1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2017 Intel Corporation 3 */ 4 5 #include <arpa/inet.h> 6 #include <getopt.h> 7 #include <linux/if_ether.h> 8 #include <linux/if_vlan.h> 9 #include <linux/virtio_net.h> 10 #include <linux/virtio_ring.h> 11 #include <signal.h> 12 #include <stdint.h> 13 #include <sys/eventfd.h> 14 #include <sys/param.h> 15 #include <unistd.h> 16 17 #include <rte_cycles.h> 18 #include <rte_ethdev.h> 19 #include <rte_log.h> 20 #include <rte_string_fns.h> 21 #include <rte_malloc.h> 22 #include <rte_net.h> 23 #include <rte_vhost.h> 24 #include <rte_ip.h> 25 #include <rte_tcp.h> 26 #include <rte_pause.h> 27 28 #include "ioat.h" 29 #include "main.h" 30 31 #ifndef MAX_QUEUES 32 #define MAX_QUEUES 128 33 #endif 34 35 /* the maximum number of external ports supported */ 36 #define MAX_SUP_PORTS 1 37 38 #define MBUF_CACHE_SIZE 128 39 #define MBUF_DATA_SIZE RTE_MBUF_DEFAULT_BUF_SIZE 40 41 #define BURST_TX_DRAIN_US 100 /* TX drain every ~100us */ 42 43 #define BURST_RX_WAIT_US 15 /* Defines how long we wait between retries on RX */ 44 #define BURST_RX_RETRIES 4 /* Number of retries on RX. */ 45 46 #define JUMBO_FRAME_MAX_SIZE 0x2600 47 48 /* State of virtio device. */ 49 #define DEVICE_MAC_LEARNING 0 50 #define DEVICE_RX 1 51 #define DEVICE_SAFE_REMOVE 2 52 53 /* Configurable number of RX/TX ring descriptors */ 54 #define RTE_TEST_RX_DESC_DEFAULT 1024 55 #define RTE_TEST_TX_DESC_DEFAULT 512 56 57 #define INVALID_PORT_ID 0xFF 58 59 /* mask of enabled ports */ 60 static uint32_t enabled_port_mask = 0; 61 62 /* Promiscuous mode */ 63 static uint32_t promiscuous; 64 65 /* number of devices/queues to support*/ 66 static uint32_t num_queues = 0; 67 static uint32_t num_devices; 68 69 static struct rte_mempool *mbuf_pool; 70 static int mergeable; 71 72 /* Enable VM2VM communications. If this is disabled then the MAC address compare is skipped. */ 73 typedef enum { 74 VM2VM_DISABLED = 0, 75 VM2VM_SOFTWARE = 1, 76 VM2VM_HARDWARE = 2, 77 VM2VM_LAST 78 } vm2vm_type; 79 static vm2vm_type vm2vm_mode = VM2VM_SOFTWARE; 80 81 /* Enable stats. */ 82 static uint32_t enable_stats = 0; 83 /* Enable retries on RX. */ 84 static uint32_t enable_retry = 1; 85 86 /* Disable TX checksum offload */ 87 static uint32_t enable_tx_csum; 88 89 /* Disable TSO offload */ 90 static uint32_t enable_tso; 91 92 static int client_mode; 93 94 static int builtin_net_driver; 95 96 static int async_vhost_driver; 97 98 static char *dma_type; 99 100 /* Specify timeout (in useconds) between retries on RX. */ 101 static uint32_t burst_rx_delay_time = BURST_RX_WAIT_US; 102 /* Specify the number of retries on RX. */ 103 static uint32_t burst_rx_retry_num = BURST_RX_RETRIES; 104 105 /* Socket file paths. Can be set by user */ 106 static char *socket_files; 107 static int nb_sockets; 108 109 /* empty vmdq configuration structure. Filled in programatically */ 110 static struct rte_eth_conf vmdq_conf_default = { 111 .rxmode = { 112 .mq_mode = ETH_MQ_RX_VMDQ_ONLY, 113 .split_hdr_size = 0, 114 /* 115 * VLAN strip is necessary for 1G NIC such as I350, 116 * this fixes bug of ipv4 forwarding in guest can't 117 * forward pakets from one virtio dev to another virtio dev. 118 */ 119 .offloads = DEV_RX_OFFLOAD_VLAN_STRIP, 120 }, 121 122 .txmode = { 123 .mq_mode = ETH_MQ_TX_NONE, 124 .offloads = (DEV_TX_OFFLOAD_IPV4_CKSUM | 125 DEV_TX_OFFLOAD_TCP_CKSUM | 126 DEV_TX_OFFLOAD_VLAN_INSERT | 127 DEV_TX_OFFLOAD_MULTI_SEGS | 128 DEV_TX_OFFLOAD_TCP_TSO), 129 }, 130 .rx_adv_conf = { 131 /* 132 * should be overridden separately in code with 133 * appropriate values 134 */ 135 .vmdq_rx_conf = { 136 .nb_queue_pools = ETH_8_POOLS, 137 .enable_default_pool = 0, 138 .default_pool = 0, 139 .nb_pool_maps = 0, 140 .pool_map = {{0, 0},}, 141 }, 142 }, 143 }; 144 145 146 static unsigned lcore_ids[RTE_MAX_LCORE]; 147 static uint16_t ports[RTE_MAX_ETHPORTS]; 148 static unsigned num_ports = 0; /**< The number of ports specified in command line */ 149 static uint16_t num_pf_queues, num_vmdq_queues; 150 static uint16_t vmdq_pool_base, vmdq_queue_base; 151 static uint16_t queues_per_pool; 152 153 const uint16_t vlan_tags[] = { 154 1000, 1001, 1002, 1003, 1004, 1005, 1006, 1007, 155 1008, 1009, 1010, 1011, 1012, 1013, 1014, 1015, 156 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 157 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 158 1032, 1033, 1034, 1035, 1036, 1037, 1038, 1039, 159 1040, 1041, 1042, 1043, 1044, 1045, 1046, 1047, 160 1048, 1049, 1050, 1051, 1052, 1053, 1054, 1055, 161 1056, 1057, 1058, 1059, 1060, 1061, 1062, 1063, 162 }; 163 164 /* ethernet addresses of ports */ 165 static struct rte_ether_addr vmdq_ports_eth_addr[RTE_MAX_ETHPORTS]; 166 167 static struct vhost_dev_tailq_list vhost_dev_list = 168 TAILQ_HEAD_INITIALIZER(vhost_dev_list); 169 170 static struct lcore_info lcore_info[RTE_MAX_LCORE]; 171 172 /* Used for queueing bursts of TX packets. */ 173 struct mbuf_table { 174 unsigned len; 175 unsigned txq_id; 176 struct rte_mbuf *m_table[MAX_PKT_BURST]; 177 }; 178 179 struct vhost_bufftable { 180 uint32_t len; 181 uint64_t pre_tsc; 182 struct rte_mbuf *m_table[MAX_PKT_BURST]; 183 }; 184 185 /* TX queue for each data core. */ 186 struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE]; 187 188 /* 189 * Vhost TX buffer for each data core. 190 * Every data core maintains a TX buffer for every vhost device, 191 * which is used for batch pkts enqueue for higher performance. 192 */ 193 struct vhost_bufftable *vhost_txbuff[RTE_MAX_LCORE * MAX_VHOST_DEVICE]; 194 195 #define MBUF_TABLE_DRAIN_TSC ((rte_get_tsc_hz() + US_PER_S - 1) \ 196 / US_PER_S * BURST_TX_DRAIN_US) 197 #define VLAN_HLEN 4 198 199 static inline int 200 open_dma(const char *value) 201 { 202 if (dma_type != NULL && strncmp(dma_type, "ioat", 4) == 0) 203 return open_ioat(value); 204 205 return -1; 206 } 207 208 /* 209 * Builds up the correct configuration for VMDQ VLAN pool map 210 * according to the pool & queue limits. 211 */ 212 static inline int 213 get_eth_conf(struct rte_eth_conf *eth_conf, uint32_t num_devices) 214 { 215 struct rte_eth_vmdq_rx_conf conf; 216 struct rte_eth_vmdq_rx_conf *def_conf = 217 &vmdq_conf_default.rx_adv_conf.vmdq_rx_conf; 218 unsigned i; 219 220 memset(&conf, 0, sizeof(conf)); 221 conf.nb_queue_pools = (enum rte_eth_nb_pools)num_devices; 222 conf.nb_pool_maps = num_devices; 223 conf.enable_loop_back = def_conf->enable_loop_back; 224 conf.rx_mode = def_conf->rx_mode; 225 226 for (i = 0; i < conf.nb_pool_maps; i++) { 227 conf.pool_map[i].vlan_id = vlan_tags[ i ]; 228 conf.pool_map[i].pools = (1UL << i); 229 } 230 231 (void)(rte_memcpy(eth_conf, &vmdq_conf_default, sizeof(*eth_conf))); 232 (void)(rte_memcpy(ð_conf->rx_adv_conf.vmdq_rx_conf, &conf, 233 sizeof(eth_conf->rx_adv_conf.vmdq_rx_conf))); 234 return 0; 235 } 236 237 /* 238 * Initialises a given port using global settings and with the rx buffers 239 * coming from the mbuf_pool passed as parameter 240 */ 241 static inline int 242 port_init(uint16_t port) 243 { 244 struct rte_eth_dev_info dev_info; 245 struct rte_eth_conf port_conf; 246 struct rte_eth_rxconf *rxconf; 247 struct rte_eth_txconf *txconf; 248 int16_t rx_rings, tx_rings; 249 uint16_t rx_ring_size, tx_ring_size; 250 int retval; 251 uint16_t q; 252 253 /* The max pool number from dev_info will be used to validate the pool number specified in cmd line */ 254 retval = rte_eth_dev_info_get(port, &dev_info); 255 if (retval != 0) { 256 RTE_LOG(ERR, VHOST_PORT, 257 "Error during getting device (port %u) info: %s\n", 258 port, strerror(-retval)); 259 260 return retval; 261 } 262 263 rxconf = &dev_info.default_rxconf; 264 txconf = &dev_info.default_txconf; 265 rxconf->rx_drop_en = 1; 266 267 /*configure the number of supported virtio devices based on VMDQ limits */ 268 num_devices = dev_info.max_vmdq_pools; 269 270 rx_ring_size = RTE_TEST_RX_DESC_DEFAULT; 271 tx_ring_size = RTE_TEST_TX_DESC_DEFAULT; 272 273 tx_rings = (uint16_t)rte_lcore_count(); 274 275 /* Get port configuration. */ 276 retval = get_eth_conf(&port_conf, num_devices); 277 if (retval < 0) 278 return retval; 279 /* NIC queues are divided into pf queues and vmdq queues. */ 280 num_pf_queues = dev_info.max_rx_queues - dev_info.vmdq_queue_num; 281 queues_per_pool = dev_info.vmdq_queue_num / dev_info.max_vmdq_pools; 282 num_vmdq_queues = num_devices * queues_per_pool; 283 num_queues = num_pf_queues + num_vmdq_queues; 284 vmdq_queue_base = dev_info.vmdq_queue_base; 285 vmdq_pool_base = dev_info.vmdq_pool_base; 286 printf("pf queue num: %u, configured vmdq pool num: %u, each vmdq pool has %u queues\n", 287 num_pf_queues, num_devices, queues_per_pool); 288 289 if (!rte_eth_dev_is_valid_port(port)) 290 return -1; 291 292 rx_rings = (uint16_t)dev_info.max_rx_queues; 293 if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE) 294 port_conf.txmode.offloads |= 295 DEV_TX_OFFLOAD_MBUF_FAST_FREE; 296 /* Configure ethernet device. */ 297 retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf); 298 if (retval != 0) { 299 RTE_LOG(ERR, VHOST_PORT, "Failed to configure port %u: %s.\n", 300 port, strerror(-retval)); 301 return retval; 302 } 303 304 retval = rte_eth_dev_adjust_nb_rx_tx_desc(port, &rx_ring_size, 305 &tx_ring_size); 306 if (retval != 0) { 307 RTE_LOG(ERR, VHOST_PORT, "Failed to adjust number of descriptors " 308 "for port %u: %s.\n", port, strerror(-retval)); 309 return retval; 310 } 311 if (rx_ring_size > RTE_TEST_RX_DESC_DEFAULT) { 312 RTE_LOG(ERR, VHOST_PORT, "Mbuf pool has an insufficient size " 313 "for Rx queues on port %u.\n", port); 314 return -1; 315 } 316 317 /* Setup the queues. */ 318 rxconf->offloads = port_conf.rxmode.offloads; 319 for (q = 0; q < rx_rings; q ++) { 320 retval = rte_eth_rx_queue_setup(port, q, rx_ring_size, 321 rte_eth_dev_socket_id(port), 322 rxconf, 323 mbuf_pool); 324 if (retval < 0) { 325 RTE_LOG(ERR, VHOST_PORT, 326 "Failed to setup rx queue %u of port %u: %s.\n", 327 q, port, strerror(-retval)); 328 return retval; 329 } 330 } 331 txconf->offloads = port_conf.txmode.offloads; 332 for (q = 0; q < tx_rings; q ++) { 333 retval = rte_eth_tx_queue_setup(port, q, tx_ring_size, 334 rte_eth_dev_socket_id(port), 335 txconf); 336 if (retval < 0) { 337 RTE_LOG(ERR, VHOST_PORT, 338 "Failed to setup tx queue %u of port %u: %s.\n", 339 q, port, strerror(-retval)); 340 return retval; 341 } 342 } 343 344 /* Start the device. */ 345 retval = rte_eth_dev_start(port); 346 if (retval < 0) { 347 RTE_LOG(ERR, VHOST_PORT, "Failed to start port %u: %s\n", 348 port, strerror(-retval)); 349 return retval; 350 } 351 352 if (promiscuous) { 353 retval = rte_eth_promiscuous_enable(port); 354 if (retval != 0) { 355 RTE_LOG(ERR, VHOST_PORT, 356 "Failed to enable promiscuous mode on port %u: %s\n", 357 port, rte_strerror(-retval)); 358 return retval; 359 } 360 } 361 362 retval = rte_eth_macaddr_get(port, &vmdq_ports_eth_addr[port]); 363 if (retval < 0) { 364 RTE_LOG(ERR, VHOST_PORT, 365 "Failed to get MAC address on port %u: %s\n", 366 port, rte_strerror(-retval)); 367 return retval; 368 } 369 370 RTE_LOG(INFO, VHOST_PORT, "Max virtio devices supported: %u\n", num_devices); 371 RTE_LOG(INFO, VHOST_PORT, "Port %u MAC: %02"PRIx8" %02"PRIx8" %02"PRIx8 372 " %02"PRIx8" %02"PRIx8" %02"PRIx8"\n", 373 port, RTE_ETHER_ADDR_BYTES(&vmdq_ports_eth_addr[port])); 374 375 return 0; 376 } 377 378 /* 379 * Set socket file path. 380 */ 381 static int 382 us_vhost_parse_socket_path(const char *q_arg) 383 { 384 char *old; 385 386 /* parse number string */ 387 if (strnlen(q_arg, PATH_MAX) == PATH_MAX) 388 return -1; 389 390 old = socket_files; 391 socket_files = realloc(socket_files, PATH_MAX * (nb_sockets + 1)); 392 if (socket_files == NULL) { 393 free(old); 394 return -1; 395 } 396 397 strlcpy(socket_files + nb_sockets * PATH_MAX, q_arg, PATH_MAX); 398 nb_sockets++; 399 400 return 0; 401 } 402 403 /* 404 * Parse the portmask provided at run time. 405 */ 406 static int 407 parse_portmask(const char *portmask) 408 { 409 char *end = NULL; 410 unsigned long pm; 411 412 errno = 0; 413 414 /* parse hexadecimal string */ 415 pm = strtoul(portmask, &end, 16); 416 if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0') || (errno != 0)) 417 return 0; 418 419 return pm; 420 421 } 422 423 /* 424 * Parse num options at run time. 425 */ 426 static int 427 parse_num_opt(const char *q_arg, uint32_t max_valid_value) 428 { 429 char *end = NULL; 430 unsigned long num; 431 432 errno = 0; 433 434 /* parse unsigned int string */ 435 num = strtoul(q_arg, &end, 10); 436 if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0') || (errno != 0)) 437 return -1; 438 439 if (num > max_valid_value) 440 return -1; 441 442 return num; 443 444 } 445 446 /* 447 * Display usage 448 */ 449 static void 450 us_vhost_usage(const char *prgname) 451 { 452 RTE_LOG(INFO, VHOST_CONFIG, "%s [EAL options] -- -p PORTMASK\n" 453 " --vm2vm [0|1|2]\n" 454 " --rx_retry [0|1] --mergeable [0|1] --stats [0-N]\n" 455 " --socket-file <path>\n" 456 " --nb-devices ND\n" 457 " -p PORTMASK: Set mask for ports to be used by application\n" 458 " --vm2vm [0|1|2]: disable/software(default)/hardware vm2vm comms\n" 459 " --rx-retry [0|1]: disable/enable(default) retries on rx. Enable retry if destintation queue is full\n" 460 " --rx-retry-delay [0-N]: timeout(in usecond) between retries on RX. This makes effect only if retries on rx enabled\n" 461 " --rx-retry-num [0-N]: the number of retries on rx. This makes effect only if retries on rx enabled\n" 462 " --mergeable [0|1]: disable(default)/enable RX mergeable buffers\n" 463 " --stats [0-N]: 0: Disable stats, N: Time in seconds to print stats\n" 464 " --socket-file: The path of the socket file.\n" 465 " --tx-csum [0|1] disable/enable TX checksum offload.\n" 466 " --tso [0|1] disable/enable TCP segment offload.\n" 467 " --client register a vhost-user socket as client mode.\n" 468 " --dma-type register dma type for your vhost async driver. For example \"ioat\" for now.\n" 469 " --dmas register dma channel for specific vhost device.\n", 470 prgname); 471 } 472 473 enum { 474 #define OPT_VM2VM "vm2vm" 475 OPT_VM2VM_NUM = 256, 476 #define OPT_RX_RETRY "rx-retry" 477 OPT_RX_RETRY_NUM, 478 #define OPT_RX_RETRY_DELAY "rx-retry-delay" 479 OPT_RX_RETRY_DELAY_NUM, 480 #define OPT_RX_RETRY_NUMB "rx-retry-num" 481 OPT_RX_RETRY_NUMB_NUM, 482 #define OPT_MERGEABLE "mergeable" 483 OPT_MERGEABLE_NUM, 484 #define OPT_STATS "stats" 485 OPT_STATS_NUM, 486 #define OPT_SOCKET_FILE "socket-file" 487 OPT_SOCKET_FILE_NUM, 488 #define OPT_TX_CSUM "tx-csum" 489 OPT_TX_CSUM_NUM, 490 #define OPT_TSO "tso" 491 OPT_TSO_NUM, 492 #define OPT_CLIENT "client" 493 OPT_CLIENT_NUM, 494 #define OPT_BUILTIN_NET_DRIVER "builtin-net-driver" 495 OPT_BUILTIN_NET_DRIVER_NUM, 496 #define OPT_DMA_TYPE "dma-type" 497 OPT_DMA_TYPE_NUM, 498 #define OPT_DMAS "dmas" 499 OPT_DMAS_NUM, 500 }; 501 502 /* 503 * Parse the arguments given in the command line of the application. 504 */ 505 static int 506 us_vhost_parse_args(int argc, char **argv) 507 { 508 int opt, ret; 509 int option_index; 510 unsigned i; 511 const char *prgname = argv[0]; 512 static struct option long_option[] = { 513 {OPT_VM2VM, required_argument, 514 NULL, OPT_VM2VM_NUM}, 515 {OPT_RX_RETRY, required_argument, 516 NULL, OPT_RX_RETRY_NUM}, 517 {OPT_RX_RETRY_DELAY, required_argument, 518 NULL, OPT_RX_RETRY_DELAY_NUM}, 519 {OPT_RX_RETRY_NUMB, required_argument, 520 NULL, OPT_RX_RETRY_NUMB_NUM}, 521 {OPT_MERGEABLE, required_argument, 522 NULL, OPT_MERGEABLE_NUM}, 523 {OPT_STATS, required_argument, 524 NULL, OPT_STATS_NUM}, 525 {OPT_SOCKET_FILE, required_argument, 526 NULL, OPT_SOCKET_FILE_NUM}, 527 {OPT_TX_CSUM, required_argument, 528 NULL, OPT_TX_CSUM_NUM}, 529 {OPT_TSO, required_argument, 530 NULL, OPT_TSO_NUM}, 531 {OPT_CLIENT, no_argument, 532 NULL, OPT_CLIENT_NUM}, 533 {OPT_BUILTIN_NET_DRIVER, no_argument, 534 NULL, OPT_BUILTIN_NET_DRIVER_NUM}, 535 {OPT_DMA_TYPE, required_argument, 536 NULL, OPT_DMA_TYPE_NUM}, 537 {OPT_DMAS, required_argument, 538 NULL, OPT_DMAS_NUM}, 539 {NULL, 0, 0, 0}, 540 }; 541 542 /* Parse command line */ 543 while ((opt = getopt_long(argc, argv, "p:P", 544 long_option, &option_index)) != EOF) { 545 switch (opt) { 546 /* Portmask */ 547 case 'p': 548 enabled_port_mask = parse_portmask(optarg); 549 if (enabled_port_mask == 0) { 550 RTE_LOG(INFO, VHOST_CONFIG, "Invalid portmask\n"); 551 us_vhost_usage(prgname); 552 return -1; 553 } 554 break; 555 556 case 'P': 557 promiscuous = 1; 558 vmdq_conf_default.rx_adv_conf.vmdq_rx_conf.rx_mode = 559 ETH_VMDQ_ACCEPT_BROADCAST | 560 ETH_VMDQ_ACCEPT_MULTICAST; 561 break; 562 563 case OPT_VM2VM_NUM: 564 ret = parse_num_opt(optarg, (VM2VM_LAST - 1)); 565 if (ret == -1) { 566 RTE_LOG(INFO, VHOST_CONFIG, 567 "Invalid argument for " 568 "vm2vm [0|1|2]\n"); 569 us_vhost_usage(prgname); 570 return -1; 571 } 572 vm2vm_mode = (vm2vm_type)ret; 573 break; 574 575 case OPT_RX_RETRY_NUM: 576 ret = parse_num_opt(optarg, 1); 577 if (ret == -1) { 578 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry [0|1]\n"); 579 us_vhost_usage(prgname); 580 return -1; 581 } 582 enable_retry = ret; 583 break; 584 585 case OPT_TX_CSUM_NUM: 586 ret = parse_num_opt(optarg, 1); 587 if (ret == -1) { 588 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tx-csum [0|1]\n"); 589 us_vhost_usage(prgname); 590 return -1; 591 } 592 enable_tx_csum = ret; 593 break; 594 595 case OPT_TSO_NUM: 596 ret = parse_num_opt(optarg, 1); 597 if (ret == -1) { 598 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tso [0|1]\n"); 599 us_vhost_usage(prgname); 600 return -1; 601 } 602 enable_tso = ret; 603 break; 604 605 case OPT_RX_RETRY_DELAY_NUM: 606 ret = parse_num_opt(optarg, INT32_MAX); 607 if (ret == -1) { 608 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-delay [0-N]\n"); 609 us_vhost_usage(prgname); 610 return -1; 611 } 612 burst_rx_delay_time = ret; 613 break; 614 615 case OPT_RX_RETRY_NUMB_NUM: 616 ret = parse_num_opt(optarg, INT32_MAX); 617 if (ret == -1) { 618 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-num [0-N]\n"); 619 us_vhost_usage(prgname); 620 return -1; 621 } 622 burst_rx_retry_num = ret; 623 break; 624 625 case OPT_MERGEABLE_NUM: 626 ret = parse_num_opt(optarg, 1); 627 if (ret == -1) { 628 RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for mergeable [0|1]\n"); 629 us_vhost_usage(prgname); 630 return -1; 631 } 632 mergeable = !!ret; 633 if (ret) { 634 vmdq_conf_default.rxmode.offloads |= 635 DEV_RX_OFFLOAD_JUMBO_FRAME; 636 vmdq_conf_default.rxmode.max_rx_pkt_len 637 = JUMBO_FRAME_MAX_SIZE; 638 } 639 break; 640 641 case OPT_STATS_NUM: 642 ret = parse_num_opt(optarg, INT32_MAX); 643 if (ret == -1) { 644 RTE_LOG(INFO, VHOST_CONFIG, 645 "Invalid argument for stats [0..N]\n"); 646 us_vhost_usage(prgname); 647 return -1; 648 } 649 enable_stats = ret; 650 break; 651 652 /* Set socket file path. */ 653 case OPT_SOCKET_FILE_NUM: 654 if (us_vhost_parse_socket_path(optarg) == -1) { 655 RTE_LOG(INFO, VHOST_CONFIG, 656 "Invalid argument for socket name (Max %d characters)\n", 657 PATH_MAX); 658 us_vhost_usage(prgname); 659 return -1; 660 } 661 break; 662 663 case OPT_DMA_TYPE_NUM: 664 dma_type = optarg; 665 break; 666 667 case OPT_DMAS_NUM: 668 if (open_dma(optarg) == -1) { 669 RTE_LOG(INFO, VHOST_CONFIG, 670 "Wrong DMA args\n"); 671 us_vhost_usage(prgname); 672 return -1; 673 } 674 async_vhost_driver = 1; 675 break; 676 677 case OPT_CLIENT_NUM: 678 client_mode = 1; 679 break; 680 681 case OPT_BUILTIN_NET_DRIVER_NUM: 682 builtin_net_driver = 1; 683 break; 684 685 /* Invalid option - print options. */ 686 default: 687 us_vhost_usage(prgname); 688 return -1; 689 } 690 } 691 692 for (i = 0; i < RTE_MAX_ETHPORTS; i++) { 693 if (enabled_port_mask & (1 << i)) 694 ports[num_ports++] = i; 695 } 696 697 if ((num_ports == 0) || (num_ports > MAX_SUP_PORTS)) { 698 RTE_LOG(INFO, VHOST_PORT, "Current enabled port number is %u," 699 "but only %u port can be enabled\n",num_ports, MAX_SUP_PORTS); 700 return -1; 701 } 702 703 return 0; 704 } 705 706 /* 707 * Update the global var NUM_PORTS and array PORTS according to system ports number 708 * and return valid ports number 709 */ 710 static unsigned check_ports_num(unsigned nb_ports) 711 { 712 unsigned valid_num_ports = num_ports; 713 unsigned portid; 714 715 if (num_ports > nb_ports) { 716 RTE_LOG(INFO, VHOST_PORT, "\nSpecified port number(%u) exceeds total system port number(%u)\n", 717 num_ports, nb_ports); 718 num_ports = nb_ports; 719 } 720 721 for (portid = 0; portid < num_ports; portid ++) { 722 if (!rte_eth_dev_is_valid_port(ports[portid])) { 723 RTE_LOG(INFO, VHOST_PORT, 724 "\nSpecified port ID(%u) is not valid\n", 725 ports[portid]); 726 ports[portid] = INVALID_PORT_ID; 727 valid_num_ports--; 728 } 729 } 730 return valid_num_ports; 731 } 732 733 static __rte_always_inline struct vhost_dev * 734 find_vhost_dev(struct rte_ether_addr *mac) 735 { 736 struct vhost_dev *vdev; 737 738 TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) { 739 if (vdev->ready == DEVICE_RX && 740 rte_is_same_ether_addr(mac, &vdev->mac_address)) 741 return vdev; 742 } 743 744 return NULL; 745 } 746 747 /* 748 * This function learns the MAC address of the device and registers this along with a 749 * vlan tag to a VMDQ. 750 */ 751 static int 752 link_vmdq(struct vhost_dev *vdev, struct rte_mbuf *m) 753 { 754 struct rte_ether_hdr *pkt_hdr; 755 int i, ret; 756 757 /* Learn MAC address of guest device from packet */ 758 pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *); 759 760 if (find_vhost_dev(&pkt_hdr->s_addr)) { 761 RTE_LOG(ERR, VHOST_DATA, 762 "(%d) device is using a registered MAC!\n", 763 vdev->vid); 764 return -1; 765 } 766 767 for (i = 0; i < RTE_ETHER_ADDR_LEN; i++) 768 vdev->mac_address.addr_bytes[i] = pkt_hdr->s_addr.addr_bytes[i]; 769 770 /* vlan_tag currently uses the device_id. */ 771 vdev->vlan_tag = vlan_tags[vdev->vid]; 772 773 /* Print out VMDQ registration info. */ 774 RTE_LOG(INFO, VHOST_DATA, 775 "(%d) mac " RTE_ETHER_ADDR_PRT_FMT " and vlan %d registered\n", 776 vdev->vid, RTE_ETHER_ADDR_BYTES(&vdev->mac_address), 777 vdev->vlan_tag); 778 779 /* Register the MAC address. */ 780 ret = rte_eth_dev_mac_addr_add(ports[0], &vdev->mac_address, 781 (uint32_t)vdev->vid + vmdq_pool_base); 782 if (ret) 783 RTE_LOG(ERR, VHOST_DATA, 784 "(%d) failed to add device MAC address to VMDQ\n", 785 vdev->vid); 786 787 rte_eth_dev_set_vlan_strip_on_queue(ports[0], vdev->vmdq_rx_q, 1); 788 789 /* Set device as ready for RX. */ 790 vdev->ready = DEVICE_RX; 791 792 return 0; 793 } 794 795 /* 796 * Removes MAC address and vlan tag from VMDQ. Ensures that nothing is adding buffers to the RX 797 * queue before disabling RX on the device. 798 */ 799 static inline void 800 unlink_vmdq(struct vhost_dev *vdev) 801 { 802 unsigned i = 0; 803 unsigned rx_count; 804 struct rte_mbuf *pkts_burst[MAX_PKT_BURST]; 805 806 if (vdev->ready == DEVICE_RX) { 807 /*clear MAC and VLAN settings*/ 808 rte_eth_dev_mac_addr_remove(ports[0], &vdev->mac_address); 809 for (i = 0; i < 6; i++) 810 vdev->mac_address.addr_bytes[i] = 0; 811 812 vdev->vlan_tag = 0; 813 814 /*Clear out the receive buffers*/ 815 rx_count = rte_eth_rx_burst(ports[0], 816 (uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST); 817 818 while (rx_count) { 819 for (i = 0; i < rx_count; i++) 820 rte_pktmbuf_free(pkts_burst[i]); 821 822 rx_count = rte_eth_rx_burst(ports[0], 823 (uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST); 824 } 825 826 vdev->ready = DEVICE_MAC_LEARNING; 827 } 828 } 829 830 static inline void 831 free_pkts(struct rte_mbuf **pkts, uint16_t n) 832 { 833 while (n--) 834 rte_pktmbuf_free(pkts[n]); 835 } 836 837 static __rte_always_inline void 838 complete_async_pkts(struct vhost_dev *vdev) 839 { 840 struct rte_mbuf *p_cpl[MAX_PKT_BURST]; 841 uint16_t complete_count; 842 843 complete_count = rte_vhost_poll_enqueue_completed(vdev->vid, 844 VIRTIO_RXQ, p_cpl, MAX_PKT_BURST); 845 if (complete_count) { 846 free_pkts(p_cpl, complete_count); 847 __atomic_sub_fetch(&vdev->pkts_inflight, complete_count, __ATOMIC_SEQ_CST); 848 } 849 850 } 851 852 static __rte_always_inline void 853 sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev, 854 struct rte_mbuf *m) 855 { 856 uint16_t ret; 857 858 if (builtin_net_driver) { 859 ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1); 860 } else { 861 ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1); 862 } 863 864 if (enable_stats) { 865 __atomic_add_fetch(&dst_vdev->stats.rx_total_atomic, 1, 866 __ATOMIC_SEQ_CST); 867 __atomic_add_fetch(&dst_vdev->stats.rx_atomic, ret, 868 __ATOMIC_SEQ_CST); 869 src_vdev->stats.tx_total++; 870 src_vdev->stats.tx += ret; 871 } 872 } 873 874 static __rte_always_inline void 875 drain_vhost(struct vhost_dev *vdev) 876 { 877 uint16_t ret; 878 uint32_t buff_idx = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid; 879 uint16_t nr_xmit = vhost_txbuff[buff_idx]->len; 880 struct rte_mbuf **m = vhost_txbuff[buff_idx]->m_table; 881 882 if (builtin_net_driver) { 883 ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit); 884 } else if (async_vhost_driver) { 885 uint32_t cpu_cpl_nr = 0; 886 uint16_t enqueue_fail = 0; 887 struct rte_mbuf *m_cpu_cpl[nr_xmit]; 888 889 complete_async_pkts(vdev); 890 ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ, 891 m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr); 892 __atomic_add_fetch(&vdev->pkts_inflight, ret - cpu_cpl_nr, __ATOMIC_SEQ_CST); 893 894 if (cpu_cpl_nr) 895 free_pkts(m_cpu_cpl, cpu_cpl_nr); 896 897 enqueue_fail = nr_xmit - ret; 898 if (enqueue_fail) 899 free_pkts(&m[ret], nr_xmit - ret); 900 } else { 901 ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ, 902 m, nr_xmit); 903 } 904 905 if (enable_stats) { 906 __atomic_add_fetch(&vdev->stats.rx_total_atomic, nr_xmit, 907 __ATOMIC_SEQ_CST); 908 __atomic_add_fetch(&vdev->stats.rx_atomic, ret, 909 __ATOMIC_SEQ_CST); 910 } 911 912 if (!async_vhost_driver) 913 free_pkts(m, nr_xmit); 914 } 915 916 static __rte_always_inline void 917 drain_vhost_table(void) 918 { 919 uint16_t lcore_id = rte_lcore_id(); 920 struct vhost_bufftable *vhost_txq; 921 struct vhost_dev *vdev; 922 uint64_t cur_tsc; 923 924 TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) { 925 vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE 926 + vdev->vid]; 927 928 cur_tsc = rte_rdtsc(); 929 if (unlikely(cur_tsc - vhost_txq->pre_tsc 930 > MBUF_TABLE_DRAIN_TSC)) { 931 RTE_LOG_DP(DEBUG, VHOST_DATA, 932 "Vhost TX queue drained after timeout with burst size %u\n", 933 vhost_txq->len); 934 drain_vhost(vdev); 935 vhost_txq->len = 0; 936 vhost_txq->pre_tsc = cur_tsc; 937 } 938 } 939 } 940 941 /* 942 * Check if the packet destination MAC address is for a local device. If so then put 943 * the packet on that devices RX queue. If not then return. 944 */ 945 static __rte_always_inline int 946 virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m) 947 { 948 struct rte_ether_hdr *pkt_hdr; 949 struct vhost_dev *dst_vdev; 950 struct vhost_bufftable *vhost_txq; 951 uint16_t lcore_id = rte_lcore_id(); 952 pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *); 953 954 dst_vdev = find_vhost_dev(&pkt_hdr->d_addr); 955 if (!dst_vdev) 956 return -1; 957 958 if (vdev->vid == dst_vdev->vid) { 959 RTE_LOG_DP(DEBUG, VHOST_DATA, 960 "(%d) TX: src and dst MAC is same. Dropping packet.\n", 961 vdev->vid); 962 return 0; 963 } 964 965 RTE_LOG_DP(DEBUG, VHOST_DATA, 966 "(%d) TX: MAC address is local\n", dst_vdev->vid); 967 968 if (unlikely(dst_vdev->remove)) { 969 RTE_LOG_DP(DEBUG, VHOST_DATA, 970 "(%d) device is marked for removal\n", dst_vdev->vid); 971 return 0; 972 } 973 974 vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE + dst_vdev->vid]; 975 vhost_txq->m_table[vhost_txq->len++] = m; 976 977 if (enable_stats) { 978 vdev->stats.tx_total++; 979 vdev->stats.tx++; 980 } 981 982 if (unlikely(vhost_txq->len == MAX_PKT_BURST)) { 983 drain_vhost(dst_vdev); 984 vhost_txq->len = 0; 985 vhost_txq->pre_tsc = rte_rdtsc(); 986 } 987 return 0; 988 } 989 990 /* 991 * Check if the destination MAC of a packet is one local VM, 992 * and get its vlan tag, and offset if it is. 993 */ 994 static __rte_always_inline int 995 find_local_dest(struct vhost_dev *vdev, struct rte_mbuf *m, 996 uint32_t *offset, uint16_t *vlan_tag) 997 { 998 struct vhost_dev *dst_vdev; 999 struct rte_ether_hdr *pkt_hdr = 1000 rte_pktmbuf_mtod(m, struct rte_ether_hdr *); 1001 1002 dst_vdev = find_vhost_dev(&pkt_hdr->d_addr); 1003 if (!dst_vdev) 1004 return 0; 1005 1006 if (vdev->vid == dst_vdev->vid) { 1007 RTE_LOG_DP(DEBUG, VHOST_DATA, 1008 "(%d) TX: src and dst MAC is same. Dropping packet.\n", 1009 vdev->vid); 1010 return -1; 1011 } 1012 1013 /* 1014 * HW vlan strip will reduce the packet length 1015 * by minus length of vlan tag, so need restore 1016 * the packet length by plus it. 1017 */ 1018 *offset = VLAN_HLEN; 1019 *vlan_tag = vlan_tags[vdev->vid]; 1020 1021 RTE_LOG_DP(DEBUG, VHOST_DATA, 1022 "(%d) TX: pkt to local VM device id: (%d), vlan tag: %u.\n", 1023 vdev->vid, dst_vdev->vid, *vlan_tag); 1024 1025 return 0; 1026 } 1027 1028 static void virtio_tx_offload(struct rte_mbuf *m) 1029 { 1030 struct rte_net_hdr_lens hdr_lens; 1031 struct rte_ipv4_hdr *ipv4_hdr; 1032 struct rte_tcp_hdr *tcp_hdr; 1033 uint32_t ptype; 1034 void *l3_hdr; 1035 1036 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK); 1037 m->l2_len = hdr_lens.l2_len; 1038 m->l3_len = hdr_lens.l3_len; 1039 m->l4_len = hdr_lens.l4_len; 1040 1041 l3_hdr = rte_pktmbuf_mtod_offset(m, void *, m->l2_len); 1042 tcp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_tcp_hdr *, 1043 m->l2_len + m->l3_len); 1044 1045 m->ol_flags |= PKT_TX_TCP_SEG; 1046 if ((ptype & RTE_PTYPE_L3_MASK) == RTE_PTYPE_L3_IPV4) { 1047 m->ol_flags |= PKT_TX_IPV4; 1048 m->ol_flags |= PKT_TX_IP_CKSUM; 1049 ipv4_hdr = l3_hdr; 1050 ipv4_hdr->hdr_checksum = 0; 1051 tcp_hdr->cksum = rte_ipv4_phdr_cksum(l3_hdr, m->ol_flags); 1052 } else { /* assume ethertype == RTE_ETHER_TYPE_IPV6 */ 1053 m->ol_flags |= PKT_TX_IPV6; 1054 tcp_hdr->cksum = rte_ipv6_phdr_cksum(l3_hdr, m->ol_flags); 1055 } 1056 } 1057 1058 static __rte_always_inline void 1059 do_drain_mbuf_table(struct mbuf_table *tx_q) 1060 { 1061 uint16_t count; 1062 1063 count = rte_eth_tx_burst(ports[0], tx_q->txq_id, 1064 tx_q->m_table, tx_q->len); 1065 if (unlikely(count < tx_q->len)) 1066 free_pkts(&tx_q->m_table[count], tx_q->len - count); 1067 1068 tx_q->len = 0; 1069 } 1070 1071 /* 1072 * This function routes the TX packet to the correct interface. This 1073 * may be a local device or the physical port. 1074 */ 1075 static __rte_always_inline void 1076 virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag) 1077 { 1078 struct mbuf_table *tx_q; 1079 unsigned offset = 0; 1080 const uint16_t lcore_id = rte_lcore_id(); 1081 struct rte_ether_hdr *nh; 1082 1083 1084 nh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *); 1085 if (unlikely(rte_is_broadcast_ether_addr(&nh->d_addr))) { 1086 struct vhost_dev *vdev2; 1087 1088 TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) { 1089 if (vdev2 != vdev) 1090 sync_virtio_xmit(vdev2, vdev, m); 1091 } 1092 goto queue2nic; 1093 } 1094 1095 /*check if destination is local VM*/ 1096 if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0)) 1097 return; 1098 1099 if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) { 1100 if (unlikely(find_local_dest(vdev, m, &offset, 1101 &vlan_tag) != 0)) { 1102 rte_pktmbuf_free(m); 1103 return; 1104 } 1105 } 1106 1107 RTE_LOG_DP(DEBUG, VHOST_DATA, 1108 "(%d) TX: MAC address is external\n", vdev->vid); 1109 1110 queue2nic: 1111 1112 /*Add packet to the port tx queue*/ 1113 tx_q = &lcore_tx_queue[lcore_id]; 1114 1115 nh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *); 1116 if (unlikely(nh->ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN))) { 1117 /* Guest has inserted the vlan tag. */ 1118 struct rte_vlan_hdr *vh = (struct rte_vlan_hdr *) (nh + 1); 1119 uint16_t vlan_tag_be = rte_cpu_to_be_16(vlan_tag); 1120 if ((vm2vm_mode == VM2VM_HARDWARE) && 1121 (vh->vlan_tci != vlan_tag_be)) 1122 vh->vlan_tci = vlan_tag_be; 1123 } else { 1124 m->ol_flags |= PKT_TX_VLAN_PKT; 1125 1126 /* 1127 * Find the right seg to adjust the data len when offset is 1128 * bigger than tail room size. 1129 */ 1130 if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) { 1131 if (likely(offset <= rte_pktmbuf_tailroom(m))) 1132 m->data_len += offset; 1133 else { 1134 struct rte_mbuf *seg = m; 1135 1136 while ((seg->next != NULL) && 1137 (offset > rte_pktmbuf_tailroom(seg))) 1138 seg = seg->next; 1139 1140 seg->data_len += offset; 1141 } 1142 m->pkt_len += offset; 1143 } 1144 1145 m->vlan_tci = vlan_tag; 1146 } 1147 1148 if (m->ol_flags & PKT_RX_LRO) 1149 virtio_tx_offload(m); 1150 1151 tx_q->m_table[tx_q->len++] = m; 1152 if (enable_stats) { 1153 vdev->stats.tx_total++; 1154 vdev->stats.tx++; 1155 } 1156 1157 if (unlikely(tx_q->len == MAX_PKT_BURST)) 1158 do_drain_mbuf_table(tx_q); 1159 } 1160 1161 1162 static __rte_always_inline void 1163 drain_mbuf_table(struct mbuf_table *tx_q) 1164 { 1165 static uint64_t prev_tsc; 1166 uint64_t cur_tsc; 1167 1168 if (tx_q->len == 0) 1169 return; 1170 1171 cur_tsc = rte_rdtsc(); 1172 if (unlikely(cur_tsc - prev_tsc > MBUF_TABLE_DRAIN_TSC)) { 1173 prev_tsc = cur_tsc; 1174 1175 RTE_LOG_DP(DEBUG, VHOST_DATA, 1176 "TX queue drained after timeout with burst size %u\n", 1177 tx_q->len); 1178 do_drain_mbuf_table(tx_q); 1179 } 1180 } 1181 1182 static __rte_always_inline void 1183 drain_eth_rx(struct vhost_dev *vdev) 1184 { 1185 uint16_t rx_count, enqueue_count; 1186 struct rte_mbuf *pkts[MAX_PKT_BURST]; 1187 1188 rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q, 1189 pkts, MAX_PKT_BURST); 1190 1191 if (!rx_count) 1192 return; 1193 1194 /* 1195 * When "enable_retry" is set, here we wait and retry when there 1196 * is no enough free slots in the queue to hold @rx_count packets, 1197 * to diminish packet loss. 1198 */ 1199 if (enable_retry && 1200 unlikely(rx_count > rte_vhost_avail_entries(vdev->vid, 1201 VIRTIO_RXQ))) { 1202 uint32_t retry; 1203 1204 for (retry = 0; retry < burst_rx_retry_num; retry++) { 1205 rte_delay_us(burst_rx_delay_time); 1206 if (rx_count <= rte_vhost_avail_entries(vdev->vid, 1207 VIRTIO_RXQ)) 1208 break; 1209 } 1210 } 1211 1212 if (builtin_net_driver) { 1213 enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ, 1214 pkts, rx_count); 1215 } else if (async_vhost_driver) { 1216 uint32_t cpu_cpl_nr = 0; 1217 uint16_t enqueue_fail = 0; 1218 struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST]; 1219 1220 complete_async_pkts(vdev); 1221 enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid, 1222 VIRTIO_RXQ, pkts, rx_count, 1223 m_cpu_cpl, &cpu_cpl_nr); 1224 __atomic_add_fetch(&vdev->pkts_inflight, enqueue_count - cpu_cpl_nr, 1225 __ATOMIC_SEQ_CST); 1226 1227 if (cpu_cpl_nr) 1228 free_pkts(m_cpu_cpl, cpu_cpl_nr); 1229 1230 enqueue_fail = rx_count - enqueue_count; 1231 if (enqueue_fail) 1232 free_pkts(&pkts[enqueue_count], enqueue_fail); 1233 1234 } else { 1235 enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ, 1236 pkts, rx_count); 1237 } 1238 1239 if (enable_stats) { 1240 __atomic_add_fetch(&vdev->stats.rx_total_atomic, rx_count, 1241 __ATOMIC_SEQ_CST); 1242 __atomic_add_fetch(&vdev->stats.rx_atomic, enqueue_count, 1243 __ATOMIC_SEQ_CST); 1244 } 1245 1246 if (!async_vhost_driver) 1247 free_pkts(pkts, rx_count); 1248 } 1249 1250 static __rte_always_inline void 1251 drain_virtio_tx(struct vhost_dev *vdev) 1252 { 1253 struct rte_mbuf *pkts[MAX_PKT_BURST]; 1254 uint16_t count; 1255 uint16_t i; 1256 1257 if (builtin_net_driver) { 1258 count = vs_dequeue_pkts(vdev, VIRTIO_TXQ, mbuf_pool, 1259 pkts, MAX_PKT_BURST); 1260 } else { 1261 count = rte_vhost_dequeue_burst(vdev->vid, VIRTIO_TXQ, 1262 mbuf_pool, pkts, MAX_PKT_BURST); 1263 } 1264 1265 /* setup VMDq for the first packet */ 1266 if (unlikely(vdev->ready == DEVICE_MAC_LEARNING) && count) { 1267 if (vdev->remove || link_vmdq(vdev, pkts[0]) == -1) 1268 free_pkts(pkts, count); 1269 } 1270 1271 for (i = 0; i < count; ++i) 1272 virtio_tx_route(vdev, pkts[i], vlan_tags[vdev->vid]); 1273 } 1274 1275 /* 1276 * Main function of vhost-switch. It basically does: 1277 * 1278 * for each vhost device { 1279 * - drain_eth_rx() 1280 * 1281 * Which drains the host eth Rx queue linked to the vhost device, 1282 * and deliver all of them to guest virito Rx ring associated with 1283 * this vhost device. 1284 * 1285 * - drain_virtio_tx() 1286 * 1287 * Which drains the guest virtio Tx queue and deliver all of them 1288 * to the target, which could be another vhost device, or the 1289 * physical eth dev. The route is done in function "virtio_tx_route". 1290 * } 1291 */ 1292 static int 1293 switch_worker(void *arg __rte_unused) 1294 { 1295 unsigned i; 1296 unsigned lcore_id = rte_lcore_id(); 1297 struct vhost_dev *vdev; 1298 struct mbuf_table *tx_q; 1299 1300 RTE_LOG(INFO, VHOST_DATA, "Procesing on Core %u started\n", lcore_id); 1301 1302 tx_q = &lcore_tx_queue[lcore_id]; 1303 for (i = 0; i < rte_lcore_count(); i++) { 1304 if (lcore_ids[i] == lcore_id) { 1305 tx_q->txq_id = i; 1306 break; 1307 } 1308 } 1309 1310 while(1) { 1311 drain_mbuf_table(tx_q); 1312 drain_vhost_table(); 1313 /* 1314 * Inform the configuration core that we have exited the 1315 * linked list and that no devices are in use if requested. 1316 */ 1317 if (lcore_info[lcore_id].dev_removal_flag == REQUEST_DEV_REMOVAL) 1318 lcore_info[lcore_id].dev_removal_flag = ACK_DEV_REMOVAL; 1319 1320 /* 1321 * Process vhost devices 1322 */ 1323 TAILQ_FOREACH(vdev, &lcore_info[lcore_id].vdev_list, 1324 lcore_vdev_entry) { 1325 if (unlikely(vdev->remove)) { 1326 unlink_vmdq(vdev); 1327 vdev->ready = DEVICE_SAFE_REMOVE; 1328 continue; 1329 } 1330 1331 if (likely(vdev->ready == DEVICE_RX)) 1332 drain_eth_rx(vdev); 1333 1334 if (likely(!vdev->remove)) 1335 drain_virtio_tx(vdev); 1336 } 1337 } 1338 1339 return 0; 1340 } 1341 1342 /* 1343 * Remove a device from the specific data core linked list and from the 1344 * main linked list. Synchonization occurs through the use of the 1345 * lcore dev_removal_flag. Device is made volatile here to avoid re-ordering 1346 * of dev->remove=1 which can cause an infinite loop in the rte_pause loop. 1347 */ 1348 static void 1349 destroy_device(int vid) 1350 { 1351 struct vhost_dev *vdev = NULL; 1352 int lcore; 1353 uint16_t i; 1354 1355 TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) { 1356 if (vdev->vid == vid) 1357 break; 1358 } 1359 if (!vdev) 1360 return; 1361 /*set the remove flag. */ 1362 vdev->remove = 1; 1363 while(vdev->ready != DEVICE_SAFE_REMOVE) { 1364 rte_pause(); 1365 } 1366 1367 for (i = 0; i < RTE_MAX_LCORE; i++) 1368 rte_free(vhost_txbuff[i * MAX_VHOST_DEVICE + vid]); 1369 1370 if (builtin_net_driver) 1371 vs_vhost_net_remove(vdev); 1372 1373 TAILQ_REMOVE(&lcore_info[vdev->coreid].vdev_list, vdev, 1374 lcore_vdev_entry); 1375 TAILQ_REMOVE(&vhost_dev_list, vdev, global_vdev_entry); 1376 1377 1378 /* Set the dev_removal_flag on each lcore. */ 1379 RTE_LCORE_FOREACH_WORKER(lcore) 1380 lcore_info[lcore].dev_removal_flag = REQUEST_DEV_REMOVAL; 1381 1382 /* 1383 * Once each core has set the dev_removal_flag to ACK_DEV_REMOVAL 1384 * we can be sure that they can no longer access the device removed 1385 * from the linked lists and that the devices are no longer in use. 1386 */ 1387 RTE_LCORE_FOREACH_WORKER(lcore) { 1388 while (lcore_info[lcore].dev_removal_flag != ACK_DEV_REMOVAL) 1389 rte_pause(); 1390 } 1391 1392 lcore_info[vdev->coreid].device_num--; 1393 1394 RTE_LOG(INFO, VHOST_DATA, 1395 "(%d) device has been removed from data core\n", 1396 vdev->vid); 1397 1398 if (async_vhost_driver) { 1399 uint16_t n_pkt = 0; 1400 struct rte_mbuf *m_cpl[vdev->pkts_inflight]; 1401 1402 while (vdev->pkts_inflight) { 1403 n_pkt = rte_vhost_clear_queue_thread_unsafe(vid, VIRTIO_RXQ, 1404 m_cpl, vdev->pkts_inflight); 1405 free_pkts(m_cpl, n_pkt); 1406 __atomic_sub_fetch(&vdev->pkts_inflight, n_pkt, __ATOMIC_SEQ_CST); 1407 } 1408 1409 rte_vhost_async_channel_unregister(vid, VIRTIO_RXQ); 1410 } 1411 1412 rte_free(vdev); 1413 } 1414 1415 /* 1416 * A new device is added to a data core. First the device is added to the main linked list 1417 * and then allocated to a specific data core. 1418 */ 1419 static int 1420 new_device(int vid) 1421 { 1422 int lcore, core_add = 0; 1423 uint16_t i; 1424 uint32_t device_num_min = num_devices; 1425 struct vhost_dev *vdev; 1426 vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE); 1427 if (vdev == NULL) { 1428 RTE_LOG(INFO, VHOST_DATA, 1429 "(%d) couldn't allocate memory for vhost dev\n", 1430 vid); 1431 return -1; 1432 } 1433 vdev->vid = vid; 1434 1435 for (i = 0; i < RTE_MAX_LCORE; i++) { 1436 vhost_txbuff[i * MAX_VHOST_DEVICE + vid] 1437 = rte_zmalloc("vhost bufftable", 1438 sizeof(struct vhost_bufftable), 1439 RTE_CACHE_LINE_SIZE); 1440 1441 if (vhost_txbuff[i * MAX_VHOST_DEVICE + vid] == NULL) { 1442 RTE_LOG(INFO, VHOST_DATA, 1443 "(%d) couldn't allocate memory for vhost TX\n", vid); 1444 return -1; 1445 } 1446 } 1447 1448 if (builtin_net_driver) 1449 vs_vhost_net_setup(vdev); 1450 1451 TAILQ_INSERT_TAIL(&vhost_dev_list, vdev, global_vdev_entry); 1452 vdev->vmdq_rx_q = vid * queues_per_pool + vmdq_queue_base; 1453 1454 /*reset ready flag*/ 1455 vdev->ready = DEVICE_MAC_LEARNING; 1456 vdev->remove = 0; 1457 1458 /* Find a suitable lcore to add the device. */ 1459 RTE_LCORE_FOREACH_WORKER(lcore) { 1460 if (lcore_info[lcore].device_num < device_num_min) { 1461 device_num_min = lcore_info[lcore].device_num; 1462 core_add = lcore; 1463 } 1464 } 1465 vdev->coreid = core_add; 1466 1467 TAILQ_INSERT_TAIL(&lcore_info[vdev->coreid].vdev_list, vdev, 1468 lcore_vdev_entry); 1469 lcore_info[vdev->coreid].device_num++; 1470 1471 /* Disable notifications. */ 1472 rte_vhost_enable_guest_notification(vid, VIRTIO_RXQ, 0); 1473 rte_vhost_enable_guest_notification(vid, VIRTIO_TXQ, 0); 1474 1475 RTE_LOG(INFO, VHOST_DATA, 1476 "(%d) device has been added to data core %d\n", 1477 vid, vdev->coreid); 1478 1479 if (async_vhost_driver) { 1480 struct rte_vhost_async_config config = {0}; 1481 struct rte_vhost_async_channel_ops channel_ops; 1482 1483 if (dma_type != NULL && strncmp(dma_type, "ioat", 4) == 0) { 1484 channel_ops.transfer_data = ioat_transfer_data_cb; 1485 channel_ops.check_completed_copies = 1486 ioat_check_completed_copies_cb; 1487 1488 config.features = RTE_VHOST_ASYNC_INORDER; 1489 config.async_threshold = 256; 1490 1491 return rte_vhost_async_channel_register(vid, VIRTIO_RXQ, 1492 config, &channel_ops); 1493 } 1494 } 1495 1496 return 0; 1497 } 1498 1499 static int 1500 vring_state_changed(int vid, uint16_t queue_id, int enable) 1501 { 1502 struct vhost_dev *vdev = NULL; 1503 1504 TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) { 1505 if (vdev->vid == vid) 1506 break; 1507 } 1508 if (!vdev) 1509 return -1; 1510 1511 if (queue_id != VIRTIO_RXQ) 1512 return 0; 1513 1514 if (async_vhost_driver) { 1515 if (!enable) { 1516 uint16_t n_pkt = 0; 1517 struct rte_mbuf *m_cpl[vdev->pkts_inflight]; 1518 1519 while (vdev->pkts_inflight) { 1520 n_pkt = rte_vhost_clear_queue_thread_unsafe(vid, queue_id, 1521 m_cpl, vdev->pkts_inflight); 1522 free_pkts(m_cpl, n_pkt); 1523 __atomic_sub_fetch(&vdev->pkts_inflight, n_pkt, __ATOMIC_SEQ_CST); 1524 } 1525 } 1526 } 1527 1528 return 0; 1529 } 1530 1531 /* 1532 * These callback allow devices to be added to the data core when configuration 1533 * has been fully complete. 1534 */ 1535 static const struct vhost_device_ops virtio_net_device_ops = 1536 { 1537 .new_device = new_device, 1538 .destroy_device = destroy_device, 1539 .vring_state_changed = vring_state_changed, 1540 }; 1541 1542 /* 1543 * This is a thread will wake up after a period to print stats if the user has 1544 * enabled them. 1545 */ 1546 static void * 1547 print_stats(__rte_unused void *arg) 1548 { 1549 struct vhost_dev *vdev; 1550 uint64_t tx_dropped, rx_dropped; 1551 uint64_t tx, tx_total, rx, rx_total; 1552 const char clr[] = { 27, '[', '2', 'J', '\0' }; 1553 const char top_left[] = { 27, '[', '1', ';', '1', 'H','\0' }; 1554 1555 while(1) { 1556 sleep(enable_stats); 1557 1558 /* Clear screen and move to top left */ 1559 printf("%s%s\n", clr, top_left); 1560 printf("Device statistics =================================\n"); 1561 1562 TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) { 1563 tx_total = vdev->stats.tx_total; 1564 tx = vdev->stats.tx; 1565 tx_dropped = tx_total - tx; 1566 1567 rx_total = __atomic_load_n(&vdev->stats.rx_total_atomic, 1568 __ATOMIC_SEQ_CST); 1569 rx = __atomic_load_n(&vdev->stats.rx_atomic, 1570 __ATOMIC_SEQ_CST); 1571 rx_dropped = rx_total - rx; 1572 1573 printf("Statistics for device %d\n" 1574 "-----------------------\n" 1575 "TX total: %" PRIu64 "\n" 1576 "TX dropped: %" PRIu64 "\n" 1577 "TX successful: %" PRIu64 "\n" 1578 "RX total: %" PRIu64 "\n" 1579 "RX dropped: %" PRIu64 "\n" 1580 "RX successful: %" PRIu64 "\n", 1581 vdev->vid, 1582 tx_total, tx_dropped, tx, 1583 rx_total, rx_dropped, rx); 1584 } 1585 1586 printf("===================================================\n"); 1587 1588 fflush(stdout); 1589 } 1590 1591 return NULL; 1592 } 1593 1594 static void 1595 unregister_drivers(int socket_num) 1596 { 1597 int i, ret; 1598 1599 for (i = 0; i < socket_num; i++) { 1600 ret = rte_vhost_driver_unregister(socket_files + i * PATH_MAX); 1601 if (ret != 0) 1602 RTE_LOG(ERR, VHOST_CONFIG, 1603 "Fail to unregister vhost driver for %s.\n", 1604 socket_files + i * PATH_MAX); 1605 } 1606 } 1607 1608 /* When we receive a INT signal, unregister vhost driver */ 1609 static void 1610 sigint_handler(__rte_unused int signum) 1611 { 1612 /* Unregister vhost driver. */ 1613 unregister_drivers(nb_sockets); 1614 1615 exit(0); 1616 } 1617 1618 /* 1619 * While creating an mbuf pool, one key thing is to figure out how 1620 * many mbuf entries is enough for our use. FYI, here are some 1621 * guidelines: 1622 * 1623 * - Each rx queue would reserve @nr_rx_desc mbufs at queue setup stage 1624 * 1625 * - For each switch core (A CPU core does the packet switch), we need 1626 * also make some reservation for receiving the packets from virtio 1627 * Tx queue. How many is enough depends on the usage. It's normally 1628 * a simple calculation like following: 1629 * 1630 * MAX_PKT_BURST * max packet size / mbuf size 1631 * 1632 * So, we definitely need allocate more mbufs when TSO is enabled. 1633 * 1634 * - Similarly, for each switching core, we should serve @nr_rx_desc 1635 * mbufs for receiving the packets from physical NIC device. 1636 * 1637 * - We also need make sure, for each switch core, we have allocated 1638 * enough mbufs to fill up the mbuf cache. 1639 */ 1640 static void 1641 create_mbuf_pool(uint16_t nr_port, uint32_t nr_switch_core, uint32_t mbuf_size, 1642 uint32_t nr_queues, uint32_t nr_rx_desc, uint32_t nr_mbuf_cache) 1643 { 1644 uint32_t nr_mbufs; 1645 uint32_t nr_mbufs_per_core; 1646 uint32_t mtu = 1500; 1647 1648 if (mergeable) 1649 mtu = 9000; 1650 if (enable_tso) 1651 mtu = 64 * 1024; 1652 1653 nr_mbufs_per_core = (mtu + mbuf_size) * MAX_PKT_BURST / 1654 (mbuf_size - RTE_PKTMBUF_HEADROOM); 1655 nr_mbufs_per_core += nr_rx_desc; 1656 nr_mbufs_per_core = RTE_MAX(nr_mbufs_per_core, nr_mbuf_cache); 1657 1658 nr_mbufs = nr_queues * nr_rx_desc; 1659 nr_mbufs += nr_mbufs_per_core * nr_switch_core; 1660 nr_mbufs *= nr_port; 1661 1662 mbuf_pool = rte_pktmbuf_pool_create("MBUF_POOL", nr_mbufs, 1663 nr_mbuf_cache, 0, mbuf_size, 1664 rte_socket_id()); 1665 if (mbuf_pool == NULL) 1666 rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n"); 1667 } 1668 1669 /* 1670 * Main function, does initialisation and calls the per-lcore functions. 1671 */ 1672 int 1673 main(int argc, char *argv[]) 1674 { 1675 unsigned lcore_id, core_id = 0; 1676 unsigned nb_ports, valid_num_ports; 1677 int ret, i; 1678 uint16_t portid; 1679 static pthread_t tid; 1680 uint64_t flags = RTE_VHOST_USER_NET_COMPLIANT_OL_FLAGS; 1681 1682 signal(SIGINT, sigint_handler); 1683 1684 /* init EAL */ 1685 ret = rte_eal_init(argc, argv); 1686 if (ret < 0) 1687 rte_exit(EXIT_FAILURE, "Error with EAL initialization\n"); 1688 argc -= ret; 1689 argv += ret; 1690 1691 /* parse app arguments */ 1692 ret = us_vhost_parse_args(argc, argv); 1693 if (ret < 0) 1694 rte_exit(EXIT_FAILURE, "Invalid argument\n"); 1695 1696 for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) { 1697 TAILQ_INIT(&lcore_info[lcore_id].vdev_list); 1698 1699 if (rte_lcore_is_enabled(lcore_id)) 1700 lcore_ids[core_id++] = lcore_id; 1701 } 1702 1703 if (rte_lcore_count() > RTE_MAX_LCORE) 1704 rte_exit(EXIT_FAILURE,"Not enough cores\n"); 1705 1706 /* Get the number of physical ports. */ 1707 nb_ports = rte_eth_dev_count_avail(); 1708 1709 /* 1710 * Update the global var NUM_PORTS and global array PORTS 1711 * and get value of var VALID_NUM_PORTS according to system ports number 1712 */ 1713 valid_num_ports = check_ports_num(nb_ports); 1714 1715 if ((valid_num_ports == 0) || (valid_num_ports > MAX_SUP_PORTS)) { 1716 RTE_LOG(INFO, VHOST_PORT, "Current enabled port number is %u," 1717 "but only %u port can be enabled\n",num_ports, MAX_SUP_PORTS); 1718 return -1; 1719 } 1720 1721 /* 1722 * FIXME: here we are trying to allocate mbufs big enough for 1723 * @MAX_QUEUES, but the truth is we're never going to use that 1724 * many queues here. We probably should only do allocation for 1725 * those queues we are going to use. 1726 */ 1727 create_mbuf_pool(valid_num_ports, rte_lcore_count() - 1, MBUF_DATA_SIZE, 1728 MAX_QUEUES, RTE_TEST_RX_DESC_DEFAULT, MBUF_CACHE_SIZE); 1729 1730 if (vm2vm_mode == VM2VM_HARDWARE) { 1731 /* Enable VT loop back to let L2 switch to do it. */ 1732 vmdq_conf_default.rx_adv_conf.vmdq_rx_conf.enable_loop_back = 1; 1733 RTE_LOG(DEBUG, VHOST_CONFIG, 1734 "Enable loop back for L2 switch in vmdq.\n"); 1735 } 1736 1737 /* initialize all ports */ 1738 RTE_ETH_FOREACH_DEV(portid) { 1739 /* skip ports that are not enabled */ 1740 if ((enabled_port_mask & (1 << portid)) == 0) { 1741 RTE_LOG(INFO, VHOST_PORT, 1742 "Skipping disabled port %d\n", portid); 1743 continue; 1744 } 1745 if (port_init(portid) != 0) 1746 rte_exit(EXIT_FAILURE, 1747 "Cannot initialize network ports\n"); 1748 } 1749 1750 /* Enable stats if the user option is set. */ 1751 if (enable_stats) { 1752 ret = rte_ctrl_thread_create(&tid, "print-stats", NULL, 1753 print_stats, NULL); 1754 if (ret < 0) 1755 rte_exit(EXIT_FAILURE, 1756 "Cannot create print-stats thread\n"); 1757 } 1758 1759 /* Launch all data cores. */ 1760 RTE_LCORE_FOREACH_WORKER(lcore_id) 1761 rte_eal_remote_launch(switch_worker, NULL, lcore_id); 1762 1763 if (client_mode) 1764 flags |= RTE_VHOST_USER_CLIENT; 1765 1766 /* Register vhost user driver to handle vhost messages. */ 1767 for (i = 0; i < nb_sockets; i++) { 1768 char *file = socket_files + i * PATH_MAX; 1769 1770 if (async_vhost_driver) 1771 flags = flags | RTE_VHOST_USER_ASYNC_COPY; 1772 1773 ret = rte_vhost_driver_register(file, flags); 1774 if (ret != 0) { 1775 unregister_drivers(i); 1776 rte_exit(EXIT_FAILURE, 1777 "vhost driver register failure.\n"); 1778 } 1779 1780 if (builtin_net_driver) 1781 rte_vhost_driver_set_features(file, VIRTIO_NET_FEATURES); 1782 1783 if (mergeable == 0) { 1784 rte_vhost_driver_disable_features(file, 1785 1ULL << VIRTIO_NET_F_MRG_RXBUF); 1786 } 1787 1788 if (enable_tx_csum == 0) { 1789 rte_vhost_driver_disable_features(file, 1790 1ULL << VIRTIO_NET_F_CSUM); 1791 } 1792 1793 if (enable_tso == 0) { 1794 rte_vhost_driver_disable_features(file, 1795 1ULL << VIRTIO_NET_F_HOST_TSO4); 1796 rte_vhost_driver_disable_features(file, 1797 1ULL << VIRTIO_NET_F_HOST_TSO6); 1798 rte_vhost_driver_disable_features(file, 1799 1ULL << VIRTIO_NET_F_GUEST_TSO4); 1800 rte_vhost_driver_disable_features(file, 1801 1ULL << VIRTIO_NET_F_GUEST_TSO6); 1802 } 1803 1804 if (promiscuous) { 1805 rte_vhost_driver_enable_features(file, 1806 1ULL << VIRTIO_NET_F_CTRL_RX); 1807 } 1808 1809 ret = rte_vhost_driver_callback_register(file, 1810 &virtio_net_device_ops); 1811 if (ret != 0) { 1812 rte_exit(EXIT_FAILURE, 1813 "failed to register vhost driver callbacks.\n"); 1814 } 1815 1816 if (rte_vhost_driver_start(file) < 0) { 1817 rte_exit(EXIT_FAILURE, 1818 "failed to start vhost driver.\n"); 1819 } 1820 } 1821 1822 RTE_LCORE_FOREACH_WORKER(lcore_id) 1823 rte_eal_wait_lcore(lcore_id); 1824 1825 /* clean up the EAL */ 1826 rte_eal_cleanup(); 1827 1828 return 0; 1829 } 1830