drivers/net/memif/rte_eth_memif.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2018-2019 Cisco Systems, Inc.  All rights reserved.
   3  */
   4
   5 #include <stdlib.h>
   6 #include <fcntl.h>
   7 #include <unistd.h>
   8 #include <sys/types.h>
   9 #include <sys/socket.h>
  10 #include <sys/un.h>
  11 #include <sys/ioctl.h>
  12 #include <sys/mman.h>
  13 #include <linux/if_ether.h>
  14 #include <errno.h>
  15 #include <sys/eventfd.h>
  16
  17 #include <rte_version.h>
  18 #include <rte_mbuf.h>
  19 #include <rte_ether.h>
  20 #include <rte_ethdev_driver.h>
  21 #include <rte_ethdev_vdev.h>
  22 #include <rte_malloc.h>
  23 #include <rte_kvargs.h>
  24 #include <rte_bus_vdev.h>
  25 #include <rte_string_fns.h>
  26 #include <rte_errno.h>
  27 #include <rte_memory.h>
  28 #include <rte_memzone.h>
  29 #include <rte_eal_memconfig.h>
  30
  31 #include "rte_eth_memif.h"
  32 #include "memif_socket.h"
  33
  34 #define ETH_MEMIF_ID_ARG                "id"
  35 #define ETH_MEMIF_ROLE_ARG              "role"
  36 #define ETH_MEMIF_PKT_BUFFER_SIZE_ARG   "bsize"
  37 #define ETH_MEMIF_RING_SIZE_ARG         "rsize"
  38 #define ETH_MEMIF_SOCKET_ARG            "socket"
  39 #define ETH_MEMIF_SOCKET_ABSTRACT_ARG   "socket-abstract"
  40 #define ETH_MEMIF_MAC_ARG               "mac"
  41 #define ETH_MEMIF_ZC_ARG                "zero-copy"
  42 #define ETH_MEMIF_SECRET_ARG            "secret"
  43
  44 static const char * const valid_arguments[] = {
  45         ETH_MEMIF_ID_ARG,
  46         ETH_MEMIF_ROLE_ARG,
  47         ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
  48         ETH_MEMIF_RING_SIZE_ARG,
  49         ETH_MEMIF_SOCKET_ARG,
  50         ETH_MEMIF_SOCKET_ABSTRACT_ARG,
  51         ETH_MEMIF_MAC_ARG,
  52         ETH_MEMIF_ZC_ARG,
  53         ETH_MEMIF_SECRET_ARG,
  54         NULL
  55 };
  56
  57 static const struct rte_eth_link pmd_link = {
  58         .link_speed = ETH_SPEED_NUM_10G,
  59         .link_duplex = ETH_LINK_FULL_DUPLEX,
  60         .link_status = ETH_LINK_DOWN,
  61         .link_autoneg = ETH_LINK_AUTONEG
  62 };
  63
  64 #define MEMIF_MP_SEND_REGION            "memif_mp_send_region"
  65
  66
  67 static int memif_region_init_zc(const struct rte_memseg_list *msl,
  68                                 const struct rte_memseg *ms, void *arg);
  69
  70 const char *
  71 memif_version(void)
  72 {
  73         return ("memif-" RTE_STR(MEMIF_VERSION_MAJOR) "." RTE_STR(MEMIF_VERSION_MINOR));
  74 }
  75
  76 /* Message header to synchronize regions */
  77 struct mp_region_msg {
  78         char port_name[RTE_DEV_NAME_MAX_LEN];
  79         memif_region_index_t idx;
  80         memif_region_size_t size;
  81 };
  82
  83 static int
  84 memif_mp_send_region(const struct rte_mp_msg *msg, const void *peer)
  85 {
  86         struct rte_eth_dev *dev;
  87         struct pmd_process_private *proc_private;
  88         const struct mp_region_msg *msg_param = (const struct mp_region_msg *)msg->param;
  89         struct rte_mp_msg reply;
  90         struct mp_region_msg *reply_param = (struct mp_region_msg *)reply.param;
  91         uint16_t port_id;
  92         int ret;
  93
  94         /* Get requested port */
  95         ret = rte_eth_dev_get_port_by_name(msg_param->port_name, &port_id);
  96         if (ret) {
  97                 MIF_LOG(ERR, "Failed to get port id for %s",
  98                         msg_param->port_name);
  99                 return -1;
 100         }
 101         dev = &rte_eth_devices[port_id];
 102         proc_private = dev->process_private;
 103
 104         memset(&reply, 0, sizeof(reply));
 105         strlcpy(reply.name, msg->name, sizeof(reply.name));
 106         reply_param->idx = msg_param->idx;
 107         if (proc_private->regions[msg_param->idx] != NULL) {
 108                 reply_param->size = proc_private->regions[msg_param->idx]->region_size;
 109                 reply.fds[0] = proc_private->regions[msg_param->idx]->fd;
 110                 reply.num_fds = 1;
 111         }
 112         reply.len_param = sizeof(*reply_param);
 113         if (rte_mp_reply(&reply, peer) < 0) {
 114                 MIF_LOG(ERR, "Failed to reply to an add region request");
 115                 return -1;
 116         }
 117
 118         return 0;
 119 }
 120
 121 /*
 122  * Request regions
 123  * Called by secondary process, when ports link status goes up.
 124  */
 125 static int
 126 memif_mp_request_regions(struct rte_eth_dev *dev)
 127 {
 128         int ret, i;
 129         struct timespec timeout = {.tv_sec = 5, .tv_nsec = 0};
 130         struct rte_mp_msg msg, *reply;
 131         struct rte_mp_reply replies;
 132         struct mp_region_msg *msg_param = (struct mp_region_msg *)msg.param;
 133         struct mp_region_msg *reply_param;
 134         struct memif_region *r;
 135         struct pmd_process_private *proc_private = dev->process_private;
 136         struct pmd_internals *pmd = dev->data->dev_private;
 137         /* in case of zero-copy slave, only request region 0 */
 138         uint16_t max_region_num = (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) ?
 139                                    1 : ETH_MEMIF_MAX_REGION_NUM;
 140
 141         MIF_LOG(DEBUG, "Requesting memory regions");
 142
 143         for (i = 0; i < max_region_num; i++) {
 144                 /* Prepare the message */
 145                 memset(&msg, 0, sizeof(msg));
 146                 strlcpy(msg.name, MEMIF_MP_SEND_REGION, sizeof(msg.name));
 147                 strlcpy(msg_param->port_name, dev->data->name,
 148                         sizeof(msg_param->port_name));
 149                 msg_param->idx = i;
 150                 msg.len_param = sizeof(*msg_param);
 151
 152                 /* Send message */
 153                 ret = rte_mp_request_sync(&msg, &replies, &timeout);
 154                 if (ret < 0 || replies.nb_received != 1) {
 155                         MIF_LOG(ERR, "Failed to send mp msg: %d",
 156                                 rte_errno);
 157                         return -1;
 158                 }
 159
 160                 reply = &replies.msgs[0];
 161                 reply_param = (struct mp_region_msg *)reply->param;
 162
 163                 if (reply_param->size > 0) {
 164                         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 165                         if (r == NULL) {
 166                                 MIF_LOG(ERR, "Failed to alloc memif region.");
 167                                 free(reply);
 168                                 return -ENOMEM;
 169                         }
 170                         r->region_size = reply_param->size;
 171                         if (reply->num_fds < 1) {
 172                                 MIF_LOG(ERR, "Missing file descriptor.");
 173                                 free(reply);
 174                                 return -1;
 175                         }
 176                         r->fd = reply->fds[0];
 177                         r->addr = NULL;
 178
 179                         proc_private->regions[reply_param->idx] = r;
 180                         proc_private->regions_num++;
 181                 }
 182                 free(reply);
 183         }
 184
 185         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
 186                 ret = rte_memseg_walk(memif_region_init_zc, (void *)proc_private);
 187                 if (ret < 0)
 188                         return ret;
 189         }
 190
 191         return memif_connect(dev);
 192 }
 193
 194 static int
 195 memif_dev_info(struct rte_eth_dev *dev __rte_unused, struct rte_eth_dev_info *dev_info)
 196 {
 197         dev_info->max_mac_addrs = 1;
 198         dev_info->max_rx_pktlen = (uint32_t)ETH_FRAME_LEN;
 199         dev_info->max_rx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
 200         dev_info->max_tx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
 201         dev_info->min_rx_bufsize = 0;
 202
 203         return 0;
 204 }
 205
 206 static memif_ring_t *
 207 memif_get_ring(struct pmd_internals *pmd, struct pmd_process_private *proc_private,
 208                memif_ring_type_t type, uint16_t ring_num)
 209 {
 210         /* rings only in region 0 */
 211         void *p = proc_private->regions[0]->addr;
 212         int ring_size = sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 213             (1 << pmd->run.log2_ring_size);
 214
 215         p = (uint8_t *)p + (ring_num + type * pmd->run.num_s2m_rings) * ring_size;
 216
 217         return (memif_ring_t *)p;
 218 }
 219
 220 static memif_region_offset_t
 221 memif_get_ring_offset(struct rte_eth_dev *dev, struct memif_queue *mq,
 222                       memif_ring_type_t type, uint16_t num)
 223 {
 224         struct pmd_internals *pmd = dev->data->dev_private;
 225         struct pmd_process_private *proc_private = dev->process_private;
 226
 227         return ((uint8_t *)memif_get_ring(pmd, proc_private, type, num) -
 228                 (uint8_t *)proc_private->regions[mq->region]->addr);
 229 }
 230
 231 static memif_ring_t *
 232 memif_get_ring_from_queue(struct pmd_process_private *proc_private,
 233                           struct memif_queue *mq)
 234 {
 235         struct memif_region *r;
 236
 237         r = proc_private->regions[mq->region];
 238         if (r == NULL)
 239                 return NULL;
 240
 241         return (memif_ring_t *)((uint8_t *)r->addr + mq->ring_offset);
 242 }
 243
 244 static void *
 245 memif_get_buffer(struct pmd_process_private *proc_private, memif_desc_t *d)
 246 {
 247         return ((uint8_t *)proc_private->regions[d->region]->addr + d->offset);
 248 }
 249
 250 /* Free mbufs received by master */
 251 static void
 252 memif_free_stored_mbufs(struct pmd_process_private *proc_private, struct memif_queue *mq)
 253 {
 254         uint16_t cur_tail;
 255         uint16_t mask = (1 << mq->log2_ring_size) - 1;
 256         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 257
 258         /* FIXME: improve performance */
 259         /* The ring->tail acts as a guard variable between Tx and Rx
 260          * threads, so using load-acquire pairs with store-release
 261          * in function eth_memif_rx for S2M queues.
 262          */
 263         cur_tail = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 264         while (mq->last_tail != cur_tail) {
 265                 RTE_MBUF_PREFETCH_TO_FREE(mq->buffers[(mq->last_tail + 1) & mask]);
 266                 /* Decrement refcnt and free mbuf. (current segment) */
 267                 rte_mbuf_refcnt_update(mq->buffers[mq->last_tail & mask], -1);
 268                 rte_pktmbuf_free_seg(mq->buffers[mq->last_tail & mask]);
 269                 mq->last_tail++;
 270         }
 271 }
 272
 273 static int
 274 memif_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *cur_tail,
 275                     struct rte_mbuf *tail)
 276 {
 277         /* Check for number-of-segments-overflow */
 278         if (unlikely(head->nb_segs + tail->nb_segs > RTE_MBUF_MAX_NB_SEGS))
 279                 return -EOVERFLOW;
 280
 281         /* Chain 'tail' onto the old tail */
 282         cur_tail->next = tail;
 283
 284         /* accumulate number of segments and total length. */
 285         head->nb_segs = (uint16_t)(head->nb_segs + tail->nb_segs);
 286
 287         tail->pkt_len = tail->data_len;
 288         head->pkt_len += tail->pkt_len;
 289
 290         return 0;
 291 }
 292
 293 static uint16_t
 294 eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 295 {
 296         struct memif_queue *mq = queue;
 297         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 298         struct pmd_process_private *proc_private =
 299                 rte_eth_devices[mq->in_port].process_private;
 300         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 301         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0;
 302         uint16_t n_rx_pkts = 0;
 303         uint16_t mbuf_size = rte_pktmbuf_data_room_size(mq->mempool) -
 304                 RTE_PKTMBUF_HEADROOM;
 305         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 306         memif_ring_type_t type = mq->type;
 307         memif_desc_t *d0;
 308         struct rte_mbuf *mbuf, *mbuf_head, *mbuf_tail;
 309         uint64_t b;
 310         ssize_t size __rte_unused;
 311         uint16_t head;
 312         int ret;
 313         struct rte_eth_link link;
 314
 315         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 316                 return 0;
 317         if (unlikely(ring == NULL)) {
 318                 /* Secondary process will attempt to request regions. */
 319                 ret = rte_eth_link_get(mq->in_port, &link);
 320                 if (ret < 0)
 321                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 322                                 mq->in_port, rte_strerror(-ret));
 323                 return 0;
 324         }
 325
 326         /* consume interrupt */
 327         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0)
 328                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 329
 330         ring_size = 1 << mq->log2_ring_size;
 331         mask = ring_size - 1;
 332
 333         if (type == MEMIF_RING_S2M) {
 334                 cur_slot = mq->last_head;
 335                 last_slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
 336         } else {
 337                 cur_slot = mq->last_tail;
 338                 last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 339         }
 340
 341         if (cur_slot == last_slot)
 342                 goto refill;
 343         n_slots = last_slot - cur_slot;
 344
 345         while (n_slots && n_rx_pkts < nb_pkts) {
 346                 mbuf_head = rte_pktmbuf_alloc(mq->mempool);
 347                 if (unlikely(mbuf_head == NULL))
 348                         goto no_free_bufs;
 349                 mbuf = mbuf_head;
 350                 mbuf->port = mq->in_port;
 351
 352 next_slot:
 353                 s0 = cur_slot & mask;
 354                 d0 = &ring->desc[s0];
 355
 356                 src_len = d0->length;
 357                 dst_off = 0;
 358                 src_off = 0;
 359
 360                 do {
 361                         dst_len = mbuf_size - dst_off;
 362                         if (dst_len == 0) {
 363                                 dst_off = 0;
 364                                 dst_len = mbuf_size;
 365
 366                                 /* store pointer to tail */
 367                                 mbuf_tail = mbuf;
 368                                 mbuf = rte_pktmbuf_alloc(mq->mempool);
 369                                 if (unlikely(mbuf == NULL))
 370                                         goto no_free_bufs;
 371                                 mbuf->port = mq->in_port;
 372                                 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 373                                 if (unlikely(ret < 0)) {
 374                                         MIF_LOG(ERR, "number-of-segments-overflow");
 375                                         rte_pktmbuf_free(mbuf);
 376                                         goto no_free_bufs;
 377                                 }
 378                         }
 379                         cp_len = RTE_MIN(dst_len, src_len);
 380
 381                         rte_pktmbuf_data_len(mbuf) += cp_len;
 382                         rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 383                         if (mbuf != mbuf_head)
 384                                 rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
 385
 386                         memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
 387                                (uint8_t *)memif_get_buffer(proc_private, d0) + src_off,
 388                                cp_len);
 389
 390                         src_off += cp_len;
 391                         dst_off += cp_len;
 392                         src_len -= cp_len;
 393                 } while (src_len);
 394
 395                 cur_slot++;
 396                 n_slots--;
 397
 398                 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
 399                         goto next_slot;
 400
 401                 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
 402                 *bufs++ = mbuf_head;
 403                 n_rx_pkts++;
 404         }
 405
 406 no_free_bufs:
 407         if (type == MEMIF_RING_S2M) {
 408                 __atomic_store_n(&ring->tail, cur_slot, __ATOMIC_RELEASE);
 409                 mq->last_head = cur_slot;
 410         } else {
 411                 mq->last_tail = cur_slot;
 412         }
 413
 414 refill:
 415         if (type == MEMIF_RING_M2S) {
 416                 /* ring->head is updated by the receiver and this function
 417                  * is called in the context of receiver thread. The loads in
 418                  * the receiver do not need to synchronize with its own stores.
 419                  */
 420                 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 421                 n_slots = ring_size - head + mq->last_tail;
 422
 423                 while (n_slots--) {
 424                         s0 = head++ & mask;
 425                         d0 = &ring->desc[s0];
 426                         d0->length = pmd->run.pkt_buffer_size;
 427                 }
 428                 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 429         }
 430
 431         mq->n_pkts += n_rx_pkts;
 432         return n_rx_pkts;
 433 }
 434
 435 static uint16_t
 436 eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 437 {
 438         struct memif_queue *mq = queue;
 439         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 440         struct pmd_process_private *proc_private =
 441                 rte_eth_devices[mq->in_port].process_private;
 442         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 443         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0, head;
 444         uint16_t n_rx_pkts = 0;
 445         memif_desc_t *d0;
 446         struct rte_mbuf *mbuf, *mbuf_tail;
 447         struct rte_mbuf *mbuf_head = NULL;
 448         int ret;
 449         struct rte_eth_link link;
 450
 451         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 452                 return 0;
 453         if (unlikely(ring == NULL)) {
 454                 /* Secondary process will attempt to request regions. */
 455                 rte_eth_link_get(mq->in_port, &link);
 456                 return 0;
 457         }
 458
 459         /* consume interrupt */
 460         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 461                 uint64_t b;
 462                 ssize_t size __rte_unused;
 463                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 464         }
 465
 466         ring_size = 1 << mq->log2_ring_size;
 467         mask = ring_size - 1;
 468
 469         cur_slot = mq->last_tail;
 470         /* The ring->tail acts as a guard variable between Tx and Rx
 471          * threads, so using load-acquire pairs with store-release
 472          * to synchronize it between threads.
 473          */
 474         last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 475         if (cur_slot == last_slot)
 476                 goto refill;
 477         n_slots = last_slot - cur_slot;
 478
 479         while (n_slots && n_rx_pkts < nb_pkts) {
 480                 s0 = cur_slot & mask;
 481
 482                 d0 = &ring->desc[s0];
 483                 mbuf_head = mq->buffers[s0];
 484                 mbuf = mbuf_head;
 485
 486 next_slot:
 487                 /* prefetch next descriptor */
 488                 if (n_rx_pkts + 1 < nb_pkts)
 489                         rte_prefetch0(&ring->desc[(cur_slot + 1) & mask]);
 490
 491                 mbuf->port = mq->in_port;
 492                 rte_pktmbuf_data_len(mbuf) = d0->length;
 493                 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 494
 495                 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 496
 497                 cur_slot++;
 498                 n_slots--;
 499                 if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
 500                         s0 = cur_slot & mask;
 501                         d0 = &ring->desc[s0];
 502                         mbuf_tail = mbuf;
 503                         mbuf = mq->buffers[s0];
 504                         ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 505                         if (unlikely(ret < 0)) {
 506                                 MIF_LOG(ERR, "number-of-segments-overflow");
 507                                 goto refill;
 508                         }
 509                         goto next_slot;
 510                 }
 511
 512                 *bufs++ = mbuf_head;
 513                 n_rx_pkts++;
 514         }
 515
 516         mq->last_tail = cur_slot;
 517
 518 /* Supply master with new buffers */
 519 refill:
 520         /* ring->head is updated by the receiver and this function
 521          * is called in the context of receiver thread. The loads in
 522          * the receiver do not need to synchronize with its own stores.
 523          */
 524         head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 525         n_slots = ring_size - head + mq->last_tail;
 526
 527         if (n_slots < 32)
 528                 goto no_free_mbufs;
 529
 530         ret = rte_pktmbuf_alloc_bulk(mq->mempool, &mq->buffers[head & mask], n_slots);
 531         if (unlikely(ret < 0))
 532                 goto no_free_mbufs;
 533
 534         while (n_slots--) {
 535                 s0 = head++ & mask;
 536                 if (n_slots > 0)
 537                         rte_prefetch0(mq->buffers[head & mask]);
 538                 d0 = &ring->desc[s0];
 539                 /* store buffer header */
 540                 mbuf = mq->buffers[s0];
 541                 /* populate descriptor */
 542                 d0->length = rte_pktmbuf_data_room_size(mq->mempool) -
 543                                 RTE_PKTMBUF_HEADROOM;
 544                 d0->region = 1;
 545                 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 546                         (uint8_t *)proc_private->regions[d0->region]->addr;
 547         }
 548 no_free_mbufs:
 549         /* The ring->head acts as a guard variable between Tx and Rx
 550          * threads, so using store-release pairs with load-acquire
 551          * in function eth_memif_tx.
 552          */
 553         __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 554
 555         mq->n_pkts += n_rx_pkts;
 556
 557         return n_rx_pkts;
 558 }
 559
 560 static uint16_t
 561 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 562 {
 563         struct memif_queue *mq = queue;
 564         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 565         struct pmd_process_private *proc_private =
 566                 rte_eth_devices[mq->in_port].process_private;
 567         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 568         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 569         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 570         memif_ring_type_t type = mq->type;
 571         memif_desc_t *d0;
 572         struct rte_mbuf *mbuf;
 573         struct rte_mbuf *mbuf_head;
 574         uint64_t a;
 575         ssize_t size;
 576         struct rte_eth_link link;
 577
 578         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 579                 return 0;
 580         if (unlikely(ring == NULL)) {
 581                 int ret;
 582
 583                 /* Secondary process will attempt to request regions. */
 584                 ret = rte_eth_link_get(mq->in_port, &link);
 585                 if (ret < 0)
 586                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 587                                 mq->in_port, rte_strerror(-ret));
 588                 return 0;
 589         }
 590
 591         ring_size = 1 << mq->log2_ring_size;
 592         mask = ring_size - 1;
 593
 594         if (type == MEMIF_RING_S2M) {
 595                 /* For S2M queues ring->head is updated by the sender and
 596                  * this function is called in the context of sending thread.
 597                  * The loads in the sender do not need to synchronize with
 598                  * its own stores. Hence, the following load can be a
 599                  * relaxed load.
 600                  */
 601                 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 602                 n_free = ring_size - slot +
 603                                 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 604         } else {
 605                 /* For M2S queues ring->tail is updated by the sender and
 606                  * this function is called in the context of sending thread.
 607                  * The loads in the sender do not need to synchronize with
 608                  * its own stores. Hence, the following load can be a
 609                  * relaxed load.
 610                  */
 611                 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 612                 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 613         }
 614
 615         while (n_tx_pkts < nb_pkts && n_free) {
 616                 mbuf_head = *bufs++;
 617                 mbuf = mbuf_head;
 618
 619                 saved_slot = slot;
 620                 d0 = &ring->desc[slot & mask];
 621                 dst_off = 0;
 622                 dst_len = (type == MEMIF_RING_S2M) ?
 623                         pmd->run.pkt_buffer_size : d0->length;
 624
 625 next_in_chain:
 626                 src_off = 0;
 627                 src_len = rte_pktmbuf_data_len(mbuf);
 628
 629                 while (src_len) {
 630                         if (dst_len == 0) {
 631                                 if (n_free) {
 632                                         slot++;
 633                                         n_free--;
 634                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 635                                         d0 = &ring->desc[slot & mask];
 636                                         dst_off = 0;
 637                                         dst_len = (type == MEMIF_RING_S2M) ?
 638                                             pmd->run.pkt_buffer_size : d0->length;
 639                                         d0->flags = 0;
 640                                 } else {
 641                                         slot = saved_slot;
 642                                         goto no_free_slots;
 643                                 }
 644                         }
 645                         cp_len = RTE_MIN(dst_len, src_len);
 646
 647                         memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
 648                                rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 649                                cp_len);
 650
 651                         mq->n_bytes += cp_len;
 652                         src_off += cp_len;
 653                         dst_off += cp_len;
 654                         src_len -= cp_len;
 655                         dst_len -= cp_len;
 656
 657                         d0->length = dst_off;
 658                 }
 659
 660                 if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 661                         mbuf = mbuf->next;
 662                         goto next_in_chain;
 663                 }
 664
 665                 n_tx_pkts++;
 666                 slot++;
 667                 n_free--;
 668                 rte_pktmbuf_free(mbuf_head);
 669         }
 670
 671 no_free_slots:
 672         if (type == MEMIF_RING_S2M)
 673                 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 674         else
 675                 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
 676
 677         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 678                 a = 1;
 679                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 680                 if (unlikely(size < 0)) {
 681                         MIF_LOG(WARNING,
 682                                 "Failed to send interrupt. %s", strerror(errno));
 683                 }
 684         }
 685
 686         mq->n_pkts += n_tx_pkts;
 687         return n_tx_pkts;
 688 }
 689
 690
 691 static int
 692 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
 693                 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
 694                 uint16_t slot, uint16_t n_free)
 695 {
 696         memif_desc_t *d0;
 697         int used_slots = 1;
 698
 699 next_in_chain:
 700         /* store pointer to mbuf to free it later */
 701         mq->buffers[slot & mask] = mbuf;
 702         /* Increment refcnt to make sure the buffer is not freed before master
 703          * receives it. (current segment)
 704          */
 705         rte_mbuf_refcnt_update(mbuf, 1);
 706         /* populate descriptor */
 707         d0 = &ring->desc[slot & mask];
 708         d0->length = rte_pktmbuf_data_len(mbuf);
 709         /* FIXME: get region index */
 710         d0->region = 1;
 711         d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 712                 (uint8_t *)proc_private->regions[d0->region]->addr;
 713         d0->flags = 0;
 714
 715         /* check if buffer is chained */
 716         if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 717                 if (n_free < 2)
 718                         return 0;
 719                 /* mark buffer as chained */
 720                 d0->flags |= MEMIF_DESC_FLAG_NEXT;
 721                 /* advance mbuf */
 722                 mbuf = mbuf->next;
 723                 /* update counters */
 724                 used_slots++;
 725                 slot++;
 726                 n_free--;
 727                 goto next_in_chain;
 728         }
 729         return used_slots;
 730 }
 731
 732 static uint16_t
 733 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 734 {
 735         struct memif_queue *mq = queue;
 736         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 737         struct pmd_process_private *proc_private =
 738                 rte_eth_devices[mq->in_port].process_private;
 739         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 740         uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
 741         struct rte_eth_link link;
 742
 743         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 744                 return 0;
 745         if (unlikely(ring == NULL)) {
 746                 /* Secondary process will attempt to request regions. */
 747                 rte_eth_link_get(mq->in_port, &link);
 748                 return 0;
 749         }
 750
 751         ring_size = 1 << mq->log2_ring_size;
 752         mask = ring_size - 1;
 753
 754         /* free mbufs received by master */
 755         memif_free_stored_mbufs(proc_private, mq);
 756
 757         /* ring type always MEMIF_RING_S2M */
 758         /* For S2M queues ring->head is updated by the sender and
 759          * this function is called in the context of sending thread.
 760          * The loads in the sender do not need to synchronize with
 761          * its own stores. Hence, the following load can be a
 762          * relaxed load.
 763          */
 764         slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 765         n_free = ring_size - slot + mq->last_tail;
 766
 767         int used_slots;
 768
 769         while (n_free && (n_tx_pkts < nb_pkts)) {
 770                 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
 771                         if ((nb_pkts - n_tx_pkts) > 8) {
 772                                 rte_prefetch0(*bufs + 4);
 773                                 rte_prefetch0(*bufs + 5);
 774                                 rte_prefetch0(*bufs + 6);
 775                                 rte_prefetch0(*bufs + 7);
 776                         }
 777                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 778                                 mask, slot, n_free);
 779                         if (unlikely(used_slots < 1))
 780                                 goto no_free_slots;
 781                         n_tx_pkts++;
 782                         slot += used_slots;
 783                         n_free -= used_slots;
 784
 785                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 786                                 mask, slot, n_free);
 787                         if (unlikely(used_slots < 1))
 788                                 goto no_free_slots;
 789                         n_tx_pkts++;
 790                         slot += used_slots;
 791                         n_free -= used_slots;
 792
 793                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 794                                 mask, slot, n_free);
 795                         if (unlikely(used_slots < 1))
 796                                 goto no_free_slots;
 797                         n_tx_pkts++;
 798                         slot += used_slots;
 799                         n_free -= used_slots;
 800
 801                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 802                                 mask, slot, n_free);
 803                         if (unlikely(used_slots < 1))
 804                                 goto no_free_slots;
 805                         n_tx_pkts++;
 806                         slot += used_slots;
 807                         n_free -= used_slots;
 808                 }
 809                 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 810                         mask, slot, n_free);
 811                 if (unlikely(used_slots < 1))
 812                         goto no_free_slots;
 813                 n_tx_pkts++;
 814                 slot += used_slots;
 815                 n_free -= used_slots;
 816         }
 817
 818 no_free_slots:
 819         /* ring type always MEMIF_RING_S2M */
 820         /* The ring->head acts as a guard variable between Tx and Rx
 821          * threads, so using store-release pairs with load-acquire
 822          * in function eth_memif_rx for S2M rings.
 823          */
 824         __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 825
 826         /* Send interrupt, if enabled. */
 827         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 828                 uint64_t a = 1;
 829                 ssize_t size = write(mq->intr_handle.fd, &a, sizeof(a));
 830                 if (unlikely(size < 0)) {
 831                         MIF_LOG(WARNING,
 832                                 "Failed to send interrupt. %s", strerror(errno));
 833                 }
 834         }
 835
 836         /* increment queue counters */
 837         mq->n_pkts += n_tx_pkts;
 838
 839         return n_tx_pkts;
 840 }
 841
 842 void
 843 memif_free_regions(struct rte_eth_dev *dev)
 844 {
 845         struct pmd_process_private *proc_private = dev->process_private;
 846         struct pmd_internals *pmd = dev->data->dev_private;
 847         int i;
 848         struct memif_region *r;
 849
 850         /* regions are allocated contiguously, so it's
 851          * enough to loop until 'proc_private->regions_num'
 852          */
 853         for (i = 0; i < proc_private->regions_num; i++) {
 854                 r = proc_private->regions[i];
 855                 if (r != NULL) {
 856                         /* This is memzone */
 857                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
 858                                 r->addr = NULL;
 859                                 if (r->fd > 0)
 860                                         close(r->fd);
 861                         }
 862                         if (r->addr != NULL) {
 863                                 munmap(r->addr, r->region_size);
 864                                 if (r->fd > 0) {
 865                                         close(r->fd);
 866                                         r->fd = -1;
 867                                 }
 868                         }
 869                         rte_free(r);
 870                         proc_private->regions[i] = NULL;
 871                 }
 872         }
 873         proc_private->regions_num = 0;
 874 }
 875
 876 static int
 877 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 878                      void *arg)
 879 {
 880         struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
 881         struct memif_region *r;
 882
 883         if (proc_private->regions_num < 1) {
 884                 MIF_LOG(ERR, "Missing descriptor region");
 885                 return -1;
 886         }
 887
 888         r = proc_private->regions[proc_private->regions_num - 1];
 889
 890         if (r->addr != msl->base_va)
 891                 r = proc_private->regions[++proc_private->regions_num - 1];
 892
 893         if (r == NULL) {
 894                 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 895                 if (r == NULL) {
 896                         MIF_LOG(ERR, "Failed to alloc memif region.");
 897                         return -ENOMEM;
 898                 }
 899
 900                 r->addr = msl->base_va;
 901                 r->region_size = ms->len;
 902                 r->fd = rte_memseg_get_fd(ms);
 903                 if (r->fd < 0)
 904                         return -1;
 905                 r->pkt_buffer_offset = 0;
 906
 907                 proc_private->regions[proc_private->regions_num - 1] = r;
 908         } else {
 909                 r->region_size += ms->len;
 910         }
 911
 912         return 0;
 913 }
 914
 915 static int
 916 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
 917 {
 918         struct pmd_internals *pmd = dev->data->dev_private;
 919         struct pmd_process_private *proc_private = dev->process_private;
 920         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 921         int ret = 0;
 922         struct memif_region *r;
 923
 924         if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 925                 MIF_LOG(ERR, "Too many regions.");
 926                 return -1;
 927         }
 928
 929         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 930         if (r == NULL) {
 931                 MIF_LOG(ERR, "Failed to alloc memif region.");
 932                 return -ENOMEM;
 933         }
 934
 935         /* calculate buffer offset */
 936         r->pkt_buffer_offset = (pmd->run.num_s2m_rings + pmd->run.num_m2s_rings) *
 937             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 938             (1 << pmd->run.log2_ring_size));
 939
 940         r->region_size = r->pkt_buffer_offset;
 941         /* if region has buffers, add buffers size to region_size */
 942         if (has_buffers == 1)
 943                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 944                         (1 << pmd->run.log2_ring_size) *
 945                         (pmd->run.num_s2m_rings +
 946                          pmd->run.num_m2s_rings));
 947
 948         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 949         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 950                  proc_private->regions_num);
 951
 952         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 953         if (r->fd < 0) {
 954                 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
 955                 ret = -1;
 956                 goto error;
 957         }
 958
 959         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 960         if (ret < 0) {
 961                 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
 962                 goto error;
 963         }
 964
 965         ret = ftruncate(r->fd, r->region_size);
 966         if (ret < 0) {
 967                 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
 968                 goto error;
 969         }
 970
 971         r->addr = mmap(NULL, r->region_size, PROT_READ |
 972                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 973         if (r->addr == MAP_FAILED) {
 974                 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
 975                 ret = -1;
 976                 goto error;
 977         }
 978
 979         proc_private->regions[proc_private->regions_num] = r;
 980         proc_private->regions_num++;
 981
 982         return ret;
 983
 984 error:
 985         if (r->fd > 0)
 986                 close(r->fd);
 987         r->fd = -1;
 988
 989         return ret;
 990 }
 991
 992 static int
 993 memif_regions_init(struct rte_eth_dev *dev)
 994 {
 995         struct pmd_internals *pmd = dev->data->dev_private;
 996         int ret;
 997
 998         /*
 999          * Zero-copy exposes dpdk memory.
1000          * Each memseg list will be represented by memif region.
1001          * Zero-copy regions indexing: memseg list idx + 1,
1002          * as we already have region 0 reserved for descriptors.
1003          */
1004         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1005                 /* create region idx 0 containing descriptors */
1006                 ret = memif_region_init_shm(dev, 0);
1007                 if (ret < 0)
1008                         return ret;
1009                 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1010                 if (ret < 0)
1011                         return ret;
1012         } else {
1013                 /* create one memory region contaning rings and buffers */
1014                 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1015                 if (ret < 0)
1016                         return ret;
1017         }
1018
1019         return 0;
1020 }
1021
1022 static void
1023 memif_init_rings(struct rte_eth_dev *dev)
1024 {
1025         struct pmd_internals *pmd = dev->data->dev_private;
1026         struct pmd_process_private *proc_private = dev->process_private;
1027         memif_ring_t *ring;
1028         int i, j;
1029         uint16_t slot;
1030
1031         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1032                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2M, i);
1033                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1034                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1035                 ring->cookie = MEMIF_COOKIE;
1036                 ring->flags = 0;
1037
1038                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1039                         continue;
1040
1041                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1042                         slot = i * (1 << pmd->run.log2_ring_size) + j;
1043                         ring->desc[j].region = 0;
1044                         ring->desc[j].offset =
1045                                 proc_private->regions[0]->pkt_buffer_offset +
1046                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1047                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1048                 }
1049         }
1050
1051         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1052                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_M2S, i);
1053                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1054                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1055                 ring->cookie = MEMIF_COOKIE;
1056                 ring->flags = 0;
1057
1058                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1059                         continue;
1060
1061                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1062                         slot = (i + pmd->run.num_s2m_rings) *
1063                             (1 << pmd->run.log2_ring_size) + j;
1064                         ring->desc[j].region = 0;
1065                         ring->desc[j].offset =
1066                                 proc_private->regions[0]->pkt_buffer_offset +
1067                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1068                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1069                 }
1070         }
1071 }
1072
1073 /* called only by slave */
1074 static int
1075 memif_init_queues(struct rte_eth_dev *dev)
1076 {
1077         struct pmd_internals *pmd = dev->data->dev_private;
1078         struct memif_queue *mq;
1079         int i;
1080
1081         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1082                 mq = dev->data->tx_queues[i];
1083                 mq->log2_ring_size = pmd->run.log2_ring_size;
1084                 /* queues located only in region 0 */
1085                 mq->region = 0;
1086                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2M, i);
1087                 mq->last_head = 0;
1088                 mq->last_tail = 0;
1089                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1090                 if (mq->intr_handle.fd < 0) {
1091                         MIF_LOG(WARNING,
1092                                 "Failed to create eventfd for tx queue %d: %s.", i,
1093                                 strerror(errno));
1094                 }
1095                 mq->buffers = NULL;
1096                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1097                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1098                                                   (1 << mq->log2_ring_size), 0);
1099                         if (mq->buffers == NULL)
1100                                 return -ENOMEM;
1101                 }
1102         }
1103
1104         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1105                 mq = dev->data->rx_queues[i];
1106                 mq->log2_ring_size = pmd->run.log2_ring_size;
1107                 /* queues located only in region 0 */
1108                 mq->region = 0;
1109                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_M2S, i);
1110                 mq->last_head = 0;
1111                 mq->last_tail = 0;
1112                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1113                 if (mq->intr_handle.fd < 0) {
1114                         MIF_LOG(WARNING,
1115                                 "Failed to create eventfd for rx queue %d: %s.", i,
1116                                 strerror(errno));
1117                 }
1118                 mq->buffers = NULL;
1119                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1120                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1121                                                   (1 << mq->log2_ring_size), 0);
1122                         if (mq->buffers == NULL)
1123                                 return -ENOMEM;
1124                 }
1125         }
1126         return 0;
1127 }
1128
1129 int
1130 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1131 {
1132         int ret;
1133
1134         ret = memif_regions_init(dev);
1135         if (ret < 0)
1136                 return ret;
1137
1138         memif_init_rings(dev);
1139
1140         ret = memif_init_queues(dev);
1141         if (ret < 0)
1142                 return ret;
1143
1144         return 0;
1145 }
1146
1147 int
1148 memif_connect(struct rte_eth_dev *dev)
1149 {
1150         struct pmd_internals *pmd = dev->data->dev_private;
1151         struct pmd_process_private *proc_private = dev->process_private;
1152         struct memif_region *mr;
1153         struct memif_queue *mq;
1154         memif_ring_t *ring;
1155         int i;
1156
1157         for (i = 0; i < proc_private->regions_num; i++) {
1158                 mr = proc_private->regions[i];
1159                 if (mr != NULL) {
1160                         if (mr->addr == NULL) {
1161                                 if (mr->fd < 0)
1162                                         return -1;
1163                                 mr->addr = mmap(NULL, mr->region_size,
1164                                                 PROT_READ | PROT_WRITE,
1165                                                 MAP_SHARED, mr->fd, 0);
1166                                 if (mr->addr == MAP_FAILED) {
1167                                         MIF_LOG(ERR, "mmap failed: %s\n",
1168                                                 strerror(errno));
1169                                         return -1;
1170                                 }
1171                         }
1172                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1173                                 /* close memseg file */
1174                                 close(mr->fd);
1175                                 mr->fd = -1;
1176                         }
1177                 }
1178         }
1179
1180         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1181                 for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1182                         mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
1183                             dev->data->tx_queues[i] : dev->data->rx_queues[i];
1184                         ring = memif_get_ring_from_queue(proc_private, mq);
1185                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1186                                 MIF_LOG(ERR, "Wrong ring");
1187                                 return -1;
1188                         }
1189                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1190                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1191                         mq->last_head = 0;
1192                         mq->last_tail = 0;
1193                         /* enable polling mode */
1194                         if (pmd->role == MEMIF_ROLE_MASTER)
1195                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1196                 }
1197                 for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1198                         mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
1199                             dev->data->rx_queues[i] : dev->data->tx_queues[i];
1200                         ring = memif_get_ring_from_queue(proc_private, mq);
1201                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1202                                 MIF_LOG(ERR, "Wrong ring");
1203                                 return -1;
1204                         }
1205                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1206                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1207                         mq->last_head = 0;
1208                         mq->last_tail = 0;
1209                         /* enable polling mode */
1210                         if (pmd->role == MEMIF_ROLE_SLAVE)
1211                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1212                 }
1213
1214                 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1215                 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1216                 dev->data->dev_link.link_status = ETH_LINK_UP;
1217         }
1218         MIF_LOG(INFO, "Connected.");
1219         return 0;
1220 }
1221
1222 static int
1223 memif_dev_start(struct rte_eth_dev *dev)
1224 {
1225         struct pmd_internals *pmd = dev->data->dev_private;
1226         int ret = 0;
1227
1228         switch (pmd->role) {
1229         case MEMIF_ROLE_SLAVE:
1230                 ret = memif_connect_slave(dev);
1231                 break;
1232         case MEMIF_ROLE_MASTER:
1233                 ret = memif_connect_master(dev);
1234                 break;
1235         default:
1236                 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1237                 ret = -1;
1238                 break;
1239         }
1240
1241         return ret;
1242 }
1243
1244 static int
1245 memif_dev_close(struct rte_eth_dev *dev)
1246 {
1247         struct pmd_internals *pmd = dev->data->dev_private;
1248         int i;
1249
1250         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1251                 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1252                 memif_disconnect(dev);
1253
1254                 for (i = 0; i < dev->data->nb_rx_queues; i++)
1255                         (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
1256                 for (i = 0; i < dev->data->nb_tx_queues; i++)
1257                         (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
1258
1259                 memif_socket_remove_device(dev);
1260         } else {
1261                 memif_disconnect(dev);
1262         }
1263
1264         rte_free(dev->process_private);
1265
1266         return 0;
1267 }
1268
1269 static int
1270 memif_dev_configure(struct rte_eth_dev *dev)
1271 {
1272         struct pmd_internals *pmd = dev->data->dev_private;
1273
1274         /*
1275          * SLAVE - TXQ
1276          * MASTER - RXQ
1277          */
1278         pmd->cfg.num_s2m_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
1279                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1280
1281         /*
1282          * SLAVE - RXQ
1283          * MASTER - TXQ
1284          */
1285         pmd->cfg.num_m2s_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
1286                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1287
1288         return 0;
1289 }
1290
1291 static int
1292 memif_tx_queue_setup(struct rte_eth_dev *dev,
1293                      uint16_t qid,
1294                      uint16_t nb_tx_desc __rte_unused,
1295                      unsigned int socket_id __rte_unused,
1296                      const struct rte_eth_txconf *tx_conf __rte_unused)
1297 {
1298         struct pmd_internals *pmd = dev->data->dev_private;
1299         struct memif_queue *mq;
1300
1301         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1302         if (mq == NULL) {
1303                 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1304                 return -ENOMEM;
1305         }
1306
1307         mq->type =
1308             (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_S2M : MEMIF_RING_M2S;
1309         mq->n_pkts = 0;
1310         mq->n_bytes = 0;
1311         mq->intr_handle.fd = -1;
1312         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1313         mq->in_port = dev->data->port_id;
1314         dev->data->tx_queues[qid] = mq;
1315
1316         return 0;
1317 }
1318
1319 static int
1320 memif_rx_queue_setup(struct rte_eth_dev *dev,
1321                      uint16_t qid,
1322                      uint16_t nb_rx_desc __rte_unused,
1323                      unsigned int socket_id __rte_unused,
1324                      const struct rte_eth_rxconf *rx_conf __rte_unused,
1325                      struct rte_mempool *mb_pool)
1326 {
1327         struct pmd_internals *pmd = dev->data->dev_private;
1328         struct memif_queue *mq;
1329
1330         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1331         if (mq == NULL) {
1332                 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1333                 return -ENOMEM;
1334         }
1335
1336         mq->type = (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_M2S : MEMIF_RING_S2M;
1337         mq->n_pkts = 0;
1338         mq->n_bytes = 0;
1339         mq->intr_handle.fd = -1;
1340         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1341         mq->mempool = mb_pool;
1342         mq->in_port = dev->data->port_id;
1343         dev->data->rx_queues[qid] = mq;
1344
1345         return 0;
1346 }
1347
1348 static void
1349 memif_queue_release(void *queue)
1350 {
1351         struct memif_queue *mq = (struct memif_queue *)queue;
1352
1353         if (!mq)
1354                 return;
1355
1356         rte_free(mq);
1357 }
1358
1359 static int
1360 memif_link_update(struct rte_eth_dev *dev,
1361                   int wait_to_complete __rte_unused)
1362 {
1363         struct pmd_process_private *proc_private;
1364
1365         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1366                 proc_private = dev->process_private;
1367                 if (dev->data->dev_link.link_status == ETH_LINK_UP &&
1368                                 proc_private->regions_num == 0) {
1369                         memif_mp_request_regions(dev);
1370                 } else if (dev->data->dev_link.link_status == ETH_LINK_DOWN &&
1371                                 proc_private->regions_num > 0) {
1372                         memif_free_regions(dev);
1373                 }
1374         }
1375         return 0;
1376 }
1377
1378 static int
1379 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1380 {
1381         struct pmd_internals *pmd = dev->data->dev_private;
1382         struct memif_queue *mq;
1383         int i;
1384         uint8_t tmp, nq;
1385
1386         stats->ipackets = 0;
1387         stats->ibytes = 0;
1388         stats->opackets = 0;
1389         stats->obytes = 0;
1390
1391         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_s2m_rings :
1392             pmd->run.num_m2s_rings;
1393         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1394             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1395
1396         /* RX stats */
1397         for (i = 0; i < nq; i++) {
1398                 mq = dev->data->rx_queues[i];
1399                 stats->q_ipackets[i] = mq->n_pkts;
1400                 stats->q_ibytes[i] = mq->n_bytes;
1401                 stats->ipackets += mq->n_pkts;
1402                 stats->ibytes += mq->n_bytes;
1403         }
1404
1405         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_m2s_rings :
1406             pmd->run.num_s2m_rings;
1407         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1408             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1409
1410         /* TX stats */
1411         for (i = 0; i < nq; i++) {
1412                 mq = dev->data->tx_queues[i];
1413                 stats->q_opackets[i] = mq->n_pkts;
1414                 stats->q_obytes[i] = mq->n_bytes;
1415                 stats->opackets += mq->n_pkts;
1416                 stats->obytes += mq->n_bytes;
1417         }
1418         return 0;
1419 }
1420
1421 static int
1422 memif_stats_reset(struct rte_eth_dev *dev)
1423 {
1424         struct pmd_internals *pmd = dev->data->dev_private;
1425         int i;
1426         struct memif_queue *mq;
1427
1428         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1429                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->tx_queues[i] :
1430                     dev->data->rx_queues[i];
1431                 mq->n_pkts = 0;
1432                 mq->n_bytes = 0;
1433         }
1434         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1435                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->rx_queues[i] :
1436                     dev->data->tx_queues[i];
1437                 mq->n_pkts = 0;
1438                 mq->n_bytes = 0;
1439         }
1440
1441         return 0;
1442 }
1443
1444 static int
1445 memif_rx_queue_intr_enable(struct rte_eth_dev *dev __rte_unused,
1446                            uint16_t qid __rte_unused)
1447 {
1448         MIF_LOG(WARNING, "Interrupt mode not supported.");
1449
1450         return -1;
1451 }
1452
1453 static int
1454 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
1455 {
1456         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
1457
1458         return 0;
1459 }
1460
1461 static const struct eth_dev_ops ops = {
1462         .dev_start = memif_dev_start,
1463         .dev_close = memif_dev_close,
1464         .dev_infos_get = memif_dev_info,
1465         .dev_configure = memif_dev_configure,
1466         .tx_queue_setup = memif_tx_queue_setup,
1467         .rx_queue_setup = memif_rx_queue_setup,
1468         .rx_queue_release = memif_queue_release,
1469         .tx_queue_release = memif_queue_release,
1470         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
1471         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
1472         .link_update = memif_link_update,
1473         .stats_get = memif_stats_get,
1474         .stats_reset = memif_stats_reset,
1475 };
1476
1477 static int
1478 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1479              memif_interface_id_t id, uint32_t flags,
1480              const char *socket_filename,
1481              memif_log2_ring_size_t log2_ring_size,
1482              uint16_t pkt_buffer_size, const char *secret,
1483              struct rte_ether_addr *ether_addr)
1484 {
1485         int ret = 0;
1486         struct rte_eth_dev *eth_dev;
1487         struct rte_eth_dev_data *data;
1488         struct pmd_internals *pmd;
1489         struct pmd_process_private *process_private;
1490         const unsigned int numa_node = vdev->device.numa_node;
1491         const char *name = rte_vdev_device_name(vdev);
1492
1493         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1494         if (eth_dev == NULL) {
1495                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1496                 return -1;
1497         }
1498
1499         process_private = (struct pmd_process_private *)
1500                 rte_zmalloc(name, sizeof(struct pmd_process_private),
1501                             RTE_CACHE_LINE_SIZE);
1502
1503         if (process_private == NULL) {
1504                 MIF_LOG(ERR, "Failed to alloc memory for process private");
1505                 return -1;
1506         }
1507         eth_dev->process_private = process_private;
1508
1509         pmd = eth_dev->data->dev_private;
1510         memset(pmd, 0, sizeof(*pmd));
1511
1512         pmd->id = id;
1513         pmd->flags = flags;
1514         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1515         pmd->role = role;
1516         /* Zero-copy flag irelevant to master. */
1517         if (pmd->role == MEMIF_ROLE_MASTER)
1518                 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1519
1520         ret = memif_socket_init(eth_dev, socket_filename);
1521         if (ret < 0)
1522                 return ret;
1523
1524         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1525         if (secret != NULL)
1526                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1527
1528         pmd->cfg.log2_ring_size = log2_ring_size;
1529         /* set in .dev_configure() */
1530         pmd->cfg.num_s2m_rings = 0;
1531         pmd->cfg.num_m2s_rings = 0;
1532
1533         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1534         rte_spinlock_init(&pmd->cc_lock);
1535
1536         data = eth_dev->data;
1537         data->dev_private = pmd;
1538         data->numa_node = numa_node;
1539         data->dev_link = pmd_link;
1540         data->mac_addrs = ether_addr;
1541         data->promiscuous = 1;
1542         data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1543
1544         eth_dev->dev_ops = &ops;
1545         eth_dev->device = &vdev->device;
1546         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1547                 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1548                 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1549         } else {
1550                 eth_dev->rx_pkt_burst = eth_memif_rx;
1551                 eth_dev->tx_pkt_burst = eth_memif_tx;
1552         }
1553
1554         rte_eth_dev_probing_finish(eth_dev);
1555
1556         return 0;
1557 }
1558
1559 static int
1560 memif_set_role(const char *key __rte_unused, const char *value,
1561                void *extra_args)
1562 {
1563         enum memif_role_t *role = (enum memif_role_t *)extra_args;
1564
1565         if (strstr(value, "master") != NULL) {
1566                 *role = MEMIF_ROLE_MASTER;
1567         } else if (strstr(value, "slave") != NULL) {
1568                 *role = MEMIF_ROLE_SLAVE;
1569         } else {
1570                 MIF_LOG(ERR, "Unknown role: %s.", value);
1571                 return -EINVAL;
1572         }
1573         return 0;
1574 }
1575
1576 static int
1577 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1578 {
1579         uint32_t *flags = (uint32_t *)extra_args;
1580
1581         if (strstr(value, "yes") != NULL) {
1582                 if (!rte_mcfg_get_single_file_segments()) {
1583                         MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1584                         return -ENOTSUP;
1585                 }
1586                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1587         } else if (strstr(value, "no") != NULL) {
1588                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1589         } else {
1590                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1591                 return -EINVAL;
1592         }
1593         return 0;
1594 }
1595
1596 static int
1597 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1598 {
1599         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1600
1601         /* even if parsing fails, 0 is a valid id */
1602         *id = strtoul(value, NULL, 10);
1603         return 0;
1604 }
1605
1606 static int
1607 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1608 {
1609         unsigned long tmp;
1610         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1611
1612         tmp = strtoul(value, NULL, 10);
1613         if (tmp == 0 || tmp > 0xFFFF) {
1614                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1615                 return -EINVAL;
1616         }
1617         *pkt_buffer_size = tmp;
1618         return 0;
1619 }
1620
1621 static int
1622 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1623 {
1624         unsigned long tmp;
1625         memif_log2_ring_size_t *log2_ring_size =
1626             (memif_log2_ring_size_t *)extra_args;
1627
1628         tmp = strtoul(value, NULL, 10);
1629         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1630                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1631                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1632                 return -EINVAL;
1633         }
1634         *log2_ring_size = tmp;
1635         return 0;
1636 }
1637
1638 /* check if directory exists and if we have permission to read/write */
1639 static int
1640 memif_check_socket_filename(const char *filename)
1641 {
1642         char *dir = NULL, *tmp;
1643         uint32_t idx;
1644         int ret = 0;
1645
1646         if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1647                 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1648                 return -1;
1649         }
1650
1651         tmp = strrchr(filename, '/');
1652         if (tmp != NULL) {
1653                 idx = tmp - filename;
1654                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1655                 if (dir == NULL) {
1656                         MIF_LOG(ERR, "Failed to allocate memory.");
1657                         return -1;
1658                 }
1659                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1660         }
1661
1662         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1663                                         W_OK, AT_EACCESS) < 0)) {
1664                 MIF_LOG(ERR, "Invalid socket directory.");
1665                 ret = -EINVAL;
1666         }
1667
1668         if (dir != NULL)
1669                 rte_free(dir);
1670
1671         return ret;
1672 }
1673
1674 static int
1675 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1676                           void *extra_args)
1677 {
1678         const char **socket_filename = (const char **)extra_args;
1679
1680         *socket_filename = value;
1681         return 0;
1682 }
1683
1684 static int
1685 memif_set_is_socket_abstract(const char *key __rte_unused, const char *value, void *extra_args)
1686 {
1687         uint32_t *flags = (uint32_t *)extra_args;
1688
1689         if (strstr(value, "yes") != NULL) {
1690                 *flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1691         } else if (strstr(value, "no") != NULL) {
1692                 *flags &= ~ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1693         } else {
1694                 MIF_LOG(ERR, "Failed to parse socket-abstract param: %s.", value);
1695                 return -EINVAL;
1696         }
1697         return 0;
1698 }
1699
1700 static int
1701 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1702 {
1703         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1704
1705         if (rte_ether_unformat_addr(value, ether_addr) < 0)
1706                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1707         return 0;
1708 }
1709
1710 static int
1711 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1712 {
1713         const char **secret = (const char **)extra_args;
1714
1715         *secret = value;
1716         return 0;
1717 }
1718
1719 static int
1720 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1721 {
1722         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1723         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1724         int ret = 0;
1725         struct rte_kvargs *kvlist;
1726         const char *name = rte_vdev_device_name(vdev);
1727         enum memif_role_t role = MEMIF_ROLE_SLAVE;
1728         memif_interface_id_t id = 0;
1729         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1730         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1731         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1732         uint32_t flags = 0;
1733         const char *secret = NULL;
1734         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1735                 sizeof(struct rte_ether_addr), 0);
1736         struct rte_eth_dev *eth_dev;
1737
1738         rte_eth_random_addr(ether_addr->addr_bytes);
1739
1740         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1741
1742         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1743                 eth_dev = rte_eth_dev_attach_secondary(name);
1744                 if (!eth_dev) {
1745                         MIF_LOG(ERR, "Failed to probe %s", name);
1746                         return -1;
1747                 }
1748
1749                 eth_dev->dev_ops = &ops;
1750                 eth_dev->device = &vdev->device;
1751                 eth_dev->rx_pkt_burst = eth_memif_rx;
1752                 eth_dev->tx_pkt_burst = eth_memif_tx;
1753
1754                 if (!rte_eal_primary_proc_alive(NULL)) {
1755                         MIF_LOG(ERR, "Primary process is missing");
1756                         return -1;
1757                 }
1758
1759                 eth_dev->process_private = (struct pmd_process_private *)
1760                         rte_zmalloc(name,
1761                                 sizeof(struct pmd_process_private),
1762                                 RTE_CACHE_LINE_SIZE);
1763                 if (eth_dev->process_private == NULL) {
1764                         MIF_LOG(ERR,
1765                                 "Failed to alloc memory for process private");
1766                         return -1;
1767                 }
1768
1769                 rte_eth_dev_probing_finish(eth_dev);
1770
1771                 return 0;
1772         }
1773
1774         ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1775         /*
1776          * Primary process can continue probing, but secondary process won't
1777          * be able to get memory regions information
1778          */
1779         if (ret < 0 && rte_errno != EEXIST)
1780                 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1781                         strerror(rte_errno));
1782
1783         /* use abstract address by default */
1784         flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1785
1786         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1787
1788         /* parse parameters */
1789         if (kvlist != NULL) {
1790                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1791                                          &memif_set_role, &role);
1792                 if (ret < 0)
1793                         goto exit;
1794                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1795                                          &memif_set_id, &id);
1796                 if (ret < 0)
1797                         goto exit;
1798                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1799                                          &memif_set_bs, &pkt_buffer_size);
1800                 if (ret < 0)
1801                         goto exit;
1802                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1803                                          &memif_set_rs, &log2_ring_size);
1804                 if (ret < 0)
1805                         goto exit;
1806                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1807                                          &memif_set_socket_filename,
1808                                          (void *)(&socket_filename));
1809                 if (ret < 0)
1810                         goto exit;
1811                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ABSTRACT_ARG,
1812                                          &memif_set_is_socket_abstract, &flags);
1813                 if (ret < 0)
1814                         goto exit;
1815                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1816                                          &memif_set_mac, ether_addr);
1817                 if (ret < 0)
1818                         goto exit;
1819                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1820                                          &memif_set_zc, &flags);
1821                 if (ret < 0)
1822                         goto exit;
1823                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1824                                          &memif_set_secret, (void *)(&secret));
1825                 if (ret < 0)
1826                         goto exit;
1827         }
1828
1829         if (!(flags & ETH_MEMIF_FLAG_SOCKET_ABSTRACT)) {
1830                 ret = memif_check_socket_filename(socket_filename);
1831                 if (ret < 0)
1832                         goto exit;
1833         }
1834
1835         /* create interface */
1836         ret = memif_create(vdev, role, id, flags, socket_filename,
1837                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1838
1839 exit:
1840         if (kvlist != NULL)
1841                 rte_kvargs_free(kvlist);
1842         return ret;
1843 }
1844
1845 static int
1846 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1847 {
1848         struct rte_eth_dev *eth_dev;
1849
1850         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1851         if (eth_dev == NULL)
1852                 return 0;
1853
1854         return rte_eth_dev_close(eth_dev->data->port_id);
1855 }
1856
1857 static struct rte_vdev_driver pmd_memif_drv = {
1858         .probe = rte_pmd_memif_probe,
1859         .remove = rte_pmd_memif_remove,
1860 };
1861
1862 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1863
1864 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1865                               ETH_MEMIF_ID_ARG "=<int>"
1866                               ETH_MEMIF_ROLE_ARG "=master|slave"
1867                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1868                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1869                               ETH_MEMIF_SOCKET_ARG "=<string>"
1870                                   ETH_MEMIF_SOCKET_ABSTRACT_ARG "=yes|no"
1871                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1872                               ETH_MEMIF_ZC_ARG "=yes|no"
1873                               ETH_MEMIF_SECRET_ARG "=<string>");
1874
1875 RTE_LOG_REGISTER(memif_logtype, pmd.net.memif, NOTICE);