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         /* ring->head is updated by the receiver and this function
 518          * is called in the context of receiver thread. The loads in
 519          * the receiver do not need to synchronize with its own stores.
 520          */
 521         head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 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         /* The ring->head acts as a guard variable between Tx and Rx
 547          * threads, so using store-release pairs with load-acquire
 548          * in function eth_memif_tx.
 549          */
 550         __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 551
 552         mq->n_pkts += n_rx_pkts;
 553
 554         return n_rx_pkts;
 555 }
 556
 557 static uint16_t
 558 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 559 {
 560         struct memif_queue *mq = queue;
 561         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 562         struct pmd_process_private *proc_private =
 563                 rte_eth_devices[mq->in_port].process_private;
 564         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 565         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 566         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 567         memif_ring_type_t type = mq->type;
 568         memif_desc_t *d0;
 569         struct rte_mbuf *mbuf;
 570         struct rte_mbuf *mbuf_head;
 571         uint64_t a;
 572         ssize_t size;
 573         struct rte_eth_link link;
 574
 575         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 576                 return 0;
 577         if (unlikely(ring == NULL)) {
 578                 int ret;
 579
 580                 /* Secondary process will attempt to request regions. */
 581                 ret = rte_eth_link_get(mq->in_port, &link);
 582                 if (ret < 0)
 583                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 584                                 mq->in_port, rte_strerror(-ret));
 585                 return 0;
 586         }
 587
 588         ring_size = 1 << mq->log2_ring_size;
 589         mask = ring_size - 1;
 590
 591         if (type == MEMIF_RING_S2M) {
 592                 /* For S2M queues ring->head is updated by the sender and
 593                  * this function is called in the context of sending thread.
 594                  * The loads in the sender do not need to synchronize with
 595                  * its own stores. Hence, the following load can be a
 596                  * relaxed load.
 597                  */
 598                 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 599                 n_free = ring_size - slot +
 600                                 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 601         } else {
 602                 /* For M2S queues ring->tail is updated by the sender and
 603                  * this function is called in the context of sending thread.
 604                  * The loads in the sender do not need to synchronize with
 605                  * its own stores. Hence, the following load can be a
 606                  * relaxed load.
 607                  */
 608                 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 609                 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 610         }
 611
 612         while (n_tx_pkts < nb_pkts && n_free) {
 613                 mbuf_head = *bufs++;
 614                 mbuf = mbuf_head;
 615
 616                 saved_slot = slot;
 617                 d0 = &ring->desc[slot & mask];
 618                 dst_off = 0;
 619                 dst_len = (type == MEMIF_RING_S2M) ?
 620                         pmd->run.pkt_buffer_size : d0->length;
 621
 622 next_in_chain:
 623                 src_off = 0;
 624                 src_len = rte_pktmbuf_data_len(mbuf);
 625
 626                 while (src_len) {
 627                         if (dst_len == 0) {
 628                                 if (n_free) {
 629                                         slot++;
 630                                         n_free--;
 631                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 632                                         d0 = &ring->desc[slot & mask];
 633                                         dst_off = 0;
 634                                         dst_len = (type == MEMIF_RING_S2M) ?
 635                                             pmd->run.pkt_buffer_size : d0->length;
 636                                         d0->flags = 0;
 637                                 } else {
 638                                         slot = saved_slot;
 639                                         goto no_free_slots;
 640                                 }
 641                         }
 642                         cp_len = RTE_MIN(dst_len, src_len);
 643
 644                         memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
 645                                rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 646                                cp_len);
 647
 648                         mq->n_bytes += cp_len;
 649                         src_off += cp_len;
 650                         dst_off += cp_len;
 651                         src_len -= cp_len;
 652                         dst_len -= cp_len;
 653
 654                         d0->length = dst_off;
 655                 }
 656
 657                 if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 658                         mbuf = mbuf->next;
 659                         goto next_in_chain;
 660                 }
 661
 662                 n_tx_pkts++;
 663                 slot++;
 664                 n_free--;
 665                 rte_pktmbuf_free(mbuf_head);
 666         }
 667
 668 no_free_slots:
 669         if (type == MEMIF_RING_S2M)
 670                 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 671         else
 672                 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
 673
 674         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 675                 a = 1;
 676                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 677                 if (unlikely(size < 0)) {
 678                         MIF_LOG(WARNING,
 679                                 "Failed to send interrupt. %s", strerror(errno));
 680                 }
 681         }
 682
 683         mq->n_pkts += n_tx_pkts;
 684         return n_tx_pkts;
 685 }
 686
 687
 688 static int
 689 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
 690                 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
 691                 uint16_t slot, uint16_t n_free)
 692 {
 693         memif_desc_t *d0;
 694         int used_slots = 1;
 695
 696 next_in_chain:
 697         /* store pointer to mbuf to free it later */
 698         mq->buffers[slot & mask] = mbuf;
 699         /* Increment refcnt to make sure the buffer is not freed before master
 700          * receives it. (current segment)
 701          */
 702         rte_mbuf_refcnt_update(mbuf, 1);
 703         /* populate descriptor */
 704         d0 = &ring->desc[slot & mask];
 705         d0->length = rte_pktmbuf_data_len(mbuf);
 706         /* FIXME: get region index */
 707         d0->region = 1;
 708         d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 709                 (uint8_t *)proc_private->regions[d0->region]->addr;
 710         d0->flags = 0;
 711
 712         /* check if buffer is chained */
 713         if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 714                 if (n_free < 2)
 715                         return 0;
 716                 /* mark buffer as chained */
 717                 d0->flags |= MEMIF_DESC_FLAG_NEXT;
 718                 /* advance mbuf */
 719                 mbuf = mbuf->next;
 720                 /* update counters */
 721                 used_slots++;
 722                 slot++;
 723                 n_free--;
 724                 goto next_in_chain;
 725         }
 726         return used_slots;
 727 }
 728
 729 static uint16_t
 730 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 731 {
 732         struct memif_queue *mq = queue;
 733         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 734         struct pmd_process_private *proc_private =
 735                 rte_eth_devices[mq->in_port].process_private;
 736         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 737         uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
 738         struct rte_eth_link link;
 739
 740         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 741                 return 0;
 742         if (unlikely(ring == NULL)) {
 743                 /* Secondary process will attempt to request regions. */
 744                 rte_eth_link_get(mq->in_port, &link);
 745                 return 0;
 746         }
 747
 748         ring_size = 1 << mq->log2_ring_size;
 749         mask = ring_size - 1;
 750
 751         /* free mbufs received by master */
 752         memif_free_stored_mbufs(proc_private, mq);
 753
 754         /* ring type always MEMIF_RING_S2M */
 755         /* For S2M queues ring->head is updated by the sender and
 756          * this function is called in the context of sending thread.
 757          * The loads in the sender do not need to synchronize with
 758          * its own stores. Hence, the following load can be a
 759          * relaxed load.
 760          */
 761         slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 762         n_free = ring_size - slot + mq->last_tail;
 763
 764         int used_slots;
 765
 766         while (n_free && (n_tx_pkts < nb_pkts)) {
 767                 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
 768                         if ((nb_pkts - n_tx_pkts) > 8) {
 769                                 rte_prefetch0(*bufs + 4);
 770                                 rte_prefetch0(*bufs + 5);
 771                                 rte_prefetch0(*bufs + 6);
 772                                 rte_prefetch0(*bufs + 7);
 773                         }
 774                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 775                                 mask, slot, n_free);
 776                         if (unlikely(used_slots < 1))
 777                                 goto no_free_slots;
 778                         n_tx_pkts++;
 779                         slot += used_slots;
 780                         n_free -= used_slots;
 781
 782                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 783                                 mask, slot, n_free);
 784                         if (unlikely(used_slots < 1))
 785                                 goto no_free_slots;
 786                         n_tx_pkts++;
 787                         slot += used_slots;
 788                         n_free -= used_slots;
 789
 790                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 791                                 mask, slot, n_free);
 792                         if (unlikely(used_slots < 1))
 793                                 goto no_free_slots;
 794                         n_tx_pkts++;
 795                         slot += used_slots;
 796                         n_free -= used_slots;
 797
 798                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 799                                 mask, slot, n_free);
 800                         if (unlikely(used_slots < 1))
 801                                 goto no_free_slots;
 802                         n_tx_pkts++;
 803                         slot += used_slots;
 804                         n_free -= used_slots;
 805                 }
 806                 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 807                         mask, slot, n_free);
 808                 if (unlikely(used_slots < 1))
 809                         goto no_free_slots;
 810                 n_tx_pkts++;
 811                 slot += used_slots;
 812                 n_free -= used_slots;
 813         }
 814
 815 no_free_slots:
 816         /* ring type always MEMIF_RING_S2M */
 817         /* The ring->head acts as a guard variable between Tx and Rx
 818          * threads, so using store-release pairs with load-acquire
 819          * in function eth_memif_rx for S2M rings.
 820          */
 821         __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 822
 823         /* Send interrupt, if enabled. */
 824         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 825                 uint64_t a = 1;
 826                 ssize_t size = write(mq->intr_handle.fd, &a, sizeof(a));
 827                 if (unlikely(size < 0)) {
 828                         MIF_LOG(WARNING,
 829                                 "Failed to send interrupt. %s", strerror(errno));
 830                 }
 831         }
 832
 833         /* increment queue counters */
 834         mq->n_pkts += n_tx_pkts;
 835
 836         return n_tx_pkts;
 837 }
 838
 839 void
 840 memif_free_regions(struct rte_eth_dev *dev)
 841 {
 842         struct pmd_process_private *proc_private = dev->process_private;
 843         struct pmd_internals *pmd = dev->data->dev_private;
 844         int i;
 845         struct memif_region *r;
 846
 847         /* regions are allocated contiguously, so it's
 848          * enough to loop until 'proc_private->regions_num'
 849          */
 850         for (i = 0; i < proc_private->regions_num; i++) {
 851                 r = proc_private->regions[i];
 852                 if (r != NULL) {
 853                         /* This is memzone */
 854                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
 855                                 r->addr = NULL;
 856                                 if (r->fd > 0)
 857                                         close(r->fd);
 858                         }
 859                         if (r->addr != NULL) {
 860                                 munmap(r->addr, r->region_size);
 861                                 if (r->fd > 0) {
 862                                         close(r->fd);
 863                                         r->fd = -1;
 864                                 }
 865                         }
 866                         rte_free(r);
 867                         proc_private->regions[i] = NULL;
 868                 }
 869         }
 870         proc_private->regions_num = 0;
 871 }
 872
 873 static int
 874 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 875                      void *arg)
 876 {
 877         struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
 878         struct memif_region *r;
 879
 880         if (proc_private->regions_num < 1) {
 881                 MIF_LOG(ERR, "Missing descriptor region");
 882                 return -1;
 883         }
 884
 885         r = proc_private->regions[proc_private->regions_num - 1];
 886
 887         if (r->addr != msl->base_va)
 888                 r = proc_private->regions[++proc_private->regions_num - 1];
 889
 890         if (r == NULL) {
 891                 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 892                 if (r == NULL) {
 893                         MIF_LOG(ERR, "Failed to alloc memif region.");
 894                         return -ENOMEM;
 895                 }
 896
 897                 r->addr = msl->base_va;
 898                 r->region_size = ms->len;
 899                 r->fd = rte_memseg_get_fd(ms);
 900                 if (r->fd < 0)
 901                         return -1;
 902                 r->pkt_buffer_offset = 0;
 903
 904                 proc_private->regions[proc_private->regions_num - 1] = r;
 905         } else {
 906                 r->region_size += ms->len;
 907         }
 908
 909         return 0;
 910 }
 911
 912 static int
 913 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
 914 {
 915         struct pmd_internals *pmd = dev->data->dev_private;
 916         struct pmd_process_private *proc_private = dev->process_private;
 917         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 918         int ret = 0;
 919         struct memif_region *r;
 920
 921         if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 922                 MIF_LOG(ERR, "Too many regions.");
 923                 return -1;
 924         }
 925
 926         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 927         if (r == NULL) {
 928                 MIF_LOG(ERR, "Failed to alloc memif region.");
 929                 return -ENOMEM;
 930         }
 931
 932         /* calculate buffer offset */
 933         r->pkt_buffer_offset = (pmd->run.num_s2m_rings + pmd->run.num_m2s_rings) *
 934             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 935             (1 << pmd->run.log2_ring_size));
 936
 937         r->region_size = r->pkt_buffer_offset;
 938         /* if region has buffers, add buffers size to region_size */
 939         if (has_buffers == 1)
 940                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 941                         (1 << pmd->run.log2_ring_size) *
 942                         (pmd->run.num_s2m_rings +
 943                          pmd->run.num_m2s_rings));
 944
 945         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 946         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 947                  proc_private->regions_num);
 948
 949         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 950         if (r->fd < 0) {
 951                 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
 952                 ret = -1;
 953                 goto error;
 954         }
 955
 956         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 957         if (ret < 0) {
 958                 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
 959                 goto error;
 960         }
 961
 962         ret = ftruncate(r->fd, r->region_size);
 963         if (ret < 0) {
 964                 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
 965                 goto error;
 966         }
 967
 968         r->addr = mmap(NULL, r->region_size, PROT_READ |
 969                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 970         if (r->addr == MAP_FAILED) {
 971                 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
 972                 ret = -1;
 973                 goto error;
 974         }
 975
 976         proc_private->regions[proc_private->regions_num] = r;
 977         proc_private->regions_num++;
 978
 979         return ret;
 980
 981 error:
 982         if (r->fd > 0)
 983                 close(r->fd);
 984         r->fd = -1;
 985
 986         return ret;
 987 }
 988
 989 static int
 990 memif_regions_init(struct rte_eth_dev *dev)
 991 {
 992         struct pmd_internals *pmd = dev->data->dev_private;
 993         int ret;
 994
 995         /*
 996          * Zero-copy exposes dpdk memory.
 997          * Each memseg list will be represented by memif region.
 998          * Zero-copy regions indexing: memseg list idx + 1,
 999          * as we already have region 0 reserved for descriptors.
1000          */
1001         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1002                 /* create region idx 0 containing descriptors */
1003                 ret = memif_region_init_shm(dev, 0);
1004                 if (ret < 0)
1005                         return ret;
1006                 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1007                 if (ret < 0)
1008                         return ret;
1009         } else {
1010                 /* create one memory region contaning rings and buffers */
1011                 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1012                 if (ret < 0)
1013                         return ret;
1014         }
1015
1016         return 0;
1017 }
1018
1019 static void
1020 memif_init_rings(struct rte_eth_dev *dev)
1021 {
1022         struct pmd_internals *pmd = dev->data->dev_private;
1023         struct pmd_process_private *proc_private = dev->process_private;
1024         memif_ring_t *ring;
1025         int i, j;
1026         uint16_t slot;
1027
1028         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1029                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2M, i);
1030                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1031                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1032                 ring->cookie = MEMIF_COOKIE;
1033                 ring->flags = 0;
1034
1035                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1036                         continue;
1037
1038                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1039                         slot = i * (1 << pmd->run.log2_ring_size) + j;
1040                         ring->desc[j].region = 0;
1041                         ring->desc[j].offset =
1042                                 proc_private->regions[0]->pkt_buffer_offset +
1043                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1044                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1045                 }
1046         }
1047
1048         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1049                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_M2S, i);
1050                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1051                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1052                 ring->cookie = MEMIF_COOKIE;
1053                 ring->flags = 0;
1054
1055                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1056                         continue;
1057
1058                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1059                         slot = (i + pmd->run.num_s2m_rings) *
1060                             (1 << pmd->run.log2_ring_size) + j;
1061                         ring->desc[j].region = 0;
1062                         ring->desc[j].offset =
1063                                 proc_private->regions[0]->pkt_buffer_offset +
1064                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1065                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1066                 }
1067         }
1068 }
1069
1070 /* called only by slave */
1071 static int
1072 memif_init_queues(struct rte_eth_dev *dev)
1073 {
1074         struct pmd_internals *pmd = dev->data->dev_private;
1075         struct memif_queue *mq;
1076         int i;
1077
1078         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1079                 mq = dev->data->tx_queues[i];
1080                 mq->log2_ring_size = pmd->run.log2_ring_size;
1081                 /* queues located only in region 0 */
1082                 mq->region = 0;
1083                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2M, i);
1084                 mq->last_head = 0;
1085                 mq->last_tail = 0;
1086                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1087                 if (mq->intr_handle.fd < 0) {
1088                         MIF_LOG(WARNING,
1089                                 "Failed to create eventfd for tx queue %d: %s.", i,
1090                                 strerror(errno));
1091                 }
1092                 mq->buffers = NULL;
1093                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1094                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1095                                                   (1 << mq->log2_ring_size), 0);
1096                         if (mq->buffers == NULL)
1097                                 return -ENOMEM;
1098                 }
1099         }
1100
1101         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1102                 mq = dev->data->rx_queues[i];
1103                 mq->log2_ring_size = pmd->run.log2_ring_size;
1104                 /* queues located only in region 0 */
1105                 mq->region = 0;
1106                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_M2S, i);
1107                 mq->last_head = 0;
1108                 mq->last_tail = 0;
1109                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1110                 if (mq->intr_handle.fd < 0) {
1111                         MIF_LOG(WARNING,
1112                                 "Failed to create eventfd for rx queue %d: %s.", i,
1113                                 strerror(errno));
1114                 }
1115                 mq->buffers = NULL;
1116                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1117                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1118                                                   (1 << mq->log2_ring_size), 0);
1119                         if (mq->buffers == NULL)
1120                                 return -ENOMEM;
1121                 }
1122         }
1123         return 0;
1124 }
1125
1126 int
1127 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1128 {
1129         int ret;
1130
1131         ret = memif_regions_init(dev);
1132         if (ret < 0)
1133                 return ret;
1134
1135         memif_init_rings(dev);
1136
1137         ret = memif_init_queues(dev);
1138         if (ret < 0)
1139                 return ret;
1140
1141         return 0;
1142 }
1143
1144 int
1145 memif_connect(struct rte_eth_dev *dev)
1146 {
1147         struct pmd_internals *pmd = dev->data->dev_private;
1148         struct pmd_process_private *proc_private = dev->process_private;
1149         struct memif_region *mr;
1150         struct memif_queue *mq;
1151         memif_ring_t *ring;
1152         int i;
1153
1154         for (i = 0; i < proc_private->regions_num; i++) {
1155                 mr = proc_private->regions[i];
1156                 if (mr != NULL) {
1157                         if (mr->addr == NULL) {
1158                                 if (mr->fd < 0)
1159                                         return -1;
1160                                 mr->addr = mmap(NULL, mr->region_size,
1161                                                 PROT_READ | PROT_WRITE,
1162                                                 MAP_SHARED, mr->fd, 0);
1163                                 if (mr->addr == MAP_FAILED) {
1164                                         MIF_LOG(ERR, "mmap failed: %s\n",
1165                                                 strerror(errno));
1166                                         return -1;
1167                                 }
1168                         }
1169                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1170                                 /* close memseg file */
1171                                 close(mr->fd);
1172                                 mr->fd = -1;
1173                         }
1174                 }
1175         }
1176
1177         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1178                 for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1179                         mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
1180                             dev->data->tx_queues[i] : dev->data->rx_queues[i];
1181                         ring = memif_get_ring_from_queue(proc_private, mq);
1182                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1183                                 MIF_LOG(ERR, "Wrong ring");
1184                                 return -1;
1185                         }
1186                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1187                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1188                         mq->last_head = 0;
1189                         mq->last_tail = 0;
1190                         /* enable polling mode */
1191                         if (pmd->role == MEMIF_ROLE_MASTER)
1192                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1193                 }
1194                 for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1195                         mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
1196                             dev->data->rx_queues[i] : dev->data->tx_queues[i];
1197                         ring = memif_get_ring_from_queue(proc_private, mq);
1198                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1199                                 MIF_LOG(ERR, "Wrong ring");
1200                                 return -1;
1201                         }
1202                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1203                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1204                         mq->last_head = 0;
1205                         mq->last_tail = 0;
1206                         /* enable polling mode */
1207                         if (pmd->role == MEMIF_ROLE_SLAVE)
1208                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1209                 }
1210
1211                 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1212                 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1213                 dev->data->dev_link.link_status = ETH_LINK_UP;
1214         }
1215         MIF_LOG(INFO, "Connected.");
1216         return 0;
1217 }
1218
1219 static int
1220 memif_dev_start(struct rte_eth_dev *dev)
1221 {
1222         struct pmd_internals *pmd = dev->data->dev_private;
1223         int ret = 0;
1224
1225         switch (pmd->role) {
1226         case MEMIF_ROLE_SLAVE:
1227                 ret = memif_connect_slave(dev);
1228                 break;
1229         case MEMIF_ROLE_MASTER:
1230                 ret = memif_connect_master(dev);
1231                 break;
1232         default:
1233                 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1234                 ret = -1;
1235                 break;
1236         }
1237
1238         return ret;
1239 }
1240
1241 static int
1242 memif_dev_close(struct rte_eth_dev *dev)
1243 {
1244         struct pmd_internals *pmd = dev->data->dev_private;
1245         int i;
1246
1247         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1248                 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1249                 memif_disconnect(dev);
1250
1251                 for (i = 0; i < dev->data->nb_rx_queues; i++)
1252                         (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
1253                 for (i = 0; i < dev->data->nb_tx_queues; i++)
1254                         (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
1255
1256                 memif_socket_remove_device(dev);
1257         } else {
1258                 memif_disconnect(dev);
1259         }
1260
1261         rte_free(dev->process_private);
1262
1263         return 0;
1264 }
1265
1266 static int
1267 memif_dev_configure(struct rte_eth_dev *dev)
1268 {
1269         struct pmd_internals *pmd = dev->data->dev_private;
1270
1271         /*
1272          * SLAVE - TXQ
1273          * MASTER - RXQ
1274          */
1275         pmd->cfg.num_s2m_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
1276                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1277
1278         /*
1279          * SLAVE - RXQ
1280          * MASTER - TXQ
1281          */
1282         pmd->cfg.num_m2s_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
1283                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1284
1285         return 0;
1286 }
1287
1288 static int
1289 memif_tx_queue_setup(struct rte_eth_dev *dev,
1290                      uint16_t qid,
1291                      uint16_t nb_tx_desc __rte_unused,
1292                      unsigned int socket_id __rte_unused,
1293                      const struct rte_eth_txconf *tx_conf __rte_unused)
1294 {
1295         struct pmd_internals *pmd = dev->data->dev_private;
1296         struct memif_queue *mq;
1297
1298         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1299         if (mq == NULL) {
1300                 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1301                 return -ENOMEM;
1302         }
1303
1304         mq->type =
1305             (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_S2M : MEMIF_RING_M2S;
1306         mq->n_pkts = 0;
1307         mq->n_bytes = 0;
1308         mq->intr_handle.fd = -1;
1309         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1310         mq->in_port = dev->data->port_id;
1311         dev->data->tx_queues[qid] = mq;
1312
1313         return 0;
1314 }
1315
1316 static int
1317 memif_rx_queue_setup(struct rte_eth_dev *dev,
1318                      uint16_t qid,
1319                      uint16_t nb_rx_desc __rte_unused,
1320                      unsigned int socket_id __rte_unused,
1321                      const struct rte_eth_rxconf *rx_conf __rte_unused,
1322                      struct rte_mempool *mb_pool)
1323 {
1324         struct pmd_internals *pmd = dev->data->dev_private;
1325         struct memif_queue *mq;
1326
1327         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1328         if (mq == NULL) {
1329                 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1330                 return -ENOMEM;
1331         }
1332
1333         mq->type = (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_M2S : MEMIF_RING_S2M;
1334         mq->n_pkts = 0;
1335         mq->n_bytes = 0;
1336         mq->intr_handle.fd = -1;
1337         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1338         mq->mempool = mb_pool;
1339         mq->in_port = dev->data->port_id;
1340         dev->data->rx_queues[qid] = mq;
1341
1342         return 0;
1343 }
1344
1345 static void
1346 memif_queue_release(void *queue)
1347 {
1348         struct memif_queue *mq = (struct memif_queue *)queue;
1349
1350         if (!mq)
1351                 return;
1352
1353         rte_free(mq);
1354 }
1355
1356 static int
1357 memif_link_update(struct rte_eth_dev *dev,
1358                   int wait_to_complete __rte_unused)
1359 {
1360         struct pmd_process_private *proc_private;
1361
1362         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1363                 proc_private = dev->process_private;
1364                 if (dev->data->dev_link.link_status == ETH_LINK_UP &&
1365                                 proc_private->regions_num == 0) {
1366                         memif_mp_request_regions(dev);
1367                 } else if (dev->data->dev_link.link_status == ETH_LINK_DOWN &&
1368                                 proc_private->regions_num > 0) {
1369                         memif_free_regions(dev);
1370                 }
1371         }
1372         return 0;
1373 }
1374
1375 static int
1376 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1377 {
1378         struct pmd_internals *pmd = dev->data->dev_private;
1379         struct memif_queue *mq;
1380         int i;
1381         uint8_t tmp, nq;
1382
1383         stats->ipackets = 0;
1384         stats->ibytes = 0;
1385         stats->opackets = 0;
1386         stats->obytes = 0;
1387
1388         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_s2m_rings :
1389             pmd->run.num_m2s_rings;
1390         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1391             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1392
1393         /* RX stats */
1394         for (i = 0; i < nq; i++) {
1395                 mq = dev->data->rx_queues[i];
1396                 stats->q_ipackets[i] = mq->n_pkts;
1397                 stats->q_ibytes[i] = mq->n_bytes;
1398                 stats->ipackets += mq->n_pkts;
1399                 stats->ibytes += mq->n_bytes;
1400         }
1401
1402         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_m2s_rings :
1403             pmd->run.num_s2m_rings;
1404         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1405             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1406
1407         /* TX stats */
1408         for (i = 0; i < nq; i++) {
1409                 mq = dev->data->tx_queues[i];
1410                 stats->q_opackets[i] = mq->n_pkts;
1411                 stats->q_obytes[i] = mq->n_bytes;
1412                 stats->opackets += mq->n_pkts;
1413                 stats->obytes += mq->n_bytes;
1414         }
1415         return 0;
1416 }
1417
1418 static int
1419 memif_stats_reset(struct rte_eth_dev *dev)
1420 {
1421         struct pmd_internals *pmd = dev->data->dev_private;
1422         int i;
1423         struct memif_queue *mq;
1424
1425         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
1426                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->tx_queues[i] :
1427                     dev->data->rx_queues[i];
1428                 mq->n_pkts = 0;
1429                 mq->n_bytes = 0;
1430         }
1431         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
1432                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->rx_queues[i] :
1433                     dev->data->tx_queues[i];
1434                 mq->n_pkts = 0;
1435                 mq->n_bytes = 0;
1436         }
1437
1438         return 0;
1439 }
1440
1441 static int
1442 memif_rx_queue_intr_enable(struct rte_eth_dev *dev __rte_unused,
1443                            uint16_t qid __rte_unused)
1444 {
1445         MIF_LOG(WARNING, "Interrupt mode not supported.");
1446
1447         return -1;
1448 }
1449
1450 static int
1451 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
1452 {
1453         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
1454
1455         return 0;
1456 }
1457
1458 static const struct eth_dev_ops ops = {
1459         .dev_start = memif_dev_start,
1460         .dev_close = memif_dev_close,
1461         .dev_infos_get = memif_dev_info,
1462         .dev_configure = memif_dev_configure,
1463         .tx_queue_setup = memif_tx_queue_setup,
1464         .rx_queue_setup = memif_rx_queue_setup,
1465         .rx_queue_release = memif_queue_release,
1466         .tx_queue_release = memif_queue_release,
1467         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
1468         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
1469         .link_update = memif_link_update,
1470         .stats_get = memif_stats_get,
1471         .stats_reset = memif_stats_reset,
1472 };
1473
1474 static int
1475 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1476              memif_interface_id_t id, uint32_t flags,
1477              const char *socket_filename,
1478              memif_log2_ring_size_t log2_ring_size,
1479              uint16_t pkt_buffer_size, const char *secret,
1480              struct rte_ether_addr *ether_addr)
1481 {
1482         int ret = 0;
1483         struct rte_eth_dev *eth_dev;
1484         struct rte_eth_dev_data *data;
1485         struct pmd_internals *pmd;
1486         struct pmd_process_private *process_private;
1487         const unsigned int numa_node = vdev->device.numa_node;
1488         const char *name = rte_vdev_device_name(vdev);
1489
1490         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1491         if (eth_dev == NULL) {
1492                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1493                 return -1;
1494         }
1495
1496         process_private = (struct pmd_process_private *)
1497                 rte_zmalloc(name, sizeof(struct pmd_process_private),
1498                             RTE_CACHE_LINE_SIZE);
1499
1500         if (process_private == NULL) {
1501                 MIF_LOG(ERR, "Failed to alloc memory for process private");
1502                 return -1;
1503         }
1504         eth_dev->process_private = process_private;
1505
1506         pmd = eth_dev->data->dev_private;
1507         memset(pmd, 0, sizeof(*pmd));
1508
1509         pmd->id = id;
1510         pmd->flags = flags;
1511         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1512         pmd->role = role;
1513         /* Zero-copy flag irelevant to master. */
1514         if (pmd->role == MEMIF_ROLE_MASTER)
1515                 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1516
1517         ret = memif_socket_init(eth_dev, socket_filename);
1518         if (ret < 0)
1519                 return ret;
1520
1521         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1522         if (secret != NULL)
1523                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1524
1525         pmd->cfg.log2_ring_size = log2_ring_size;
1526         /* set in .dev_configure() */
1527         pmd->cfg.num_s2m_rings = 0;
1528         pmd->cfg.num_m2s_rings = 0;
1529
1530         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1531         rte_spinlock_init(&pmd->cc_lock);
1532
1533         data = eth_dev->data;
1534         data->dev_private = pmd;
1535         data->numa_node = numa_node;
1536         data->dev_link = pmd_link;
1537         data->mac_addrs = ether_addr;
1538         data->promiscuous = 1;
1539
1540         eth_dev->dev_ops = &ops;
1541         eth_dev->device = &vdev->device;
1542         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1543                 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1544                 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1545         } else {
1546                 eth_dev->rx_pkt_burst = eth_memif_rx;
1547                 eth_dev->tx_pkt_burst = eth_memif_tx;
1548         }
1549
1550         rte_eth_dev_probing_finish(eth_dev);
1551
1552         return 0;
1553 }
1554
1555 static int
1556 memif_set_role(const char *key __rte_unused, const char *value,
1557                void *extra_args)
1558 {
1559         enum memif_role_t *role = (enum memif_role_t *)extra_args;
1560
1561         if (strstr(value, "master") != NULL) {
1562                 *role = MEMIF_ROLE_MASTER;
1563         } else if (strstr(value, "slave") != NULL) {
1564                 *role = MEMIF_ROLE_SLAVE;
1565         } else {
1566                 MIF_LOG(ERR, "Unknown role: %s.", value);
1567                 return -EINVAL;
1568         }
1569         return 0;
1570 }
1571
1572 static int
1573 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1574 {
1575         uint32_t *flags = (uint32_t *)extra_args;
1576
1577         if (strstr(value, "yes") != NULL) {
1578                 if (!rte_mcfg_get_single_file_segments()) {
1579                         MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1580                         return -ENOTSUP;
1581                 }
1582                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1583         } else if (strstr(value, "no") != NULL) {
1584                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1585         } else {
1586                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1587                 return -EINVAL;
1588         }
1589         return 0;
1590 }
1591
1592 static int
1593 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1594 {
1595         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1596
1597         /* even if parsing fails, 0 is a valid id */
1598         *id = strtoul(value, NULL, 10);
1599         return 0;
1600 }
1601
1602 static int
1603 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1604 {
1605         unsigned long tmp;
1606         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1607
1608         tmp = strtoul(value, NULL, 10);
1609         if (tmp == 0 || tmp > 0xFFFF) {
1610                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1611                 return -EINVAL;
1612         }
1613         *pkt_buffer_size = tmp;
1614         return 0;
1615 }
1616
1617 static int
1618 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1619 {
1620         unsigned long tmp;
1621         memif_log2_ring_size_t *log2_ring_size =
1622             (memif_log2_ring_size_t *)extra_args;
1623
1624         tmp = strtoul(value, NULL, 10);
1625         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1626                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1627                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1628                 return -EINVAL;
1629         }
1630         *log2_ring_size = tmp;
1631         return 0;
1632 }
1633
1634 /* check if directory exists and if we have permission to read/write */
1635 static int
1636 memif_check_socket_filename(const char *filename)
1637 {
1638         char *dir = NULL, *tmp;
1639         uint32_t idx;
1640         int ret = 0;
1641
1642         if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1643                 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1644                 return -1;
1645         }
1646
1647         tmp = strrchr(filename, '/');
1648         if (tmp != NULL) {
1649                 idx = tmp - filename;
1650                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1651                 if (dir == NULL) {
1652                         MIF_LOG(ERR, "Failed to allocate memory.");
1653                         return -1;
1654                 }
1655                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1656         }
1657
1658         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1659                                         W_OK, AT_EACCESS) < 0)) {
1660                 MIF_LOG(ERR, "Invalid socket directory.");
1661                 ret = -EINVAL;
1662         }
1663
1664         if (dir != NULL)
1665                 rte_free(dir);
1666
1667         return ret;
1668 }
1669
1670 static int
1671 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1672                           void *extra_args)
1673 {
1674         const char **socket_filename = (const char **)extra_args;
1675
1676         *socket_filename = value;
1677         return memif_check_socket_filename(*socket_filename);
1678 }
1679
1680 static int
1681 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1682 {
1683         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1684
1685         if (rte_ether_unformat_addr(value, ether_addr) < 0)
1686                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1687         return 0;
1688 }
1689
1690 static int
1691 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1692 {
1693         const char **secret = (const char **)extra_args;
1694
1695         *secret = value;
1696         return 0;
1697 }
1698
1699 static int
1700 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1701 {
1702         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1703         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1704         int ret = 0;
1705         struct rte_kvargs *kvlist;
1706         const char *name = rte_vdev_device_name(vdev);
1707         enum memif_role_t role = MEMIF_ROLE_SLAVE;
1708         memif_interface_id_t id = 0;
1709         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1710         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1711         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1712         uint32_t flags = 0;
1713         const char *secret = NULL;
1714         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1715                 sizeof(struct rte_ether_addr), 0);
1716         struct rte_eth_dev *eth_dev;
1717
1718         rte_eth_random_addr(ether_addr->addr_bytes);
1719
1720         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1721
1722         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1723                 eth_dev = rte_eth_dev_attach_secondary(name);
1724                 if (!eth_dev) {
1725                         MIF_LOG(ERR, "Failed to probe %s", name);
1726                         return -1;
1727                 }
1728
1729                 eth_dev->dev_ops = &ops;
1730                 eth_dev->device = &vdev->device;
1731                 eth_dev->rx_pkt_burst = eth_memif_rx;
1732                 eth_dev->tx_pkt_burst = eth_memif_tx;
1733
1734                 if (!rte_eal_primary_proc_alive(NULL)) {
1735                         MIF_LOG(ERR, "Primary process is missing");
1736                         return -1;
1737                 }
1738
1739                 eth_dev->process_private = (struct pmd_process_private *)
1740                         rte_zmalloc(name,
1741                                 sizeof(struct pmd_process_private),
1742                                 RTE_CACHE_LINE_SIZE);
1743                 if (eth_dev->process_private == NULL) {
1744                         MIF_LOG(ERR,
1745                                 "Failed to alloc memory for process private");
1746                         return -1;
1747                 }
1748
1749                 rte_eth_dev_probing_finish(eth_dev);
1750
1751                 return 0;
1752         }
1753
1754         ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1755         /*
1756          * Primary process can continue probing, but secondary process won't
1757          * be able to get memory regions information
1758          */
1759         if (ret < 0 && rte_errno != EEXIST)
1760                 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1761                         strerror(rte_errno));
1762
1763         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1764
1765         /* parse parameters */
1766         if (kvlist != NULL) {
1767                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1768                                          &memif_set_role, &role);
1769                 if (ret < 0)
1770                         goto exit;
1771                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1772                                          &memif_set_id, &id);
1773                 if (ret < 0)
1774                         goto exit;
1775                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1776                                          &memif_set_bs, &pkt_buffer_size);
1777                 if (ret < 0)
1778                         goto exit;
1779                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1780                                          &memif_set_rs, &log2_ring_size);
1781                 if (ret < 0)
1782                         goto exit;
1783                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1784                                          &memif_set_socket_filename,
1785                                          (void *)(&socket_filename));
1786                 if (ret < 0)
1787                         goto exit;
1788                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1789                                          &memif_set_mac, ether_addr);
1790                 if (ret < 0)
1791                         goto exit;
1792                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1793                                          &memif_set_zc, &flags);
1794                 if (ret < 0)
1795                         goto exit;
1796                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1797                                          &memif_set_secret, (void *)(&secret));
1798                 if (ret < 0)
1799                         goto exit;
1800         }
1801
1802         /* create interface */
1803         ret = memif_create(vdev, role, id, flags, socket_filename,
1804                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1805
1806 exit:
1807         if (kvlist != NULL)
1808                 rte_kvargs_free(kvlist);
1809         return ret;
1810 }
1811
1812 static int
1813 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1814 {
1815         struct rte_eth_dev *eth_dev;
1816
1817         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1818         if (eth_dev == NULL)
1819                 return 0;
1820
1821         rte_eth_dev_close(eth_dev->data->port_id);
1822
1823         return 0;
1824 }
1825
1826 static struct rte_vdev_driver pmd_memif_drv = {
1827         .probe = rte_pmd_memif_probe,
1828         .remove = rte_pmd_memif_remove,
1829 };
1830
1831 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1832
1833 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1834                               ETH_MEMIF_ID_ARG "=<int>"
1835                               ETH_MEMIF_ROLE_ARG "=master|slave"
1836                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1837                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1838                               ETH_MEMIF_SOCKET_ARG "=<string>"
1839                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1840                               ETH_MEMIF_ZC_ARG "=yes|no"
1841                               ETH_MEMIF_SECRET_ARG "=<string>");
1842
1843 RTE_LOG_REGISTER(memif_logtype, pmd.net.memif, NOTICE);