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 <ethdev_driver.h>
  21 #include <ethdev_vdev.h>
  22 #include <rte_malloc.h>
  23 #include <rte_kvargs.h>
  24 #include <rte_bus_vdev.h>
  25 #include <rte_string_fns.h>
  26 #include <rte_errno.h>
  27 #include <rte_memory.h>
  28 #include <rte_memzone.h>
  29 #include <rte_eal_memconfig.h>
  30
  31 #include "rte_eth_memif.h"
  32 #include "memif_socket.h"
  33
  34 #define ETH_MEMIF_ID_ARG                "id"
  35 #define ETH_MEMIF_ROLE_ARG              "role"
  36 #define ETH_MEMIF_PKT_BUFFER_SIZE_ARG   "bsize"
  37 #define ETH_MEMIF_RING_SIZE_ARG         "rsize"
  38 #define ETH_MEMIF_SOCKET_ARG            "socket"
  39 #define ETH_MEMIF_SOCKET_ABSTRACT_ARG   "socket-abstract"
  40 #define ETH_MEMIF_MAC_ARG               "mac"
  41 #define ETH_MEMIF_ZC_ARG                "zero-copy"
  42 #define ETH_MEMIF_SECRET_ARG            "secret"
  43
  44 static const char * const valid_arguments[] = {
  45         ETH_MEMIF_ID_ARG,
  46         ETH_MEMIF_ROLE_ARG,
  47         ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
  48         ETH_MEMIF_RING_SIZE_ARG,
  49         ETH_MEMIF_SOCKET_ARG,
  50         ETH_MEMIF_SOCKET_ABSTRACT_ARG,
  51         ETH_MEMIF_MAC_ARG,
  52         ETH_MEMIF_ZC_ARG,
  53         ETH_MEMIF_SECRET_ARG,
  54         NULL
  55 };
  56
  57 static const struct rte_eth_link pmd_link = {
  58         .link_speed = ETH_SPEED_NUM_10G,
  59         .link_duplex = ETH_LINK_FULL_DUPLEX,
  60         .link_status = ETH_LINK_DOWN,
  61         .link_autoneg = ETH_LINK_AUTONEG
  62 };
  63
  64 #define MEMIF_MP_SEND_REGION            "memif_mp_send_region"
  65
  66
  67 static int memif_region_init_zc(const struct rte_memseg_list *msl,
  68                                 const struct rte_memseg *ms, void *arg);
  69
  70 const char *
  71 memif_version(void)
  72 {
  73         return ("memif-" RTE_STR(MEMIF_VERSION_MAJOR) "." RTE_STR(MEMIF_VERSION_MINOR));
  74 }
  75
  76 /* Message header to synchronize regions */
  77 struct mp_region_msg {
  78         char port_name[RTE_DEV_NAME_MAX_LEN];
  79         memif_region_index_t idx;
  80         memif_region_size_t size;
  81 };
  82
  83 static int
  84 memif_mp_send_region(const struct rte_mp_msg *msg, const void *peer)
  85 {
  86         struct rte_eth_dev *dev;
  87         struct pmd_process_private *proc_private;
  88         const struct mp_region_msg *msg_param = (const struct mp_region_msg *)msg->param;
  89         struct rte_mp_msg reply;
  90         struct mp_region_msg *reply_param = (struct mp_region_msg *)reply.param;
  91         uint16_t port_id;
  92         int ret;
  93
  94         /* Get requested port */
  95         ret = rte_eth_dev_get_port_by_name(msg_param->port_name, &port_id);
  96         if (ret) {
  97                 MIF_LOG(ERR, "Failed to get port id for %s",
  98                         msg_param->port_name);
  99                 return -1;
 100         }
 101         dev = &rte_eth_devices[port_id];
 102         proc_private = dev->process_private;
 103
 104         memset(&reply, 0, sizeof(reply));
 105         strlcpy(reply.name, msg->name, sizeof(reply.name));
 106         reply_param->idx = msg_param->idx;
 107         if (proc_private->regions[msg_param->idx] != NULL) {
 108                 reply_param->size = proc_private->regions[msg_param->idx]->region_size;
 109                 reply.fds[0] = proc_private->regions[msg_param->idx]->fd;
 110                 reply.num_fds = 1;
 111         }
 112         reply.len_param = sizeof(*reply_param);
 113         if (rte_mp_reply(&reply, peer) < 0) {
 114                 MIF_LOG(ERR, "Failed to reply to an add region request");
 115                 return -1;
 116         }
 117
 118         return 0;
 119 }
 120
 121 /*
 122  * Request regions
 123  * Called by secondary process, when ports link status goes up.
 124  */
 125 static int
 126 memif_mp_request_regions(struct rte_eth_dev *dev)
 127 {
 128         int ret, i;
 129         struct timespec timeout = {.tv_sec = 5, .tv_nsec = 0};
 130         struct rte_mp_msg msg, *reply;
 131         struct rte_mp_reply replies;
 132         struct mp_region_msg *msg_param = (struct mp_region_msg *)msg.param;
 133         struct mp_region_msg *reply_param;
 134         struct memif_region *r;
 135         struct pmd_process_private *proc_private = dev->process_private;
 136         struct pmd_internals *pmd = dev->data->dev_private;
 137         /* in case of zero-copy client, only request region 0 */
 138         uint16_t max_region_num = (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) ?
 139                                    1 : ETH_MEMIF_MAX_REGION_NUM;
 140
 141         MIF_LOG(DEBUG, "Requesting memory regions");
 142
 143         for (i = 0; i < max_region_num; i++) {
 144                 /* Prepare the message */
 145                 memset(&msg, 0, sizeof(msg));
 146                 strlcpy(msg.name, MEMIF_MP_SEND_REGION, sizeof(msg.name));
 147                 strlcpy(msg_param->port_name, dev->data->name,
 148                         sizeof(msg_param->port_name));
 149                 msg_param->idx = i;
 150                 msg.len_param = sizeof(*msg_param);
 151
 152                 /* Send message */
 153                 ret = rte_mp_request_sync(&msg, &replies, &timeout);
 154                 if (ret < 0 || replies.nb_received != 1) {
 155                         MIF_LOG(ERR, "Failed to send mp msg: %d",
 156                                 rte_errno);
 157                         return -1;
 158                 }
 159
 160                 reply = &replies.msgs[0];
 161                 reply_param = (struct mp_region_msg *)reply->param;
 162
 163                 if (reply_param->size > 0) {
 164                         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 165                         if (r == NULL) {
 166                                 MIF_LOG(ERR, "Failed to alloc memif region.");
 167                                 free(reply);
 168                                 return -ENOMEM;
 169                         }
 170                         r->region_size = reply_param->size;
 171                         if (reply->num_fds < 1) {
 172                                 MIF_LOG(ERR, "Missing file descriptor.");
 173                                 free(reply);
 174                                 return -1;
 175                         }
 176                         r->fd = reply->fds[0];
 177                         r->addr = NULL;
 178
 179                         proc_private->regions[reply_param->idx] = r;
 180                         proc_private->regions_num++;
 181                 }
 182                 free(reply);
 183         }
 184
 185         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
 186                 ret = rte_memseg_walk(memif_region_init_zc, (void *)proc_private);
 187                 if (ret < 0)
 188                         return ret;
 189         }
 190
 191         return memif_connect(dev);
 192 }
 193
 194 static int
 195 memif_dev_info(struct rte_eth_dev *dev __rte_unused, struct rte_eth_dev_info *dev_info)
 196 {
 197         dev_info->max_mac_addrs = 1;
 198         dev_info->max_rx_pktlen = (uint32_t)ETH_FRAME_LEN;
 199         dev_info->max_rx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
 200         dev_info->max_tx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
 201         dev_info->min_rx_bufsize = 0;
 202
 203         return 0;
 204 }
 205
 206 static memif_ring_t *
 207 memif_get_ring(struct pmd_internals *pmd, struct pmd_process_private *proc_private,
 208                memif_ring_type_t type, uint16_t ring_num)
 209 {
 210         /* rings only in region 0 */
 211         void *p = proc_private->regions[0]->addr;
 212         int ring_size = sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 213             (1 << pmd->run.log2_ring_size);
 214
 215         p = (uint8_t *)p + (ring_num + type * pmd->run.num_c2s_rings) * ring_size;
 216
 217         return (memif_ring_t *)p;
 218 }
 219
 220 static memif_region_offset_t
 221 memif_get_ring_offset(struct rte_eth_dev *dev, struct memif_queue *mq,
 222                       memif_ring_type_t type, uint16_t num)
 223 {
 224         struct pmd_internals *pmd = dev->data->dev_private;
 225         struct pmd_process_private *proc_private = dev->process_private;
 226
 227         return ((uint8_t *)memif_get_ring(pmd, proc_private, type, num) -
 228                 (uint8_t *)proc_private->regions[mq->region]->addr);
 229 }
 230
 231 static memif_ring_t *
 232 memif_get_ring_from_queue(struct pmd_process_private *proc_private,
 233                           struct memif_queue *mq)
 234 {
 235         struct memif_region *r;
 236
 237         r = proc_private->regions[mq->region];
 238         if (r == NULL)
 239                 return NULL;
 240
 241         return (memif_ring_t *)((uint8_t *)r->addr + mq->ring_offset);
 242 }
 243
 244 static void *
 245 memif_get_buffer(struct pmd_process_private *proc_private, memif_desc_t *d)
 246 {
 247         return ((uint8_t *)proc_private->regions[d->region]->addr + d->offset);
 248 }
 249
 250 /* Free mbufs received by server */
 251 static void
 252 memif_free_stored_mbufs(struct pmd_process_private *proc_private, struct memif_queue *mq)
 253 {
 254         uint16_t cur_tail;
 255         uint16_t mask = (1 << mq->log2_ring_size) - 1;
 256         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 257
 258         /* FIXME: improve performance */
 259         /* The ring->tail acts as a guard variable between Tx and Rx
 260          * threads, so using load-acquire pairs with store-release
 261          * in function eth_memif_rx for C2S queues.
 262          */
 263         cur_tail = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 264         while (mq->last_tail != cur_tail) {
 265                 RTE_MBUF_PREFETCH_TO_FREE(mq->buffers[(mq->last_tail + 1) & mask]);
 266                 /* Decrement refcnt and free mbuf. (current segment) */
 267                 rte_mbuf_refcnt_update(mq->buffers[mq->last_tail & mask], -1);
 268                 rte_pktmbuf_free_seg(mq->buffers[mq->last_tail & mask]);
 269                 mq->last_tail++;
 270         }
 271 }
 272
 273 static int
 274 memif_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *cur_tail,
 275                     struct rte_mbuf *tail)
 276 {
 277         /* Check for number-of-segments-overflow */
 278         if (unlikely(head->nb_segs + tail->nb_segs > RTE_MBUF_MAX_NB_SEGS))
 279                 return -EOVERFLOW;
 280
 281         /* Chain 'tail' onto the old tail */
 282         cur_tail->next = tail;
 283
 284         /* accumulate number of segments and total length. */
 285         head->nb_segs = (uint16_t)(head->nb_segs + tail->nb_segs);
 286
 287         tail->pkt_len = tail->data_len;
 288         head->pkt_len += tail->pkt_len;
 289
 290         return 0;
 291 }
 292
 293 static uint16_t
 294 eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 295 {
 296         struct memif_queue *mq = queue;
 297         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 298         struct pmd_process_private *proc_private =
 299                 rte_eth_devices[mq->in_port].process_private;
 300         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 301         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0;
 302         uint16_t n_rx_pkts = 0;
 303         uint16_t mbuf_size = rte_pktmbuf_data_room_size(mq->mempool) -
 304                 RTE_PKTMBUF_HEADROOM;
 305         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 306         memif_ring_type_t type = mq->type;
 307         memif_desc_t *d0;
 308         struct rte_mbuf *mbuf, *mbuf_head, *mbuf_tail;
 309         uint64_t b;
 310         ssize_t size __rte_unused;
 311         uint16_t head;
 312         int ret;
 313         struct rte_eth_link link;
 314
 315         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 316                 return 0;
 317         if (unlikely(ring == NULL)) {
 318                 /* Secondary process will attempt to request regions. */
 319                 ret = rte_eth_link_get(mq->in_port, &link);
 320                 if (ret < 0)
 321                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 322                                 mq->in_port, rte_strerror(-ret));
 323                 return 0;
 324         }
 325
 326         /* consume interrupt */
 327         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0)
 328                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 329
 330         ring_size = 1 << mq->log2_ring_size;
 331         mask = ring_size - 1;
 332
 333         if (type == MEMIF_RING_C2S) {
 334                 cur_slot = mq->last_head;
 335                 last_slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
 336         } else {
 337                 cur_slot = mq->last_tail;
 338                 last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 339         }
 340
 341         if (cur_slot == last_slot)
 342                 goto refill;
 343         n_slots = last_slot - cur_slot;
 344
 345         while (n_slots && n_rx_pkts < nb_pkts) {
 346                 mbuf_head = rte_pktmbuf_alloc(mq->mempool);
 347                 if (unlikely(mbuf_head == NULL))
 348                         goto no_free_bufs;
 349                 mbuf = mbuf_head;
 350                 mbuf->port = mq->in_port;
 351
 352 next_slot:
 353                 s0 = cur_slot & mask;
 354                 d0 = &ring->desc[s0];
 355
 356                 src_len = d0->length;
 357                 dst_off = 0;
 358                 src_off = 0;
 359
 360                 do {
 361                         dst_len = mbuf_size - dst_off;
 362                         if (dst_len == 0) {
 363                                 dst_off = 0;
 364                                 dst_len = mbuf_size;
 365
 366                                 /* store pointer to tail */
 367                                 mbuf_tail = mbuf;
 368                                 mbuf = rte_pktmbuf_alloc(mq->mempool);
 369                                 if (unlikely(mbuf == NULL))
 370                                         goto no_free_bufs;
 371                                 mbuf->port = mq->in_port;
 372                                 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 373                                 if (unlikely(ret < 0)) {
 374                                         MIF_LOG(ERR, "number-of-segments-overflow");
 375                                         rte_pktmbuf_free(mbuf);
 376                                         goto no_free_bufs;
 377                                 }
 378                         }
 379                         cp_len = RTE_MIN(dst_len, src_len);
 380
 381                         rte_pktmbuf_data_len(mbuf) += cp_len;
 382                         rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 383                         if (mbuf != mbuf_head)
 384                                 rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
 385
 386                         memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
 387                                (uint8_t *)memif_get_buffer(proc_private, d0) + src_off,
 388                                cp_len);
 389
 390                         src_off += cp_len;
 391                         dst_off += cp_len;
 392                         src_len -= cp_len;
 393                 } while (src_len);
 394
 395                 cur_slot++;
 396                 n_slots--;
 397
 398                 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
 399                         goto next_slot;
 400
 401                 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
 402                 *bufs++ = mbuf_head;
 403                 n_rx_pkts++;
 404         }
 405
 406 no_free_bufs:
 407         if (type == MEMIF_RING_C2S) {
 408                 __atomic_store_n(&ring->tail, cur_slot, __ATOMIC_RELEASE);
 409                 mq->last_head = cur_slot;
 410         } else {
 411                 mq->last_tail = cur_slot;
 412         }
 413
 414 refill:
 415         if (type == MEMIF_RING_S2C) {
 416                 /* ring->head is updated by the receiver and this function
 417                  * is called in the context of receiver thread. The loads in
 418                  * the receiver do not need to synchronize with its own stores.
 419                  */
 420                 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 421                 n_slots = ring_size - head + mq->last_tail;
 422
 423                 while (n_slots--) {
 424                         s0 = head++ & mask;
 425                         d0 = &ring->desc[s0];
 426                         d0->length = pmd->run.pkt_buffer_size;
 427                 }
 428                 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 429         }
 430
 431         mq->n_pkts += n_rx_pkts;
 432         return n_rx_pkts;
 433 }
 434
 435 static uint16_t
 436 eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 437 {
 438         struct memif_queue *mq = queue;
 439         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 440         struct pmd_process_private *proc_private =
 441                 rte_eth_devices[mq->in_port].process_private;
 442         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 443         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0, head;
 444         uint16_t n_rx_pkts = 0;
 445         memif_desc_t *d0;
 446         struct rte_mbuf *mbuf, *mbuf_tail;
 447         struct rte_mbuf *mbuf_head = NULL;
 448         int ret;
 449         struct rte_eth_link link;
 450
 451         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 452                 return 0;
 453         if (unlikely(ring == NULL)) {
 454                 /* Secondary process will attempt to request regions. */
 455                 rte_eth_link_get(mq->in_port, &link);
 456                 return 0;
 457         }
 458
 459         /* consume interrupt */
 460         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 461                 uint64_t b;
 462                 ssize_t size __rte_unused;
 463                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 464         }
 465
 466         ring_size = 1 << mq->log2_ring_size;
 467         mask = ring_size - 1;
 468
 469         cur_slot = mq->last_tail;
 470         /* The ring->tail acts as a guard variable between Tx and Rx
 471          * threads, so using load-acquire pairs with store-release
 472          * to synchronize it between threads.
 473          */
 474         last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 475         if (cur_slot == last_slot)
 476                 goto refill;
 477         n_slots = last_slot - cur_slot;
 478
 479         while (n_slots && n_rx_pkts < nb_pkts) {
 480                 s0 = cur_slot & mask;
 481
 482                 d0 = &ring->desc[s0];
 483                 mbuf_head = mq->buffers[s0];
 484                 mbuf = mbuf_head;
 485
 486 next_slot:
 487                 /* prefetch next descriptor */
 488                 if (n_rx_pkts + 1 < nb_pkts)
 489                         rte_prefetch0(&ring->desc[(cur_slot + 1) & mask]);
 490
 491                 mbuf->port = mq->in_port;
 492                 rte_pktmbuf_data_len(mbuf) = d0->length;
 493                 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 494
 495                 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 496
 497                 cur_slot++;
 498                 n_slots--;
 499                 if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
 500                         s0 = cur_slot & mask;
 501                         d0 = &ring->desc[s0];
 502                         mbuf_tail = mbuf;
 503                         mbuf = mq->buffers[s0];
 504                         ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 505                         if (unlikely(ret < 0)) {
 506                                 MIF_LOG(ERR, "number-of-segments-overflow");
 507                                 goto refill;
 508                         }
 509                         goto next_slot;
 510                 }
 511
 512                 *bufs++ = mbuf_head;
 513                 n_rx_pkts++;
 514         }
 515
 516         mq->last_tail = cur_slot;
 517
 518 /* Supply server with new buffers */
 519 refill:
 520         /* ring->head is updated by the receiver and this function
 521          * is called in the context of receiver thread. The loads in
 522          * the receiver do not need to synchronize with its own stores.
 523          */
 524         head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 525         n_slots = ring_size - head + mq->last_tail;
 526
 527         if (n_slots < 32)
 528                 goto no_free_mbufs;
 529
 530         ret = rte_pktmbuf_alloc_bulk(mq->mempool, &mq->buffers[head & mask], n_slots);
 531         if (unlikely(ret < 0))
 532                 goto no_free_mbufs;
 533
 534         while (n_slots--) {
 535                 s0 = head++ & mask;
 536                 if (n_slots > 0)
 537                         rte_prefetch0(mq->buffers[head & mask]);
 538                 d0 = &ring->desc[s0];
 539                 /* store buffer header */
 540                 mbuf = mq->buffers[s0];
 541                 /* populate descriptor */
 542                 d0->length = rte_pktmbuf_data_room_size(mq->mempool) -
 543                                 RTE_PKTMBUF_HEADROOM;
 544                 d0->region = 1;
 545                 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 546                         (uint8_t *)proc_private->regions[d0->region]->addr;
 547         }
 548 no_free_mbufs:
 549         /* The ring->head acts as a guard variable between Tx and Rx
 550          * threads, so using store-release pairs with load-acquire
 551          * in function eth_memif_tx.
 552          */
 553         __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 554
 555         mq->n_pkts += n_rx_pkts;
 556
 557         return n_rx_pkts;
 558 }
 559
 560 static uint16_t
 561 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 562 {
 563         struct memif_queue *mq = queue;
 564         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 565         struct pmd_process_private *proc_private =
 566                 rte_eth_devices[mq->in_port].process_private;
 567         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 568         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 569         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 570         memif_ring_type_t type = mq->type;
 571         memif_desc_t *d0;
 572         struct rte_mbuf *mbuf;
 573         struct rte_mbuf *mbuf_head;
 574         uint64_t a;
 575         ssize_t size;
 576         struct rte_eth_link link;
 577
 578         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 579                 return 0;
 580         if (unlikely(ring == NULL)) {
 581                 int ret;
 582
 583                 /* Secondary process will attempt to request regions. */
 584                 ret = rte_eth_link_get(mq->in_port, &link);
 585                 if (ret < 0)
 586                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 587                                 mq->in_port, rte_strerror(-ret));
 588                 return 0;
 589         }
 590
 591         ring_size = 1 << mq->log2_ring_size;
 592         mask = ring_size - 1;
 593
 594         if (type == MEMIF_RING_C2S) {
 595                 /* For C2S queues ring->head is updated by the sender and
 596                  * this function is called in the context of sending thread.
 597                  * The loads in the sender do not need to synchronize with
 598                  * its own stores. Hence, the following load can be a
 599                  * relaxed load.
 600                  */
 601                 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 602                 n_free = ring_size - slot +
 603                                 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 604         } else {
 605                 /* For S2C queues ring->tail is updated by the sender and
 606                  * this function is called in the context of sending thread.
 607                  * The loads in the sender do not need to synchronize with
 608                  * its own stores. Hence, the following load can be a
 609                  * relaxed load.
 610                  */
 611                 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 612                 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 613         }
 614
 615         while (n_tx_pkts < nb_pkts && n_free) {
 616                 mbuf_head = *bufs++;
 617                 mbuf = mbuf_head;
 618
 619                 saved_slot = slot;
 620                 d0 = &ring->desc[slot & mask];
 621                 dst_off = 0;
 622                 dst_len = (type == MEMIF_RING_C2S) ?
 623                         pmd->run.pkt_buffer_size : d0->length;
 624
 625 next_in_chain:
 626                 src_off = 0;
 627                 src_len = rte_pktmbuf_data_len(mbuf);
 628
 629                 while (src_len) {
 630                         if (dst_len == 0) {
 631                                 if (n_free) {
 632                                         slot++;
 633                                         n_free--;
 634                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 635                                         d0 = &ring->desc[slot & mask];
 636                                         dst_off = 0;
 637                                         dst_len = (type == MEMIF_RING_C2S) ?
 638                                             pmd->run.pkt_buffer_size : d0->length;
 639                                         d0->flags = 0;
 640                                 } else {
 641                                         slot = saved_slot;
 642                                         goto no_free_slots;
 643                                 }
 644                         }
 645                         cp_len = RTE_MIN(dst_len, src_len);
 646
 647                         memcpy((uint8_t *)memif_get_buffer(proc_private, d0) + dst_off,
 648                                rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 649                                cp_len);
 650
 651                         mq->n_bytes += cp_len;
 652                         src_off += cp_len;
 653                         dst_off += cp_len;
 654                         src_len -= cp_len;
 655                         dst_len -= cp_len;
 656
 657                         d0->length = dst_off;
 658                 }
 659
 660                 if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 661                         mbuf = mbuf->next;
 662                         goto next_in_chain;
 663                 }
 664
 665                 n_tx_pkts++;
 666                 slot++;
 667                 n_free--;
 668                 rte_pktmbuf_free(mbuf_head);
 669         }
 670
 671 no_free_slots:
 672         if (type == MEMIF_RING_C2S)
 673                 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 674         else
 675                 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
 676
 677         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 678                 a = 1;
 679                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 680                 if (unlikely(size < 0)) {
 681                         MIF_LOG(WARNING,
 682                                 "Failed to send interrupt. %s", strerror(errno));
 683                 }
 684         }
 685
 686         mq->n_pkts += n_tx_pkts;
 687         return n_tx_pkts;
 688 }
 689
 690
 691 static int
 692 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
 693                 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
 694                 uint16_t slot, uint16_t n_free)
 695 {
 696         memif_desc_t *d0;
 697         int used_slots = 1;
 698
 699 next_in_chain:
 700         /* store pointer to mbuf to free it later */
 701         mq->buffers[slot & mask] = mbuf;
 702         /* Increment refcnt to make sure the buffer is not freed before server
 703          * receives it. (current segment)
 704          */
 705         rte_mbuf_refcnt_update(mbuf, 1);
 706         /* populate descriptor */
 707         d0 = &ring->desc[slot & mask];
 708         d0->length = rte_pktmbuf_data_len(mbuf);
 709         mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 710         /* FIXME: get region index */
 711         d0->region = 1;
 712         d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 713                 (uint8_t *)proc_private->regions[d0->region]->addr;
 714         d0->flags = 0;
 715
 716         /* check if buffer is chained */
 717         if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 718                 if (n_free < 2)
 719                         return 0;
 720                 /* mark buffer as chained */
 721                 d0->flags |= MEMIF_DESC_FLAG_NEXT;
 722                 /* advance mbuf */
 723                 mbuf = mbuf->next;
 724                 /* update counters */
 725                 used_slots++;
 726                 slot++;
 727                 n_free--;
 728                 goto next_in_chain;
 729         }
 730         return used_slots;
 731 }
 732
 733 static uint16_t
 734 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 735 {
 736         struct memif_queue *mq = queue;
 737         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 738         struct pmd_process_private *proc_private =
 739                 rte_eth_devices[mq->in_port].process_private;
 740         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 741         uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
 742         struct rte_eth_link link;
 743
 744         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 745                 return 0;
 746         if (unlikely(ring == NULL)) {
 747                 /* Secondary process will attempt to request regions. */
 748                 rte_eth_link_get(mq->in_port, &link);
 749                 return 0;
 750         }
 751
 752         ring_size = 1 << mq->log2_ring_size;
 753         mask = ring_size - 1;
 754
 755         /* free mbufs received by server */
 756         memif_free_stored_mbufs(proc_private, mq);
 757
 758         /* ring type always MEMIF_RING_C2S */
 759         /* For C2S queues ring->head is updated by the sender and
 760          * this function is called in the context of sending thread.
 761          * The loads in the sender do not need to synchronize with
 762          * its own stores. Hence, the following load can be a
 763          * relaxed load.
 764          */
 765         slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 766         n_free = ring_size - slot + mq->last_tail;
 767
 768         int used_slots;
 769
 770         while (n_free && (n_tx_pkts < nb_pkts)) {
 771                 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
 772                         if ((nb_pkts - n_tx_pkts) > 8) {
 773                                 rte_prefetch0(*bufs + 4);
 774                                 rte_prefetch0(*bufs + 5);
 775                                 rte_prefetch0(*bufs + 6);
 776                                 rte_prefetch0(*bufs + 7);
 777                         }
 778                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 779                                 mask, slot, n_free);
 780                         if (unlikely(used_slots < 1))
 781                                 goto no_free_slots;
 782                         n_tx_pkts++;
 783                         slot += used_slots;
 784                         n_free -= used_slots;
 785
 786                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 787                                 mask, slot, n_free);
 788                         if (unlikely(used_slots < 1))
 789                                 goto no_free_slots;
 790                         n_tx_pkts++;
 791                         slot += used_slots;
 792                         n_free -= used_slots;
 793
 794                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 795                                 mask, slot, n_free);
 796                         if (unlikely(used_slots < 1))
 797                                 goto no_free_slots;
 798                         n_tx_pkts++;
 799                         slot += used_slots;
 800                         n_free -= used_slots;
 801
 802                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 803                                 mask, slot, n_free);
 804                         if (unlikely(used_slots < 1))
 805                                 goto no_free_slots;
 806                         n_tx_pkts++;
 807                         slot += used_slots;
 808                         n_free -= used_slots;
 809                 }
 810                 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 811                         mask, slot, n_free);
 812                 if (unlikely(used_slots < 1))
 813                         goto no_free_slots;
 814                 n_tx_pkts++;
 815                 slot += used_slots;
 816                 n_free -= used_slots;
 817         }
 818
 819 no_free_slots:
 820         /* ring type always MEMIF_RING_C2S */
 821         /* The ring->head acts as a guard variable between Tx and Rx
 822          * threads, so using store-release pairs with load-acquire
 823          * in function eth_memif_rx for C2S rings.
 824          */
 825         __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 826
 827         /* Send interrupt, if enabled. */
 828         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 829                 uint64_t a = 1;
 830                 ssize_t size = write(mq->intr_handle.fd, &a, sizeof(a));
 831                 if (unlikely(size < 0)) {
 832                         MIF_LOG(WARNING,
 833                                 "Failed to send interrupt. %s", strerror(errno));
 834                 }
 835         }
 836
 837         /* increment queue counters */
 838         mq->n_pkts += n_tx_pkts;
 839
 840         return n_tx_pkts;
 841 }
 842
 843 void
 844 memif_free_regions(struct rte_eth_dev *dev)
 845 {
 846         struct pmd_process_private *proc_private = dev->process_private;
 847         struct pmd_internals *pmd = dev->data->dev_private;
 848         int i;
 849         struct memif_region *r;
 850
 851         /* regions are allocated contiguously, so it's
 852          * enough to loop until 'proc_private->regions_num'
 853          */
 854         for (i = 0; i < proc_private->regions_num; i++) {
 855                 r = proc_private->regions[i];
 856                 if (r != NULL) {
 857                         /* This is memzone */
 858                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
 859                                 r->addr = NULL;
 860                                 if (r->fd > 0)
 861                                         close(r->fd);
 862                         }
 863                         if (r->addr != NULL) {
 864                                 munmap(r->addr, r->region_size);
 865                                 if (r->fd > 0) {
 866                                         close(r->fd);
 867                                         r->fd = -1;
 868                                 }
 869                         }
 870                         rte_free(r);
 871                         proc_private->regions[i] = NULL;
 872                 }
 873         }
 874         proc_private->regions_num = 0;
 875 }
 876
 877 static int
 878 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 879                      void *arg)
 880 {
 881         struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
 882         struct memif_region *r;
 883
 884         if (proc_private->regions_num < 1) {
 885                 MIF_LOG(ERR, "Missing descriptor region");
 886                 return -1;
 887         }
 888
 889         r = proc_private->regions[proc_private->regions_num - 1];
 890
 891         if (r->addr != msl->base_va)
 892                 r = proc_private->regions[++proc_private->regions_num - 1];
 893
 894         if (r == NULL) {
 895                 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 896                 if (r == NULL) {
 897                         MIF_LOG(ERR, "Failed to alloc memif region.");
 898                         return -ENOMEM;
 899                 }
 900
 901                 r->addr = msl->base_va;
 902                 r->region_size = ms->len;
 903                 r->fd = rte_memseg_get_fd(ms);
 904                 if (r->fd < 0)
 905                         return -1;
 906                 r->pkt_buffer_offset = 0;
 907
 908                 proc_private->regions[proc_private->regions_num - 1] = r;
 909         } else {
 910                 r->region_size += ms->len;
 911         }
 912
 913         return 0;
 914 }
 915
 916 static int
 917 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
 918 {
 919         struct pmd_internals *pmd = dev->data->dev_private;
 920         struct pmd_process_private *proc_private = dev->process_private;
 921         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 922         int ret = 0;
 923         struct memif_region *r;
 924
 925         if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 926                 MIF_LOG(ERR, "Too many regions.");
 927                 return -1;
 928         }
 929
 930         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 931         if (r == NULL) {
 932                 MIF_LOG(ERR, "Failed to alloc memif region.");
 933                 return -ENOMEM;
 934         }
 935
 936         /* calculate buffer offset */
 937         r->pkt_buffer_offset = (pmd->run.num_c2s_rings + pmd->run.num_s2c_rings) *
 938             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 939             (1 << pmd->run.log2_ring_size));
 940
 941         r->region_size = r->pkt_buffer_offset;
 942         /* if region has buffers, add buffers size to region_size */
 943         if (has_buffers == 1)
 944                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 945                         (1 << pmd->run.log2_ring_size) *
 946                         (pmd->run.num_c2s_rings +
 947                          pmd->run.num_s2c_rings));
 948
 949         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 950         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 951                  proc_private->regions_num);
 952
 953         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 954         if (r->fd < 0) {
 955                 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
 956                 ret = -1;
 957                 goto error;
 958         }
 959
 960         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 961         if (ret < 0) {
 962                 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
 963                 goto error;
 964         }
 965
 966         ret = ftruncate(r->fd, r->region_size);
 967         if (ret < 0) {
 968                 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
 969                 goto error;
 970         }
 971
 972         r->addr = mmap(NULL, r->region_size, PROT_READ |
 973                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 974         if (r->addr == MAP_FAILED) {
 975                 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
 976                 ret = -1;
 977                 goto error;
 978         }
 979
 980         proc_private->regions[proc_private->regions_num] = r;
 981         proc_private->regions_num++;
 982
 983         return ret;
 984
 985 error:
 986         if (r->fd > 0)
 987                 close(r->fd);
 988         r->fd = -1;
 989
 990         return ret;
 991 }
 992
 993 static int
 994 memif_regions_init(struct rte_eth_dev *dev)
 995 {
 996         struct pmd_internals *pmd = dev->data->dev_private;
 997         int ret;
 998
 999         /*
1000          * Zero-copy exposes dpdk memory.
1001          * Each memseg list will be represented by memif region.
1002          * Zero-copy regions indexing: memseg list idx + 1,
1003          * as we already have region 0 reserved for descriptors.
1004          */
1005         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1006                 /* create region idx 0 containing descriptors */
1007                 ret = memif_region_init_shm(dev, 0);
1008                 if (ret < 0)
1009                         return ret;
1010                 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1011                 if (ret < 0)
1012                         return ret;
1013         } else {
1014                 /* create one memory region contaning rings and buffers */
1015                 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1016                 if (ret < 0)
1017                         return ret;
1018         }
1019
1020         return 0;
1021 }
1022
1023 static void
1024 memif_init_rings(struct rte_eth_dev *dev)
1025 {
1026         struct pmd_internals *pmd = dev->data->dev_private;
1027         struct pmd_process_private *proc_private = dev->process_private;
1028         memif_ring_t *ring;
1029         int i, j;
1030         uint16_t slot;
1031
1032         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1033                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_C2S, i);
1034                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1035                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1036                 ring->cookie = MEMIF_COOKIE;
1037                 ring->flags = 0;
1038
1039                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1040                         continue;
1041
1042                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1043                         slot = i * (1 << pmd->run.log2_ring_size) + j;
1044                         ring->desc[j].region = 0;
1045                         ring->desc[j].offset =
1046                                 proc_private->regions[0]->pkt_buffer_offset +
1047                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1048                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1049                 }
1050         }
1051
1052         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1053                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2C, i);
1054                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1055                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1056                 ring->cookie = MEMIF_COOKIE;
1057                 ring->flags = 0;
1058
1059                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1060                         continue;
1061
1062                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1063                         slot = (i + pmd->run.num_c2s_rings) *
1064                             (1 << pmd->run.log2_ring_size) + j;
1065                         ring->desc[j].region = 0;
1066                         ring->desc[j].offset =
1067                                 proc_private->regions[0]->pkt_buffer_offset +
1068                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1069                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1070                 }
1071         }
1072 }
1073
1074 /* called only by client */
1075 static int
1076 memif_init_queues(struct rte_eth_dev *dev)
1077 {
1078         struct pmd_internals *pmd = dev->data->dev_private;
1079         struct memif_queue *mq;
1080         int i;
1081
1082         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1083                 mq = dev->data->tx_queues[i];
1084                 mq->log2_ring_size = pmd->run.log2_ring_size;
1085                 /* queues located only in region 0 */
1086                 mq->region = 0;
1087                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_C2S, i);
1088                 mq->last_head = 0;
1089                 mq->last_tail = 0;
1090                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1091                 if (mq->intr_handle.fd < 0) {
1092                         MIF_LOG(WARNING,
1093                                 "Failed to create eventfd for tx queue %d: %s.", i,
1094                                 strerror(errno));
1095                 }
1096                 mq->buffers = NULL;
1097                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1098                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1099                                                   (1 << mq->log2_ring_size), 0);
1100                         if (mq->buffers == NULL)
1101                                 return -ENOMEM;
1102                 }
1103         }
1104
1105         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1106                 mq = dev->data->rx_queues[i];
1107                 mq->log2_ring_size = pmd->run.log2_ring_size;
1108                 /* queues located only in region 0 */
1109                 mq->region = 0;
1110                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2C, i);
1111                 mq->last_head = 0;
1112                 mq->last_tail = 0;
1113                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1114                 if (mq->intr_handle.fd < 0) {
1115                         MIF_LOG(WARNING,
1116                                 "Failed to create eventfd for rx queue %d: %s.", i,
1117                                 strerror(errno));
1118                 }
1119                 mq->buffers = NULL;
1120                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1121                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1122                                                   (1 << mq->log2_ring_size), 0);
1123                         if (mq->buffers == NULL)
1124                                 return -ENOMEM;
1125                 }
1126         }
1127         return 0;
1128 }
1129
1130 int
1131 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1132 {
1133         int ret;
1134
1135         ret = memif_regions_init(dev);
1136         if (ret < 0)
1137                 return ret;
1138
1139         memif_init_rings(dev);
1140
1141         ret = memif_init_queues(dev);
1142         if (ret < 0)
1143                 return ret;
1144
1145         return 0;
1146 }
1147
1148 int
1149 memif_connect(struct rte_eth_dev *dev)
1150 {
1151         struct pmd_internals *pmd = dev->data->dev_private;
1152         struct pmd_process_private *proc_private = dev->process_private;
1153         struct memif_region *mr;
1154         struct memif_queue *mq;
1155         memif_ring_t *ring;
1156         int i;
1157
1158         for (i = 0; i < proc_private->regions_num; i++) {
1159                 mr = proc_private->regions[i];
1160                 if (mr != NULL) {
1161                         if (mr->addr == NULL) {
1162                                 if (mr->fd < 0)
1163                                         return -1;
1164                                 mr->addr = mmap(NULL, mr->region_size,
1165                                                 PROT_READ | PROT_WRITE,
1166                                                 MAP_SHARED, mr->fd, 0);
1167                                 if (mr->addr == MAP_FAILED) {
1168                                         MIF_LOG(ERR, "mmap failed: %s\n",
1169                                                 strerror(errno));
1170                                         return -1;
1171                                 }
1172                         }
1173                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1174                                 /* close memseg file */
1175                                 close(mr->fd);
1176                                 mr->fd = -1;
1177                         }
1178                 }
1179         }
1180
1181         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1182                 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1183                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1184                             dev->data->tx_queues[i] : dev->data->rx_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_SERVER)
1196                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1197                 }
1198                 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1199                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1200                             dev->data->rx_queues[i] : dev->data->tx_queues[i];
1201                         ring = memif_get_ring_from_queue(proc_private, mq);
1202                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1203                                 MIF_LOG(ERR, "Wrong ring");
1204                                 return -1;
1205                         }
1206                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1207                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1208                         mq->last_head = 0;
1209                         mq->last_tail = 0;
1210                         /* enable polling mode */
1211                         if (pmd->role == MEMIF_ROLE_CLIENT)
1212                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1213                 }
1214
1215                 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1216                 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1217                 dev->data->dev_link.link_status = ETH_LINK_UP;
1218         }
1219         MIF_LOG(INFO, "Connected.");
1220         return 0;
1221 }
1222
1223 static int
1224 memif_dev_start(struct rte_eth_dev *dev)
1225 {
1226         struct pmd_internals *pmd = dev->data->dev_private;
1227         int ret = 0;
1228
1229         switch (pmd->role) {
1230         case MEMIF_ROLE_CLIENT:
1231                 ret = memif_connect_client(dev);
1232                 break;
1233         case MEMIF_ROLE_SERVER:
1234                 ret = memif_connect_server(dev);
1235                 break;
1236         default:
1237                 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1238                 ret = -1;
1239                 break;
1240         }
1241
1242         return ret;
1243 }
1244
1245 static int
1246 memif_dev_close(struct rte_eth_dev *dev)
1247 {
1248         struct pmd_internals *pmd = dev->data->dev_private;
1249         int i;
1250
1251         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1252                 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1253                 memif_disconnect(dev);
1254
1255                 for (i = 0; i < dev->data->nb_rx_queues; i++)
1256                         (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
1257                 for (i = 0; i < dev->data->nb_tx_queues; i++)
1258                         (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
1259
1260                 memif_socket_remove_device(dev);
1261         } else {
1262                 memif_disconnect(dev);
1263         }
1264
1265         rte_free(dev->process_private);
1266
1267         return 0;
1268 }
1269
1270 static int
1271 memif_dev_configure(struct rte_eth_dev *dev)
1272 {
1273         struct pmd_internals *pmd = dev->data->dev_private;
1274
1275         /*
1276          * CLIENT - TXQ
1277          * SERVER - RXQ
1278          */
1279         pmd->cfg.num_c2s_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1280                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1281
1282         /*
1283          * CLIENT - RXQ
1284          * SERVER - TXQ
1285          */
1286         pmd->cfg.num_s2c_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1287                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1288
1289         return 0;
1290 }
1291
1292 static int
1293 memif_tx_queue_setup(struct rte_eth_dev *dev,
1294                      uint16_t qid,
1295                      uint16_t nb_tx_desc __rte_unused,
1296                      unsigned int socket_id __rte_unused,
1297                      const struct rte_eth_txconf *tx_conf __rte_unused)
1298 {
1299         struct pmd_internals *pmd = dev->data->dev_private;
1300         struct memif_queue *mq;
1301
1302         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1303         if (mq == NULL) {
1304                 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1305                 return -ENOMEM;
1306         }
1307
1308         mq->type =
1309             (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_C2S : MEMIF_RING_S2C;
1310         mq->n_pkts = 0;
1311         mq->n_bytes = 0;
1312         mq->intr_handle.fd = -1;
1313         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1314         mq->in_port = dev->data->port_id;
1315         dev->data->tx_queues[qid] = mq;
1316
1317         return 0;
1318 }
1319
1320 static int
1321 memif_rx_queue_setup(struct rte_eth_dev *dev,
1322                      uint16_t qid,
1323                      uint16_t nb_rx_desc __rte_unused,
1324                      unsigned int socket_id __rte_unused,
1325                      const struct rte_eth_rxconf *rx_conf __rte_unused,
1326                      struct rte_mempool *mb_pool)
1327 {
1328         struct pmd_internals *pmd = dev->data->dev_private;
1329         struct memif_queue *mq;
1330
1331         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1332         if (mq == NULL) {
1333                 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1334                 return -ENOMEM;
1335         }
1336
1337         mq->type = (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_S2C : MEMIF_RING_C2S;
1338         mq->n_pkts = 0;
1339         mq->n_bytes = 0;
1340         mq->intr_handle.fd = -1;
1341         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1342         mq->mempool = mb_pool;
1343         mq->in_port = dev->data->port_id;
1344         dev->data->rx_queues[qid] = mq;
1345
1346         return 0;
1347 }
1348
1349 static void
1350 memif_queue_release(void *queue)
1351 {
1352         struct memif_queue *mq = (struct memif_queue *)queue;
1353
1354         if (!mq)
1355                 return;
1356
1357         rte_free(mq);
1358 }
1359
1360 static int
1361 memif_link_update(struct rte_eth_dev *dev,
1362                   int wait_to_complete __rte_unused)
1363 {
1364         struct pmd_process_private *proc_private;
1365
1366         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1367                 proc_private = dev->process_private;
1368                 if (dev->data->dev_link.link_status == ETH_LINK_UP &&
1369                                 proc_private->regions_num == 0) {
1370                         memif_mp_request_regions(dev);
1371                 } else if (dev->data->dev_link.link_status == ETH_LINK_DOWN &&
1372                                 proc_private->regions_num > 0) {
1373                         memif_free_regions(dev);
1374                 }
1375         }
1376         return 0;
1377 }
1378
1379 static int
1380 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1381 {
1382         struct pmd_internals *pmd = dev->data->dev_private;
1383         struct memif_queue *mq;
1384         int i;
1385         uint8_t tmp, nq;
1386
1387         stats->ipackets = 0;
1388         stats->ibytes = 0;
1389         stats->opackets = 0;
1390         stats->obytes = 0;
1391
1392         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_c2s_rings :
1393             pmd->run.num_s2c_rings;
1394         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1395             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1396
1397         /* RX stats */
1398         for (i = 0; i < nq; i++) {
1399                 mq = dev->data->rx_queues[i];
1400                 stats->q_ipackets[i] = mq->n_pkts;
1401                 stats->q_ibytes[i] = mq->n_bytes;
1402                 stats->ipackets += mq->n_pkts;
1403                 stats->ibytes += mq->n_bytes;
1404         }
1405
1406         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_s2c_rings :
1407             pmd->run.num_c2s_rings;
1408         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1409             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1410
1411         /* TX stats */
1412         for (i = 0; i < nq; i++) {
1413                 mq = dev->data->tx_queues[i];
1414                 stats->q_opackets[i] = mq->n_pkts;
1415                 stats->q_obytes[i] = mq->n_bytes;
1416                 stats->opackets += mq->n_pkts;
1417                 stats->obytes += mq->n_bytes;
1418         }
1419         return 0;
1420 }
1421
1422 static int
1423 memif_stats_reset(struct rte_eth_dev *dev)
1424 {
1425         struct pmd_internals *pmd = dev->data->dev_private;
1426         int i;
1427         struct memif_queue *mq;
1428
1429         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1430                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->tx_queues[i] :
1431                     dev->data->rx_queues[i];
1432                 mq->n_pkts = 0;
1433                 mq->n_bytes = 0;
1434         }
1435         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1436                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->rx_queues[i] :
1437                     dev->data->tx_queues[i];
1438                 mq->n_pkts = 0;
1439                 mq->n_bytes = 0;
1440         }
1441
1442         return 0;
1443 }
1444
1445 static int
1446 memif_rx_queue_intr_enable(struct rte_eth_dev *dev __rte_unused,
1447                            uint16_t qid __rte_unused)
1448 {
1449         MIF_LOG(WARNING, "Interrupt mode not supported.");
1450
1451         return -1;
1452 }
1453
1454 static int
1455 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
1456 {
1457         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
1458
1459         return 0;
1460 }
1461
1462 static const struct eth_dev_ops ops = {
1463         .dev_start = memif_dev_start,
1464         .dev_close = memif_dev_close,
1465         .dev_infos_get = memif_dev_info,
1466         .dev_configure = memif_dev_configure,
1467         .tx_queue_setup = memif_tx_queue_setup,
1468         .rx_queue_setup = memif_rx_queue_setup,
1469         .rx_queue_release = memif_queue_release,
1470         .tx_queue_release = memif_queue_release,
1471         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
1472         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
1473         .link_update = memif_link_update,
1474         .stats_get = memif_stats_get,
1475         .stats_reset = memif_stats_reset,
1476 };
1477
1478 static int
1479 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1480              memif_interface_id_t id, uint32_t flags,
1481              const char *socket_filename,
1482              memif_log2_ring_size_t log2_ring_size,
1483              uint16_t pkt_buffer_size, const char *secret,
1484              struct rte_ether_addr *ether_addr)
1485 {
1486         int ret = 0;
1487         struct rte_eth_dev *eth_dev;
1488         struct rte_eth_dev_data *data;
1489         struct pmd_internals *pmd;
1490         struct pmd_process_private *process_private;
1491         const unsigned int numa_node = vdev->device.numa_node;
1492         const char *name = rte_vdev_device_name(vdev);
1493
1494         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1495         if (eth_dev == NULL) {
1496                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1497                 return -1;
1498         }
1499
1500         process_private = (struct pmd_process_private *)
1501                 rte_zmalloc(name, sizeof(struct pmd_process_private),
1502                             RTE_CACHE_LINE_SIZE);
1503
1504         if (process_private == NULL) {
1505                 MIF_LOG(ERR, "Failed to alloc memory for process private");
1506                 return -1;
1507         }
1508         eth_dev->process_private = process_private;
1509
1510         pmd = eth_dev->data->dev_private;
1511         memset(pmd, 0, sizeof(*pmd));
1512
1513         pmd->id = id;
1514         pmd->flags = flags;
1515         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1516         pmd->role = role;
1517         /* Zero-copy flag irelevant to server. */
1518         if (pmd->role == MEMIF_ROLE_SERVER)
1519                 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1520
1521         ret = memif_socket_init(eth_dev, socket_filename);
1522         if (ret < 0)
1523                 return ret;
1524
1525         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1526         if (secret != NULL)
1527                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1528
1529         pmd->cfg.log2_ring_size = log2_ring_size;
1530         /* set in .dev_configure() */
1531         pmd->cfg.num_c2s_rings = 0;
1532         pmd->cfg.num_s2c_rings = 0;
1533
1534         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1535         rte_spinlock_init(&pmd->cc_lock);
1536
1537         data = eth_dev->data;
1538         data->dev_private = pmd;
1539         data->numa_node = numa_node;
1540         data->dev_link = pmd_link;
1541         data->mac_addrs = ether_addr;
1542         data->promiscuous = 1;
1543         data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1544
1545         eth_dev->dev_ops = &ops;
1546         eth_dev->device = &vdev->device;
1547         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1548                 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1549                 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1550         } else {
1551                 eth_dev->rx_pkt_burst = eth_memif_rx;
1552                 eth_dev->tx_pkt_burst = eth_memif_tx;
1553         }
1554
1555         rte_eth_dev_probing_finish(eth_dev);
1556
1557         return 0;
1558 }
1559
1560 static int
1561 memif_set_role(const char *key __rte_unused, const char *value,
1562                void *extra_args)
1563 {
1564         enum memif_role_t *role = (enum memif_role_t *)extra_args;
1565
1566         if (strstr(value, "server") != NULL) {
1567                 *role = MEMIF_ROLE_SERVER;
1568         } else if (strstr(value, "client") != NULL) {
1569                 *role = MEMIF_ROLE_CLIENT;
1570         } else if (strstr(value, "master") != NULL) {
1571                 MIF_LOG(NOTICE, "Role argument \"master\" is deprecated, use \"server\"");
1572                 *role = MEMIF_ROLE_SERVER;
1573         } else if (strstr(value, "slave") != NULL) {
1574                 MIF_LOG(NOTICE, "Role argument \"slave\" is deprecated, use \"client\"");
1575                 *role = MEMIF_ROLE_CLIENT;
1576         } else {
1577                 MIF_LOG(ERR, "Unknown role: %s.", value);
1578                 return -EINVAL;
1579         }
1580         return 0;
1581 }
1582
1583 static int
1584 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1585 {
1586         uint32_t *flags = (uint32_t *)extra_args;
1587
1588         if (strstr(value, "yes") != NULL) {
1589                 if (!rte_mcfg_get_single_file_segments()) {
1590                         MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1591                         return -ENOTSUP;
1592                 }
1593                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1594         } else if (strstr(value, "no") != NULL) {
1595                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1596         } else {
1597                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1598                 return -EINVAL;
1599         }
1600         return 0;
1601 }
1602
1603 static int
1604 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1605 {
1606         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1607
1608         /* even if parsing fails, 0 is a valid id */
1609         *id = strtoul(value, NULL, 10);
1610         return 0;
1611 }
1612
1613 static int
1614 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1615 {
1616         unsigned long tmp;
1617         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1618
1619         tmp = strtoul(value, NULL, 10);
1620         if (tmp == 0 || tmp > 0xFFFF) {
1621                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1622                 return -EINVAL;
1623         }
1624         *pkt_buffer_size = tmp;
1625         return 0;
1626 }
1627
1628 static int
1629 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1630 {
1631         unsigned long tmp;
1632         memif_log2_ring_size_t *log2_ring_size =
1633             (memif_log2_ring_size_t *)extra_args;
1634
1635         tmp = strtoul(value, NULL, 10);
1636         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1637                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1638                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1639                 return -EINVAL;
1640         }
1641         *log2_ring_size = tmp;
1642         return 0;
1643 }
1644
1645 /* check if directory exists and if we have permission to read/write */
1646 static int
1647 memif_check_socket_filename(const char *filename)
1648 {
1649         char *dir = NULL, *tmp;
1650         uint32_t idx;
1651         int ret = 0;
1652
1653         if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1654                 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1655                 return -1;
1656         }
1657
1658         tmp = strrchr(filename, '/');
1659         if (tmp != NULL) {
1660                 idx = tmp - filename;
1661                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1662                 if (dir == NULL) {
1663                         MIF_LOG(ERR, "Failed to allocate memory.");
1664                         return -1;
1665                 }
1666                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1667         }
1668
1669         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1670                                         W_OK, AT_EACCESS) < 0)) {
1671                 MIF_LOG(ERR, "Invalid socket directory.");
1672                 ret = -EINVAL;
1673         }
1674
1675         if (dir != NULL)
1676                 rte_free(dir);
1677
1678         return ret;
1679 }
1680
1681 static int
1682 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1683                           void *extra_args)
1684 {
1685         const char **socket_filename = (const char **)extra_args;
1686
1687         *socket_filename = value;
1688         return 0;
1689 }
1690
1691 static int
1692 memif_set_is_socket_abstract(const char *key __rte_unused, const char *value, void *extra_args)
1693 {
1694         uint32_t *flags = (uint32_t *)extra_args;
1695
1696         if (strstr(value, "yes") != NULL) {
1697                 *flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1698         } else if (strstr(value, "no") != NULL) {
1699                 *flags &= ~ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1700         } else {
1701                 MIF_LOG(ERR, "Failed to parse socket-abstract param: %s.", value);
1702                 return -EINVAL;
1703         }
1704         return 0;
1705 }
1706
1707 static int
1708 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1709 {
1710         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1711
1712         if (rte_ether_unformat_addr(value, ether_addr) < 0)
1713                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1714         return 0;
1715 }
1716
1717 static int
1718 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1719 {
1720         const char **secret = (const char **)extra_args;
1721
1722         *secret = value;
1723         return 0;
1724 }
1725
1726 static int
1727 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1728 {
1729         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1730         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1731         int ret = 0;
1732         struct rte_kvargs *kvlist;
1733         const char *name = rte_vdev_device_name(vdev);
1734         enum memif_role_t role = MEMIF_ROLE_CLIENT;
1735         memif_interface_id_t id = 0;
1736         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1737         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1738         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1739         uint32_t flags = 0;
1740         const char *secret = NULL;
1741         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1742                 sizeof(struct rte_ether_addr), 0);
1743         struct rte_eth_dev *eth_dev;
1744
1745         rte_eth_random_addr(ether_addr->addr_bytes);
1746
1747         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1748
1749         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1750                 eth_dev = rte_eth_dev_attach_secondary(name);
1751                 if (!eth_dev) {
1752                         MIF_LOG(ERR, "Failed to probe %s", name);
1753                         return -1;
1754                 }
1755
1756                 eth_dev->dev_ops = &ops;
1757                 eth_dev->device = &vdev->device;
1758                 eth_dev->rx_pkt_burst = eth_memif_rx;
1759                 eth_dev->tx_pkt_burst = eth_memif_tx;
1760
1761                 if (!rte_eal_primary_proc_alive(NULL)) {
1762                         MIF_LOG(ERR, "Primary process is missing");
1763                         return -1;
1764                 }
1765
1766                 eth_dev->process_private = (struct pmd_process_private *)
1767                         rte_zmalloc(name,
1768                                 sizeof(struct pmd_process_private),
1769                                 RTE_CACHE_LINE_SIZE);
1770                 if (eth_dev->process_private == NULL) {
1771                         MIF_LOG(ERR,
1772                                 "Failed to alloc memory for process private");
1773                         return -1;
1774                 }
1775
1776                 rte_eth_dev_probing_finish(eth_dev);
1777
1778                 return 0;
1779         }
1780
1781         ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1782         /*
1783          * Primary process can continue probing, but secondary process won't
1784          * be able to get memory regions information
1785          */
1786         if (ret < 0 && rte_errno != EEXIST)
1787                 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1788                         strerror(rte_errno));
1789
1790         /* use abstract address by default */
1791         flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1792
1793         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1794
1795         /* parse parameters */
1796         if (kvlist != NULL) {
1797                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1798                                          &memif_set_role, &role);
1799                 if (ret < 0)
1800                         goto exit;
1801                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1802                                          &memif_set_id, &id);
1803                 if (ret < 0)
1804                         goto exit;
1805                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1806                                          &memif_set_bs, &pkt_buffer_size);
1807                 if (ret < 0)
1808                         goto exit;
1809                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1810                                          &memif_set_rs, &log2_ring_size);
1811                 if (ret < 0)
1812                         goto exit;
1813                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1814                                          &memif_set_socket_filename,
1815                                          (void *)(&socket_filename));
1816                 if (ret < 0)
1817                         goto exit;
1818                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ABSTRACT_ARG,
1819                                          &memif_set_is_socket_abstract, &flags);
1820                 if (ret < 0)
1821                         goto exit;
1822                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1823                                          &memif_set_mac, ether_addr);
1824                 if (ret < 0)
1825                         goto exit;
1826                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1827                                          &memif_set_zc, &flags);
1828                 if (ret < 0)
1829                         goto exit;
1830                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1831                                          &memif_set_secret, (void *)(&secret));
1832                 if (ret < 0)
1833                         goto exit;
1834         }
1835
1836         if (!(flags & ETH_MEMIF_FLAG_SOCKET_ABSTRACT)) {
1837                 ret = memif_check_socket_filename(socket_filename);
1838                 if (ret < 0)
1839                         goto exit;
1840         }
1841
1842         /* create interface */
1843         ret = memif_create(vdev, role, id, flags, socket_filename,
1844                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1845
1846 exit:
1847         if (kvlist != NULL)
1848                 rte_kvargs_free(kvlist);
1849         return ret;
1850 }
1851
1852 static int
1853 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1854 {
1855         struct rte_eth_dev *eth_dev;
1856
1857         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1858         if (eth_dev == NULL)
1859                 return 0;
1860
1861         return rte_eth_dev_close(eth_dev->data->port_id);
1862 }
1863
1864 static struct rte_vdev_driver pmd_memif_drv = {
1865         .probe = rte_pmd_memif_probe,
1866         .remove = rte_pmd_memif_remove,
1867 };
1868
1869 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1870
1871 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1872                               ETH_MEMIF_ID_ARG "=<int>"
1873                               ETH_MEMIF_ROLE_ARG "=server|client"
1874                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1875                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1876                               ETH_MEMIF_SOCKET_ARG "=<string>"
1877                                   ETH_MEMIF_SOCKET_ABSTRACT_ARG "=yes|no"
1878                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1879                               ETH_MEMIF_ZC_ARG "=yes|no"
1880                               ETH_MEMIF_SECRET_ARG "=<string>");
1881
1882 RTE_LOG_REGISTER_DEFAULT(memif_logtype, NOTICE);