xref: /dpdk/examples/vhost/main.c (revision ca7036b4af3a82d258cca914e71171434b3d0320)
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(&eth_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,
374 			vmdq_ports_eth_addr[port].addr_bytes[0],
375 			vmdq_ports_eth_addr[port].addr_bytes[1],
376 			vmdq_ports_eth_addr[port].addr_bytes[2],
377 			vmdq_ports_eth_addr[port].addr_bytes[3],
378 			vmdq_ports_eth_addr[port].addr_bytes[4],
379 			vmdq_ports_eth_addr[port].addr_bytes[5]);
380 
381 	return 0;
382 }
383 
384 /*
385  * Set socket file path.
386  */
387 static int
388 us_vhost_parse_socket_path(const char *q_arg)
389 {
390 	char *old;
391 
392 	/* parse number string */
393 	if (strnlen(q_arg, PATH_MAX) == PATH_MAX)
394 		return -1;
395 
396 	old = socket_files;
397 	socket_files = realloc(socket_files, PATH_MAX * (nb_sockets + 1));
398 	if (socket_files == NULL) {
399 		free(old);
400 		return -1;
401 	}
402 
403 	strlcpy(socket_files + nb_sockets * PATH_MAX, q_arg, PATH_MAX);
404 	nb_sockets++;
405 
406 	return 0;
407 }
408 
409 /*
410  * Parse the portmask provided at run time.
411  */
412 static int
413 parse_portmask(const char *portmask)
414 {
415 	char *end = NULL;
416 	unsigned long pm;
417 
418 	errno = 0;
419 
420 	/* parse hexadecimal string */
421 	pm = strtoul(portmask, &end, 16);
422 	if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0') || (errno != 0))
423 		return 0;
424 
425 	return pm;
426 
427 }
428 
429 /*
430  * Parse num options at run time.
431  */
432 static int
433 parse_num_opt(const char *q_arg, uint32_t max_valid_value)
434 {
435 	char *end = NULL;
436 	unsigned long num;
437 
438 	errno = 0;
439 
440 	/* parse unsigned int string */
441 	num = strtoul(q_arg, &end, 10);
442 	if ((q_arg[0] == '\0') || (end == NULL) || (*end != '\0') || (errno != 0))
443 		return -1;
444 
445 	if (num > max_valid_value)
446 		return -1;
447 
448 	return num;
449 
450 }
451 
452 /*
453  * Display usage
454  */
455 static void
456 us_vhost_usage(const char *prgname)
457 {
458 	RTE_LOG(INFO, VHOST_CONFIG, "%s [EAL options] -- -p PORTMASK\n"
459 	"		--vm2vm [0|1|2]\n"
460 	"		--rx_retry [0|1] --mergeable [0|1] --stats [0-N]\n"
461 	"		--socket-file <path>\n"
462 	"		--nb-devices ND\n"
463 	"		-p PORTMASK: Set mask for ports to be used by application\n"
464 	"		--vm2vm [0|1|2]: disable/software(default)/hardware vm2vm comms\n"
465 	"		--rx-retry [0|1]: disable/enable(default) retries on rx. Enable retry if destintation queue is full\n"
466 	"		--rx-retry-delay [0-N]: timeout(in usecond) between retries on RX. This makes effect only if retries on rx enabled\n"
467 	"		--rx-retry-num [0-N]: the number of retries on rx. This makes effect only if retries on rx enabled\n"
468 	"		--mergeable [0|1]: disable(default)/enable RX mergeable buffers\n"
469 	"		--stats [0-N]: 0: Disable stats, N: Time in seconds to print stats\n"
470 	"		--socket-file: The path of the socket file.\n"
471 	"		--tx-csum [0|1] disable/enable TX checksum offload.\n"
472 	"		--tso [0|1] disable/enable TCP segment offload.\n"
473 	"		--client register a vhost-user socket as client mode.\n"
474 	"		--dma-type register dma type for your vhost async driver. For example \"ioat\" for now.\n"
475 	"		--dmas register dma channel for specific vhost device.\n",
476 	       prgname);
477 }
478 
479 enum {
480 #define OPT_VM2VM               "vm2vm"
481 	OPT_VM2VM_NUM = 256,
482 #define OPT_RX_RETRY            "rx-retry"
483 	OPT_RX_RETRY_NUM,
484 #define OPT_RX_RETRY_DELAY      "rx-retry-delay"
485 	OPT_RX_RETRY_DELAY_NUM,
486 #define OPT_RX_RETRY_NUMB       "rx-retry-num"
487 	OPT_RX_RETRY_NUMB_NUM,
488 #define OPT_MERGEABLE           "mergeable"
489 	OPT_MERGEABLE_NUM,
490 #define OPT_STATS               "stats"
491 	OPT_STATS_NUM,
492 #define OPT_SOCKET_FILE         "socket-file"
493 	OPT_SOCKET_FILE_NUM,
494 #define OPT_TX_CSUM             "tx-csum"
495 	OPT_TX_CSUM_NUM,
496 #define OPT_TSO                 "tso"
497 	OPT_TSO_NUM,
498 #define OPT_CLIENT              "client"
499 	OPT_CLIENT_NUM,
500 #define OPT_BUILTIN_NET_DRIVER  "builtin-net-driver"
501 	OPT_BUILTIN_NET_DRIVER_NUM,
502 #define OPT_DMA_TYPE            "dma-type"
503 	OPT_DMA_TYPE_NUM,
504 #define OPT_DMAS                "dmas"
505 	OPT_DMAS_NUM,
506 };
507 
508 /*
509  * Parse the arguments given in the command line of the application.
510  */
511 static int
512 us_vhost_parse_args(int argc, char **argv)
513 {
514 	int opt, ret;
515 	int option_index;
516 	unsigned i;
517 	const char *prgname = argv[0];
518 	static struct option long_option[] = {
519 		{OPT_VM2VM, required_argument,
520 				NULL, OPT_VM2VM_NUM},
521 		{OPT_RX_RETRY, required_argument,
522 				NULL, OPT_RX_RETRY_NUM},
523 		{OPT_RX_RETRY_DELAY, required_argument,
524 				NULL, OPT_RX_RETRY_DELAY_NUM},
525 		{OPT_RX_RETRY_NUMB, required_argument,
526 				NULL, OPT_RX_RETRY_NUMB_NUM},
527 		{OPT_MERGEABLE, required_argument,
528 				NULL, OPT_MERGEABLE_NUM},
529 		{OPT_STATS, required_argument,
530 				NULL, OPT_STATS_NUM},
531 		{OPT_SOCKET_FILE, required_argument,
532 				NULL, OPT_SOCKET_FILE_NUM},
533 		{OPT_TX_CSUM, required_argument,
534 				NULL, OPT_TX_CSUM_NUM},
535 		{OPT_TSO, required_argument,
536 				NULL, OPT_TSO_NUM},
537 		{OPT_CLIENT, no_argument,
538 				NULL, OPT_CLIENT_NUM},
539 		{OPT_BUILTIN_NET_DRIVER, no_argument,
540 				NULL, OPT_BUILTIN_NET_DRIVER_NUM},
541 		{OPT_DMA_TYPE, required_argument,
542 				NULL, OPT_DMA_TYPE_NUM},
543 		{OPT_DMAS, required_argument,
544 				NULL, OPT_DMAS_NUM},
545 		{NULL, 0, 0, 0},
546 	};
547 
548 	/* Parse command line */
549 	while ((opt = getopt_long(argc, argv, "p:P",
550 			long_option, &option_index)) != EOF) {
551 		switch (opt) {
552 		/* Portmask */
553 		case 'p':
554 			enabled_port_mask = parse_portmask(optarg);
555 			if (enabled_port_mask == 0) {
556 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid portmask\n");
557 				us_vhost_usage(prgname);
558 				return -1;
559 			}
560 			break;
561 
562 		case 'P':
563 			promiscuous = 1;
564 			vmdq_conf_default.rx_adv_conf.vmdq_rx_conf.rx_mode =
565 				ETH_VMDQ_ACCEPT_BROADCAST |
566 				ETH_VMDQ_ACCEPT_MULTICAST;
567 			break;
568 
569 		case OPT_VM2VM_NUM:
570 			ret = parse_num_opt(optarg, (VM2VM_LAST - 1));
571 			if (ret == -1) {
572 				RTE_LOG(INFO, VHOST_CONFIG,
573 					"Invalid argument for "
574 					"vm2vm [0|1|2]\n");
575 				us_vhost_usage(prgname);
576 				return -1;
577 			}
578 			vm2vm_mode = (vm2vm_type)ret;
579 			break;
580 
581 		case OPT_RX_RETRY_NUM:
582 			ret = parse_num_opt(optarg, 1);
583 			if (ret == -1) {
584 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry [0|1]\n");
585 				us_vhost_usage(prgname);
586 				return -1;
587 			}
588 			enable_retry = ret;
589 			break;
590 
591 		case OPT_TX_CSUM_NUM:
592 			ret = parse_num_opt(optarg, 1);
593 			if (ret == -1) {
594 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tx-csum [0|1]\n");
595 				us_vhost_usage(prgname);
596 				return -1;
597 			}
598 			enable_tx_csum = ret;
599 			break;
600 
601 		case OPT_TSO_NUM:
602 			ret = parse_num_opt(optarg, 1);
603 			if (ret == -1) {
604 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for tso [0|1]\n");
605 				us_vhost_usage(prgname);
606 				return -1;
607 			}
608 			enable_tso = ret;
609 			break;
610 
611 		case OPT_RX_RETRY_DELAY_NUM:
612 			ret = parse_num_opt(optarg, INT32_MAX);
613 			if (ret == -1) {
614 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-delay [0-N]\n");
615 				us_vhost_usage(prgname);
616 				return -1;
617 			}
618 			burst_rx_delay_time = ret;
619 			break;
620 
621 		case OPT_RX_RETRY_NUMB_NUM:
622 			ret = parse_num_opt(optarg, INT32_MAX);
623 			if (ret == -1) {
624 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for rx-retry-num [0-N]\n");
625 				us_vhost_usage(prgname);
626 				return -1;
627 			}
628 			burst_rx_retry_num = ret;
629 			break;
630 
631 		case OPT_MERGEABLE_NUM:
632 			ret = parse_num_opt(optarg, 1);
633 			if (ret == -1) {
634 				RTE_LOG(INFO, VHOST_CONFIG, "Invalid argument for mergeable [0|1]\n");
635 				us_vhost_usage(prgname);
636 				return -1;
637 			}
638 			mergeable = !!ret;
639 			if (ret) {
640 				vmdq_conf_default.rxmode.offloads |=
641 					DEV_RX_OFFLOAD_JUMBO_FRAME;
642 				vmdq_conf_default.rxmode.max_rx_pkt_len
643 					= JUMBO_FRAME_MAX_SIZE;
644 			}
645 			break;
646 
647 		case OPT_STATS_NUM:
648 			ret = parse_num_opt(optarg, INT32_MAX);
649 			if (ret == -1) {
650 				RTE_LOG(INFO, VHOST_CONFIG,
651 					"Invalid argument for stats [0..N]\n");
652 				us_vhost_usage(prgname);
653 				return -1;
654 			}
655 			enable_stats = ret;
656 			break;
657 
658 		/* Set socket file path. */
659 		case OPT_SOCKET_FILE_NUM:
660 			if (us_vhost_parse_socket_path(optarg) == -1) {
661 				RTE_LOG(INFO, VHOST_CONFIG,
662 				"Invalid argument for socket name (Max %d characters)\n",
663 				PATH_MAX);
664 				us_vhost_usage(prgname);
665 				return -1;
666 			}
667 			break;
668 
669 		case OPT_DMA_TYPE_NUM:
670 			dma_type = optarg;
671 			break;
672 
673 		case OPT_DMAS_NUM:
674 			if (open_dma(optarg) == -1) {
675 				RTE_LOG(INFO, VHOST_CONFIG,
676 					"Wrong DMA args\n");
677 				us_vhost_usage(prgname);
678 				return -1;
679 			}
680 			async_vhost_driver = 1;
681 			break;
682 
683 		case OPT_CLIENT_NUM:
684 			client_mode = 1;
685 			break;
686 
687 		case OPT_BUILTIN_NET_DRIVER_NUM:
688 			builtin_net_driver = 1;
689 			break;
690 
691 		/* Invalid option - print options. */
692 		default:
693 			us_vhost_usage(prgname);
694 			return -1;
695 		}
696 	}
697 
698 	for (i = 0; i < RTE_MAX_ETHPORTS; i++) {
699 		if (enabled_port_mask & (1 << i))
700 			ports[num_ports++] = i;
701 	}
702 
703 	if ((num_ports ==  0) || (num_ports > MAX_SUP_PORTS)) {
704 		RTE_LOG(INFO, VHOST_PORT, "Current enabled port number is %u,"
705 			"but only %u port can be enabled\n",num_ports, MAX_SUP_PORTS);
706 		return -1;
707 	}
708 
709 	return 0;
710 }
711 
712 /*
713  * Update the global var NUM_PORTS and array PORTS according to system ports number
714  * and return valid ports number
715  */
716 static unsigned check_ports_num(unsigned nb_ports)
717 {
718 	unsigned valid_num_ports = num_ports;
719 	unsigned portid;
720 
721 	if (num_ports > nb_ports) {
722 		RTE_LOG(INFO, VHOST_PORT, "\nSpecified port number(%u) exceeds total system port number(%u)\n",
723 			num_ports, nb_ports);
724 		num_ports = nb_ports;
725 	}
726 
727 	for (portid = 0; portid < num_ports; portid ++) {
728 		if (!rte_eth_dev_is_valid_port(ports[portid])) {
729 			RTE_LOG(INFO, VHOST_PORT,
730 				"\nSpecified port ID(%u) is not valid\n",
731 				ports[portid]);
732 			ports[portid] = INVALID_PORT_ID;
733 			valid_num_ports--;
734 		}
735 	}
736 	return valid_num_ports;
737 }
738 
739 static __rte_always_inline struct vhost_dev *
740 find_vhost_dev(struct rte_ether_addr *mac)
741 {
742 	struct vhost_dev *vdev;
743 
744 	TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
745 		if (vdev->ready == DEVICE_RX &&
746 		    rte_is_same_ether_addr(mac, &vdev->mac_address))
747 			return vdev;
748 	}
749 
750 	return NULL;
751 }
752 
753 /*
754  * This function learns the MAC address of the device and registers this along with a
755  * vlan tag to a VMDQ.
756  */
757 static int
758 link_vmdq(struct vhost_dev *vdev, struct rte_mbuf *m)
759 {
760 	struct rte_ether_hdr *pkt_hdr;
761 	int i, ret;
762 
763 	/* Learn MAC address of guest device from packet */
764 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
765 
766 	if (find_vhost_dev(&pkt_hdr->s_addr)) {
767 		RTE_LOG(ERR, VHOST_DATA,
768 			"(%d) device is using a registered MAC!\n",
769 			vdev->vid);
770 		return -1;
771 	}
772 
773 	for (i = 0; i < RTE_ETHER_ADDR_LEN; i++)
774 		vdev->mac_address.addr_bytes[i] = pkt_hdr->s_addr.addr_bytes[i];
775 
776 	/* vlan_tag currently uses the device_id. */
777 	vdev->vlan_tag = vlan_tags[vdev->vid];
778 
779 	/* Print out VMDQ registration info. */
780 	RTE_LOG(INFO, VHOST_DATA,
781 		"(%d) mac %02x:%02x:%02x:%02x:%02x:%02x and vlan %d registered\n",
782 		vdev->vid,
783 		vdev->mac_address.addr_bytes[0], vdev->mac_address.addr_bytes[1],
784 		vdev->mac_address.addr_bytes[2], vdev->mac_address.addr_bytes[3],
785 		vdev->mac_address.addr_bytes[4], vdev->mac_address.addr_bytes[5],
786 		vdev->vlan_tag);
787 
788 	/* Register the MAC address. */
789 	ret = rte_eth_dev_mac_addr_add(ports[0], &vdev->mac_address,
790 				(uint32_t)vdev->vid + vmdq_pool_base);
791 	if (ret)
792 		RTE_LOG(ERR, VHOST_DATA,
793 			"(%d) failed to add device MAC address to VMDQ\n",
794 			vdev->vid);
795 
796 	rte_eth_dev_set_vlan_strip_on_queue(ports[0], vdev->vmdq_rx_q, 1);
797 
798 	/* Set device as ready for RX. */
799 	vdev->ready = DEVICE_RX;
800 
801 	return 0;
802 }
803 
804 /*
805  * Removes MAC address and vlan tag from VMDQ. Ensures that nothing is adding buffers to the RX
806  * queue before disabling RX on the device.
807  */
808 static inline void
809 unlink_vmdq(struct vhost_dev *vdev)
810 {
811 	unsigned i = 0;
812 	unsigned rx_count;
813 	struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
814 
815 	if (vdev->ready == DEVICE_RX) {
816 		/*clear MAC and VLAN settings*/
817 		rte_eth_dev_mac_addr_remove(ports[0], &vdev->mac_address);
818 		for (i = 0; i < 6; i++)
819 			vdev->mac_address.addr_bytes[i] = 0;
820 
821 		vdev->vlan_tag = 0;
822 
823 		/*Clear out the receive buffers*/
824 		rx_count = rte_eth_rx_burst(ports[0],
825 					(uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST);
826 
827 		while (rx_count) {
828 			for (i = 0; i < rx_count; i++)
829 				rte_pktmbuf_free(pkts_burst[i]);
830 
831 			rx_count = rte_eth_rx_burst(ports[0],
832 					(uint16_t)vdev->vmdq_rx_q, pkts_burst, MAX_PKT_BURST);
833 		}
834 
835 		vdev->ready = DEVICE_MAC_LEARNING;
836 	}
837 }
838 
839 static inline void
840 free_pkts(struct rte_mbuf **pkts, uint16_t n)
841 {
842 	while (n--)
843 		rte_pktmbuf_free(pkts[n]);
844 }
845 
846 static __rte_always_inline void
847 complete_async_pkts(struct vhost_dev *vdev)
848 {
849 	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
850 	uint16_t complete_count;
851 
852 	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
853 					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
854 	if (complete_count)
855 		free_pkts(p_cpl, complete_count);
856 }
857 
858 static __rte_always_inline void
859 sync_virtio_xmit(struct vhost_dev *dst_vdev, struct vhost_dev *src_vdev,
860 	    struct rte_mbuf *m)
861 {
862 	uint16_t ret;
863 
864 	if (builtin_net_driver) {
865 		ret = vs_enqueue_pkts(dst_vdev, VIRTIO_RXQ, &m, 1);
866 	} else {
867 		ret = rte_vhost_enqueue_burst(dst_vdev->vid, VIRTIO_RXQ, &m, 1);
868 	}
869 
870 	if (enable_stats) {
871 		__atomic_add_fetch(&dst_vdev->stats.rx_total_atomic, 1,
872 				__ATOMIC_SEQ_CST);
873 		__atomic_add_fetch(&dst_vdev->stats.rx_atomic, ret,
874 				__ATOMIC_SEQ_CST);
875 		src_vdev->stats.tx_total++;
876 		src_vdev->stats.tx += ret;
877 	}
878 }
879 
880 static __rte_always_inline void
881 drain_vhost(struct vhost_dev *vdev)
882 {
883 	uint16_t ret;
884 	uint32_t buff_idx = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
885 	uint16_t nr_xmit = vhost_txbuff[buff_idx]->len;
886 	struct rte_mbuf **m = vhost_txbuff[buff_idx]->m_table;
887 
888 	if (builtin_net_driver) {
889 		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
890 	} else if (async_vhost_driver) {
891 		uint32_t cpu_cpl_nr = 0;
892 		uint16_t enqueue_fail = 0;
893 		struct rte_mbuf *m_cpu_cpl[nr_xmit];
894 
895 		complete_async_pkts(vdev);
896 		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ,
897 					m, nr_xmit, m_cpu_cpl, &cpu_cpl_nr);
898 
899 		if (cpu_cpl_nr)
900 			free_pkts(m_cpu_cpl, cpu_cpl_nr);
901 
902 		enqueue_fail = nr_xmit - ret;
903 		if (enqueue_fail)
904 			free_pkts(&m[ret], nr_xmit - ret);
905 	} else {
906 		ret = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
907 						m, nr_xmit);
908 	}
909 
910 	if (enable_stats) {
911 		__atomic_add_fetch(&vdev->stats.rx_total_atomic, nr_xmit,
912 				__ATOMIC_SEQ_CST);
913 		__atomic_add_fetch(&vdev->stats.rx_atomic, ret,
914 				__ATOMIC_SEQ_CST);
915 	}
916 
917 	if (!async_vhost_driver)
918 		free_pkts(m, nr_xmit);
919 }
920 
921 static __rte_always_inline void
922 drain_vhost_table(void)
923 {
924 	uint16_t lcore_id = rte_lcore_id();
925 	struct vhost_bufftable *vhost_txq;
926 	struct vhost_dev *vdev;
927 	uint64_t cur_tsc;
928 
929 	TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
930 		vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE
931 						+ vdev->vid];
932 
933 		cur_tsc = rte_rdtsc();
934 		if (unlikely(cur_tsc - vhost_txq->pre_tsc
935 				> MBUF_TABLE_DRAIN_TSC)) {
936 			RTE_LOG_DP(DEBUG, VHOST_DATA,
937 				"Vhost TX queue drained after timeout with burst size %u\n",
938 				vhost_txq->len);
939 			drain_vhost(vdev);
940 			vhost_txq->len = 0;
941 			vhost_txq->pre_tsc = cur_tsc;
942 		}
943 	}
944 }
945 
946 /*
947  * Check if the packet destination MAC address is for a local device. If so then put
948  * the packet on that devices RX queue. If not then return.
949  */
950 static __rte_always_inline int
951 virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
952 {
953 	struct rte_ether_hdr *pkt_hdr;
954 	struct vhost_dev *dst_vdev;
955 	struct vhost_bufftable *vhost_txq;
956 	uint16_t lcore_id = rte_lcore_id();
957 	pkt_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
958 
959 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);
960 	if (!dst_vdev)
961 		return -1;
962 
963 	if (vdev->vid == dst_vdev->vid) {
964 		RTE_LOG_DP(DEBUG, VHOST_DATA,
965 			"(%d) TX: src and dst MAC is same. Dropping packet.\n",
966 			vdev->vid);
967 		return 0;
968 	}
969 
970 	RTE_LOG_DP(DEBUG, VHOST_DATA,
971 		"(%d) TX: MAC address is local\n", dst_vdev->vid);
972 
973 	if (unlikely(dst_vdev->remove)) {
974 		RTE_LOG_DP(DEBUG, VHOST_DATA,
975 			"(%d) device is marked for removal\n", dst_vdev->vid);
976 		return 0;
977 	}
978 
979 	vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE + dst_vdev->vid];
980 	vhost_txq->m_table[vhost_txq->len++] = m;
981 
982 	if (enable_stats) {
983 		vdev->stats.tx_total++;
984 		vdev->stats.tx++;
985 	}
986 
987 	if (unlikely(vhost_txq->len == MAX_PKT_BURST)) {
988 		drain_vhost(dst_vdev);
989 		vhost_txq->len = 0;
990 		vhost_txq->pre_tsc = rte_rdtsc();
991 	}
992 	return 0;
993 }
994 
995 /*
996  * Check if the destination MAC of a packet is one local VM,
997  * and get its vlan tag, and offset if it is.
998  */
999 static __rte_always_inline int
1000 find_local_dest(struct vhost_dev *vdev, struct rte_mbuf *m,
1001 	uint32_t *offset, uint16_t *vlan_tag)
1002 {
1003 	struct vhost_dev *dst_vdev;
1004 	struct rte_ether_hdr *pkt_hdr =
1005 		rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1006 
1007 	dst_vdev = find_vhost_dev(&pkt_hdr->d_addr);
1008 	if (!dst_vdev)
1009 		return 0;
1010 
1011 	if (vdev->vid == dst_vdev->vid) {
1012 		RTE_LOG_DP(DEBUG, VHOST_DATA,
1013 			"(%d) TX: src and dst MAC is same. Dropping packet.\n",
1014 			vdev->vid);
1015 		return -1;
1016 	}
1017 
1018 	/*
1019 	 * HW vlan strip will reduce the packet length
1020 	 * by minus length of vlan tag, so need restore
1021 	 * the packet length by plus it.
1022 	 */
1023 	*offset  = VLAN_HLEN;
1024 	*vlan_tag = vlan_tags[vdev->vid];
1025 
1026 	RTE_LOG_DP(DEBUG, VHOST_DATA,
1027 		"(%d) TX: pkt to local VM device id: (%d), vlan tag: %u.\n",
1028 		vdev->vid, dst_vdev->vid, *vlan_tag);
1029 
1030 	return 0;
1031 }
1032 
1033 static void virtio_tx_offload(struct rte_mbuf *m)
1034 {
1035 	struct rte_net_hdr_lens hdr_lens;
1036 	struct rte_ipv4_hdr *ipv4_hdr;
1037 	struct rte_tcp_hdr *tcp_hdr;
1038 	uint32_t ptype;
1039 	void *l3_hdr;
1040 
1041 	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
1042 	m->l2_len = hdr_lens.l2_len;
1043 	m->l3_len = hdr_lens.l3_len;
1044 	m->l4_len = hdr_lens.l4_len;
1045 
1046 	l3_hdr = rte_pktmbuf_mtod_offset(m, void *, m->l2_len);
1047 	tcp_hdr = rte_pktmbuf_mtod_offset(m, struct rte_tcp_hdr *,
1048 		m->l2_len + m->l3_len);
1049 
1050 	m->ol_flags |= PKT_TX_TCP_SEG;
1051 	if ((ptype & RTE_PTYPE_L3_MASK) == RTE_PTYPE_L3_IPV4) {
1052 		m->ol_flags |= PKT_TX_IPV4;
1053 		m->ol_flags |= PKT_TX_IP_CKSUM;
1054 		ipv4_hdr = l3_hdr;
1055 		ipv4_hdr->hdr_checksum = 0;
1056 		tcp_hdr->cksum = rte_ipv4_phdr_cksum(l3_hdr, m->ol_flags);
1057 	} else { /* assume ethertype == RTE_ETHER_TYPE_IPV6 */
1058 		m->ol_flags |= PKT_TX_IPV6;
1059 		tcp_hdr->cksum = rte_ipv6_phdr_cksum(l3_hdr, m->ol_flags);
1060 	}
1061 }
1062 
1063 static __rte_always_inline void
1064 do_drain_mbuf_table(struct mbuf_table *tx_q)
1065 {
1066 	uint16_t count;
1067 
1068 	count = rte_eth_tx_burst(ports[0], tx_q->txq_id,
1069 				 tx_q->m_table, tx_q->len);
1070 	if (unlikely(count < tx_q->len))
1071 		free_pkts(&tx_q->m_table[count], tx_q->len - count);
1072 
1073 	tx_q->len = 0;
1074 }
1075 
1076 /*
1077  * This function routes the TX packet to the correct interface. This
1078  * may be a local device or the physical port.
1079  */
1080 static __rte_always_inline void
1081 virtio_tx_route(struct vhost_dev *vdev, struct rte_mbuf *m, uint16_t vlan_tag)
1082 {
1083 	struct mbuf_table *tx_q;
1084 	unsigned offset = 0;
1085 	const uint16_t lcore_id = rte_lcore_id();
1086 	struct rte_ether_hdr *nh;
1087 
1088 
1089 	nh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1090 	if (unlikely(rte_is_broadcast_ether_addr(&nh->d_addr))) {
1091 		struct vhost_dev *vdev2;
1092 
1093 		TAILQ_FOREACH(vdev2, &vhost_dev_list, global_vdev_entry) {
1094 			if (vdev2 != vdev)
1095 				sync_virtio_xmit(vdev2, vdev, m);
1096 		}
1097 		goto queue2nic;
1098 	}
1099 
1100 	/*check if destination is local VM*/
1101 	if ((vm2vm_mode == VM2VM_SOFTWARE) && (virtio_tx_local(vdev, m) == 0))
1102 		return;
1103 
1104 	if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) {
1105 		if (unlikely(find_local_dest(vdev, m, &offset,
1106 					     &vlan_tag) != 0)) {
1107 			rte_pktmbuf_free(m);
1108 			return;
1109 		}
1110 	}
1111 
1112 	RTE_LOG_DP(DEBUG, VHOST_DATA,
1113 		"(%d) TX: MAC address is external\n", vdev->vid);
1114 
1115 queue2nic:
1116 
1117 	/*Add packet to the port tx queue*/
1118 	tx_q = &lcore_tx_queue[lcore_id];
1119 
1120 	nh = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1121 	if (unlikely(nh->ether_type == rte_cpu_to_be_16(RTE_ETHER_TYPE_VLAN))) {
1122 		/* Guest has inserted the vlan tag. */
1123 		struct rte_vlan_hdr *vh = (struct rte_vlan_hdr *) (nh + 1);
1124 		uint16_t vlan_tag_be = rte_cpu_to_be_16(vlan_tag);
1125 		if ((vm2vm_mode == VM2VM_HARDWARE) &&
1126 			(vh->vlan_tci != vlan_tag_be))
1127 			vh->vlan_tci = vlan_tag_be;
1128 	} else {
1129 		m->ol_flags |= PKT_TX_VLAN_PKT;
1130 
1131 		/*
1132 		 * Find the right seg to adjust the data len when offset is
1133 		 * bigger than tail room size.
1134 		 */
1135 		if (unlikely(vm2vm_mode == VM2VM_HARDWARE)) {
1136 			if (likely(offset <= rte_pktmbuf_tailroom(m)))
1137 				m->data_len += offset;
1138 			else {
1139 				struct rte_mbuf *seg = m;
1140 
1141 				while ((seg->next != NULL) &&
1142 					(offset > rte_pktmbuf_tailroom(seg)))
1143 					seg = seg->next;
1144 
1145 				seg->data_len += offset;
1146 			}
1147 			m->pkt_len += offset;
1148 		}
1149 
1150 		m->vlan_tci = vlan_tag;
1151 	}
1152 
1153 	if (m->ol_flags & PKT_RX_LRO)
1154 		virtio_tx_offload(m);
1155 
1156 	tx_q->m_table[tx_q->len++] = m;
1157 	if (enable_stats) {
1158 		vdev->stats.tx_total++;
1159 		vdev->stats.tx++;
1160 	}
1161 
1162 	if (unlikely(tx_q->len == MAX_PKT_BURST))
1163 		do_drain_mbuf_table(tx_q);
1164 }
1165 
1166 
1167 static __rte_always_inline void
1168 drain_mbuf_table(struct mbuf_table *tx_q)
1169 {
1170 	static uint64_t prev_tsc;
1171 	uint64_t cur_tsc;
1172 
1173 	if (tx_q->len == 0)
1174 		return;
1175 
1176 	cur_tsc = rte_rdtsc();
1177 	if (unlikely(cur_tsc - prev_tsc > MBUF_TABLE_DRAIN_TSC)) {
1178 		prev_tsc = cur_tsc;
1179 
1180 		RTE_LOG_DP(DEBUG, VHOST_DATA,
1181 			"TX queue drained after timeout with burst size %u\n",
1182 			tx_q->len);
1183 		do_drain_mbuf_table(tx_q);
1184 	}
1185 }
1186 
1187 static __rte_always_inline void
1188 drain_eth_rx(struct vhost_dev *vdev)
1189 {
1190 	uint16_t rx_count, enqueue_count;
1191 	struct rte_mbuf *pkts[MAX_PKT_BURST];
1192 
1193 	rx_count = rte_eth_rx_burst(ports[0], vdev->vmdq_rx_q,
1194 				    pkts, MAX_PKT_BURST);
1195 
1196 	if (!rx_count)
1197 		return;
1198 
1199 	/*
1200 	 * When "enable_retry" is set, here we wait and retry when there
1201 	 * is no enough free slots in the queue to hold @rx_count packets,
1202 	 * to diminish packet loss.
1203 	 */
1204 	if (enable_retry &&
1205 	    unlikely(rx_count > rte_vhost_avail_entries(vdev->vid,
1206 			VIRTIO_RXQ))) {
1207 		uint32_t retry;
1208 
1209 		for (retry = 0; retry < burst_rx_retry_num; retry++) {
1210 			rte_delay_us(burst_rx_delay_time);
1211 			if (rx_count <= rte_vhost_avail_entries(vdev->vid,
1212 					VIRTIO_RXQ))
1213 				break;
1214 		}
1215 	}
1216 
1217 	if (builtin_net_driver) {
1218 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
1219 						pkts, rx_count);
1220 	} else if (async_vhost_driver) {
1221 		uint32_t cpu_cpl_nr = 0;
1222 		uint16_t enqueue_fail = 0;
1223 		struct rte_mbuf *m_cpu_cpl[MAX_PKT_BURST];
1224 
1225 		complete_async_pkts(vdev);
1226 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
1227 					VIRTIO_RXQ, pkts, rx_count,
1228 					m_cpu_cpl, &cpu_cpl_nr);
1229 		if (cpu_cpl_nr)
1230 			free_pkts(m_cpu_cpl, cpu_cpl_nr);
1231 
1232 		enqueue_fail = rx_count - enqueue_count;
1233 		if (enqueue_fail)
1234 			free_pkts(&pkts[enqueue_count], enqueue_fail);
1235 
1236 	} else {
1237 		enqueue_count = rte_vhost_enqueue_burst(vdev->vid, VIRTIO_RXQ,
1238 						pkts, rx_count);
1239 	}
1240 
1241 	if (enable_stats) {
1242 		__atomic_add_fetch(&vdev->stats.rx_total_atomic, rx_count,
1243 				__ATOMIC_SEQ_CST);
1244 		__atomic_add_fetch(&vdev->stats.rx_atomic, enqueue_count,
1245 				__ATOMIC_SEQ_CST);
1246 	}
1247 
1248 	if (!async_vhost_driver)
1249 		free_pkts(pkts, rx_count);
1250 }
1251 
1252 static __rte_always_inline void
1253 drain_virtio_tx(struct vhost_dev *vdev)
1254 {
1255 	struct rte_mbuf *pkts[MAX_PKT_BURST];
1256 	uint16_t count;
1257 	uint16_t i;
1258 
1259 	if (builtin_net_driver) {
1260 		count = vs_dequeue_pkts(vdev, VIRTIO_TXQ, mbuf_pool,
1261 					pkts, MAX_PKT_BURST);
1262 	} else {
1263 		count = rte_vhost_dequeue_burst(vdev->vid, VIRTIO_TXQ,
1264 					mbuf_pool, pkts, MAX_PKT_BURST);
1265 	}
1266 
1267 	/* setup VMDq for the first packet */
1268 	if (unlikely(vdev->ready == DEVICE_MAC_LEARNING) && count) {
1269 		if (vdev->remove || link_vmdq(vdev, pkts[0]) == -1)
1270 			free_pkts(pkts, count);
1271 	}
1272 
1273 	for (i = 0; i < count; ++i)
1274 		virtio_tx_route(vdev, pkts[i], vlan_tags[vdev->vid]);
1275 }
1276 
1277 /*
1278  * Main function of vhost-switch. It basically does:
1279  *
1280  * for each vhost device {
1281  *    - drain_eth_rx()
1282  *
1283  *      Which drains the host eth Rx queue linked to the vhost device,
1284  *      and deliver all of them to guest virito Rx ring associated with
1285  *      this vhost device.
1286  *
1287  *    - drain_virtio_tx()
1288  *
1289  *      Which drains the guest virtio Tx queue and deliver all of them
1290  *      to the target, which could be another vhost device, or the
1291  *      physical eth dev. The route is done in function "virtio_tx_route".
1292  * }
1293  */
1294 static int
1295 switch_worker(void *arg __rte_unused)
1296 {
1297 	unsigned i;
1298 	unsigned lcore_id = rte_lcore_id();
1299 	struct vhost_dev *vdev;
1300 	struct mbuf_table *tx_q;
1301 
1302 	RTE_LOG(INFO, VHOST_DATA, "Procesing on Core %u started\n", lcore_id);
1303 
1304 	tx_q = &lcore_tx_queue[lcore_id];
1305 	for (i = 0; i < rte_lcore_count(); i++) {
1306 		if (lcore_ids[i] == lcore_id) {
1307 			tx_q->txq_id = i;
1308 			break;
1309 		}
1310 	}
1311 
1312 	while(1) {
1313 		drain_mbuf_table(tx_q);
1314 		drain_vhost_table();
1315 		/*
1316 		 * Inform the configuration core that we have exited the
1317 		 * linked list and that no devices are in use if requested.
1318 		 */
1319 		if (lcore_info[lcore_id].dev_removal_flag == REQUEST_DEV_REMOVAL)
1320 			lcore_info[lcore_id].dev_removal_flag = ACK_DEV_REMOVAL;
1321 
1322 		/*
1323 		 * Process vhost devices
1324 		 */
1325 		TAILQ_FOREACH(vdev, &lcore_info[lcore_id].vdev_list,
1326 			      lcore_vdev_entry) {
1327 			if (unlikely(vdev->remove)) {
1328 				unlink_vmdq(vdev);
1329 				vdev->ready = DEVICE_SAFE_REMOVE;
1330 				continue;
1331 			}
1332 
1333 			if (likely(vdev->ready == DEVICE_RX))
1334 				drain_eth_rx(vdev);
1335 
1336 			if (likely(!vdev->remove))
1337 				drain_virtio_tx(vdev);
1338 		}
1339 	}
1340 
1341 	return 0;
1342 }
1343 
1344 /*
1345  * Remove a device from the specific data core linked list and from the
1346  * main linked list. Synchonization  occurs through the use of the
1347  * lcore dev_removal_flag. Device is made volatile here to avoid re-ordering
1348  * of dev->remove=1 which can cause an infinite loop in the rte_pause loop.
1349  */
1350 static void
1351 destroy_device(int vid)
1352 {
1353 	struct vhost_dev *vdev = NULL;
1354 	int lcore;
1355 	uint16_t i;
1356 
1357 	TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
1358 		if (vdev->vid == vid)
1359 			break;
1360 	}
1361 	if (!vdev)
1362 		return;
1363 	/*set the remove flag. */
1364 	vdev->remove = 1;
1365 	while(vdev->ready != DEVICE_SAFE_REMOVE) {
1366 		rte_pause();
1367 	}
1368 
1369 	for (i = 0; i < RTE_MAX_LCORE; i++)
1370 		rte_free(vhost_txbuff[i * MAX_VHOST_DEVICE + vid]);
1371 
1372 	if (builtin_net_driver)
1373 		vs_vhost_net_remove(vdev);
1374 
1375 	TAILQ_REMOVE(&lcore_info[vdev->coreid].vdev_list, vdev,
1376 		     lcore_vdev_entry);
1377 	TAILQ_REMOVE(&vhost_dev_list, vdev, global_vdev_entry);
1378 
1379 
1380 	/* Set the dev_removal_flag on each lcore. */
1381 	RTE_LCORE_FOREACH_WORKER(lcore)
1382 		lcore_info[lcore].dev_removal_flag = REQUEST_DEV_REMOVAL;
1383 
1384 	/*
1385 	 * Once each core has set the dev_removal_flag to ACK_DEV_REMOVAL
1386 	 * we can be sure that they can no longer access the device removed
1387 	 * from the linked lists and that the devices are no longer in use.
1388 	 */
1389 	RTE_LCORE_FOREACH_WORKER(lcore) {
1390 		while (lcore_info[lcore].dev_removal_flag != ACK_DEV_REMOVAL)
1391 			rte_pause();
1392 	}
1393 
1394 	lcore_info[vdev->coreid].device_num--;
1395 
1396 	RTE_LOG(INFO, VHOST_DATA,
1397 		"(%d) device has been removed from data core\n",
1398 		vdev->vid);
1399 
1400 	if (async_vhost_driver)
1401 		rte_vhost_async_channel_unregister(vid, VIRTIO_RXQ);
1402 
1403 	rte_free(vdev);
1404 }
1405 
1406 /*
1407  * A new device is added to a data core. First the device is added to the main linked list
1408  * and then allocated to a specific data core.
1409  */
1410 static int
1411 new_device(int vid)
1412 {
1413 	int lcore, core_add = 0;
1414 	uint16_t i;
1415 	uint32_t device_num_min = num_devices;
1416 	struct vhost_dev *vdev;
1417 	vdev = rte_zmalloc("vhost device", sizeof(*vdev), RTE_CACHE_LINE_SIZE);
1418 	if (vdev == NULL) {
1419 		RTE_LOG(INFO, VHOST_DATA,
1420 			"(%d) couldn't allocate memory for vhost dev\n",
1421 			vid);
1422 		return -1;
1423 	}
1424 	vdev->vid = vid;
1425 
1426 	for (i = 0; i < RTE_MAX_LCORE; i++) {
1427 		vhost_txbuff[i * MAX_VHOST_DEVICE + vid]
1428 			= rte_zmalloc("vhost bufftable",
1429 				sizeof(struct vhost_bufftable),
1430 				RTE_CACHE_LINE_SIZE);
1431 
1432 		if (vhost_txbuff[i * MAX_VHOST_DEVICE + vid] == NULL) {
1433 			RTE_LOG(INFO, VHOST_DATA,
1434 			  "(%d) couldn't allocate memory for vhost TX\n", vid);
1435 			return -1;
1436 		}
1437 	}
1438 
1439 	if (builtin_net_driver)
1440 		vs_vhost_net_setup(vdev);
1441 
1442 	TAILQ_INSERT_TAIL(&vhost_dev_list, vdev, global_vdev_entry);
1443 	vdev->vmdq_rx_q = vid * queues_per_pool + vmdq_queue_base;
1444 
1445 	/*reset ready flag*/
1446 	vdev->ready = DEVICE_MAC_LEARNING;
1447 	vdev->remove = 0;
1448 
1449 	/* Find a suitable lcore to add the device. */
1450 	RTE_LCORE_FOREACH_WORKER(lcore) {
1451 		if (lcore_info[lcore].device_num < device_num_min) {
1452 			device_num_min = lcore_info[lcore].device_num;
1453 			core_add = lcore;
1454 		}
1455 	}
1456 	vdev->coreid = core_add;
1457 
1458 	TAILQ_INSERT_TAIL(&lcore_info[vdev->coreid].vdev_list, vdev,
1459 			  lcore_vdev_entry);
1460 	lcore_info[vdev->coreid].device_num++;
1461 
1462 	/* Disable notifications. */
1463 	rte_vhost_enable_guest_notification(vid, VIRTIO_RXQ, 0);
1464 	rte_vhost_enable_guest_notification(vid, VIRTIO_TXQ, 0);
1465 
1466 	RTE_LOG(INFO, VHOST_DATA,
1467 		"(%d) device has been added to data core %d\n",
1468 		vid, vdev->coreid);
1469 
1470 	if (async_vhost_driver) {
1471 		struct rte_vhost_async_features f;
1472 		struct rte_vhost_async_channel_ops channel_ops;
1473 
1474 		if (dma_type != NULL && strncmp(dma_type, "ioat", 4) == 0) {
1475 			channel_ops.transfer_data = ioat_transfer_data_cb;
1476 			channel_ops.check_completed_copies =
1477 				ioat_check_completed_copies_cb;
1478 
1479 			f.async_inorder = 1;
1480 			f.async_threshold = 256;
1481 
1482 			return rte_vhost_async_channel_register(vid, VIRTIO_RXQ,
1483 				f.intval, &channel_ops);
1484 		}
1485 	}
1486 
1487 	return 0;
1488 }
1489 
1490 /*
1491  * These callback allow devices to be added to the data core when configuration
1492  * has been fully complete.
1493  */
1494 static const struct vhost_device_ops virtio_net_device_ops =
1495 {
1496 	.new_device =  new_device,
1497 	.destroy_device = destroy_device,
1498 };
1499 
1500 /*
1501  * This is a thread will wake up after a period to print stats if the user has
1502  * enabled them.
1503  */
1504 static void *
1505 print_stats(__rte_unused void *arg)
1506 {
1507 	struct vhost_dev *vdev;
1508 	uint64_t tx_dropped, rx_dropped;
1509 	uint64_t tx, tx_total, rx, rx_total;
1510 	const char clr[] = { 27, '[', '2', 'J', '\0' };
1511 	const char top_left[] = { 27, '[', '1', ';', '1', 'H','\0' };
1512 
1513 	while(1) {
1514 		sleep(enable_stats);
1515 
1516 		/* Clear screen and move to top left */
1517 		printf("%s%s\n", clr, top_left);
1518 		printf("Device statistics =================================\n");
1519 
1520 		TAILQ_FOREACH(vdev, &vhost_dev_list, global_vdev_entry) {
1521 			tx_total   = vdev->stats.tx_total;
1522 			tx         = vdev->stats.tx;
1523 			tx_dropped = tx_total - tx;
1524 
1525 			rx_total = __atomic_load_n(&vdev->stats.rx_total_atomic,
1526 				__ATOMIC_SEQ_CST);
1527 			rx         = __atomic_load_n(&vdev->stats.rx_atomic,
1528 				__ATOMIC_SEQ_CST);
1529 			rx_dropped = rx_total - rx;
1530 
1531 			printf("Statistics for device %d\n"
1532 				"-----------------------\n"
1533 				"TX total:              %" PRIu64 "\n"
1534 				"TX dropped:            %" PRIu64 "\n"
1535 				"TX successful:         %" PRIu64 "\n"
1536 				"RX total:              %" PRIu64 "\n"
1537 				"RX dropped:            %" PRIu64 "\n"
1538 				"RX successful:         %" PRIu64 "\n",
1539 				vdev->vid,
1540 				tx_total, tx_dropped, tx,
1541 				rx_total, rx_dropped, rx);
1542 		}
1543 
1544 		printf("===================================================\n");
1545 
1546 		fflush(stdout);
1547 	}
1548 
1549 	return NULL;
1550 }
1551 
1552 static void
1553 unregister_drivers(int socket_num)
1554 {
1555 	int i, ret;
1556 
1557 	for (i = 0; i < socket_num; i++) {
1558 		ret = rte_vhost_driver_unregister(socket_files + i * PATH_MAX);
1559 		if (ret != 0)
1560 			RTE_LOG(ERR, VHOST_CONFIG,
1561 				"Fail to unregister vhost driver for %s.\n",
1562 				socket_files + i * PATH_MAX);
1563 	}
1564 }
1565 
1566 /* When we receive a INT signal, unregister vhost driver */
1567 static void
1568 sigint_handler(__rte_unused int signum)
1569 {
1570 	/* Unregister vhost driver. */
1571 	unregister_drivers(nb_sockets);
1572 
1573 	exit(0);
1574 }
1575 
1576 /*
1577  * While creating an mbuf pool, one key thing is to figure out how
1578  * many mbuf entries is enough for our use. FYI, here are some
1579  * guidelines:
1580  *
1581  * - Each rx queue would reserve @nr_rx_desc mbufs at queue setup stage
1582  *
1583  * - For each switch core (A CPU core does the packet switch), we need
1584  *   also make some reservation for receiving the packets from virtio
1585  *   Tx queue. How many is enough depends on the usage. It's normally
1586  *   a simple calculation like following:
1587  *
1588  *       MAX_PKT_BURST * max packet size / mbuf size
1589  *
1590  *   So, we definitely need allocate more mbufs when TSO is enabled.
1591  *
1592  * - Similarly, for each switching core, we should serve @nr_rx_desc
1593  *   mbufs for receiving the packets from physical NIC device.
1594  *
1595  * - We also need make sure, for each switch core, we have allocated
1596  *   enough mbufs to fill up the mbuf cache.
1597  */
1598 static void
1599 create_mbuf_pool(uint16_t nr_port, uint32_t nr_switch_core, uint32_t mbuf_size,
1600 	uint32_t nr_queues, uint32_t nr_rx_desc, uint32_t nr_mbuf_cache)
1601 {
1602 	uint32_t nr_mbufs;
1603 	uint32_t nr_mbufs_per_core;
1604 	uint32_t mtu = 1500;
1605 
1606 	if (mergeable)
1607 		mtu = 9000;
1608 	if (enable_tso)
1609 		mtu = 64 * 1024;
1610 
1611 	nr_mbufs_per_core  = (mtu + mbuf_size) * MAX_PKT_BURST /
1612 			(mbuf_size - RTE_PKTMBUF_HEADROOM);
1613 	nr_mbufs_per_core += nr_rx_desc;
1614 	nr_mbufs_per_core  = RTE_MAX(nr_mbufs_per_core, nr_mbuf_cache);
1615 
1616 	nr_mbufs  = nr_queues * nr_rx_desc;
1617 	nr_mbufs += nr_mbufs_per_core * nr_switch_core;
1618 	nr_mbufs *= nr_port;
1619 
1620 	mbuf_pool = rte_pktmbuf_pool_create("MBUF_POOL", nr_mbufs,
1621 					    nr_mbuf_cache, 0, mbuf_size,
1622 					    rte_socket_id());
1623 	if (mbuf_pool == NULL)
1624 		rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n");
1625 }
1626 
1627 /*
1628  * Main function, does initialisation and calls the per-lcore functions.
1629  */
1630 int
1631 main(int argc, char *argv[])
1632 {
1633 	unsigned lcore_id, core_id = 0;
1634 	unsigned nb_ports, valid_num_ports;
1635 	int ret, i;
1636 	uint16_t portid;
1637 	static pthread_t tid;
1638 	uint64_t flags = RTE_VHOST_USER_NET_COMPLIANT_OL_FLAGS;
1639 
1640 	signal(SIGINT, sigint_handler);
1641 
1642 	/* init EAL */
1643 	ret = rte_eal_init(argc, argv);
1644 	if (ret < 0)
1645 		rte_exit(EXIT_FAILURE, "Error with EAL initialization\n");
1646 	argc -= ret;
1647 	argv += ret;
1648 
1649 	/* parse app arguments */
1650 	ret = us_vhost_parse_args(argc, argv);
1651 	if (ret < 0)
1652 		rte_exit(EXIT_FAILURE, "Invalid argument\n");
1653 
1654 	for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
1655 		TAILQ_INIT(&lcore_info[lcore_id].vdev_list);
1656 
1657 		if (rte_lcore_is_enabled(lcore_id))
1658 			lcore_ids[core_id++] = lcore_id;
1659 	}
1660 
1661 	if (rte_lcore_count() > RTE_MAX_LCORE)
1662 		rte_exit(EXIT_FAILURE,"Not enough cores\n");
1663 
1664 	/* Get the number of physical ports. */
1665 	nb_ports = rte_eth_dev_count_avail();
1666 
1667 	/*
1668 	 * Update the global var NUM_PORTS and global array PORTS
1669 	 * and get value of var VALID_NUM_PORTS according to system ports number
1670 	 */
1671 	valid_num_ports = check_ports_num(nb_ports);
1672 
1673 	if ((valid_num_ports ==  0) || (valid_num_ports > MAX_SUP_PORTS)) {
1674 		RTE_LOG(INFO, VHOST_PORT, "Current enabled port number is %u,"
1675 			"but only %u port can be enabled\n",num_ports, MAX_SUP_PORTS);
1676 		return -1;
1677 	}
1678 
1679 	/*
1680 	 * FIXME: here we are trying to allocate mbufs big enough for
1681 	 * @MAX_QUEUES, but the truth is we're never going to use that
1682 	 * many queues here. We probably should only do allocation for
1683 	 * those queues we are going to use.
1684 	 */
1685 	create_mbuf_pool(valid_num_ports, rte_lcore_count() - 1, MBUF_DATA_SIZE,
1686 			 MAX_QUEUES, RTE_TEST_RX_DESC_DEFAULT, MBUF_CACHE_SIZE);
1687 
1688 	if (vm2vm_mode == VM2VM_HARDWARE) {
1689 		/* Enable VT loop back to let L2 switch to do it. */
1690 		vmdq_conf_default.rx_adv_conf.vmdq_rx_conf.enable_loop_back = 1;
1691 		RTE_LOG(DEBUG, VHOST_CONFIG,
1692 			"Enable loop back for L2 switch in vmdq.\n");
1693 	}
1694 
1695 	/* initialize all ports */
1696 	RTE_ETH_FOREACH_DEV(portid) {
1697 		/* skip ports that are not enabled */
1698 		if ((enabled_port_mask & (1 << portid)) == 0) {
1699 			RTE_LOG(INFO, VHOST_PORT,
1700 				"Skipping disabled port %d\n", portid);
1701 			continue;
1702 		}
1703 		if (port_init(portid) != 0)
1704 			rte_exit(EXIT_FAILURE,
1705 				"Cannot initialize network ports\n");
1706 	}
1707 
1708 	/* Enable stats if the user option is set. */
1709 	if (enable_stats) {
1710 		ret = rte_ctrl_thread_create(&tid, "print-stats", NULL,
1711 					print_stats, NULL);
1712 		if (ret < 0)
1713 			rte_exit(EXIT_FAILURE,
1714 				"Cannot create print-stats thread\n");
1715 	}
1716 
1717 	/* Launch all data cores. */
1718 	RTE_LCORE_FOREACH_WORKER(lcore_id)
1719 		rte_eal_remote_launch(switch_worker, NULL, lcore_id);
1720 
1721 	if (client_mode)
1722 		flags |= RTE_VHOST_USER_CLIENT;
1723 
1724 	/* Register vhost user driver to handle vhost messages. */
1725 	for (i = 0; i < nb_sockets; i++) {
1726 		char *file = socket_files + i * PATH_MAX;
1727 
1728 		if (async_vhost_driver)
1729 			flags = flags | RTE_VHOST_USER_ASYNC_COPY;
1730 
1731 		ret = rte_vhost_driver_register(file, flags);
1732 		if (ret != 0) {
1733 			unregister_drivers(i);
1734 			rte_exit(EXIT_FAILURE,
1735 				"vhost driver register failure.\n");
1736 		}
1737 
1738 		if (builtin_net_driver)
1739 			rte_vhost_driver_set_features(file, VIRTIO_NET_FEATURES);
1740 
1741 		if (mergeable == 0) {
1742 			rte_vhost_driver_disable_features(file,
1743 				1ULL << VIRTIO_NET_F_MRG_RXBUF);
1744 		}
1745 
1746 		if (enable_tx_csum == 0) {
1747 			rte_vhost_driver_disable_features(file,
1748 				1ULL << VIRTIO_NET_F_CSUM);
1749 		}
1750 
1751 		if (enable_tso == 0) {
1752 			rte_vhost_driver_disable_features(file,
1753 				1ULL << VIRTIO_NET_F_HOST_TSO4);
1754 			rte_vhost_driver_disable_features(file,
1755 				1ULL << VIRTIO_NET_F_HOST_TSO6);
1756 			rte_vhost_driver_disable_features(file,
1757 				1ULL << VIRTIO_NET_F_GUEST_TSO4);
1758 			rte_vhost_driver_disable_features(file,
1759 				1ULL << VIRTIO_NET_F_GUEST_TSO6);
1760 		}
1761 
1762 		if (promiscuous) {
1763 			rte_vhost_driver_enable_features(file,
1764 				1ULL << VIRTIO_NET_F_CTRL_RX);
1765 		}
1766 
1767 		ret = rte_vhost_driver_callback_register(file,
1768 			&virtio_net_device_ops);
1769 		if (ret != 0) {
1770 			rte_exit(EXIT_FAILURE,
1771 				"failed to register vhost driver callbacks.\n");
1772 		}
1773 
1774 		if (rte_vhost_driver_start(file) < 0) {
1775 			rte_exit(EXIT_FAILURE,
1776 				"failed to start vhost driver.\n");
1777 		}
1778 	}
1779 
1780 	RTE_LCORE_FOREACH_WORKER(lcore_id)
1781 		rte_eal_wait_lcore(lcore_id);
1782 
1783 	/* clean up the EAL */
1784 	rte_eal_cleanup();
1785 
1786 	return 0;
1787 }
1788