
  19 #include <rte_launch.h>
  20 #include <rte_per_lcore.h>
  21 #include <rte_lcore.h>
  22 #include <rte_branch_prediction.h>
  23 #include <rte_atomic.h>
  24 #include <rte_ring.h>
  25 #include <rte_log.h>
  26 #include <rte_debug.h>
  27 #include <rte_mempool.h>
  28 #include <rte_memcpy.h>
  29 #include <rte_mbuf.h>
  30 #include <rte_ether.h>
  31 #include <rte_interrupts.h>
  32 #include <rte_ethdev.h>
  33 #include <rte_byteorder.h>
  34 #include <rte_malloc.h>
  35 #include <rte_string_fns.h>
  36 #include <rte_efd.h>
  37 #include <rte_ip.h>
  38
  39 #include "common.h"
  40 #include "args.h"
  41 #include "init.h"
  42
  43 /*
  44  * When doing reads from the NIC or the node queues,
  45  * use this batch size
  46  */
  47 #define PACKET_READ_SIZE 32
  48
  49 /*
  50  * Local buffers to put packets in, used to send packets in bursts to the
  51  * nodes
  52  */
  53 struct node_rx_buf {
  54         struct rte_mbuf *buffer[PACKET_READ_SIZE];
  55         uint16_t count;
  56 };
  57
  58 struct efd_stats {
  59         uint64_t distributed;
  60         uint64_t drop;
  61 } flow_dist_stats;
  62
  63 /* One buffer per node rx queue - dynamically allocate array */
  64 static struct node_rx_buf *cl_rx_buf;
  65
  66 static const char *
  67 get_printable_mac_addr(uint16_t port)
  68 {
  69         static const char err_address[] = "00:00:00:00:00:00";
  70         static char addresses[RTE_MAX_ETHPORTS][sizeof(err_address)];
  71         struct ether_addr mac;
  72
  73         if (unlikely(port >= RTE_MAX_ETHPORTS))
  74                 return err_address;
  75         if (unlikely(addresses[port][0] == '\0')) {
  76                 rte_eth_macaddr_get(port, &mac);
  77                 snprintf(addresses[port], sizeof(addresses[port]),
  78                                 "%02x:%02x:%02x:%02x:%02x:%02x\n",
  79                                 mac.addr_bytes[0], mac.addr_bytes[1],
  80                                 mac.addr_bytes[2], mac.addr_bytes[3],
  81                                 mac.addr_bytes[4], mac.addr_bytes[5]);
  82         }
  83         return addresses[port];
  84 }
  85
  86 /*
  87  * This function displays the recorded statistics for each port
  88  * and for each node. It uses ANSI terminal codes to clear
  89  * screen when called. It is called from a single non-master
  90  * thread in the server process, when the process is run with more
  91  * than one lcore enabled.
  92  */
  93 static void
  94 do_stats_display(void)
  95 {
  96         unsigned int i, j;
  97         const char clr[] = {27, '[', '2', 'J', '\0'};
  98         const char topLeft[] = {27, '[', '1', ';', '1', 'H', '\0'};
  99         uint64_t port_tx[RTE_MAX_ETHPORTS], port_tx_drop[RTE_MAX_ETHPORTS];
 100         uint64_t node_tx[MAX_NODES], node_tx_drop[MAX_NODES];
 101
 102         /* to get TX stats, we need to do some summing calculations */
 103         memset(port_tx, 0, sizeof(port_tx));
 104         memset(port_tx_drop, 0, sizeof(port_tx_drop));
 105         memset(node_tx, 0, sizeof(node_tx));
 106         memset(node_tx_drop, 0, sizeof(node_tx_drop));
 107
 108         for (i = 0; i < num_nodes; i++) {
 109                 const struct tx_stats *tx = &info->tx_stats[i];
 110
 111                 for (j = 0; j < info->num_ports; j++) {
 112                         const uint64_t tx_val = tx->tx[info->id[j]];
 113                         const uint64_t drop_val = tx->tx_drop[info->id[j]];
 114
 115                         port_tx[j] += tx_val;
 116                         port_tx_drop[j] += drop_val;
 117                         node_tx[i] += tx_val;
 118                         node_tx_drop[i] += drop_val;
 119                 }
 120         }
 121
 122         /* Clear screen and move to top left */
 123         printf("%s%s", clr, topLeft);
 124
 125         printf("PORTS\n");
 126         printf("-----\n");
 127         for (i = 0; i < info->num_ports; i++)
 128                 printf("Port %u: '%s'\t", (unsigned int)info->id[i],
 129                                 get_printable_mac_addr(info->id[i]));
 130         printf("\n\n");
 131         for (i = 0; i < info->num_ports; i++) {
 132                 printf("Port %u - rx: %9"PRIu64"\t"
 133                                 "tx: %9"PRIu64"\n",
 134                                 (unsigned int)info->id[i], info->rx_stats.rx[i],
 135                                 port_tx[i]);
 136         }
 137
 138         printf("\nSERVER\n");
 139         printf("-----\n");
 140         printf("distributed: %9"PRIu64", drop: %9"PRIu64"\n",
 141                         flow_dist_stats.distributed, flow_dist_stats.drop);
 142
 143         printf("\nNODES\n");
 144         printf("-------\n");
 145         for (i = 0; i < num_nodes; i++) {
 146                 const unsigned long long rx = nodes[i].stats.rx;
 147                 const unsigned long long rx_drop = nodes[i].stats.rx_drop;
 148                 const struct filter_stats *filter = &info->filter_stats[i];
 149
 150                 printf("Node %2u - rx: %9llu, rx_drop: %9llu\n"
 151                                 "            tx: %9"PRIu64", tx_drop: %9"PRIu64"\n"
 152                                 "            filter_passed: %9"PRIu64", "
 153                                 "filter_drop: %9"PRIu64"\n",
 154                                 i, rx, rx_drop, node_tx[i], node_tx_drop[i],
 155                                 filter->passed, filter->drop);
 156         }
 157
 158         printf("\n");
 159 }
 160
 161 /*
 162  * The function called from each non-master lcore used by the process.
 163  * The test_and_set function is used to randomly pick a single lcore on which
 164  * the code to display the statistics will run. Otherwise, the code just
 165  * repeatedly sleeps.
 166  */
 167 static int
 168 sleep_lcore(__attribute__((unused)) void *dummy)
 169 {
 170         /* Used to pick a display thread - static, so zero-initialised */
 171         static rte_atomic32_t display_stats;
 172
 173         /* Only one core should display stats */
 174         if (rte_atomic32_test_and_set(&display_stats)) {
 175                 const unsigned int sleeptime = 1;
 176
 177                 printf("Core %u displaying statistics\n", rte_lcore_id());
 178
 179                 /* Longer initial pause so above printf is seen */
 180                 sleep(sleeptime * 3);
 181
 182                 /* Loop forever: sleep always returns 0 or <= param */
 183                 while (sleep(sleeptime) <= sleeptime)
 184                         do_stats_display();
 185         }
 186         return 0;
 187 }
 188
 189 /*
 190  * Function to set all the node statistic values to zero.
 191  * Called at program startup.
 192  */
 193 static void
 194 clear_stats(void)
 195 {
 196         unsigned int i;
 197
 198         for (i = 0; i < num_nodes; i++)
 199                 nodes[i].stats.rx = nodes[i].stats.rx_drop = 0;
 200 }
 201
 202 /*
 203  * send a burst of traffic to a node, assuming there are packets
 204  * available to be sent to this node
 205  */
 206 static void
 207 flush_rx_queue(uint16_t node)
 208 {
 209         uint16_t j;
 210         struct node *cl;
 211
 212         if (cl_rx_buf[node].count == 0)
 213                 return;
 214
 215         cl = &nodes[node];
 216         if (rte_ring_enqueue_bulk(cl->rx_q, (void **)cl_rx_buf[node].buffer,
 217                         cl_rx_buf[node].count, NULL) != cl_rx_buf[node].count){
 218                 for (j = 0; j < cl_rx_buf[node].count; j++)
 219                         rte_pktmbuf_free(cl_rx_buf[node].buffer[j]);
 220                 cl->stats.rx_drop += cl_rx_buf[node].count;
 221         } else
 222                 cl->stats.rx += cl_rx_buf[node].count;
 223
 224         cl_rx_buf[node].count = 0;
 225 }
 226
 227 /*
 228  * marks a packet down to be sent to a particular node process
 229  */
 230 static inline void
 231 enqueue_rx_packet(uint8_t node, struct rte_mbuf *buf)
 232 {
 233         cl_rx_buf[node].buffer[cl_rx_buf[node].count++] = buf;
 234 }
 235
 236 /*
 237  * This function takes a group of packets and routes them
 238  * individually to the node process. Very simply round-robins the packets
 239  * without checking any of the packet contents.
 240  */
 241 static void
 242 process_packets(uint32_t port_num __rte_unused, struct rte_mbuf *pkts[],
 243                 uint16_t rx_count, unsigned int socket_id)
 244 {
 245         uint16_t i;
 246         uint8_t node;
 247         efd_value_t data[RTE_EFD_BURST_MAX];
 248         const void *key_ptrs[RTE_EFD_BURST_MAX];
 249
 250         struct ipv4_hdr *ipv4_hdr;
 251         uint32_t ipv4_dst_ip[RTE_EFD_BURST_MAX];
 252
 253         for (i = 0; i < rx_count; i++) {
 254                 /* Handle IPv4 header.*/
 255                 ipv4_hdr = rte_pktmbuf_mtod_offset(pkts[i], struct ipv4_hdr *,
 256                                 sizeof(struct ether_hdr));
 257                 ipv4_dst_ip[i] = ipv4_hdr->dst_addr;
 258                 key_ptrs[i] = (void *)&ipv4_dst_ip[i];
 259         }
 260
 261         rte_efd_lookup_bulk(efd_table, socket_id, rx_count,
 262                                 (const void **) key_ptrs, data);
 263         for (i = 0; i < rx_count; i++) {
 264                 node = (uint8_t) ((uintptr_t)data[i]);
 265
 266                 if (node >= num_nodes) {
 267                         /*
 268                          * Node is out of range, which means that
 269                          * flow has not been inserted
 270                          */
 271                         flow_dist_stats.drop++;
 272                         rte_pktmbuf_free(pkts[i]);
 273                 } else {
 274                         flow_dist_stats.distributed++;
 275                         enqueue_rx_packet(node, pkts[i]);
 276                 }
 277         }
 278
 279         for (i = 0; i < num_nodes; i++)
 280                 flush_rx_queue(i);
 281 }
 282
 283 /*
 284  * Function called by the master lcore of the DPDK process.
 285  */
 286 static void
 287 do_packet_forwarding(void)
 288 {
 289         unsigned int port_num = 0; /* indexes the port[] array */
 290         unsigned int socket_id = rte_socket_id();
 291
 292         for (;;) {
 293                 struct rte_mbuf *buf[PACKET_READ_SIZE];
 294                 uint16_t rx_count;
 295
 296                 /* read a port */
 297                 rx_count = rte_eth_rx_burst(info->id[port_num], 0,
 298                                 buf, PACKET_READ_SIZE);
 299                 info->rx_stats.rx[port_num] += rx_count;
 300
 301                 /* Now process the NIC packets read */
 302                 if (likely(rx_count > 0))
 303                         process_packets(port_num, buf, rx_count, socket_id);
 304
 305                 /* move to next port */
 306                 if (++port_num == info->num_ports)
 307                         port_num = 0;
 308         }
 309 }
 310
 311 int
 312 main(int argc, char *argv[])
 313 {
 314         /* initialise the system */
 315         if (init(argc, argv) < 0)
 316                 return -1;
 317         RTE_LOG(INFO, APP, "Finished Process Init.\n");
 318
 319         cl_rx_buf = calloc(num_nodes, sizeof(cl_rx_buf[0]));
 320
 321         /* clear statistics */
 322         clear_stats();
 323
 324         /* put all other cores to sleep bar master */
 325         rte_eal_mp_remote_launch(sleep_lcore, NULL, SKIP_MASTER);
 326
 327         do_packet_forwarding();
 328         return 0;
 329 }