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                         rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *,
 387                                                            dst_off),
 388                                 (uint8_t *)memif_get_buffer(proc_private, d0) +
 389                                 src_off, cp_len);
 390
 391                         src_off += cp_len;
 392                         dst_off += cp_len;
 393                         src_len -= cp_len;
 394                 } while (src_len);
 395
 396                 cur_slot++;
 397                 n_slots--;
 398
 399                 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
 400                         goto next_slot;
 401
 402                 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
 403                 *bufs++ = mbuf_head;
 404                 n_rx_pkts++;
 405         }
 406
 407 no_free_bufs:
 408         if (type == MEMIF_RING_C2S) {
 409                 __atomic_store_n(&ring->tail, cur_slot, __ATOMIC_RELEASE);
 410                 mq->last_head = cur_slot;
 411         } else {
 412                 mq->last_tail = cur_slot;
 413         }
 414
 415 refill:
 416         if (type == MEMIF_RING_S2C) {
 417                 /* ring->head is updated by the receiver and this function
 418                  * is called in the context of receiver thread. The loads in
 419                  * the receiver do not need to synchronize with its own stores.
 420                  */
 421                 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 422                 n_slots = ring_size - head + mq->last_tail;
 423
 424                 while (n_slots--) {
 425                         s0 = head++ & mask;
 426                         d0 = &ring->desc[s0];
 427                         d0->length = pmd->run.pkt_buffer_size;
 428                 }
 429                 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 430         }
 431
 432         mq->n_pkts += n_rx_pkts;
 433         return n_rx_pkts;
 434 }
 435
 436 static uint16_t
 437 eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 438 {
 439         struct memif_queue *mq = queue;
 440         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 441         struct pmd_process_private *proc_private =
 442                 rte_eth_devices[mq->in_port].process_private;
 443         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 444         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0, head;
 445         uint16_t n_rx_pkts = 0;
 446         memif_desc_t *d0;
 447         struct rte_mbuf *mbuf, *mbuf_tail;
 448         struct rte_mbuf *mbuf_head = NULL;
 449         int ret;
 450         struct rte_eth_link link;
 451
 452         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 453                 return 0;
 454         if (unlikely(ring == NULL)) {
 455                 /* Secondary process will attempt to request regions. */
 456                 rte_eth_link_get(mq->in_port, &link);
 457                 return 0;
 458         }
 459
 460         /* consume interrupt */
 461         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 462                 uint64_t b;
 463                 ssize_t size __rte_unused;
 464                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 465         }
 466
 467         ring_size = 1 << mq->log2_ring_size;
 468         mask = ring_size - 1;
 469
 470         cur_slot = mq->last_tail;
 471         /* The ring->tail acts as a guard variable between Tx and Rx
 472          * threads, so using load-acquire pairs with store-release
 473          * to synchronize it between threads.
 474          */
 475         last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 476         if (cur_slot == last_slot)
 477                 goto refill;
 478         n_slots = last_slot - cur_slot;
 479
 480         while (n_slots && n_rx_pkts < nb_pkts) {
 481                 s0 = cur_slot & mask;
 482
 483                 d0 = &ring->desc[s0];
 484                 mbuf_head = mq->buffers[s0];
 485                 mbuf = mbuf_head;
 486
 487 next_slot:
 488                 /* prefetch next descriptor */
 489                 if (n_rx_pkts + 1 < nb_pkts)
 490                         rte_prefetch0(&ring->desc[(cur_slot + 1) & mask]);
 491
 492                 mbuf->port = mq->in_port;
 493                 rte_pktmbuf_data_len(mbuf) = d0->length;
 494                 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 495
 496                 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 497
 498                 cur_slot++;
 499                 n_slots--;
 500                 if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
 501                         s0 = cur_slot & mask;
 502                         d0 = &ring->desc[s0];
 503                         mbuf_tail = mbuf;
 504                         mbuf = mq->buffers[s0];
 505                         ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 506                         if (unlikely(ret < 0)) {
 507                                 MIF_LOG(ERR, "number-of-segments-overflow");
 508                                 goto refill;
 509                         }
 510                         goto next_slot;
 511                 }
 512
 513                 *bufs++ = mbuf_head;
 514                 n_rx_pkts++;
 515         }
 516
 517         mq->last_tail = cur_slot;
 518
 519 /* Supply server with new buffers */
 520 refill:
 521         /* ring->head is updated by the receiver and this function
 522          * is called in the context of receiver thread. The loads in
 523          * the receiver do not need to synchronize with its own stores.
 524          */
 525         head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 526         n_slots = ring_size - head + mq->last_tail;
 527
 528         if (n_slots < 32)
 529                 goto no_free_mbufs;
 530
 531         ret = rte_pktmbuf_alloc_bulk(mq->mempool, &mq->buffers[head & mask], n_slots);
 532         if (unlikely(ret < 0))
 533                 goto no_free_mbufs;
 534
 535         while (n_slots--) {
 536                 s0 = head++ & mask;
 537                 if (n_slots > 0)
 538                         rte_prefetch0(mq->buffers[head & mask]);
 539                 d0 = &ring->desc[s0];
 540                 /* store buffer header */
 541                 mbuf = mq->buffers[s0];
 542                 /* populate descriptor */
 543                 d0->length = rte_pktmbuf_data_room_size(mq->mempool) -
 544                                 RTE_PKTMBUF_HEADROOM;
 545                 d0->region = 1;
 546                 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 547                         (uint8_t *)proc_private->regions[d0->region]->addr;
 548         }
 549 no_free_mbufs:
 550         /* The ring->head acts as a guard variable between Tx and Rx
 551          * threads, so using store-release pairs with load-acquire
 552          * in function eth_memif_tx.
 553          */
 554         __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 555
 556         mq->n_pkts += n_rx_pkts;
 557
 558         return n_rx_pkts;
 559 }
 560
 561 static uint16_t
 562 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 563 {
 564         struct memif_queue *mq = queue;
 565         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 566         struct pmd_process_private *proc_private =
 567                 rte_eth_devices[mq->in_port].process_private;
 568         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 569         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 570         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 571         memif_ring_type_t type = mq->type;
 572         memif_desc_t *d0;
 573         struct rte_mbuf *mbuf;
 574         struct rte_mbuf *mbuf_head;
 575         uint64_t a;
 576         ssize_t size;
 577         struct rte_eth_link link;
 578
 579         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 580                 return 0;
 581         if (unlikely(ring == NULL)) {
 582                 int ret;
 583
 584                 /* Secondary process will attempt to request regions. */
 585                 ret = rte_eth_link_get(mq->in_port, &link);
 586                 if (ret < 0)
 587                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 588                                 mq->in_port, rte_strerror(-ret));
 589                 return 0;
 590         }
 591
 592         ring_size = 1 << mq->log2_ring_size;
 593         mask = ring_size - 1;
 594
 595         if (type == MEMIF_RING_C2S) {
 596                 /* For C2S queues ring->head is updated by the sender and
 597                  * this function is called in the context of sending thread.
 598                  * The loads in the sender do not need to synchronize with
 599                  * its own stores. Hence, the following load can be a
 600                  * relaxed load.
 601                  */
 602                 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 603                 n_free = ring_size - slot +
 604                                 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 605         } else {
 606                 /* For S2C queues ring->tail is updated by the sender and
 607                  * this function is called in the context of sending thread.
 608                  * The loads in the sender do not need to synchronize with
 609                  * its own stores. Hence, the following load can be a
 610                  * relaxed load.
 611                  */
 612                 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 613                 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 614         }
 615
 616         while (n_tx_pkts < nb_pkts && n_free) {
 617                 mbuf_head = *bufs++;
 618                 mbuf = mbuf_head;
 619
 620                 saved_slot = slot;
 621                 d0 = &ring->desc[slot & mask];
 622                 dst_off = 0;
 623                 dst_len = (type == MEMIF_RING_C2S) ?
 624                         pmd->run.pkt_buffer_size : d0->length;
 625
 626 next_in_chain:
 627                 src_off = 0;
 628                 src_len = rte_pktmbuf_data_len(mbuf);
 629
 630                 while (src_len) {
 631                         if (dst_len == 0) {
 632                                 if (n_free) {
 633                                         slot++;
 634                                         n_free--;
 635                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 636                                         d0 = &ring->desc[slot & mask];
 637                                         dst_off = 0;
 638                                         dst_len = (type == MEMIF_RING_C2S) ?
 639                                             pmd->run.pkt_buffer_size : d0->length;
 640                                         d0->flags = 0;
 641                                 } else {
 642                                         slot = saved_slot;
 643                                         goto no_free_slots;
 644                                 }
 645                         }
 646                         cp_len = RTE_MIN(dst_len, src_len);
 647
 648                         rte_memcpy((uint8_t *)memif_get_buffer(proc_private,
 649                                                                d0) + dst_off,
 650                                 rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 651                                 cp_len);
 652
 653                         mq->n_bytes += cp_len;
 654                         src_off += cp_len;
 655                         dst_off += cp_len;
 656                         src_len -= cp_len;
 657                         dst_len -= cp_len;
 658
 659                         d0->length = dst_off;
 660                 }
 661
 662                 if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 663                         mbuf = mbuf->next;
 664                         goto next_in_chain;
 665                 }
 666
 667                 n_tx_pkts++;
 668                 slot++;
 669                 n_free--;
 670                 rte_pktmbuf_free(mbuf_head);
 671         }
 672
 673 no_free_slots:
 674         if (type == MEMIF_RING_C2S)
 675                 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 676         else
 677                 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
 678
 679         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 680                 a = 1;
 681                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 682                 if (unlikely(size < 0)) {
 683                         MIF_LOG(WARNING,
 684                                 "Failed to send interrupt. %s", strerror(errno));
 685                 }
 686         }
 687
 688         mq->n_pkts += n_tx_pkts;
 689         return n_tx_pkts;
 690 }
 691
 692
 693 static int
 694 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
 695                 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
 696                 uint16_t slot, uint16_t n_free)
 697 {
 698         memif_desc_t *d0;
 699         int used_slots = 1;
 700
 701 next_in_chain:
 702         /* store pointer to mbuf to free it later */
 703         mq->buffers[slot & mask] = mbuf;
 704         /* Increment refcnt to make sure the buffer is not freed before server
 705          * receives it. (current segment)
 706          */
 707         rte_mbuf_refcnt_update(mbuf, 1);
 708         /* populate descriptor */
 709         d0 = &ring->desc[slot & mask];
 710         d0->length = rte_pktmbuf_data_len(mbuf);
 711         mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 712         /* FIXME: get region index */
 713         d0->region = 1;
 714         d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 715                 (uint8_t *)proc_private->regions[d0->region]->addr;
 716         d0->flags = 0;
 717
 718         /* check if buffer is chained */
 719         if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 720                 if (n_free < 2)
 721                         return 0;
 722                 /* mark buffer as chained */
 723                 d0->flags |= MEMIF_DESC_FLAG_NEXT;
 724                 /* advance mbuf */
 725                 mbuf = mbuf->next;
 726                 /* update counters */
 727                 used_slots++;
 728                 slot++;
 729                 n_free--;
 730                 goto next_in_chain;
 731         }
 732         return used_slots;
 733 }
 734
 735 static uint16_t
 736 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 737 {
 738         struct memif_queue *mq = queue;
 739         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 740         struct pmd_process_private *proc_private =
 741                 rte_eth_devices[mq->in_port].process_private;
 742         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 743         uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
 744         struct rte_eth_link link;
 745
 746         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 747                 return 0;
 748         if (unlikely(ring == NULL)) {
 749                 /* Secondary process will attempt to request regions. */
 750                 rte_eth_link_get(mq->in_port, &link);
 751                 return 0;
 752         }
 753
 754         ring_size = 1 << mq->log2_ring_size;
 755         mask = ring_size - 1;
 756
 757         /* free mbufs received by server */
 758         memif_free_stored_mbufs(proc_private, mq);
 759
 760         /* ring type always MEMIF_RING_C2S */
 761         /* For C2S queues ring->head is updated by the sender and
 762          * this function is called in the context of sending thread.
 763          * The loads in the sender do not need to synchronize with
 764          * its own stores. Hence, the following load can be a
 765          * relaxed load.
 766          */
 767         slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 768         n_free = ring_size - slot + mq->last_tail;
 769
 770         int used_slots;
 771
 772         while (n_free && (n_tx_pkts < nb_pkts)) {
 773                 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
 774                         if ((nb_pkts - n_tx_pkts) > 8) {
 775                                 rte_prefetch0(*bufs + 4);
 776                                 rte_prefetch0(*bufs + 5);
 777                                 rte_prefetch0(*bufs + 6);
 778                                 rte_prefetch0(*bufs + 7);
 779                         }
 780                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 781                                 mask, slot, n_free);
 782                         if (unlikely(used_slots < 1))
 783                                 goto no_free_slots;
 784                         n_tx_pkts++;
 785                         slot += used_slots;
 786                         n_free -= used_slots;
 787
 788                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 789                                 mask, slot, n_free);
 790                         if (unlikely(used_slots < 1))
 791                                 goto no_free_slots;
 792                         n_tx_pkts++;
 793                         slot += used_slots;
 794                         n_free -= used_slots;
 795
 796                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 797                                 mask, slot, n_free);
 798                         if (unlikely(used_slots < 1))
 799                                 goto no_free_slots;
 800                         n_tx_pkts++;
 801                         slot += used_slots;
 802                         n_free -= used_slots;
 803
 804                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 805                                 mask, slot, n_free);
 806                         if (unlikely(used_slots < 1))
 807                                 goto no_free_slots;
 808                         n_tx_pkts++;
 809                         slot += used_slots;
 810                         n_free -= used_slots;
 811                 }
 812                 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 813                         mask, slot, n_free);
 814                 if (unlikely(used_slots < 1))
 815                         goto no_free_slots;
 816                 n_tx_pkts++;
 817                 slot += used_slots;
 818                 n_free -= used_slots;
 819         }
 820
 821 no_free_slots:
 822         /* ring type always MEMIF_RING_C2S */
 823         /* The ring->head acts as a guard variable between Tx and Rx
 824          * threads, so using store-release pairs with load-acquire
 825          * in function eth_memif_rx for C2S rings.
 826          */
 827         __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 828
 829         /* Send interrupt, if enabled. */
 830         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 831                 uint64_t a = 1;
 832                 ssize_t size = write(mq->intr_handle.fd, &a, sizeof(a));
 833                 if (unlikely(size < 0)) {
 834                         MIF_LOG(WARNING,
 835                                 "Failed to send interrupt. %s", strerror(errno));
 836                 }
 837         }
 838
 839         /* increment queue counters */
 840         mq->n_pkts += n_tx_pkts;
 841
 842         return n_tx_pkts;
 843 }
 844
 845 void
 846 memif_free_regions(struct rte_eth_dev *dev)
 847 {
 848         struct pmd_process_private *proc_private = dev->process_private;
 849         struct pmd_internals *pmd = dev->data->dev_private;
 850         int i;
 851         struct memif_region *r;
 852
 853         /* regions are allocated contiguously, so it's
 854          * enough to loop until 'proc_private->regions_num'
 855          */
 856         for (i = 0; i < proc_private->regions_num; i++) {
 857                 r = proc_private->regions[i];
 858                 if (r != NULL) {
 859                         /* This is memzone */
 860                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
 861                                 r->addr = NULL;
 862                                 if (r->fd > 0)
 863                                         close(r->fd);
 864                         }
 865                         if (r->addr != NULL) {
 866                                 munmap(r->addr, r->region_size);
 867                                 if (r->fd > 0) {
 868                                         close(r->fd);
 869                                         r->fd = -1;
 870                                 }
 871                         }
 872                         rte_free(r);
 873                         proc_private->regions[i] = NULL;
 874                 }
 875         }
 876         proc_private->regions_num = 0;
 877 }
 878
 879 static int
 880 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 881                      void *arg)
 882 {
 883         struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
 884         struct memif_region *r;
 885
 886         if (proc_private->regions_num < 1) {
 887                 MIF_LOG(ERR, "Missing descriptor region");
 888                 return -1;
 889         }
 890
 891         r = proc_private->regions[proc_private->regions_num - 1];
 892
 893         if (r->addr != msl->base_va)
 894                 r = proc_private->regions[++proc_private->regions_num - 1];
 895
 896         if (r == NULL) {
 897                 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 898                 if (r == NULL) {
 899                         MIF_LOG(ERR, "Failed to alloc memif region.");
 900                         return -ENOMEM;
 901                 }
 902
 903                 r->addr = msl->base_va;
 904                 r->region_size = ms->len;
 905                 r->fd = rte_memseg_get_fd(ms);
 906                 if (r->fd < 0)
 907                         return -1;
 908                 r->pkt_buffer_offset = 0;
 909
 910                 proc_private->regions[proc_private->regions_num - 1] = r;
 911         } else {
 912                 r->region_size += ms->len;
 913         }
 914
 915         return 0;
 916 }
 917
 918 static int
 919 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
 920 {
 921         struct pmd_internals *pmd = dev->data->dev_private;
 922         struct pmd_process_private *proc_private = dev->process_private;
 923         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 924         int ret = 0;
 925         struct memif_region *r;
 926
 927         if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 928                 MIF_LOG(ERR, "Too many regions.");
 929                 return -1;
 930         }
 931
 932         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 933         if (r == NULL) {
 934                 MIF_LOG(ERR, "Failed to alloc memif region.");
 935                 return -ENOMEM;
 936         }
 937
 938         /* calculate buffer offset */
 939         r->pkt_buffer_offset = (pmd->run.num_c2s_rings + pmd->run.num_s2c_rings) *
 940             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 941             (1 << pmd->run.log2_ring_size));
 942
 943         r->region_size = r->pkt_buffer_offset;
 944         /* if region has buffers, add buffers size to region_size */
 945         if (has_buffers == 1)
 946                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 947                         (1 << pmd->run.log2_ring_size) *
 948                         (pmd->run.num_c2s_rings +
 949                          pmd->run.num_s2c_rings));
 950
 951         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 952         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 953                  proc_private->regions_num);
 954
 955         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 956         if (r->fd < 0) {
 957                 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
 958                 ret = -1;
 959                 goto error;
 960         }
 961
 962         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 963         if (ret < 0) {
 964                 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
 965                 goto error;
 966         }
 967
 968         ret = ftruncate(r->fd, r->region_size);
 969         if (ret < 0) {
 970                 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
 971                 goto error;
 972         }
 973
 974         r->addr = mmap(NULL, r->region_size, PROT_READ |
 975                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 976         if (r->addr == MAP_FAILED) {
 977                 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
 978                 ret = -1;
 979                 goto error;
 980         }
 981
 982         proc_private->regions[proc_private->regions_num] = r;
 983         proc_private->regions_num++;
 984
 985         return ret;
 986
 987 error:
 988         if (r->fd > 0)
 989                 close(r->fd);
 990         r->fd = -1;
 991
 992         return ret;
 993 }
 994
 995 static int
 996 memif_regions_init(struct rte_eth_dev *dev)
 997 {
 998         struct pmd_internals *pmd = dev->data->dev_private;
 999         int ret;
1000
1001         /*
1002          * Zero-copy exposes dpdk memory.
1003          * Each memseg list will be represented by memif region.
1004          * Zero-copy regions indexing: memseg list idx + 1,
1005          * as we already have region 0 reserved for descriptors.
1006          */
1007         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1008                 /* create region idx 0 containing descriptors */
1009                 ret = memif_region_init_shm(dev, 0);
1010                 if (ret < 0)
1011                         return ret;
1012                 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1013                 if (ret < 0)
1014                         return ret;
1015         } else {
1016                 /* create one memory region contaning rings and buffers */
1017                 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1018                 if (ret < 0)
1019                         return ret;
1020         }
1021
1022         return 0;
1023 }
1024
1025 static void
1026 memif_init_rings(struct rte_eth_dev *dev)
1027 {
1028         struct pmd_internals *pmd = dev->data->dev_private;
1029         struct pmd_process_private *proc_private = dev->process_private;
1030         memif_ring_t *ring;
1031         int i, j;
1032         uint16_t slot;
1033
1034         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1035                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_C2S, i);
1036                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1037                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1038                 ring->cookie = MEMIF_COOKIE;
1039                 ring->flags = 0;
1040
1041                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1042                         continue;
1043
1044                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1045                         slot = i * (1 << pmd->run.log2_ring_size) + j;
1046                         ring->desc[j].region = 0;
1047                         ring->desc[j].offset =
1048                                 proc_private->regions[0]->pkt_buffer_offset +
1049                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1050                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1051                 }
1052         }
1053
1054         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1055                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2C, i);
1056                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1057                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1058                 ring->cookie = MEMIF_COOKIE;
1059                 ring->flags = 0;
1060
1061                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1062                         continue;
1063
1064                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1065                         slot = (i + pmd->run.num_c2s_rings) *
1066                             (1 << pmd->run.log2_ring_size) + j;
1067                         ring->desc[j].region = 0;
1068                         ring->desc[j].offset =
1069                                 proc_private->regions[0]->pkt_buffer_offset +
1070                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1071                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1072                 }
1073         }
1074 }
1075
1076 /* called only by client */
1077 static int
1078 memif_init_queues(struct rte_eth_dev *dev)
1079 {
1080         struct pmd_internals *pmd = dev->data->dev_private;
1081         struct memif_queue *mq;
1082         int i;
1083
1084         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1085                 mq = dev->data->tx_queues[i];
1086                 mq->log2_ring_size = pmd->run.log2_ring_size;
1087                 /* queues located only in region 0 */
1088                 mq->region = 0;
1089                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_C2S, i);
1090                 mq->last_head = 0;
1091                 mq->last_tail = 0;
1092                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1093                 if (mq->intr_handle.fd < 0) {
1094                         MIF_LOG(WARNING,
1095                                 "Failed to create eventfd for tx queue %d: %s.", i,
1096                                 strerror(errno));
1097                 }
1098                 mq->buffers = NULL;
1099                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1100                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1101                                                   (1 << mq->log2_ring_size), 0);
1102                         if (mq->buffers == NULL)
1103                                 return -ENOMEM;
1104                 }
1105         }
1106
1107         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1108                 mq = dev->data->rx_queues[i];
1109                 mq->log2_ring_size = pmd->run.log2_ring_size;
1110                 /* queues located only in region 0 */
1111                 mq->region = 0;
1112                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2C, i);
1113                 mq->last_head = 0;
1114                 mq->last_tail = 0;
1115                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1116                 if (mq->intr_handle.fd < 0) {
1117                         MIF_LOG(WARNING,
1118                                 "Failed to create eventfd for rx queue %d: %s.", i,
1119                                 strerror(errno));
1120                 }
1121                 mq->buffers = NULL;
1122                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1123                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1124                                                   (1 << mq->log2_ring_size), 0);
1125                         if (mq->buffers == NULL)
1126                                 return -ENOMEM;
1127                 }
1128         }
1129         return 0;
1130 }
1131
1132 int
1133 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1134 {
1135         int ret;
1136
1137         ret = memif_regions_init(dev);
1138         if (ret < 0)
1139                 return ret;
1140
1141         memif_init_rings(dev);
1142
1143         ret = memif_init_queues(dev);
1144         if (ret < 0)
1145                 return ret;
1146
1147         return 0;
1148 }
1149
1150 int
1151 memif_connect(struct rte_eth_dev *dev)
1152 {
1153         struct pmd_internals *pmd = dev->data->dev_private;
1154         struct pmd_process_private *proc_private = dev->process_private;
1155         struct memif_region *mr;
1156         struct memif_queue *mq;
1157         memif_ring_t *ring;
1158         int i;
1159
1160         for (i = 0; i < proc_private->regions_num; i++) {
1161                 mr = proc_private->regions[i];
1162                 if (mr != NULL) {
1163                         if (mr->addr == NULL) {
1164                                 if (mr->fd < 0)
1165                                         return -1;
1166                                 mr->addr = mmap(NULL, mr->region_size,
1167                                                 PROT_READ | PROT_WRITE,
1168                                                 MAP_SHARED, mr->fd, 0);
1169                                 if (mr->addr == MAP_FAILED) {
1170                                         MIF_LOG(ERR, "mmap failed: %s\n",
1171                                                 strerror(errno));
1172                                         return -1;
1173                                 }
1174                         }
1175                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1176                                 /* close memseg file */
1177                                 close(mr->fd);
1178                                 mr->fd = -1;
1179                         }
1180                 }
1181         }
1182
1183         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1184                 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1185                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1186                             dev->data->tx_queues[i] : dev->data->rx_queues[i];
1187                         ring = memif_get_ring_from_queue(proc_private, mq);
1188                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1189                                 MIF_LOG(ERR, "Wrong ring");
1190                                 return -1;
1191                         }
1192                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1193                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1194                         mq->last_head = 0;
1195                         mq->last_tail = 0;
1196                         /* enable polling mode */
1197                         if (pmd->role == MEMIF_ROLE_SERVER)
1198                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1199                 }
1200                 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1201                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1202                             dev->data->rx_queues[i] : dev->data->tx_queues[i];
1203                         ring = memif_get_ring_from_queue(proc_private, mq);
1204                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1205                                 MIF_LOG(ERR, "Wrong ring");
1206                                 return -1;
1207                         }
1208                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1209                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1210                         mq->last_head = 0;
1211                         mq->last_tail = 0;
1212                         /* enable polling mode */
1213                         if (pmd->role == MEMIF_ROLE_CLIENT)
1214                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1215                 }
1216
1217                 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1218                 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1219                 dev->data->dev_link.link_status = ETH_LINK_UP;
1220         }
1221         MIF_LOG(INFO, "Connected.");
1222         return 0;
1223 }
1224
1225 static int
1226 memif_dev_start(struct rte_eth_dev *dev)
1227 {
1228         struct pmd_internals *pmd = dev->data->dev_private;
1229         int ret = 0;
1230
1231         switch (pmd->role) {
1232         case MEMIF_ROLE_CLIENT:
1233                 ret = memif_connect_client(dev);
1234                 break;
1235         case MEMIF_ROLE_SERVER:
1236                 ret = memif_connect_server(dev);
1237                 break;
1238         default:
1239                 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1240                 ret = -1;
1241                 break;
1242         }
1243
1244         return ret;
1245 }
1246
1247 static int
1248 memif_dev_close(struct rte_eth_dev *dev)
1249 {
1250         struct pmd_internals *pmd = dev->data->dev_private;
1251         int i;
1252
1253         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1254                 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1255                 memif_disconnect(dev);
1256
1257                 for (i = 0; i < dev->data->nb_rx_queues; i++)
1258                         (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
1259                 for (i = 0; i < dev->data->nb_tx_queues; i++)
1260                         (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
1261
1262                 memif_socket_remove_device(dev);
1263         } else {
1264                 memif_disconnect(dev);
1265         }
1266
1267         rte_free(dev->process_private);
1268
1269         return 0;
1270 }
1271
1272 static int
1273 memif_dev_configure(struct rte_eth_dev *dev)
1274 {
1275         struct pmd_internals *pmd = dev->data->dev_private;
1276
1277         /*
1278          * CLIENT - TXQ
1279          * SERVER - RXQ
1280          */
1281         pmd->cfg.num_c2s_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1282                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1283
1284         /*
1285          * CLIENT - RXQ
1286          * SERVER - TXQ
1287          */
1288         pmd->cfg.num_s2c_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1289                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1290
1291         return 0;
1292 }
1293
1294 static int
1295 memif_tx_queue_setup(struct rte_eth_dev *dev,
1296                      uint16_t qid,
1297                      uint16_t nb_tx_desc __rte_unused,
1298                      unsigned int socket_id __rte_unused,
1299                      const struct rte_eth_txconf *tx_conf __rte_unused)
1300 {
1301         struct pmd_internals *pmd = dev->data->dev_private;
1302         struct memif_queue *mq;
1303
1304         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1305         if (mq == NULL) {
1306                 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1307                 return -ENOMEM;
1308         }
1309
1310         mq->type =
1311             (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_C2S : MEMIF_RING_S2C;
1312         mq->n_pkts = 0;
1313         mq->n_bytes = 0;
1314         mq->intr_handle.fd = -1;
1315         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1316         mq->in_port = dev->data->port_id;
1317         dev->data->tx_queues[qid] = mq;
1318
1319         return 0;
1320 }
1321
1322 static int
1323 memif_rx_queue_setup(struct rte_eth_dev *dev,
1324                      uint16_t qid,
1325                      uint16_t nb_rx_desc __rte_unused,
1326                      unsigned int socket_id __rte_unused,
1327                      const struct rte_eth_rxconf *rx_conf __rte_unused,
1328                      struct rte_mempool *mb_pool)
1329 {
1330         struct pmd_internals *pmd = dev->data->dev_private;
1331         struct memif_queue *mq;
1332
1333         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1334         if (mq == NULL) {
1335                 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1336                 return -ENOMEM;
1337         }
1338
1339         mq->type = (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_S2C : MEMIF_RING_C2S;
1340         mq->n_pkts = 0;
1341         mq->n_bytes = 0;
1342         mq->intr_handle.fd = -1;
1343         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1344         mq->mempool = mb_pool;
1345         mq->in_port = dev->data->port_id;
1346         dev->data->rx_queues[qid] = mq;
1347
1348         return 0;
1349 }
1350
1351 static void
1352 memif_queue_release(void *queue)
1353 {
1354         struct memif_queue *mq = (struct memif_queue *)queue;
1355
1356         if (!mq)
1357                 return;
1358
1359         rte_free(mq);
1360 }
1361
1362 static int
1363 memif_link_update(struct rte_eth_dev *dev,
1364                   int wait_to_complete __rte_unused)
1365 {
1366         struct pmd_process_private *proc_private;
1367
1368         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1369                 proc_private = dev->process_private;
1370                 if (dev->data->dev_link.link_status == ETH_LINK_UP &&
1371                                 proc_private->regions_num == 0) {
1372                         memif_mp_request_regions(dev);
1373                 } else if (dev->data->dev_link.link_status == ETH_LINK_DOWN &&
1374                                 proc_private->regions_num > 0) {
1375                         memif_free_regions(dev);
1376                 }
1377         }
1378         return 0;
1379 }
1380
1381 static int
1382 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1383 {
1384         struct pmd_internals *pmd = dev->data->dev_private;
1385         struct memif_queue *mq;
1386         int i;
1387         uint8_t tmp, nq;
1388
1389         stats->ipackets = 0;
1390         stats->ibytes = 0;
1391         stats->opackets = 0;
1392         stats->obytes = 0;
1393
1394         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_c2s_rings :
1395             pmd->run.num_s2c_rings;
1396         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1397             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1398
1399         /* RX stats */
1400         for (i = 0; i < nq; i++) {
1401                 mq = dev->data->rx_queues[i];
1402                 stats->q_ipackets[i] = mq->n_pkts;
1403                 stats->q_ibytes[i] = mq->n_bytes;
1404                 stats->ipackets += mq->n_pkts;
1405                 stats->ibytes += mq->n_bytes;
1406         }
1407
1408         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_s2c_rings :
1409             pmd->run.num_c2s_rings;
1410         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1411             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1412
1413         /* TX stats */
1414         for (i = 0; i < nq; i++) {
1415                 mq = dev->data->tx_queues[i];
1416                 stats->q_opackets[i] = mq->n_pkts;
1417                 stats->q_obytes[i] = mq->n_bytes;
1418                 stats->opackets += mq->n_pkts;
1419                 stats->obytes += mq->n_bytes;
1420         }
1421         return 0;
1422 }
1423
1424 static int
1425 memif_stats_reset(struct rte_eth_dev *dev)
1426 {
1427         struct pmd_internals *pmd = dev->data->dev_private;
1428         int i;
1429         struct memif_queue *mq;
1430
1431         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1432                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->tx_queues[i] :
1433                     dev->data->rx_queues[i];
1434                 mq->n_pkts = 0;
1435                 mq->n_bytes = 0;
1436         }
1437         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1438                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->rx_queues[i] :
1439                     dev->data->tx_queues[i];
1440                 mq->n_pkts = 0;
1441                 mq->n_bytes = 0;
1442         }
1443
1444         return 0;
1445 }
1446
1447 static int
1448 memif_rx_queue_intr_enable(struct rte_eth_dev *dev __rte_unused,
1449                            uint16_t qid __rte_unused)
1450 {
1451         MIF_LOG(WARNING, "Interrupt mode not supported.");
1452
1453         return -1;
1454 }
1455
1456 static int
1457 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
1458 {
1459         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
1460
1461         return 0;
1462 }
1463
1464 static const struct eth_dev_ops ops = {
1465         .dev_start = memif_dev_start,
1466         .dev_close = memif_dev_close,
1467         .dev_infos_get = memif_dev_info,
1468         .dev_configure = memif_dev_configure,
1469         .tx_queue_setup = memif_tx_queue_setup,
1470         .rx_queue_setup = memif_rx_queue_setup,
1471         .rx_queue_release = memif_queue_release,
1472         .tx_queue_release = memif_queue_release,
1473         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
1474         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
1475         .link_update = memif_link_update,
1476         .stats_get = memif_stats_get,
1477         .stats_reset = memif_stats_reset,
1478 };
1479
1480 static int
1481 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1482              memif_interface_id_t id, uint32_t flags,
1483              const char *socket_filename,
1484              memif_log2_ring_size_t log2_ring_size,
1485              uint16_t pkt_buffer_size, const char *secret,
1486              struct rte_ether_addr *ether_addr)
1487 {
1488         int ret = 0;
1489         struct rte_eth_dev *eth_dev;
1490         struct rte_eth_dev_data *data;
1491         struct pmd_internals *pmd;
1492         struct pmd_process_private *process_private;
1493         const unsigned int numa_node = vdev->device.numa_node;
1494         const char *name = rte_vdev_device_name(vdev);
1495
1496         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1497         if (eth_dev == NULL) {
1498                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1499                 return -1;
1500         }
1501
1502         process_private = (struct pmd_process_private *)
1503                 rte_zmalloc(name, sizeof(struct pmd_process_private),
1504                             RTE_CACHE_LINE_SIZE);
1505
1506         if (process_private == NULL) {
1507                 MIF_LOG(ERR, "Failed to alloc memory for process private");
1508                 return -1;
1509         }
1510         eth_dev->process_private = process_private;
1511
1512         pmd = eth_dev->data->dev_private;
1513         memset(pmd, 0, sizeof(*pmd));
1514
1515         pmd->id = id;
1516         pmd->flags = flags;
1517         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1518         pmd->role = role;
1519         /* Zero-copy flag irelevant to server. */
1520         if (pmd->role == MEMIF_ROLE_SERVER)
1521                 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1522
1523         ret = memif_socket_init(eth_dev, socket_filename);
1524         if (ret < 0)
1525                 return ret;
1526
1527         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1528         if (secret != NULL)
1529                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1530
1531         pmd->cfg.log2_ring_size = log2_ring_size;
1532         /* set in .dev_configure() */
1533         pmd->cfg.num_c2s_rings = 0;
1534         pmd->cfg.num_s2c_rings = 0;
1535
1536         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1537         rte_spinlock_init(&pmd->cc_lock);
1538
1539         data = eth_dev->data;
1540         data->dev_private = pmd;
1541         data->numa_node = numa_node;
1542         data->dev_link = pmd_link;
1543         data->mac_addrs = ether_addr;
1544         data->promiscuous = 1;
1545         data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1546
1547         eth_dev->dev_ops = &ops;
1548         eth_dev->device = &vdev->device;
1549         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1550                 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1551                 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1552         } else {
1553                 eth_dev->rx_pkt_burst = eth_memif_rx;
1554                 eth_dev->tx_pkt_burst = eth_memif_tx;
1555         }
1556
1557         rte_eth_dev_probing_finish(eth_dev);
1558
1559         return 0;
1560 }
1561
1562 static int
1563 memif_set_role(const char *key __rte_unused, const char *value,
1564                void *extra_args)
1565 {
1566         enum memif_role_t *role = (enum memif_role_t *)extra_args;
1567
1568         if (strstr(value, "server") != NULL) {
1569                 *role = MEMIF_ROLE_SERVER;
1570         } else if (strstr(value, "client") != NULL) {
1571                 *role = MEMIF_ROLE_CLIENT;
1572         } else if (strstr(value, "master") != NULL) {
1573                 MIF_LOG(NOTICE, "Role argument \"master\" is deprecated, use \"server\"");
1574                 *role = MEMIF_ROLE_SERVER;
1575         } else if (strstr(value, "slave") != NULL) {
1576                 MIF_LOG(NOTICE, "Role argument \"slave\" is deprecated, use \"client\"");
1577                 *role = MEMIF_ROLE_CLIENT;
1578         } else {
1579                 MIF_LOG(ERR, "Unknown role: %s.", value);
1580                 return -EINVAL;
1581         }
1582         return 0;
1583 }
1584
1585 static int
1586 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1587 {
1588         uint32_t *flags = (uint32_t *)extra_args;
1589
1590         if (strstr(value, "yes") != NULL) {
1591                 if (!rte_mcfg_get_single_file_segments()) {
1592                         MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1593                         return -ENOTSUP;
1594                 }
1595                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1596         } else if (strstr(value, "no") != NULL) {
1597                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1598         } else {
1599                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1600                 return -EINVAL;
1601         }
1602         return 0;
1603 }
1604
1605 static int
1606 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1607 {
1608         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1609
1610         /* even if parsing fails, 0 is a valid id */
1611         *id = strtoul(value, NULL, 10);
1612         return 0;
1613 }
1614
1615 static int
1616 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1617 {
1618         unsigned long tmp;
1619         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1620
1621         tmp = strtoul(value, NULL, 10);
1622         if (tmp == 0 || tmp > 0xFFFF) {
1623                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1624                 return -EINVAL;
1625         }
1626         *pkt_buffer_size = tmp;
1627         return 0;
1628 }
1629
1630 static int
1631 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1632 {
1633         unsigned long tmp;
1634         memif_log2_ring_size_t *log2_ring_size =
1635             (memif_log2_ring_size_t *)extra_args;
1636
1637         tmp = strtoul(value, NULL, 10);
1638         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1639                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1640                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1641                 return -EINVAL;
1642         }
1643         *log2_ring_size = tmp;
1644         return 0;
1645 }
1646
1647 /* check if directory exists and if we have permission to read/write */
1648 static int
1649 memif_check_socket_filename(const char *filename)
1650 {
1651         char *dir = NULL, *tmp;
1652         uint32_t idx;
1653         int ret = 0;
1654
1655         if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1656                 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1657                 return -1;
1658         }
1659
1660         tmp = strrchr(filename, '/');
1661         if (tmp != NULL) {
1662                 idx = tmp - filename;
1663                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1664                 if (dir == NULL) {
1665                         MIF_LOG(ERR, "Failed to allocate memory.");
1666                         return -1;
1667                 }
1668                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1669         }
1670
1671         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1672                                         W_OK, AT_EACCESS) < 0)) {
1673                 MIF_LOG(ERR, "Invalid socket directory.");
1674                 ret = -EINVAL;
1675         }
1676
1677         if (dir != NULL)
1678                 rte_free(dir);
1679
1680         return ret;
1681 }
1682
1683 static int
1684 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1685                           void *extra_args)
1686 {
1687         const char **socket_filename = (const char **)extra_args;
1688
1689         *socket_filename = value;
1690         return 0;
1691 }
1692
1693 static int
1694 memif_set_is_socket_abstract(const char *key __rte_unused, const char *value, void *extra_args)
1695 {
1696         uint32_t *flags = (uint32_t *)extra_args;
1697
1698         if (strstr(value, "yes") != NULL) {
1699                 *flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1700         } else if (strstr(value, "no") != NULL) {
1701                 *flags &= ~ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1702         } else {
1703                 MIF_LOG(ERR, "Failed to parse socket-abstract param: %s.", value);
1704                 return -EINVAL;
1705         }
1706         return 0;
1707 }
1708
1709 static int
1710 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1711 {
1712         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1713
1714         if (rte_ether_unformat_addr(value, ether_addr) < 0)
1715                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1716         return 0;
1717 }
1718
1719 static int
1720 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1721 {
1722         const char **secret = (const char **)extra_args;
1723
1724         *secret = value;
1725         return 0;
1726 }
1727
1728 static int
1729 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1730 {
1731         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1732         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1733         int ret = 0;
1734         struct rte_kvargs *kvlist;
1735         const char *name = rte_vdev_device_name(vdev);
1736         enum memif_role_t role = MEMIF_ROLE_CLIENT;
1737         memif_interface_id_t id = 0;
1738         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1739         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1740         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1741         uint32_t flags = 0;
1742         const char *secret = NULL;
1743         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1744                 sizeof(struct rte_ether_addr), 0);
1745         struct rte_eth_dev *eth_dev;
1746
1747         rte_eth_random_addr(ether_addr->addr_bytes);
1748
1749         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1750
1751         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1752                 eth_dev = rte_eth_dev_attach_secondary(name);
1753                 if (!eth_dev) {
1754                         MIF_LOG(ERR, "Failed to probe %s", name);
1755                         return -1;
1756                 }
1757
1758                 eth_dev->dev_ops = &ops;
1759                 eth_dev->device = &vdev->device;
1760                 eth_dev->rx_pkt_burst = eth_memif_rx;
1761                 eth_dev->tx_pkt_burst = eth_memif_tx;
1762
1763                 if (!rte_eal_primary_proc_alive(NULL)) {
1764                         MIF_LOG(ERR, "Primary process is missing");
1765                         return -1;
1766                 }
1767
1768                 eth_dev->process_private = (struct pmd_process_private *)
1769                         rte_zmalloc(name,
1770                                 sizeof(struct pmd_process_private),
1771                                 RTE_CACHE_LINE_SIZE);
1772                 if (eth_dev->process_private == NULL) {
1773                         MIF_LOG(ERR,
1774                                 "Failed to alloc memory for process private");
1775                         return -1;
1776                 }
1777
1778                 rte_eth_dev_probing_finish(eth_dev);
1779
1780                 return 0;
1781         }
1782
1783         ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1784         /*
1785          * Primary process can continue probing, but secondary process won't
1786          * be able to get memory regions information
1787          */
1788         if (ret < 0 && rte_errno != EEXIST)
1789                 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1790                         strerror(rte_errno));
1791
1792         /* use abstract address by default */
1793         flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1794
1795         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1796
1797         /* parse parameters */
1798         if (kvlist != NULL) {
1799                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1800                                          &memif_set_role, &role);
1801                 if (ret < 0)
1802                         goto exit;
1803                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1804                                          &memif_set_id, &id);
1805                 if (ret < 0)
1806                         goto exit;
1807                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1808                                          &memif_set_bs, &pkt_buffer_size);
1809                 if (ret < 0)
1810                         goto exit;
1811                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1812                                          &memif_set_rs, &log2_ring_size);
1813                 if (ret < 0)
1814                         goto exit;
1815                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1816                                          &memif_set_socket_filename,
1817                                          (void *)(&socket_filename));
1818                 if (ret < 0)
1819                         goto exit;
1820                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ABSTRACT_ARG,
1821                                          &memif_set_is_socket_abstract, &flags);
1822                 if (ret < 0)
1823                         goto exit;
1824                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1825                                          &memif_set_mac, ether_addr);
1826                 if (ret < 0)
1827                         goto exit;
1828                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1829                                          &memif_set_zc, &flags);
1830                 if (ret < 0)
1831                         goto exit;
1832                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1833                                          &memif_set_secret, (void *)(&secret));
1834                 if (ret < 0)
1835                         goto exit;
1836         }
1837
1838         if (!(flags & ETH_MEMIF_FLAG_SOCKET_ABSTRACT)) {
1839                 ret = memif_check_socket_filename(socket_filename);
1840                 if (ret < 0)
1841                         goto exit;
1842         }
1843
1844         /* create interface */
1845         ret = memif_create(vdev, role, id, flags, socket_filename,
1846                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1847
1848 exit:
1849         if (kvlist != NULL)
1850                 rte_kvargs_free(kvlist);
1851         return ret;
1852 }
1853
1854 static int
1855 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1856 {
1857         struct rte_eth_dev *eth_dev;
1858
1859         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1860         if (eth_dev == NULL)
1861                 return 0;
1862
1863         return rte_eth_dev_close(eth_dev->data->port_id);
1864 }
1865
1866 static struct rte_vdev_driver pmd_memif_drv = {
1867         .probe = rte_pmd_memif_probe,
1868         .remove = rte_pmd_memif_remove,
1869 };
1870
1871 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1872
1873 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1874                               ETH_MEMIF_ID_ARG "=<int>"
1875                               ETH_MEMIF_ROLE_ARG "=server|client"
1876                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1877                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1878                               ETH_MEMIF_SOCKET_ARG "=<string>"
1879                                   ETH_MEMIF_SOCKET_ABSTRACT_ARG "=yes|no"
1880                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1881                               ETH_MEMIF_ZC_ARG "=yes|no"
1882                               ETH_MEMIF_SECRET_ARG "=<string>");
1883
1884 RTE_LOG_REGISTER_DEFAULT(memif_logtype, NOTICE);