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