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